text
stringlengths 2.85k
2.55M
| label
class label 11
classes |
|---|---|
Programmable Restoration Granularity in
Constraint Programming
Yong Lin and Martin Henz
arXiv:1601.06517v2 [cs.PL] 4 Feb 2016
School of Computing, National University of Singapore, Singapore
{linyong, henz}@comp.nus.edu.sg
Abstract. In most constraint programming systems, a limited number
of search engines is offered while the programming of user-customized
search algorithms requires low-level efforts, which complicates the deployment of such algorithms. To alleviate this limitation, concepts such
as computation spaces have been developed. Computation spaces provide
a coarse-grained restoration mechanism, because they store all information contained in a search tree node. Other granularities are possible, and
in this paper we make the case for dynamically adapting the restoration
granularity during search. In order to elucidate programmable restoration granularity, we present restoration as an aspect of a constraint programming system, using the model of aspect-oriented programming. A
proof-of-concept implementation using Gecode shows promising results.
1
Introduction
Constraint Programming (CP) solves combinatorial problems through constraintbased search, which explores a problem-defined search tree in a certain order.
The exploring of search tree may encounter a failure, which signals a false search
direction. To recover search from such a failure, restoration service is called to
provide an ever visited state to switch search direction.
In CP systems, restoration has been implemented in several manners. Mozart/Oz
system copies each ever accessed state for direct retrieval, in a format of computation space (space for short); Constraint Logic Programming (CLP) systems
such as CHIP [3], clp(FD) [2] and ECLi PSe [1] record state change information
in a trail data strucutre to roll back; Gecode [10] system maintains generated
constraints to recompute [8]; our newly proposed recollection [6] stores constraint
propagation affected variable domains to recollect.
For all above implemented techniques, they differ on the granularity of the
stored information. Copying is coarse-grained since a computation space encapsulates all speculative computation structures; on the other hand, recomputation does not store any specific data structures, but the instructions committed
for guiding search. Both trailing and recollection are finer-grained than space
copying since they store only part information of a state. Actually, the granularity of a restored technique determines its advantages and limitations. If one
can program to synthetically use these techniques in a system, more efficient
2
Yong Lin and Martin Henz
restoration is expected; an early effort was conducted by the proposal of computation space, which enables user to place state copies in the search tree as they
intend. However, as we have claimed, space is coarse-grained, while recomputation is only aware of committed constraints. Therefore, we advocate that it is
a non-trivial study to include other finer-grained techniques to program a more
powerful restoration scheme.
In this paper, we propose to program restoration granularities, aiming at
improving constraint programming system performance. The initial effort has
been conducted by integrating coarse-grained copying, finer-grained recollection
and constraint-aware recomputation. Especially, we propose to program restoration granularity in aspect-oriented programming for a more modular and flexible
restoration implementation. The rest of this paper is organized as follows: Section 2 briefly goes through the notions in CP systen implementation; in Section 3,
we present a prototype of programming restoration granularity ; subsequently,
in Section 4, we employ the concept aspect to illustrate the landscape of more
modular and flexible restoration implementation, taking Gecode as an example;
last, we conclude in Section 5.
2
Basics
In CP systems, constraint propagation is demonstrated insufficient to solve a
problem when it reaches a fix point. Search then comes into play; it branches
on the fix point to generate constraints and then commits one of generated constraints to proceed. This interleaving of constraint propagation and branching
creates a search tree, where each node is a computation state. In such a search
tree, branches are constraints, internal nodes are fix points and leaf nodes are
either solutions or search failures. A search failure indicates a false search direction, which requires the restoration of a previously visited internal state to
switch search directions.
In constraint programming systems, state restoration can be implemented in
several ways. Trailing-based method stores how one states was modified to reach
other conjunct state in a trail structure to roll back previous performed operations [4]; this method is efficient for the problems imposing weak propagation [7].
However, trailing couples tightly with propagation engine and search facilities,
which potentially limits parallel exploration. Copying approach duplicates each
ever accessed fix point state and restoration is a simple direct state retrieval. This
technique causes neither runtime nor memory issues for small and medium size
problem. As for large problems, copying may introduce memory management
penalty on account of its substantially occupied memory [7]. Recomputation
barely maintains previously committed constraints so that it can recompute a
state. This approach consumes rather limited memory, but its runtime may be
dragged significantly if a problem is computation expensive. Our proposed recollection memorizes the variables that were changed in fix point reasoning steps
at a moderate extra memory investment; it conducts restoration by updating
a higher level state, using the stored variables. For all these restoration tech-
Programmable Restoration Granularity in Constraint Programming
3
niques, we can observe that their ways to conduct restoration are determined
by the granularity of information they have stored during search. Specifically,
copying is coarse-grained since it stores all information of states; by contrast,
trailing and recollection are finer-grained. As for recomputation, it merely stores
the meta-information (constraints) that was used to instruct search directions.
3
A Prototype
In this section, we would present a proof-of-concept that programs restoration
granularity. The organization of this section is as follows: Section 3.1 highlights the traits of a few problem, which spurs our investigation in programming
restoration granularity and we describe the prototype in Section 3.2; evaluation
is conducted in Section 3.3.
3.1
Motivation
A deep search tree may be expanded for solving a hard problem. Table 1 illustrates the search tree statistics of four problems exploring for the first solution,
where the Queens problem is modeled by either a set of disequality constraints
or three global constraints of the “all-different constraints” family (denoted as
Queens-S). In this table, the column failures counts the total number of failures
during search; first signals the tree level where the first search failure emerges,
while peak is the value of the deepest tree level. [1, first ) accumulates the number
of failures occurs between the root and the first level, and [first, peak ] records
the number of failures between first and peak.
Problem
failures first peak [1, first) [first, peak ]
Queens(200)
146,838 164
Queens-S(200)
146,838 164
Knights(22)
19,877 386
Sport-League(22) 1,035
62
200
200
451
249
0
0
0
5
146,838
146,838
19,877
1,030
Table 1. Search Tree Statistics of Problem Search Trees
Although any of restoration techniques is capable of accomplishing its duty
correctly, they respectively exhibit advantage towards certain case, as we have
claimed. From these search tree statistics, we perceive that the emergence of
the first failure can be an important signal for more intensive search failures.
Therefore, if copying is employed as an underlying restoration to support the
solving of these problems, then the space copies maintained between the root
f irst − 1 level cannot contribute at all. This indicates that an ideal restoration
should customize problem search traits.
4
Yong Lin and Martin Henz
3.2
Programming Granularity
To deal with the skewed failure distribution as displayed in Table 1, a straightforward method is storing restoration information of difference granularity as
search proceeds, and restoration then adapt the stored information to conduct
relevant process. We initially program a prototype by integrating copying, recomputation and recollection. In this prototype, we take the search tree level
where the first failure emerges as a border level, which horizontally divides the
search tree into upper and bottom two parts. For search in the upper part of the
search tree, only generated constraints are kept, while neither coarse-grained
state copies nor finer-grained propagation modified variables is stored. When
search explores beyond the border level, both copying and recollection will be
activated to collaborate: the placing of a coarse-grained state copy alternates n
finer-grained recollection explorations (n is configurable and set to eithgt ). Because of this change of information granularity, restoration now need to switch
between different code segment. Specifically, when a state in upper part search
tree is expected, the restoration routine switches to recomputation. When a state
in the bottom part is required, the restoration routine first attempts to recollect
from the nearest state copy; if this effort fails, it then recomputes to restore
the border state first and then update this border state to the requested one
by recollection. We can take an optimization measure by maintaining a border
state copy to avoid the possible recomputation when restoring a bottom state.
Similarly, one can also employ other measures to optimize the recomputation in
upper part.
3.3
Evaluation
We evaluate our programmed prototype over the four problems, where we sought
the motivation to program restoration granularity. Both recomputation and recollection enable adaptive function and set copying distance to eight. The evaluations were conducted on a Intel Core 2 Quad processor PC system, running
an Ubuntu operating system 11.10 with four Gigabyte main memory. We built
our sample on top of Gecode version 3.7.3, which also served as the reference
instance of comparisons. Each collected runtime1 value is an arithmetic mean
of 20 runs with a variation coefficient less than 2%; memory numbers are peak
memory consumption.
We compare the prototype with both adaptive recomputation and adaptive
recollection, and Table 2 depicts the evaluations results. These numbers reveal
that our prototype can significantly save memory than the other two restoration
alternatives; but for Sport-League problem, the memory saving is not as significant, which can be on account of it first failure comes earlier (at level 62 of a tree
with peak depth 249). Meanwhile, Queens problems expose better runtime performances by adaptive recomputation. Nevertheless, it deserves to highlight that
1
We take wall clock time.
Programmable Restoration Granularity in Constraint Programming
Problems
Queens(200)
Queens-S(200)
Knights(22)
Sport-League(22)
5
Recomputation
Recollection
Prototype
Time(ms) Mem(KB) Time(ms) Mem(KB) Time(ms) Mem(KB)
4,330
2,156
1,858
352
25,748
1,485
4,460
7,710
4,578
2,473
1,704
331
28,238
3,974
4,592
7,937
4,601
2,469
1,744
339
6,244
542
2,333
6,109
Table 2. Comparisons of Prototype with Recomputation and Recollection
Knights problem almost halves memory consumption than the other two techniques, while marginally improves its runtime than recomputation. This promising result confirms the opportunities of programming restoration granularity to
seek better performance.
4
Programmability
As mentioned in previous section, the way to implement restoration is actually determined by the granularity of storing information for restoration. This
information is implemented to store by search engine at exploration steps; on
the other hand, restoration is a service provided for search engine. This observation claims that the maintained restoration information cuts across the two
abstractions: search engine and restoration.
In developing applications, the occurrence of crosscutting abstraction is not
rare; transactions, security-related operation, logging etc all exemplify crosscutting abstractions. To facilitate the programming of crosscutting abstraction,
aspect-oriented programming [5](AOP) proposes to encapsulate a crosscutting
abstraction as an aspect. The implementation of an aspect mainly consists of
two tasks: advice and pointcut. An advice is a means of specifying the code to
run at a joint point, where additional behaviors attempt to add in the program;
a pointcut determines the matches of executing specified advice at a joint point.
To program restoration granularity, one should correctly implement the switching between restoration technique code segments. In our prototype, we achieve
the communication between code segments by introducing a signal. However,
this signal couples tightly with a specific program; suppose one redefines the
criteria to switch restoration techniques, the code is quite likely to change, probably drastically. Therefore, it is of great significance to implement a more flexible
and modular design to program restoration granularity, and we claim that programming restoration as an aspect can achieve this design goal.
In this section, we propose to adopt aspect-oriented programming to program
restoration granularity in Gecode system for a modular and flexible implementation. Specifically, Section 4.1 briefly explains the design of search facilities
in Gecode system; In Section 4.2, we discuss the issues related with deploying
aspect in Gecode search facilities.
6
Yong Lin and Martin Henz
4.1
Search Facilities
In the Gecode system, it defines a class of Edge to keep the restoration information of each visited fixpoint; all Edge objects will be pushed into a stack
structure Path by search engine. Program 1 outlines the class declaration of
Edge. In the Edge, the Choice object encapsulates fix point generated two constraints2 , which are respectively represented by 0 and 1; the currently committed
constraint alternative is recorded by the integer variable alternative. Choice object is a compulsory information component in each created Edge object since
it instructs search direction by committing constraint, whereas the space copy
space is optional. Specifically, if search engine wraps a space in each Edge,
then it is a copying-based restoration; if none of Edge stores space copy, it is
a recomputation-based restoration; hybrid variants can be obtained by placing
space copies in Edges occasionally such as adaptive recomputation. In Gecode,
all restoration related services are defined on the structure path, which collects
the Edge objects generated between root and current node.
Program 1 Class Edge
class Edge
{
Space * space;
/* Space copy */
Choice * choice;
/* fixpoint generated Choice */
/* committed choice alternative */
unsigned int altternative;
vector¡int¿ doms;
/* variable domains */
public:
Edge(Space * s, Choice * c, vector¡int¿ doms):
{
alternative = 0;
/* commit to the first alternative */
...
/* other initialization statements */
}
...
}
As we have claimed, space is coarse-grained data, while constraints are metainformation for instructing search. We introduce a finer-grained data structure
dom, which collects the variable domains that were affected in the fix point
reasoning3. This vector contains sufficient information to conduct recollection.
One can also introduce other data structures to further explore other restoration
possibilities.
In Gecode, search engine is constructed by programming on computation
spaces. The space provides an interface status() to trigger internal constraint
propagation, which returns a status value; search engine then switches to the
relevant code segment according to the space status, as depicted in Program2.
If the space return a constraint propagation results of failure, the search engine
2
3
We restrict our discussion on binary search tree
We insist on Finite Domain Integer Domain problems in this discussion
Programmable Restoration Granularity in Constraint Programming
7
first requests to adjust search direction by calling Adjust()4 , and then enter the
code segment where restoration is defined5 .
Program 2 Search Engine
while true do
switch (Status(space))
/* query space status */
case fixpoint :
Choice ch←Choice(space)
/* return solution space */
Push(Edge(ch . . .))
Commit(space, ch)
/* commit a constraint to space */
case solution:
return space
/* return solution space */
case failure:
if not Adjust(. . .) then
break
/* The problem is not solvable */
end if
Restore(. . .) {
Pointcut1
Recomputation code
/* originally programmed recomputation */
Pointcut2 }
end switch
end while
4.2
Restoration as an Aspect
To program in aspect-oriented programming paradigm, it requires support from
the underlying programming language compiler/system. Currently, quite a few
programming languages have implemented AOP, within the language, or as an
external library; AspectC++ [9] has provided a set of extensions to facilitate the
use of AOP in C++. Gecode is an open source constraint programming library
constructed in C++; thus, a following exploration is provisioning restoration
service of Gecode with the aspect-oriented programming paradigm.
To deploy restoration as an aspect, it requires to address the definition of
advice and pointcut. Specifically, we define advice to exploit the information
of a specific granularity so that it can individually, or collaboratively with the
recomputation code, restore ever visited fix point; this collaboration requires
the pointcuts (advice can be matched and applied for execution) either before
and after the orignal recomputation code segment, as illustrated in Program 2.
In our prototype, we place the code of copying and recollection at the place
marked by pointcut2 ; but as we have stressed, we currently use a specific signal
to coordinate the code switching. Thus, we conjecture that a more modular and
flexible restoration can be built by using AOP.
4
5
It will return a Boolean false if another search direction is impossible
Restore() method is not actually defined in search engine, we expose it body in
search engine for facilitating explanation
8
5
Yong Lin and Martin Henz
Conclusion
Restoration is a key component of search facilities in constraint programming
systems; its implementation can significantly impact on the performance and
construction of a search engine. Although the proposal of computation space
facilitates the programming of restoration, it is coarse-grained since a space
encapsulates all data structures of a state in the search tree.
In this paper, we have proposed to program restoration granularity. We implemented such a prototype by integrates coarse-grained copying, finer-grained
recollection and constraint-based recomputation; its evaluation gives promising results. In the prototype, we employed a specific signal to coordinate the
execution of between restoration code segments, which couples tightly with a
program. Thus, we have further propose to provision restoration with aspectoriented programming. The significance of this prospective effort is, on the one
hand, modularizing restoration code segments that implement different techniques while flexing their assemble. On the other hand, it also potentially provides more provisions to construct search engines that run a wide spectrum of
search algorithms.
References
1. Abderrahamane Aggoun, David Chan, Pierre Dufresne, Eamon Falvey, Hugh
Grant, Warwick Harvey, Alexander Herold, Geoffrey Macartney, Micha Meier,
David Miller, et al. Eclipse user manual, 1997.
2. Philippe Codognet and Daniel Diaz. Compiling constraints in clp (fd). The Journal
of Logic Programming, 27(3):185–226, 1996.
3. Mehmet Dincbas, Pascal Van Hentenryck, Helmut Simonis, Abderrahmane Aggoun, Thomas Graf, and Françoise Berthier. The constraint logic programming
language chip. In Proceeding of the International Conference on Fifth Generation
Computer Science FGCS-88, pages 693–702, 1988.
4. Joxan Jaffar and Michael J. Maher. Constraint logic programming: a survey. The
Journal of Logic Programming, pages 503 – 581, 1994. Special Issue: Ten Years of
Logic Programming.
5. Gregor Kiczales, John Lamping, Anurag Mendhekar, Chris Maeda, Cristina Lopes,
Jean-Marc Loingtier, and John Irwin. Aspect-oriented programming. Springer,
1997.
6. Yong Lin.
Prototypical implementation of recollection based on Gecode.
www.comp.nus.edu.sg/~ henz/recollection, May 2013.
7. Christian Schulte. Comparing trailing and copying for constraint programming.
In Proceedings of the 1999 International Conference on Logic Programming, pages
275–289, Cambridge, MA, USA, 1999. Massachusetts Institute of Technology.
8. Christian Schulte. Programming Constraint Services, volume 2302 of Lecture Notes
in Artificial Intelligence. Springer-Verlag, 2002.
9. AspectC++ Project Team. The Homepage of AspectC++. www.aspectc.org,
2012.
10. Gecode Project Team.
Generic COnstraint Development Environment.
www.gecode.org, June 2012.
| 6cs.PL
|
Exploring Spatial Context for 3D Semantic Segmentation of Point Clouds
Francis Engelmann† , Theodora Kontogianni† , Alexander Hermans and Bastian Leibe
Computer Vision Group, Visual Computing Institute
RWTH Aachen University
arXiv:1802.01500v1 [cs.CV] 5 Feb 2018
{engelmann,kontogianni,hermans,leibe}@vision.rwth-aachen.de
Abstract
Deep learning approaches have made tremendous
progress in the field of semantic segmentation over the past
few years. However, most current approaches operate in
the 2D image space. Direct semantic segmentation of unstructured 3D point clouds is still an open research problem. The recently proposed PointNet architecture presents
an interesting step ahead in that it can operate on unstructured point clouds, achieving encouraging segmentation results. However, it subdivides the input points into a grid of
blocks and processes each such block individually. In this
paper, we investigate the question how such an architecture
can be extended to incorporate larger-scale spatial context.
We build upon PointNet and propose two extensions that
enlarge the receptive field over the 3D scene. We evaluate
the proposed strategies on challenging indoor and outdoor
datasets and show improved results in both scenarios.
Input-Level Context
Figure 1: We explore mechanisms to extend the spatial context for 3D semantic segmentation of point clouds.
can be obtained from visual SLAM approaches operating on
the vehicle’s cameras [13]. Finding approaches that can directly operate on point cloud data is highly desirable, since
it avoids costly preprocessing and format conversion steps.
However, the question what is the best network architecture
to process unstructured 3D point clouds is still largely open.
In this paper, we take inspiration from the recent PointNet work by Qi et al. [25], which currently defines the state
of the art in 3D semantic segmentation. PointNet learns a
higher dimensional spatial feature representation for each
3D point and then aggregates all the points within a small
3D volume (typically an occupancy grid cell) in order to
bring in some form of 3D neighborhood context. However,
this neighborhood context is very restricted, as each grid
cell is processed independently.
In this paper, we investigate possible mechanisms to incorporate context into a point cloud processing architecture.
We focus on spatial context, which has been identified as
being very important for semantic segmentation [19, 24].
We introduce two mechanisms to add spatial context to an
existing PointNet. The first mechanism incorporates neighborhood information by processing input data from multiple scales or multiple adjacent regions together (inputlevel context). The second mechanism operates on the estimated point descriptors and aims at consolidating them by
exchanging information over a larger spatial neighborhood
1. Introduction
Semantic segmentation is an important capability for
intelligent vehicles, such as autonomous cars or mobile
robots. Identifying the semantic meaning of the observed
3D structure around the vehicle is a prerequisite for solving
subsequent tasks such as navigation or reconstruction [5, 6].
Consequently, the problem has attracted a lot of attention,
and notable successes have been achieved with the help of
deep learning techniques. However, most state-of-the-art
semantic segmentation approaches operate on 2D images,
which naturally lend themselves to processing with Convolutional Neural Networks (CNNs) [16, 3, 28, 22].
Processing unstructured 3D point clouds, such as those
obtained from LiDAR or stereo sensors, is a much harder
problem, and it is only recently that first successful deep
learning approaches have been proposed for this task [18,
12, 31, 25]. Such point clouds can be obtained from LiDAR sensors mounted on top of a recording vehicle or they
† Both
Output-Level Context
authors contributed equally.
1
2. Related Work
Unstructured Point Clouds. A varied number of sensors and setups exist which help to obtain unstructured point
clouds: areal data from airborne laser scanners, laser scanners mounted on dynamic setups in a push-broom configuration [17], rotating lasers e.g. Velodyne [8], or static lasers
[9]. Additionally, indoor spaces can be scanned using devices such as the Microsoft Kinect [21] or Matterport cameras [1]. All these devices produce point clouds of different
quality and density. We apply our method to indoor data
from [1] and to synthetic urban outdoor data from [7].
Traditional Methods. Hackel et al. [10] use traditional
random forest classifiers with 3D features (without color).
Their method is based on eigenvalues and eigenvectors of
covariance tensors created by the nearest neighbors of the
points. Their main contribution is an efficient approximate
nearest neighbors computation at different scales. Munoz
et al. [20] follow a similar approach but replace the random
forest classifier with an associative Markov network. Random forest classifiers are also used in [32] to classify data
from 2D images and 3D point clouds, which they later fuse.
Similarly, Xu et al. [30] fuse camera and LiDAR sensor
data. Xiong et al. [29] propose a sequential parsing procedure that learns the spatial relationships of objects. Lai et
al. [15] introduce a hierarchical sparse coding technique for
learning features from synthetic data. Vosselman et al. [27]
combine multiple segmentation and post-processing methods to achieve useful point cloud segmentations.
Deep-learning Methods. In a deep learning context,
point clouds can be represented in a regular volumetric grid
in order to apply 3D convolutions [18, 12]. Alternatively,
3D points can be mapped to a 2D representation followed
by 2D convolutions [26]. In [2], the authors are performing 2D convolutions in 2D snapshots of a 3D point cloud
and then project the labels back to 3D space. In [23] a deep
learning framework learns semantic segmentation by tracking point clouds. Yi et al. [31] use spectral CNNs on 3D
models represented as shape graphs for shape part segmentation. Recent methods operate directly on raw point clouds
with KD-trees [14] or fully convolutional layers [25].
MLP
Point
Features
M
1 x D’
Global
Feature
S
C
MLP
NxM
N x D’
NxD
(output-level context). For both mechanisms, we explore
several possible realizations and compare them experimentally. As our results show, both mechanisms improve semantic segmentation quality.
Contributions. The key contributions of our work can
be summarized as follows: (1) We present two mechanisms
that can be used to incorporate spatial context into semantic 3D point cloud segmentation. (2) We show how these
mechanisms can be incorporated into the PointNet pipeline.
(3) We verify experimentally that our proposed extensions
achieve improved results on challenging indoor and outdoor
datasets.
N x D+D’
Figure 2: Simplified PointNet Architecture. In this work,
we build upon the PointNet architecture for semantic segmentation. In short, it computes a global feature which
summarizes a set of input points. Specifically, the network
takes N points as input, applies a series of multi-layerperceptrons transformations and aggregates the point features by max pooling them into a global feature. Global
and local features are concatenated and the per point class
scores are returned. (MLP): Multi-Layer-Perception, (M):
Max-Pool, (S): Vertical Stack, (C): Concatenate. See text
and Qi et al. [25] for more details.
3. Method
In this section we start by reviewing the PointNet model,
then we introduce our mechanisms of extending context and
finish by describing our two exemplary architectures.
3.1. PointNet
PointNet [25] is a deep neural network that, when used
for semantic segmentation, takes as input a point cloud and
outputs the per point semantic class labels. First, it splits
a point cloud into 3D blocks, then it takes N points inside a block and after a series of Multi-Layer-Perceptrons
(MLP) per point, the points are mapped into a higher dimensional space D0 , these are called local point-features.
Max-pooling is applied to aggregate information from all
the points resulting in a common global-feature invariant to
input permutations. The global-feature is then concatenated
with all the point-features. After another series of MLPs
these combined features are used to predict the M output
class scores. Figure 2 shows a simplified model.
Caveats. The global-features in PointNet summarize
the context of a single block (block-feature), as a result the
aggregated information is passed only among points inside
the same block.
Context outside a block is equally important and could help
make more informed class label predictions. Therefore
we introduce two mechanisms to add context: input-level
context – which operates directly on the input point clouds
– and output-level context – which consolidates the output
from the input-level context.
3.2. Input-Level Context
In this straightforward addition, we increase the context of the network by considering a group of blocks simultaneously instead of one individual block at a time as
done in PointNet. Context is shared among all blocks in
a group. These groups of blocks are selected either from
NxD
Input-Level Context
MLP (64,128)
Output Score
Multi-Scale
Block Feature
M
NxD
MLP (64,128)
M
NxD
MLP (64,128)
M
C
S
C
O=256
O=128
MLP (M)
:block feature
Consolidation Unit (CU)
NxO
Multi-Scale Blocks
same position
different scales
NxM
N x 128+384
1 x 384
MLP (O)
N x 2zO
1xO
M
S
C
M
: max pool
S
: stack
C
: concatenate
Figure 3: Architecture with multi-scale input blocks and consolidation units (MS-CU). The network takes as input three
blocks from multiple scales, each one containing N D-dimensional points. Separately, for each scale, it learns a blockfeature similarly to the PointNet mechanism. The concatenated block-features are appended to the input-features and then
transformed by a sequence of consolidation units (see Section 3.3). The network outputs per point scores. Shaded fields
represent block-features.
the same position but at multiple different scales (MultiScale Blocks, see Figure 3, left) or from neighboring cells
in a regular grid (Grid Blocks, see Figure 4, left). For each
input block, we compute a block-feature using the mechanism from PointNet. For the multi-scale version, we train a
block-descriptor for each scale individually to obtain scaledependent block-features. In the case of grid blocks, all
block features are computed by a shared single-scale blockdescriptor. In the end, both approaches output a set of
block-features corresponding to the input blocks.
3.3. Output-Level Context
At this stage, we further consolidate the block-features
obtained from the previous stage. Here, we differ between
two consolidation approaches:
Consolidation Units (CU) consume a set of point features,
transform them into a higher dimensional space using
MLPs and apply max-pooling to generate a common
block-feature which is again concatenated with each of the
high dimensional input features (see Figure 3, blue box).
This procedure is similar to the block-feature mechanism
of PointNet. The key point is that CUs can be chained together into a sequence CUs forming a deeper network. The
intuition behind this setup is as follows: In the beginning
each point sees only its own features. After appending the
block-features, each point is additionally informed about
the features of its neighboring points. By applying CUs
multiple times, this shared knowledge is reinforced.
Recurrent Consolidation Units (RCU) are the second type of context consolidation we employ. RCUs
take as input a sequence of block-features originating
from spatially nearby blocks and return a sequence of
corresponding updated block-features. The core idea
is to create block-features that take into consideration
neighboring blocks as well. In more detail, RCUs are
implemented as RNNs, specifically GRUs [4], which are
a simpler variation of standard LSTMs [11]. GRUs have
the capability to learn long range dependencies. That range
can either be over time (as in speech recognition) or over
space as in our case. The cells of the unrolled GRU are
connected in an unsynchronized many-to-many fashion
(see Figure 4, blue box). This means that the updated
block-features are returned only after the GRU has seen the
whole input sequence of block-features. Intuitively, GRU
retain relevant information about the scene in their internal
memory and update it according to new observations.
We use this memory mechanism to consolidate and share
the information across all input blocks. For example,
the decision about whether a point belongs to a chair is
changed if the network remembers that it has seen a table
further down in the room.
In the following, we describe two exemplary architectures which combine the previously introduced components. For those, we provide a detailed evaluation and report
improved results in Section 4.
3.4. Multi-Scale (MS) Architecture
The full MS architecture is displayed in Figure 3. The
learned block-features from the multi-scale blocks, (see
Section 3.2) are concatenated into one multi-scale blockfeature. This multi-scale block-feature is further concatenated with the transformed input point-features and passed
through a series of CUs (see Section 3.3). Applying a final
MLP results in output scores for each input point.
Specific for this architecture is the sampling procedure to
select the positions of the multi-scale blocks: We randomly
pick a D-dimensional point from the input point cloud as
the center of the blocks and we group together N randomly
selected points that fall within a specified radius. This procedure is repeated at the same point for multiple radii.
Output Score
Input-Level Context
NxM
M
shared
…
M
MLP (128,64)
shared
…
MLP (128,64)
M
shared
…
MLP (128,64)
M
NxD
NxD
MLP (128,64)
NxD
NxD
S
Grid Blocks
different positions
same scale
1 x 64
1 x 64
1 x 64
1 x 64
1 x 64
GRU-RNN
1 x 64
1 x 64
1 x 64
S
S
S
S
…
…
…
C
MLP (M)
C
MLP (M)
C
MLP (M)
C
MLP (M)
shared
shared
shared
Recurrent Consolidation Unit (RCU)
:block feature
GRU RNN (unrolled)
M
: max pool
S
: stack
C
: concatenate
Figure 4: Architecture with grid input blocks and a recurrent consolidation unit (GB-RCU). The network takes as
input four blocks from a grid structure, each one containing N D-dimensional points. It then learns the block-features
using the same MLP weights for each block. All block-features are passed through a recurrent consolidation unit (see
Section 3.3) which shares the spatial context among all blocks and returns updated block-features. The updated blockfeatures are appended to the input-features together with the original block-features and used to compute the output per point
scores. Shaded fields represent block-features. Some skip-connections are omitted for clarity.
3.5. Grid (G) Architecture
Figure 4 shows the pipeline of the architecture with grid
input blocks. It consists of the following components: The
input level context is a group of four blocks from a 2x2
grid-neighborhood (see Section 3.2) is fed into a series
of MLPs that transform the point features, with weights
shared among all blocks. These block-features are passed
to an RCU that updates the individual block-features with
common context from all neighboring blocks. The updated block-features are then concatenated with the original
block-features. They are then used, along with the local features, for class predictions. After a series of fully connected
layers the output of class scores is computed for each point.
4. Experiments
For experimental evaluation, we compare our two architectures with PointNet [25], the current state-of-the-art
semantic segmentation method directly operating on point
clouds. We produce quantitative results for our models and
the baseline on two challenging datasets: Stanford LargeScale 3D Indoor Spaces (S3DIS) [1] and on Virtual KITTI
(vKITTI) [7]. Additionally, we provide qualitative results
on point clouds obtained from a Velodyne HDL-64E LiDAR scanner from the KITTI dataset [8]. We will now describe these datasets in more detail.
Stanford Large Scale 3D Indoor Scenes. This dataset
is composed of 6 different large scale indoor areas, mainly
conference rooms, personal offices and open spaces. It contains dense 3D point clouds scanned using a Matterport
camera. Each point is labeled with one of the 13 semantic
classes listed in Table 1. Using the reference implementation of PointNet, we were able to reproduce the results reported by Qi et al. [25], see Table 4. Throughout the paper,
we follow the same evaluation protocol used in [25], which
is a 6-fold cross validation over all the areas.
Virtual KITTI. Due to the lack of semantically annotated large-scale outdoor datasets, we rely on the photorealistic synthetic vKITTI dataset which closely mimics the
real-world KITTI dataset. It consists of 5 different monocular video sequences in urban settings, fully annotated with
depth and pixel-level semantic labels. In total there are
13 semantic class, listed in Table 2. For our purposes, we
project the given 2D depth into 3D space to obtain semantically annotated 3D point clouds. Conveniently, this procedure results in point clouds that resemble the varying density of real world point clouds obtained by Velodyne LiDAR
scanners (see Figure 5). For test and training, we split the
original sequences into 6 non-overlapping subsequences.
The final train-test sets are created by choosing point clouds
from each subsequence at regular time-intervals. For evaluation, we also follow the 6-fold cross validation protocol.
4.1. Evaluation Measures
As in [25], we evaluate on the intersection over union
(IoU), the average per class accuracy and overall accuracy.
Intersection over union is computed as:
IoU =
TP
TP + FP + FN
(1)
where TP is the number of true positives, FP the number of
false positives and FN the number of false negatives.
4.2. Quantitative Results
In this section, we analyze the effectiveness of the inputblock schemes and the consolidation units exemplary on the
two previously introduced models. As input features, we
differentiate between geometry (XYZ) and geometry with
color (XYZ+RGB).
Geometry with Color. First, we compare the gridblocks in combination with a recurrent consolidation block
(G+RCU) to the original PointNet. Using the same evaluation setup as described in [25] we are able to show improved
results over PointNet, see Table 4 and Table 1. This proves
our hypothesis that RCU are able to convey context among
blocks and thus improving results. During training, each
room is split into blocks of 1x1 m on the ground plane.
Each block extends over the whole room height. Neighboring blocks are overlapping by 0.5 meters in both directions. We select four blocks simultaneously from a 2x2
grid-neighborhood (see Figure 4, left). Each block contains
4096 points. The unrolled GRU is 8 cells long (4 input, 4
output). It’s memory size is 64. During testing, the room is
split into non-overlapping blocks and evaluated on all 2x2
groups of blocks. Each block is evaluated only once.
Next, we take a look at the multi-scale input block with
consolidation units (MS-CU) model. To build the multiscale blocks, we follow the process described in Section 3.4.
As radii, we choose [0.25, 0.5, 1.0] m. As distance metric
we choose the Chebyshev-distance which generates axisaligned rectangular blocks. The middle scale block is equal
to the PointNet block regarding shape and size.
By using sampling (necessary for the multi-scale block
construction), we diverge from the previous training procedure so we re-run all experiments under these new conditions.
We validate the influence of each of the architecture’s
components by adding them one-by-one to our pipeline and
evaluating after each step, see Table 4 and Table 1. First,
we only consider the input-level context i.e the multi-scale
block feature (MS) as input to our pipeline while skipping
the consolidation units. This shows some performance benefits over PointNet but not as much as one would expect
Figure 6: Qualitative results on laser point clouds.
Dataset: Velodyne HDL-64E laser scans from KITTI Raw
[8]. We trained our model on vKITTI point clouds without
color and applied it on real-world laser point clouds. So far,
only classes like road, building and car give decent results.
considering the enlarged input context. Next, we take only
single-scale input blocks and add one consolidation unit
(SS+CU(1)). The results show that the CU outperforms the
MS input blocks. It also shows that CUs provide a simple
technique to boost the network’s performance. Finally, we
combine both the MS blocks and the CU while appending
another CU to the network (MS+CC(2)). This full model is
depicted in Figure 3.
Geometry only.
Until now, each input point
was described by a 9-dimensional feature vector
[X, Y, Z, R, G, B, X 0 , Y 0 , Z 0 ] where [X, Y, Z] are the
spatial coordinates of a point, [R, G, B] its color and
[X 0 , Y 0 , Z 0 ] the normalized coordinated based on the size
of the environment, see [25] for further details. Without
doubt, color is a very strong input feature in the context
of semantic segmentation. In this section, we pose the
question what will happen if no color information is
available like it is the case with point clouds obtained from
laser scanners. To simulate the missing colors, we simply
discard the color information from the input feature and
re-run the experiments. Table 3 and 2 show the obtained
results. See caption for discussion of the results.
4.3. Qualitative Results
We present qualitative results of our models applied to
indoor scenarios in Figure 7 and outdoor results in Figure ??
along with a short discussion. Additionally, we applied our
pre-trained geometry-only model (vKITTI) to real-world
laser data. The results are shown in Figure 6 and Figure ??.
5. Conclusion
Figure 5: We train our network on synthetic point clouds
generated from vKITTI [7] (left) and apply it onto realworld Velodyne LiDAR point clouds (right). The structure
and the varying density are comparable.
In this work, we investigated the question how to incorporate spatial context into a neural network architecture for
3D semantic segmentation. Building upon PointNet, we
proposed two extension (Input-level context and Outputlevel context) which we successfully applied onto indoor
and outdoor datasets. Still, numerous other combinations
remain possible. The full exploration of the design space is
left for future work.
Floo
Wal
l
Bea
m
Colu
mn
Win
dow
Doo
r
Tab
le
Cha
ir
Sofa
Boo
Boa
rd
Clut
ter
kcas
e
Ceil
i ng
*PointNet [25]
*MS
*MS + RCU
*SS + CU(1)
*MS + CU(2)
43.5
44.4
45.5
45.9
47.8
81.5
82.2
83.6
88.6
88.6
86.7
86.9
86.9
92.6
95.8
64.8
64.2
67.5
66.3
67.3
29.4
33.8
40.5
36.2
36.9
16.3
22.8
17.1
23.6
24.9
39.1
43.3
37.0
47.1
48.6
48.1
52.0
48.8
51.2
52.3
52.5
51.0
53.9
50.2
51.9
42.5
38.6
42.3
36.9
45.1
5.4
9.2
6.8
12.6
10.6
37.6
36.1
39.7
33.7
36.8
30.4
23.6
32.8
22.7
24.7
31.4
33.7
34.2
35.3
37.5
PointNet [25]
G + RCU
47.6
49.7
88.0
90.3
88.7
92.1
69.3
67.9
42.4
44.7
23.1
24.2
47.5
52.3
51.6
51.2
54.1
58.1
42.0
47.4
9.6
6.9
38.2
39.0
29.4
30.0
35.2
41.9
r
mean IoU
S3DIS Dataset [1]
rd
Boa
Clut
te
26.4
30.0
12.5
16.8
25.5
31.2
Car
Van
3.5
5.3
0.7
2.2
1.5
3.6
25.1
43.0
3.4
13.3
r
kcas
Boo
3.7
10.6
11.6
18.7
Truck
2.8
9.4
Sofa
Traffi
c
3.6
12.4
42.4
46.1
Misc
Traffi
cSign
49.9
58.4
ir
ail
Guar
dR
17.7
44.0
Cha
Tab
le
51.6
47.2
Pole
Doo
r
35.3
36.7
Light
Win
dow
29.3
28.8
mn
Colu
11.9
14.6
19.6
25.6
Road
76.4
87.1
m
32.9
38.9
Bea
vKITTI Dataset [7]
*PointNet [25]
*MS + CU(2)
37.0
37.7
ing
57.9
58.8
Build
Wal
l
87.2
94.9
ion
84.0
86.5
Vege
tat
mean IoU
17.9
26.4
Floo
r
40.0
43.0
Tree
*PointNet [25]
*MS + CU(2)
mean IoU
Ceil
S3DIS Dataset [1]
Terra
in
ing
e
Table 1: IoU per semantic class. S3DIS dataset with XYZ-RGB input features. We compare our models with different
components against the original PointNet baseline. By adding different components, we can see an improvement of mean
IoU. We obtain state-of-the-art results in mean IoU and all individual class IoU. Entries marked with * use random sampling
for input block selection instead of discrete positions on a regular grid.
Table 2: IoU per semantic class. S3DIS and vKITTI datasets both with XYZ input features (no color). Our methods
not only outperform PointNet consistently on two datasets, the improvements in mean IoU are also more considerable when
no color is available. This suggests that our network architectures are able to learn improved geometric features and are more
robust to varying point densities as they occur in the outdoor vKITTI dataset.
mean
IoU
overall
accuracy
avg. class
accuracy
S3DIS Dataset [1] – no RGB
*PointNet [25]
*MS + CU(2)
40.0
43.0
72.1
75.4
52.9
55.2
vKITTI Dataset [7] – no RGB
*PointNet [25]
*MS + CU(2)
17.9
26.4
63.3
73.2
29.9
40.9
Table 3: S3DIS and vKITTI datasets with only XYZ input
features, without RGB. We show improved results on indoor
(S3DIS) and outdoor (vKITTI) datasets. Our presented mechanisms are even more important when no color is available.
S3DIS Dataset [1]
XYZ-RGB
mean
IoU
overall
accuracy
avg. class
accuracy
*PointNet [25]
*MS
*MS + RCU
*SS + CU(1)
*MS + CU(2)
43.5
44.4
45.5
45.9
47.8
75.0
75.5
77.2
77.8
79.2
55.5
57.6
57.2
57.7
59.7
PointNet [25]
G + RCU
47.6
49.7
78.5
81.1
66.2
66.4
Table 4: S3DIS Dataset with XYZ-RGB input features. Comparison of different context expansion techniques on input- and
output-level (see Sections 3.2–3.3). MS: Multi-Scale, SS: SingleScale, G: Grid, CU: Consolidation Unit, RCU: Recurrent Consolidation Unit. Entries marked with * use random sampling for input
block selection instead of discrete positions on a regular grid.
Ceiling
Floor
XYZ-RGB Input
Wall
Beam
Column
PointNet[25]
Window
Door
Table
Ours, G-RCU
Chair
Sofa
Bookcase
Ours, MS-CU(2)
Board
Clutter
Ground Truth
Figure 7: Indoor qualitative results. Dataset: S3DIS [1] with XYZ-RGB input features. From left to right: input point
cloud, baseline method PointNet, our results using the G-RCU model (see Figure 4), our results using the MS-CU(2) model
(see Figure 3), ground truth semantic labels. Our models produce more consistent and less noisy labels.
Terrain Tree Vegetation Building Road Car Truck Van GuardRail TrafficSign TrafficLight Pole Misc
Ours, MS-CU(2)
PointNet [25]
Ground Truth
Figure 8: Outdoor qualitative results. Dataset: Virtual KITTI [7]. Results were obtained using only XYZ coordinates as input, no color information was used. Left: baseline method PointNet. Center: our results using the MS-CU model
as illustrated in Figure 3. Right: ground truth semantic labels. The outputs of our method are less fragmented (cars, houses)
and finer structures like street lights and poles are recognized better.
Tree
Grass
Topiary
Ground
Obstacle
Unknown
6. Acknowledgment
We are grateful to our colleagues for providing valuable
feedback on the paper and having fruitful discussions, especially with Umer Rafi and Paul Voigtlaender. This work
was supported by the ERC Starting Grant project CVSUPER (ERC-2012-StG-307432).
Ground truth
Labels: top
Our prediction
Labels: below
Terrain Tree Vegetation GuardRail TrafficSign TrafficLight
Figure 9: Qualitative results on 3DRMS’17 Challenge.
We trained our model on vKITTI point clouds without
color and applied it to the 3DRMS laser data. Training and
test datasets do not have the same semantic labels. Despite
that, common classes like trees are successfully segmented
and plausible ones are given otherwise (e.g. terrain instead
of grass, guardrail instead of obstacle).
References
[1] I. Armeni, O. Sener, A. R. Zamir, H. Jiang, I. Brilakis,
M. Fischer, and S. Savarese. 3D Semantic Parsing of LargeScale Indoor Spaces. In CVPR, 2016.
[2] A. Boulch, B. L. Saux, and N. Audebert. Unstructured Point
Cloud Semantic Labeling Using Deep Segmentation Networks. In Eurographics Workshop on 3D Object Retrieval,
2017.
[3] L. Chen, G. Papandreou, I. Kokkinos, K. Murphy, and A. L.
Yuille. DeepLab: Semantic Image Segmentation with Deep
Convolutional Nets, Atrous Convolution, and Fully Connected CRFs. arXiv preprint arXiv:1606.00915, 2016.
[4] K. Cho, B. van Merrienboer, Ç. Gülçehre, F. Bougares,
H. Schwenk, and Y. Bengio. Learning Phrase Representations using RNN Encoder-Decoder for Statistical Machine
Translation. In EMNLP, 2014.
[5] F. Engelmann, J. Stückler, and B. Leibe. Joint object pose
estimation and shape reconstruction in urban street scenes
using 3D shape priors. In Proc. of the German Conference
on Pattern Recognition (GCPR), 2016.
[6] F. Engelmann, J. Stückler, and B. Leibe. SAMP: shape and
motion priors for 4d vehicle reconstruction. In WACV, 2017.
[7] A. Gaidon, Q. Wang, Y. Cabon, and E. Vig. Virtual Worlds as
Proxy for Multi-Object Tracking Analysis. In CVPR, 2016.
[8] A. Geiger, P. Lenz, C. Stiller, and R. Urtasun. Vision meets
Robotics: The KITTI Dataset. IJRR, 32(11), 2013.
[9] T. Hackel, N. Savinov, L. Ladicky, J. D. Wegner,
K. Schindler, and M. Pollefeys. Semantic3D.net: A new
Large-scale Point Cloud Classification Benchmark. arXiv
preprint arXiv:1704.03847, 2017.
[10] T. Hackel, J. D. Wegner, and K. Schindler. Fast Semantic
Segmentation of 3D Points Clouds with Strongly Varying
Density. ISPRS, 3(3), 2016.
[11] S. Hochreiter and J. Schmidhuber. Long Short-Term Memory. Neural computation, 9(8), 1997.
[12] J. Huang and S. You. Point cloud labeling using 3D Convolutional Neural Network. In ICPR, 2016.
[13] A. Kasyanov, F. Engelmann, J. Stückler, and B. Leibe.
Keyframe-Based Visual-Inertial Online SLAM with Relocalization. In IROS, 2017.
[14] R. Klokov and V. S. Lempitsky. Escape from Cells: Deep
Kd-Networks for The Recognition of 3D Point Cloud Models. arXiv preprint arXiv:1704.01222, 2017.
[15] K. Lai, L. Bo, and D. Fox. Unsupervised feature learning for
3D scene labeling. In ICRA, 2014.
[16] J. Long, E. Shelhamer, and T. Darrell. Fully Convolutional
Networks for Semantic Segmentation. In CVPR, 2015.
[17] W. Maddern, G. Pascoe, C. Linegar, and P. Newman. 1 Year,
1000km: The Oxford RobotCar Dataset. IJRR, 36(1), 2017.
[18] D. Maturana and S. Scherer. VoxNet: A 3D Convolutional
Neural Network for Real-Time Object Recognition. In IROS,
2015.
[19] R. Mottaghi, X. Chen, X. Liu, N.-G. Cho, S.-W. Lee, S. Fidler, R. Urtasun, and A. Yuille. The Role of Context for
Object Detection and Semantic Segmentation in the Wild. In
CVPR, 2014.
[20] D. Munoz, N. Vandapel, and M. Hebert. Directional Associative Markov Network for 3-D Point Cloud Classification.
In 3DPVT, 2008.
[21] P. K. Nathan Silberman, Derek Hoiem and R. Fergus. Indoor
segmentation and support inference from rgbd images. In
ECCV, 2012.
[22] H. Noh, S. Hong, and B. Han. Learning Deconvolution Network for Semantic Segmentation. In ICCV, 2015.
[23] P. Ondruska, J. Dequaire, D. Zeng Wang, and I. Posner. Endto-End Tracking and Semantic Segmentation Using Recurrent Neural Networks. In RSS, Workshop on Limits and Potentials of Deep Learning in Robotics, 2016.
[24] T. Pohlen, A. Hermans, M. Mathias, and B. Leibe. FullResolution Residual Networks for Semantic Segmentation in
Street Scenes. In CVPR, 2017.
[25] C. R. Qi, H. Su, K. Mo, and L. J. Guibas. PointNet: Deep
Learning on Point Sets for 3D Classification and Segmentation. In CVPR, 2017.
[26] C. R. Qi, H. Su, M. Niener, A. Dai, M. Yan, and L. J. Guibas.
Volumetric and Multi-View CNNs for Object Classification
on 3D Data. In CVPR, 2016.
[27] G. Vosselman. Point Cloud Segmentation for Urban Scene
Classification. ISPRS, 1, 2013.
[28] Z. Wu, C. Shen, and A. van den Hengel. High-performance
Semantic Segmentation Using Very Deep Fully Convolutional Networks. arXiv preprint arXiv:1604.04339, 2016.
[29] X. Xiong, D. Munoz, J. A. Bagnell, and M. Hebert. 3-D
Scene Analysis via Sequenced Predictions over Points and
Regions. In ICRA, 2011.
[30] P. Xu, F. Davoine, J. Bordes, H. Zhao, and T. Denoeux. Information Fusion on Oversegmented Images: An Application
for Urban Scene Understanding. In MVA, 2013.
[31] L. Yi, H. Su, X. Guo, and L. J. Guibas. SyncSpecCNN: Synchronized Spectral CNN for 3D Shape Segmentation. arXiv
preprint arXiv:1612.00606, 2016.
[32] R. Zhang, S. A. Candra, K. Vetter, and A. Zakhor. Sensor Fusion for Semantic Segmentation of Urban Scenes. In ICRA,
2015.
| 1cs.CV
|
arXiv:1505.01034v2 [math.OC] 12 Apr 2016
A probabilistic interpretation of set-membership filtering:
application to polynomial systems through polytopic bounding ⋆
Alessio Benavoli a , Dario Piga b ,
a
b
IDSIA Dalle Molle Institute for Artificial Intelligence SUPSI-USI, Manno, Switzerland.
IMT Institute for Advanced Studies Lucca, Piazza San Francesco 19, 55100 Lucca, Italy.
Abstract
Set-membership estimation is usually formulated in the context of set-valued calculus and no probabilistic calculations are necessary.
In this paper, we show that set-membership estimation can be equivalently formulated in the probabilistic setting by employing sets of
probability measures. Inference in set-membership estimation is thus carried out by computing expectations with respect to the updated
set of probability measures P as in the probabilistic case. In particular, it is shown that inference can be performed by solving a particular
semi-infinite linear programming problem, which is a special case of the truncated moment problem in which only the zero-th order
moment is known (i.e., the support). By writing the dual of the above semi-infinite linear programming problem, it is shown that, if the
nonlinearities in the measurement and process equations are polynomial and if the bounding sets for initial state, process and measurement
noises are described by polynomial inequalities, then an approximation of this semi-infinite linear programming problem can efficiently be
obtained by using the theory of sum-of-squares polynomial optimization. We then derive a smart greedy procedure to compute a polytopic
outer-approximation of the true membership-set, by computing the minimum-volume polytope that outer-bounds the set that includes all
the means computed with respect to P.
Key words: State estimation; Filtering; Set-membership estimation; set of probability measures; Sum-of-squares polynomials.
1 Introduction
Inferring the value of the state of a dynamical system at the
various time instants is a classical problem in control and estimation theory. The state is estimated based on noisy signal
observations and on a state transition model, which in turn
is affected by two sources of uncertainty (namely, process
disturbance and uncertainty on the initial state conditions).
In the literature, there are two main approaches for dealing
with the uncertainties and noises acting on the system:
• the stochastic (probabilistic) approach that assumes that
the noises and the uncertainties are unknown but they can
be described by known probability distributions.
• the set-membership approach that assumes that the noises
and the uncertainties are unknown but bounded in some
compact sets.
The probabilistic approach is grounded on Bayesian filtering, whose aim is to update with the measurements and
⋆ This paper was not presented at any IFAC meeting. Corresponding author Alessio Benavoli. Tel. 0041 58 666 6509
Email addresses: [email protected] (Alessio Benavoli),
[email protected] (Dario Piga).
Preprint submitted to Automatica
propagate up on time the probability density function (PDF)
of the state. Inferences are then carried out by computing
expectations with respect to this PDF, i.e., mean, variance,
credible regions. It is well known that, for linear discretetime dynamical systems corrupted by Gaussian noises, the
Bayesian filter reduces to the Kalman filter.
The set-membership approach is instead based on the
construction of a compact set which is guaranteed to include the state values of the system that are consistent
with the measured output and the assumed bounds on
the noises/disturbances [1,2,3,4,5,6]. This compact set is
propagated in time and updated recursively with the output observations. In set-membership estimation, computing
inferences thus means to determine this compact set. Setmembership estimation was first proposed in [7,8], where
an ellipsoidal bounding of the state of linear dynamical
systems is computed. The application of ellipsoidal sets
to the state estimation problem has also been studied by
other authors, for example [9,10], and, independently, in the
communications and signal processing community, starting
from the works [11,12,13,14]. In order to improve the estimation accuracy, the use of a convex polytope instead of an
ellipsoid has been proposed in [15,16]. Unfortunately such
a polytope may be extremely complex and the correspond-
11 November 2017
about the bounding region as well as the probabilistic moments (mean and variance) of the noises or that are able
to deal with a Gaussian measurement noise and a bounded,
with known moments, process noise etc.. Moreover, it can
allow us to compute credible regions (Bayesian confidence
intervals) that takes into account of both deterministic and
probabilistic uncertainty, as well as it allows us to make decisions by choosing the action that minimizes the expectation of some loss function (this is important, for instance, in
control design). In the context of this paper a first attempt
in combining deterministic and probabilistic uncertainty has
been proposed in [29], while [31] has proposed a joint Zonotopic and Gaussian Kalman filter for discrete-time LTV systems simultaneously subject to bounded disturbances and
Gaussian noises. The work [32] instead proposes a Bayesian
approach to set-membership estimation imposing a uniform
distribution on the membership-set similar to the idea proposed in [33,34]. We will show that this approach is different from set-membership estimation, since set-membership
estimation cannot be interpreted in the Bayesian framework,
but only in the framework of set of probability measures.
Second, under this probabilistic interpretation, inferences in
set-membership estimation are carried out by computing expectations with respect to the set P as in the probabilistic
case. In particular, we show that the membership set X (i.e.,
the set that includes the state with guarantee) can be obtained
by computing the union of the supports of the probability
measures in P. Moreover, we prove that a minimum volume
convex outer-approximation of X can simply be obtained
by computing the set M that includes all the means computed with respect to the probabilities in P. The proof is not
constructive, hence we do not have a convenient description
of M. However we show that we can determine the least
conservative half-space H that includes M , by solving a
semi-infinite linear programming problem. This problem is
a special case of the truncated moment problem [35,36,37]
in which only the zero-th order moment is known (i.e., the
support).
Third, by writing the dual of the above semi-infinite linear
programming problem, we show that, if the nonlinearities in
the measurement and process equations are polynomial and
if the bounding sets for initial state, process and measurement noises are described by polynomial inequalities, then
a feasible solution of the dual can be obtained by simply
checking the non-negativity of a polynomial on a compact
set described by polynomial inequalities. An approximation
of this semi-infinite linear programming problem can be obtained by reformulating it as semidefinite programming by
using the theory of sum-of-squares (SOS) polynomial optimization. We prove that the approximate solution is robust,
in the sense that the computed half-space H is guaranteed
to include M, and so the membership set X .
Fourth, we provide a procedure to determine the minimumvolume polytope S bounding M. This procedure is based
on a refinement of the algorithm originally proposed in [38]
to compute an approximation of the minimum-volume polytope containing a given semialgebraic set. In particular, we
use a Monte Carlo integration approach to compute an approximation of the volume of a polytope, and a greedy pro-
ing polytopic updating algorithms may require an excessive
amount of calculations and storage (without any approximations, the number of vertices of the polytope increases
exponentially in time). For this reason, it has been suggested to outer approximate the true polytope with a simpler
polytope, i.e. possessing a limited number of vertices or,
equivalently, faces [17]. In this respect, a parallelotopic
approximation of the set-membership set was presented in
[18,19]. A parallelotope is the generalisation of the parallelogram to Rn . Minimum-volume bounding parallelotopes
are then used to estimate the state of a discrete-linear dynamical system with polynomial complexity. Zonotopes
have been proposed to reduce the conservativeness of parallelotopes. Intuitively zonotopes are polytopes with parallel
faces, for a more precise definition see [20, Ch. 2]. A parallelotope is thus a special zonotope. Zonotopes are used in
[21,22,23] to build a state bounding observer in the context
of linear discrete systems.
Zonotopes are also employed to address the problem of
set-membership estimation for non-linear discrete-time systems with a bounded description of noise and uncertainties [24]. At each sample time, a guaranteed bound of the
uncertain state trajectory of the system is calculated using interval arithmetic applied to the nonlinear functions
through the mean interval extension theorem. This outer
bound is represented by a zonotope. Similar approaches for
set-membership estimation for nonlinear systems are presented in [25,26,27], where ellipsoids are used instead of
zonotopes. Recently, randomized methods are used in [28]
to approximate, with probabilistic guarantees, the uncertain
state trajectory with polynomial sublevel sets.
The aim of this paper is to address the problem of the estimation of the state of a discrete-time non-linear dynamical system (characterized by polynomial non-linearities) in
which initial state and noises are unknown but bounded by
some compact sets (defined by polynomial inequalities). We
are therefore in the context of set-membership estimation,
but we will address this problem in a very different way
from the approaches presented above. We reformulate setmembership in the probabilistic setting and solve it using
the theory of moments and positive polynomials. More precisely the contributions are the following.
First, by exploiting recent results on filtering with sets of
probability measures [29,30], we show that set-membership
estimation can be equivalently formulated in a probabilistic setting by employing sets of probability measures. In
particular, we show that the prediction and updating steps
of set-membership estimation can be obtained by applying
Chapman-Kolmogorov equation and Bayes’ rule point-wise
to the elements of this set of probability measures P . This
unifies the probabilistic approach (Bayes filter) and the setmembership approach to state estimation. This result can
have an enormous impact, because it finally can allow us to
combine set-membership and classical probabilistic uncertainty in order to obtain hybrid filters, i.e., stochastic (probabilistic) filters that are for instance able to use information
2
and therefore n = 2, m = 1, d = 2 and s(d) =
cedure to determine an outer-bounding polytope S as the
intersection of a pre-specified number of half-spaces Hj ,
where each half-space Hj is added to the description of S
so to minimize the volume of the polytope including M.
This allows us to solve the set-membership estimation problem for polynomial non-linear systems very efficiently and
through convex optimization.
Finally, by means of a numerical example involving the
Lotka Volterra prey-predator model, we show the effectiveness of our approach.
x(0) ∈ X0 , w(k) ∈ Wk , v(k) ∈ Vk ,
y(k) = cd (x(k), k) + v(k),
Wk = {w(k) ∈ Rn : hw
i (w(k), k) ≤ 0, i = 1, . . . , tw } ,
(5)
where hw
i (with i = 1, . . . , tw , tw ∈ N) are polynomial
functions in the variable w(k). The sets X0 , Vk are described
in a similar manner.
This paper addresses a set-membership filtering problem,
which aims at recursively estimating, at each time sample
k = 1, 2, . . . , To , (an outer approximation of) the state uncertainty set Xk , defined as the set of all values x(k) compatible with the available information, namely the system
equations (1), the bounds on the initial state and on the noises
(4), and the output observations y(1), y(2), . . . , y(To ). Formally, the set-membership filtering problem is defined as
follows.
(1)
where x(k) = [x1 (k), . . . , xn (k)]⊤ ∈ Rn is the state of the
system at the time k, y(k) ∈ Rm is the measured output
vector, w(k − 1) ∈ Rn is the process noise and v(k) ∈
Rm is the measurement noise. In this paper, we consider
polynomial non-linearities ad (x(k), k) and cd (x(k), k), i.e.,
ad (x(k − 1), k − 1) =Ak−1 qd (x(k − 1)),
cd (x(k), k) =Ck qd (x(k)),
(4)
where X0 , Wk , Vk are compact basic semi-algebraic sets,
i.e., compact sets described by the polynomial inequalities:
Consider an uncertain non-linear discrete-time dynamical
system described by the difference equations:
x(k) = ad (x(k − 1), k − 1) + w(k − 1),
= 6.
We further assume that the only available information about
the initial state x(0) and the noises w(k), v(k) is:
2 Problem Description
(
n+d
d
Problem 1 [Set-membership filtering]
Given the system equations (1), the observations, the bounding sets for the noises Wk , Vk and the initial state uncertainty set X0 , compute recursively the state uncertainty set
Xk defined as:
(2a)
(2b)
with
qd (x) =
Xk = { x(k) ∈ Rn: x(k)−ad (x(k − 1), k − 1) ∈ Wk−1 ,
[1, x1 , . . . , xn , x21 , x1 x2 , . . . , xn−1 xn , x2n , . . . , xd1 , . . . , xdn ]⊤
(3)
being the vector of all monomials of degrees
less
than
or
equal to d, which has dimension s(d) = n+d
,
and
A
∈
k−1
d
Rn×s(d) , Ck ∈ Rm×s(d) are known time-variant coefficient
matrices. The resulting system will be referred in the paper
as uncertain time-variant polynomial system of degree d.
y(k) − cd (x(k), k) ∈ Vk ,
x(k − 1) ∈ Xk−1 }
for each k = 1, 2, . . . , To .
Note that, in general, the sets Xk might be nonconvex and
their representation can become more and more complicated
as the time index k increases. Under the assumption that
Xk is bounded, algorithms for computing simple sets (e.g.,
boxes, parallelotopes, zonotops or ellipsoidal regions) outerbounding the state uncertainty sets Xk have been then proposed to reduce this complexity. After formulating the setmembership filtering problem in a probabilistic setting, this
paper presents an algorithm for computing (an approximation of) the minimum-volume polytope outer-bounding the
sets Xk .
Example 1 Let us consider the discrete-time polynomial
system:
x1 (k) = x1 (k − 1) (2 − x1 (k − 1)) + w1 (k − 1),
x2 (k) = x1 (k − 1)x2 (k − 1) + 0.5x2 (k − 1) + w2 (k − 1),
The output equation is given by: y(k) = x1 (k) + x2 (k) +
v(k). We can rewrite this system as in (2a)–(2b):
qd (x) = [1, x1 , x2 , x21 , x1 x2 , x22 ]⊤
"
#
0 2 0 −1 0 0
Ak−1 =
0 0 0.5 0 1 0
h
i
Ck−1 = 0 1 1 0 0 0
3 A probabilistic framework for set-membership estimation
Set-membership estimation is usually formulated in the context of set-valued calculus. We will show in the following
3
upper bound for the expectation of g is given by the solution
of the optimization problem:
paragraph that set-membership estimation can be equivalently formulated in the probabilistic setting by employing
sets of probability measures. Consider the set-membership
constraint x ∈ X (the time index is dropped for brevity of
notation) with X ⊂ Rn . This constraint can be translated
in a probabilistic setting by saying that the only probabilistic information on the value x of the variable X is that it
belongs to the set X , or equivalently,
sup
P
X
(7)
Problem (7), i.e., determining an upper bound for the expectation of g with respect to the probability measure Pr given
the knowledge of its support X , is a special case of the truncated moment problem [35,36,37] in which only the zero-th
order moment is known (i.e., the support). Hence, we have
the following result [39], [40, Lemma 3.1]:
The support does not uniquely define a probability measure,
as there are an indefinite number of probability measures
with support X . 3 Hence, x ∈ X is equivalent to the constraint that the probability measure of x belongs to the set
PX (X), that is the set of all probability measures on the
variable X with support X . Let us define with P the Cumulative Distribution Function (CDF) of the probability measure Pr. For instance on R we have that P (x) = Pr(−∞, x]
(this definition can easily be extended to Rn ). Then we can
easily characterize the set of probability measures PX (X)
as follows:
Proposition 1 The optimum of (7) is obtained by an atomic
measure 4 Pr = δx̂ , where x̂ = arg supx∈X g(x).
Note in fact that, ∀ Pr ∈ PX (X), with associated CDF P ,
E[g] =
dP (x) = 1 ,
g(x)dP (x),
which is a semi-infinite linear program, since it has a finite
number constraints and an infinite dimensional variable (the
probability measure Pr). Note that we use “sup” instead
of “max” to indicate that an optimal solution might not be
attained. The lower bound of the expectation can be obtained
by replacing sup with inf.
where Pr is a probability measure on X . 1 More precisely
Pr is a nonnegative Borel measure on X . 2 In other words,
this means that we only know the support of the probability
measure of the variable X.
R
X
s.t. P ∈ PX (X),
Pr(X ∈ X ) = Pr(X ) = 1,
PX (X) = P :
R
Z
g(x)dP (x) ≤
X
(6)
Z
g(x)δx̂ (dx) = g(x̂),
X
where g(x̂), by definition of x̂, is the supremum of g on
X . The first integral must be understood as a LebesgueStieltjes integral with respect to the cumulative distribution
of an atomic measure on Rn . This means that dP (x) denotes
the distributional derivative of the cumulative distribution
of an atomic measure, that are in our case Dirac measures
δx̂ (dx) (hence the second integral). From this result, it follows that the probability measures that gives the lower and
upper bounds for the expectation of g are atomic (discrete)
measures.
In order to formulate the set-membership filtering problem
in a probabilistic framework it is useful to exploit a result derived by Karr in [39], where it is proven that the set of probability measures PX (X) which are feasible for the semiinfinite linear program problem (7) is convex and compact
with respect to the weak∗ topology. As a result, PX (X) can
be expressed as the convex hull of its extreme points and,
according to Proposition 1, these extreme points are atomic
measures on X , i.e.:
where the integral is a Lebesgue-Stieltjes integral with respect to P . Hence, because of the equivalence between Borel
probability measures and cumulative distributions, hereafter
we will use interchangeably Pr and P .
3.1 Inference on the state
In state estimation, we are interested in making inferences
about X or, equivalently, computing expectations of realvalued functions g of X. Since there are an indefinite number
of probability measures with support X , we cannot compute
a single expectation of g. However, we can compute upper
and lower bounds for the expectation of g with respect to the
probability measures Pr with support X . For instance, the
1
To clarify this aspect, consider the experiment of rolling a dice.
Assume that the probability Pr of the outcomes x of the dice
is completely unknown, then the only knowledge about the experiment is that x ∈ X = {1, 2, 3, 4, 5, 6}, or, equivalently, that
Pr({1, 2, 3, 4, 5, 6}) = 1. Therefore, the statement Pr(X ) = 1 is
a model for our (epistemic) uncertainty about the probabilities of
the dice outcomes. We only know that x ∈ {1, 2, 3, 4, 5, 6}.
2
The sample space is Rn and we are considering the Borel σalgebra. X is assumed to be an element of the σ-algebra.
3
The uniform distribution is one of them, but it is not the only
one. So by considering only the uniform distribution as in [32],
we loose the full equivalence with set-membership.
PX (X) ≡ Co {δx̂ : x̂ ∈ X } ,
(8)
where ≡ means equivalent in terms of inferences (expectations). Summing up what we have obtained so far:
4
An atomic measure in Rn is a measure which accepts as an
argument a subset A of Rn , and returns δx (A) = 1 if x ∈ A,
zero otherwise.
4
(1) the set-membership constraint x ∈ X is equivalent to
(6);
(2) for the inferences, PX (X) is equivalent to the convex
hull of all atomic measures on X , (8).
by applying the Chapman-Kolmogorov equation point-wise
to the probability measures Pr(·|x(k−1)) in P(X(k)|X(k−
1)) and Qr in P(X(k − 1)) we obtain
Z
Pr(x(k)) =
Ix(k) (x′ )dP (x′ |x(k − 1))dQ(x(k − 1))
n
n
R
R
Z Z
=
Ix(k) (x′ )δad (x(k−1),k−1)+ŵ (dx′ )δx̂ (dx(k − 1))
n
n
R
R
Z
δad (x(k−1),k−1)+ŵ (x(k))δx̂ (dx(k − 1))
=
Hence, we can derive the prediction and updating step
for set-membership estimation by applying the ChapmanKolmogorov equation and Bayes’ rule to the set of probability measures in (8). This means that, by reformulating
set-membership constraints in a probabilistic way, we can
reformulate set-membership estimation in the realm of
stochastic (probabilistic) filtering applied to set of probability measures.
Rn
= δad (x̂,k−1)+ŵ (x(k))
(12)
where Ix(k) (x′ ) denotes the indicator function 5 and with
x̂ ∈ Xk−1 and ŵ
R ∈ Wk−1 and where we have exploited the fact that Rn Ix(k) (x′ )δad (x(k−1),k−1)+ŵ (dx′ ) =
Ix(k) (ad (x(k −1), k −1)+ ŵ) = δad (x(k−1),k−1)+ŵ (x(k)).
From (8), (12) and the definition of X̂k , the theorem follows.
3.2 Propagating in time and updating set of distributions
We start by deriving the set-membership filtering prediction
step by applying the Chapman-Kolmogorov equation.
Theorem 1 (Prediction) Consider the system equation in
(1) with w(k − 1) ∈ Wk−1 and assume that the only probabilistic knowledge about X(k − 1) is the support Xk−1 .
Then it follows that the probability measure Pr on the value
x(k) of the state at time k belongs to the set
o
n
P̂X̂k (X(k)) ≡ Co δx̂ : x̂ ∈ Xˆk ,
Z
Theorem 1 shows that, by applying the ChapmanKolmogorov equation point-wise to the probability measures in P(X(k)|x(k − 1)) and PXk−1 (X(k − 1)), we can
obtain a set of probability measures PX̂k (X(k)), which is
completely defined by its support and whose support coincides with the one obtained in set-membership estimation
after the prediction step.
(9)
with
We now derive a similar result for the updating step.
n
Xˆk = x(k) : x(k) = ad (x(k − 1), k − 1) + w(k − 1)
o
with x(k − 1) ∈ Xk−1 , w(k − 1) ∈ Wk−1 ,
Theorem 2 (Updating) Consider the measurement equation in (1) with v(k) ∈ Vk and assume that the only probabilistic knowledge about x(k) is described by (9)–(10). Then
it follows that the updated probability measures P on the
value x(k) of the state at time k belongs to the set:
(10)
or equivalently:
n
Xˆk = x(k) : x(k) − ad (x(k − 1), k − 1) ∈ Wk−1
o
with x(k − 1) ∈ Xk−1 . (11)
PXk (X(k)) ≡ Co {δx̂ : x̂ ∈ Xk } ,
(13)
Xk = Xˆk ∩ Yk ,
(14)
Yk = {x(k) : y(k) − cd (x(k), k) ∈ Vk }.
(15)
where
Proof: Let us consider the time instant k. From the system
equation in (1), w(k − 1) ∈ Wk−1 and (8), it follows that
with
P(X(k)|x(k − 1))
≡ Co δad (x(k−1),k−1)+ŵ : ŵ ∈ Wk−1 ,
Proof: Observe that, at each time k,
P(Y(k)|x(k)) ≡ Co δcd (x(k),k)+v̂ : v̂ ∈ Vk .
this is the conditional set of probability measures for the variable X(k) given the value x(k − 1) of the variable X(k − 1)
Hence, since X(k − 1) ∈ Xk−1 and so the set of probability
measures for the variable X(k − 1) is
Then, the updating step consists of applying Bayes’rule
to the probability measures P(Y(k)|x(k)) and to Q in
PXk−1 (X(k − 1)) ≡ Co {δx̂ : x̂ ∈ Xk−1 } ,
5
5
Ix(k) (x′ ) = 1 when x(k) = x′ and zero otherwise.
by Xk , i.e.,
P̂X̂k (X(k)):
R
)dP (y′ |x(k))dQ(x(k))
dP (x(k)|y(k)) = R
I
(y′ )dP (y′ |x(k))dQ(x(k))
Rn Rm y(k)
Pr(y(k)|x(k))dQ(x(k))
= R
,
Rn Pr(y(k)|x(k))dQ(x(k))
RRm Iy(k) (y
′
Pr(y(k)|x(k)) =
Iy(k) (y′ )dP (y′ |x(k)).
Rm
Rn
=
R
A1.1 Initialize PX0 (X(0)) ≡ Co {δx̂ : x̂ ∈ X0 }.
A1.2 For k = 1, . . . , To :
o
n
A1.2.1 P̂X̂k (X(k)) ≡ Co δx̂ : x̂ ∈ X̂k with X̂k defined in (10);
A1.2.2 PXk (X(k)) ≡ Co {δx̂ : x̂ ∈ Xk } with Xk defined in (14).
Pr(y(k)|x(k))dQ(x(k))
Rn
δcd (x(k),k)+v̂ (y(k))δx̂ (dx(k)) > 0.
steps A1.2.1 and A1.2.1 are the prediction and the updating
steps, respectively. Note that the set of probability measures
PXk (X(k)) (or P̂X̂k (X(k))) is computed by taking into account all the observations yk = {y(1), y(2), . . . , y(k)} (respectively yk−1 ). Hence, it should be more correctly denoted
as PXk (X(k)|yk ) (respectively PX̂k (X(k)|yk−1 )). We have
omitted this notation for brevity.
Hence, the above inequality holds if and only if x̂ and v̂ are
chosen, at time k, such that:
cd (x̂, k) + v̂ = y(k).
(16)
Bayes’ rule is only defined for those probability measures
for which the denominator is strictly positive, that implies
that the above equality must be satisfied. 6 The equality (16)
can be satisfied only if x̂ ∈ Yk which, together with the
constraint x̂ ∈ X̂k , implies that
Remark 1 Under the assumptions (2a),(2b) and (5), the set
Xk is a semialgebraic set in Rn , described by the intersections of the semialgebraic sets X̂k (Eq. (11)) and Yk (Eq.
(15)). Formally, Xk is the projection in the space of x(k) of
the set
x̂ ∈ Xˆk ∩ Yk .
Under the constraint (16), it follows that δcd (x̂,k)+v̂ (y(k)) =
1 and, thus, the denominator is equal to one. Hence, we have
that
R
(17)
Algorithm 1: prediction and updating
Note that the probability of a point on Rn can be nonzero
since Pr is an atomic measure. In order to apply Bayes’ rule
we need to ensure that the denominator is strictly greater
than zero:
R
dP (x(k)) = 1,
Xk
where Xk is given by (14), or equivalently by (6). In other
words, the support of the probability measure Pr of the value
of the state x(k) given the output observation y(k) and the
system equations (1) is nothing but Xk . This is in accordance
with the set-membership formulation, which claims that
x(k) belongs to state uncertainty set Xk defined in (6). Then
we can solve set-membership filtering by applying recursively Theorems 1 and 2, as described in Algorithm 1. The
where we have exploited the fact that
Z
Z
′
X˜k = x̃ ∈ R2n : hs (x̃(k)) ≤ 0, s = 1, . . . , m ,
(18)
where x̃(k) is the augmented state vector x̃(k) =
⊤
⊤
x (k) x⊤ (k − 1) and hs (x̃(k)) (with s = 1, . . . , m)
are the polynomial functions in x(k) and x(k −1) (or equivalently in x̃(k)) defining Xˆk and Yk . In the rest of the paper,
we will use the following notation to describe the set Xk :
′
dP (x(k)|y(k)) = Rm Iy(k) (y )dP (y |x(k))dQ(x(k))
R
= Rm Iy(k) (y′ )δcd (x(k),k)+v̂ (y(k))δx̂ (dx(k))
= δcd (x(k),k)+v̂ (y(k))δx̂ (dx(k))
= δcd (x̂,k)+v̂ (y(k))δx̂ (dx(k))
Xk = {x(k) ∈ Rn : hs (x̃(k)) ≤ 0, s = 1, . . . , m} . (19)
= δx̂ (dx(k))
Remark 2 The reformulation of set-membership in the
probabilistic framework is important for two main reasons.
First, it allows us to reinterpret the operations performed
in set-membership estimation and justifies them in terms
of a probabilistic framework. We have just seen the reinterpretation of prediction and updating in terms of the
Chapman-Kolmogorov equation and Bayes’ rule. We will
further investigate this interpretation in the next sections.
In particular, in Section 4, we will show that the convex
membership set computed in set-membership estimation can
also be interpreted as the set of posterior means calculated
with x̂ ∈ Xˆk ∩ Yk . Hence, the updated probability measure Pr(·|y(k)) on the values of the state at time k is
Pr(·|y(k)) = δx̂ , which proves the theorem.
From Theorem 2, the support of the updated probability
measure Pr on the value x(k) of the state at time k is given
6
This way of updating set of probability measures has been
proposed by Walley [41, Appendix J] under the name of regular
extension.
6
Theorem 3 Assume that Xk is compact and that Ω1 ⊆ Rn
is a convex set defined as follows:
with respect to the posterior set of probability measures
PXk (X(k)) (in the Bayesian setting, we know that the
posterior mean is the optimal estimate with respect to a
quadratic loss function – a similar result holds for the set
of posterior means [42, Sec.5]). This result can also now
be applied to set-membership estimation because, after this
probabilistic interpretation, we are now able to compute
expectations. Moreover, in Section 5 we will also highlight
the connection between set-membership estimation and the
theory of moments (through duality).
Second, we are now potentially able to combine setmembership and classical probabilistic uncertainty in order
to obtain hybrid filters, i.e., stochastic (probabilistic) filters that are for instance able to use information about
the bounding region as well as the probabilistic moments
(mean and variance) of the noises or that are able to deal
with a Gaussian measurement noise and a bounded, with
known moments, process noise etc.. A first attempt in this
direction is described in [29] for scalar systems. We plan to
further investigate this direction in future work by using the
theory of SOS polynomial optimization, that we also use in
the next sections.
Ω1 = arg
Ω⊆R Ω
dx(k)
Then, it results that Ω1 = M, with
Z
M=
x(k)dP (x(k)) : P ∈ PXk (X(k)) .
(21)
(22)
Xk
Proof: From (21) it follows that Ω1 is the minimum volume
convex set that includes Xk . Thus, if Xk is convex, then
Ω1 = Xk . Hence, from (8), the equality
Z
x(k)δx̂(k) (dx(k)) = x̂(k),
Xk
and (22), it immediately follows that M = Ω1 . Conversely
assume that Xk is not convex, then Ω1 ⊃ Xk . Since Ω1
is the minimum volume convex set that includes Xk , then
Ω1 must be equal to the convex-hull of Xk . This means
that for each x̂ ∈ Ω1 , there exist z1 , z2 ∈ Xk such that
wz1 + (1 − w)z2 = x̂ for some w ∈ [0, 1] (by definition of
convex hull). Then, consider the probability measure
wδz1 + (1 − w)δz2 .
(23)
Because of (8), it holds:
dx(k)
s.t.
R
dP (x(k)) = 1, ∀P ∈ PXk (X(k)).
Ω⊆R ,Ω conv. Ω
Ω
In the probabilistic formulation of filtering, all the available
information at time k is encoded in the posterior probability
distribution of the state x(k) given all the observations yk ).
In the set-membership setting, this information is encoded in
the updated set of probability measures PXk (X(k)). Inferences can then be expressed in terms of expectations computed with respect to this set. The set-membership estimation problem can, for instance, be reformulated as follows:
R
R
s.t.
R
dP (x(k)) = 1, ∀P ∈ PXk (X(k))
4 Computing the support as an inference on the set of
probability measures
Ω∗ = arg minn
inf
n
wδz1 + (1 − w)δz2 ∈ PXk (X(k)),
(20)
and
Z
Ω
The solution of (20) is the minimum-volume set Ω ⊆ Rn ,
such that Pr(x(k) ∈ Ω) = 1 for all probability measures Pr
in PXk (X(k)) (i.e., with support Xk ). 7 Thus, Ω∗ coincides
with Xk . Since Xk may be not convex, the problem (20) is
in general difficult to solve. However, the problem can be
simplified by restricting Ω to be convex, thus computing a
convex outer-approximation of Xk .
(24)
x(k) (wδz1 (dx(k)) + (1 − w)δz2 (dx(k))) = x̂. (25)
Xk
Thus, x̂ belongs to M, and vice versa.
Theorem 3 has the following fundamental implications:
• a convex outer-bounding of the set of all the possible
means computed with respect to the probability measures
in PXk (X(k)) (i.e., the set M) is also a convex outerbounding of the support Xk of the set of probability measures PXk (X(k)).
• the tightest convex outer-bounding of the support Xk of
the set of probability distributions PXk (X(k)) is the set
of the means computed with respect to the probability
measure in PXk (X(k)).
The following theorem shows that computing the minimumvolume convex set Ω such that P (x(k) ∈ Ω) = 1 is
equivalent to obtain the set that includes all the possible
means computed with respect to the probability measure in
PXk (X(k)).
7
It is thus the union of all the supports of the probability measures
in PXk (X(k)).
7
M (or equivalently, including X ), is obtained for ν = ν ∗ ,
with
R
ν ∗ = max ω ⊤ xdP (x)
R P
(28)
s.t. dP (x) = 1.
We can thus use M as an outer-approximation of Xk . Algorithm 1 is therefore modified to include the following additional steps.
Refinement of Algorithm 1: outer-approximation step
X
A1.1.3 Outer-approximate Xk with M defined in (22).
A1.1.4 Redefine PXk (X(k)) ≡ Co {δx̂ : x̂ ∈ M}.
R
Proof: Let ρ = xdP (x) be a point belonging to M. Let
us first prove that if ν ≥ ν ∗ , then M ⊆ H. First, note that:
Z
ω ⊤ ρ ≤ ν ∗ = sup ω ⊤ xdP (x)
P
Z
s.t.
dP (x) = 1
Unfortunately, Theorem 3 does not provide a constructive
way to find the set M. However, by restricting the outerapproximation of the support Xk to have a simple form (e.g.,
a polytope), Theorem 3 can be still exploited to determine an
outer-bounding set of Xk . The following theorem provides
results to compute an outer-bounding box of Xk .
X
Therefore,
for ν ≥ ν ∗ , ω ⊤ ρ ≤ ν ∗ ≤ ν, which means that
R
ρ = xdP (x) also belongs to H for all ρ ∈ M. Thus, H
contains M. By choosing ν = ν ∗ , we obtain the tightest
half-space defined by the normal vector ω that includes M.
Theorem 4 (Box approximation) The minimum volume
box that includes Xk can be found by solving the following
family of optimization problems
R
x∗i (k) = opt xi (k)dP (x(k))
P
R
dP (x(k)) = 1.
s.t.
It can be observed that (28) reduces to (26) when ω = ei
for i = 1, . . . , n, where ei is an element of the natural
basis of Rn . Note that, in Problem (28): (i) the optimization
variables are the amount of non-negative mass assigned to
each point x in X (i.e., the measure P r(x)); (ii) the objective
function and the constraint are linear in the optimization
variables. Therefore, (28) is a semi-infinite linear program
(i.e., infinite number of decision variables but finite number
of constraints). By exploiting duality of semi-infinite linear
program (see for instance [43]), we can write the dual of
(28), which is defined as:
(26)
Xk
for i = 1, . . . , n, where byRselecting opt to be min or max
we obtain theR half-spaces xi (k)dP (x(k)) ≥ x∗i (k) and,
respectively, xi (k)dP (x(k)) ≤ x∗i (k) which define the
box.
The proof of Theorem 4 is provided together with the proof
of Theorem 5. Based on Theorem 4, by computing the lower
and upper means of the components x1 (k), . . . , xn (k) of the
vector x(k), the tightest box that outer-approximates Xk is
obtained. In the following we will discuss how to efficiently
solve optimization problems similar to (26) and how to find
an outer-approximation of Xk that is less conservative than
a box. For simplicity of notation, in the rest of the paper, the
dependence of the state x(k) and of the set Xk on the time
index k will be dropped, and only used when necessary.
ν ∗ = inf ν
ν
s.t. ν ≥ ω ⊤ x, ∀x ∈ X ,
which is also a semi-infinite linear program (i.e., finite number of decision variables (ν) but infinite number of constraints). A solution ν is feasible for Problem (29) provided
that:
ν − ω ⊤ x ≥ 0, ∀x ∈ X .
Hence, checking the feasibility of ν is equivalent to check
the non-negativity of the polynomial ν − ω ⊤ x in the set X .
5 Exploiting duality
Remark 3 The probabilistic formulation of the setmembership estimation described so far is general enough,
and it is valid also when the dynamical system in (1) is
not a polynomial system and when the uncertainty sets
X0 , Wk , Vk in (4) are not semialgebraic, but just compact sets. The assumptions of polynomiality are used in
the following to efficiently solve the semi-infinite linear
programming problem (29) through convex optimization.
In this section we discuss how to efficiently solve optimization problems similar to (26). In particular, we slightly modify (26) in order to be able to determine the more general
half-space
H = ρ ∈ Rn : ω ⊤ ρ ≤ ν ,
(27)
R
where ω ∈ Rn , ν ∈ R and ρ = xdP (x). 8
5.1 Sum-of-squares polynomials
Theorem 5 Let us fix the normal vector ω defining the halfspace H in (27). Then, the tightest half-space H including
8
(29)
A sufficient condition for a polynomial to be non-negative
over a semialgebraic set is that it can be written in terms of
sum-of-squares (SOS) polynomials (see, e.g., [44]).
The half-space H lies on the space of the means.
8
Definition 1 A polynomial σ(x̃), with x̃ ∈ R2n , of degree
2d is a sum-of-squares polynomial, denoted by σ(x̃) ∈ Σ[x̃],
if and only if it can be written as:
σ(x̃) = qd (x̃)⊤ Qqd (x̃),
to linear equalities in ν and in the matrix coefficients Qs (with s = 1, . . . , m). Besides, enforcing
σ0 (x̃), σ1 (x̃), . . . , σm (x̃) to be sum of square polynomials leads to linear matrix inequality (LMI) constraints in the coefficients of σ0 (x̃), σ1 (x̃), . . . , σm (x̃)
(i.e., Qs 0).
(2) For ν = ν ∗∗ , the robust constraint ν ∗∗ − ω ⊤ x ≥
0 ∀x ∈ X appearing in Problem (29) is guaranteed to
be satisfied. As matter of fact, for all x̃ ∈ X˜ , hs (x̃) ≤ 0
(with s = 1, . . . , m) by definition of X˜ . Furthermore,
the SOS polynomials σs (x̃) = qd (x̃)⊤ Qs qd (x̃) (with
s = 0, . . . , m) are always nonnegative over R2n as
Qs 0. Thus, both the left and the right side of the
equation in Problem (33) are nonnegative for all x̃ ∈ X̃ .
(3) Since the equality constraint in (33) gives only a sufficient condition for the non-negativity of ν − ω⊤ x on
X , it follows that ν ∗ ≤ ν ∗∗ . Therefore, conservativeness is introduced in solving (33) instead of (29), as
highlighted in Corollary 1.
(4) However, according to the Putinar’s Positivstellensatz
(see, e.g., [46] and [47, Ch. 3]), a polynomial which is
nonnegative over a compact semialgebraic set X can
exactly always be written as a combination of SOS
polynomials, provided that the degree of the SOS polynomials σ0 (x̃), . . . , σm (x̃) is large enough. In other
words, we can make ν ∗∗ close to ν ∗ by increasing the
degree of the SOS. However, in practice it often happens that the relaxed solution ν ∗∗ and the optimal one
ν ∗ coincide with each other for small values of the SOS
degree 2d.
(30)
where Q is a real symmetric
positive semidefinite matrix
of dimension 2n+d
.
The
vector
of monomials qd (x̃) is
d
defined as in (3). The set of SOS polynomials of degree less
then or equal to 2d is denoted as Σ2d [x̃].
Then, for a given integer d ≥ 1, a sufficient condition for
ν − ω ⊤ x to be non-negative in X is (see for instance [37,
Ch. 4]):
ν − ω ⊤ x = σ0 (x̃) −
m
P
σs (x̃)hs (x̃) ∀ x̃ ∈ R2n
s=1
(31)
σ0 (x̃), σ1 (x̃), . . . , σm (x̃) ∈ Σ2d [x̃],
where hs (x̃) (with s = 1, . . . , m) are the polynomial nonpositive inequality constraints defining the semialgebraic set
X . In order to avoid confusion, we would like to stress
that also ν − ω⊤ x is a polynomial in the variable x̃. In
fact, we remind that the augmented state x̃(k) is defined as:
⊤
x̃(k) = x⊤ (k) x⊤ (k − 1) .
The following (more conservative) optimization problem can
be then solved instead of (29):
ν ∗∗ = inf ν
Corollary 1 The set M is guaranteed to belong to the halfspace H : ω⊤ x ≤ ν ∗∗ , i.e.
ν,σs
ν − ω⊤ x = σ0 (x̃) −
m
X
σs (x̃)hs (x̃), ∀ x̃ ∈ R2n
s=1
M ⊆ H.
σ0 (x̃), σ1 (x̃), . . . , σm (x̃) ∈ Σ2d [x̃].
(34)
(32)
Proof: The proof straightforwardly follows from Theorem 5
and ν ∗ ≤ ν ∗∗ .
Note that, by rewriting the SOS polynomials σs (x̃) (with
s = 0, . . . , m) as in (30), Problem (32) can be also rewritten
as:
Example 2 Let us consider the discrete-time polynomial
system described by the difference equations:
ν ∗∗ = inf ν
ν,Qs
⊤
ν − ω x =qd (x̃)⊤ Q0 qd (x̃)+
m
X
−
qd (x̃)⊤ Qs qd (x̃)hs (x̃), ∀ x̃ ∈ R2n
x1 (k)=x1(k−1)x2 (k−1)(x1(k−1) + x2(k−1))+w1 (k−1),
x2 (k)=x1(k−1)x2 (k−1)(2x1(k−1) + x2(k−1))+w2 (k−1).
(35)
The output equation is given by: y(k) = x1 (k) + x2 (k) +
v(k). The following conditions are assumed: (i) the initial state x(0) belongs to X0 = {x(0) : kx(0)k2 ≤ 0.2},
the process noise w(k) = [w1 (k) w2 (k)]⊤ is bounded by
kw(k)k2 ≤ 0.4, and the measurement noise by kv(k)k∞ ≤
0.5. The observed output y(k) at time k = 1 is y(k) = 0.
We are interested in computing an half-space H : ω ⊤ ρ ≤ ν
containing the state uncertainty set Xk (or equivalently M)
at time k = 1. The normal vector ω characterizing H is
fixed and it is equal to ω = [−1 − 0.5]⊤ . In order to compute the constant parameter ν defining H, the SDP Problem
s=1
Qs 0, s = 0, . . . , m.
(33)
Some remarks:
(1) Problem (33) is a semidefinite programming (SDP)
problem [44,45], thus convex. In fact, checking if the
polynomial ν − ω T x is equal to qd (x̃)⊤ Q0 qd (x̃) −
P
m
⊤
2n
leads
s=1 qd (x̃) Qs qd (x̃)hs (x̃) for all x̃ ∈ R
9
Now consider the following family of half-spaces:
Hj = ρ ∈ Rn : ω ⊤
j ρ ≤ νj ,
for j = 1, . . . , J with J ≥ n + 1. Our goal is to choose the
normal vectors ωj , along with the constant parameters νj ,
defining the half-spaces Hj such that
TJ
(1) M ⊆ S = j=1 Hj ;
(2) the polytope S has minimum volume.
In other words, now also the normal vectors ω j for j =
1, . . . , J have to be optimized. Then, we can formulate the
problem we aim to solve as:
Fig. 1. True state uncertainty set X1 (dark grey region) and half-space H : −ρ1 − 0.5ρ2 ≤ 0.45 (light gray region).
inf
S
(33) with x̃(1) = [xT (1) xT (0)]T and
Z
dx s.t. M ⊆ S,
(38)
S
where S in (38) is constrained to be a polytope. There are
two main aspects making (38) a challenging problem, i.e.,
h1 (x̃(1)) :x1 (0)2 + x2 (0)2 − 0.22 ≤ 0,
(36)
2
(1) the minimum-volume polytope outer-approximating a
generic compact set in Rn might not exist. For instance,
if M is an ellipsoid, its convex hull is described by an
infinite number of half-spaces, namely all the supporting hyperplanes at every boundary point of M.
R
(2) the problem of computing the exact volume S dx of
a polytope S in Rn is #P -hard (see, e.g. [50,51]. The
interested reader is also referred to [52] for details on
#P -hard problems). Although several algorithms have
been proposed in the literature to compute the volume
of a polytope S through triangulation [53,54,55,56],
Gram’s relation [57], Laplace transform [58] or randomized methods [59,60,61], all the approaches mentioned above require an exact description of the polytope S in terms of its half-space or vertex representation. However, in our case, the parameters ω j , νj defining the half-spaces Hj are unknown, as determining
ωj , νj is part of the problem itself.
h2 (x̃(1)) :(x1 (1)−x1(0)x2 (0)(x1(0) + x2(0))) +
|
{z
}
w12 (0)
2
(x1 (1)−x1(0)x2 (0)(2x1(0) + x2(0))) − 0.42 ≤ 0,
{z
}
|
w22 (0)
h3 (x̃(1)) :y(1) − x1 (1) − x2 (1) − 0.5 ≤ 0,
{z
}
|
v(1)
h4 (x̃(1)) : − y(1) − x1 (1) − x2 (1) − 0.5 ≤ 0,
|
{z
}
v(1)
(37)
is solved for a SOS degree 2d = 4. The SOStools [48] has
been used to easily handle the SOS polynomials appearing
in (33). The CPU time taken by the solver SeDuMi [49]
to compute a solution of the SDP Problem (33) on a 2.40GHz Intel Pentium IV with 3 GB of RAM is 2.1 seconds.
The computed half-space H is plotted in Fig. 1, along with
the true state uncertainty set X1 . According to Theorem 5
and Corollary 1, X1 is included in the half-space H. Note
also that, although the original robust optimization problem
(29) has been replaced with the SDP problem (33), the computed parameter ν ∗∗ defining H is such that the hyperplane
ω⊤ x = ν ∗∗ is “almost” tangent to the set X1 . Thus, only a
small level of conservativeness is introduced in using SOS.
In the following paragraph we present a greedy algorithm to
evaluate an approximation of the minimum-volume polytope
outer-approximating the set M.
6.1 Approximation of the objective function
As already pointed out in the previous paragraph, one of the
main problems in solving (38) is that an analytical expression
for the computation of the volume of a polytope S in Rn is
not available and the polytope S is unknown, as computing
S is part of the problem itself. In order to overcome such
a problem, a Monte Carlo integration approach [62] is used
here to approximate the volume of S. Specifically, given an
outer-bounding box B of the set M (which can be computed
as discussed in Theorem 4) and a sequence of N random
points {pi }N
i=1 independent and uniformly distributed in B,
6 Computation of the minimum-volume polytope containing M
In the previous section, given the normal vector ω defining
the half-space H in (27), we have shown how to compute,
through convex optimization, the constant parameter ν such
that M ⊂ H.
10
the integral
R
Algorithm 2: Polytopic outer approximation S of M
[input ] List L = {pi }N
i=1 of N random points uniformly
distributed in the box B.
A2.1 Set j = 1.
A2.2 Compute the half-space Hj , defined as Hj : ω ⊤
j ρ−
νj ≤ 0 (with ωj 6= 0), that contains the minimum number
of points in the list L and such that M is included in Hj ,
i.e.,
dx can be approximated as:
S
Z
dx ≈ V ol(B)
S
N
1 X
I{S} (pi ),
N i=1
(39)
where V ol(B) is the volume of the box B and I{S} (pi ) is the
indicator function of the (unknown) polytope S defined as
I{S} (pi ) =
(
1 if pi ∈ S
0 otherwise
ω ∗j , νj∗ =arg minn
(40)
ω j ∈R
νj ∈R i=1
#
N
1 X
I{S} (pi ) = V ol(S),
E V ol(B)
N i=1
"
N →∞
(41)
(corresponding to the objective function of problem (43)).
Then, the new half-space H2 that minimizes an approximation of the volume of the polytope B ∩H1 ∩H2 is generated.
In order to approximate the volume of B ∩ H1 ∩ H2 , all
the points pi of the list L = {pi }N
i=1 that do not belong
to the polytope B ∩ H1 are discarded, and all and only
the points belonging to B ∩ H1 are collected in a new list
1
The volume of B ∩ H1 ∩ H2
L1 = {pi }N
i=1 (step A2.3).
PN1
is then approximated by i=1 I{H2 } (pi ), with pi ∈ L1 .
The procedure is repeated until NJ+1 = NJ (step A2.4),
which means that the number of samples pi belonging to
the polytope B ∩ H1 ∩ . . . ∩ HJ+1 is equal to the number
of samples pi belonging to the polytope B ∩ H1 ∩ . . . ∩ HJ .
Note that, because of the constraint M ⊆ Hj appearing in
optimization problem (50), the half-spaces H1 , . . . , HJ are
TJ
guaranteed to contain the set M, and thus S = B ∩ j=1 Hj
is an outer approximation of M. Finally, we would like
to remark that, in case we are interested also in bounding
the maximum number of half-spaces defining the polytopic
outer approximation S , Algorithm 2 can be stopped after
an a-priori specified number of iterations.
where w.p. 1 is for with probability 1. For finite samples N ,
the level of accuracy of the approximation in (39) depends
on the shape of the set S as well as on the volume of the
outer box B. The reader is referred to as [62] for details on
Monte Carlo integration methods.
On the basis of (39), the volume minimization of problem
(38) can be then approximated as
min
S∈S
N
X
I{S} (pi ) s.t. M ⊆ S
(43)
A2.3 Collect all the points pi ∈ L belonging to the halfspace Hj (computed through (43)) in a list Lj . Let Nj be
the number of elements of Lj .
A2.4 If Nj < N , then L ← Lj , N ← Nj , j ← j + 1 and
go to step A2.2. Otherwise, set J = j − 1 and go to step
A2.5.
TJ
A2.5 Define the polytope S as S = B ∩ j=1 Hj .
[output ] Polytope S.
where the expectation is taken with respect to the random
variable pi . Furthermore, because of the strong law of large
numbers,
N
1 X
I{S} (pi ) = V ol(S) w.p. 1,
N i=1
I{Hj } (pi )
s.t.
ω j 6= 0
M ⊆ Hj
pi ∈ L, i = 1, . . . , N
Remark 4 It is worth remarking that:
lim V ol(B)
N
X
(42)
i=1
In the following subsection, we describe a greedy procedure
aiming at computing an approximation of the minimization
problem (42).
6.2 A greedy approach for solving (42)
The key steps of the approach proposed in this section to
compute a polytopic outer-approximation S of the set M
are summarized in Algorithm 2.
Example 3 Let us consider again Example 2. The first steps
of Algorithm 2 are visualized in Fig. 2. An outer-bounding
box B of the true state uncertainty set (dark gray region) is
first computed (Fig. (a)). A set of 80 random points (black
dots) uniformly distributed in B is generated (Fig. (b)). The
half-space H1 containing the true state uncertainty set and
the minimum number of points is computed. The points which
do not belong to H1 are discarded (gray dots in Fig. (c)). A
new half-space H2 containing the true state uncertainty set
Algorithm 2 generates a sequence of half-spaces H1 , . . . , HJ
as follows. First, the half-space H1 that minimizes an
approximation of the volume of the polytope B ∩ H1 is
computed. The approximation is due to Rthe fact that the
volume of B ∩ H1 , given by the integral B∩H1 dx, is apPN
proximated (up to the constant V ol(B)
) by i=1 I{H1 } (pi )
N
11
and the minimum number of black dots is computed (Fig.
(d)). Again, the points that do not belong to H1 ∩ H2 are
discarded (gray dots in Fig. (d)). The procedure terminates
when no more black points can be discarded.
RHj (pi )
✻
❅
❅
❅
❅
q 1
IHj (pi )
❅
❅
❅
❅
✲
ω⊤
j p i − νj
Fig. 3. Indicator function IHj (pi ) (black solid line) and approximate function R{Hj } (pi ) (gray thin line). When ω ⊤
j pi − νj > 0,
I{Hj } (pi ) and R{Hj } (pi ) are overlapped and they are equal to 0.
tions I{Hj } (pi ) with the convex functions R{Hj } (pi ), i.e.,
(a)
(b)
ω̃ ∗j , ν̃j∗ =arg minn
N
X
ω j ∈R
νj ∈R i=1
R{Hj } (pi )
s.t.
ω j 6= 0
M ⊆ Hj
pi ∈ L, i = 1, . . . , N.
(c)
(46)
Theorem 6 If (i) there exists at least one point pi in the list
⊤
L such that ω̃ ∗j pi − ν̃j∗ < 0 and (ii) ω̃ ∗j , ν̃j∗ is the optimal
(d)
⊤
solution of problem (46), then the hyperplane ω̃ ∗j ρ−ν̃j∗ = 0
is a supporting hyperplane for the set M.
Fig. 2. First steps of Algorithm 2.
Proof: Theorem 6 is proved by contradiction. Let H̃j∗ be the
Technical details of step A2.2, which is the core of Algorithm 2, are provided in the following sections.
⊤
half-space defined as H̃j∗ : ω̃ ∗j ρ − ν̃j∗ ≤ 0. Let us suppose
that ω̃∗j , ν̃j∗ is a feasible solution of problem (46) such that
⊤
ω̃∗j ρ − ν̃j∗ = 0 is not a supporting hyperplane for M, that
6.3 Approximation of the indicator functions
⊤
is, for some ε > 0, H̃j : ω̃ ∗j ρ − ν̃j∗ + ε ≤ 0 for all x ∈ M.
Let us define ν̃j as ν̃j = ν̃j∗ − ε. Note that {ω̃∗j , ν̃j } is
still a feasible solution of problem (46) and H̃j ⊆ H̃j∗ . Let
P
V∗ = N
i=1 R{H̃∗ } (pi ) be the value of the cost function of
Note that the objective function of problem (43) is noncontinuous and nonconvex since it is the sum of the indicator
functions I{Hj } (pi ) defined as
j
I{Hj } (pi ) =
(
1 if
0 if
ω⊤
j pi
ω⊤
j pi
− νj ≤ 0,
− νj > 0.
Problem (46) obtained for ω = ω̃∗j and ν = ν̃j∗ . R{H̃∗ } (pi )
j
is then given by
(44)
R{H̃∗ } (pi ) =
We then transform it in a convex objective function. Each
indicator function I{Hj } (pi ) is here approximated by the
convex function R{Hj } (xi ) defined as
R{Hj } (pi ) =
(
⊤
−ω⊤
j pi + νj if ω j pi − νj ≤ 0,
0
if ω⊤
j pi − νj > 0.
j
(
⊤
⊤
−ω̃∗j pi + ν̃j∗ if ω̃∗j pi − ν̃j∗ ≤ 0
⊤
if ω̃∗j pi − ν̃j∗ > 0
0
(47)
PN
Similarly, let Ṽ =
i=1 R{H̃j } (pi ) be the value of the
cost function of Problem (46) obtained when ω = ω̃∗j and
ν = ν̃j . The term R{H̃j } (pi ) is the given by
(45)
R{H̃j } (pi ) =
A plot of the functions I{Hj } (pi ) and R{Hj } (pi ) is given in
Fig. 3.
(
⊤
⊤
−ω̃ ∗j pi + ν̃j if ω̃∗j pi − ν̃j ≤ 0
⊤
if ω̃∗j pi − ν̃j > 0
0
(48)
Since H̃j ⊆ H̃j∗ , then when R{H̃∗ } (pi ) = 0, also R{H̃j } (pi )
Problem (43) is thus relaxed by replacing the indicator func-
j
12
is equal to zero. On the other hand, when R{H̃∗ } (pi ) =
following SDP problems:
j
⊤
−ω̃∗j pi
+ ν̃j∗ > 0, then R{H̃j } (pi ) can be equal either to
⊤
zero or to −ω̃∗j pi + ν̃j
⊤
−ω̃∗j pi + ν̃j∗ −ε
ω∗j , ν ∗j = arg minn
⊤
−ω̃ ∗j pi + ν̃j∗ .
=
≤
On the basis of the above considerations, it follows:
R
= R{H̃j } (pi ) if ω̃ ∗j pi − ν̃j∗ ≥ 0
{H̃∗
} (pi )
j
> R{H̃j } (pi ) if ω̃ ∗j pi − ν̃j∗ < 0
R
ω j ∈R
νj ∈R i=1
Qs
⊤
(49)
νj − ω j x = qd (x̃)⊤ Q0 qd (x̃)+
m
X
−
qd (x̃)⊤ Qs qd (x̃)hs (x̃), ∀ x̃ ∈ R2n
Since by hypothesis (i) there exists at least one point pi in
⊤
the list L such that ω̃∗j pi − ν̃j∗ < 0, it follows that V ∗ > Ṽ .
Therefore, ω̃ ∗j , ν̃j∗ is not the optimal solution of problem
(46). This contradicts hypothesis (ii).
s=1
ω ∗j , ν ∗j
νj − ω j x = qd (x̃)⊤ Q0 qd (x̃)+
m
X
−
qd (x̃)⊤ Qs qd (x̃)hs (x̃), ∀ x̃ ∈ R2n
R{Hj } (pi )
(51b)
s=1
Qs 0, s = 0, . . . , m.
pi ∈ L, i = 1, . . . , N
with ω j,1 denoting the first component of vector ω j . The
optimizer {ω∗j , νj∗ } of Problem (50) is the given by the pair
{ω∗j , ν ∗j } or {ω∗j , ν ∗j } that provides the minimum value of
PN
the objective function i=1 R{Hj } (pi ).
Remark 5 For a fixed degree 2d of the SOS polynomials,
the number of optimization variables of Problems (51) increases polynomially with the state dimension n and linearly with the number m of constraints hs (x̃) defining the
set X . Specifically, the number of optimization variables of
Problem (51) is O(mn2d ). In fact, the number of free decision variables
in the
matrices Qs (with s = 0, . . . , m) is
2n+d
2n+d
1+ d
d
= O(n2d ). On the other hand, for a
2
fixed n, the size of the matrices Qs increases exponentially
with the degree 2d of the SOS polynomials. In order not to
obtain too conservative results, practical experience of the
authors suggests to take d ≥ ⌈ d2 ⌉ + 1, where ⌈·⌉ denotes
the ceiling operator. We remind that d is the degree of the
considered polynomial system in (1). Roughly speaking, because of memory requirement issues, the relaxed SDP problems (51) can be solved in commercial workstations and
with general purpose SDP solvers like SeDuMi in case of
polynomial systems with 4 state variables and of degree d
not greater than 6. Systems with more state variables can
be considered in case of smaller values of d. Similarly, systems of higher degree can be considered in case of a smaller
number of state variables.
R{Hj } (pi )
s.t.
ω j 6= 0
ω j ∈R
νj ∈R i=1
Qs
νj − ω j x = qd (x̃)⊤ Q0 qd (x̃)+
m
X
−
qd (x̃)⊤ Qs qd (x̃)hs (x̃), ∀ x̃ ∈ R2n
The constraints M ⊆ Hj can be handled through the SOSbased approach already discussed in Section 5.1. Specifically, by introducing a SOS relaxation, Problem (46) is replaced by:
ω j ∈R
νj ∈R i=1
Qs
= arg minn
N
X
s.t.
ω j,1 = −1
6.4 Handling the constraint M ⊆ Hj
N
X
(51a)
Qs 0, s = 0, . . . , m.
pi ∈ L, i = 1, . . . , N
Theorem 6 has the following interpretation. Among all
the half-spaces defined by the normal vector ω̃ ∗j and containing the set M, the optimization problem (46) provides
⊤
the half-space Hj∗ : ω̃ ∗j ρ − ν̃j∗ ≤ 0 which minimizes the
volume Rof the polytope B ∩ H1∗ ∩ . . . ∩ Hj∗ , even if the
integral B∩H∗ ∩...H∗ dx is approximated (up to a constant)
1
j
PN
by i=1 I{H∗j } (pi ) and the indicator functions I{H∗j } (pi )
are replaced by the convex functions R{H∗j } (pi ).
ω ∗j , νj∗ = arg minn
R{Hj } (pi )
s.t.
ωj,1 = 1
⊤
{H̃∗
} (pi )
j
N
X
(50)
s=1
Qs 0, s = 0, . . . , m.
pi ∈ L, i = 1, . . . , N
Note that, as already discussed in Section 5.1, the constraint
satisfied for all x ∈ X . Therefore, the
νj − ω ⊤
j x ≥ 0 is
is guaranteed to
half-space: Hj = ρ ∈ Rn : ω ⊤
j ρ ≤ νj
contain X . Thus, also the set M is included in Hj . Finally,
note that, in order to deal with the nonconvex constraint
ωj 6= 0 in (50), Problem (50) can be splitted into the two
13
Remark 6 As already discussed, Algorithm 2 computes, at
⊤
each iteration, an half-space Hj : ωj x̃ − νj ≤ 0 containing
the set X (thus also M), i.e.,
⊤
ωj x̃ − νj ≤ 0
∀ x̃ ∈ X .
(52)
The parameters ωj and νj are then computed by solving
Problem (50), and replacing the robust constraint (52) with
a SOS constraint (see Problem (50)). Note that the same
principles of Algorithm 2 and of the SOS-based relaxation
discussed in this section can be used to compute, instead of
an half-space Hj , a more complex semialgebraic set (e.g.,
an ellipsoid) described by the polynomial inequality:
ω ⊤ q(x̃) ≤ 0
∀ x̃ ∈ X ,
Fig. 4. Final polytope after running Algorithm 2.
(53)
• Approximation of the robust constraint ν − ω⊤ x ≥
0 ∀x ∈ X with the
Pmconvex conservative constraint
ν − ω ⊤ x = σ0 (x̃) − s=1 σs (x̃)hs (x̃).
with q(x̃) being a vector of monomials in the variable x̃. The
parameters ω can be then computed by properly modifying
the SOS-relaxed Problem (50). For instance, in case we are
interested in computing an ellipsoidal outer approximation
of X , the function ω ⊤ q(x̃) should have a quadratic form,
and its Hessian should be enforced to be positive definite.
The latter source of approximation can be reduced by increasing the degree 2d of the SOS polynomials. In fact,
as already discussed in Section 5.1, according to the Putinar’s Positivstellensatz each function ν − ω ⊤ x such that
ν − ω ⊤ xP≥ 0 ∀x ∈ X can be written as ν − ω⊤ x =
σ0 (x̃) − m
s=1 σs (x̃)hs (x̃) provided that the degree of the
SOS polynomials σ0 , σ1 , . . . , σm is large enough. On the
other hand, there is no theoretical result concerning the accuracy of the approximation of the indicator functions in
Problem (43) with the convex functions R{Hj } (pi ) appearing in Problem (51). Because of that, the polytope S obtained
by solving convex problems (51) (for j = 1, . . . , J) is not
guaranteed to minimize the original nonconvex optimization
problem (42). Algorithm 3 can then be used to refine the
polytopic outer approximation S provided by Algorithm 2.
Example 4 Let us continue with Example 2. Fig. 4 shows
the polytope obtained by applying Algorithm 2 solving Problems (51) instead of the nonconvex optimization in A2.2.
The SDP Problems (51) are solved for a degree of the SOS
polynomials equal to 2d = 4. The solution is a polytope
S that outer-bounds X1 . It can be observed that because
of the approximations introduced (SOS and the approximation of the indicator functions), which are necessary to efficiently solve the optimizations, the half-spaces bounding
X1 are not tangent to it and the computed region S still include two black points. Therefore, the computed polytope is
not the minimum-volume polytope. However, it is already a
very good outer-approximation of it. In the next section, we
describe a further refinement of Algorithm 2 aiming to computing a tighter polytope S. According to the steps A1.1.3
and A1.1.4 of Algorithm 1, we outer-approximate M (and
so X1 ) with S. At the next time step (k = 2) of the setmembership filter, we repeat the procedure to compute a new
polytope outer-bounding X2 . The difference is now that instead of h1 (·) in (36), we have the 9 linear inequalities that
define the polytope in Fig. 4. This procedure is repeated recursively in time.
The main principle of Algorithm 3 is to process, one by one,
all the points belonging to the polytopic outer-approximation
S initially given by Algorithm 2. For each of such points
⊤
pi , an half-space Hi : ω∗i x̃ − νi∗ ≤ 0 including the set
X (i.e., X ⊆ Hi ) and at the same not containing the point
⊤
pi (i.e., pi 6∈ Hi , or equivalently −ω∗i pi + νi∗ < 0) is
seeked. In this way, all the points pi which do not belong to
the minimum volume polytopic outer approximation of X
are discarded. Thus, a tighter (but more complex) polytopic
outer approximation of X is obtained.
6.5 Refinement of the polytope S
An important feature enjoyed by the refined polytope S ∗ is
given by the following theorem.
Summarizing, an approximate solution of the robust optimization problem (43) is computed by solving the convex
SDP problems (51), and, on the basis of Algorithm 2, the
polytopic-outer approximation S of the set M is then defined as S = B ∩ H1 ∩ . . . ∩ HJ .
Note that, in solving (51) instead of (43), two different
sources of approximation are introduced:
Theorem 7 The polytope S ∗ computed with Algorithm 3 is
a global minimizer of problem (42).
Proof: Let S̃ be a polytope belonging to the set of feasibility
of problem (42) (i.e., M ⊆ S̃) which does not minimize
(42). This means that there exists a polytope S̃˜ such that
M ⊆ S̃˜ ⊆ S̃ and a point p̄ given as input of Algorithm 2
• Approximation of the indicator functions I{Hj } (pi ) with
the convex functions R{Hj } (pi ) (see Fig. 3);
14
Algorithm 3: Refinement of the polytope S
[input] Sequence of the random points pi provided as input
of Algorithm 2 and such that pi ∈ S. Let Ñ be the number
of points pi belonging to S.
A3.1 S ∗ ← S
A3.2 for i = 1 : Ñ
A3.2.1 Compute the solution of the following optimization problem
ω∗i , νi∗ =arg
min −ω⊤ pi + ν
ω ∈ Rn
ν∈R
(54)
s.t.
ω 6= 0
Fig. 5. Exampe 1: hyperplanes defining the polytope S1∗ (black
lines) and true state uncertainty set X1 (gray region).
ν − ω ⊤ x ≥ 0 ∀x ∈ X .
problems (51) in the format required by the used SDP solver.
7 Numerical examples
A3.2.2 S ∗ ← S ∗ ∩ Hi .
[output] Polytope S ∗ .
Let us consider the discrete-time Lotka Volterra preypredator model [63] described by the difference equations:
˜ Thus, for p = p̄, the optimal
such that: p̄ ∈ S̃ and p̄ 6∈ S̃.
i
⊤
solution {ω∗i , νi∗ } of Problem (54) is such that ω ∗i pi −νi∗ >
⊤
0. Let Hi be the half-space defined as Hi : ω ∗i x − νi∗ ≤ 0.
Obviously, p̄ 6∈ Hi . Besides, the output S ∗ of Algorithm 3
is contained in the hyperspace Hi . Therefore, since p̄ 6∈ Hi
and S ∗ ⊆ Hi , it follows that the point p̄ 6∈ S ∗ . Then, a
polytope S̃ that does not minimize the optimization problem
(42) can not be the output of Algorithm 3.
x1(k)=x1 (k−1)(r+1−rx1 (k−1)−bx2 (k−1))+w1 (k−1),
x2(k)=cx1 (k−1)x2 (k−1) + (1 − d)x2 (k−1)+w2 (k−1),
(55)
where x1 (k) and x2 (k) denote the prey and the predator
population size, respectively. In the example, the following
values of the parameters are considered: r = 0.25, b = 0.95,
c = 1.1 and d = 0.55. The observed output is the sum of
the population of the prey and predator densities, i.e.,
Theorem 7 mainly says that there exists no polytope including M and containing less randomly generated points pi
than S ∗ . However, it is worth remarking that only an approximated solution of Problem (54) can be computed, as
the robust constraint ν − ω ⊤ x ≥ 0 ∀x ∈ X appearing in
(54) has to be handled with the SOS-based techniques described in the previous section. Thus, conservativeness could
be added at this step. Therefore, the main interpretation to
be given to Theorem 7 is that Algorithm 3 cancels the effect
of approximating the indicator function I{Hj } (pi ) with the
convex function R{Hj } (pi ).
y(k) = x1 (k) + x2 (k) + v(k),
(56)
where the measurement noise v(k) is bounded and such that
kv(k)k∞ ≤ 0.05. The initial prey and predator sizes x(0) =
⊤
[x1 (0) x2 (0)] are known to belong to the box X0 =
[0.28 0.32] × [0.78 0.82] and the noise process w(k) =
[w1 (k) w2 (k)]⊤ is bounded by kw(k)k∞ ≤ 0.001. The
data are obtained by simulating the model with initial conditions x1 (0) = 0.8 and x2 (0) = 0.3, and by corrupting
the output observations with a random noise v(k) uniformly
distributed within the interval [−0.05 0.05].
Example 5 Let us continue with Example 2. Fig. 5 shows
the computed polytope S1∗ , along with the true state uncertainty set X1 . The CPU taken by the proposed algorithm to
compute the 54 hyper-spaces that define the polytope S1∗ is
about 830 seconds. However, only 80 out of 830 seconds
are spent by the solver SeDuMi to solve 108 (i.e., 54 × 2)
SDP problems of the type (51). The other 750 seconds are
required by the SOStools interface to formulate, 108 times,
the SDP problems (51) in the format used by SeDuMi. Therefore, the computational time required to compute the polytope S1∗ can be drastically reduced not only by using more
efficient SDP solvers, but also directly formulating the SDP
Polytopic outer approximations Sk∗ of the state uncertainty
sets Xk (with k = 1, . . . , 40) are computed through Algorithm 2. N = 20 random points are used to approximate the
volume of the polytope Sk∗ (as described in Section 6.1). In
order to limit the complexity in the description of the polytopes Sk∗ , the maximum number of halfspaces describing Sk∗
is set to 8. This means that Algorithm 2 is stopped after at
most 4 iterations (we remind that the initial outer-bounding
box Bk is already described by 4 half-spaces). When the
15
0.4
0.4
0.3
0.3
x2
0.5
x2
0.5
0.2
0.2
0.1
0.1
0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0
0.1
0.8
0.2
0.3
x1
0.4
0.5
0.6
0.7
0.8
x1
Fig. 6. Example 2: outer-bounding polytopes (gray) and true state
trajectory (black dots).
Fig. 7. Example 2: outer-bounding boxes (gray) and true state
trajectory (black dots).
output of Algorithm 2 is a polytope Sk∗ described by less
than 8 half-spaces, Algorithms 3 is used to refine the polytopic outer approximation Sk∗ . Fig. 6 shows the computed
polytopes Sk∗ outer approximating the state uncertainty sets
Xk (with k = 1, . . . , 40), along with the true state trajectory.
The Hybrid toolbox [64] has been used to plot the polytopes in Fig. 6. The average CPU time required to compute
a polytope Sk∗ is 28 seconds (not including the time required
by the SOStools interface to formulate the SDP problems
(51) in the format used by the solver SeDuMi). For the sake
of comparison, Fig. 7 shows the outer-bounding approximations of the state uncertainty sets Xk when boxes, instead of
polytopes, are propagated over time. For a better comparison, in Fig. 8 the bounds on the time-trajectory of each state
variable obtained by propagating boxes and polytopes are
plotted. The obtained results show that, as expected, propagating polytopic uncertainty sets instead of boxes provides
a more accurate state estimation. Finally, we would like to
remark that a small uncertainty on the noise process is assumed (i.e., kw(k)k∞ ≤ 0.001) since, for larger bounds on
kw(k)k∞ , it would not be possible to clearly visualize the
uncertainty boxes in Fig. 7.
0.9
0.8
0.7
x1 , x2
0.6
x1
0.5
0.4
0.3
0.2
x2
0.1
0
0
10
20
30
40
time
Fig. 8. Example 2: bounds on state trajectories obtained by propagating boxes (black line); bounds on state trajectories obtained
by propagating polytopes (gray line); true state trajectory (black
dots).
lem can be obtained by using the theory of sum-of-squares
polynomial optimization. We have finally derived a procedure to compute a polytopic outer-approximation of the true
membership-set, by computing the minimum-volume polytope that outer-bounds the set that includes all the means
computed with respect to P. It is worth remarking that the
set-membership filtering approach discussed in the paper can
be extended to handle noise-corrupted input signal observations and uncertainty in the model parameters, provided
that the corresponding state uncertainty set Xk remains a
semi-algebraic set. As future works, we aim first to speed up
the proposed state estimation algorithm in order to be able
to use it in real-time applications in systems with fast dynamics. To this aim, dedicated numerical algorithms, written
in Fortran and C++, for solving the formulated SDP opti-
8 Conclusions
In this paper we have shown that set-membership estimation can be equivalently formulated in a probabilistic setting by employing sets of probability measures. Inferences
in set-membership estimation are thus carried out by computing expectations with respect to the updated set of probability measures P, as in the probabilistic case, and they can
be formulated as a semi-infinite linear programming problem. We have further shown that, if the nonlinearities in the
measurement and process equations are polynomial and if
the bounding sets for initial state, process and measurement
noises are described by polynomial inequalities, then an approximation of this semi-infinite linear programming prob-
16
mization problems will be developed. Furthermore, the SDP
problems will be directly formulated in the format required
by the SDP solver, thus avoiding the use of interfaces like
SOStools. An open source toolbox will be then released.
Second, by exploiting the probabilistic interpretation of setmembership estimation, we plan to reformulate it using the
theory of moments developed by Lasserre. This will allow
us to ground totally set-membership estimation in the realm
of the probabilistic setting, which will give us the possibility
of combining the two approaches in order to obtain hybrid
filters, i.e., filters that include both classical probabilistic uncertainties and set-membership uncertainties.
[15] H. Piet-Lahanier and E. Walter, “Further results on recursive
polyhedral description of parameter uncertainty in the bounded-error
context,” in Proceedings of the 28th IEEE Conference on Decision
and Control, Tampa, Florida, USA, pp. 1964 –1966, 1989.
[16] S. Mo and J. Norton, “Fast and robust algorithm to compute
exact polytope parameter bounds,” Mathematics and computers in
simulation, vol. 32, no. 5-6, pp. 481–493, 1990.
[17] V. Broman and M. Shensa, “A compact algorithm for the intersection
and approximation of N-dimensional polytopes,” Mathematics and
computers in simulation, vol. 32, no. 5-6, pp. 469–480, 1990.
[18] L. Chisci, A. Garulli, and G. Zappa, “Recursive state bounding by
parallelotopes,” Automatica, vol. 32:7, pp. 1049–1055, 1996.
[19] L. Chisci, A. Garulli, A. Vicino, and G. Zappa, “Block
recursive parallelotopic bounding in set membership identification,”
Automatica, vol. 34, no. 1, pp. 15–22, 1998.
References
[20] V. T. H. Le, C. Stoica, T. Alamo, E. F. Camacho, and D. Dumur,
Zonotopes: From Guaranteed State-estimation to Control. John Wiley
& Sons, 2013.
[1] M. Milanese and A. Vicino, “Optimal estimation theory for dynamic
sistems with set membership uncertainty: an overview,” Automatica,
vol. 27, no. 6, pp. 997–1009, 1991.
[21] V. Puig, J. Saludes, and J. Quevedo, “Worst-case simulation of
discrete linear time-invariant interval dynamic systems,” Reliable
Computing, vol. 9, no. 4, pp. 251–290, 2003.
[2] P. L. Combettes, “The foundations of set theoretic estimation,”
Proceedings of the IEEE, vol. 81, pp. 182–208, Feb 1993.
[22] C. Combastel, “A state bounding observer based on zonotopes,” in
European Control Conference, 2003.
[3] M. Milanese, J. Norton, H. Piet-Lahanier, and E. Walter, eds.,
Bounding approaches to system identification. New York: Plenum
Press, 1996.
[23] V. T. H. Le, T. Alamo, E. F. Camacho, C. Stoica, and D. Dumur,
“A new approach for guaranteed state estimation by zonotopes,” in
Proceedings of the 18th IFAC World Congress, Milano, Italy, vol. 28,
2011.
[4] M. Milanese and C. Novara, “Unified set membership theory
for identification, prediction and filtering of nonlinear systems,”
Automatica, vol. 47, no. 10, pp. 2141–2151, 2011.
[5] V. Cerone, D. Piga, and D. Regruto, “Set-membership error-invariables identification through convex relaxation techniques,” IEEE
Transactions on Automatic Control, vol. 57, pp. 517–522, 2012.
[24] T. Alamo, J. Bravo, and E. Camacho, “Guaranteed state estimation
by zonotopes,” in Proceedings of the 42nd IEEE Conference on
Decision and Control, Maui, Hawaii, USA, pp. 5831 – 5836, dec.
2003.
[6] M. Casini, A. Garulli, and A. Vicino, “Feasible parameter set
approximation for linear models with bounded uncertain regressors,”
IEEE Transactions on Automatic Control, vol. 50, no. 11, pp. 2910–
2920, 2014.
[25] G. Calafiore, “Reliable localization using set-valued nonlinear filters,”
IEEE Transactions on Systems, Man and Cybernetics, Part A: Systems
and Humans, vol. 35, no. 2, pp. 189–197, 2005.
[26] L. El Ghaoui and G. Calafiore, “Robust filtering for discrete-time
systems with bounded noise and parametric uncertainty,” IEEE
Transactions on Automatic Control, vol. 46, no. 7, pp. 1084–1089,
2001.
[7] F. C. Schweppe, “Recursive state estimation: Unknown but bounded
errors and system inputs,” in Adaptive Processes, Sixth Symposium
on, vol. 6, pp. 102 –107, oct. 1967.
[8] D. Bertsekas and I. Rhodes, “Recursive state estimation for a
set-membership description of uncertainty,” IEEE Transactions on
Automatic Control, vol. 16, pp. 117 – 128, apr 1971.
[27] C. Maier and F. Allgöwer, “A set-valued filter for discrete
time polynomial systems using sum of squares programming,” in
Proceedings of the 48th IEEE Conference on Decision and Control,
Shanghai, China, pp. 223–228, 2009.
[9] V. Kuntsevich and M. Lychak, Guaranteed estimates, adaptation and
robustness in control systems. Springer-Verlag, 1992.
[28] F. Dabbene, D. Henrion, C. Lagoa, and P. Shcherbakov, “Randomized
approximations of the image set of nonlinear mappings with
applications to filtering,” in IFAC Symposium on Robust Control
Design (ROCOND 2015), Bratislava (Russia), 2015.
[10] A. Savkin and I. Petersen, “Robust state estimation and model
validation for discrete-time uncertain systems with a deterministic
description of noise and uncertainty,” Automatica, vol. 34, no. 2,
pp. 271–274, 1998.
[29] A. Benavoli, “The generalized moment-based filter,” IEEE
Transactions on Automatic Control, vol. 58, no. 10, pp. 2642–2647,
2013.
[11] E. Fogel and Y. Huang, “On the value of information in system
identificationbounded noise case,” Automatica, vol. 18, no. 2, pp. 229
– 238, 1982.
[30] A. Benavoli, M. Zaffalon, and E. Miranda, “Robust filtering through
coherent lower previsions,” IEEE Transactions on Automatic Control,
2011.
[12] J. R. Deller and S. F. Odeh, “Implementing the optimal bounding
ellipsoid algorithm on a fast processor,” in Acoustics, Speech, and
Signal Processing, 1989. ICASSP-89., 1989 International Conference
on, pp. 1067–1070, IEEE, 1989.
[13] J. R. Deller and T. C. Luk, “Linear prediction analysis of speech
based on set-membership theory,” Computer Speech & Language,
vol. 3, no. 4, pp. 301 – 327, 1989.
[31] C. Combastel, “Merging kalman filtering and zonotopic state
bounding for robust fault detection under noisy environment,” IFACPapersOnLine, vol. 48, no. 21, pp. 289 – 295, 2015. 9th {IFAC}
Symposium on Fault Detection, Supervision and Safety for Technical
Processes {SAFEPROCESS}, Paris, 2015.
[14] J. R. Deller, M. Nayeri, and M. S. Liu, “Unifying the
landmark developments in optimal bounding ellipsoid identification,”
International Journal of Adaptive Control and Signal Processing,
vol. 8, no. 1, pp. 43–60, 1994.
[32] R. Fernandez-Canti, J. Blesa, and V. Puig, “Set-membership
identification and fault detection using a bayesian framework,” in
Control and Fault-Tolerant Systems (SysTol), 2013 Conference on,
pp. 572–577, Oct 2013.
17
[54] E. Allgöwer and P. Schmidt, “Computing volumes of polyhedra.,”
Math. Comput., vol. 46, no. 173, pp. 171–174, 1986.
[33] A. Gning, L. Mihaylova, and F. Abdallah, “Mixture of uniform
probability density functions for non linear state estimation using
interval analysis,” in Information Fusion (FUSION), 2010 13th
Conference on, pp. 1–8, 2010.
[55] J. B. Lasserre, “An analytical expression and an algorithm for the
volume of a convex polyhedron in Rn ,” Journal of optimization
theory and applications, vol. 39, no. 3, pp. 363–377, 1983.
[34] A. Gning, B. Ristic, and L. Mihaylova, “A box particle filter
for stochastic and set-theoretic measurements with association
uncertainty,” in Information Fusion (FUSION), 2011 Proceedings of
the 14th International Conference on, pp. 1–8, 2011.
[56] A. Bemporad, C. Filippi, and F. Torrisi, “Inner and outer
approximations of polytopes using boxes,” Computational Geometry,
vol. 27, no. 2, pp. 151–178, 2004.
[35] J. Shohat and J. Tamarkin, The problem of moments. American
Mathematical Society, 1950.
[57] J. Lawrence, “Polytope volume computation,” Mathematics of
Computation, vol. 57, no. 195, pp. 259–271, 1991.
[36] M. Krein and A. Nudelḿan, The Markov moment problem and
extremal problems, vol. 50. Amer Mathematical Society, 1977.
[58] J. B. Lasserre and E. S. Zeron, “A Laplace transform algorithm for
the volume of a convex polytope,” Journal of the ACM, vol. 48,
no. 6, pp. 1126–1140, 2001.
[37] J. Lasserre, Moments, positive polynomials and their applications,
vol. 1 of Imperial College Press Optimization Series. World
Scientific, 2009.
[59] R. Smith, “Efficient Monte Carlo procedures for generating points
uniformly distributed over bounded regions,” Operations Research,
pp. 1296–1308, 1984.
[38] V. Cerone, D. Piga, and D. Regruto, “Polytopic outer approximations
of semialgebraic sets,” in IEEE 51st Annual Conference on Decision
and Control, Maui, Hawaii, USA, pp. 7793–7798, 2012.
[60] M. Dyer, A. Frieze, and R. Kannan, “A random polynomial-time
algorithm for approximating the volume of convex bodies,” Journal
of the ACM, vol. 38, no. 1, pp. 1–17, 1991.
[39] A. Karr, “Extreme points of certain sets of probability measures, with
applications,” Mathematics of Operations Research, vol. 8, no. 1,
pp. 74–85, 1983.
[61] S. Wiback, I. Famili, H. Greenberg, and B. Palsson, “Monte Carlo
sampling can be used to determine the size and shape of the steadystate flux space,” Journal of theoretical biology, vol. 228, no. 4,
pp. 437–447, 2004.
[40] A. Shapiro, “On duality theory of conic linear problems,” in in SemiInfinite Programming: Nonconvex Optimization and Its Applications,
pp. 135–165, 2001.
[62] C. Robert and G. Casella, Monte Carlo statistical methods. Springer
Science, 2004.
[41] P. Walley, Statistical Reasoning with Imprecise Probabilities. New
York: Chapman and Hall, 1991.
[63] M. R. S. Raj, A. G. M. Selvam, and R. Janagaraj, “Stability in a
discrete prey-predator model,” Internation Journal of Latest Research
in Science and Technology, vol. 2, no. 1, pp. 482–485, 2013.
[42] A. Benavoli and M. Zaffalon, “Density-ratio robustness in dynamic
state estimation,” Mechanical Systems and Signal Processing, vol. 37,
no. 12, pp. 54 – 75, 2013.
[64] A. Bemporad, “Hybrid Toolbox - User’s Guide,” 2004.
http://cse.lab.imtlucca.it/ bemporad/hybrid/toolbox.
[43] J. B. Lasserre, “Global optimization with polynomials and the
problem of moments,” SIAM Journal on Optimization, vol. 11,
pp. 796–817, 2001.
[44] P. Parrilo, “Semidefinite programming relaxations for semialgebraic
problems,” Mathematical Programming, vol. 96, pp. 293–320, 2003.
[45] G. Chesi, A. Garulli, A. Tesi, and A. Vicino, “Solving quadratic
distance problems: an LMI-based approach,” IEEE Trans. Automatic
Control, vol. 48, no. 2, pp. 200–212, 2003.
[46] M. Putinar, “Positive polynomials on compact semi-algebraic sets,”
Indiana University Mathematics Journal, vol. 42, pp. 969–984, 1993.
[47] M. Laurent, “Sums of squares, moment matrices and optimization
over polynomials,” Emerging Applications of Algebraic Geometry,
Vol. 149 of IMA Volumes in Mathematics and its Applications, M.
Putinar and S. Sullivant (eds.), pp. 157–270, 2009.
[48] A.
Papachristodoulou,
J. Anderson, G. Valmorbida, S. Prajna, P. Seiler, and P. A. Parrilo,
SOSTOOLS: Sum of squares optimization toolbox for MATLAB, 2013.
url: http://www.eng.ox.ac.uk/control/sostools.
[49] J. F. Sturm, “Using SeDuMi 1.02, a MATLAB Toolbox for
optimization over symmetric cones,” Optim. Methods Software,
vol. 11, no. 12, pp. 625–653, 1999.
[50] M. Dyer and A. Frieze, “On the complexity of computing the volume
of a polyhedron,” SIAM Journal on Computing, vol. 17, pp. 967–
974, 1988.
[51] B. Bueler, A. Enge, and K. Fukuda, “Exact volume computation for
polytopes: a practical study,” in DMV SEMINAR, vol. 29, pp. 131–
154, Springer, 2000.
[52] S. Arora and B. Barak, Computational complexity: a modern
approach. Cambridge University Press, 2009.
[53] J. Cohen and T. Hickey, “Two algorithms for determining volumes
of convex polyhedra,” Journal of the ACM, vol. 26, no. 3, pp. 401–
414, 1979.
18
url:
0.4
yo − ŷ
0.2
0
−0.2
−0.4
20
40
60
80
100 120 140 160 180 200
Sample
0.4
yo − ŷ
0.2
0
−0.2
−0.4
20
40
60
80
100 120 140 160 180 200
Sample
| 3cs.SY
|
Evolving Deep Convolutional Neural Networks
for Image Classification
Yanan Sun, Bing Xue and Mengjie Zhang
arXiv:1710.10741v2 [cs.NE] 31 Oct 2017
School of Engineering and Computer Science, Victoria University of Wellington
PO Box 600, Wellington 6140, New Zealand
e-mail: {Yanan.Sunm, Bing.Xue, Mengjie.Zhang}@ecs.vuw.ac.nz
where X and Y are the input data and corresponding label,
respectively, F (·) denotes the architecture choosing function
with the given data, G(·) refers to the initialization method
of the connection weights Weight based on the chosen architecture, and L(·) measures the differences between the
true label and the label predicted by the CNN using X
and Weight . Typically, the Gradient Descend (GD)-based approaches, e.g., Stochastic GD (SGD), are utilized to minimize
L(X, Weight , Y ) within the given number of epochs, where
the connection weight values are iteratively updated. Although
L(·) is not differentiable in all occasions, GD-based methods
are preferred due to their effectiveness and good scalability
as the number of connection weights increases. A CNN
commonly has a huge number of connection weights. However,
F (·) and G(·) are countable functions that are discrete and
neither convex or concave, and they are not well addressed in
practice. Furthermore, because the gradient-based optimizers
are heavily dependent on the initial values of the weights
(including biases), it is essential to choose a suitable G(·)
that can help the consecutive GD-based approaches to escape
from local minima. Furthermore, the performance of assigned
architectures cannot be evaluated until the minimization of L(·)
is finished, while the minimization is a progress of multiple
iterations, which in turn increases the difficulty of choosing the
I. I NTRODUCTION
potential optimal F (·). Therefore, the architecture design and
Convolutional neural networks (CNNs) have demonstrated
connection weight initialization strategy should be carefully
their exceptional superiority in visual recognition tasks, such
treated in CNNs.
as traffic sign recognition [1], biological image segmentaTypically, most of the architecture design approaches were
tion [2], and image classification [3]. CNNs are originally
initially developed for the deep learning algorithms in the early
motivated by the computational model of the cat visual cortex
date (e.g., the Stacked Auto-Encoders (SAEs) [10], [11] and
specializing in processing vision and signal related tasks [4].
the Deep Belief Networks (DBNs) [12]), such as Grid Search
Since LetNet-5 was proposed in 1989 [5], [6], which is an
(GS), Random Search (RS) [13], Bayesian-based Gaussion
implementation of CNNs and whose connection weights are
Process (BGP) [14], [15], Tree-structured Parzen Estimators
optimized by the Back-Propagation (BP) algorithm [7], various
(TPE) [16], and Sequential Model-Based Global Optimizavariants of CNNs have been developed, such as VGGNet [8]
tion (SMBO) [17]. Theoretically, GS exhaustively tests all
and ResNet [9]. These variants significantly improve the clascombinations of the parameters to expectedly seize the best
sification accuracies of the best rivals in image classification
one. However, GS cannot evaluate all combinations within an
tasks. Diverse variants of CNNs differ from their architectures
acceptable time in reality. Moreover, GS is difficult to optimize
and weight connections.
the parameters of continuous values [13]. RS could reduce the
Mathematically, a CNN can be formulated by (1) in the
exhaustive adverse of GS, but the absolute “random” severely
context of an image classification task with the input (X, Y ),
challenges the sampling behavior in the search space [18]–[20].
In addition, BGP incurs extra parameters (i.e., the kernels) that
Architecture = F (X, Y )
Weight = G(Architecture )
(1) are arduous to tune. TPE treats each parameter independently,
while the most key parameters in CNNs are dependent (e.g.,
minimize L(X, Weight , Y )
Abstract—Evolutionary computation methods have been successfully applied to neural networks since two decades ago, while
those methods cannot scale well to the modern deep neural networks due to the complicated architectures and large quantities
of connection weights. In this paper, we propose a new method
using genetic algorithms for evolving the architectures and
connection weight initialization values of a deep convolutional
neural network to address image classification problems. In the
proposed algorithm, an efficient variable-length gene encoding
strategy is designed to represent the different building blocks and
the unpredictable optimal depth in convolutional neural networks.
In addition, a new representation scheme is developed for effectively initializing connection weights of deep convolutional neural
networks, which is expected to avoid networks getting stuck into
local minima which is typically a major issue in the backward
gradient-based optimization. Furthermore, a novel fitness evaluation method is proposed to speed up the heuristic search
with substantially less computational resource. The proposed
algorithm is examined and compared with 22 existing algorithms
on nine widely used image classification tasks, including the stateof-the-art methods. The experimental results demonstrate the
remarkable superiority of the proposed algorithm over the stateof-the-art algorithms in terms of classification error rate and the
number of parameters (weights).
Index Terms—Genetic algorithms, convolutional neural network, image classification, deep learning.
the convolutional layer size and its stride, more details can be
seen in Subsection II-A). The methods mentioned above have
shown their good performance in most SAEs and DBNs, but
are not suitable to CNNs. Their success in SAEs and DBNs is
due largely to the architecture construction approaches, which
are greedy layer-wised by stacking a group of building blocks
with the same structures (i.e., the three-layer neural networks).
In each building block, these architecture-search methods are
utilized for only optimizing the parameters, such as the number
of neurons in the corresponding layer. However in CNNs, the
layer-wised method cannot be applied due to their architecture
characteristics of non-stacking routine, and we must confirm
the entire architectures at a time. Furthermore, multiple different building blocks exist in CNNs, and different orders
of them would result in significantly different performance.
Therefore, the architecture design in CNNs should be carefully
treated. Recently, Baker et al. [21] proposed an architecture
design approach for CNNs based on reinforcement learning,
named MetaQNN, which employed 10 Graphic Processing
Unit (GPU) cards with 8-10 days for the experiments on the
CIFAR-10 dataset.
Due to the drawbacks of existing methods and limited
computational resources available to interested researchers,
most of these works in CNNs are typically performed by
experts with rich domain knowledge [13]. Genetic Algorithms
(GAs), which are a paradigm of the evolutionary algorithms
that do not require rich domain knowledge [22], [23], adapt
the meta-heuristic pattern motivated by the process of natural
selection [24] for optimization problems. GAs are preferred in
various fields due to their characteristics of gradient-free and
insensitivity to local minima [25]. These promising properties
are collectively achieved by a repeated series of the selection, mutation, and crossover operators. Therefore, it can be
naturally utilized for the optimization of architecture design
and the connection weight initialization for CNNs. Indeed,
GAs for evolving neural networks can be traced back to
1989 [26]. In 1999, Yao [25] presented a survey about these
different approaches, which are largely for the optimization
of connection weights in the fixed architecture topologies.
In 2002, Stanley and Miikkulainen proposed the NeuronEvolution Augmenting Topology (NEAT) [27] algorithm to
evolve the architecture and connection weights of a small
scale neural network. Afterwards, the HyperNEAT [28], i.e.,
NEAT combined with the compositional pattern producing
networks [29], was proposed to evolve a larger scale neural
network with an indirect encoding strategy. Motivated by
the HyperNEAT, multiple variants [30]–[32] were proposed
to evolve even larger scale neural networks. However, the
major deficiencies of the HyperNEAT-based approaches are:
1) they are only suitable for evolving deep neural networks
with global connection and single building blocks, such as
SAEs and DBNs, but not CNNs where local connection exists
and multiple different building blocks need to be evolved
simultaneously, and 2) hybrid weigh connections (e.g., connections between two layers that are non-adjacent) may be
evolved, which are contrast to the architectures of CNNs.
Indeed, the views have been many years that evolutionary
algorithms are incapable of evolving the architecture and
connection weights in CNNs due to the tremendous number
of related parameters [33]–[35]. Until very recently in 2017,
Google showed their Large Evolution for Image Classification
(LEIC) method specializing at the architecture optimization
of CNNs [36]. LETC is materialized by GAs without the
crossover operator, implemented on 250 high-end computers,
and archives competitive performance against the state-ofthe-art on the CIFAR-10 dataset by training for about 20
days. Actually, by directly using GAs for the architecture
design of CNNs, several issues would raise in nature: 1)
the best architecture is unknown until the performance is
received based on it. However, evaluating the performance of
one individual takes a long time, and appears more severely
for the entire population. This would require much more
computational resources for speeding up the evolution; 2)
the optimal depth of CNNs for one particular problem is
unknown, therefore it is hard to constrain the search space
for the architecture optimization. In this regard, a variablelength gene encoding strategy may be the best choice for
both 1) and 2), but how to assign the crossover operation for
different building blocks is a newly resulted problem; and 3)
the weight initialization values heavily affect the performance
of the confirmed architecture, but addressing this problem
involves a good gene encoding strategy and the optimization
of hundreds and thousands decision variables.
A. Goal
The aim of this paper is to design and develop an effective
and efficient GA method to automatically discover good architectures and corresponding connection weight initialization
values of CNNs (i.e., the first two formulae in (1)) without
manual intervention. To achieve this goal, the objectives below
have been specified:
1) Design a flexible gene encoding scheme of the architecture, which does not constrain the maximal length of
the building blocks in CNNs. With this gene encoding
scheme, the evolved architecture is expected to benefit
CNNs to achieve good performance in solving different
tasks at hand.
2) Investigate the connection weight encoding strategy,
which is capable of representing tremendous numbers of
the connection weights in an economy way. With this
encoding approach, the weight connection initialization
problem in CNNs is expected to be effectively optimized
by the proposed GA.
3) Develop associated selection (including the environmental selection), crossover, and mutation operators that can
cope with the designed gene encoding strategies of both
architectures and connection weights.
4) Propose an effective fitness measure of the individuals
representing different CNNs, which does not require
intensive computational resources.
Input
24x24
Feature maps
4@20x20
Feature maps
4@10x10
Feature maps
8@8x8
Feature maps
8@4x4
Full connection layer
128x1
Flatten
Convolution
Pooling
Convolution
Pooling
Fig. 1. An general architecture of the Convolutional Neural Network.
5) Investigate whether the new approach significantly outperform the existing methods in both classification accuracy
and number of weights.
B. Organization
the feature map and the input data through padding zeros,
the convolutional operations are categorized into two types:
the VALID (without padding) and the SAME (with padding).
Specifically, each element in the feature map is the sum of the
products of each element from the filter and the corresponding
elements this filter overlaps. If the input data is with multiple
channels, say 3, one feature map will also require 3 different
filters, and each filter convolves on each channel, then the
results are summed element-wised.
1
0
0
1
1
0
1
0
1
1
1
1
0
1
0
0
1
2
2
0
0
1
1
0
1
0
1
1
1
1
0
1
2
2
The reminder of this paper is organized as follows: the Filter {width:2, height:2}
2
2
1
1
1
1
1
1
1
1
1
Feature map
background of the CNNs, the related works on the architecture Stride {width:1, height:1}
Input 4x4
3x3
Input 4x4
design and weight initialization approaches of CNNs are
reviewed in Section II. The framework and the details of each
Fig. 2. An illustration of convolutional operation.
step in the proposed algorithm are elaborated in Section III.
The experiment design and experimental results of the proAn example of the VALID convolutional operation is illusposed algorithm are shown in Sections IV and V, respectively. trated by Fig. 2, where the filter has the same width and height
Next, further discussions are made in Section VI. Finally, the equal to 2, the input data is with the size of 4 × 4, and the
conclusions and future work are detailed in Section VII.
stride has the same height and width equal to 1. As shown in
Fig. 2, the generated feature map is with the size of 3 × 3, the
II. BACKGROUND AND R ELATED W ORK
shadow areas in the input data with different colors refer to
A. Architecture of Convolutional Neural Network
the overlaps with the filter at different positions of the input
1) Bone of CNNs: Fig. 1 exhibits an extensive architecture data, the shadow areas in the feature map are the respective
of one CNN, where there are two convolutional operations, resulted convolutional outcomes, and numbers in the filter are
two pooling operations, the resulted four groups of feature the connection weight values. Generally, convolutional results
maps, and the full connection layer in the tail. The last layer, regarding each filter are updated by adding a bias term and
which is a full connection layer, receives the input data by then through a nonlinearity, such as the Rectifier Linear Unit
flattening all elements of the fourth group of feature maps. (ReLU) [37], before they are stored into the feature map. ObviGenerally, the convolutional layers and the pooling layers can ously, the involved parameters in one convolutional operation
be mixed to stack together in the head of the architecture, are the filter width, the filter height, the number of feature
while the full connection layers are constantly stacked with maps, the stride width, the stride height, the convolutional
each other in the tail of the architecture. The numbers in Fig. 1 type, and the connection weight in the filter.
refer to the sizes of the corresponding layer. Particularly, the
3) Pooling: Intuitively, the pooling operation resembles the
input data is with 24 × 24, the output is with 128 × 1, and convolution operation except for the element-wised product
the other numbers denote the feature map configurations. For and the resulted values of the corresponding feature map.
example, 4@20 × 20 implies there are 4 feature maps, each Briefly, the pooling operation employs a predefined window
with the size of 4 × 4.
(i.e., the kernel) to collect the average value or the maximal
In the following, the details of the convolutional layer and value of the elements where it slides, and the slide size is
the pooling layer, which are associated with the convolution also called “stride” as in the convolutional operation. For
and the pooling operations, respectively, are documented in better understanding, an example of the pooling operation is
detail, while the full connection layer is not intended to illustrated by Fig. 3, where the kernel is with the size of 2 × 2,
describe here because it is well-known.
and the both stride width and height are 2, the input data
2) Convolution: Given an input image with the size of is with the size of 4 × 4, the shadows with different colors
n × n, in order to receive a feature map generated by the refer to the two slide positions and the resulted pooling values.
convolutional operations, a filter must be defined in advance. In this example, the maximal pooling operation is employed.
Actually, a filter (it can also be simply seen as a matrix) is Evidently in the pooling operation, the involved parameters
randomly initialized with a predefined size (i.e., the filter width are the kernel width, the kernel height, the stride width, the
and the filter height). Then, this filter travels from the leftmost stride height, and the pooling type.
to the rightmost of the input data with the step size equal to
a stride (i.e., the stride width), and then travels again after B. CNN Architecture Design Algorithm
In this subsection, only LEIC [36] is concerned due to its
moving downward with the step size equal to a stride (i.e.,
the stride height), until reach the right bottom of the input specific intention for the architecture design of CNNs. To this
image. Depending on whether to keep the same sizes between end, we will review the algorithm and then point out the
first initialization method. In the second initialization approach,
the shortage of the first one has been solved, but the major diffi0
1
1
1
0
1
1
1
1
1
culty exists in choosing the parameter of the distribution, such
1
1
0
1
1
1
0
1
1
1
as the range of the uniform distribution, and the mean value as
Kernel {width:2, height:2}
1
1
1
1
1
1
1
1
Stride {width:2, height:2}
Feature map
well as the standard derivation of the Gaussian distribution. To
Input 4x4
2x2
Input 4x4
solve this problem, the Xavier initializer presented a range for
uniform sampling based on the neuron saturation prior using
Fig. 3. An illustration of pooling operation.
the sigmoid activation function [39]. Supposed the number of
neurons in two adjacent layers are n1 and n2 , the values of
limitations, which are used to highlight the necessity of the the weights connecting
these two layers are initialized
within
i
h p
p
corresponding work in the proposed algorithm.
the range of − 6/(n1 + n2 ), 6/(n1 + n2 ) by uniformly
LEIC employed a GA to evolve the architectures of CNNs, sampling. Although the Xavier initializer works better than the
where individuals were evolved from scratch, and each individ- initialization methods from the other two categories, a couple
ual was encoded with a variable-length chromosome. During of major issues exist: 1) It highly relies on the architectures of
the phase of evolution, different kinds of layers can be incor- CNNs, particularity the number of neurons in each layer in the
porated into the individuals by the mutation operation, with networks (e.g., n and n in its formulation). If the optimal
1
2
the expectation that individuals with promising performance architectures of the networks are not found, the resulted
will be generated. Note that, the crossover operation, which initialized parameters perform badly as well, then there is no
are designed for local search in GAs, were not investigated way to evaluate the desired performance of the architectures
in the main part of LEIC. Without any crossover operator, and may mislead the adjustment of the architectures. 2) The
GAs typically require a large population size for reducing Xavier initializer is presented on the usage of the sigmoid
the adverse impact. Therefore, LETC adopted a population activation, while the widely used activation function in CNNs
size of 103 , while that of a magnitude with order 102 is is the ReLU [3], [9], [40], [41].
a general setting in the GA community. In addition, 250
To the best of our knowledge, there has not been any existhigh-end computers were employed for LEIC, which was ing evolutionary algorithm for searching for the connection
caused by the fitness assignment approach utilized in LEIC. weight initialization of deep learning algorithms including
In LEIC, each individual was evaluated with the final image CNNs. The main reason is the tremendous numbers of weights,
classification accuracy, which typically took a long time due to which are difficult to be effectively encoded into the chromothe iterative nature of GD-based optimizers. However, LEIC somes and to be efficiently optimized due to its large-scale
reported an external experiment on the crossover operation, global optimization nature [42].
which was used only for exchanging the mutation probabilities
III. T HE PROPOSED ALGORITHM
and the connection weights trained by SGD. Although it is
In this section, the proposed Evolving deep Convolutional
difficult to reach the actual reason why the crossover has not
been used for evolving the architectures in LETC, it is obvious Neural Networks (EvoCNN) for image classification is docuat least that the crossover operations are not easy to achieve mented in detail.
for chromosomes with different lengths. This would be more A. Algorithm Overview
complicated in CNNs that have multiple different building
blocks.
In summary, the main deficiency of LEIC is its high com- Algorithm 1: Framework of EvoCNN
putational complexity mainly caused by the fitness evaluation 1 P0 ←Initialize the population with the proposed gene
encoding strategy;
strategy it adopts as well as without crossover operations. The
2
t
← 0;
huge computational resource that LEIC requires makes it very
3
while
termination criterion is not satisfied do
intractable in academic environment.
4
Evaluate the fitness of individuals in Pt ;
C. Connection Weight Initialization
5
S ← Select parent solutions with the developed slack
binary tournament selection;
Typically, the initialization methods are classified into three
6
Q
t ← Generate offsprings with the designed genetic
different categorises. The first employs constant numbers to
operators
from S;
initialize all connection weights, such as the zero initializer,
7
P
←Environmental
selection from Pt ∪ Qt ;
t+1
one initializer, and other fixed value initializer. The second is
8
t
←
t
+
1;
the distribution initializer, such as using the Gaussion distribution or uniform distribution to initialize the weights. The third 9 end
covers the initialization approaches with some prior knowl- 10 Select the best individual from Pt and decode it to the
corresponding convolutional neural network.
edge, and the famous Xavier initializer [38] belongs to this
category. Because of the numerous connection weights existAlgorithm 1 outlines the framework of the proposed
ing in CNNs, it is not necessary that all the connection weights
start with the same values, which is the major deficiency of the EvoCNN method. Firstly, the population is initialized based on
max(1,0,1,1)
0
1
1
0
max(0,1,1,0)
0
1
1
0
C
P
C
C
C
P
C
P
C
Convolutional
C
C
P
F
...
F
F
F
...
P
...
C
Pooling
P
Full connection
F
Fig. 4. An example of three chromosomes with different lengths in EvoCNN.
TABLE I
T HE ENCODED INFORMATION IN E VO CNN.
Unit Type
convolutional
layer
pooling layer
full connection
layer
Encoded Information
the filter width, the filter height, the number of feature
maps, the stride width, the stride height, the convolutional type, the standard deviation and the mean value
of filter elements
the kernel width, the kernel height, the stride width, the
stride height, and the pooling type (i.e., the average or
the maximal)
the number of neurons, the standard deviation of
connection weights, and the mean value of connection
weights
the proposed flexible gene encoding strategy (line 1). Then, the
evolution begins to take effect until a predefined termination
criterion, such as the maximum number of the generations,
has been satisfied (lines 3-9). Finally, the best individual is
selected and decoded to the corresponding CNN (line 10) for
final training.
During the evolution, all individuals are evaluated first based
on the proposed efficient fitness measurement (line 4). After
that, parent solutions are selected by the developed slack
binary tournament selection (line 5), and new offspring are
generated with the designed genetic operators (line 6). Next,
representatives are selected from the existing individuals and
the generated offsprings to form the population in the next
generation to participate subsequent evolution (line 7). In
the following subsections, the key steps in Algorithm 1 are
narrated in detail.
B. Gene Encoding Strategy
Table I. Commonly, hundreds of thousands connection weights
may exist in one convolutional or full connection layer, which
cannot be all explicitly represented by a chromosome and
effectively optimized by GAs. Therefore, in EvoCNN, we use
only two statistical real numbers, the standard derivation and
mean value of the connection weights, to represent the numerous weight parameters, which can be easily implemented by
GAs. When the optimal mean value and the standard derivation
are received, the connection weights are sampled from the
corresponding Gaussian distribution. The details of population
initialization in EvoCNN are given in the next subsection
based on the gene encoding strategy introduced above.
C. Population Initialization
Algorithm 2: Population Initialization
Input: The population size N , the maximal number of
convolutional and pooling layers Ncp , and the
maximal number of full connection layers Nf .
Output: Initialized population P0 .
1 P0 ← ∅;
2 while |P0 | ≤ N do
3
part1 ← ∅;
4
ncp ← Uniformaly generate an integer between
[1, Ncp ];
5
while |part1 | ≤ ncp do
6
r ← Uniformly generated a number between
[0, 1];
7
if r ≤ 0.5 then
8
l ← Initialize a convolutional layer with
random settings;
9
else
10
l ← Initialize a pooling layer with random
settings;
11
end
12
part1 ← part1 ∪ l;
13
end
14
part2 ← ∅;
15
nf ← Uniformaly generate an integer between
[1, Nf ];
16
while |part2 | ≤ nf do
17
l ← Initialize a full connection layer with random
settings;
18
part2 ← part2 ∪ l;
19
end
20
P0 ← P0 ∪ (part1 ∪ part2 );
21 end
22 Return P0 .
As introduced in Subsection II-A, three different building
blocks, i.e., the convolutional layer, the pooling layer, and the
full connection layer, exist in the architectures of CNNs. Therefore, they should be encoded in parallel into one chromosome
for evolution. Because the optimal depth of a CNN in solving
one particular problem is unknown prior to confirming its
architecture, the variable-length gene encoding strategy, which
is very suitable for this occasion, is employed in the proposed
EvoCNN method. Furthermore, because the performance of
CNNs is highly affected by their depths [8], [43]–[45], this
For convenience of the elaboration, each chromosome is
variable-length gene encoding strategy makes the proposed
EvoCNN method have chances to reach the best result due separated into two parts. The first part includes the convolutional layers and the pooling layers, and the other part is the
to no constrains on the architecture search space.
In particular, an example of three chromosomes with dif- full connection layers. Based on the convention of the CNN
ferent lengths from EvoCNN is illustrated by Fig. 4, and all architectures, the first part starts with one convolutional layer.
represented information in these three layers are detailed in The second part can be added to only at the tail of the first
part. In addition, the length of each part is set by randomly evaluation in EvoCNN. Because EvoCNN concerns on solving
choosing a number within a predefined range.
image classification tasks, the classification error is the best
Algorithm 2 shows the major steps of the population initial- strategy to assign their fitness. The number of connection
ization, where |·| is a cardinality operator, lines 3-13 show the weights is also chosen as an additional indicator to measure
generation of the first part of one chromosome, and lines 14- the individual’s quality based on the principle of Occam’s
19 show that of the second part. During the initialization of razor [46].
the first part, a convolutional layer is added first. Then, a
With the conventions, each represented CNN is trained
convolutional layer or a pooling layer is determined by the on the training set Dtrain , and the fitness is estimated on
once coin tossing probability and then added to the end, which another dataset Df itness 1 . CNNs are frequently with deep
is repeated until the predefined length of this part is met. For architectures, thus thoroughly training them for receiving the
the second part, full connection layers are chosen and then final classification error would take considerable expenditure
added. Note here that, convolutional layers, pooling layers, of computing resource and a very long time due to the
and full connection layers are initialized with the random large number of training epochs required (>100 epochs are
settings, i.e., the information encoded into them are randomly invariably treated in fully training CNNs). This will make it
specified before they are stored into the corresponding part. much more impracticable here due to the population-based
After these two parts finished, they are combined and returned GAs with multiple generations (i.e., each individual will take a
as one chromosome. With the same approach, a population of full training in each generation). We have designed an efficient
individuals are generated.
method to address this concern. In this method, each individual
is trained with only a small number of epochs, say 5 or 10
D. Fitness Evaluation
epochs, based on their architectures and connection weight
initialization values, and then the mean value and the standard
derivation
of classification error are calculated on each batch
Algorithm 3: Fitness Evaluation
of
D
f itness in the last epoch. Both the mean value and the
Input: The population Pt , the training epoch number k
standard
derivation of classification errors, are employed as the
for measuring the accuracy tendency, the training
fitness
of
one individual. Obviously, the smaller mean value,
set Dtrain , the fitness evaluation dataset Df itness ,
the
better
individual.
When the compared individuals are with
and the batch size num of batch.
the
same
mean
values,
the less standard derivation indicates
Output: The population with fitness Pt .
the
better
one.
1 for each individual s in Pt do
In summary, three indicators are used in the fitness evalu2
i ← 1;
ation,
which are the mean value, standard derivation, and the
3
eval steps ← |Df itness |/num of batch;
number
of parameters. There are several motivations behind
4
while i ≤ k do
this
fitness
evaluation strategy: 1) It is sufficient to investigat5
Train the connection weights of the CNN
ing
only
the
tendency of the performance. If individuals are
represented by individual s;
with
better
performance
in the first several training epochs of
6
if i == k then
CNNs,
they
will
probably
still have the better performance in
7
accy list ← ∅;
the
following
training
epochs
with greater confidence. 2) The
8
j ← 1;
mean
value
and
the
standard
derivation are statistical signifi9
while j ≤ eval steps do
cance
indicators,
thus
suitable
for investigating this tendency,
10
accyj ← Evaluate the classification error
and
the
final
classification
error
can be received by optimizing
on the j-th batch data from Df itness ;
only
the
best
individual
evolved
by the proposed EvoCNN
11
accy list ← accy list ∪ accyj ;
method.
3)
CNN
models
with
less
number of connection
12
j ← j + 1;
weights
are
preferred
by
smart
devices
(more details are
13
end
discussed
in
Section
VI).
14
Calculate the number of parameters in s, the
mean value and standard derivation from
E. Slack Binary Tournament Selection
accy list, assign them to individual s, and
update s from Pt ;
We develop one slack version of the standard binary tour15
end
nament selection, which are documented in Algorithm 4, to
16
i ← i + 1;
select parent solutions for the crossover operations in the
17
end
proposed EvoCNN method. Briefly, two sets of comparisons
18 end
are employed. The comparisons between the mean values of
19 Return Pt .
individuals involves a threshold α, and that comparisons between the parameter numbers involves another threshold β. If
Fitness evaluation aims at giving a quantitative measure
determining which individuals qualify for serving as parent
solutions. Algorithm 3 manifests the framework of the fitness
1 The original training set is randomly split into D
train and Df itness ,
where Df itness is unseen to the CNN training phase, which can give a good
indication of the generalization accuracy on the test set.
units, i.e., the convolutional layer, the pooling layer, and the
full connection layer, are firstly collected into three different
lists based on their orders in the corresponding chromosome,
which refers to the Unit Collection (UC) phase. Then, these
three lists are aligned at the top, and units at the same positions
are performed the crossover. This phase is named the UA
and crossover phase. Finally, the Unit Restore (UR) phase is
employed, i.e., when the crossover operation is completed, the
units in these lists are restored to their original positions of
the associated chromosomes. With these three consequential
phases (i.e., the UC, the UA and crossover, and the UR), two
chromosomes with different lengths could easily exchange
their gene information for crossover. Because the crossover
operation is performed on the unit lists where only the units
with the same types are loaded, this proposed UA crossover
operation is natural (because they have the same origins).
For the remaining units, which do not perform crossover
operations due to no paired ones, are kept at the same position.
Mutation operations may perform on each position of the
units from one chromosome. For a selected mutation point, a
unit could be added, deleted, or modified, which is determined
by a probability of 1/3. In the case of unit addition, a convolutional layer, a pooling layer, or a full connection layer is added
by taking a probability of 1/3. If the mutation is to modify an
existing unit, the particular modification is dependent on the
unit type, and all the encoded information in the unit would be
changed (encoded information on each unit type can be seen in
Table I). Note that all the encoded formation is denoted by real
the parent solution cannot be selected with these comparisons,
numbers, therefore the Simulated Crossover (SBX) [47] and
the individual with smaller standard derivation is chosen.
the polynomial mutation [48] are employed in the proposed
In practice, tremendous number of parameters exist in deep
EvoCNN method due to their notable show in real number
CNNs, which would easily cause the overfitting problem.
gene representations.
Therefore, when the difference between the mean values of
two individuals is smaller than the threshold α, we further con- G. Environmental Selection
sider the number of connection weights. The slight change of
the parameter numbers will not highly affect the performance
Algorithm 5: Environmental Selection
of CNNs. Consequently, β is also introduced.
Input: The elistsm fraction γ, and the current population
By iteratively performing this selection, parent solutions
Pt ∪ Qt .
are selected and stored into a mating pool. In the proposed
Output:
The selected population Pt+1 .
EvoCNN method, the size of the mating pool is set to be the
1 a ← Calculate the number of elites based on γ and the
same of the population size.
predefined population size N from Algorithm 2;
F. Offspring Generation
2 Pt+1 ← Select a individuals that have the best mean
values from Pt ∪ Qt ;
The steps for generating offspring are given as follows:
3 Pt ∪ Qt ← Pt ∪ Qt − Pt+1 ;
step 1): randomly select two parent solutions from the mating
4 while |Pt+1 | < N do
pool;
5
s1 , s2 ← Randomly select two individuals from
step 2): use crossover operator on the selected solutions to
Pt ∪ Qt ;
generate offspring;
6
s ← Employe Algorithm 4 to select one individual
step 3): use mutation operator on the generated offspring;
from s1 and s2 ;
step 4): store offspring, remove the parent solutions from the
7
Pt+1 ← Pt+1 ∪ s;
mating pool, and perform steps 1-3 until the mating
8 end
pool is empty.
9 Return Pt+1 .
The proposed crossover operation can be seen in Fig. 5. To
achieve crossover, we design a method called Unit Alignment
(UA) for recombining two individuals with different chromoThe environmental selection is shown in Algorithm 5. Dursome lengths. During the crossover operation, three different ing the environmental selection, the elitism and the diversity
Algorithm 4: Slack Binary Tournament Selection
Input: Two compared individuals, the mean value
threshold α, and the paramemter number
threshold β.
Output: The selected individual.
1 s1 ← The individual with larger mean value;
2 s2 ← The other individual;
3 µ1 , µ2 ← The mean values of s1 , s2 ;
4 std1 , std2 ← The standard derivations of s1 , s2 ;
5 c1 , c2 ← The parameter numbers of s1 , s2 ;
6 if µ1 − µ2 > α then
7
Return s1 .
8 else
9
if c1 − c2 > β then
10
Return s2 .
11
else
12
if std1 < std2 then
13
Return s1 .
14
else if std1 > std2 then
15
Return s2 .
16
else
17
Return randon one from {s1 , s2 }.
18
end
19
end
20 end
Convolutional
Pooling
C
Full connection
P
Chromosome #1
C1
P1
C2
Convolutional
unit list
C3
F
Chromosome #2
P2
Pooling
unit list
F1
F2
F3
C1
Full connection
unit list
P1
C2
P2
Convolutional
unit list
P3
F1
Pooling
unit list
F2
F3
Full connection
unit list
C1
P1
F1
C1
P1
F1
C2
P2
F2
C2
P2
F2
P3
F3
C3
F3
F4
F4
(a) Unit Collection
crossover
crossover
C1
C1
crossover
P1
crossover
P1
F1
P2
F2
P3
F3
F1
crossover
C2
C2
crossover
P2
C3
F2
crossover
F3
F4
(b) Unit Aligh and Crossover
Offspring #2
Offspring #1
C1
P1
C2
C3
P2
F1
F2
F3
C1
P1
C2
P2
P3
F1
F2
F3
F4
(c) Unit Restore
Fig. 5. An example to illustrate the entire crossover process. In this example, the first chromosome is with length 8 including three convolutional layers, two
pooling layers, and three full connection layers; the other one is with length 9 including two convolutional layers, three pooling layers, and four full connection
layers. In the first step of crossover (i.e., the unit collection), the convolutiona layers, the pooling layers, and the full connection layers are collected from
each chromosome and stacked with the same orders to them in each chromosome (see Fig. 5a). In the second step, the unit lists with the same unit types
are aligned at the top, i.e., the two convolutional layer lists are picked and aligned, and the same operations on the other two lists. When these unit lists
finish the alignment, units at the same positions from the each two lists are paired and performed crossover operation, which can be shown in Fig. 5b. At last,
units from these unit lists are restored based on the positions where they are from (see Fig. 5b). Note in Fig. 5b, the units that have experienced crossover
operations are highlighted with italics and underline fonts, while the units that do not perform the crossover operations remain the same.
are explicitly and elaborately addressed. To be specific, a fraction of individuals with promising mean values are chosen first,
and then the remaining individuals are selected by the modified
binary tournament selection demonstrated in Subsection III-F.
By these two strategies, the elitism and the diversity are
considered simultaneously, which are expected to collectively
improve the performance of the proposed EvoCNN method.
Note here that, the selected elites are removed before the
tournament selection gets start to work (line 3 of Algorithm 5),
while the individuals selected for the purpose of diversity are
kept in the current population for the next round of tournament
selection (lines 4-8 of Algorithm 5) based on the convention
in the community.
and then the decoded CNN will be deeply trained with a larger
number of epochs by SGD for future usage.
IV. E XPERIMENT D ESIGN
In order to quantify the performance of the proposed
EvoCNN, a series of experiments are designed and performed
on the chosen image classification benchmark datasets, which
are further compared to state-of-the-art peer competitors. In the
following, these benchmark datasets are briefly introduced at
first. Then, the peer competitors are given. Finally, parameter
settings of the proposed EvoCNN method participating into
these experiments are documented.
A. Benchmark Datasets
H. Best Individual Selection and Decoding
In these experiments, nine widely used image classification
At the end of evolution, multiple individuals are with benchmark datasets are used to examine the performance of
promising mean values but different architectures and connec- the proposed EvoCNN method. They are the Fashion [49], the
tion weight initialization values. In this regard, there will be Rectangle [50], the Rectangle Images (RI) [50], the Convex
multiple choices to select the “Best Individual”. For example, Sets (CS) [50], the MNIST Basic (MB) [50], the MNIST
if we are only concerned with the best performance, we with Background Images (MBI) [50], Random Background
could neglect their architecture configurations and consider (MRB) [50], Rotated Digits (MRD) [50], and with RD plus
only the classification accuracy. Otherwise, if we emphasis Background Images (MRDBI) [50] benchmarks.
Based on the types of the classification objects, these
the smaller number of parameters, corresponding decision
could be made. Once the “Best Individual” is confirmed, benchmarks are classified into three different categories. The
the corresponding CNN is decoded based on the encoded first category includes only Fashion, which is for recognizing
architecture and connection weight initialization information, 10 fashion objects (e.g., trousers, coats, and so on) in the
50, 000 training images and 10, 000 test images. The second final classification accuracy and is invisible to public. As for
one is composed of the MNIST [6] variants including the other architecture design approaches introduced in Section I,
MB, MBI, MRB, MRD, and the MRDBI benchmarks for such as GS, RS, among others, they are not scalable to CNNs
classifying 10 hand-written digits (i.e., 0-9). Because the and not suitable to be directly compared as well.
MNIST has been easily achieved 97%, these MNIST variants
In the experiments, state-of-the-art algorithms that have
are arbitrarily added into different barriers (e.g., random back- reported promising classification errors on the chosen benchgrounds, rotations) from the MNIST to improve the complexity marks are collected as the peer competitors of the proof classification algorithms. Furthermore, these variants have posed EvoCNN method. To be specific, the peer com12, 000 training images and 50, 000 test images, which further petitors on the Fashion benchmark are collected from the
challenges the classification algorithms due to the mush less dataset homepage2. They are 2C1P2F+Dropout, 2C1P, 3C2F,
training data while more test data. The third category is for 3C1P2F+Dropout, GRU+SVM+Dropout, GoogleNet [41],
recognizing the shapes of objects (i.e., the rectangle or not for AlexNet [3], SqueezeNet-200 [51], MLP 256-128-64, and
the Rectangle and RI benchmarks, and the convex or not for VGG16 [52], which perform the experiments on the raw
the Convex benchmark). Obviously, this category covers the dataset without any preprocessing. The peer competitors on
Rectangle, the RI, and the CS benchmarks that contain 1, 200, other benchmarks are CAE-2 [53], TIRBM [54], PGBM+DN12, 000, and 8, 000 training images, respectively, and all of 1 [55], ScatNet-2 [56], RandNet-2 [57], PCANet-2 (softthem include 50, 000 test images. Compared to the Rectangle max) [57], LDANet-2 [57], SVM+RBF [50], SVM+Poly [50],
benchmark, the RI is generated by adding randomly sampled NNet [50], SAA-3 [50], and DBN-3 [50], which are from
backgrounds from the MNIST for increasing the difficulties the literature [57] recently published and the provider of the
of classification algorithms.
benchmarks3.
In addition, each image in these benchmarks is with the size
28 × 28, and examples from these benchmarks are shown in C. Parameter Settings
Fig. 6 for reference. Furthermore, another reason for using
All the parameter settings are set based on the conventions
these benchmark datasets is that different algorithms have in the communities of evolutionary algorithms [58] and deep
reported their promising results, which is convenient for the learning [59], in addition to the maximal length of each basic
comparisons on the performance of the proposed EvoCNN layer. Specifically, the population size and the total generation
method and these host algorithms (details can be seen Subsec- number are set to be 100. The distribution index of SBX and
tion IV-B).
Polynomial mutation are both set to be 20, and their associated
probabilities are specified as 0.9 and 0.1, respectively. The
maximum lengths of the convolutional layers, the pooling
layers, and the full connection layers are set to be the same
(a) Examples from the first category. From left to right, they are T-shirt, trouser, as 5 (i.e., the maximal depths of CNNs in these experiments
pullover, dress, coat, sandal, shirt, sneaker, bag, and ankle boot.
are 15). Moreover, the proportion of the elitism is set to be
20% based on the Pareto principle. Note that the 20% data
are randomly selected from the training images as the fitness
evaluation dataset. For the implementation of the proposed
(b) Examples from the second category. From left to right, each two images
EvoCNN method, we constrain the same values of the width
are from one group, and each group is from MBi, MRB, MRD, MRDBI, and
MB, respectively. These images refer to the hand-written digits 0, 4, 2, 6, 0, and height for filter, stride, and kernel. The width and height
5, 7, 5, 9, and 6, respectively.
for stride in the convolutional layer are set to be 1, those in
the pooling layer are specified at the same value to its kernel,
and the convolutional type is fixed to “SAME”.
The proposed EvoCNN method is implemented by Tensor(c) Examples from the thrid category, From left to right, the first three images
flow [60], and each copy of the code runs in a computer
are from the Rectangle benchmark, the following four ones are from the RI
benchmark, and the remaingings are from the Convex benchmark. Specifically, equipped with two GPU cards with the identical model number
these examples with the index 1, 2, 6, 7, and 11 are positive samples.
GTX1080. During the final training phase, each individual
Fig. 6. Examples from the benchmarks chosen.
is imposed by the BatchNorm [61] for speeding up and the
weight decay with an unified number for preventing from
B. Peer Competitors
the overfitting. Due to the heuristic nature of the proposed
Ideally, the algorithms for neural network architecture de- EvoCNN method, 30 independent runs are performed on each
sign discussed in Section I should be collected here as the peer benchmark dataset, and the mean results are estimated for
competitors. However, because MetaQNN [21] and LEIC [36] the comparisons unless otherwise specified. Furthermore, the
highly rely on the computational resources, it is impossible experiments take 4 days on the Fashion benchmark for each
to reproduce the experimental results in an academic envi- run, 2–3 days on other benchmarks.
ronment. Furthermore, the benchmark dataset investigated in
2 https://github.com/zalandoresearch/fashion-mnist
these two algorithms employed different data preprocessing
3 http://www.iro.umontreal.ca/∼ lisa/twiki/bin/view.cgi/Public/DeepVsShallowComparisonIC
and augmentation techniques, which will highly enhance the
V. E XPERIMENTAL R ESULTS AND A NALYSIS
classification error rates, where the difference is only 0.2%.
The
proposed EvoCNN method further decreases the error rate
In this section, the experimental results and analysis of the
by
0.83%
to 5.47%. Furthermore, the mean performance of
proposed EvoCNN method against peer competitors are shown
EvoCNN
is
even better than the best performance of eight
in Subsection V-A. Then, the weight initialization method in
competitors,
and
only a little worse than the best of GoogleNet
the proposed EvoCNN method are specifically investigated in
and
VGG16.
However,
EvoCNN has a much smaller number
Subsection V-B.
of connection weights — EvoCNN uses 6.52 million weights
A. Overall Results
while GoogleNet uses 101 million and VGG uses 26 million
Experimental results on the Fashion benchmark dataset are weights. EvoCNN also employs only half of the numbers of
shown in Table II, and those on the MB, MRD, MRB, MBI, training epochs used in the VGG16. The results show that the
MRDBI, Rectangle, RI, and Convex benchmark datasets are proposed EvoCNN method obtains much better performance in
shown in Table III. In Tables II and III, the last two rows the architecture design and connection weight initialization of
denote the best and mean classification errors received from CNNs on the Fashion benchmark, which significantly reduces
the proposed EvoCNN method, respectively, and other rows the classification error of the evolved CNN generated by the
show the best classification errors reported by peer competi- proposed EvoCNN method.
According to Table III, EvoCNN is the best one among all
tors4 . In order to conveniently investigate the comparisons, the
terms “(+)” and “(-)” are provided to denote whether the best the 13 different methods. Specifically, EvoCNN achieves the
result generated by the proposed EvoCNN method is better or best performance on five of the eight datasets, and the second
worse that the best result obtained by the corresponding peer best on the other three datasets — the MB, MRD, and Convex
competitor. The term “—” means there is no available result datasets, where LDANet-2, TIRBM, and PCANet-2(softmax)
reported from the provider or cannot be counted. Most informa- achieve the lowest classification error rate, respectively. Furtion of the number of parameters and number of training epoch thermore, comparing the mean performance of EvoCNN with
from the peer competitors on the Fashion benchmark dataset the best of the other 12 methods, EvoCNN is the best on four
is available, therefore, such information from EvoCNN is also datasets (MRB, MBI, Rectangle and RI), second best on two
shown in Table II for multi-view comparisons. However, for (MRD and Convex), third best and fourth best on the other two
the benchmarks in Table III, such information is not presented datasets (MRDBI and MB). Particularly on MRB and MBI,
because they are not available from the peer competitors. the lowest classification error rates of the others are 6.08%
Note that all the results from the peer competitors and the and 11.5%, and the mean error rates of EvoCNN are 3.59%
proposed EvoCNN method are without any data augmentation and 4.62%, respectively. In summary, the best classification
error of the proposed EvoCNN method wins 80 out of the
preprocessing on the benchmarks.
84 comparisons against the best results from the 12 peer
competitors, and the mean classification error of EvoCNN is
TABLE II
T HE CLASSIFICATION ERRORS OF THE PROPOSED E VO CNN METHOD
better than the best error of the 12 methods on 75 out of the
AGAINST THE PEER COMPETITORS ON THE FASHION BENCHMARK
84 comparisons.
DATASET.
B. Performance Regarding Weight Initialization
It is clearly shown in Table II that by comparing the best
performance, the proposed EvoCNN method outperforms all
the ten peer competitors. The two state-of-the-art algorithms,
GoogleNet and VGG16, achieve respectively 6.5% and 6.3%
4 It is a convention in deep learning community that only the best result is
reported.
60
weight initializaed by EvoCNN
weight initialized by Xavier
50
40
30
0
Fashion
MB
MRD
MRB
MBI
MRDBI
Benchmark datasets
Rectangle
5.39
6.33
10
5.97
8.27
20
0.01
0.23
300
30
—
150
100
—
—
200
25
200
100
100
37.38
41.88
3.27M
100K
—
7.14M
—
101M
60M
500K
41K
26M
6.68M
6.52M
4.62
5.8
8.40(+)
7.50(+)
9.30(+)
7.40(+)
10.30(+)
6.30(+)
10.10(+)
10.00(+)
10.00(+)
6.50(+)
5.47
7.28
3.59
3.93
2C1P2F+Drouout
2C1P
3C2F
3C1P2F+Dropout
GRU+SVM+Dropout
GoogleNet [41]
AlexNet [3]
SqueezeNet-200 [51]
MLP 256-128-64
VGG16 [52]
EvoCNN (best)
EvoCNN (mean)
5.46
6.24
# epochs
1.28
2.45
# parameters
7.28
9.28
error(%)
Classification error (%)
classifier
RI
Convex
Fig. 7. Performance comparison between the proposed EvoCNN method and
the CNN using the evolved architecture and the Xavier weight initializer.
Further experiments are performed to examine the effectiveness of the connection weight initialization method in the
proposed EvoCNN approach. By comparing different weight
initializers, we would also investigate whether the architecture or the connection weight initialization values will affect
the classification performance. To achieve this, we initialize
another group of CNNs with the same architectures as that of
TABLE III
T HE CLASSIFICATION ERRORS OF THE PROPOSED E VO CNN METHOD AGAINST THE PEER COMPETITORS ON THE MB, MRD, MRB, MBI, MRDBI,
R ECTANGLE , RI, AND C ONVEX BENCHMARK DATASETS
classifier
MB
MRD
MRB
MBI
MRDBI
Rectangle
RI
Convex
CAE-2 [53]
TIRBM [54]
PGBM+DN-1 [55]
ScatNet-2 [56]
RandNet-2 [57]
PCANet-2 (softmax) [57]
LDANet-2 [57]
SVM+RBF [50]
SVM+Poly [50]
NNet [50]
SAA-3 [50]
DBN-3 [50]
EvoCNN (best)
EvoCNN (mean)
2.48(+)
—
—
1.27(+)
1.25(+)
1.40(+)
1.05(-)
3.03(+)
3.69(+)
4.69(+)
3.46(+)
3.11(+)
1.18
1.28
9.66(+)
4.20(-)
—
7.48(+)
8.47(+)
8.52(+)
7.52(+)
11.11(+)
15.42(+)
18.11(+)
10.30(+)
10.30(+)
5.22
5.46
10.90(+)
—
6.08(+)
12.30(+)
13.47(+)
6.85(+)
6.81(+)
14.58(+)
16.62(+)
20.04(+)
11.28(+)
6.73(+)
2.80
3.59
15.50(+)
—
12.25(+)
18.40(+)
11.65(+)
11.55(+)
12.42(+)
22.61(+)
24.01(+)
27.41(+)
23.00(+)
16.31(+)
4.53
4.62
45.23(+)
35.50(+)
36.76(+)
50.48(+)
43.69(+)
35.86(+)
38.54(+)
55.18(+)
56.41(+)
62.16(+)
51.93(+)
47.39(+)
35.03
37.38
1.21(+)
—
—
0.01(=)
0.09(+)
0.49(+)
0.14(+)
2.15(+)
2.15(+)
7.16(+)
2.41(+)
2.61(+)
0.01
0.01
21.54(+)
—
—
8.02(+)
17.00(+)
13.39(+)
16.20(+)
24.04(+)
24.05(+)
33.20(+)
24.05(+)
22.50(+)
5.03
5.97
—
—
—
6.50(+)
5.45(+)
4.19(-)
7.22(+)
19.13(+)
19.82(+)
32.25(+)
18.41(+)
18.63(+)
4.82
5.39
the evolved EvoCNN, but their weights are initialized with the
widely used Xavier initializer [38] (details in Subsection II-C).
The comparisons are illustrated in Fig. 7, where the xaxis denotes the benchmark datasets, and y-axis denotes the
classification error rates. Fig. 7 clearly shows that the proposed
connection weight initialization strategy in EvoCNN improves
the classification performance on all the benchmarks over the
widely used Xavier initializer. To be specific, the proposed
weight initialization method reaches ≈1.5% classification accuracy improvement on the Fashion, MB, MRD, MBI, RI, and
Convex benchmarks, and 4.5% on the MRDBI benchmark. By
comparing with the results in Tables II and III, it also can be
concluded that the architectures of CNNs contribute to the
classification performance more than that of the connection
weight initialization. The proposed EvoCNN method gained
promising performance by automatically evolving both the
initial connection weights and the architectures of CNNs.
operation in this context (because crossover operations can
not be easily performed between the genes from different
origins). Therefore, a new crossover operation is introduced
to the proposed EvoCNN method. UC, UA, as well as UR are
designed to complete the crossover operation, which enhances
the communications between the encoded information, and expectedly leverages the performance in pursuing the promising
architectures of CNNs.
The typically used approaches in optimizing the weights of
CNNs are based on the gradient information. It is well known
that the gradient-based optimizers are sensitive to the starting
position of the parameters to be optimized. Without a better
starting position, the gradient-based algorithms are easy to be
trapped into local minima. Intuitively, finding a better starting
position of the connection weights by GAs is intractable due
to the huge numbers of parameters. As we have elaborated, a
huge number of parameters can not be efficiently encoded into
the chromosomes, nor effectively optimized. In the proposed
VI. F URTHER D ISCUSSIONS
EvoCNN method, an indirect encoding approach is employed,
which
only encodes the means and standard derivations of
In this section, we will further discuss the encoding stratthe
weights
in each layer. In fact, employing the standard
egy of the architectures and the weights related parameters,
derivation
and
the mean to initialize the starting position of
and the fitness evaluation in the proposed EvoCNN method.
connection
weights
is a compromised approach and can be
Extra findings from the experimental results are discussed as
ubiquitously
seen
from
deep learning libraries [60], [62], [63],
well, which could provide insights on the applications of the
which
is
the
motivation
of
this design in the proposed EvoCNN
proposed EvoCNN method.
method.
With
this
strategy,
only two real numbers are used
In GAs, crossover operators play the role of exploitation
to
denote
the
hundreds
of
thousands
of parameters, which
search (i.e., the local search), and mutation operators act
could
save
much
computational
resource
for encoding and
the exploration search (i.e., the global search). Because the
optimization.
local search and the global search should complement to each
other, only well designing both of them could remarkably
Existing techniques for searching for the architectures of
promote the performance. Generally, the crossover operators CNNs typically take the final classification accuracy as the
operate on the chromosomes with the same lengths, while fitness of individuals. A final classification accuracy typically
the proposed EvoCNN method employs the variable-length requires many more epochs of the training, which is a very
ones. Furthermore, multiple different basic layers exist in time-consuming process. In order to complete the architecture
CNNs, which improve the difficulties of designing crossover design with this fitness evaluation approach, it is natural
to employ a lot of computation resource to perform this
task in parallel to speed up the design. However, computational resource is not necessarily available to all interested
researchers. Furthermore, it also requires extra professional
assistances, such as the task schedule and synchronization
in the multi-thread context, which is beyond the expertise
of most researchers/users. In fact, it is not necessary to
check individual final classification accuracy, but a tendency
that could predict their future quality would be sufficient.
Therefore, we only employ a small number of epochs to
train these individuals. With this kind of fitness measurement,
the proposed EvoCNN method does not highly reply on the
computational resource. In summary, with the well-designed
yet simplicity encoding strategies and this fitness evaluation
method, researchers without rich domain knowledge could
also design the promising CNN models for addressing specific
tasks in their (academic) environment, where is the computational resources are typically limited.
TABLE IV
T HE CLASSIFICATION ACCURACY AND THE PARAMETER NUMBERS FROM
THE INDIVIDUALS ON THE FASHION BENCHMARK DATASET.
classification accuracy
# parameters
99%
569,250
98%
98,588
97%
7,035
96%
3,205
95%
955
Moreover, interesting findings have also been discovered
when the proposed EvoCNN method terminates, i.e., multiple
individuals are with the similar performance but significant
different numbers of connection weights (See Table IV), which
are produced due to population-based nature of the proposed
EvoCNN method. Recently, various applications have been
developed, such as the auto-driving car and some interesting
real-time mobile applications. Due to the limited processing
capacity and battery in these devices, trained CNNs with
similar performance but fewer parameters are much more
preferred because they require less computational resource
and the energy consumption. In addition, similar performance
of the individuals are also found with different lengths of
the basic layers, and there are multiple pieces of hardware
that have been especially designed to speed the primitive
operations in CNNs, such as the specialized hardware for
convolutional operations. In this regard, the proposed EvoCNN
method could give manufacturers more choices to make a
decision based on their own preferences. If the traditional
approaches are employed here, it is difficult to find out such
models at a single run.
VII. C ONCLUSIONS
AND
F UTURE W ORK
The goal of the paper was to develop a new evolutionary approach to automatically evolving the architectures and weights
of CNNs for image classification problems. This goal has
been successfully achieved by proposing a new representation
for weight initialization strategy, a new encoding scheme
of variable-length chromosomes, a new genetic operator, a
slacked binary tournament selection, and an efficient fitness
evaluation method. This approach was examined and compared with 22 peer competitors including the most state-ofthe-are algorithms on nine benchmark datasets commonly
used in deep learning. The experimental results show that
the proposed EvoCNN method significantly outperforms all
of these existing algorithms on almost all these datasets in
terms of their best classification performance. Furthermore,
the mean classification error rate of the proposed algorithm
is even better than the best of others in many cases. In
addition, the model optimized by EvoCNN is with a much
smaller number of parameters yet promising best classification
performance. Specifically, the model optimized by EvoCNN
employs 100 epochs to reach the lowest classification error
rate of 5.47% on the Fasion dataset, while the state-of-theart VGG16 employs 200 epochs but 6.50% classification error
rate. Findings from the experimental results also suggest that
the proposed EvoCNN method could provide more options
for the manufacturers that are interested in integrating CNNs
into their products with limited computational and battery
resources.
In this paper, we investigate the proposed EvoCNN method
only on the commonly used middle-scale benchmarks. However, large-scale data are also widely exist in the current era
of big data. Because evaluating only one epoch on these largescale data would require significant computational resource
and take a long period, the proposed fitness evaluation method
is not suitable for them unless a huge amount computational
resources are available. In the future, we will put effort on
efficient fitness evaluation techniques. In addition, we will also
investigate evolutionary algorithms for recurrent neural networks, which are powerful tools for addressing time-dependent
data, such as the voice and video data.
R EFERENCES
[1] D. CireşAn, U. Meier, J. Masci, and J. Schmidhuber, “Multi-column
deep neural network for traffic sign classification,” Neural Networks,
vol. 32, pp. 333–338, 2012.
[2] F. Ning, D. Delhomme, Y. LeCun, F. Piano, L. Bottou, and P. E. Barbano,
“Toward automatic phenotyping of developing embryos from videos,”
IEEE Transactions on Image Processing, vol. 14, no. 9, pp. 1360–1371,
2005.
[3] A. Krizhevsky, I. Sutskever, and G. E. Hinton, “Imagenet classification
with deep convolutional neural networks,” in Advances in Neural Information Processing Systems, 2012, pp. 1097–1105.
[4] D. H. Hubel and T. N. Wiesel, “Receptive fields, binocular interaction
and functional architecture in the cat’s visual cortex,” The Journal of
Physiology, vol. 160, no. 1, pp. 106–154, 1962.
[5] Y. LeCun, B. Boser, J. S. Denker, D. Henderson, R. E. Howard,
W. Hubbard, and L. D. Jackel, “Backpropagation applied to handwritten
zip code recognition,” Neural Computation, vol. 1, no. 4, pp. 541–551,
1989.
[6] Y. LeCun, L. Bottou, Y. Bengio, and P. Haffner, “Gradient-based learning
applied to document recognition,” Proceedings of the IEEE, vol. 86,
no. 11, pp. 2278–2324, 1998.
[7] D. E. Rumelhart, G. E. Hinton, and R. J. Williams, “Learning representations by back-propagating errors,” Cognitive Modeling, vol. 5, no. 3,
p. 1, 1988.
[8] K. Simonyan and A. Zisserman, “Very deep convolutional networks for
large-scale image recognition,” arXiv preprint arXiv:1409.1556, 2014.
[9] K. He, X. Zhang, S. Ren, and J. Sun, “Deep residual learning for image
recognition,” in Proceedings of the IEEE Conference on Computer Vision
and Pattern Recognition, 2016, pp. 770–778.
[10] H. Bourlard and Y. Kamp, “Auto-association by multilayer perceptrons
and singular value decomposition,” Biological Cybernetics, vol. 59, no.
4-5, pp. 291–294, 1988.
[11] G. E. Hinton and R. S. Zemel, “Autoencoders, minimum description
length, and helmholtz free energy,” Advances in Neural Information
Processing Systems, pp. 3–3, 1994.
[12] G. E. Hinton and R. R. Salakhutdinov, “Reducing the dimensionality of
data with neural networks,” Science, vol. 313, no. 5786, pp. 504–507,
2006.
[13] J. Bergstra and Y. Bengio, “Random search for hyper-parameter optimization,” Journal of Machine Learning Research, vol. 13, no. Feb, pp.
281–305, 2012.
[14] C. E. Rasmussen and C. K. Williams, Gaussian processes for machine
learning. MIT press Cambridge, 2006, vol. 1.
[15] J. Močkus, “On bayesian methods for seeking the extremum,” in
Optimization Techniques IFIP Technical Conference. Springer, 1975,
pp. 400–404.
[16] J. S. Bergstra, R. Bardenet, Y. Bengio, and B. Kégl, “Algorithms
for hyper-parameter optimization,” in Advances in Neural Information
Processing Systems, 2011, pp. 2546–2554.
[17] F. Hutter, H. H. Hoos, and K. Leyton-Brown, “Sequential model-based
optimization for general algorithm configuration.” LION, vol. 5, pp. 507–
523, 2011.
[18] J. Beck and T. Fiala, “integer-making theorems,” Discrete Applied
Mathematics, vol. 3, no. 1, pp. 1–8, 1981.
[19] J. Halton and G. Smith, “Radical inverse quasi-random point sequence,
algorithm 247,” Communications of the ACM, vol. 7, p. 701, 1964.
[20] E. I. Atanassov, A. Karaivanova, and S. Ivanovska, “Tuning the generation of sobol sequence with owen scrambling.” in Large-Scale Scientific
Computing. Springer, 2009, pp. 459–466.
[21] B. Baker, O. Gupta, N. Naik, and R. Raskar, “Designing neural network
architectures using reinforcement learning,” International Conference on
Learning Representations, 2017.
[22] D. Ashlock, Evolutionary computation for modeling and optimization.
Springer Science & Business Media, 2006.
[23] T. Back, Evolutionary algorithms in theory and practice: evolution
strategies, evolutionary programming, genetic algorithms.
Oxford
university press, 1996.
[24] M. Mitchell, An introduction to genetic algorithms. MIT press, 1998.
[25] X. Yao, “Evolving artificial neural networks,” Proceedings of the IEEE,
vol. 87, no. 9, pp. 1423–1447, 1999.
[26] G. Miller, “Designing neural networks using genetic algorithms,” Proceedings of ICGA-89, 1989.
[27] K. O. Stanley and R. Miikkulainen, “Evolving neural networks through
augmenting topologies,” Evolutionary computation, vol. 10, no. 2, pp.
99–127, 2002.
[28] K. O. Stanley, D. B. D’Ambrosio, and J. Gauci, “A hypercube-based
encoding for evolving large-scale neural networks,” Artificial life, vol. 15,
no. 2, pp. 185–212, 2009.
[29] K. O. Stanley, “Compositional pattern producing networks: A novel
abstraction of development,” Genetic programming and evolvable machines, vol. 8, no. 2, pp. 131–162, 2007.
[30] J. K. Pugh and K. O. Stanley, “Evolving multimodal controllers with
hyperneat,” in Proceedings of the 15th Annual Conference on Genetic
and Evolutionary Computation. ACM, 2013, pp. 735–742.
[31] M. Kim and L. Rigazio, “Deep clustered convolutional kernels,” in
Feature Extraction: Modern Questions and Challenges, 2015, pp. 160–
172.
[32] C. Fernando, D. Banarse, M. Reynolds, F. Besse, D. Pfau, M. Jaderberg,
M. Lanctot, and D. Wierstra, “Convolution by evolution: Differentiable
pattern producing networks,” in Proceedings of the 2016 on Genetic and
Evolutionary Computation Conference. ACM, 2016, pp. 109–116.
[33] B. Zoph and Q. V. Le, “Neural architecture search with reinforcement
learning,” arXiv preprint arXiv:1611.01578, 2016.
[34] B. Baker, O. Gupta, N. Naik, and R. Raskar, “Designing neural
network architectures using reinforcement learning,” arXiv preprint
arXiv:1611.02167, 2016.
[35] P. Verbancsics and J. Harguess, “Generative neuroevolution for deep
learning,” arXiv preprint arXiv:1312.5355, 2013.
[36] E. Real, S. Moore, A. Selle, S. Saxena, Y. L. Suematsu, Q. Le, and
A. Kurakin, “Large-scale evolution of image classifiers,” arXiv preprint
arXiv:1703.01041, 2017.
[37] X. Glorot, A. Bordes, and Y. Bengio, “Deep sparse rectifier neural
networks,” in Proceedings of the Fourteenth International Conference
on Artificial Intelligence and Statistics, 2011, pp. 315–323.
[38] X. Glorot and Y. Bengio, “Understanding the difficulty of training deep
feedforward neural networks,” in Proceedings of the 13th International
Conference on Artificial Intelligence and Statistics, 2010, pp. 249–256.
[39] A. L. Hodgkin and A. F. Huxley, “A quantitative description of membrane current and its application to conduction and excitation in nerve,”
The Journal of physiology, vol. 117, no. 4, pp. 500–544, 1952.
[40] M. D. Zeiler and R. Fergus, “Visualizing and understanding convolutional networks,” in European Conference on Computer Vision.
Springer, 2014, pp. 818–833.
[41] C. Szegedy, W. Liu, Y. Jia, P. Sermanet, S. Reed, D. Anguelov, D. Erhan,
V. Vanhoucke, and A. Rabinovich, “Going deeper with convolutions,” in
Proceedings of the IEEE Conference on Computer Vision and Pattern
Recognition, 2015, pp. 1–9.
[42] M. N. Omidvar, X. Li, Y. Mei, and X. Yao, “Cooperative co-evolution
with differential grouping for large scale optimization,” IEEE Transactions on Evolutionary Computation, vol. 18, no. 3, pp. 378–393, 2014.
[43] Y. Bengio, A. Courville, and P. Vincent, “Representation learning: A
review and new perspectives,” IEEE Transactions on Pattern Analysis
and Machine Intelligence, vol. 35, no. 8, pp. 1798–1828, 2013.
[44] Y. Bengio and O. Delalleau, “On the expressive power of deep architectures,” in International Conference on Algorithmic Learning Theory.
Springer, 2011, pp. 18–36.
[45] O. Delalleau and Y. Bengio, “Shallow vs. deep sum-product networks,”
in Advances in Neural Information Processing Systems, 2011, pp. 666–
674.
[46] A. Blumer, A. Ehrenfeucht, D. Haussler, and M. K. Warmuth, “Occam’s
razor,” Information Processing Letters, vol. 24, no. 6, pp. 377–380, 1987.
[47] K. Deb and R. B. Agrawal, “Simulated binary crossover for continuous
search space,” Complex Systems, vol. 9, no. 3, pp. 1–15, 1994.
[48] K. Deb, Multi-objective optimization using evolutionary algorithms.
John Wiley & Sons, 2001, vol. 16.
[49] H. Xiao, K. Rasul, and R. Vollgraf, “Fashion-mnist: a novel image
dataset for benchmarking machine learning algorithms,” arXiv preprint
arXiv:1708.07747, 2017.
[50] H. Larochelle, D. Erhan, A. Courville, J. Bergstra, and Y. Bengio,
“An empirical evaluation of deep architectures on problems with many
factors of variation,” in Proceedings of the 24th International Conference
on Machine Learning. ACM, 2007, pp. 473–480.
[51] F. N. Iandola, S. Han, M. W. Moskewicz, K. Ashraf, W. J. Dally,
and K. Keutzer, “Squeezenet: Alexnet-level accuracy with 50x fewer
parameters and < 0.5 mb model size,” arXiv preprint arXiv:1602.07360,
2016.
[52] K. Simonyan and A. Zisserman, “Very deep convolutional networks for
large-scale image recognition,” arXiv preprint arXiv:1409.1556, 2014.
[53] S. Rifai, P. Vincent, X. Muller, X. Glorot, and Y. Bengio, “Contractive
auto-encoders: Explicit invariance during feature extraction,” in Proceedings of the 28th International Conference on Machine Learning, 2011,
pp. 833–840.
[54] K. Sohn and H. Lee, “Learning invariant representations with local
transformations,” arXiv preprint arXiv:1206.6418, 2012.
[55] K. Sohn, G. Zhou, C. Lee, and H. Lee, “Learning and selecting features
jointly with point-wise gated boltzmann machines,” in International
Conference on Machine Learning, 2013, pp. 217–225.
[56] J. Bruna and S. Mallat, “Invariant scattering convolution networks,”
IEEE transactions on Pattern Analysis and Machine Intelligence, vol. 35,
no. 8, pp. 1872–1886, 2013.
[57] T.-H. Chan, K. Jia, S. Gao, J. Lu, Z. Zeng, and Y. Ma, “Pcanet: A simple
deep learning baseline for image classification?” IEEE Transactions on
Image Processing, vol. 24, no. 12, pp. 5017–5032, 2015.
[58] K. Deb, A. Pratap, S. Agarwal, and T. Meyarivan, “A fast and elitist
multiobjective genetic algorithm: Nsga-ii,” IEEE Transactions on Evolutionary Computation, vol. 6, no. 2, pp. 182–197, 2002.
[59] G. E. Hinton, “A practical guide to training restricted boltzmann
machines,” in Neural Networks: Tricks of the Trade. Springer, 2012,
pp. 599–619.
[60] M. Abadi, A. Agarwal, P. Barham, E. Brevdo, Z. Chen, C. Citro, G. S.
Corrado, A. Davis, J. Dean, M. Devin et al., “Tensorflow: Large-scale
machine learning on heterogeneous distributed systems,” arXiv preprint
arXiv:1603.04467, 2016.
[61] S. Ioffe and C. Szegedy, “Batch normalization: Accelerating deep
network training by reducing internal covariate shift,” in International
Conference on Machine Learning, 2015, pp. 448–456.
[62] Y. Jia, E. Shelhamer, J. Donahue, S. Karayev, J. Long, R. Girshick,
S. Guadarrama, and T. Darrell, “Caffe: Convolutional architecture for
fast feature embedding,” in Proceedings of the 22nd ACM international
conference on Multimedia. ACM, 2014, pp. 675–678.
[63] T. Chen, M. Li, Y. Li, M. Lin, N. Wang, M. Wang, T. Xiao, B. Xu,
C. Zhang, and Z. Zhang, “Mxnet: A flexible and efficient machine
learning library for heterogeneous distributed systems,” arXiv preprint
arXiv:1512.01274, 2015.
| 9cs.NE
|
arXiv:1702.00143v2 [math.GR] 16 Feb 2017
ISOMORPHISM CLASSES OF FOUR DIMENSIONAL
NILPOTENT ASSOCIATIVE ALGEBRAS OVER A FIELD
MARCO ANTONIO PELLEGRINI
Abstract. In this paper we classify the isomorphism classes of four dimensional nilpotent associative algebras over a field F, studying regular subgroups
of the affine group AGL4 (F). In particular we provide explicit representatives
for such classes when F is a finite field, the real field R or an algebraically
closed field.
1. Introduction
The aim of this paper is the classification of the isomorphism classes of four
dimensional nilpotent associative algebras over a field F. This result will be achieved
exploiting the properties of the regular subgroups of the affine group AGL4 (F).
It is worth recalling that the affine group AGLn (F) can be identified with the subgroup of GLn+1 (F) consisting of the invertible matrices having (1, 0, . . . , 0)T as first
column. It follows that AGLn (F) acts on the right on the set A = {(1, v) : v ∈ Fn }
of affine points. A subgroup R of AGLn (F) is called regular if it acts regularly on
A, namely if for every v ∈ Fn there exists a unique element in R having (1, v) as
first row. Writing the elements r of a regular subgroup R ≤ AGLn (F) as
1
v
(1.1)
r=
= µR (v),
0 In + δR (v)
we give rise to the functions δR : Fn → Matn (F) and µR : Fn → R. For instance,
the translation subgroup Tn of AGLn (F) is a regular subgroup, corresponding to
the choice δTn equal to the zero function.
We focus our attention on the set ∆n (F) of the regular subgroups R of AGLn (F)
with the property that δR is a linear function. Notice that if R is abelian, then
R ∈ ∆n (F) (see [1, 8]); however, for n ≥ 3, the set ∆n (F) also contains nonabelian
groups (see [5]). We also remark that if R ∈ ∆n (F), then R is unipotent and
contains nontrivial translations (see [4, 6]). Our main motivation for studying the
set ∆n (F) is that there exists a bijection between the set of isomorphism classes of
nilpotent associative algebras of dimension n over a field F and the set of conjugacy
classes of subgroups in ∆n (F) (see [4, Proposition 2.2]). In fact, our classification
of four dimensional nilpotent associative F-algebras (Theorem 1.1) will follow from
the classification of the conjugacy classes of regular subgroups in ∆4 (F).
The paper is organized as follows. In Section 2 we recall some useful results
concerning ∆n (F) and in the following sections we classify, up to conjugation, the
regular subgroups in ∆4 (F), see Theorems 4.2 and 5.12. More in detail, in Section
4 we deal with abelian regular subgroups and in Section 5 we consider nonabelian
2010 Mathematics Subject Classification. 16Z05, 16N40, 15A21, 20B35.
Key words and phrases. Regular subgroup; congruent matrices; nilpotent algebra; finite field.
1
2
MARCO ANTONIO PELLEGRINI
subgroups. Corollaries 4.3 and 4.5 and Propositions 5.15 and 5.23 give the explicit
classification for algebraically closed fields or finite fields. The classification for the
real field is given in Corollary 4.4 and Proposition 5.20.
Before to describe the results we obtained, we need to fix some notation. The sets
ΠS (F) and ΠA (F) are complete sets of representatives of projective congruent classes
for, respectively, symmetric and asymmetric matrices in Mat3 (F), see Definition
2.2. Given a field F, we denote by F a transversal for (F∗ )2 in F∗ . Also, for
any ρ ∈ F , we denote by Aρ (F) a fixed set of representatives for the equivalence
classes with respect to the following relation >ρ defined on A(F) = (F \ {1}) × F:
given (β1 , β2 ), (β3 , β4 ) ∈ A(F), we write (β1 , β2 ) >ρ (β3 , β4 ) if and only if either
(β3 , β4 ) = (β1 , ±β2 ) or
(1.2)
β3 =
β 1 t2 + β 2 t − ρ
t2 + β2 t − ρβ1
and
β4 = ±
β2 t2 − 2ρ(β1 + 1)t − ρβ2
t2 + β2 t − ρβ1
for some t ∈ F such that (t2 + ρ)(t2 + β2 t − ρβ1 ) 6= 0. If char F = 2, we fix two
additional sets B(F) and C(F). The set B(F) is a transversal for the subgroup
B(F) = {λ2 + λ : λ ∈ F} in F. The set C(F) is a set of representatives for the
equivalence classes with respect to the following relation ⋆ defined on F : given
x2 +x2 ρ
ρ1 , ρ2 ∈ F , we write ρ1 ⋆ ρ2 if and only if ρ2 = x12 +x22 ρ11 for some x1 , x2 , x3 , x4 ∈ F
3
4
such that x1 x4 6= x2 x3 . For convenience, we assume 1 ∈ F and 0 ∈ B(F).
We can now state our main result.
Theorem 1.1. The distinct isomorphism classes of four dimensional nilpotent associative algebras over a field F can be represented by the algebras described in the
second column of Table 1 (abelian case) and Table 2 (nonabelian case).
In particular, there are exactly
9
if F = F is algebraically closed and char F 6= 2,
10 if F = F is algebraically closed and char F = 2,
11 if F = Fq is finite,
12 if F = R,
nonisomorphic abelian nilpotent associative F-algebras of dimension four. We point
out that the classification given in Theorem 1.1 for the abelian case was already
obtained, using other techniques, by De Graaf [2] (if F = Fq or R) and by Poonen
[7] (for algebraically closed fields). It is also interesting to observe that if F = Fq
is a finite field, then there are exactly 5q + 9 (if q is odd) or 5q + 6 (if q is even)
nonisomorphic nonabelian nilpotent associative Fq -algebras of dimension four (see
Proposition 5.23).
2. Preliminary results
Our classification of regular subgroups in ∆4 (F) basically relies on the key results
of [4, 5] that we recall in this section.
Proposition 2.1. [5, Theorems 2.4 and 4.4] Let R ∈ ∆n (F). Then,
(a) R is a unipotent subgroup;
(b) up to conjugation under AGLn (F), the center Z(R) of R contains a nontrivial element z that coincides with its Jordan form diag (Jn1 , . . . , Jnk ) having
upper unitriangular Jordan blocks Jni of respective sizes ni ≥ ni+1 for all
i ≥ 1.
ISOMORPHISM CLASSES OF FOUR DIMENSIONAL NILP. ASSOC. ALGEBRAS
R
R − I5
0 x1 x2 x3 x4
0
0
x
x
x
1
2
3
0
0
0
x
x
1
2
0
0
0
0
x1
0
0
0
0
0
0 x1 x2 x3 x4
0
0
x1 x2
0
0
0
0
x1
0
0
0
0
0
0
0
0
0
0
0
0 x1 x2 x3 x4
0
0
x1
0
x2
0
0
0
0
x1
0
0
0
0
x3
0
0
0
0
0
0 x1 x2 x3 x4
0
0
0
x1 x2
0
0
0
0
x1
0
0
0
0
0
0
0
0
0
0
0 x1 x2
x3
x4
0
0
0
x1
x2
0
0
0
ρx2 x1
0
0
0
0
0
0
0
0
0
0
0 x1 x2
x3
x4
0
0
0
x1
x2
0
0
0
ǫx
x
+ x2
2
1
0
0
0
0
0
0
0
0
0
0
x1 x2 x3
x4
0
0
0
α1 x1 + α2 x2 + α3 x3
0
0
0
α2 x1 + β2 x2 + β3 x3
0
0
0
α3 x1 + β3 x2 + γ3 x3
3
Conditions
S(5)
S(4,1)
♯
S(3,2)
U1 (0, 0)
U1 (0, ρ)
U1 (1, ǫ)
RD
0
0
0
0
0
0
0
0
char F 6= 2
ρ ∈ F
ρ ∈ C(F)
if char F 6= 2
if char F = 2
char F = 2, ǫ ∈ B(F)
α1
α2
α3
α2
β2
β3
α3
β3 ∈ ΠS (F)
γ3
0
Table 1. Four dimensional abelian nilpotent associative Falgebras and abelian regular subgroups of AGL4 (F).
If δR is a linear function then a regular subset R of AGLn (F) is a subgroup if
and only if
δR (vδR (w)) = δR (v)δR (w),
for all v, w ∈ Fn ,
see [5]. Also, given v, w ∈ Fn ,
µR (v)µR (w) = µR (w)µR (v)
if and only if
vδR (w) = wδR (v).
We also recall that two subgroups R1 , R2 ∈ ∆n (F) are conjugate in AGLn (F) if
and only if there exists a matrix g = diag(1, ḡ) with ḡ ∈ GLn (F) such that R1g = R2
(see [5, Proposition 3.4]). On the other hand it is useful, mainly to prove that two
regular subgroups are not conjugate, to introduce the following three parameters.
4
MARCO ANTONIO PELLEGRINI
R
0
0
0
0
0
0
0
0
0
0
U2 (0, 1, 0, 1, 1)
U2 (0, 1, 0, 0, 1)
U3 (0, 1, 0)
U3 (0, 1, 1)
U3 (1, λ, 0)
U4 (ρ, β1 , β2 )
U5 (1, 1, ǫ)
RD
x1
0
0
0
0
x1
0
0
0
0
R − I5
x2 x3
x4
x1
0
x2 + x3
0
0
x1
0
0
x3
0
0
0
x2 x3
x4
x1
0
x2 + x3
0
0
x1
0
0
0
0
0
0
x1 x2 x3 x4
0
x1
0
0
0
0
0
0
0
0
0
x1
0
0
0
0
0
0
0
0
0
0 x1 x2 x3
x4
0
0
x1
0
0
0
0
0
0
0
0
0
0
0
x1 + x3
0
0
0
0
0
0 x1 x2 x3
x4
0
0
x1
0
x3
0
0
0
0
0
0
0
0
0
λx1
0
0
0
0
0
1 x1 x2
x3
x4
0
1
0
x1
x2
0
0
1
ρx
β
x
+
β
x
2
1
1
2
2
0
0
0
1
0
0
0
0
0
1
0 x1 x2 x3
x4
0
0
0
x2 x1 + x2
0
0
0
x1
ǫx2
0
0
0
0
0
0
0
0
0
0
0 x1 x2 x3
x4
0
0
0
0
α1 x1 + α2 x2 + α3 x3
0
0
0
0
β 1 x1 + β 2 x2 + β 3 x3
0
0
0
0
γ1 x1 + γ2 x2 + γ3 x3
0
0
0
0
Conditions
λ ∈ F \ {1}
ρ ∈ F , (β1 , β2 ) ∈ Aρ (F)
char F = 2, ǫ ∈ B(F)
α1
β1
γ1
α2
β2
γ2
α3
β3 ∈ ΠA (F)
γ3
0
Table 2. Four dimensional nonabelian nilpotent associative Falgebras and nonabelian regular subgroups in ∆4 (F).
Let H ∈ {R, Z(R)}, where R ∈ ∆n (F). Then we may define:
d(H) =
r(H) =
k(H) =
max{deg minF (h − In+1 ) : h ∈ H};
max{rk(h − In+1 ) : h ∈ H};
dim Ker (δR |H ).
The last parameter is justified since the set {v ∈ Fn : µR (v) ∈ Z(R)} is a subspace
of Fn (see [4, Proposition 2.1]).
We will proceed considering the possible values of (d(Z(R)), r(Z(R)) for R ∈
∆4 (F) and working in the centralizer of a Jordan form as described in Proposition
2.1(b). Unfortunately, the case r(Z(R)) = 2 will require a different approach, based
ISOMORPHISM CLASSES OF FOUR DIMENSIONAL NILP. ASSOC. ALGEBRAS
5
on the classification of the regular subgroups in ∆3 (F) obtained in [5]. In particular,
we will make use of the following observation.
e ∈ ∆n (F), n ≥ 2, can be written as
Any regular subgroup R
X
xn
1
e = R(R, D) = 0 In−1 + δR (X) DX T : X ∈ Fn−1 , xn ∈ F ,
(2.1)
R
0
0
1
for some R ∈ ∆n−1 (F) and some square matrix D ∈ Matn−1 (F) such that
(2.2)
T
DδR (ei )T eT
j = δR (ej )Dei
for all i, j = 1, . . . , n − 1
({e1 , . . . , en−1 } denotes the canonical basis of Fn−1 ), see [4]. In particular, we can
take R = Tn−1 : in this case any square matrix D ∈ Matn−1 satisfies (2.2). We set
X
xn
1
(2.3)
RD := R(Tn−1 , D) = 0 In−1 DX T : X ∈ Fn−1 , xn ∈ F .
0
0
1
The study of these subgroups RD is justified mainly because of their connection
with projectively congruent matrices.
Definition 2.2. [9] Two matrices A, B ∈ Matn (F) are said to be projectively congruent if P AP T = λB, for some non-zero element λ ∈ F∗ and some invertible
matrix P ∈ GLn (F).
Lemma 2.3. [4, Lemma 3.2] Given two matrices A, B ∈ Matn−1 (F), the subgroups
RA and RB are conjugate in AGLn (F) if and only if A and B are projectively
congruent.
Furthermore, it is easy to see that
2 if DT = −D and D has zero diagonal,
d(RD ) =
3 otherwise,
2 if D =
6 0,
r(RD ) =
1 if D = 0,
k(RD ) =
n − rk(D).
A given nilpotent associative algebra N of dimension n can be embedded into
Matn+1 (F) via
0 mB
,
m 7→
0 δ(m)
where, for any m ∈ N , mB and δ(m) denote, respectively, the coordinate row
vector of m and the matrix of the right multiplication by m with respect to a
fixed basis B of N over F. Identifying N with its image, L = FIn+1 + N is a
split local subalgebra of Matn+1 (F) with Jacobson radical J(L) = N . The subset
R = {In+1 + m : m ∈ N } ⊆ L consists of invertible matrices and is closed under
multiplication, since δ(m1 δ(m2 )) = δ(m1 )δ(m2 ) for all m1 , m2 ∈ N . Hence, R is a
regular subgroup lying in ∆n (F), by [5, Lemma 2.1]. On the other hand, given a
regular subgroup R ∈ ∆n (F), we have that
(2.4)
LR = FIn+1 + R
is a split local F-algebra of dimension n, by [5, Theorem 3.3]. Notice that set
0
v
n
N = J(LR ) = R − In+1 =
:v∈F
0 δR (v)
6
MARCO ANTONIO PELLEGRINI
is a nilpotent associative F-algebra of dimension n.
Proposition 2.4. [4, 5] Let R1 , R2 ∈ ∆n (F) and LR1 , LR2 be as in (2.4). Then
the following conditions are equivalent:
(a) the subgroups R1 and R2 are conjugate in AGLn (F);
(b) the split local algebras LR1 and LR2 are isomorphic;
(c) the nilpotent associative algebras J(LR1 ) and J(LR2 ) are isomorphic.
For sake of brevity, we write
x1 x2 . . . xn
1
R=
0 In + δR (x1 , x2 , . . . , xn )
to indicate the regular subgroup
x1 x2 . . . xn
1
R=
: x1 , . . . , xn ∈ F .
0 In + δR (x1 , x2 , . . . , xn )
For all i ≤ n, we denote by Xi the matrix of R − In+1 obtained taking xi = 1 and
xj = 0 for all j 6= i. Finally, Ei,j is the elementary matrix having 1 at position
(i, j), 0 elsewhere.
3. The regular subgroups in ∆4 (F)
For classifying the regular subgroups R ∈ ∆4 (F), where F is any field, we consider
the possible values of (d(Z(R)), r(Z(R))). By Proposition 2.1 the subgroup R is
unipotent and its center contains an element conjugate in AGL4 (F) to one of the
following Jordan forms z:
z
d(Z(R)) r(Z(R))
J5
5
4
diag(J4 , I1 )
4
3
diag(J3 , J2 )
3
3
diag(J3 , I2 )
3
2
diag(J2 , J2 , I1 )
2
2
diag(J2 , I3 )
2
1
Some of these cases have been already studied in [5]: we recall here the results
obtained in Lemmas 5.2, 5.4 and 7.4. First of all, if d(Z(R)) ≥ 4, then R is abelian.
In particular, if d(R) = 5, then R is conjugate to
1 x1 x2 x3 x4
0 1 x1 x2 x3
1 x1 x2
S(5) =
0 0
.
0 0
0
1 x1
0 0
0
0
1
If d(R) = 4, we
the subgroup R
where
1
0
S(4,1) =
0
0
0
have two conjugacy classes of regular subgroups: when k(R) = 2
♯
is conjugate to S(4,1) and when k(R) = 1 it is conjugate to S(3,2)
,
x1
1
0
0
0
x2
x1
1
0
0
x3
x2
x1
1
0
x4
0
0
0
1
and
♯
S(3,2)
1
0
=
0
0
0
x1
1
0
0
0
x2
x1
1
0
0
x3
0
0
1
0
x4
x2
x1
.
x3
1
ISOMORPHISM CLASSES OF FOUR DIMENSIONAL NILP. ASSOC. ALGEBRAS
7
Now, if d(Z(R)) = r(Z(R)) = 3, then again R is abelian. Conjugating the subgroups
obtained in [5, Lemma 7.4] by g = diag(I2 , E1,2 + E2,1 , 1), we obtain that R is
conjugate to
1 x1 x2 x3
x4
0 1
0
x1
x2
(3.1)
U1 (α, β) = 0 0
1 βx2 x1 + αx2
0 0
0
1
0
0 0
0
0
1
for some α, β ∈ F.
If d(Z(R)) = 3 and r(Z(R)) = 2, we may assume that z = diag(J3 , I2 )g ∈ Z(R),
where g = diag(I2 , E1,3 + E2,1 + E3,2 ). So, working in CAGL4 (F) (z) and using the
linearity of δR we obtain that R is conjugate to a subgroup RD , where
1 0
0
D = 0 β2 β3 ,
0 γ2 γ3
for some β2 , β3 , γ2 , γ3 ∈ F.
If d(Z(R)) = r(Z(R)) = 2, we may assume that the center of R contains the
element z = diag(J2 , J2 , I1 )g , where g = diag(1, E1,2 + E2,1 , E1,2 + E2,1 ). Again,
working in CAGL4 (F) (z) and using the linearity of δR we obtain that R is conjugate
to the subgroup
1 x1 x2 x3
x4
0 1 ζx1 0 α1 x1 + x2 + α3 x3
1
0
x1
(3.2)
U2 (α1 , α3 , γ1 , γ3 , ζ) =
0 0
0 0
0
1
γ1 x1 + γ3 x3
0 0
0
0
1
for some α1 , α3 , γ1 , γ3 , ζ ∈ F.
Finally, if R is abelian and (d(R), r(R)) = (2, 1), then R = T4 by [5, Lemma 5.3].
4. The abelian case
In this section we assume that R ∈ ∆4 (F) is abelian. In view of the results
recalled in Section 3, we are left to determine the conjugacy classes of subgroups
of type U1 (α, β), U2 (α1 , α3 , γ1 , γ3 , ζ) and RD where D ∈ Mat3 (F) is a symmetric
matrix. We start with the subgroups U1 (α, β).
Lemma 4.1. Let R be an abelian regular subgroup of AGL4 (F) such that d(R) =
r(R) = 3. Then R is conjugate to exactly one of the following subgroups:
U1 (0, ρ) with ρ ∈ F ∪ {0}
if char F 6= 2;
U1 (1, ǫ) with ǫ ∈ B(F)
or
U1 (0, ρ) with ρ ∈ C(F)
if char F = 2.
Proof. By the previous argument, we can assume that R = U1 (α, β), as in (3.1), for
some α, β ∈ F. First, suppose that char F 6= 2 and set ∆ = α2 + 4β. If ∆ 6= 0, write
∆ = ω 2 ρ for some ω ∈ F∗ and a unique ρ ∈ F . We can define an epimorphism (of
algebras) Ψ : F[t1 , t2 ] → LU1 (α,β) by setting
Ψ(t1 ) = ωρX1 ,
Ψ(t2 ) = αX1 − 2X2 .
8
MARCO ANTONIO PELLEGRINI
We have Ker (Ψ) = ht31 , t32 , t21 −ρt22 i. The subgroup U1 (α, β) is conjugate to U1 (0, ρ)
by Proposition 2.4. Now, it is quite easy to see that two subgroups U1 (0, ρ1 ) and
U1 (0, ρ2 ) (ρ1 , ρ2 ∈ F ) are conjugate in AGL4 (F) if and only if ρ1 = ρ2 .
If α2 + 4β = 0, we consider the epimorphism Ψ : F[t1 , t2 ] → LU1 (α,β) defined by
Ψ(t1 ) = X1 ,
ht31 ,
Ψ(t2 ) = αX1 − 2X2 ,
t21 t2 ,
whose kernel is Ker (Ψ) =
t22 i. Again by Proposition 2.4, U1 (α, β) is
conjugate to U1 (0, 0) (which is not conjugate to U1 (0, ρ), ρ ∈ F ).
Next, suppose that char F = 2. If α 6= 0, take ǫ ∈ B(F) such that ǫ + αβ2 ∈ B(F).
Then, there exists r ∈ F such that r2 + αr + β + α2 ǫ = 0 (a polynomial t2 + t + λ is
reducible in F[t] if and only if λ ∈ B(F)). In this case, we can define an epimorphism
Ψ : F[t1 , t2 ] → LU1 (α,β) by setting
Ψ(t1 ) = rX1 + X2 ,
ht31 ,
t32 ,
t21
Ψ(t2 ) = αX1 .
ǫt22 i,
+ t1 t2 +
proving that U1 (α, β) is conjugate
We obtain Ker (Ψ) =
to U1 (1, ǫ). Now, U1 (1, ǫ1 ) and U1 (1, ǫ2 ) are conjugate in AGL4 (F) if and only if
ǫ1 + ǫ2 ∈ B(F). It follows that U (α, β) is conjugate to exactly one subgroup U1 (1, ǫ)
with ǫ ∈ B(F).
Finally, assume α = 0. For β = 0 observe that U1 (0, 0)g = U1 (0, 1), where
g = I5 + E3,2 + E5,4 . If β 6= 0, write β = ω 2 ρ for some ω ∈ F∗ and a unique ρ ∈ F .
An epimorphism Ψ : F[t1 , t2 ] → LU1 (0,β) can be defined by taking
Ψ(t1 ) = X2 ,
ht31 ,
t32 ,
t21
Ψ(t2 ) = ωX1 .
ρt22 i
and so U1 (0, β) is conjugate to U1 (0, ρ).
+
We have Ker (Ψ) =
Furthermore, given ρ1 , ρ2 ∈ F , the subgroup U1 (0, ρ1 ) is conjugate to U1 (0, ρ2 ) if
x2 +x2 ρ
and only if ρ1 ⋆ ρ2 , that is if and only if ρ2 = x21 +x22 ρ11 for some x1 , x2 , x3 , x4 ∈ F
3
4
such that x1 x4 6= x2 x3 (see the notation described in the Introduction).
To conclude we observe that subgroups U1 (0, ρ), ρ ∈ C(F), are not conjugate to
subgroups U1 (1, ǫ), ǫ ∈ B(F).
We can now classify the abelian regular subgroups of AGL4 (F).
Theorem 4.2. Let F be a field. The distinct conjugacy classes of abelian regular
subgroups of AGL4 (F) can be represented by the subgroups described in the first
column of Table 1.
Proof. Let R ≤ AGL4 (F) be an abelian regular subgroup. Clearly, 2 ≤ d(R) ≤ 5.
As seen in Section 3, if d(R) = 5 then R is conjugate to S(5) , and if d(R) = 4 then
♯
R is conjugate either to S(4,1) or to S(3,2)
.
Suppose that d(R) = 3. Then 2 ≤ r(R) ≤ 3. If r(R) = 3, we apply Lemma 4.1:
if char F 6= 2, R is conjugate to U1 (0, ρ) for a unique ρ ∈ F ∪ {0}; if char F = 2,
it is conjugate either to U1 (1, ǫ) for a unique ǫ ∈ B(F) or to U1 (0, ρ) for a unique
ρ ∈ C(F). If r(R) = 2, then R is conjugate to a subgroup RD for some D ∈ Mat3 (F).
Suppose d(R) = 2. Then 1 ≤ r(R) ≤ 2. If r(R) = 2 we apply [5, Lemma
7.7], obtaining that char F = 2 and R is conjugate to U2 (0, 0, 0, 0, 0) = RD , where
D = E1,2 + E2,1 . If r(R) = 1, then U = T4 = R0 , where 0 ∈ Mat3 (F) is the zero
matrix.
Notice that in three cases, R is conjugate to a subgroup RD . By Lemma 2.3 the
statement of the theorem is proved considering a set ΠS (F) of representatives for
projective congruent classes of symmetric matrices in Mat3 (F).
ISOMORPHISM CLASSES OF FOUR DIMENSIONAL NILP. ASSOC. ALGEBRAS
9
We now provide explicit representatives for algebraically closed fields, for R and
for finite fields.
Corollary 4.3. Let F = F be an algebraically closed field. If char F 6= 2, there are
exactly 9 distinct conjugacy classes of abelian regular subgroups of AGL4 (F), that
can be represented by
♯
S(5) , S(4,1) , S(3,2)
, U1 (0, 0), U1 (0, 1), RE1,1 , RE1,1 +E2,2 , RI3 , T4 .
If char F = 2, there are exactly 10 distinct conjugacy classes of abelian regular
subgroups of AGL4 (F), that can be represented by
♯
S(5) , S(4,1) , S(3,2)
, U1 (0, 0), U1 (1, 0), RE1,1 , RE1,1 +E2,2 , RE2,3 +E3,2 , RI3 , T4 .
Proof. Since F is algebraically closed, F = {1} and, when char F = 2, B(F) = {0}
and C(F) = {1}. The set ΠS (F) can be obtained applying, for instance, [3, Theorem
2.1]. Hence, the statement is proved noticing that, as observed in the proof of
Lemma 4.1, when char F = 2 the subgroups U1 (0, 0) and U1 (0, 1) are conjugate.
Corollary 4.4. There are exactly 12 distinct conjugacy classes of abelian regular
subgroups of AGL4 (R), that can be represented by
♯
S(5) , S(4,1) , S(3,2)
, U1 (0, 0), U1 (0, 1), U1 (0, −1),
RE1,1 , RE1,1 +E2,2 , RE1,1 −E2,2 , RE1,1 +E2,2 −E3,3 , RI3 , T4 .
Proof. First of all, we can take R = {1, −1}. The set ΠS (F) can be obtained
again applying [3, Theorem 2.1], noticing that two matrices A, B ∈ Matn (F) are
projectively congruent if and only if A is congruent to ±B.
Corollary 4.5. Let F = Fq be a finite field. There are exactly 11 distinct conjugacy
classes of abelian regular subgroups of AGL4 (q), that can be represented by:
(a) for q odd,
♯
S(5) , S(4,1) , S(3,2)
, U1 (0, 0), U1 (0, 1), U1 (0, ξ), RD ,
where x2 − ξ is a fixed irreducible polynomial of Fq [x] and
D ∈ ΠS (Fq ) = {0, E1,1 , E1,1 + E2,2 , E1,1 + ξE2,2 , I3 };
(b) for q even,
♯
S(5) , S4,1 , S(3,2)
, U1 (0, 0), U1 (1, 0), U1 (1, ξ), RD ,
where x2 + x + ξ is a fixed irreducible polynomial of Fq [x] and
D ∈ ΠS (Fq ) = {0, E1,1 , E1,1 + E2,2 , E2,3 + E3,2 , I3 }.
Proof. For q odd we have F = {1, ξ} and for q even we have F = C(F) = {1}
and B(F) = {0, ξ}. The set ΠS (Fq ) has been determined in [9, Theorem 4].
10
MARCO ANTONIO PELLEGRINI
5. The nonabelian case
In this section we assume that R ∈ ∆4 (F) is nonabelian. We are reduced to
study the cases when (d(Z(R)), r(Z(R))) = (3, 2), (2, 2) or (2, 1). We recall that if
(d(Z(R)), r(Z(R))) = (3, 2) then R is conjugate to a subgroup of shape RD .
If d(Z(R)) = r(Z(R)) = 2, then R is conjugate to U2 (α1 , α3 , γ1 , γ3 , ζ) as in (3.2).
Clearly, U2 (α1 , α3 , γ1 , γ3 , 0) = RD , where
α1 1 α3
D = 1 0 0 .
γ1 0 γ3
So, we have to consider now the case ζ 6= 0. Notice that since R is nonabelian, we
must have γ1 6= α3 . Furthermore, we have d(R) = 4 and r(R) = 3.
Lemma 5.1. There are exactly 2 conjugacy classes of nonabelian regular subgroups R ∈ ∆4 (F) such that d(Z(R)) = r(Z(R)) = 2, not conjugate to RA for
any A ∈ Mat3 (F). Such classes can be represented by the subgroups U2 (0, 1, 0, 1, 1)
and U2 (0, 1, 0, 0, 1).
Proof. By the previous considerations, we may assume that R = U2 (α1 , α3 , γ1 , γ3 , ζ)
where ζ 6= 0 and γ1 6= α3 .
First, suppose γ3 6= 0: in this case k(R) = 1. Consider the algebra L =
LU2 (0,1,0,1,1) defined in (2.4):
L = Span(t1 , t2 ) such that: t31 = t22 = t1 t2 ;
t2 t1 = 0.
The function Ψ : L → LR defined by
Ψ(t1 ) =
(α3 − γ1 )2
X1 ,
γ3 ζ
Ψ(t2 ) =
−γ1 (α3 − γ1 )3
(α3 − γ1 )3
X2 +
X3 ,
2
γ3 ζ
γ32 ζ
is an isomorphism: by Proposition 2.4, R is conjugate to U2 (0, 1, 0, 1, 1).
Suppose now that γ3 = 0: in this case k(R) = 2 and so R is not conjugate to
U2 (0, 1, 0, 1, 1). Consider the algebra L = LU2 (0,1,0,0,1) :
L = Span(t1 , t2 ) such that: t31 = t1 t2 ;
t22 = t2 t1 = 0.
The function Ψ : L → LR defined by
Ψ(t1 ) = (α3 − γ1 )X1 ,
Ψ(t2 ) = −ζγ1 (α3 − γ1 )X2 + ζ(α3 − γ1 )X3
is an isomorphism and so R is conjugate to U2 (0, 1, 0, 0, 1).
We now consider the case when d(Z(R)) = 2 and r(Z(R)) = 1. Instead of
working in CAGL4 (F) (diag(J2 , I3 )), it is more convenient to write R as R(R̃, D), for
some R̃ ∈ ∆3 (F) and D ∈ Mat3 (F), see (2.1). The regular subgroups in ∆3 (F) have
been classified in [5].
Theorem 5.2. [5, Section 7.2] Let F be a field. The distinct conjugacy classes of
regular subgroups in ∆3 (F) are represented by
S(4) ,
S(3,1) ,
Rρ (ρ ∈ F ),
U13 (if char F = 2),
T3 ,
N1 ,
N3,λ (λ ∈ F∗ ),
N2 (if char F 6= 2).
ISOMORPHISM CLASSES OF FOUR DIMENSIONAL NILP. ASSOC. ALGEBRAS
11
Suppose that R = R(S(4) , D) for some D ∈ Mat3 (F). By (2.2), D = α1 E1,1 +
α2 (E1,2 E2,1 ) + α3 (E1,3 + E2,2 + E3,1 ) for some α1 , α2 , α3 ∈ F. It follows that R is
abelian. Furthermore, R(T3 , D) = RD . Hence, we have to consider the six subcases
corresponding to R̃ = S(3,1) , Rρ , U13 , N1 , N2 , N3,λ .
Lemma 5.3. There are exactly |F| + 1 distinct conjugacy classes of nonabelian
subgroups R = R(S(3,1) , D) ∈ ∆4 (F) such that d(Z(R)) = 2 and r(Z(R)) = 1. Such
classes can be represented by the subgroups corresponding to
D
d(R) r(R)
E3,1
3
3
E3,1 + E3,3
3
3
E1,3 + λE3,1 (λ ∈ F \ {0, 1})
3
3
E1,3
3
2
k(R)
3
2
2
2
These subgroups are not conjugate to RA for any A ∈ Mat3 (F).
Proof. Let R = R(S(3,1) , D) for some D ∈ Mat3 (F) such that d(Z(R)) = 2 and
r(Z(R))
= 1.
By (2.2) and the hypothesis that R is nonabelian, we obtain D =
ᾱ1 ᾱ2 ᾱ3
ᾱ2 0
0 with γ̄1 6= ᾱ3 . Moreover, the condition r(Z(R)) = 1 implies ᾱ2 = 0.
γ̄1 0 γ̄3
Then, R is conjugate via g = I5 − ᾱ1 E3,5 to
1 x1 x2 x3
x4
0 1 x1 0
α3 x3
,
1
0
0
U3 (α3 , γ1 , γ3 ) = 0 0
0 0
0
1 γ1 x1 + γ3 x3
0 0
0
0
1
for some α3 , γ1 , γ3 ∈ F such that γ1 6= α3 .
Suppose that γ3 6= 0. Then d(R) = r(R) = 3 and k(R) = 2. Consider the algebra
L = LU3 (0,1,1) :
L = Span(t1 , t2 ) such that: t2 t1 = t22 ;
t31 = t1 t2 = 0.
The function Ψ : L → LR defined by
Ψ(t1 ) = γ3 X1 − α3 X3 ,
Ψ(t2 ) = (γ1 − α3 )X3
is an isomorphism and so R is conjugate to U3 (0, 1, 1).
Suppose now that γ3 = α3 = 0 (and so γ1 6= 0). Then d(R) = r(R) = k(R) = 3.
In this case, consider the algebra L = LU3 (0,1,0) :
L = Span(t1 , t2 ) such that: t31 = t22 = t1 t2 = t2 t21 = 0.
The function Ψ : L → LR defined by
Ψ(t1 ) = X1 ,
Ψ(t2 ) = X3
is an isomorphism: it follows that R is conjugate to U3 (0, 1, 0).
Finally, suppose that γ3 = 0 and α3 6= 0. For any λ ∈ F \ {1}, consider the
algebra Lλ = LU3 (1,λ,0) :
Lλ = Span(t1 , t2 ) such that: t31 = t22 = t21 t2 = 0;
t2 t1 = λt1 t2 .
12
MARCO ANTONIO PELLEGRINI
The function Ψ : Lλ → LR defined by
γ1
α3
is an isomorphism and hence R is conjugate to U3 (1, λ, 0) for some λ 6= 1. Now,
notice that if λ = 0 then d(U3 (1, 0, 0)) = 3 and r(U3 (1, 0, 0)) = k(U3 (1, 0, 0)) = 2.
Otherwise, d(U3 (1, λ, 0)) = r(U3 (1, λ, 0)) = 3 and k(U3 (1, λ, 0)) = 2. Furthermore,
U3 (1, 0, 0) is not conjugate to any RA (direct computations) and the same holds for
the other subgroups, since r(RA ) ≤ 2. The subgroups U3 (1, λ1 , 0) and U3 (1, λ2 , 0)
are conjugate if and only if λ1 = λ2 . To conclude, we observe that U3 (1, λ, 0) is not
conjugate to U3 (0, 1, 1).
Ψ(t1 ) = X1 ,
Ψ(t2 ) = X3 ,
λ=
Before to consider the next subcase, we recall the following notation given in the
Introduction. For any ρ ∈ F , we denote by Aρ (F) a fixed set of representatives
for the equivalence classes with respect to the following relation >ρ defined on
A(F) = (F \ {1}) × F: given (β1 , β2 ), (β3 , β4 ) ∈ A(F), we write (β1 , β2 ) >ρ (β3 , β4 ) if
and only if either (β3 , β4 ) = (β1 , ±β2 ) or
(5.1)
β3 =
β 1 t2 + β 2 t − ρ
t2 + β2 t − ρβ1
and
β4 = ±
β2 t2 − 2ρ(β1 + 1)t − ρβ2
t2 + β2 t − ρβ1
for some t ∈ F such that (t2 + ρ)(t2 + β2 t − ρβ1 ) 6= 0.
Lemma 5.4. The distinct conjugacy classes of nonabelian subgroups R = R(Rρ , D)
∈ ∆4 (F) such that d(Z(R)) = 2 and r(Z(R)) = 1 can be represented by
R(Rρ , E1,2 + β1 E2,1 + β2 E2,2 )
with ρ ∈ F and (β1 , β2 ) ∈ Aρ (F). Furthermore, d(R) = r(R) = 3 and k(R) = 2.
Proof. Let R = R(Rρ , D) ∈ ∆4 (F), with ρ ∈ F and D ∈ Mat3 (F), be a nonabelian
subgroup such that d(Z(R)) = 2 andr(Z(R)) = 1.
By (2.2) and the hypothesis
ᾱ1 ᾱ2 0
that R is nonabelian, we have D = β̄1 β̄2 0, with β̄1 6= ᾱ2 . Taking g =
0
0 0
1 −ᾱ1 ᾱ2−1
0 1
ρ −β̄2 β̄1−1
diag I3 ,
if ᾱ2 6= 0 and g = diag 1,
,
if
ρ 0
0
ᾱ−1
0
ρβ̄1−1
2
ᾱ2 = 0, we obtain that R is conjugate via g to the subgroup
1 x1 x2 x3
x4
0 1
0
x1
x2
0
0
1
ρx
β
x
+
β
x
where β1 6= 1.
U4 (ρ, β1 , β2 ) =
2
1 1
2 2 ,
0 0
0
1
0
0 0
0
0
1
Let ρ1 , ρ2 ∈ F , β1 , β2 , β3 , β4 ∈ F with β1 , β3 6= 1. It is quite easy to verify that
the subgroups U4 (ρ1 , β1 , β2 ) and U4 (ρ2 , β3 , β4 ) are conjugate in AGL4
(F) if and
a1 a2
,
only if there exists an element g = diag(1, A, B) ∈ GL5 (F), with A =
a3 a4
b1 b2
B =
, such that U4 (ρ1 , β1 , β2 )g = U4 (ρ2 , β3 , β4 ). This holds if and only
b3 b4
if b1 = ρ2 a22 + a21 , b2 = (β3 + 1)a1 a2 + β4 a22 , b3 = ρ2 a2 a4 + a1 a3 , b4 = β3 a2 a3 +
β4 a2 a4 + a1 a4 and
(5.2)
a1 a3 + a2 a4 ρ2 = 0,
ISOMORPHISM CLASSES OF FOUR DIMENSIONAL NILP. ASSOC. ALGEBRAS
13
(5.3)
(β1 β3 − 1)a2 a3 + β4 (β1 − 1)a2 a4 + (β1 − β3 )a1 a4 = 0,
ρ1 ρ2 a22 + ρ1 a21 − ρ2 a24 − a23 = 0,
β2 (β3 a2 a3 + β4 a2 a4 + a1 a4 ) + β4 (a22 ρ1 − a24 ) + (β3 + 1)(a1 a2 ρ1 − a3 a4 ) = 0
provided that
(a1 a4 − a2 a3 )(a21 + a1 a2 β4 − a22 ρ2 β3 ) 6= 0.
(5.4)
Now, if a2 = 0, from (5.2) and (5.4) we get a3 = 0 and in this case (5.3) gives
β3 = β1 , ρ2 = ρ1 and β4 = ±β2 (notice that, taking g = diag(1, 1, −1, 1, −1),
we have U4 (ρ1 , β1 , β2 )g = U4 (ρ1 , β1 , −β2 )). Hence, we may suppose a2 6= 0 and
a4 = −a1 a3 (a2 ρ2 )−1 by (5.2). In this case (5.3) gives ρ2 = ρ1 , a3 = ±a2 ρ1 , f0 = 0
and f+ = 0 or f− = 0, where
f0
= a21 (β1 − β3 ) + a1 a2 β4 (β1 − 1) − a22 ρ1 (β1 β3 − 1),
f±
= a21 (β2 ± β4 ) + a1 a2 (β2 β4 ∓ 2ρ1 (β3 + 1)) − a22 ρ1 (β2 β3 ± β4 )
provided that
(5.5)
(a21 + a22 ρ1 )(a21 + a1 a2 β4 − a22 ρ1 β3 ) 6= 0.
Taking a2 = ±1, it follows that U4 (ρ1 , β1 , β2 ) is conjugate to U4 (ρ2 , β3 , β4 ) if and
only if ρ2 = ρ1 and either (β3 , β4 ) = (β1 , ±β2 ) or
2
β2 t2 − 2ρ1 (β1 + 1)t − ρ1 β2
β1 t + β2 t − ρ1
, ±
(β3 , β4 ) = 2
t + β2 t − ρ1 β1
t2 + β 2 t − ρ 1 β 1
for some t ∈ F such that (t2 + ρ)(t2 + β2 t − ρ1 β1 ) 6= 0. In other words, this holds if
and only if ρ2 = ρ1 and (β1 , β2 ) >ρ1 (β3 , β4 ).
Remark 5.5. Let ρ ∈ F , β1 , β3 ∈ F \ {1} and β2 , β4 ∈ F. Then:
(a) U4 (ρ, β1 , β2 ) and U4 (ρ, −1, 0) are conjugate if and only if (β1 , β2 ) = (−1, 0);
(b) U4 (ρ, β1 , β2 ) and U4 (ρ, β1 , β4 ) are conjugate if and only if β4 = ±β2 ;
(c) U4 (ρ, β1 , 0) and U4 (ρ, β3 , 0) are conjugate if and only if either β3 = β1 or
β1 6= 0 and β3 = β1−1 ;
(d) U4 (ρ, β1 , 0) and U4 (ρ, 0, β4 ) are conjugate if and only if β1 = ρω 2 and
2
6= ±1.
β4 = ρω2ρω
2 −1 for some ω ∈ F such that ρω
Remark 5.6. It will be useful to give the presentation of the following algebras:
(a) LU4 (ρ,0,λ) = Span(t1 , t2 ) such that: t31 = t2 t1 = 0; t22 = ρt21 + λt1 t2 ;
(b) LU4 (ρ,−1,0) = Span(t1 , t2 ) such that: t31 = t1 t2 + t2 t1 = 0; t22 = ρt21 .
Lemma 5.7. Assume char F = 2. There are exactly |B(F)| distinct conjugacy
classes of nonabelian subgroups R = R(U13 , D) ∈ ∆4 (F) such that d(Z(R)) = 2 and
r(Z(R)) = 1. Such classes can be represented by the subgroups corresponding to
D = E1,1 + E1,2 + λE2,2 (λ ∈ B(F)).
Furthermore, d(R) = r(R) = 3 and k(R) = 2.
Proof. Let R = R(U13 , D) for some
3 (F). By (2.2) and the hypothesis that
D ∈ Mat
ᾱ1 ᾱ2 0
R is nonabelian, we have D = β̄1 β̄2 0 with β̄1 6= ᾱ2 . Then, R is conjugate
0
0 0
14
MARCO ANTONIO PELLEGRINI
via g = I5 + β̄1 E4,5 to
1 x1
0 1
U5 (α1 , α2 , β2 ) =
0 0
0 0
0 0
x2
0
1
0
0
x3
x2
x1
1
0
x4
α1 x1 + α2 x2
,
β2 x2
0
1
where α2 6= 0. Then d(R) = r(R) = 3 and k(U ) = 2.
For any λ ∈ F, consider the algebra Lλ = LU5 (1,1,λ) :
Lλ = Span(t1 , t2 ) such that: t31 = t32 = t1 t2 + t2 t1 + t21 = 0;
t22 = λt21
(clearly, when λ = 0, the condition t32 = 0 can be omitted). We now define an
isomorphism of split local algebras Ψ : Lλ → LR in the following way:
Ψ(t1 ) = α2 X1 + (α1 + α2 )X2 , Ψ(t2 ) = α2 X2 ,
if β2 = 0,
Ψ(t1 ) = α2 X2 ,
Ψ(t2 ) = β2 X1 ,
if β2 6= 0,
where λ =
α1 β2
.
α22
Hence, R is conjugate to U5 (1, 1, λ) for some λ ∈ F. Now,
U5 (1, 1, λ1 ) and U5 (1, 1, λ2 ) are conjugate if and only if λ1 +λ2 ∈ B(F). We conclude
that are exactly |B(F)| conjugacy classes.
Lemma 5.8. Let R = R(N1 , D) ∈ ∆4 (F) be a nonabelian subgroup such that
d(Z(R)) = 2 and r(Z(R)) = 1. Then R is conjugate to one of the following
subgroups: RE2,1 , U3 (1, 0, 0), U3 (0, 1, 0), U3 (0, 1, 1), U4 (ρ, −1, 0), U4 (ρ, 0, λ) or
U5 (1, 1, 0), for some ρ ∈ F and λ ∈ F.
ᾱ1 ᾱ2 0
Proof. Let R = R(N1 , D) for some D ∈ Mat3 (F). By (2.2), D = β̄1 β̄2 0.
0
0 0
Then, R is conjugate via g = I5 − β̄1 E4,5 to
1 x1 x2 x3
x4
0 1
0
0 α1 x1 + α2 x2
,
U6 (α1 , α2 , β2 ) = 0 0
1 x1
β2 x2
0 0
0
1
0
0 0
0
0
1
for some α1 , α2 , β2 ∈ F. It is easy to see that U6 (α1 , α2 , β2 ) is conjugate to some
RA if and only if α1 = α2 = β2 = 0. More in detail, U6 (0, 0, 0)g = RE2,1 , where
g = diag(I3 , E1,2 + E2,1 ).
If α1 = α2 = 0 and β2 6= 0, then d(R) = 3 and r(R) = k(R) = 2. The function
Ψ : LU3 (1,0,0) → LR defined by
Ψ(t1 ) = X2 ,
Ψ(t2 ) = X1
is an isomorphism. By Proposition 2.4, R is conjugate to the subgroup U3 (1, 0, 0).
If α1 6= 0 and α2 = β2 = 0, then d(R) = r(R) = k(R) = 3. The function
Ψ : LU3 (0,1,0) → LR defined by
Ψ(t1 ) = X1 ,
Ψ(t2 ) = X2
is an isomorphism and so R is conjugate to U3 (0, 1, 0).
ISOMORPHISM CLASSES OF FOUR DIMENSIONAL NILP. ASSOC. ALGEBRAS
15
For the other values of α1 , α2 , β2 we have d(R) = r(R) = 3 and k(R) = 2. If
α1 , β2 6= 0 and α2 = 0, write αβ21 = ω 2 ρ for some ω ∈ F∗ and a unique ρ ∈ F : the
function Ψ : LU4 (ρ,0,0) → LR defined by
Ψ(t1 ) = ωX2 ,
Ψ(t2 ) = X1
is an isomorphism. Hence, R is conjugate to U4 (ρ, 0, 0).
We are left to consider the case where α2 6= 0. Suppose α1 = β2 = 0. If
char F 6= 2, write −1 = ω 2 ρ for some ω ∈ F∗ and a unique ρ ∈ F : the function
Ψ : LU4 (ρ,−1,0) → LR defined by
Ψ(t1 ) = ωX1 + ωX2 ,
Ψ(t2 ) = X1 − X2
is an isomorphism and hence R is conjugate to U4 (ρ, −1, 0). If char F = 2, the
function Ψ : LU5 (1,1,0) → LR defined by
Ψ(t1 ) = X1 + X2 ,
Ψ(t2 ) = X2
is an isomorphism and hence R is conjugate to U5 (1, 1, 0).
If (α1 , β2 ) 6= (0, 0), set ∆ = α1 β2 − α22 . If ∆ = 0 (and so β2 6= 0) then the
function Ψ : LU3 (0,1,1) → LR defined by
Ψ(t1 ) = β2 X1 ,
Ψ(t2 ) = β2 X1 − α2 X2
is an isomorphism. Hence, U is conjugate to U3 (0, 1, 1). If ∆ 6= 0, write
∆
α22
= ω2ρ
for some ω ∈ F∗ and a unique ρ ∈ F . The function Ψ : LU4 (ρ,0,ω−1 ) → LR defined
by
Ψ(t1 ) = α2 ωX2 ,
Ψ(t2 ) = −β2 X1 + α2 X2 if β2 6= 0,
Ψ(t1 ) = α2 ωX1 − α1 ωX2 , Ψ(t2 ) = α2 X1
if β2 = 0,
is an isomorphism. Hence, R is conjugate to U4 (ρ, 0, ω −1 ).
Lemma 5.9. Assume char F 6= 2. Let R = R(N2 , D) ∈ ∆4 (F) be a nonabelian
subgroup such that d(Z(R)) = 2 and r(Z(R)) = 1. Then R is conjugate to one of
the following subgroups: R(E1,2 −E2,1 ) , U3 (1, −1, 0) or U4 (ρ, −1, 0) for some ρ ∈ F .
Proof.
ᾱ1
β̄1
0
Let
ᾱ2
β̄2
0
R= R(N2 , D) for some D ∈ Mat3 (F). By (2.2) we must have D =
0
0. Then, R is conjugate via g = I5 − β̄1 E4,5 to
0
1 x1 x2 x3
x4
0 1
0 −x2 α1 x1 + α2 x2
0
0
1
x1
β2 x2
U7 (α1 , α2 , β2 ) =
0 0
0
1
0
0 0
0
0
1
for some α1 , α2 , β2 ∈ F. It is easy to see that U7 (α1 , α2 , β2 ) is conjugate to some
RA if and only if α1 = α2 = β2 = 0. In this case d(R) = r(R) = k(R) = 2 and
U7 (0, 0, 0)g = R(E1,2 −E2,1 ) where diag(I3 , (E2,1 − E1,2 )).
Suppose now (α1 , α2 , β2 ) 6= (0, 0, 0) and let ∆ = 4α1 β2 − α22 . Notice that in this
case d(R) = r(R) = 3 and k(R) = 2. Assume ∆ = 0 and consider the function
Ψ : LU3 (1,−1,0) → LR given by
Ψ(t1 ) = X1 , Ψ(t2 ) = α2 X1 − 2α1 X2 if α1 6= 0,
Ψ(t1 ) = X2 , Ψ(t2 ) = X1
if α1 = α2 = 0.
16
MARCO ANTONIO PELLEGRINI
Since Ψ is an isomorphism, we obtain that U7 (α1 , α2 , β2 ) is conjugate to U3 (1, −1, 0).
Now, assume ∆ 6= 0 and write ∆ = ω 2 ρ for some ω ∈ F∗ and a unique ρ ∈ F . The
function Ψ : LU4 (ρ,−1,0) → LR given by
Ψ(t1 ) = ωX1 ,
Ψ(t2 ) = α2 X1 − 2α1 X2
if α1 6= 0,
Ψ(t1 ) = ω(β2 + 1)X1 − α2 ωX2 , Ψ(t2 ) = α2 (β2 − 1)X1 − α22 X2 if α1 = 0,
is an isomorphism: R is conjugate to U4 (ρ, −1, 0).
Lemma 5.10. Let R = R(N3,λ , D) ∈ ∆4 (F), λ ∈ F∗ , be a nonabelian subgroup
such that d(Z(R)) = 2 and r(Z(R)) = 1. Then, R is conjugate to one of the
following subgroups: RA , where A = E1,1 + E2,1 + λE2,2 , U3 (0, 1, 1), U3 (1, µ, 0) for
some µ ∈ F \ {0, 1}, U4 (ρ, β1 , β2 ) for some ρ ∈ F and β1 , β2 ∈ F, or U5 (1, 1, ω)
where ω ∈ F is such that ω + λ ∈ B(F).
Proof. Let
ᾱ1
D = β̄1
0
∗
R = R(N
3,λ , D) for some λ ∈ F and some D ∈ Mat3 (F). By (2.2)
ᾱ2 0
β̄2 0. Then, R is conjugate via g = I5 − β̄λ2 E4,5 to
0 0
1 x1 x2
x3
x4
0 1
0 x1 + x2 α1 x1 + α2 x2
0
0
1
λx2
β1 x1
Ũ8 (λ, α1 , α2 , β1 ) =
0 0
0
1
0
0 0
0
0
1
for some α1 , α2 , β1 ∈ F. It is easy to see that Ũ8 (λ, α1 , α2 , β1 ) is conjugate to some
RA if and only if α1 = α2 = β1 = 0. In this case d(R) = 3 and r(R) = k(R) = 2
and Ũ8 (λ, 0, 0, 0)g = RA , where g = E1,1 + E2,3 + E5,4 − λE2,2 + λE3,3 + λ2 E4,5
and A = E1,1 + E2,1 + λE2,2 .
So, suppose (α1 , α2 , β1 ) 6= (0, 0, 0): we have d(R) = r(R) = 3 and k(R) = 2. We
may also assume β1 = 1. Namely, if β1 6= 0 it suffices to conjugate Ũ8 (λ, α1 , α2 , β1 )
by g
= diag(1,
β1
, β1
, β12 , β1 ). If β1 = 0 and
α2 6= 0, then we conjugate by the matrix
1 −1
λ (λα1 − α2 )α−1
2
diag 1,
,
and if α2 = β1 = 0, we conjugate by
λ 0
0
−λα−1
2 2
λα1 −λα21
0
α1
.
,
the matrix diag 1,
0
λα1
−λα1 α1
Now, set U8 (λ, α1 , α2 ) = Ũ8 (λ, α1 , α2 , 1) and ∆ = −λ(α2 − 1)(α1 − α2 + 1) − 1.
Suppose α2 = 1. In this case ∆ = −1 that we write as −1 = ω 2 ρ for some ω ∈ F∗
α21 λ−α1 +2
1
,
, ω(α
and a unique ρ ∈ F . If α1 6= 0, 1, then U8 (λ, α1 , 1)g = U4 ρ, 1−α
1
1 −1)
where
!!
ωα21
α1 0
0 α1 −1
g = diag 1,
,
.
ωα1
1 ω
α1 1−α
1
, where
If α1 = 0 and char F 6= 2, then U8 (λ, 0, 1)g = U4 ρ, 1 − λ2 , − 2(λ+1)
ωλ
−1
4
0
− ωλ
ω
1
g = diag 1,
,
. If α1 = 0 and char F = 2, then
2(λ+2)
−1
−2
ω
−1
2ω
ωλ
1 0
0 1
g
2
U8 (λ, 0, 1) = U5 (1, 1, λ + µ + µ ), where g = diag 1,
,
. If
µ 1
1 µ
ISOMORPHISM CLASSES OF FOUR DIMENSIONAL NILP. ASSOC. ALGEBRAS
R
d(R) r(R)
U3 (0, 1, 0)
3
3
U3 (0, 1, 1)
3
3
U3 (1, λ, 0)
3
3
U4 (ρ, β1 , β2 )
3
3
U5 (1, 1, ǫ)
3
3
U3 (1, 0, 0)
3
2
k(R)
3
2
2
2
2
2
17
Conditions
λ ∈ F \ {0, 1}
ρ ∈ F , (β1 , β2 ) ∈ Aρ (F)
char F = 2, ǫ ∈ B(F)
Table 3. Representatives for conjugacy classes of nonabelian subgroups having d(Z(R)) = 2 and r(Z(R)) = 1.
α1 = 1, then U8 (λ, 1, 1)g = U4 ρ, 0, −(λ + 1)ω −1 , where
0
1
0
−ω −1
g = diag 1,
,
.
ω −1 1
−ω −2 −λω −1
Hence, we may now suppose α2 6= 1. If ∆ 6= 0, we write ∆ = ρω 2 for some ω ∈ F∗
and a unique ρ ∈ F . Then U8 (λ, α1 , α2 )g = U4 (ρ, α2 , −(α1 α2 λ−α1 λ+α2 +1)ω −1 )
where
1
ω
λ(1 − α2 )2
0
g = diag 1,
,
.
λ(1 − α2 ) 0
λ(1 − α2 ) ωλ(1 − α2 )
2
2 −1) −1
2
If ∆ = 0, i.e. α1 = λ(α
λ(α2 −1) , we need to distinguish two cases. If λ(α2 −1) +α2 6=
0, then U8 (λ, α1 , α2 )g = U3 (0, 1, 1), where
1
0
λ(1 − α2 ) − 1
0
λ(1 − α2 )
0
−λ
0
,
g = diag
2
1,
0
λ(α2 − 1)
0
λα2
2
0
λ(1 − α2 )
0
λ (α2 − 1)
and if λ =
−α2
(α2 −1)2 ,
then U8 (λ, α1 , α2 )g = U3 (1, α2 , 0), where
1
0
0
0
2
αα−1
0
1
0
2
g = diag
2 .
1, 0
−α2 0 1 − α2
α2
0
α2
0
α2 −1
Corollary 5.11. Let R ∈ ∆4 (F) be a nonabelian subgroup such that d(Z(R)) = 2
and r(Z(R)) = 1. Suppose that R is not conjugate to any RA , A ∈ Mat3 (F). Then
R is conjugate to exactly one of the subgroups listed in Table 3.
Proof. In view of Lemmas 5.3, 5.4, 5.7, 5.8, 5.9 and 5.10, R is conjugate to one of the
subgroups listed in Table 3. Since d(RA ) ≤ 2, it follows from Lemma 5.3 that these
subgroups are not conjugate to RA for any A ∈ Mat3 (F). Direct computations
show that no pair of subgroups U3 (0, 1, 1), U3 (1, λ, 0), U4 (ρ, β1 , β2 ) and U5 (1, 1, ǫ)
is conjugate in AGL4 (F).
We can now give the classification of the nonabelian subgroups in ∆4 (F).
18
MARCO ANTONIO PELLEGRINI
Theorem 5.12. Let F be a field. The distinct conjugacy classes of nonabelian
regular subgroups in ∆4 (F) can be represented by the subgroups described in the
first column of Table 2.
Proof. Let R ∈ ∆4 (F) be a nonabelian regular subgroup. By the considerations given in Section 3 we have the following possibilities: (d(Z(R)), r(Z(R))) =
(3, 2), (2, 2) or (2, 1). If (d(Z(R)), r(Z(R))) = (2, 2) then R is conjugate either to
U2 (0, 1, 0, 1, 1) or to U2 (0, 1, 0, 0, 1), by Lemma 5.1. If (d(Z(R)), r(Z(R))) = (2, 1),
then R is conjugate to either one of the subgroups of Table 3 or to a subgroup RA ,
by Corollary 5.11. If (d(Z(R)), r(Z(R))) = (3, 2), then R is conjugate to a subgroup
RA . The statement of the theorem is proved considering a set ΠA (F) of representatives for projective congruent classes of asymmetric matrices in Mat3 (F).
We would like to give explicit sets Aρ , at least when F is algebraically closed, the
real field R or a finite field. Observe that, for every field F of characteristic 6= 2,
(−1, 0) ∈ Aρ for all ρ ∈ F , by Remark 5.5(a).
Fixing ρ ∈ F , β1 ∈ F \ {1} and β2 ∈ F such that (β1 , β2 ) 6= (−1, 0), define the
following polynomials in F[t]:
(5.6)
f1 (t)
g1 (t)
= β1 t2 + β2 t − ρ,
= t2 + β2 t − ρβ1 ,
f2 (t) =
g2 (t) =
β2 t2 − 2ρ(β1 + 1)t − ρβ2 ,
t2 + ρ.
Lemma 5.13. Suppose that β1 6= 1 and (β1 , β2 ) 6= (−1, 0). If either β1 = 0 or
f1 (t) is reducible, then U4 (ρ, β1 , β2 ) is conjugate to U4 (ρ, 0, λ) for some λ ∈ F.
Proof. Clearly, the statement holds if β1 = 0. So, assume β1 6= 0. Since f1 (t) is
reducible, we can take a ∈ F such
that f1 (a) = 0 and from (5.1) we obtain that
U4 (ρ, β1 , β2 ) is conjugate to U4 ρ, 0, fg21 (a)
(a) provided that g1 (a)g2 (a) 6= 0. Suppose
that f1 (a) = g1 (a)g2 (a) = 0. Observe that f1 (a) = g1 (a) = 0 implies g2 (a) = 0,
since f1 (t) = g1 (t) + (β1 − 1)g2 (t). So, we may suppose that any root of f1 (t) is
also a root of g2 (t). Clearly, this implies that ρ = −ω 2 for some ω ∈ F. Take
a = ai = (−1)i ω, i = 0, 1. If f1 (ai ) = 0 for some i, then β2 = (−1)i+1 ω(β1 + 1).
In this case, f1 (t) = (t − ai )(β1 t − (−1)i ω) implies that β1 (−1)j ω − (−1)i ω = 0 for
some j = 0, 1, whence β1 = −1 and (i, j) ∈ {(0, 1), (1, 0)}. It follows that β2 = 0,
contradicting our assumption (β1 , β2 ) 6= (−1, 0).
In the following, given a subset S of F∗ , we denote by −S and S −1 the sets
{−s : s ∈ S} and {s−1 : s ∈ S}, respectively.
Corollary 5.14. Let F be a field with no quadratic extensions.
(a) If char F 6= 2, then A1 = {(−1, 0), (0, 0)} ∪ {(0, λ) : λ ∈ N(F)}, where N(F)
is a fixed subset of F∗ such that F∗ = N(F) ⊔ −N(F).
(b) If char F = 2, then A1 = {(0, λ) : λ ∈ F}.
Proof. It follows from Lemma 5.13 and Remark 5.5.
Proposition 5.15. Let F = F be an algebraically closed field. The distinct conjugacy classes of nonabelian regular subgroups in ∆4 (F) can be represented by
U2 (0, 1, 0, 1, 1),
U2 (0, 1, 0, 0, 1),
U3 (1, λ, 0) (λ ∈ F \ {1}),
U3 (0, 1, 0),
U3 (0, 1, 1),
RD (D ∈ ΠA (F))
ISOMORPHISM CLASSES OF FOUR DIMENSIONAL NILP. ASSOC. ALGEBRAS
and
U4 (1, −1, 0), U4 (1, 0, 0), U4 (1, 0, β) (β ∈ N(F))
U4 (1, 0, β) (β ∈ F), U5 (1, 1, 0)
19
if char F 6= 2,
if char F = 2,
where, if char F 6= 2,
ΠA (F) =
{E2,3 , E1,2 + E2,3 , E1,3 + E2,2 + E3,1 + E3,2 , E3,2 + E3,3 − E2,3 ,
E1,1 + E2,3 , E1,1 + E3,2 + E3,3 − E2,3 , E1,1 + E2,3 − E3,2 ,
E2,3 − E3,2 , E2,3 + µE3,2 , E1,1 + E2,3 + µE3,2 : µ ∈ M(F)},
and M(F) and N(F) are subsets of F∗ such that F \ {0, ±1} = M(F) ⊔ M(F)−1 and
F∗ = N(F) ⊔ −N(F); if char F = 2,
ΠA (F)
= {E2,3 , E1,1 + E2,3 , E1,2 + E2,3 , E1,3 + E2,2 + E3,1 + E3,2 ,
E2,3 + µE3,2 : µ ∈ M(F)}
and M(F) is a subset of F∗ such that F \ {0, 1} = M(F) ⊔ M(F)−1 .
Proof. By Theorem 5.12 and Corollary 5.14 we are left to determine the set ΠA (F).
Applying [3, Theorem 2.1], the elements of ΠA (F) are the asymmetric matrices
obtained as diagonal sum of the following matrices:
0 1 0
0 0 1
0 1
0 1
0 −1
0 , 1 ,
,
,
, 0 0 1 , 0 1 0 ,
0 0
µ 0
1 1
0 0 0
1 1 0
where µ ∈ M(F). If char F 6= 2 we also allow µ = −1; if char F = 2 we must omit
the matrices
0 1
diag 1,
.
µ 0
Lemma 5.16. Let ρ = 1 and char F 6= 2. Suppose that there exists an element
ι ∈ F such that ι2 = −1. Then
A1 = {(−1, 0), (0, 0)} ∪ {(λ + 1, 2ι) : λ ∈ F∗ }.
Proof. We show that every subgroup U4 (1, β1 , β2 ), β1 6= 1, is conjugate to a unique
subgroup U4 (1, β3 , β4 ) with (β3 , β4 ) ∈ A1 as in the statement. This clearly holds
if β4 = ±2ι, so suppose β4 6= ±2ι. Keeping the notation of (5.6), let h1 (t) =
2
f2 (t) − 2ιg1 (t) ∈ F[t]. Then, a = 2ββ12 +ιβ
−2ι is a root of h1 (t). It follows that R is
conjugate to U4 (1, λ + 1, 2ι) for some λ ∈ F∗ , provided that g1 (a)g2 (a) 6= 0. Notice
that g1 (a)g2 (a) = 0 if and only if (β1 + ιβ2 + 1)(β22 + 4β1 ) = 0.
Suppose that β22 + 4β1 = 0. If β2 6= 0, take b = 2β2−1 . In this case, f1 (b) =
(β 2 +4)2
β 2 +4
f2 (b) = 0, g1 (b) = 24β 2
6= 0 and g2 (b) = 2β 2 6= 0. It follows that R is
2
2
conjugate to U4 (1, 0, 0) (this clearly holds even if β2 = 0). Now, suppose that
2 +2
β1 + ιβ2 + 1 = 0 and let h2 (t) = f2 (t) + 2ιg1 (t). Take c = − 3ιβ
β2 +2ι . Then h2 (c) = 0,
2
2 (β2 −2ι) )
2 (β2 −2ι)
g1 (c) = −2ιβ
6= 0 and g2 (c) = −8β
6= 0. It follows that R is
(β2 +2ι)2
(β2 +2ι)2
conjugate to U4 (1, −3, −2ι) and then to U4 (1, −3, 2ι), by Remark 5.5.
Finally, it is an easy computation to verify that U4 (1, λ1 + 1, 2ι) is conjugate to
U4 (1, λ2 + 1, 2ι) if and only if λ1 = λ2 .
Clearly the previous lemma gives an alternative set A1 for algebraically closed
field of characteristic 6= 2. In a very similar way we can prove the following.
20
MARCO ANTONIO PELLEGRINI
Lemma 5.17. Let ρ = −1, char F 6= 2 and suppose that −1 6∈ (F)2 . Then
A−1 = {(−1, 0), (0, 0)} ∪ {(λ + 1, 2) : λ ∈ F∗ }.
Lemma 5.18. Suppose that char F 6= 2, β1 6= 0, 1, β2 6= 0 and (β1 , β2 ) 6= (−1, 0). If
f1 (t) is irreducible and f2 (t) is reducible, then U4 (ρ, β1 , β2 ) is conjugate to U4 (ρ, λ, 0)
for some λ ∈ F \ {0, 1}.
Proof. Since f2 (t) is reducible, we can take a ∈ F such that f2 (a)
= 0 and from
f1 (a)
(5.1) we obtain that U4 (ρ, β1 , β2 ) is conjugate to U4 ρ, g1 (a) , 0 provided that
g1 (a)g2 (a) 6= 0. So, first, suppose that f2 (a) = g2 (a) = 0 for some a ∈ F. As
in Lemma 5.13, this implies that ρ = −ω 2 for some ω ∈ F∗ and a = ai = (−1)i ω
for some i = 0, 1. In particular, we get β2 = (−1)i+1 ω(β1 + 1) (whence β1 6= −1).
In this case, we obtain that f1 (ai ) = 0, an absurd, since f1 (t) is irreducible. Next,
suppose that any root of f2 (t) is a root of g1 (t) = 0 (but not a root of g2 (t)).
Let r be the resultant between f2 (t) and g1 (t) with respect to t. From r = 0
we get (β22 + 4ρβ1 )(ρ(β1 + 1)2 + β22 ) = 0. Since f1 (t) is irreducible, this implies
ρ(β1 + 1)2 + β22 = 0 and so ρ = −ω 2 for some ω ∈ F∗ and β2 = ±ω(β1 + 1). As
before, this condition implies that f1 (t) is reducible, a contradiction.
Corollary 5.19. Let F = R. Then
A1 = {(0, λ) : λ ∈ [0, +∞)} ∪ {(λ, 0) : λ ∈ [−1, 0)}.
Proof. The result follows from Lemmas 5.13 and 5.18 and Remark 5.5, observing
that f2 (t) ∈ R[t] is reducible provided that β2 6= 0.
From Theorem 5.12 and [3, Theorem 2.1] we obtain the following classification.
Proposition 5.20. The distinct conjugacy classes of nonabelian regular subgroups
in ∆4 (R) can be represented by:
U2 (0, 1, 0, 1, 1),
U2 (0, 1, 0, 0, 1),
U3 (0, 1, 1),
U3 (0, 1, 0),
RD (D ∈ ΠA (R)),
∗
U3 (1, λ + 1, 0) (λ ∈ R ), U4 (1, 0, λ) (λ ∈ [0, +∞)), U4 (1, λ, 0) (λ ∈ [−1, 0)),
U4 (−1, 0, 0), U4 (−1, −1, 0), U4 (−1, λ + 1, 2) (λ ∈ R∗ ),
where
ΠA (R)
= {E1,2 + E2,3 , E1,3 + E2,2 + E3,1 + E3,2 , E1,2 ± E3,3 , E1,2 − E2,1 ,
E1,2 − E2,1 ± E3,3 , E1,2 + µE2,1 , E1,2 − E2,1 − E2,2 ± E3,3 , E1,2 ,
E1,2 + µE2,1 ± E3,3 , νE1,1 − E1,2 + E2,1 + νE2,2 , E1,2 − E2,1 − E2,2 ,
νE1,1 − E1,2 + E2,1 + νE2,2 ± E3,3 : µ ∈ (0, 1), ν ∈ R∗ }.
Lemma 5.21. Suppose that char F 6= 2, β1 6= 0, 1 and (β1 , β2 ) 6= (−1, 0). Let
F = {1, ξ} and let ϑ ∈ F such that ϑ2 + 1 = σ 2 ξ for some σ ∈ F∗ . If f1 (t)
and f2 (t) are both irreducible, then U4 (1, β1 , β2 ) and U4 (ξ, β1 , β2 ) are conjugate,
respective, to U4 (1, λ1 , ϑ(λ1 + 1)) and U4 (ξ, −1, λ2 ) for some λ1 ∈ F \ {±1} and
λ2 ∈ F∗ .
Proof. Since f1 (t) and f2 (t) are both irreducible polynomials, then β22 +4ρβ1 = ω 2 ξ
and (β1 + 1)2 ρ2 + ρβ22 = ν 2 ξ for some ω, ν ∈ F∗ . First, consider ρ = 1. The
statement is obvious if β2 = ϑ(β1 + 1). So, assume β2 6= ϑ(β1 + 1) and let h(t) =
f2 (t)−ϑ(f1 (t)+g1 (t)). Since the discriminant of h(t) is 4(ϑ2 +1)ν 2 ξ = (2σνξ)2 , there
exists a ∈ F such that h(a) = 0. Now, denote by ri the resultant between h(t) and
ISOMORPHISM CLASSES OF FOUR DIMENSIONAL NILP. ASSOC. ALGEBRAS
21
gi (t) with respect to t: we obtain r1 = ν 2 ξ((β1 −1)2 ϑ2 −ω 2 ξ) and r2 = 4(ϑ2 +1)ν 2 ξ.
From r1 = 0 we get ξ = ((β1 − 1)ϑω −1 )2 , an absurd; from r2 = 0 we obtain
ϑ2 + 1 = 0, which is in contradiction with the initial
assumption on ϑ. Hence,
(a) ϑ(f1 (a)+g1 (a))
.
,
g1 (a)g2 (a) 6= 0 and so U4 (1, β1 , β2 ) is conjugate to U4 1, gf11 (a)
g1 (a)
Next, consider ρ = ξ. The statement is obvious if β1 = −1. So, assume β1 6= −1
and let h(t) = f1 (t) + g1 (t). Since the discriminant of h(t) is 4ν 2 , there exists a ∈ F
such that h(a) = 0. Clearly, if h(a) = g1 (a) = 0, then f1 (a) = 0, contradicting the
assumption that f1 (t) is irreducible. So, suppose that h(a) = g2 (a) = 0 and denote
2
by r the resultant between h(t) and g2 (t) with respect to t. Since
ξ 6= 0, we
r = 4ν
f2 (a)
obtain that g2 (a) 6= 0 and so U4 (ξ, β1 , β2 ) is conjugate to U4 ξ, −1, g1 (a) .
Lemma 5.22. Let F = Fq be a finite field and keep the previous notation.
(a) Suppose q even. Then we can take F = {1} and
−1
: a ∈ F \ {0, 1} ,
A1 = {(0, λ) : λ ∈ F} ∪ 1 + κ−1
a , aκa
where, for every a ∈ F\{0, 1}, κa ∈ F∗ is a fixed element such that κ2a +κa 6∈
a2 B(F). In particular, |A1 | = 2q − 2.
(b) Suppose q ≡ 1 (mod 4). Then we can take F = {1, ξ} and
A1
=
{(−1, 0), (0, 0)} ∪ {(λ + 1, 2ι) : λ ∈ F∗q },
Aξ
=
{(−1, 0), (0, 0)} ∪ {(0, λ) : λ ∈ N(Fq )} ∪ {(λ, 0) : λ ∈ P(Fq )} ∪
∪{(−1, λ) : λ ∈ Q(Fq )},
where ι2 = −1 and
F∗q \ {±1} ∪ {ω 2 ξ : ω ∈ Fq } = P(Fq ) ⊔ −P(Fq )
and
F∗q \ a + a−1 ξ : a ∈ F∗q = Q(Fq ) ⊔ −Q(Fq ).
In particular, |A1 | + |Aξ | = 2q + 1.
(c) Suppose q ≡ 3 (mod 4). Then we can take F = {1, −1} and
A1
=
{(−1, 0), (0, 0)} ∪ {(0, λ) : λ ∈ N(Fq )} ∪ {(λ, 0) : λ ∈ P(Fq )} ∪
A−1
=
∪{(λ, ϑ(λ + 1)) : λ ∈ S(Fq )},
{(−1, 0), (0, 0)} ∪ {(λ + 1, 2) : λ ∈ F∗q },
where ϑ is such that ϑ2 + σ 2 + 1 = 0 for some σ ∈ F∗ and
F∗q \ {±1} ∪ {ω 2 : ω ∈ F∗q } = P(Fq ) ⊔ −P(Fq )
and
F∗q
1 − aϑ
∗
\ {±1} ∪
: a ∈ Fq \ {−ϑ}
= S(Fq ) ⊔ S(Fq )−1 .
a(a + ϑ)
In particular, |A1 | + |A−1 | = 2q + 1.
Proof. Let f1 (t), f2 (t), g1 (t), g2 (t) as in (5.6). Suppose q even. If either β1 6= 0 or
f1 (t) is reducible, then U4 (1, β1 , β2 ) is conjugate to U4 (1, 0, λ) for a unique λ ∈ F, by
Lemma 5.13 and Remark 5.5. Now, suppose β1 6= 0, 1 and f1 (t) irreducible. Then
β2 6= 0, β1 β2−2 6∈ B(F) and g1 (t) is also irreducible. We obtain that U4 (1, β1 , β2 )
is conjugate to U4 (1, β3 , β4 ) (β3 6= 0, 1, β4 6= 0, β3 β4−2 6∈ B(F)) if and only if
−1
β2 (β3 +1) = β4 (β1 +1). So, U4 (1, β1 , β2 ) is conjugate to U1 (1, 1+κ−1
a , aκa ), where
−1
∗
2
a = β2 (β1 + 1) 6= 0, 1 and κa ∈ F is chosen in such a way that κa + κa 6∈ a2 B(F).
22
MARCO ANTONIO PELLEGRINI
Suppose q ≡ 1 (mod 4). The set A1 is given in Lemma 5.16. So, consider the
case ρ = ξ with (β1 , β2 ) 6= (−1, 0). If either β1 = 0 or f1 (t) is reducible, then
R = U4 (ξ, β1 , β2 ) is conjugate to U4 (ξ, 0, λ) for a unique λ ∈ N(Fq ), by Lemma 5.13
and Remark 5.5. So, assume β1 6= 0 and f1 (t) irreducible. If f2 (t) is reducible, then
R is conjugate to U4 (ξ, λ, 0) for some λ 6= 0, 1, by Lemma 5.18. Recall that, by
Remark 5.5, U4 (ξ, λ1 , 0) is conjugate to U4 (ξ, λ2 , 0) if and only if either λ2 = λ1 or
2
λ2 = λ−1
1 ; U4 (ξ, λ1 , 0) is conjugate to U4 (ξ, 0, λ2 ) if and only if λ1 = ξω and λ2 =
2
−1
2ξω(ξω − 1) for some ω ∈ F; U4 (ξ, −1, λ1 ) is conjugate to U4 (ξ, −1, λ2 ) if and
only if λ2 = ±λ1 . Finally, if β1 6= 0, 1 and both f1 (t) and f2 (t) are reducible, then
R is conjugate to U4 (ξ, −1, λ) for some λ ∈ F, by Lemma 5.21. Now, U4 (ξ, −1, λ1 )
is not conjugate to U4 (ξ, λ2 , 0) if λ1 6= 0. Furthermore, U4 (ξ, −1, λ1 ) is conjugate
2
to U4 (ξ, 0, λ2 ) if and only if λ1 = a + ξa−1 and λ2 = ±a2a−ξ for some a ∈ F∗ .
Suppose q ≡ 3 (mod 4). The set A−1 is given in Lemma 5.17. So, consider the
case ρ = 1 with (β1 , β2 ) 6= (−1, 0). Proceeding in a way similar to what we did
for the case q ≡ 1 (mod 4), R = U4 (1, β1 , β2 ) is conjugate to a subgroup having
one of the following shape U4 (1, 0, λ) for a unique λ ∈ N(Fq ), U4 (1, λ, 0) for some
λ 6= 0, 1, or U4 (1, λ, ϑ(λ + 1)) for some λ 6= 1. Now, U4 (1, λ1 , 0) is conjugate to
U4 (1, λ2 , 0) if and only if either λ2 = λ1 or λ2 = λ−1
1 ; U4 (1, λ1 , 0) is conjugate to
U4 (1, 0, λ2 ) if and only if λ1 = ω 2 and λ2 = 2ω(ω 2 − 1)−1 for some ω ∈ F \ {±1};
U4 (1, λ1 , ϑ(λ1 + 1)) is conjugate to U4 (1, λ2 , ϑ(λ2 + 1)) if and only if λ2 = λ±1
1 .
Furthermore, U4 (1, λ1 , ϑ(λ1 + 1)) is conjugate to U4 (1, 0, λ2 ) if and only if λ1 =
±1
2
1−aϑ
−1
; U4 (1, λ, ϑ(λ1 + 1)) is not conjugate
, λ2 = ∓ aa2ϑ−2a−ϑ
a(a+ϑ)
+2aϑ−1 , a 6= 0, −ϑ, ϑ
to U4 (1, λ2 , 0).
Proposition 5.23. Let F = Fq be a finite field. If q is odd, there are exactly 5q + 9
distinct conjugacy classes of nonabelian regular subgroups in ∆4 (Fq ), that can be
represented by:
U2 (0, 1, 0, 1, 1), U2 (0, 1, 0, 0, 1), U3 (0, 1, 1), U3 (0, 1, 0), U3 (1, λ, 0) (1 6= λ ∈ Fq ),
RD (D ∈ ΠA (Fq )), U4 (1, β1 , β2 ) ((β1 , β2 ) ∈ A1 ), U4 (ξ, β1 , β2 ) ((β1 , β2 ) ∈ Aξ ),
where
ΠA (Fq )
= {E2,3 − E3,2 , E1,1 + E2,3 − E3,2 , E2,2 + E2,3 + λE3,3 ,
E1,1 + E2,2 + E2,3 + λE3,3 , ξE1,1 + E2,2 + 2E2,3 + E3,3 ,
E1,3 + E2,3 + E3,2 , E1,1 + E1,3 + E2,3 + E3,2 : λ ∈ Fq }
and x2 − ξ is a fixed irreducible polynomial of Fq [x] (take ξ = −1 if q ≡ 3 (mod 4)).
If q is even, there are exactly 5q + 6 distinct conjugacy classes of nonabelian
regular subgroups in ∆4 (Fq ), that can be represented by:
U2 (0, 1, 0, 1, 1), U2 (0, 1, 0, 0, 1), U3 (0, 1, 1), U3 (0, 1, 0), U3 (1, λ, 0) (1 6= λ ∈ Fq ),
U5 (1, 1, 0), U5 (1, 1, ξ), RD (D ∈ ΠA (Fq )), U4 (1, β1 , β2 ) ((β1 , β2 ) ∈ A1 ),
where
ΠA (Fq ) = {E2,2 + E2,3 + λE3,3 , E1,1 + E2,2 + E2,3 + λE3,3 ,
E1,3 + E2,3 + E3,2 , E1,1 + E1,3 + E2,3 + E3,2 ,
ξE1,1 + E1,3 + E2,3 + E3,2 + E3,3 : λ ∈ Fq },
2
and x + x + ξ is a fixed irreducible polynomial of Fq [x].
ISOMORPHISM CLASSES OF FOUR DIMENSIONAL NILP. ASSOC. ALGEBRAS
23
Proof. The set ΠS (Fq ) have been determined in [9, Theorem 4], so the result follows
from Theorem 5.12 and Lemma 5.22.
We conclude recalling that Theorem 1.1 follows immediately from Theorems 4.2
and 5.12, applying Proposition 2.4.
References
[1] A. Caranti, F. Dalla Volta and M. Sala, Abelian regular subgroups of the affine group and
radical rings, Publ. Math. Debrecen 69 (2006) 297–308.
[2] W. De Graaf, Classification of nilpotent associative algebras of small dimension,
arXiv:1009.5339, 27 Sep 2010.
[3] R.A. Horn and V.V. Sergeichuk, Canonical matrices of bilinear and sesquilinear forms, Linear
Algebra Appl. 428 (2008), 193–223.
[4] M.A. Pellegrini, Regular subgroups, nilpotent algebras and projectively congruent matrices,
to appear in Int. J. Group Theory.
[5] M.A. Pellegrini and M.C. Tamburini Bellani, More on regular subgroups of the affine group,
Linear Algebra Appl. 505 (2016), 126–151.
[6] M.A. Pellegrini and M.C. Tamburini Bellani, Regular subgroups of the affine group with no
translations, to appear in J. Algebra, http://dx.doi.org/10.1016/j.jalgebra.2017.01.045.
[7] B. Poonen, Isomorphism types of commutative algebras of finite rank over an algebraically
closed field, in: Computational arithmetic geometry, 111–120, Contemp. Math., 463, Amer.
Math. Soc., Providence, RI, 2008.
[8] M.C. Tamburini Bellani, Some remarks on regular subgroups of the affine group, Int. J.
Group Theory 1 (2012) 17–23.
[9] G.D. Williams, Projective congruence in M3 (Fq ), Results Math. 41 (2002), 396–402.
Dipartimento di Matematica e Fisica, Università Cattolica del Sacro Cuore, Via
Musei 41, I-25121 Brescia, Italy
E-mail address: [email protected]
| 4math.GR
|
E cient textualrepresentation of structure
Brenton Chapin
arXiv:1706.00862v1 [cs.PL] 2 Jun 2017
ABSTRACT
This paper attempts a more formal approach to the legibility of text
based programming languages, presenting, with proof, minimum
possible ways of representing structure in text interleaved with
information. This presumes that a minimalist approach is best for
purposes of human readability, data storage and transmission, and
machine evaluation.
Several proposals are given for improving the expression of interleaved hierarchical structure. For instance, a single colon can
replace a pair of brackets, and bracket types do not need to be repeated in both opening and closing symbols or words. Historic and
customary uses of punctuation symbols guided the chosen form
and nature of the improvements.
KEYWORDS
programming language design, structured programming, human
readability, syntax, notation, history, data compression, minification
ACM Reference format:
Brenton Chapin. 2016. Efficient textual representation of structure. In
Proceedings of ACM Conference, Washington, DC, USA, July 2017 (Conference’17), 11 pages.
DOI: 10.1145/nnnnnnn.nnnnnnn
1
INTRODUCTION
Information is almost always more useful when organized, and
structure is key to that. Therefore efficient and clear representation of structure is of paramount importance. Structured programming languages are only one use of structure to organize one kind
of information, source code, but they make such unprecedentedly
elaborate use of structure that they have exposed deficiencies in
our methods of expression.
The languages for programming and math developed in an evolutionary manner, with much borrowing from earlier work, and
the reasons for various decisions became buried in custom and
history. Studies on choices of characters and legibility have been
patchy, with some questions overlooked because they were thought
relatively unimportant. Pioneers of programming languages hurriedly made expedient adaptations of existing notations for similar
problems. Most heavily borrowed was mathematical notation.
With many important questions needing settling with the creation of the first programming languages, issues of symbology were
mostly passed over as unimportant and arbitrary. As Bob Bemer
Permission to make digital or hard copies of all or part of this work for personal or
classroom use is granted without fee provided that copies are not made or distributed
for profit or commercial advantage and that copies bear this notice and the full citation on the first page. Copyrights for components of this work owned by others than
ACM must be honored. Abstracting with credit is permitted. To copy otherwise, or republish, to post on servers or to redistribute to lists, requires prior specific permission
and/or a fee. Request permissions from [email protected].
Conference’17, Washington, DC, USA
© 2016 ACM. 978-x-xxxx-xxxx-x/YY/MM. . . $15.00
DOI: 10.1145/nnnnnnn.nnnnnnn
noted, “much documentation is lost, and it was characteristic of
the times that nobody seemed to think character sets a very important feature of computers [12].” There are many studies on the
readability and other properties of various fonts and color combinations, but when discussed in relation to source code, the term
“readability” refers more to comprehensibility [21].
Punctuation is the class of symbol most closely associated with
showy, interleaved structure. Positioning is the other major method
used to indicate structure, and among the other intended purposes,
“control characters” attempted to provide ways to position text.
Currently, Python is the most popular programming language that
relies on text positioning rather than punctuation to indicate structure. Visual programming goes further yet, replacing textual indicators of structure and flow with graphical ones.
The programming language wars are still hot today, with new
languages emerging and gaining followers. One cause of the passionate debates is the tendency of language designers to resort to
an evangelical approach to justify their choices of design elements
for which they have little compelling technical reason. Sometimes
the designers make overlarge and unsubstantiated claims [19]. For
many programming languages, one of the defining features is the
choice and usage of symbols. These choices are not modifiable by
the programmers, so that if such changes are desired, a whole new
programming language may need to be created, another factor in
the very proliferation of programming languages that the ALGOL
designers were hoping to avoid.
The ideas in this paper aim at the foundation, the symbolic representation of the structure. Structure is chosen as the crucial concept that must be addressed to improve representation. Minimalism is the chosen guide.
Too much minimalism is certainly possible, for instance by expecting people to work with information that has been minimized
by data compression techniques which transform the data into a
very compact but human unreadable form. The minification techniques of removing unnecessary whitespace and shortening variable names is another example. The target of these minimization
efforts is the representation of the structure of the source code, not
the source code itself. Further, this is not about rewriting and rearranging to place information in more efficient structures, this is
about making the representation more efficient regardless of the
structure chosen. Punctuation has always been minimal, using
smaller and less obtrusive symbols than those used to represent
the letters of a language, and the syntax of structural elements follows that pattern.
Some more items to note are splits between textual representations used for source code, versus those used for markup, as in
HTML, and data organization, as in XML and YAML. Within programming languages, there is the dot (or arrow) notation of Object
Oriented Programming and the completely different notations for
Structured Programming, such as the curly braces. Yet those splits
Conference’17, July 2017, Washington, DC, USA
seem artificial, as hierarchical structure is used in all. Many programming languages are needlessly poor at expressing data. Several of the improvements in C++11 and C++14 touch on this issue, allowing more flexible constructions of constants and initializations of arrays and objects. One intent of JSON is to bridge this
divide.
Also, these are all interleaved formats, meaning the symbols
that denote the structure are interleaved with the symbols of the
data. Goals in data storage are minimal size, and fast access. An
obvious and common method to achieve both is to exclude all complex structure from the data, using an external representation. The
disadvantage is that they require some connection, often expressed
in fixed sizes and padding, which can end up using more space
than an interleaved method. Over the years, the attention paid to
brevity has varied with the cost and availability of storage.
By 1963, the American Standard Code for Information Interchange (ASCII) was set to a fixed size of 7 bits, though at least
two variable codes, Morse code (1844) and Huffman coding (1952)
existed at the time.
One of the goals of XML was human readability. Minimalism
was thought orthogonal or possibly even antithetical to the goal
of human readability, and the resulting language ironically suffers
from excessive verbosity that obscures essentials, rendering the result less human readable. XML and COBOL show that a negative
attitude towards minimalism (”10. Terseness in XML markup is
of minimal importance. [14]”), that regarding minimalism as unrelated or even an impediment to comprehension, is not correct.
Minimalism is also central to Information Theory, in which it was
demonstrated that the crude redundancy of repeating information
over and over, is very wasteful and poor at preserving the fidelity
of data against errors in the transmission. If repetition is a poor
means of ensuring the fidelity of data, perhaps it is also a poor
means of representing structure in ways easy for humans to read.
Another demonstration of the limited usefulness of repetition is
the FAT file system, which despite allocating room for a copy of
the directories and file names, is actually one of the most fragile
and easily corrupted file systems currently in use.
Of particular note is the C programming language. So many
programming languages adopted the C syntax that they have been
tagged with a moniker of their own, the “curly-brace” languages.
Perhaps one of the reasons curly brace syntax eclipsed Pascal and
ALGOL is the use of a single character each, rather than the words
BEGIN and END, to delimit blocks. The designers of C did not restrict themselves to curly braces only, they also used square brackets, parentheses, and even angle brackets, for array indexing, lists
of function parameters, and macros respectively. Why that choice
of symbol assignment? Why not use parentheses for all, and rely
on context or some other means to distinguish between a parameter list and a block of code? If there is any doubt that it is possible
to use only parentheses, the LISP programming language is proof.
Or, why not copy FORTRAN in the use of parentheses for array
indices? One kind of answer is that in C these different kinds of
brackets serve as sigils, to distinguish between identifiers for functions, arrays, and variables. But that only begs the question of why
have sigils? And it still does not answer why any particular symbol
was chosen for a particular use.
Brenton Chapin
2 HISTORY
For answers, one must dig into the history of computation and
mathematics. In the case of C, the chain of preceding languages is
roughly B, BCPL (Basic CPL), CPL (Combined Programming Language), and finally ALGOL (Algorithmic Language). The paper on
ALGOL 58 [6] says of the choice to use square brackets to delimit
array indices, only that “subscripted variables” (the term used in
ALGOL for what today we call an array variable, or simply an array), “designate quantities which are components of multidimensional arrays” and that “The complete list of subscripts is enclosed
in the subscript brackets [].” But why did they pick square brackets? FORTRAN, the oldest programming language to achieve wide
acceptance, uses parentheses, not square brackets.
For that matter, why use any bracket at all? No one says. It
seems likely that they would rather have used actual subscripted
text, just like in mathematical notation, but early computers could
not do it. Square brackets was a notational device to indicate subscription without actually presenting the text so. Apart from computer limitations, a big problem with subscripting is that the notation doesn’t nest well, at 3 or more levels becoming too small for
the human eye to read. One can surmise from the use of the term
“subscript” that this was another borrowing, from linear algebra in
which a matrix is denoted with square brackets. And indeed the
original name of ALGOL 58, is International Algebraic Language.
The only deviation in the use of square brackets for array indexes
from ALGOL to C was BCPL, which among the many simplifications of CPL it introduced, attempted to repurpose square brackets
for code blocks, using only pointer arithmetic to access array elements [8].
ASCII codified the glyphs used for nearly all programming languages. A notable exception is APL, which makes use of mathematical symbols, mainly from Set Theory and Vector calculus, that
were not put in ASCII [7]. Unlike EBCDIC, ASCII at least organized the alphabet into a contiguous block. But the exact set of
punctuation symbols is unclear, ranging from all symbols that are
not letters, numbers, and control characters, to only those used to
clarify the structure and meaning of sentences and text. There are
no formal, ordered, centuries old lists of punctuation symbols.
The ASCII ordering and choice of punctuation is derived from
the QWERTY keyboard layout, which dates to the late 19th century. The notion that QWERTY was deliberately arranged to slow
typists down is a popular but wrong myth [23]. Morse Code and
many other factors were considered, and over the years small changes
have been made to accommodate new uses. For instance, “shift-2”
is the double quote mark on many older keyboards, but today is
‘@’ on most keyboards.
We could go further back, and ask why mathematical notation
uses parentheses for functions, and square brackets for matrices.
Why is y = fx the customary, canonical expression for a function,
and why in particular the use of parentheses to bracket the independent variable x? In A History of Mathematical Notations [3],
Cajori credits Euler (1707-1783) with the first use of parentheses
to bracket the variable of a function, in a 1734 paper. That paper is
E44 [1] in the numbering scheme created to refer to Euler’s works.
However, examining E44 and several others of Euler’s papers, one
finds no such use of parentheses, and the exact phrase and formula
Efficient textual representation of structure
Cajori quoted is not present. Euler uses parentheses to group parts
of equations, but not to separate function names and variables. Euler’s notation is y = fx, and it is up to the reader to understand
that x and y are variables, and f is a function.
Note also the choice of the letter f because it is the first letter of
the word “function”, a custom followed in many places, such as the
decision in FORTRAN to use the first letter of a variable name to
indicate integer (name begins with “I” for integer, through “Z”) or
floating point (name begins with “A” through “H”). This desire to
match functionality to the first letter of an appropriate term was
taken to extremes, so that more than one early game employed
a lucky placement of keys on the QWERTY keyboard,’W’, ’E’, ’S’,
plus ’3’, to refer to west, east, south, and north respectively.
By 1837, in a major work on Number Theory which is regarded
as also an important paper on the modern definition of a function, Dirichlet (1805-1859) used parentheses around the independent variable [2]. But why did mathematicians pick those symbols,
that format? They too engaged in expedience, adopting the idea of
parentheses from still earlier scholars. Mathematical notation has
a long evolutionary history, and while fascinating, the main point
here is that many choices of symbols and syntax were made long
before any possible use in programming languages was conceived.
While 1837 is also the year that Babbage proposed the Analytical
Engine, arguably the first computer, functioning computation machinery would not be built until many years later. Therefore symbols and syntax certainly could not have been chosen based on
experiences in computer programming.
That was about as far as the early pioneers went in exploring
questions of how best to symbolize code and data. None of the
terms and areas of study, not semiotics, symbology, linguistics,
grammar, lexicology, punctuation, readability, typography, legibility, notation, expressiveness, or rubrication, quite address these
questions. Studies of notation and syntax get the closest, but even
there syntax is confined to issues of context.
Most programming languages use a hierarchical structure to organize code. Possibly the earliest and simplest formally specified
language for expressing hierarchy is Dyck Language. Object Oriented Programming and Functional Programming did not abandon
this fundamental organization, they only added to it. Declarative
programming, as represented in Prolog and SQL, at first glance
seems not to need much structure. A point of confusion is order vs
structure vs hierarchy. Declarative programming needs structure,
but not order and not necessarily hierarchy. Hierarchic structure,
of programs and data, can be more efficiently represented with several changes.
The advent of markup languages revived interest in hierarchical data storage, which was introduced in the 1960s, before the
relational database model. No longer were interleaving structural
symbols just for programs, they were harnessed to organize data.
Traditionally, data has been organized into fixed size elements so
that no symbols need be reserved for explicit denotation of structure, and, even more importantly, so that random access is quick,
taking O 1 time to retrieve any one element. This is also true of the
pre-computer era, which used tables extensively, carefully lining
up columns to aid the human eye. Where one-size-fits-all is inadequate, the expedient method used is to have a small fixed size field
Conference’17, July 2017, Washington, DC, USA
to hold a value for the size of a variable length field. Packet networking is an example of this organization of data. The roughly
analogous method in writing is the technique of employing any
of a variety of superscripted symbols such as an asterisk, *, or a
dagger, y, to indicate there is a footnote.
XML and HTML are the most well known of these markup languages, and like programming languages, their history is also evolutionary. Both trace back to Standard Generalized Markup Language (SGML) which was standardized in 1986, predating the World
Wide Web. Like so many other decisions in languages, the creators
of the Web seized upon SGML out of expediency. SGML in turn
descends from GML, an IBM effort to manage technical documentation and data, based upon ideas first articulated circa 1966 [16].
But as many have complained over the years, these markup
languages have undesirable features, and among the biggest is extreme verbosity. The rules they force upon users, to come closer to
the goal of “human readability”, often have the opposite effect. On
the scales of minimalism, XML and relatives are extremely poor because their representations are highly redundant. Not only must
brackets be balanced in “proper” HTML and XML, but the matching tags must repeat the tag name. Why did the designers do it?
Ironically, those rules have done much to add clutter and thereby
reduce the human readability that was their intended goal. YAML
(YAML Ain’t Markup Language) was motivated in part by recognition that XML is burdened with design constraints that have little
purpose in data serialization [17]. Lightweight markup languages
such as Markdown are an acknowledgment that the human readability of HTML could be better.
Most popular programming languages are poor at expressing
data. Here are some examples to illustrate this. A list of the first 10
chemical elements can be encoded in a JavaScript array like this:
c o n s t CE = [ ” ? ” , ”H” , ” He ” , ” L i ” , ” Be ” , ” B ” ,
”C ” , ” N ” , ” O ” , ” F ” , ” Ne ” ] ;
A simple trick yields a much cleaner representation:
c o n s t CE = ” ? H He L i Be B C N O F Ne ”
. split (” ”);
But this is the very sort of trick that makes programming needlessly difficult for professional programmers unfamiliar with the
arcana of a particular language.
One problem is that the default, unquoted meaning of an alphanumeric sequence is to treat it as the name of a variable. The
double quote mark changes the mode, but that mode has no support for structural elements, so only a simple string can be encoded.
The programmer is forced to change modes over and over, entering
string mode to give a short string, leaving string mode to impart a
tiny amount of structure, then entering string mode again to give
the next string. Or the programmer can use a clever trick such as
the split function, or create a function to parse a string into a
complicated object, or even employ a library such as YAML.
Another example, of a family tree, in Python:
c l a s s t n : # t n means ” t r e e node ”
def
init
( s e l f , name , c h i l d =None ) :
i f c h i l d ==None : s e l f . c = [ ]
else : self . c = child
Conference’17, July 2017, Washington, DC, USA
s e l f . n = name
familytree =
[ tn ( ” grandmother ” ,
[ tn ( ” older uncle ” ,
[ tn ( ” o l d e st 1 st cousin ” ) ,
t n ( ” 2 nd o l d e s t 1 s t c o u s i n ” ) ] ) ,
tn ( ” fat her ” ,
[ tn ( ” older s i s t e r ” ,
[ tn ( ” niece ” ) ,
t n ( ” nephew ” ) ] ) ,
t n ( ” you ” ,
[ t n ( ” son ” ,
[ tn ( ” granddaughter ” ) ] ) ,
tn ( ” daughter ” ,
[ tn ( ” grandson ” ) ] ) ] ) ,
t n ( ” younger b r o t h e r ” ) ] ) ,
...
This terrible encoding is littered with alternating brackets of 2
kinds, as well as double quote marks and commas. This shows
that Python can be even worse than LISP, for those who thought
Python’s use of indentation lead to clean code in all cases, and that
LISP had too many parentheses. To get clean looking code, the expert programmer resorts to using functions to read a simple string
(which may be a data file) into a complicated object. Employing a
data serialization library such as YAML, is a common method of
handling this issue. Should it be the preferred method? Shouldn’t
programming languages be able to do better with their native syntax? After all, native handling of regular expressions is what made
Perl popular. Improvements in the representation of structure are
applicable both to coding and to data representation.
3
ELIMINATING RUNS OF BRACKETS
The first change addresses a problem most languages have, but
which is perhaps most obvious in LISP, and for which it has been
criticized in the “backronym” of Lots of Idiotic Spurious Parentheses. Often, brackets cluster, as several structures all start or end
simultaneously. They can add to the visual clutter without adding
to the ease of comprehension.
There are many solutions to this problem, among them operator
precedence, and postfix notation, also known as Reverse Polish notation, first conceived in 1924[4]. A limitation of these Polish notations is that to make brackets unnecessary, the number of operands
must be fixed, an inflexibility that is insufficiently general for the
structures used in programming.
A popular short cut is use of context and knowledge about the
permitted or sensible content of subtrees. For instance, in HTML
the paragraph indicator, <p>, cannot be nested. This is often used
to omit the matching closing bracket, </p>, when the next structure is another paragraph, or something else that cannot be inside
a paragraph, such as a header. Such omissions are not officially
sanctioned in HTML, but are so popular that web browsers had to
support them anyway. Obvious problems with this approach are
that knowledge of every exception to the rules for indicating the
nesting may be very large, and may change.
Brenton Chapin
The approach taken in Perl 6 is to allow all kinds of shortcuts
that do not greatly complicate the parser. Compared to Perl 5,
some brackets are no longer required. In particular, the parentheses of the if and for statements are optional [18]. Effectively, this
change is a recognition that if and for are enough by themselves
to indicate structure, that they are in fact now part of the set of
symbols used to denote structure.
One could employ 2 sets of brackets, perhaps () and [], in a
scheme in which a closing bracket closes its matching opening
bracket, and all the open brackets of the other kind in between.
For example, [a [b]] becomes [a (b], [d [e [f]]] becomes
(d [e [f). This idea can work in the other direction. [[g] h]
becomes [g) h]. It even works in both directions at once, with
((j)(k)) becoming [j)(k]. However, the best this idea can do
for ((m)) is [(m].
An issue is that 2 more symbols are needed. We can employ
only one more symbol, eliminating only one of the excess opening or closing brackets, and still clean up most of clutter. Call a 3
symbol system that eliminates excess closing brackets a “closing 3”,
and a 3 symbol system that eliminates excess opening brackets an
“opening 3”. Using colon, :, for this 3rd symbol in a closing 3 system, because that approximately matches the traditional use of the
colon in written natural languages, changes (a (b)) into (a : b).
((m)) becomes (:m), (((n))) becomes (::n), and ((j)(k)) becomes ((j):k). Additionally, the brackets are still balanced, with
equal numbers of opening and closing brackets in all the systems.
For a slightly larger example, consider this Ackermann function,
from the classic textbook Structure and Interpretation of Computer
Programs, exercise 1.8 [10]:
( d e f i n e (A x y )
( cond ( ( = y 0 )
((= x 0)
((= y 1)
( e l s e (A
0)
(∗ 2 y ))
2)
(
x 1)
(A x (
y 1))))))
Employing a closing 3 system as suggested above, gives this:
( d e f i n e (A x y )
: cond ( ( = y 0 )
((= x 0)
((= y 1)
: else :A
0)
:∗ 2 y)
2)
(
x 1)
:A x :
y 1)
6 colons have replaced 6 opening brackets. The 6 matching closing brackets have been removed. Indeed, there is never a need for
multiple adjacent closing brackets, as proven next.
Theorem 3.1. Given a sequence S of arbitrary symbols over an
alphabet A in which 2 symbols, an “opening” and a “closing” symbol,
are reserved to denote hierarchy in a format that interleaves data
and structure, and S is properly balanced, the hierarchy can always
be represented in a system with 3 reserved symbols in which there are
no runs (sequences of length 2 or greater) of the closing symbol.
Proof. WLOG, let ‘(’ and ‘)’, the parentheses, represent the
opening and closing symbols in both systems, and let ‘:’, the colon,
Efficient textual representation of structure
represent the 3rd symbol in the 3 symbol system. To allow elimination of all runs of 2 or more closing symbols, assign ‘:’ the same
meaning as ‘(’, the opening of a subtree, except that the matching
closing symbol for ‘:’ is an already necessary ‘)’ that matches an
existing ‘’ which precedes the ’:’.
Then, instances of the sequence “( s1 ( s2 ))” in which s1
and s2 are arbitrary sequences which may include balanced occurrences of ‘(’ and ‘)’ and ‘:’, may be replaced with “( s1 : s2
)”.
The replacement symbols are sufficient to represent all the relationships. The symbols still indicate that s1 is the parent of s2 ,
preserve all relationships s1 and s2 have with all other sequences
before and after because none of them need change and no additional context is needed, and preserve all relationships contained
within and between s1 and s2 also because none of them change,
nor add any contextual dependencies.
This replacement can be applied repeatedly, to reduce any number of adjacent closing brackets to 1 closing bracket. Each replacement preserves the property of balance for all remaining parentheses, as exactly one pair of matched parentheses is replaced with a
single colon.
The corollary that no runs of the opening bracket are needed in
an opening 3 system, is obvious.
A pushdown automaton can easily transform from an opening
3 to a 2, or from a 2 to a closing 3, if the data is processed in reverse order. Of course, a pushdown automaton can easily reverse
a string. In practice, the C++ style of comment delimited by a 2
slashes takes more work to detect from back to front. Nor can the
start of a C style comment be simply determined working from
back to front, because “/*” can be within a comment.
A natural question is why not use a 4 symbol system, as originally outlined above with the 2 sets of bracket symbols, and eliminate all runs of opening and closing brackets? Simply put, the
additional savings is not significant, as can be seen in that it is no
help at all on the examples of ((m)) and (((n))).
As to why, it is not possible to employ any finite set of symbols to represent infinitely many numbers with just 1 symbol each,
no matter what form the representation takes. If the representation takes the form of n opening brackets followed by n closing
brackets, all of one kind of bracket can be collapsed, because n is
preserved in the other. If both are collapsed, then n must be represented some other way. That is why the idea of using 2 sets of
brackets does not work to reduce all runs of opening and of closing
brackets to 1.
Thus we see that the idea of replacing each run of closing brackets with a single closing bracket is really the removal of a redundancy, the redundancy of specifying the depth twice, first with
opening brackets, then with an equal number of closing brackets.
That redundancy is no longer available to remove once one kind
of bracket has been reduced.
The 3 symbol system need not be exclusive, can mix with 2 symbol usage as in (a (b : c)). In practice, in coding it will likely be
preferable to use the 3rd symbol only for subtrees that are known
in advance to be the terminal child. For other uses, such as minification of JavaScript or JSON, may want to use the 3rd symbol
everywhere possible.
Conference’17, July 2017, Washington, DC, USA
Removing the redundancies of the 2 symbol system can be of
some value in data compression. Since the amount of information
encoded is the same, an ideal data compression algorithm should
produce the same size compressed output whether a 2 or a 3 symbol system is used. In practice, the output sizes vary, sometimes
better for the 3 symbol system, and sometimes worse. To better test
whether the more efficient representation helps with data compression, can try a much larger example. Biologists have organized millions of species into a Tree of Life [28], using Newick format [11],
an interleaved hierarchical format. Tests upon grafted solution.tre
from version 9.1, the file with the highest percentage of interleaved
structural symbols relative to data, containing 100,734 parentheses in 721,324 characters total, show an “opening 3” system does
reduce size even after compression.
compression
none
gzip
bzip2
xz
system
original 2 symbol
721,324
250,142
218,717
211,812
opening 3
690,077
241,169
213,341
203,724
A final note about whether to prefer an opening 3 or a closing
3 system. The closing 3 is the better fit with our customs. For instance, in curly brace languages, the name of an array is given before the index of the desired element. It is arr[39] not [39]arr. It
is the same with function names and parameters– the name comes
first.
4 UNIVERSAL BRACKET
A sequence such as “[x(y]z)” in which 2 different sets of brackets are interwoven, is almost always an error, not valid in any
mainstream language. An analogous sequence in HTML could be
“<b>x<i>y</b>z</i>”, which is not valid, even though its meaning can in this case make sense. The HTML specification calls this
“misnesting”. This invalidity is used in an ad hoc fashion to reduce some of HTML’s redundancy. A common case is the closing of an outer structure that implies an inner structure must also
close, as in this example: “<tr>x<td>y</tr>”. Some omissions require knowledge that some structure is not allowed. For instance,
“<p><p></p></p>” is not valid because the ‘p’ element (p for paragraph) can’t be the direct child of another ‘p’ element. Therefore
“<p>x<p>” always implies a closing tag preceding the 2nd opening
tag: “<p>x</p><p>”. This usage is acknowledged in HTML5, but
still recommended against: “..the closing tag is considered optional.
Never rely on this. It might produce unexpected results and/or errors if you forget the end tag.” [25]
A combination opening and closing tag in one, called a “selfclosing” tag, is meant for an empty element, and has been in XML
from the start [14]. As of version 5, HTML has adopted a variation
of this idea. The XML self-closing tag requires a slash character immediately before the closing angle bracket. In HTML5, 15 element
types were singled out as making sense only as empty (void), and
HTML does not use or permit the penultimate slash character in
those tags.
Another solution to some of HTML’s verbosity is to omit the
name from the end tags, using only “</>”, which works fine since
misnesting is not allowed or often sensible anyway. SGML has
Conference’17, July 2017, Washington, DC, USA
this feature in its SHORTTAG constructs, calling it the empty end
tag. But HTML does not allow it. This idea of a universal closing
bracket or, alternatively, a universal opening bracket, can be employed in any language containing 2 or more sets of bracket symbols and in which interweaving is invalid. It eliminates misnesting,
as interweaving is no longer possible. And it reduces the alphabet
size.
If we choose the square bracket for the universal closing symbol, then a sequence such as “(x[y]z)” could become “(x[y]z]”,
and the closing parenthesis symbol would be unused, and could
be repurposed. (Note that this change does not reduce the number
of closing brackets, there are still 2 in the example. It reduces the
required size of the alphabet.)
There can still be a need for other closing characters, such as
an “unindent” invisible control character. The universal closing
bracket could still be used for that, but would want it to be invisible
in that case.
Converting back and forth between a representation that uses
closing brackets that match the opening brackets, and a representation that uses only a universal closing bracket is easily done with
a pushdown automaton. The type of the node is preserved in the
choice of opening bracket, and having the type repeated with a
matching closing bracket is merely redundant.
Having established that a universal bracket symbol is workable,
several more questions naturally arise. Does it make code easier
to understand, more human readable? Many have expressed the
sentiment that requiring a closing tag to repeat the type given in
the opening tag helps prevent human mistakes, and is therefore
good. The issue is confused by the practice of entirely omitting
tags in specific situation. With a means of representing structure
that is not so tiresomely redundant, these ugly short cuts can be
made unnecessary.
5
TYPES FOR NODES
Often the representational capability of a hierarchical structure is
enhanced by adding some means of representing different kinds of
children. An example is the “red–black tree” in which tree nodes
have been assigned an additional property, a color. This can be
and is often done independently of the structure, by means of an
additional data item. Another very popular method is sigils in the
form of different kinds of brackets. ASCII has 3 sets of symbols
meant solely for brackets: the parenthesis, square bracket, and
curly braces. One more set, the angle brackets, doubles as the mathematical symbols “greater than” and “less than”, and for that reason
was used gingerly. Further sets can be contrived, for instance ‘n’
and ‘/’, and for that matter of course any two arbitrary characters
could be chosen to serve as brackets. The obvious complaint is that
4 sets is far too few. Even if a dozen plus from Unicode are added,
it still isn’t enough.
In any case, programming language designers used all the ASCII
symbols meant for brackets early on. The curly brace languages
employ curly braces to denote blocks of code, square brackets to
denote array indices, and parentheses for parameter lists. The dual
purpose symbols for angle brackets did not go unused as brackets,
being employed in the C preprocessor, and later in the markup
languages SGML, XML, and HTML.
Brenton Chapin
These SGML markup languages expanded the number of bracket
types infinitely, by allowing multiple character brackets. Although
that solves the problems caused by finite quantities of different
bracket symbols, the means and requirements chosen add greatly
to the verbosity, a common criticism often expressed in abuse of
the rules rather than in words. It is possible that the desire for a
visual match between the opening and closing bracket led to the
SGML requirement that both the opening and closing brackets contain copies of the string employed to give them a type, despite the
obvious redundancy.
An efficient way is to designate one of the bracket sets as ”typed”,
the start of a multicharacter bracket. That allows the other brackets
to remain bare to be used same as traditionally used in programming languages, and still allows infinite bracket types. Which symbol is best employed, and where should it be positioned? Between
bracket and name, or after the name? Or, should it be a combined
symbol, a bracket that indicates a name is to follow, since there
is more than one kind of bracket available? Possibly the most efficient use of existing symbols is to keep parentheses as is, bare,
and designate the square bracket or curly brace as the indicator
for a child structure with a type, with the name to follow, similar
to HTML.
Another method is to reserve a symbol to indicate a type name
only, no structure. ‘$’ is often used similarly.
Whichever method is chosen to indicate the start of a type name,
how is the name to be ended? The name could be just like variable
names in programming languages, with only letters and numbers
(and the underscore character) allowed in the name so that any
other symbol, such as the space character, automatically ends the
name. The method of using a special symbol, as done with the
closing angle bracket of HTML, is also workable.
But the designers of HTML did not let tag names be only names.
They crammed additional structured information into “attributes”.
An example is “<ul class="vs1" id="item1"> content </ul>”.
This information could have been in the same system, for instance
something like “<ul> <attr> class = "vs1" id = "item1"
</attr> content </ul>”, or even “<ul> <attr> <nm>class </nm>
<val> vs1 </val> <nm> id </nm> <val> item1 </val> </attr>
content </ul>”. The only purposes this alternate subsystem really
serves is visual distinction and less verbosity, though it’s claimed
to maintain the distinction between data and metadata. HTML has
evolved towards lighter use of attributes, moving much formatting
information from the tags to CSS, where it is also less redundant.
6 REPRESENTING SIBLINGS AND COUSINS
The list is well known and has a long history. Each item in a list
can be considered a sibling of each other item. Traditionally, each
item is on its own line, or is separated by a comma. LISP means
“LISt Processor”, and is built around the idea of making lists the
fundamental building block with which to organize both data and
code. Comma Separated Values (CSV) notation [20] is a simple
data format based on one list with items separated by, of course,
commas. One of the most notorious departures from the use of
commas is multidimensional arrays in C, in which the syntax to
access an element at index x;y is not x;y, it is xy.
Efficient textual representation of structure
The idea of separating items in a list with a single symbol (or
word) seems simple, but turns out to have several surprisingly
tricky issues.
Consider how to represent a list in a language that does not have
any symbol analogous to the comma, Dyck Language interleaved
with data. How is the sibling relationship expressed? (First, note
the convention is to place the parent before the child, as in p(c),
although the opposite, (c)p is just as expressive.) One way is to
wrap brackets around each individual data item. Then the number
of brackets needed to represent a relationship must be increased
by 1 for all depths, so that (a)(b) means a and b are siblings, and
((c))((d)) means c and d are 1st cousins. A 2x2 array would be
((p)(q))((r)(s)). Although it works, it is far more verbose. Additionally it spoils the abbreviation of allowing siblings to be separated by a child, as in a(e)b, which must instead be (a(e))(b).
So, a better way is to always separate siblings with a child, using
a null child if the older sibling has no children, as in a()b. Then a
2x2 array can be represented with (p()q)(r()s).
Expanding to cousins is still a problem. With the addition of the
comma as a sibling separator, (p()q)(r()s) becomes (p,q)(r,s).
The sequence still has a “)(”, which the comma does not help reduce. An obvious extension is to introduce another symbol, say
semicolon, to separate 1st cousins. Then the sequence can become
(p,q;r,s).
What to do for 2nd cousins? Just add brackets to the semicolon, as in );(? Or employ yet another symbol to replace ))((?
How many symbols should be so employed? The ASCII committee
settled on 4, ASCII characters 28 through 31, though 8 were proposed [12]. They were called Information Separators [9]. We can
do better than that.
There are several issues with having 2 or more Information Separators that merit careful consideration.
First, consider the sequence p(q,r;s). q and r are siblings to
each other, and descendants of p, and s is 1st cousin to q and r.
There are several different more precise meanings this could have.
The semicolon can be given higher precedence than the brackets,
that is, all three of q, r, and s are children of p. In that case, this
particular sequence is invalid, because r and s cannot be children
of p and 1st cousins to each other. All children of the same parent
must be siblings.
Another interpretation is to allow a single opening bracket to
separate a variable number of generations instead of always one
generation. Then, since grandchildren of p can be 1st cousins to
one another, all 3 of q, r, and s must be grandchildren of p. But
this idea has the big disadvantage of adding context dependency
to the grammar. Whether q is a child or a grandchild of p cannot
be known until all the characters between the opening and closing
brackets are scanned. If a semicolon is found on the same level as
q, then q is a grandchild of p. If there are even deeper separators,
q is a great grandchild or even more distant descendant of p. If
none are found, then q is a child of p.
Best is to consider the semicolon as a combined open and close
bracket, )(, having the same precedence as any other bracket. In
that case, sis not a descendant of p, sis a nephew of p. That meaning does not add context. This does have more invalid strings, for
instance the simple sequence r;s is invalid because the brackets
are not balanced.
Conference’17, July 2017, Washington, DC, USA
Second, consider how to combine separators with colons. The
colon is a bracket, and should have the same precedence. Then a
sequence such as (p:q;r) means that p is parent to q, and not
parent or sibling to r. p is uncle to r, q is 1st cousin to r, and
r’s parent is null. Perhaps the easiest way to see this is to reverse
the colon transform to get (p(q;r)), then reverse the separator
transform to get (p(q)(r)). If p and rare supposed to be siblings,
and q a child of p, the correct way to represent that is not to use
semicolon or colon, it is p(q)r.
The 2 transforms, colon and separator, are mostly complementary, but in some cases can compete to reduce the same redundancies. The following table shows the results of transforming each of
the 14 Dyck words of length 8 (replacing ][ with a comma rather
than a semicolon, for greater visual clarity.)
Dyck word colon
separator both
1 [[[[]]]]
[:::]
[[[[]]]]
[:::]
2 [][[[]]]
[][::]
[,[[]]]
[,::]
3 [[][[]]]
[[]::]
[[,[]]]
[:,:]
4 [[]][[]]
[:][:]
[[],[]]
[[],:]
5 [[[][]]]
[:[]:]
[[[,]]]
[::,]
6 [[[]][]]
[[:]:]
[[[],]]
[:[],]
7 [[[]]][]
[::][]
[[[]],]
[[:],]
8 [][[][]]
[][[]:]
[,[,]]
[,:,]
9 [][[]][]
[][:][]
[,[],]
[,[],]
10 [[][][]]
[[][]:]
[[,,]]
[:,,]
11 [[][]][]
[[]:][]
[[,],]
[[,],]
12 [[]][][]
[:][][]
[[],,]
[[],,]
13 [][][[]]
[][][:]
[,,[]]
[,,:]
14 [][][][]
[][][][] [,,,]
[,,,]
The last column shows the result of applying the semicolon transform, followed by the colon transform. Applied second, the transform to colon can be blind to the presence of any separators, and
be correct and achieve maximum reduction. A separator acts as a
bridge, so that a colon can start a list, a natural looking use, rather
than opening the last item of a list.
If the separator transform is second, then to achieve maximum
reduction, as well as a correct transformation, it has to be done
with awareness of colons. A colon may be opening the last item
in a list, and it can be moved to the head. [[]:] can become [:,]
by replacing ]:, which is an open close pair of brackets, with a
separator, and then, replacing the opening bracket of the previous
item in the list with a colon. This can be repeated until the colon
has migrated to the front of the list. If the separator transform is
done blindly on a sequence with colons, it can be incorrect. [[]][]
is [:][] but then replacing the ][ with a separator gives [:,],
which is not correct. Correct is [[],]. Undoing [:,] shows that
sequence is actually [[][]], a list of 2 items.
Applying both transforms to the Ackermann function given earlier replaces a total of 11 bracket pairs with either a single separator
(the comma was used in this example) or a single colon:
( define :A x
cond : ( = y
(= x
(= y
else
y,
0)
0,
1)
:A
0,
∗ 2 y) ,
2,
:
x 1,
A x :
y 1)
Conference’17, July 2017, Washington, DC, USA
Third, what of types? Should the separated items be the same
types? The traditional meaning of a comma is as a separator only,
of untyped data.
Or, should text adjacent to a comma be interpreted as a type
name? A way to resolve that question is to provide another means
to add a type if desired, and let separators remain separators only.
For example, as mentioned in the section on types, the ‘$’ character could be used to indicate an alphanumeric sequence is a type
name. Deeper separators would need a list of types, or restrictions
on elements for which they can specify a new type, and while notations for that can of course be invented, there is little point when
opening brackets can accomplish that with reasonable efficiency
and without any additional rules.
Fourth, there are different potential meanings for runs of a separator symbol. 2 adjacent semicolons could mean that there is an
empty element in the middle, like for(;;i++) in C. Or, it could
mean that the separation between the data elements on either side
is deeper, that is, they are 2nd cousins instead of 1st cousins. What
should 2 semicolons mean, )()( or ))((? The former is the more
widely used meaning. The latter is accomplished by the limited
method of having more Information Separator symbols, which cannot neatly handle great depths. It seems useful to have clear and
concise ways to express either meaning. One way to do this is to
have 2 Information Separators, one for siblings and one for cousins.
The wrinkle is that repetition of these symbols would have the 2
different meanings. n of the sibling separator can mean there are
n 1 siblings who have been omitted from the list, while n of the
cousin separator can mean the cousins are nth cousins, being 1st
cousins only when n = 1. This approach, combined with an efficient way to express quantities, discussed next, can express both
meanings.
However, another way is not to use the system for expressing
quantities, and then assign different meanings to those quantities,
but to use typing. A semicolon could be followed by an integer
to indicate the depth of the divide, eg. “;3” means the adjacent
elements are 3rd cousins. Then a run of n semicolons can mean
that there are n 1 1st cousins in the middle, same as a run of n
commas means n 1 middle siblings. This makes it slightly harder
to support typed separators, but of course it can still be done, That
point is moot if sticking with the traditional meaning of separators
being typeless.
A minor matter is that separators have an inherent off-by-one
issue. A comma separated list usually contains one fewer commas
than data items. Often, specifications allow a meaningless trailing
comma to be present, for the convenience of programmers.
A big reason to support efficient representation of a cousin relationship and even reserve symbols especially for it rather than rely
on brackets is that it is a natural way to map multidimensional arrays to a hierarchical structure. Another reason is that people are
familiar with and like separators.
7
EFFICIENT REPRESENTATION OF
ARBITRARY QUANTITIES
Infinitely many numbers cannot be represented with single symbols from a finite set of symbols.
Brenton Chapin
Though we can’t collapse arbitrary quantities to single symbols,
we can however do better than using n symbols to represent n,
by employing the same principle used in the Arabic numbering
system that replaced unary numbering systems such as the Roman
one and hash marks. All this is well known, as is that a binary
numbering system has the minimum number of symbols needed
to represent quantities of n with log n symbols.
Can we do even better than log n, represent any arbitrary quantity of size n with even fewer symbols? No. For this question, the
Pigeonhole principle applies. As in data compression, to be able
to represent some quantities of amount n with fewer than log n
symbols (from a finite set of symbols), other quantities must be
represented with more than log n symbols. When the amounts are
averaged over all quantities n, the size is log n, or greater.
Numbering systems can be employed to represent structure. Rather
than come up with more and more symbols to represent greater
and greater quantities, as the ASCII committee did with their 4
separator symbols, can employ 2 symbols in a binary code.
Obviously any one symbol which may be repeated can be made
one member of a set of 2 symbols to be used in a binary encoding. But if there are many symbols which may be repeated, finding
enough symbols becomes a problem.
Since quantities are so useful, and unused symbols so precious, a
better idea is to reserve 2 symbols for a binary code for quantities
only, for any other symbol that may be repeated. For example,
instead of using 2 kinds of open bracket symbol in a binary code
as in something like [(([ to represent 9 open brackets, have 1001(
mean 9 open brackets, 1101* mean 13 asterisks, and so on.
Still better is to use an escape character and a decimal representation. The backslash can be used for this, as the only backslash
escape sequence that uses a number is n0, to mean the NULL character, ASCII 0. Then 9 open brackets can be represented with n9(.
One desirable additional symbol to allow is the minus sign, for negative quantities. If only integers are allowed, then there is no need
to overload the meaning of an escaped period for a decimal point
character.
This sort of representation is the well known idea of run-length
encoding (RLE) [5]. RLE is simple and easy, even relatively easy
for a person to understand without computer aid.
Of course there is the minor problem that the numeric symbols
themselves cannot be the object of a RLE escape sequence. There
are several easy ways to resolve that issue. Easiest is to simply not
support repetition of the digit characters, forcing the use of the
traditional method if repetition of a digit is wanted. Perhaps next
easiest is to employ a terminal symbol. To keep the representation
one character shorter, the terminal symbol can be optional, used
only if needed to remove ambiguity.
But RLE is very limited in what it can express. That keeps it
dirt simple and easy to read, but perhaps more expressiveness is
desirable, for such uses as repeating patterns, not only single characters. One simple example of such a sequence is the CR/LF pair.
With a trivial amount of additional syntax, it is possible to efficiently encode repeating patterns. A further use is as a repetition
of an escape. Suppose one has a string consisting of many, perhaps over half, of characters that must be escaped. One traditional
method is to inflate by up to double the quantity of characters by
preceding each special character with an escape character. That
Efficient textual representation of structure
can get difficult for a programmer to read, as seen in Perl’s regular expressions. A quantity that can be applied to indicate how
many characters are to be escaped can supersede the traditional escape character method. This notion is fairly obvious and has been
proposed on a number of occasions, for instance by Rivest in his
draft for S-expressions [15], for what he called “Verbatim representation”, and with Hollerith constants in FORTRAN 66 [13]. Perl’s
regular expressions has a similar mechanism.
While it is trivial to extend run length encoding to handle repeating patterns, there are still many other highly redundant strings
that this extended RLE cannot encode, yet are simple to describe.
The question is how far to go, how much complexity is really useful and can still be easily encoded? And, would it still be human
readable?
Perhaps an efficient way to represent “))(())((” and larger
combinations is also desirable? To encode such things, more is
required. A simple idea is to support the encoding of lists of quantities, a vector, rather than a single quantity. The escape character
can be employed as a separator. Then what is needed is agreement
on the meanings to assign to the multiple quantities. For example,
to encode 5 repetitions of a string of length 4, “abcd”, should it be
“n4n5abcd” or “n5n4abcd” or something else?
But if a vector of quantities is such a good idea, why not a tree
of quantities? Takes only 2 symbols to represent the structure of
a tree. However, the additional complexity is almost certainly too
much to remain human readable, and there’s the question of what
uses could we make of a tree of quantities?
One use for a vector of quantities is for the sizes of the dimensions of a multidimensional array. Such a usage creeps into the domain of external representation of structure. The interleaving can
be reduced to a single character used as a separator, or removed entirely. For instance, a 2x3 array with variable sized elements could
be notated as n2n3n? 1a,1b,1c,2a,2b,2c, using the same separator symbol every time, with the division between 1c and 2a known
to be deeper than the rest only because that info was given in the
vector of quantities. Or that 2x3 array with fixed sized elements
could be notated as n2n3n2 1a1b1c2a2b2c.
If means to represent something analogous to a Hollerith constant are provided, some probably will use it for very complicated
objects. Just serialize the data, and use the total length of the resulting string as the size. Supporting runs of the same symbol, and
blocks of data analogous to Hollerith constants, provides enough
for further elaboration if desired, while keeping the notation simpler.
We get away with unary representations, because we stick to
small and simple structures. If we seldom count higher than 3 or 5,
and almost never higher than 10, tally marks work fine. A check of
the Firefox source code reveals that only a few programs reached
a nesting depth of 15, with most never exceeding 10, so RLE for
opening brackets and colons, and quantities to indicate the depths
of separators are not going to remove much clutter. But perhaps
flatter structuring has been chosen to avoid the clutter that would
result from deeper nestings. And block escapes are still a viable
use of quantities.
Conference’17, July 2017, Washington, DC, USA
8 REPRESENTING STRUCTURE WITH
POSITIONING
The only ASCII control characters still really useful are the 2 for
indicating a new line of text. Next most used is tab, which has
ambiguous meaning and is easily and often replaced with spaces,
except in a few special cases such as Makefiles. The rest of the
ASCII control characters are very seldom seen, and when present,
modern applications may simply ignore their meanings [27]. Of
the 132,231 text files in the Firefox 50 source code, just 121 have
ASCII control characters other than the 3 most common: LF, tab,
and CR. A mere 5 files use ANSI escape sequences, which start with
the Escape character (ctrl-[, ASCII 27), and that only to set text
colors.
ASCII’s minimal means of positioning text is sufficient but not
efficient or neat. One of the worst inefficiencies is the very repetitive use of spaces to indent lines of text. Some ANSI escape sequences address this issue, but not well. The VT100 series text
terminals became popular in large part because they adopted and
extended the ANSI escape sequences. Yet they have not been much
used outside of terminal control. They did not grow beyond that
niche to become common within text files. Colored text is the ANSI
escape sequence most used, yet it is rare. One of the most common uses of colored text, highlighting of source code, does not use
ANSI at all, even though editing may still be done in a terminal
that supports ANSI. Rather, text editors parse the source code being edited to compute which colors to assign, updating in real time
as the user makes changes. HTML and CSS can specify colors directly, and are not limited to a tiny 16 color palette. That and word
processor options have become the way to set text and other colors in documents. ASCII and ANSI must use a fixed width font to
position text accurately, and consequently, source code is almost
always viewed in such fonts.
What sort of positioning information would be most useful?
Means of clear, easy, and minimal description of position that best
supports useful structures should be leading contenders. Indentation is the most popular way to express hierarchy through position
alone. It is so common that even though curly brace languages do
not use indentation, coders are exhorted to use “proper indentation” anyway, so that their source code is more readable. Perhaps
the most prominent and distinctive feature of the Python programming language is the use of pure positioning to indicate code structure. Another major use is the alignment of columns, usually for
tables. The ASCII tab character does not do either of these well.
Superscripting and subscripting can be considered a kind of positioning. It has a major limitation in that it does not scale. Each
successively deeper nesting requires progressively smaller text, which
soon becomes too small to read.
A proposal is to reassign 4 ASCII control characters for indentation. 3 of them can be increase indent (push), revert indent (pop),
and boost indent, analogous to the 2 brackets and colon in a closing 3 system. These characters can be invisible and have a width of
zero, not directly affecting the position of text. They only change
the level of indentation. The 4th character can mean forward to
the next indentation, replacing the leading spaces on all indented
lines of text. It could also mean advance to the next line, but that
Conference’17, July 2017, Washington, DC, USA
would be less flexible, wouldn’t support situations in which text
such as a line number is wanted before the indentation.
These characters do not specify the size of an indentation, only
the number of levels. This would allow individual users to set the
indentation size themselves without affecting others. It could also
make variable width fonts usable, as the problem of what unit to
use to specify indentation sizes is entirely avoided. It does add one
item to the state a text editor or viewer must maintain: a stack of
indentation levels.
Indentation characters could ease editing of source code. There
would be no more need to shift blocks of code several spaces right
or left by changing the number of leading spaces on each line,
whether by manually adding or deleting each space, or by using
some smart editor function such as that assigned to the tab key in
EMACS.
For the columns of tables, need better tab functionality. It would
be desirable not to rely on the use of a monospace font to recreate
the intended horizontal alignments. A limitation to dump is any
sort of tiny maximum distance of 8 spaces. Further, it should be
independent of any indentation setting. The C1 set of control characters contains several intended for tabular positioning, but they
do not do enough. One problem is that the state they set up is
global. Another is that they still implicitly depend upon a fixed
font, using the cursor position for fixing the location of tab stops.
It is basically a copy of the ideas and means used in the most advanced typewriters, with all their limitations.
The means HTML provides for laying out tables is fairly comprehensive and flexible, and proven over many years of use. If
better handling of tables is desired in plain text, copying HTML’s
handling and a subset of capabilities concerning tables into control character functionality seems a good approach. Lightweight
markup languages such as Markdown [26] and Bulletin Board Code
(bbcode) [22], arose to satisfy the desire to be able to create lists
and tables in text based forums more easily than with HTML. This
shows that many users like text editing to have such capabilities.
9
CONCLUSION
This paper proposed several changes in standard textual notations
to eliminate redundancy that may be hampering the human readability of structured documents such as source code. Proving that
human readability is improved was not attempted. Instead, the paper surmises that some kinds of redundancy merely add clutter,
showed where and how redundancy lurks, and proposed ways to
eliminate it. Good answers to Lots of Idiotic Spurious Parentheses
have been desired for a long time, and perhaps until now have not
been satisfactory.
Notation that scales and adds expressiveness, and allows much
more brevity without sacrificing clarity, is especially preferred. Punctuation with long histories in natural languages, especially English,
was tapped as a guide, in part because those uses are familiar to
people literate in those languages.
The first proposed change was to add a 3rd kind of bracket symbol roughly equivalent to the meaning of the colon in English, so
that a parent–child relationship represented as “(p(c))” is instead
represented as “(p:c)”. Proof was given that this 3 symbol system
Brenton Chapin
can collapse all runs of 2 or more closing brackets to a single closing bracket.
The idea of a universal closing bracket was presented. ”fa (b
[c] d) eg” can be represented as “fa (b [c] d] e]”, reducing the number of different symbols required, as ‘)’ and ‘g’ are no
longer needed.
More use of separators was proposed to replace sequences of
closing brackets followed by opening brackets. “((a()b)(c()d))”
can be represented with “((a,b;c,d))”.
Ways of adding types to the structure were discussed.
Positioning was recognized as an important way of denoting
structure. It is observed that means of expressing position have
been neglected. Markup languages limit themselves to data, and
are not much used for writing of other structured information such
as source code. Moreover, by using visible text to express position,
and requiring translation with special tools such as a web browser,
they fail at the goal of using position alone to express structure.
The means provided in ASCII work only with monospace fonts,
and require much wasteful redundancy. Repurposing some of the
unused control characters to better support indentation and tabular structure was proposed.
Together, these changes reduced the number and quantity of
symbols needed. They improved the amount of data compression
obtained by general purpose data compression programs. They reduced the size of the source code. Whether the goal of greater
human readability was also achieved was not studied, but it was
surmised that removing redundancies in the notation does help
with readability.
REFERENCES
[1] Leonhard Euler. “De infinitis curvis eiusdem generis seu methodus inveniendi aequationes pro infinitis curvis eiusdem generis,” Commentarii
academiae scientiarum Petropolitanae 7, 1740, pp. 174-189. Original available
at http://eulerarchive.maa.org/docs/originals/E044.pdf, English translation at
http://www.17centurymaths.com/contents/euler/e044tr.pdf .
[2] Dirichlet, P. G. L., “Beweis des Satzes, dass jede unbegrenzte arithmetische Progression, deren erstes Glied und Differenz ganze Zahlen
ohne gemeinschaftlichen Factor sind, unendlich viele Primzahlen
enthlt”, Abhand. Ak. Wiss. Berlin, 1837. English translation available at
http://arxiv.org/pdf/0808.1408v2.pdf.
[3] Florian Cajori. “A History of Mathematical Notations.” 1928-1929. Paragraph
643, Vol II, pg. 268.
[4] Jan Lukasiewicz. “Uwagi o aksjomacie Nicoda i ’dedukcji uoglniajfbcej’ ”
Ksiga pamifbtkowa Polskiego Towarzystwa Filozoficznego, Lww 1931.
[5] Oliver, B. M. “Efficient Coding.” Bell System Technical Journal, 31: 724-750.
doi:10.1002/j.1538-7305.1952.tb01403.x
[6] A. J. Perlis and K. Samelson. “Preliminary Report – International Algebraic
Language,” Communications of the ACM, July 1958.
[7] Kenneth E. Iverson. A Programming Language. John Wiley & Sons, Inc., New
York, NY, USA. 1962.
[8] Richards, Martin. “BCPL: A tool for compiler writing and system programming”
[9] Winett, J., ”EBCDIC Codes and Their Mapping to ASCII”, RFC 183, DOI
10.17487/RFC0183, July 1971, http://www.rfc-editor.org/info/rfc183
[10] Abelson, Harold and Sussman, Gerald Jay, with Sussman, Julie. “Structure
and Interpretation of Computer Programs.” 1985.
[11] Felsenstein,
Joe
et
al.
The
Newick
tree
format.
http://evolution.genetics.washington.edu/phylip/newicktree.html
[12] Bemer,
Robert.
“The
Great
Curly
Brace
Trace
Chase”
http://www.bobbemer.com/BRACES.HTM
[13] FORTRAN 77 4.0 Reference Manual. SunSoft. Part No.: 802-2998-10 Revision
A, November 1995. p. 38. https://archive.org/details/Fortran77Manual
[14] Tim Bray,
C. M.
Sperberg-McQueen,
et.
al.
Extensible
Markup Language (XML),
W3C
Working
Draft 14-Nov-96.
https://www.w3.org/TR/WD-xml-961114
[15] Rivest, R. “S-Expressions” Network Working Group, Internet Draft, May 4,
1997. Section 4.1. http://people.csail.mit.edu/rivest/Sexp.txt
Efficient textual representation of structure
[16] Goldfarb, Charles F. “SGML: The Reason Why and the First Published Hint.”
Journal of the American Society for Information Science, volume 48, number
7 (July 1997). Annotated reprint of Goldfarb, Charles F., Mosher, Edward J.,
and Peterson, Theodore I. “An Online System for Integrated Text Processing”
Journal of the American Society for Information Science, volume 7 (Oct 1970).
http://www.sgmlsource.com/history/jasis.htm.
[17] Oren Ben-Kiki, Clark Evans, Brian Ingerson. YAML Ain’t Markup Language
(YAML) 1.0. Final Draft 2004-JAN-29. http://yaml.org/spec/1.0/
[18] Wall,
Larry.
“Synopsis
4:
Blocks
and
Statements”
http://design.perl6.org/S04.html
[19] Markstrum, Shane. “Staking Claims: A History of Programming Language
Design Claims and Evidence.”
[20] Shafranovich, Y., ”Common Format and MIME Type for Comma-Separated
Values (CSV) Files”, RFC 4180, DOI 10.17487/RFC4180, October 2005,
http://www.rfc-editor.org/info/rfc4180
[21] Buse, Raymond P.L. and Weimer, Westley R. “A Metric for Software Readability.” Proceedings of the 2008 International Symposium on Software Testing
and Analysis, pp. 121-130.
[22] anonymous authors. Bulletin Board Code. http://www.bbcode.org
[23] Yasuoka, Kiochi and Yasuoka, Motoko. “On the Prehistory of QWERTY.” ZINBUN (2011), 42: 161-174. https://doi.org/10.14989/139379
[24] ECMA
404
The
JSON
Data
Interchange
Format.
https://www.ecma-international.org/publications/files/ECMA-ST/ECMA-404.pdf
[25] HTML Elements. http://www.w3schools.com/html/html elements.asp
[26] MacFarlane, John. CommonMark Spec. http://spec.commonmark.org/
[27] Leonard, S., ”Guidance on Markdown: Design Philosophies, Stability Strategies, and Select Registrations”, RFC 7764, DOI 10.17487/RFC7764, March 2016.
http://www.rfc-editor.org/info/rfc7764
[28] Open Tree of Life, v9.1. http://files.opentreeoflife.org/synthesis/opentree9.1/opentree9.1 tree.tgz
Conference’17, July 2017, Washington, DC, USA
| 6cs.PL
|
Some theoretical results on tensor elliptical
distribution
arXiv:1709.00801v1 [math.ST] 4 Sep 2017
M. Arashi
Department of Statistics, School of Mathematical Sciences
Shahrood University of Technology, Shahrood, Iran
Abstract: The multilinear normal distribution is a widely used tool in tensor analysis
of magnetic resonance imaging (MRI). Diffusion tensor MRI provides a statistical
estimate of a symmetric 2nd -order diffusion tensor, for each voxel within an imaging
volume. In this article, tensor elliptical (TE) distribution is introduced as an extension to the multilinear normal (MLN) distribution. Some properties including the
characteristic function and distribution of affine transformations are given. An integral representation connecting densities of TE and MLN distributions is exhibited
that is used in deriving the expectation of any measurable function of a TE variate.
Key words and phrases: Characteristic generator; Inverse Laplace transform; Stochastic representation; Tensor; Vectorial operator.
AMS Classification: Primary: 62E15, 60E10 Secondary: 53A45, 15A69
1
Introduction
Nowadays, analysis of matrix-valued data sets is become quite common in medical sciences, since
the collected data are of multiple-way (multiple-component) arrays. For example in medical
imaging, it has become possible to collect magnetic resonance imaging (MRI) data that can be
used to infer the apparent diffusivity of water in tissue in vivo. In this regard, there is a need
to consider parallel extensions of bilinear forms1 , namely tensor matrices. Tensor matrices have
been commonly used to approximate the diffusivity profile of images. This approximation yields
a diffusion tensor magnetic resonance imaging (DT-MRI) data set. Processing of DT-MRI data
sets has scientific significance in clinical sciences. Figure 1 shows the tensor filed in a diffusion
MRI image.
In image analysis, the characteristic or precision matrix of the underlying model for tensor observations and distribution of eigenvalues play deterministic roles. Hence, the underlying
tensor distribution influences the respective inference. The use of tensor and associated distributional structure in Statistics dates back to McCullagh (1987). McCullagh (1984) had already
introduced tensor notation in statistics with particular reference to the computation of polynomial cumulants. For a selective papers about tensors and applications in statistics, we refer to
Sakata (2016).
In all pronounced studies in statistical tensor analysis, tensor normal (or multilinear normal)
distribution is employed for the underlying distribution of observations. However, a slight change
in the specification of the distribution, as pointed by Basser and Pajevic (2003), may play havoc
1
Bilinear form is a two-way (two-component) array, with each component represents a vector of observations
1
Figure 1: Visualization of tensor filed of a brain
Figure 2: Visualization of a 3-dimensional data set as a 3rd -order tensor.
on the resulting inferences. To broaden the scope of the distributions and achieve reasonable
inferential conclusions, and in order to accommodate the heavier tailed distributions in a reasonable way and produce robust inference procedures for applications, tensor t-distribution can
be employed in related analysis. From a broader view point, one may define the class of tensor
elliptical distributions which includes the latter distribution as special one. In this article, we
define a new class of tensor elliptical distributions and study some of its statistical properties.
2
Preliminaries
In this section we introduce related notation to our study and give some definitions. We adhere
to the notation of Ohlson et al. (2013).
Let X be a tensor of order k (kth -order tensor, in tensor parlance), with the dimension
p = (p1 , p2 , . . . , pk ) in the x = (x1 , x2 , . . . , xk ) direction. Figure 2 shows the special case when
k = 3. Indeed 2nd -order tensor is matrix, 1st -order tensor is vector, and 0th -order tensor is
scalar.
In connection with Figure 2, Figure 3 shows that the collected data can be interpreted as
tensor, where the assessment of cardiac ventricular with helical structure is done by DT-MRI.
2
Figure 3: Helical structure of the cardiac ventricular anatomy
A vectorial representation of a tensor, make the related inference much simpler. Let vec X
denote the vectorization of tensor X = (xi1 i2 ···ik ), according to the definition of Kolda and Bader
(2009) given by
vec X
=
=
p1
X
···
pk
X
xi1 i2 ···ik e1i1 ⊗ · · · ⊗ ekik ,
ik =1
i1 =1
X
xi1 i2 ···ik ep1:k ,
(2.1)
Ip
, ..., e1i1 are the unit basis vectors of size pk , pk−1 , ..., p1 , respectively, ep1:k =
where ekik , eik−1
k−1
e1i1 ⊗ · · · ⊗ ekik , where ⊗ denotes the Kronecker product, Ip is the index set defined as Ip =
{i1 , . . . , ik : 1 ≤ ij ≤ pj , 1 ≤ j ≤ k}. In Ohlson et al. (2012), the authors concentrated
on the estimation of a Kronecker structured covariance matrix of order three (k = 3), the so
called double separable covariance matrix, generalizing the work of Srivastava et al. (2008), for
multilinear normal (MLN) distributions.
Let T p denote P
the space of all vectors x = vec X , where X is a tensor of order k, i.e.,
p
p
T = {x : x =
Ip xi1 i2 ···ik e1:k }. Note that this tensor space is described using vectors.
However, we can define tensor spaces using matrices. This is given in the following definition.
Definition 1 Let
o
n
P
(i) T pq = X : X = Ip ∪Iq xi1 ,··· ,ik ,j1 ,··· ,jl ep1:k (dq1:l )′ , Iq = {j1 , . . . , jl : 1 ≤ ji ≤ pi , 1 ≤ i ≤ l}
pq
(ii) T pq
: X = X 1 ⊗ . . . ⊗ X k , X i : pi × q i }
⊗ = {X ∈ T
(iii) T p⊗ = X ∈ T pp
⊗ : X = X 1 ⊗ . . . ⊗ X k , X i : pi × pi
Theorem 1 (Ohlson et al., 2013) A tensor X is MLN of order k, denoted by X ∼ Np (µ, Σ)
1
if x = µ + Σ 2 u, where x, µ ∈ T p , Σ ∈ T p⊗ , p = (p1 , . . . , pk ), and the elements of u ∈ T p are
independent standard normally distributed.
Note that Σ ∈ T p⊗ can be written as Kronecker product Σ = Σ1 ⊗ . . . ⊗ Σk
3
Indeed, Theorem 1 configures the MLN distribution using the stochastic representation of
vector x ∈ T p . This methodology can be mimicked to extend the above result for elliptical
models. Before revealing the main result of this paper, we need to consider the definition of
matrix elliptical distributions.
3
Tensor Elliptical Distributions
Q
∗
Let u(p ) , p∗ = ki=1 pi , denote a random vector distributed uniformly on the unit sphere
∗
surface in Rp , with characteristic function (cf) Ωp∗ (·). Hereafter, using Theorem 2.2 of Fang
et al. (1990), we propose a definition for tensor elliptical (TE) distribution. The methodology
behind our definition of TE distribution comes from two facts: (1) a random matrix X has
matrix elliptical distribution if and only if vec X has vector-variate elliptical distribution which
will be used for tensor (see Gupta et al., 2013) (2) the difference between vector-variate elliptical
and TE lies in the structure of the parameter space generated by µ and Σ.
Definition 2 A random tensor X is TE of order k, denoted by X ∼ Ep (µ, Σ, ψ), if
1
∗
x = vec(X ) = µ + RΣ 2 u(p ) ,
1
(3.1)
∗
where x, µ ∈ T p , Σ 2 ∈ T p⊗ is any square root, p = (p1 , . . . , pk ), R ≥ 0 is independent of u(p ) ,
and R ∼ F , for some cumulative distribution function (cdf ) F (·) over [0, ∞), is related to ψ by
the following relation
Z
ψ(x) =
R+
Ωp∗ (xr 2 )dF (r).
(3.2)
The question arises whether the parameters in Definition 2 are uniquely defined. TheQanswer
is no. To see this, assume that ai , i = 1, . . . , k are positive constants such that a∗ = kj=1 aj ,
∗
1
∗
Σj = aj Σj , j = 1, . . . , k and ψ (x) = ψ p∗ x . Then Ep (µ, Σ, ψ) and Ep (µ, Σ∗ , ψ ∗ ), where
Σ∗ = Σ∗1 ⊗ . . . ⊗ Σ∗k , define the same tensor elliptical distribution.
Using vector representation, x ∈ T p , we can conveniently write the probability distribution
function (pdf) of a TE extending the pdf of MLN distribution. The following result gives the
pdf of a random tensor elliptical if it possesses a density, as an extension to Ohlson et al. (2013).
Theorem 2 Under the assumptions of Definition 2, the pdf of the TE distribution is given by
1
fX (x) = |Σ|− 2 g (x − µ)′ Σ−1 (x − µ) ,
where g(·) is a non-negative function (density generator, say) satisfying
Z
1 ∗
y 2 p −1 g(y)dy < ∞.
R+
We designate X ∼ Ep (µ, Σ, g).
In a similar fashion, we have the following result.
Theorem 3 Let X ∼ Ep (µ, Σ, ψ). Then, its characteristic function has form
′
φX (S) = eiS µ ψ(S ′ ΣS),
Remark 1 Since
1
|Σ|− 2
1
= |Σ1 ⊗ Σ2 ⊗ . . . ⊗ Σk |− 2
4
S ∈ T p.
(3.3)
=
=
|Σ1 |
k
Y
i=1
p∗
p1
− 1
2
× |Σ2 |
∗
p
− 2p
|Σi |
p∗
p2
− 1
2
× . . . × |Σk |
p∗
pk
− 1
2
i
1 ∗
taking g(y) = (2π)− 2 p exp − 12 y in Definition 2, gives the pdf of MLN distribution (as given
in Theorem 1 of Ohlson et al., 2013) as
!
k
Y
p∗
1 ∗
1
− 2p
′
−1
−2p
(3.4)
fX (x) = (2π)
|Σi | i exp − (x − µ) Σ (x − µ) ,
2
i=1
where Σ is positive definite, x, µ ∈ T p , Σ ∈ T p⊗ , and p∗ =
Qk
i=1 pi .
The following result gives the distribution of affine transformations for TE variates.
Theorem 4 Let X ∼ Ep (µ, Σ, ψ), with vec X ∈ T p , A ∈ T qp is nonsingular, and B ∈ T q .
Then, AX + B ∼ Eq (Aµ + B, AΣA′ ), where Aµ + B ∈ T q and AΣA′ ∈ T q⊗ .
Proof. Let y = Ax + B, where x = vec(X ). From the stochastic representation in Definition
1
∗
2, the proof directly follows from y = (Aµ + B) + R(AΣ 2 )u(p ) .
The following result is a direct consequent of Theorem 2.16 of Gupta et al. (2013) for tensor
elliptical distributions.
Theorem 5 Under the assumptions of Definition 2, the pdf of R has from
1 ∗
2π 2 p p∗ −1
2
r
g
r
,
hR (r) =
Γ 21 p∗
r ≥ 0.
The following theorem reveals the distribution of quadratic form for a special case.
Theorem 6 Let X ∼ Ep (0, Σ(1) , ψ), where Σ(1) = Σ ⊗ I p2 ⊗ . . . ⊗ I pk ∈ T p⊗ ,
Then, the pdf of A = X X ′ is given by
Qk
j=2 pj .
1
1 (1)
π p |Σ|− 2 p1
p −p1 −1
2
g(tr Σ−1 A)
f (A) =
1 (1) |A|
Γp 1 2 p
∗
where p(1) =
p = (p1 , . . . , pk ).
In the forthcoming section, we provide a weighting representation of the pdf of TE variate
using the Laplace operator.
4
Weighting Representation
Although the proposed theorems in previous section are obtained conventionally, it is not easy to
achieve other statistical properties of the TE distributions from Definition 2 straightforwardly.
However, under a mild conditions, one can make connection between densities of TE and MLN
pdfs and derive other properties of the TE distributions using MLN distributions. In this
section, we propose a weighting representation which connects densities of the TE and MLN
distributions. This result is given in the following theorem.
5
Theorem 7 Let X ∼ Ep (µ, Σ, g), where µ ∈ T p , Σ ∈ T p⊗ and g : R+ → R+ . Also assume that
g(s2 ) is differentiable when s2 is sufficiently large, and g(s2 ) vanishes faster than s−k ; k > 1 as
s → ∞. Then, the pdf of X can be represented as an integral of series of MLN pdfs given by
Z
W(t)fNp (µ,t−1 Σ) (x)dt,
fX (x) =
R+
where fNp (µ,t−1 Σ) (·) is the pdf of Np (µ, t−1 Σ) and W(·) is a weighting function.
Proof. Let s2 = 12 (x − µ)′ Σ−1 (x − µ) and
1 ∗
W(t) = (2π) 2 p t−
p∗
2
L−1 g 2s2 ,
where L is the Laplace transform operator. It should be noted that under the regularity condition
∗
Q
1
−p
on g(s2 ), the inverse Laplace transform exists. Then, from |Σ|− 2 = ki=1 |Σi | 2pi , we have
h
∗i
1
1 ∗ p
1
f (x) = |Σ|− 2 g(2s2 ) = |Σ|− 2 L W(t)(2π)− 2 p t 2
"
#
k
∗ Y
p∗
− 2p
− 12 p∗ p2
= L W(t)(2π)
t
|Σi | i
i=1
=
=
=
Z
Z
Z
− 12 p∗
R+
W(t)(2π)
t
p∗
2
k
Y
i=1
− 12 p∗
k
Y
R+
R+
W(t)fNp (µ,t−1 Σ) (x)dt.
i=1
|Σi |
− k1
W(t)(2π)
|t
∗
p
− 2p
i
!
∗
p
− 2p
Σi |
i
2
e−ts dt
!
1
′
−1 Σ)−1 (x−µ)
e− 2 (x−µ) (t
dt
The proof is complete.
Thus, a TE variable is an integral over all MLN variables having the same covariance subject
to different scales.
Since fX (·) is the pdf of X , using Fubini’s theorem, we obtain
Z Z
Z
W(t)fNp (µ,t−1 Σ) (x)dtdx
fX (x)dx =
1=
χ
Z χ R+ Z
W(t) fNp (µ,t−1 Σ) (x)dxdt
=
+
χ
R
Z
W(t)dt
(4.1)
=
R+
where χ is the sample space. Hence, for positive weighting functions W(·), the weighting representation of TE distributions can be interpreted as an scale mixture of MLN distributions.
However, sometimes, W(·) can be negative. Note that a TE distribution is completely defined
by the matrix Σ ∈ T p⊗ and the scalar weighting function W(·).
Theorem 7 enables us to describe more properties of TE distributions via MLN distributions.
This can be done using the following important result.
Theorem 8 Let x ∼ Ep (µ, Σ, g), µ ∈ T p , Σ ∈ T p⊗ and g : R+ → R+ with weighting function
W(·), and B(x) be any Borel measurable function of x ∈ T p . Then, if E[B(x)] exists, we have
Z
W(t)ENp (µ,t−1 Σ) [B(x)]dt
E[B(x)] =
R+
6
5
Examples
In this section, we provide some examples of TE distributions based on Definition 2 with respective weighting function, as defined in Theorem 7.
Firstly, we consider some examples in which the weighting function W(.) is always positive,
resulting to scale mixture of multilinear normal distributions.
(i) Multilinear normal distribution (Ohlson et al., 2013)
The weighting function has form
W(t) = δ(t − 1),
where δ(·) is the dirac delta or impulse function having the property
R
f (x)δ(x)dx = f (0),
R
for every Borel-measurable function f (·).
(ii) Multilinear ε-contaminated normal distribution
We say the random tensor X ∈ T p has multilinear ε-contaminated normal distribution if
it has the following density
!
k
Y
p∗
1
1
− 2p
′
−1
|Σi | i
fX (x) =
(1 − ε) exp − (x − µ) Σ (x − µ)
1 ∗
2
(2π) 2 p
i=1
1
ε
+ p∗ exp − 2 (x − µ)′ Σ−1 (x − µ) .
σ
2σ
Then it can be concluded that the weighting function is given by
W(t) = (1 − ε)δ(t − 1) + εδ(t − σ 2 ).
(iii) Tensor t-distribution
We say the random matrix X ∈ T p has tensor t-distribution if it has the following density
∗
p∗
−(p∗ +ν)
ν 2 Γ p 2+ν
1
′ −1
.
(5.1)
fX (x) =
1 + (x − µ) Σ (x − µ)
1 ∗
ν
π 2 p Γ( ν2 )
ν
The corresponding weighting function has form W(t) =
tν
( tν2 ) 2 e− 2
.
tΓ( ν2 )
The tensor Cauchy distribution is obtained by setting ν = 1 in (5.1).
It is of much interest to consider cases in which the weighting function W(.) is not always positive. Such kind of distributions are not scale mixture of multilinear normal
distributions. The item below is not a tensor distribution, however it is 0th -order tensor
distribution.
(iv) The one-dimensional distribution with the following density
√
x 4 −1
2
,
1+
f (x) =
πσ
σ
where the weighting function is given by W(t) = √1tπ sin 2t .
7
6
Inference
Theorem 9 Suppose that tensor variables X 1 , . . . , X n are jointly distributed with the following
pdf
k
n
Y
X
p∗
−
|Σi | 2pi g
x′j Σ−1 xj , Σ = Σ1 ⊗ Σ2 ⊗ . . . ⊗ Σk
i=1
j=1
(r)
(2)
(3)
(k)
such that σp2 p2 = σp3 p3 = . . . = σpk pk = 1, where Σr = σij . Further, suppose g(·) is such
that g(x′ x) is a pdf in Rp and y p /2 g(y) has a finite positive maximum yg (see Anderson et al.,
1986 for the existence of yg ). Suppose that Σ̃ is an estimator which obtains from solving the
following equations (see Ohlson et al., 2013)
∗
∗
Σ̃1 =
Σ̃r
1
p∗2:k n
n
X
x′j Σ−1
2:k xj
j=1
and, for r = 2, . . . , k
n
−1
X
1
2,r(r)
2,r(r) ′
2,r
xi
,
Σ
x
=
i
1:k\r
p∗1:r−1 p∗r+1:k n
j=1
where
2,r(r)
=
xj
X
lp
pr
xi1 ,...ik e2,r
i1 :ik \ir eir
′
Σ2,r
1:k\r = Σ2 ⊗ . . . ⊗ Σr−1 ⊗ Σ1 ⊗ . . . ⊗ Σk
p
Then, the MLE of Σ is given by
Σ̂ =
− p1∗
Proof. Let A = |Σ|
p
r−1
r+1
= epi11 ⊗ epi33 ⊗ . . . ⊗ eir−1
⊗ eir+1
⊗ . . . ⊗ epikk
e2,r
i1 :ik \ir
p∗
Σ̃
yg
Σ. Also for any j = 1, . . . , n write
1
dj = x′j Σ−1 xj = |Σ|− p∗ x′j A−1 xj .
1
Since |Σ|− 2 =
Qk
∗
p
− 2p
i=1 |Σi |
(6.1)
, the likelihood can be written as
n
X
1
dj
L = |Σ|− 2 g
i
j=1
=
|Σ|
=
1
p∗
n
X
j=1
− p∗
2
n
X
j=1
p∗
dj
− p∗
x′j A−1 xj
2
2
d
p∗
2
g
n
X
j=1
g(d),
dj
(6.2)
P
where d = nj=1 dj .
The maximum of (6.2) is attained at  = à and dˆ = yg . Then the MLE of Σ is given by
1
Σ̂ = |Σ̂|
1
p∗
 =
|Σ̂| p∗
1
|Σ̃| p∗
8
Σ̃.
(6.3)
On the other hand, from (6.1) we get
|Σ̂|
|Σ̃|
1
p∗
1
p∗
=
=
′ −1
j=1 xj  xj
Pn ˆ
j=1 dj
Pn
′ −1
j=1 xj à xj
Pn ˜
j=1 dj
Pn
′ −1
j=1 xj à xj
=
Pn
=
Pn
dˆ
′ −1
j=1 xj à xj
d˜
′ −1
j=1 xj à xj
=
Pn
=
Pn
yg
′ −1
j=1 xj à xj
p∗
(6.4)
Substituting (6.4) in (6.3) and using Theorem 4.1 of Ohlson et al. (2013) gives the result.
7
Conclusion
In this article, for the purpose of robust inferring on diffusion tensor magnetic resonance imaging
(DT-MRI) observations, we proposed a class of tensor elliptical (TE) distributions. This class
includes many heavier tail distributions than the tensor normal or multilinear normal (MLN)
distribution. Important statistical properties including the characteristic function along with
the distribution of affine transformations derived. A weighting representations also exhibited
that connects densities of TE and MLN distributions.
References
Basser, P. J. and Pajevic, S. (2003). A normal distribution for tensor-valued random variables:
applications to diffusion tensor MRI. IEEE Transactions on Medical Imaging, 22(7):785794.
Fang, K.T., Kotz, S., Ng, K.W. (1990) Symmetric Multivariate and Related Distributions. Chapman and Hall, London.
Gupta, A.K. and Varga, T., and Bondar, T. (2013) Elliptically Contoured Models in Statistics
and Portfolio Theory, 2rd Ed., Springer, New York. Kolda, T.G. and Bader, B.W. (2009).
Tensor decompositions and applications. SIAM Review, 51(3):455-500.
McCullagh, P. (1987). Tensor Methods in Statistics, Chapman & Hall, London.
McCullagh, P. (1984). Tensor notation and cumulants of polynomials. Biometrika, 71(3):461476.
Ohlson, M., Rauf Ahmad, M., and von Rosen, D. (2012). More on the Kronecker structured
covariance matrix. Communications in Statistics Theory and Methods, 41:2512-2523.
Ohlson, M., Rauf Ahmad, M., and von Rosen, D. (2013). The multilinear normal distribution:
Introduction and some basic properties. Journal of Multivariate Analysis, 113:37-47.
Sakata, T. (2016). Applied Matrix and Tensor Variate Data Analysis, Springer, Japan.
Srivastava, M., von Rosen, T., and von Rosen, D. (2008). Models with a Kronecker product
covariance structure: estimation and testing. Mathematical Methods in Statistics, 17:357370.
9
| 10math.ST
|
arXiv:1711.04260v1 [cs.CV] 12 Nov 2017
Evaluation of trackers for Pan-Tilt-Zoom Scenarios
Yucao Tang
Guillaume-Alexandre Bilodeau
University of Electronic Science and Technology of China
Chengdu, China
tang [email protected]
LITIV Lab.
Polytechnique Montréal
Montréal, Canada
[email protected]
Abstract—Tracking with a Pan-Tilt-Zoom (PTZ) camera has
been a research topic in computer vision for many years.
Compared to tracking with a still camera, the images captured
with a PTZ camera are highly dynamic in nature because the
camera can perform large motion resulting in quickly changing
capture conditions. Furthermore, tracking with a PTZ camera
involves camera control to position the camera on the target. For
successful tracking and camera control, the tracker must be fast
enough, or has to be able to predict accurately the next position
of the target. Therefore, standard benchmarks do not allow to
assess properly the quality of a tracker for the PTZ scenario. In
this work, we use a virtual PTZ framework to evaluate different
tracking algorithms and compare their performances. We also
extend the framework to add target position prediction for the
next frame, accounting for camera motion and processing delays.
By doing this, we can assess if predicting can make long-term
tracking more robust as it may help slower algorithms for keeping
the target in the field of view of the camera. Results confirm that
both speed and robustness are required for tracking under the
PTZ scenario.
Index Terms—Pan-Tilt-Zoom tracking, performance evaluation, tracking algorithms
I. I NTRODUCTION
Tracking with a Pan-Tilt-Zoom (PTZ) camera has been a
research topic in computer vision for many years. Compared
to tracking with a still camera, the images captured with a PTZ
camera are highly changing in nature because the camera can
perform large motion resulting in quickly changing capture
conditions. Furthermore, tracking with a PTZ camera involves
controlling a camera and it is an online process. Therefore,
standard benchmarks, e.g. [1], do not allow to evaluate the
performance of a tracker for the PTZ scenario because they do
not account for camera control, neither for drop frames caused
by spending too much time for processing a frame to track a
target. A specific benchmark is required for evaluation in this
scenario. In this work, we used the virtual PTZ framework
that has been developed by [2] to evaluate recent trackers
for the PTZ tracking scenario. The goal of this evaluation is
to assess the performance of trackers in an online tracking
scenario where a combination of speed and robustness is
required. PTZ scenarios include many changes in scale and
in illumination and large motion. We evaluated 19 tracking
algorithms and compared and analyzed their performances.
Many of these trackers were tested in the VOT 2016 [1]
benchmark. We also extended the framework to add target
This work was conducted while Yucao Tang was doing a MITACS Globalink internship at Polytechnique Montréal
position prediction for the next frame, accounting for camera
motion and processing delays. By doing this we can assess if
prediction can help to make long-term tracking more robust
and help slower algorithm for keeping the target in the field
of view of the camera.
II. E VALUATION FRAMEWORK
AND METRICS
Chen et al. [2] proposed a C++ framework to evaluate in
a reproducible way trackers in PTZ scenarios. Because of the
online nature of this scenario, the authors proposed the use of
a PTZ camera simulator that pans, tilts and zooms based on a
spherical video captured offline. The simulator also includes
relevant delays that result in drop frames if tracking takes too
long and if the target has a large motion in the image plane.
These delays are categorized into execution delay, motion
delay and communication delay.
A. Simulator Configuration
We use the evaluation framework as indicated by Chen et
al. [2]. However, since some tracker codes are in C++ and
others in Matlab, we adjusted how the delays are calculated
to ensure fairness in execution time evaluation. We used the
chronometer functions based on execution time in C++ 11 to
calculate the time elapsed for processing frames by the trackers
instead of the real-time clock. Some of the program running
time is spent to read the image from disk drive and it was not
included in the execution time. The motion delay is calculated
by the time it takes for the simulated camera to tilt and/or pan.
We decided that there would be no communication delays by
supposing that the camera is not networked.
Some codes are written originally in Matlab and we had
to use the Matlab engine to call the Matlab function, or in
other word, tracker interfaces. Minor changes have been made
to the Matlab source codes such as eliminating the display
and drawing functions since those will affect the speed of
processing but are totally unnecessary and are thus taken as
irrelevant delays. Besides, after practical experiments, it turns
out that calling Matlab engine from C++ is actually in a time
scale of milliseconds and this overhead can be neglected since
the time to process frame by trackers is much longer than that
delay.
B. Performance Evaluation
Chen et al. defined four performance metrics to evaluate the
trackers [2]. These metrics are calculated in the image plane
for the current camera viewpoint (i.e. the viewed subregion
on the image sphere projected on the camera image plane).
Let CGT and CP T be the ground-truth target center and the
predicted target center, and AGT and AP T be the ground-truth
target bounding box and the predicted bounding box, respectively. CF OV is the center of the camera image plane, or in
other words, field of view (FOV). T P E t (Target Point Error)
and BORt (Box Overlapping Ratio) evaluate the quality of
target localization and are defined as
T P E = |CGT − CP T |
move. We experimented with three motion models to obtain
the target motion (∆d ) between two instants:
• Model 1: Object is moving at a constant speed and uses
the velocity of last instant
∆d = V1 × ∆t
•
and
•
BORt =
AGT ∩ AP T
AGT ∪ AP T
(2)
T P Ot (Target Point Offset) and T F t (Track Fragmentation)
evaluate the quality of the camera control and are defined as
t
T P O = |CF OV − CGT |
(3)
and
t
TF =
1, if T P E t is invalid
0,
otherwise
(4)
T F t indicates whether the target is inside the camera FOV.
T P E t and T P Ot are invalid and assigned -1 if the target is
outside the FOV. The overall metrics T P E, T P O, BOR are
the average metrics of the valid tracked frames. T F is the
sum of T F t divided by the number of processed frames. In
the experiments, we report only T F and BOR as they are
the most significant metrics. Besides T P E and BOR have
similar purpose, and so are T P O and T F .
III. TARGET P OSITION P REDICTION
In [2], the authors made the remark that for a tracker to be
successful, it should be either very fast, or should use some
kind of target position prediction to keep the target close to
the FOV of the camera. To assess the practicality of predicting
a target position based on previous track information, we
implemented three target motion models. Since there will be
a delay between frames, it is necessary to predict the object
position on the image sphere and move the camera accordingly
so that the target appears in its FOV. Therefore, the prediction
must account for the motion of the target and the motion of the
camera. For calculation, let a target in the first frame appear
at point P0 (on the spherical image). At next frame, the target
moves to P1 (again on the spherical image). We can calculate
its speed as V0 = P1 − P0 /t1 − t0 . Then in the third frame
if the target moved from P1 to P2 its speed would then be
V1 = P2 − P1 /t2 − t1 . Thus a basic classical mechanics model
can be used to estimate the next position in the fourth frame.
The position prediction should locate the target near the
image center as much as possible. By knowing the motion of
the target it is possible to predict where it will be later after
a delay ∆t . ∆t should account for the processing time of the
current frame in addition to the time it takes for the camera to
Model 2: Object is moving at a constant speed and uses
mean velocity in last two instant.
(V1 + V0 )
× ∆t
2
Model 3: Object can accelerate
∆d =
(1)
(5)
A × ∆t 2
+ V1 × ∆t
2
where A = V1 − V0 /t1 − t0 is the acceleration.
∆d =
(6)
(7)
IV. T ESTED T RACKERS
Among the 19 trackers we tested, six trackers are variations
of correlation filters: KCF, SRDCF, SWCF, DSST, DFST and
sKCF. Two trackers combine correlation filter outputs with
color: STAPLE and STAPLE+. One is based on structured
SVM: STRUCK. Two trackers are based purely on color:
DAT and ASMS. One tracker is based on normalized crosscorrelation: NCC. Two trackers are based on boosting: MIL
and BOOSTING. One tracker is based on optical flow (MEDIANFLOW) and one tracker includes a detector (TLD). Two
trackers can be categorized as part-based: DPCF and CTSE.
Another one combines many trackers in an ensemble: KFEBT. Below, we briefly describe the trackers. More details
can be found in the original papers describing each of them.
1) Kernelized Correlation Filter tracker (KCF) [3] KCF
is operating on HOG features. It localizes target with
the equivalent of a kernel ridge regression trained with
sample patches around the object at different translations.
This version of the KCF tracker includes multi-scale
support, sub-cell peak estimation and a model update by
linear interpolation.
2) Spatially Regularized Discriminative Correlation Filter Tracker (SRDCF) [4] This tracker is derived from
KCF. It introduces a spatial regularization function that
penalizes filter coefficients residing outside the target
area, thus solving the problems arising from assumptions
of periodicity in learning the correlation filters. The size
of the training and detection samples can be increased
without affecting the effective filter size. By selecting the
spatial regularization function to have a sparse discrete
Fourier spectrum, the filter is optimized directly in the
Fourier domain. SRDCF employs also Color Names and
greyscale features.
3) Spatial Windowing for Correlation Filter-Based Visual Tracking (SWCF) [5] This tracker is derived from
KCF. It predicts a spatial window for the observation of
the object so that the correlation output of the correlation
filter as well as the windowed observation are improved.
Moreover, the estimated spatial window of the object
4)
5)
6)
7)
8)
9)
10)
11)
patch indicates the object regions that are useful for
correlation.
Discriminative Scale Space Tracker (DSST) [6] DSST
extends the Minimum Output Sum of Squared Errors
(MOSSE)[7] tracker with robust scale estimation. DSST
also learns a one-dimensional discriminative scale filter
which is used to predict the target size. The intensity
features used in MOSSE [7] tracker are combined with a
pixel-dense representation of HOG features.
Dynamic Feature Selection Tracker (DFST) [8] DFST
is a visual tracking algorithm based on the real-time
selection of locally and temporally discriminative features. DFST provides a significant gain in accuracy and
precision with respect to KCF by the use of a dynamic set
of features. A further improvement is given by making
micro-shifts at the predicted position according the best
template matching.
Scalable Kernel Correlation Filter with Sparse Feature Integration (sKCF) [9] This tracker is derived
from KCF. It introduces an adjustable Gaussian window
function and a keypoint-based model for scale estimation.
It deals with the fixed window size limitation in KCF.
Sum of Template And Pixel-wise LEarners (STAPLE)
[10] STAPLE combines two image patch representations
that are sensitive to complementary factors to learn a
model that is robust to both color changes and deformations. It combines the scores of two models in a dense
window translation search. The scores of the two models
are indicative of their reliability.
An improved STAPLE tracker with multiple feature integration (STAPLE+) STAPLE+ is based on the
STAPLE tracker and improves it by integrating multiple
features. It extracts HOG features from color probability
map to exploit color information better. The final response
map is thus a fusion of scores obtained with different
features.
STRUCtured output tracking with Kernels
(STRUCK) [11] This is a framework for adaptive
visual object tracking. It applies a support vector
machine which is learned online. It introduces a
budgeting mechanism that prevents the unbounded
growth in the number of support vectors that would
otherwise occur during tracking.
Distractor Aware Tracker (DAT) [12] This is a trackingby-detection approach based on appearance. To distinguish the object from the surrounding areas, a discriminative model using color histograms is applied. It adapts
the object representation beforehand so that distractors
are suppressed and the risk of drifting is reduced.
Scale Adaptive Mean Shift (ASMS) [13] This is a
mean-shift tracker [14] optimizing the Hellinger distance
between a template histogram and the target in the image.
The optimization is done by a gradient descent. ASMS
addresses the problem of scale adaptation and scale
estimation. It also introduces two improvements over the
original mean-shift [14] to make the scale estimation
12)
13)
14)
15)
16)
17)
18)
19)
more robust in the presence of background clutter: 1)
a histogram color weighting and 2) a forward-backward
consistency check.
Normalized Cross-Correlation (NCC) [15] NCC follows the basic idea of tracking by searching for the best
match between a static grayscale template and the image
using normalized cross-correlation.
Multiple Instance Learning tracker (MIL) [16] MIL
uses a tracking-by-detection approach with multiple instance learning instead of traditional supervised learning
methods. It shows improved robustness to inaccuracies of
the tracker and to incorrectly labeled training samples.
BOOSTING [17] It is based on MIL[16]. This is a realtime object tracker based on a novel on-line version of the
AdaBoost algorithm. The classifier uses the surrounding
background as negative examples in the update step to
avoid the drifting problem.
MEDIANFLOW [18] This tracker uses optical flow to
match points between frames. The tracking is performed
forward and backward in time and the discrepancies between these two trajectories are measured. The proposed
error enables reliable detection of tracking failures and
selection of reliable trajectories in video sequences.
Tracking-learning-detection(TLD) [19] It combines
both a tracker and a detector. The tracker follows the
object from frame to frame using MEDIANFLOW [18].
The detector localizes the target using all appearances
that have been observed so far and corrects the tracker if
necessary. The learning estimates the detector errors and
updates it to avoid these errors in the future.
Deformable Part-based Tracking by Coupled Global
and Local Correlation Filters (DPCF) [20] This tracker
that is derived from KCF relies on joint interactions
between a global filter and local part filters. The local
filters provide an initial estimate which is used by the
global filter as a reference to determine the final result.
The global filter provides a feedback to the part filters. In
this way, it handles both partial occlusion (with the part
filters) but also scale changes (with the global filter).
Contextual Object Tracker with Structural Encoding
(CTSE)[21] This tracker uses contextual and structural
information (that is specific to a target object) into the
appearance model. This is first achieved by including
features from complementary region having correlated
motion with the target object. Second, a local structure
that represents the spatial constraints between features
within the target object is included. SIFT keypoints are
used as features to encode both these information.
Kalman filter ensemble-based tracker (KF-EBT) This
tracker combines the result of two other trackers:
ASMS[13] using a color histogram and KCF. Using a
Kalman filter, the tracker works in cycles of prediction
and correction. Firstly, a motion model predicts the target
next position. Secondly, the trackers results are fused with
the predicted position and the model is updated.
V. R ESULTS
AND ANALYSIS
In this section, we report our results on the 36 video
sequences of [2]. The video sequences that consist in tracking
persons, faces and objects include difficulties such as motion
blur, scale change, out-of-plane rotation, fast motion, cluttered
background, illumination variation, low resolution, occlusion,
presence of distractors and articulated objects. Results are
reported for the whole dataset.
A. Ranking method
During testing, we discovered that since different trackers
have different tracking speed, using only the four metrics in
section II-B is not enough. For example, some trackers have
T P E, T P O and BOR metrics the are good because they
track slowly in real-time simulation, which means they just
track a few frames correctly and all other frames are invalid
and are ignored for the calculation of the metrics. Under this
circumstance, the tracker succeed in tracking every processed
frames (which are only the first frames). Only T F can capture
to some extent this lack of performance as it verifies if the
target is in the FOV or not. Thus we consider another essential
metric: processed frame ratio (P R):
PR =
FN P
,
FT O
(8)
where FN P is the number of processed frames and FT O is
the total number of frames.
T F contains part of P R information since it stands for
whether the object is inside the FOV in the processed frames.
If P R is low, T F will be high since the tracker will not be able
to track the object correctly in the processed frames because
of the long interval between them. However, a high T F can
be caused also by poor robustness.
After considering the PTZ camera nature and the tracker test
results, we formulated a ranking formula. The formula stands
for the Euclidean distance between the point defined by the
pair (BOR, T F ) obtained by a tracker and the ideal tracker
(top-right point in figure 1). The score is thus:
p
(9)
Score = (1 − BOR)2 + T F 2 .
We selected T F and BOR because we consider that T P E
conveys similar information as BOR and T F conveys similar
information as T P O.
B. Results without processing delays
We have first set the execution ratio to zero in order to
compare different trackers for their performances for in-plane
rotations, out-of-plane rotations and drastic scene changes
caused by the camera motion. We are looking for the most
robust trackers, neglecting their processing times (they are set
to all perform at the same speed). Note that in this experiment,
the camera motion delays are considered as they reflects the
robustness of the trackers. If a tracker performs poorly, this
will cause unnecessary camera motion that will result in drop
frames.
TABLE I
T RACKER PERFORMANCES WITH EXECUTION RATIO SET TO 0 ON THE 36
VIDEO SEQUENCES OF [2]. T RACKERS ARE RANKED BASED ON THEIR
SCORE ( EQUATION 9).
Tracker Names
ASMS
DPCF [MATLAB]
TLD
DAT [MATLAB]
Staple+ [MATLAB]
Staple [MATLAB]
KF-EBT
DSST
STRUCK
SKCF
KCF
SWCF [MATLAB]
MIL
DFST [MATLAB]
BOOSTING
MEDIANFLOW
SRDCF [MATLAB]
NCC
CTSE
Score
0.65
0.71
0.72
0.72
0.72
0.72
0.74
0.78
0.85
0.87
0.88
0.88
0.89
0.90
0.92
0.98
1.01
1.03
1.21
BOR
0.42
0.40
0.56
0.35
0.37
0.36
0.36
0.38
0.31
0.30
0.28
0.31
0.28
0.27
0.26
0.27
0.73
0.24
0.04
TF
0.30
0.38
0.57
0.31
0.35
0.34
0.37
0.48
0.50
0.52
0.51
0.55
0.52
0.53
0.54
0.65
0.97
0.69
0.74
PR
0.92
0.80
0.41
0.77
1.00
1.00
0.98
0.92
0.94
0.91
0.95
0.75
0.85
0.93
0.62
0.95
0.14
0.90
0.44
Table I and Figure 1(a) give the results for the 19 trackers
based on their ranking. In the PTZ camera scenario, the
difficulties with in-plane and out-of-plane rotations will be amplified because of the application dynamic nature. Surprisingly,
trackers which adopted a scale adaptation function such as
SRDCF do not necessarily perform better than other trackers
due in great part to their slow execution speed. ASMS and
DPCF are the best performers in this experiment. In the VOT
2016 benchmark [1] there is no drop frames caused by delays
and less viewpoint changes caused by camera motion. Thus
it is reasonable that there is difference between the ranking
of our framework and that of the VOT ranking. However, our
ranking is still quite similar to that of VOT 2016 benchmark.
Trackers like Staple, Staple+, KCF-EBT and DAT are good
both in VOT benchmark and our benchmark. However, in
our benchmark the performance of ASMS is surprisingly the
best when it just ranked in the middle in VOT. And some
trackers like SRDCF do not behave well in our PTZ framework
probably because they do not output bounding boxes when the
tracking fails and as a result, the PTZ camera is not controlled
correctly. The VOT system will check every five frames to
verify whether the tracker has failed. If it has failed, it is
reinitialized. In our framework, we do not intervene in the
tracking process at all. Early failures are thus more penalized.
C. Results with processing delays
We then set the execution ratio to 1 to track objects. In the
real world the objects will keep moving while the tracker is
processing a frame. As a result, the task of tracking is harder
in this case since the intervals between processed frames are
caused both by the time for processing the frame and the
camera motion delay. P R should decline and the trackers
TABLE II
T RACKER PERFORMANCES WITH EXECUTION RATIO SET TO 1 ON THE 36
VIDEO SEQUENCES OF [2]. T RACKERS ARE RANKED BASED ON THEIR
SCORE ( EQUATION 9).
Tracker Names
ASMS
Staple [MATLAB]
KF-EBT
DAT [MATLAB]
TLD
DSST
KCF
STRUCK
SKCF
BOOSTING
Staple+ [MATLAB]
SRDCF [MATLAB]
DPCF [MATLAB]
DFST [MATLAB]
MEDIANFLOW
SWCF [MATLAB]
NCC
MIL
CTSE
Score
0.64
0.65
0.72
0.75
0.83
0.85
0.85
0.85
0.86
0.89
0.91
0.92
0.93
0.95
0.96
1.01
1.03
1.04
1.22
BOR
0.43
0.40
0.38
0.36
0.70
0.32
0.30
0.30
0.30
0.29
0.34
0.73
0.34
0.31
0.27
0.20
0.24
0.20
0.07
TF
0.29
0.25
0.36
0.39
0.77
0.51
0.49
0.49
0.50
0.53
0.62
0.88
0.65
0.65
0.62
0.61
0.69
0.66
0.79
PR
0.88
0.35
0.71
0.21
0.13
0.48
0.63
0.15
0.90
0.54
0.08
0.03
0.08
0.09
0.85
0.23
0.88
0.15
0.11
should lose targets more easily. Table II and Figure 1(b) give
the results for the 19 trackers.
Compared to the previous case, the ranking of trackers
changes. ASMS still ranks first, but DPCF degrades because
it is very slow and its P R value declines to 0.08 where it
was previously at 0.8. The other trackers relative rankings do
not change much, but the average score of trackers is higher,
which means that their performances are worse because of
the execution delay in the tracking process. Still, compared
to the VOT 2016 benchmark [1], the performance of ASMS
is still surprisingly the best. This means that this tracker is
particularly good for handling viewpoint changes. Finally, performance in VOT are better than for our task because we are
testing online tracking and camera control. The PTZ scenario
requires tracking people with different illumination variation,
different scale and from rapidly changing viewpoints. All of
those reasons make our results unique compared to other
benchmarks.
TABLE III
T RACKER PERFORMANCES WITH EXECUTION RATIO SET TO 1 AND WITH
VARIOUS PREDICTION METHODS ON THE 36 VIDEO SEQUENCES OF [2].
Tracker Names
KCF
KCF
KCF
KCF
BOOSTING
BOOSTING
BOOSTING
BOOSTING
Prediction
None
Model 1
Model 2
Model 3
None
Model 1
Model 2
Model 3
Score
0.85
1.06
0.97
1.05
0.89
1.03
1.10
1.08
BOR
0.30
0.20
0.24
0.24
0.29
0.26
0.17
0.18
TF
0.49
0.69
0.61
0.73
0.53
0.71
0.72
0.70
PR
0.63
0.14
0.15
0.12
0.54
0.12
0.12
0.11
tracker will lose the target. If the target cannot be tracked, its
speed will not be updated leading to an even worse situation
where the camera just rotate more or less randomly. The highspeed trackers are affected the most by wrong prediction. Their
process ratio decline from above 0.8 to below 0.2.
Therefore, we can conclude from this experiment that
although appealing in theory, compensating slow tracking by
a position prediction is not easy to apply in practice. It may
work for objects that are far away and that mostly move in
a plane, but it cannot work for target that are closer and that
are moving toward or away from the camera. In such cases,
the motion of the target cannot be predicted in 2D. For best
results, it is thus preferable to design a fast tracker.
VI. C ONCLUSION
This paper presented a benchmark of recent trackers for the
PTZ tracking scenario. Surprisingly, high-speed trackers, such
as MEDIANFLOW and NCC, do not necessarily behave better
than others. However, since predicting the target position was
shown to be difficult, slow trackers should be avoided for the
PTZ tracking task. The results of our test indicate that the top
performing tracker for the PTZ scenario is the ASMS tracker.
This tracker performed very well in accuracy as well as in robustness in our tests. It is impossible to conclusively determine
whether the performance of ASMS over other trackers come
from its image features or its approach. Nevertheless, results
of top trackers show that features play a significant role in the
final performance.
D. Target Position Prediction
Finally, we tested target position prediction to investigate if
it can help trackers to perform better. Table III gives results
for two trackers. Results are similar for all the other trackers.
The proposed models (see section III) for predicting the next
position of the target are not improving results. This is due
to the fact that the speed of the target is difficult to estimate
because it moves in 3D, but we estimate its motion in 2D.
Therefore, the predicted speed is not very accurate. After
calculating the speed, the framework will use this speed to
predict the target position in the next frame. It may cause
unnecessary large motion by the camera. For example, the
predicted target motion may be too large so the camera, by
rotating to this wrongly predicted position, will add delays
in the tracking process. This adds to the possibility that the
R EFERENCES
[1] M. Kristan, J. Matas, A. Leonardis, M. Felsberg, L. Cehovin, G. Fernández, T. Vojir, G. Hager, G. Nebehay,
and R. Pflugfelder, “The visual object tracking vot2016
challenge results,” in Proceedings of the IEEE international conference on computer vision workshops, 2016,
pp. 1–23.
[2] G. Chen, P.-L. St-Charles, W. Bouachir, G.-A. Bilodeau,
and R. Bergevin, “Reproducible evaluation of pan-tiltzoom tracking,” in Image Processing (ICIP), 2015 IEEE
International Conference on. IEEE, 2015, pp. 2055–
2059.
[3] J. F. Henriques, R. Caseiro, P. Martins, and J. Batista,
“High-speed tracking with kernelized correlation filters,”
process ratio = 0
BOR
0.8
0.6
0.4
0.2
0
1
0.5
0
kcf
boosting
medianflow
mil
dfst
srdcf
stapleplus
colorkcf
dpcf
swcf
dat
asms
tld
staple
struck
dsst
skcf
ncc
CTSE
TF
(a) Without processing delay
1
process ratio = 1
0.8
BOR
1
0.6
0.4
0.2
0
1
0.5
0
TF
(b) With processing delay
Fig. 1. BOR vs T F scatter plot without (left) and with processing delays (right) on the 36 video sequences of [2]. The legend that indicates the symbols
representing the trackers is the same for both plots. Ideal results are BOR = 1 and T F = 0.
[4]
[5]
[6]
[7]
[8]
[9]
[10]
[11]
IEEE Transactions on Pattern Analysis and Machine
Intelligence, vol. 37, no. 3, pp. 583–596, 2015.
M. Danelljan, G. Hager, F. Shahbaz Khan, and M. Felsberg, “Learning spatially regularized correlation filters
for visual tracking,” in Proceedings of the IEEE International Conference on Computer Vision, 2015, pp. 4310–
4318.
E. Gundogdu and A. A. Alatan, “Spatial windowing
for correlation filter based visual tracking,” in Image
Processing (ICIP), 2016 IEEE International Conference
on. IEEE, 2016, pp. 1684–1688.
M. Danelljan, G. Häger, F. Khan, and M. Felsberg,
“Accurate scale estimation for robust visual tracking,” in
British Machine Vision Conference, Nottingham, September 1-5, 2014. BMVA Press, 2014.
D. S. Bolme, J. R. Beveridge, B. A. Draper, and Y. M.
Lui, “Visual object tracking using adaptive correlation
filters,” in Computer Vision and Pattern Recognition
(CVPR), 2010 IEEE Conference on. IEEE, 2010, pp.
2544–2550.
G. Roffo and S. Melzi, “Online feature selection for
visual tracking.” in BMVC, 2016.
A. S. Montero, J. Lang, and R. Laganiere, “Scalable
kernel correlation filter with sparse feature integration,”
in Computer Vision Workshop (ICCVW), 2015 IEEE
International Conference on. IEEE, 2015, pp. 587–594.
L. Bertinetto, J. Valmadre, S. Golodetz, O. Miksik, and
P. H. Torr, “Staple: Complementary learners for realtime tracking,” in Proceedings of the IEEE Conference
on Computer Vision and Pattern Recognition, 2016, pp.
1401–1409.
S. Hare, S. Golodetz, A. Saffari, V. Vineet, M.-M. Cheng,
S. L. Hicks, and P. H. Torr, “Struck: Structured output
tracking with kernels,” IEEE transactions on pattern
analysis and machine intelligence, vol. 38, no. 10, pp.
2096–2109, 2016.
[12] H. Possegger, T. Mauthner, and H. Bischof, “In defense
of color-based model-free tracking,” in Proceedings of
the IEEE Conference on Computer Vision and Pattern
Recognition, 2015, pp. 2113–2120.
[13] T. Vojir, J. Noskova, and J. Matas, “Robust scale-adaptive
mean-shift for tracking,” Pattern Recognition Letters,
vol. 49, pp. 250–258, 2014.
[14] D. Comaniciu, V. Ramesh, and P. Meer, “Real-time tracking of non-rigid objects using mean shift,” in Computer
Vision and Pattern Recognition, 2000. Proceedings. IEEE
Conference on, vol. 2. IEEE, 2000, pp. 142–149.
[15] J. P. Lewis, “Fast normalized cross-correlation,” in Vision
interface, vol. 10, no. 1, 1995, pp. 120–123.
[16] B. Babenko, M.-H. Yang, and S. Belongie, “Robust
object tracking with online multiple instance learning,”
IEEE transactions on pattern analysis and machine intelligence, vol. 33, no. 8, pp. 1619–1632, 2011.
[17] H. Grabner, M. Grabner, and H. Bischof, “Real-time
tracking via on-line boosting.”
[18] Z. Kalal, K. Mikolajczyk, and J. Matas, “Forwardbackward error: Automatic detection of tracking failures,” in Pattern recognition (ICPR), 2010 20th international conference on. IEEE, 2010, pp. 2756–2759.
[19] ——, “Tracking-learning-detection,” IEEE transactions
on pattern analysis and machine intelligence, vol. 34,
no. 7, pp. 1409–1422, 2012.
[20] O. Akin, E. Erdem, A. Erdem, and K. Mikolajczyk, “Deformable part-based tracking by coupled global and local
correlation filters,” Journal of Visual Communication and
Image Representation, vol. 38, pp. 763–774, 2016.
[21] T. Chakravorty, G.-A. Bilodeau, and E. Granger, “Contextual object tracker with structure encoding,” in Image
Processing (ICIP), 2015 IEEE International Conference
on. IEEE, 2015, pp. 4937–4941.
| 1cs.CV
|
Language-Based Image Editing with Recurrent attentive Models
arXiv:1711.06288v1 [cs.CV] 16 Nov 2017
Jianbo Chen∗ , Yelong Shen† , Jianfeng Gao† , Jingjing Liu† , Xiaodong Liu†
University of California, Berkeley∗ and Microsoft Research†
[email protected]
yeshen, jfgao, jingjl, [email protected]
Abstract
We investigate the problem of Language-Based Image Editing (LBIE) in this work. Given a source
image and a natural language description, we want to generate a target image by editing the source image based on the description. We propose a generic modeling framework for two sub-tasks of LBIE:
language-based image segmentation and image colorization. The framework uses recurrent attentive
models to fuse image and language features. Instead of using a fixed step size, we introduce for each region of the image a termination gate to dynamically determine in each inference step whether to continue
extrapolating additional information from the textual description. The effectiveness of the framework has
been validated on three datasets. First, we introduce a synthetic dataset, called CoSaL, to evaluate the
end-to-end performance of our LBIE system. Second, we show that the framework leads to state-of-theart performance on image segmentation on the ReferIt dataset. Third, we present the first language-based
colorization result on the Oxford-102 Flowers dataset, laying the foundation for future research.
1
Introduction
In this work, we aim to develop an automatic Language-Based Image Editing (LBIE) system. Given a
source image, which can be a sketch, a grayscale image or a natural image, the system will automatically
generate a target image by editing the source image following natural language instructions provided by
users. Such a system has a wide range of applications from Computer-Aided Design (CAD) to Virtual
Reality (VR). As illustrated in Figure 1, we can envision that a fashion designer presents a sketch of a pair
of new shoes (i.e., the source image) to a customer, who can provide modifications on the style and color
in verbal description, which can then be taken by the LBIE system to change the original design. The final
output (i.e., the target image) is the revised and enriched design that would meet the customers requirement.
Figure 2 showcases the use of LBIE for VR. While most VR systems are still using button-controlled or
touchscreen interface, LBIE provides a natural user interface for future VR systems, where users can easily
modify the base environment via natural language instructions.
LBIE can cover a broad range of tasks in image generation: shape, color, size, texture, position, etc.
This paper focuses on the two basic sub-tasks of LBIE: language-based segmentation and colorization for
shapes and colors. For example, in Figure 3, given the grayscale image and the expression “The flower
has red petals with yellow stigmas in the middle”, the segmentation model will identify each region of the
image as “petals”, “stigmas”, and the colorization model will paint each pixel with the suggested color. In
this intertwined task of segmentation and colorization, the distribution of target images can be multi-modal
in the sense that each pixel will have a definitive ground truth on segmentation, but not necessarily on color.
For example, the pixels on petals in Figure 3 should be red based on the textual description, but the specific
numeric values of the red color in RGB space is not uniquely specified. The system is required to colorize
the petals based on real-world knowledge. Another uncertainty lies in the fact that the input description
1
Figure 1: In an interactive design interface, a sketch of shoes is presented to a customer, who then gives
a verbal instruction on how to modify the design: “The insole of the shoes should be brown. The vamp
and the heel should be purple and shining”. The system will colorize the sketch following the customer’s
instruction. (images from [11]).
Figure 2: The image on the left is an initial virtual environment. The user provides a textual description:
“The afternoon light flooded the little room from the window, shining the ground in front of a brown bookshelf made of wood. Besides the bookshelf lies a sofa with light-colored cushions. There is a blue carpet in
front of the sofa, and a clock with dark contours above it...”. The system then modifies the virtual environment into the right image.
might not cover every detail of the image. Undescribed regions such as the leaves in the given example,
should be rendered based on past memory. In summary, the ultimate goal is to generate images guided by
the constraint of natural language expressions, but should also align with common sense of the real world.
Language-based image segmentation has been studied previously in [9]. However, our task is far more
challenging because the textual description in our task often contains multiple sentences (as in Figure 2),
while in [9] most of the expressions are simple phrases. To the best of our knowledge, language-based
colorization has not been studied systematically before. In most previous work, images are generated either
solely with natural language expressions [20],[30] or based on another image [11],[3],[31]. Instead, we want
to generate a target image based on both natural language expressions and source image. Related tasks will
be discussed in detail in Section 2.
A unique challenge in language-based image editing is the complexity of natural language expressions
and their correlation with the source images. Like the example shown in Figure 2, the description usually
consists of multiple sentences. Each sentence might refer to multiple objects in the source image. Many
details also need to be inferred from multiple sentences. When human edits the source image based on a
textual description, we often keep in mind which sentences are related to which region/object in the image,
and go back to the sentences multiple times while editing that region. This behavior of “going back” often
varies from region to region, depending on the complexity of the description for that region. An investigation
of this problem is carried out on CoSaL, a dataset introduced for the purpose, detailed in Section 4.
2
Figure 3: Left: sketch image. Middle: grayscale image. Right: color image (from [17]). A language-based
image editing system will take either of the first two images as the input, and generate the third color image
following a natural language expression: “The flower has red petals with yellow stigmas in the middle”,.
Our goal is to design a generic framework for the two sub-tasks in language-based image editing. A
diagram of the model is shown in Figure 4. Inspired by the observation aforementioned, we introduce a
recurrent attentive fusion module in our framework. The fusion module takes in image features encoding
the source image via a convolutional neural network, and textual features encoding the natural language
expression via an LSTM, and outputs the fused features to be upsampled by a deconvolutional network
into target images. In the fusion module, recurrent attentive models are employed to extract distinct textual
features based on the spatial features from different parts of an image. And a termination gate is introduced
for each region to control the number of steps it interacts with the textual features. A Gumbel-Softmax
reparametrization trick [12] is used for end-to-end training of the entire network. Details of the models and
the training process are described in Section 3.
Our contributions are summarized as follows:
• We define a new task of language-based image editing (LBIE).
• We present a generic modeling framework based on recurrent attentive models for two sub-tasks of
LBIE: language-based image segmentation and colorization.
• We introduce a synthetic dataset CoSaL to evaluate the end-to-end performance of the LBIE system.
• We achieve new state-of-the-art performance on language-based image segmentation on the ReferIt
dataset.
• We present the first language-based colorization result on the Oxford-102 Flowers dataset, laying the
foundation for future research.
2
Related Work
While the task of language-based image editing has not been studied, the community has taken significant
steps in several related areas, including Language Based object detection and Segmentation (LBS) [9],[10],
Image-to-Image Translation (IIT) [11], Generating Images from Text (GIT) [19], [30], Image Captioning
(IC) [13], [24], [28], Visual Question Answering (VQA) [2], [29], Machine Reading Comprehension (MRC)
[8], etc. We summarize the types of inputs and outputs for these related tasks in Table 1.
3
Recurrent Attentive-Fusion Module
Termination Gate 𝜏 2
Termination Gate 𝜏 1
Spatial Feature Map 𝑆 0 =V
Decoder
Spatial Feature Map 𝑆 1
Fusion
Spatial Feature Map 𝑆 𝑇
Output
𝑂
Fusion
Image Encoder
Attention
Attention
Attentive Feature Map 𝑈
Attentive Feature Map 𝑈1
0
Language Feature Vectors 𝑈
The flower has red petals with
yellow stigmas in the middle
Language Encoder
Figure 4: A high-level diagram of our model, composed of a convolutional image encoder, an LSTM text
encoder, a fusion module, a deconvolutional upsampling layer, with an optional convolutional discriminator.
MRC
VQA
IIT
IC
GIT
LBS
LBIE
Inputs
Text Image
YES
NO
YES YES
NO
YES
NO
YES
NO
YES
YES YES
YES YES
Outputs
Text Image
YES
NO
YES
NO
NO
YES
YES
NO
NO
YES
NO
YES
NO
YES
Table 1: The types of inputs and outputs for related tasks
Recurrent attentive models
Recurrent attentive models have been applied to visual question answering (VQA) to fuse language and
image features [29]. The stacked attention network proposed in [29] identifies the image regions that are
relevant to the question via multiple attention layers, which can progressively filter out noises and pinpoint
the regions relevant to the answer. In image generation, a sequential variational auto-encoder framework,
such as DRAW[7], has shown substantial improvement over standard variational auto-encoders (VAE) [15].
Similar ideas have also been explored for machine reading comprehension, where models can take multiple
iterations to infer an answer based on the given query and document [4], [26], [25], [27]. In [22] and [21],
a novel neural network architecture called ReasoNet is proposed for reading comprehension. ReasoNet
4
performs multi-step inference where the number of steps is determined by a termination gate according to
the difficulty of the problem. ReasoNet is trained using policy gradient methods.
Segmentation from language expressions
The task of language-based image segmentation is first proposed in [9]. Given an image and a natural language description, the system will identify the regions of the image that correspond to the visual entities
described in the text. The authors in [9] proposed an end-to-end approach which is composed of three
main components: a convolutional network to encode source images, an LSTM network to encode natural language descriptions, and a fully convolutional classification and upsampling network for pixel-wise
segmentation.
One of the key differences between their approach and ours is the way of integrating image and text
features. In [9], for each region in the image, the extracted spatial features are concatenated with the same
textual features. Inspired by the alignment model of [13], in our approach, each spatial feature is aligned
with different textual features based on attention models. Our approach yields superior segmentation results
than that of [9] on a benchmark dataset.
Conditional GANs in image generation
Generative adversarial networks (GANs) [6] have been widely used as a powerful tool for image generation.
Conditional GANs [16] are often employed when there are constraints that a generated image needs to
satisfy. For example, deep convolutional conditional GANs [18] have been used to synthesize images based
on textual descriptions [20] [30]. [11] proposed the use of conditional GANs for image-to-image translation.
Different from these tasks, LBIE takes both image and text as input, presenting an additional challenge of
fusing the features of the source image and the textual description.
3
The Framework
Overview The proposed modeling framework, as shown in 4, is based on neural networks, and is generic
to both the language-based image segmentation and colorization tasks. The framework is composed of a
convolutional image encoder, an LSTM text encoder, a fusion network that generates a fusion feature map
by integrating image and text features, a deconvolutional network that generates pixel-wise outputs (the
target image) by upsampling the fusion feature map, and an optional convolutional discriminator used for
training colorization models.
Image encoder The image encoder is a multi-layer convolutional neural network (CNN). Given a source
image of size H × W , the CNN encoder produces a M × N spatial feature map, with each position on the
feature map containing a D-dimensional feature vector (D channels), V = {vi : i = 1, . . . , M × N }, vi ∈
RD .
Language encoder The language encoder is a recurrent Long Short-Term Memory (LSTM) network.
Given a natural language expression of length L, we first embed each word into a vector through a word
embedding matrix, then use LSTM to produce for each word a contextual vector that encodes its contextual
information such as word order and word-word dependencies. The resulting language feature map is U =
{ui : i = 1, . . . , L}, ui ∈ RK .
5
Recurrent attentive fusion module The fusion network fuses text information in U into the M ×N image
feature map V , and outputs an M × N fusion feature map, with each position (image region) containing an
editing feature vector, O = {oi : i = 1, . . . , M × N }, oi ∈ RD .
The fusion network is devised to mimic the human image editing process. For each region in the source
image vi , the fusion network reads the language feature map U repeatedly with attention on different parts
each time until enough editing information is collected to generate the target image region. The number of
steps varies from region to region.
Internal state The internal state at time step t is denoted as S t = {sti , i = 1, . . . , M × N }, which
is a spatial feature map, with each poition (image region) containing a vector representation of the editing
information state. The initial state is the spatial feature map from the source image, S 0 = V . The sequence
of internal states is modeled by Convolutional Gated Recurrent Units (C-GRUs) which will be described
below.
Attention The attention at time step t is denoted as Û t = {ûti , i = 1, . . . , M × N }, which is a spatial
feature map generated based on the current internal state S t and the language feature map U :
Û t = Attention(U, S t ; θa )
where Attention(.) is implemented as follows
T
βij ∝ exp{sti W uj }
ûti =
L
X
βij uj .
j=1
C-GRUs C-GRUs update the current internal state S t by infusing the attention feature map Û t :
S t+1 = C-GRUs(S t , Û t ; θc ).
where C-GRUs(.) is implemented as follows
z = σ(W1 ⊗ S t + W2 ⊗ Û t + b1 ),
r = σ(W3 ⊗ S t + W4 ⊗ Û t + b2 ),
c = ReLU(W5 ⊗ (r
Ôt = h = (1 − z)
S t ) + W6 ⊗ Û t + b),
St + z
c,
S t+1 = W7 ⊗ h,
where is the elementwise-product, and ⊗ is the convolutional operator. Note that Ôt is the intermediate
output of the fusion feature map at time step t.
6
Termination gates There are M × N termination gates, each for one image region vi in V . Each
termination gate generates a binary random variable according to the current internal state of its image
region: τit ∼ p(·ftg (sti ; θtg )). If τit = 1, the fusion process for the image region vi stops at t, and the editing
feature vector for this image region is set as oi = ôti . When all terminate gates are true, the fusion process
for the entire image is completed, and the fustion network outputs the fusion feature map O.
We define ζ = (ζ1 , ζ2 , . . . , ζM ×N ), where ζi = (τi1 , τi2 , . . . , τiT ), a categorical distribution with p(ζi =
et ) = βit , where
Y
βit = ftg (sti ; θtg ) (1 − ftg (ski ; θtg )).
k<t
the probability of ith feature map stopping at time t.
Inference Algorithm 1 describes the stochastic inference process of the fusion network. The state
sequence S (1:T ) is hidden and dynamic, chained through attention and C-GRU in a recurrent fashion. The
fusion network outputs for each image region vi an editing feature vector oi at the ti -th step, where ti is
controlled by the ith termination gate, which varies from region to region.
Algorithm 1 Stochastic Inference of the Fusion Network
Require: V ∈ RD×(M ×N ) : Spatial feature map of image.
Require: U ∈ RK×L : Language feature map of expression.
Ensure: Fusion feature map O ∈ RD×(M ×N ) .
function F USION(V, U )
Initialize S 0 = V .
for all t = 0 to tmax − 1 do
Û t = Attention(U, S t ; θa )
S t+1 , Ôt = C-GRUs(S t , Û t ; θc )
Sample τ t+1 ∼ p(·|ftg (S t+1 ; θtg ))
if τit+1 = 1 and τis = 0 for s ≤ t then
Set Oi = Ôit+1 .
end if
end for
for all i = 1 to M × N do
if τi = 0 then
Set oi = ôitmax −1
end if
end for
end function
Image decoder The image decoder is a multi-layer deconvolutional network. It takes as input the M × N
fusion feature map O produced by the fusion module, and unsamples from O to produce a H × W × De
editing map E of the same size as the target image, where De is the number of classes in segmentation and
2 (ab channels) in colorization.
Discriminator The discriminator Dφ (E) takes in a generated image and outputs the probability of the
image being realistic. The structure of the discriminator follows that in [20], which uses a convolutional
7
neural network to extract features from the image. In our model, the discriminator serves the function of
judging whether an image looks natural, but not whether it is compatible with language descriptions. The
latter constraint is left to an L1 loss to be described later.
Loss and training Denote the loss as L(θ) = Eζ [l(E(ζζ , θ), Y )], where the expectation is taken over the
categorical variables ζ generated by termination gates, and lθ (ζζ ) = l(E(ζζ , θ), Y ) is the loss of output at ζ .
Because the sample space is of exponential size T M ×N , it is intractable to sum over the entire sample space.
We denote the density of ζ by pθ (ζζ ). A naive approach to approximation is to subsample the loss and update
parameters via the gradient of Monte Carlo estimate of loss:
∇θ L(θ) = ∇θ Eζ [lθ (ζζ )]
= ∇θ Eζ [lθ (ζζ )]
X
pθ (ζζ ) lθ (ζζ )∇θ log pθ (ζζ ) + ∇θ lθ (ζζ )
=
ζ
≈
1 X
lθ (ζζ )∇θ log pθ (ζζ ) + ∇θ lθ (ζζ ),
|S|
ζ ∈S
where S is a subset of ζ sampled from the distributon pθ (ζζ ). In experiments, we found the above Monte
Carlo estimate suffers from high variance. To resolve this issue, we employ the Gumbel-Softmax reparameterization trick [12], which replaces every ζi ∈ {0, 1}T sampled from Cat(β1 , β2 , . . . , βT ) by another
random variable zi generated from Gumbel-softmax distribution:
exp((log βit + εti )/λ)
zit = PT
,
t
t
k=1 exp((log βi + εi )/λ)
where λ is a temperature annealed via a fixed schedule and the auxiliary random variables ε1i , . . . , εTi are
i.i.d. samples drawn from Gumbel(0, 1) independent of the parameters βi :
εti = − log(− log uti ), uti ∼ Unif(0, 1).
Define z(εε, θ) = (z1 , z2 , . . . , zM N ). The loss can be rewritten as L(θ) = Eε [lθ (zz (εε, θ))], and the update is
approximated by taking the gradient of Monte Carlo estimates of the loss obtained from sampling ε .
We use two different losses for segmentation and colorization respectively.
Segmentation In segmentation, we assume there is a unique answer for each pixel on whether or not it
is being referred in the stage of segmentation. The response map E is of size H ×W ×De , which produces a
log probability for each class for each pixel. We use a pixel-wise softmax cross-entropy loss during training:
l(E, Y ) = Cross-Entropy(Softmax(E), Y ).
Colorization In colorization, the high-level goal is to generate realistic images under the constraint of
natural language expressions and input scene representations, we introduce a mixture of GAN loss and L1
loss for optimization as in [11]. A discriminator Dφ parametrized by φ is introduced for constructing the
GAN loss.
8
The response map E is the predicted ab color channels. It is combined with the grayscale source image
to produce a generated color image E 0 . The generator loss is a GAN loss taking E 0 as input, and L1 loss
between the ab channels of the target image Y and the response map E:
l(E, Y ) = log(1 − Dφ (E)) + γkE − Y k1 (γ = 0.01).
The discriminator Dφ is trained by first generating a sample E via Algorithm 1, combined with the
grayscale source image to produce E 0 , and optimize the following loss over φ:
log(Dφ (E 0 )) + log(1 − Dφ (Y )).
The generator loss and the discriminator loss are optimized alternatively in the training stage.
4
Experiments
We conducted three experiments to validate the performance of the proposed framework. A new dataset
“CoSaL” was introduced to test the capability of understanding multi-sentence descriptions and associating
the inferred textual features with visual features. Our framework also yielded state-of-the-art performance
on the benchmark dataset ReferIt [14] for image segmentation. A third experiment was carried out on the
Oxford-102 Flowers dataset [17], for the language-based colorization task.
4.1
Experiments on CoSaL
Dataset A synthetic dataset CoSAL - Colorizing Shapes with Artificial Language is introduced for studying the correlation between images and multi-sentence descriptions. Each image in the dataset consists of
nine shapes, paired with a textual description of the image. The goal of the task is defined as: given a blackwhite image and its corresponding description, train a model that can automatically colorize the nine shapes
following the textual description. Figure 5 shows an example of this task, which requires sophisticated
coreference resolution, multi-step inference and logical reasoning.
The dataset was created as follows: first, we divide a white-background image into 3 × 3 regions.
Each region contains a shape randomly sampled from a set of S shapes (e.g., squares, fat rectangles, tall
rectangles, circles, fat ellipses, tall ellipses, diamonds, etc.) Each shape is then filled with one of C color
choices, chosen at random. The position and the size of each shape are generated by uniform random
variables. As illustrated in Figure 5, the difficulty of this task increases with the number of color choices. In
our experiments, we specify C = 3 as a start point.
The descriptive sentences for each image can be divided into two categories: direct descriptions and
relational descriptions. The former prescribes the color of a certain shape (e.g., Diamond is red), and the
latter depicts one shape conditional of another (e.g., The shape left to Diamond is blue). To understand direct
descriptions, the model needs to associate a specified shape with its textual features. Relational description
adds another degree of difficulty, which calls for advanced inference capability of relational/multi-step reasoning. The ratio of direct descriptions to relational descriptions varies among different images, and all the
colors and shapes in each image are uniquely determined by the description. In our experiment, we randomly generated 50, 000 images with corresponding descriptions for training purpose, and 10, 000 images
with descriptions for testing.
9
Metric For this task, we use average IOU over nine shapes and the background as the evaluation metric.
Specifically, for each region, we compute the intersection-over-union (IOU), which is the ratio of the total
intersection area to the total union area of predicted colors and ground truth colors. We also compute the
IOU for the background (white) of each image. The IOU for 10 classes (9 shapes + 1 background) are
computed over the entire test set and then averaged.
Model Implementation A six-layer convolutional network is implemented as the image feature extractor.
Each layer has a 3 × 3 kernel with stride 1 and output dimension 4, 4, 8, 8, 16, 16. ReLU is used for nonlinearity after each layer, and a max-pooling layer with a kernel of size 2 is inserted after every two layers. Each
sentence in the textual description is encoded with bidirectional LSTMs that share parameters. The LSTMs
have 16 units. In the fusion network, the attention model has 16 units, the GRU cells use 16 units, and the
termination gate uses a linear map on top of the hidden state of each GRU cell. Two convolutional layers of
kernel size 1 × 1 with the output dimension of 16, 7 are put on top of the fused features as a classifier. Then
an upsampling layer is implemented on top of it, with a single-layer deconvolutional network of kernel size
16, stride 8 to upsample the classifier to the original resolution. The upsampling layer is initialized with
bilinear transforms. The maximum of T of termination steps vary from 1to4. When T = 1, the model is
reduced to simply concatenating features extracted from the convolutional network with the last vector from
LSTM.
Results Table 2 shows the average IoU of two models, with T = 1 and T = 4, respectively. Results
show that the model with attention and T = 4 achieves a better performance when there are more relational
descriptions in the dataset. When there are more direct descriptions, the two models achieve similar performance. This demonstrates the framework’s capability of interpreting multiple sentences and associating
them with the source image.
T
Attention
1
1
4
No
Yes
Yes
Number of direct descriptions
4
6
8
0.2107
0.4030
0.5033
0.2499
0.5220
0.5313
0.3186
0.7097
0.7017
Table 2: The average IoU of two models, without attention at T = 1 and with attention at T = 1, 4.
Performance varies among datasets with different ratios of direct to relational descriptions.
Figure 5 illustrates how the model with T = 3 interprets the nine sentences during each inference step.
Sentences in red are attended to in the first step. Those in yellow and green are attended in the next two
consecutive steps. One observation from the experiment was that the model first extracted information from
direct descriptions, and then focused on relational descriptions for additional reasoning.
4.2
Experiments on ReferIt
Dataset The ReferIt dataset is composed of 19, 894 photographs of real world scenes, along with 130, 525
natural language descriptions on 96, 654 distinct objects in those photographs [14]. The dataset contains
238 different object categories, including animals, people, buildings, objects and background elements (e.g.,
grass, sky). Both training and development datasets include 10, 000 images.
10
Figure 5: Right: ground truth image. Left: illustration of which sentences are attended to at each time step.
Red, yellow and green represent the first, second and third time step, respectively.
Model
[email protected]
[email protected]
[email protected]
[email protected]
[email protected]
IoU
SCRC bbox [10]
GroundeR bbox [5]
Hu, etc.[9]
Our model
9.73%
11.08%
34.02%
32.53%
4.43%
6.20%
26.71%
27.9%
1.51%
2.74%
19.32%
18.76%
0.27%
0.78%
11.63%
12.37%
0.03%
0.20%
3.92%
4.37%
21.72%
20.50%
48.03%
50.09%
Table 3: The results of previous models and our model on the ReferIt dataset.
Metric Following [9], we use two metrics for evaluation: 1) overall intersection-over-union (overall IoU)
of the predicted and ground truth of each region, averaged over the entire test set; 2) precision@threshold,
the percentage of test data whose (per image) IoU between prediction and ground truth is above the threshold. Thresholds are set to 0.5, 0.6, 0.7, 0.8, 0.9.
Model Implementation A VGG-16 architecture [23] is used as the image encoder for images of size 512×
512. Textual descriptions are encoded with an LSTM of 1, 024 units. In the fusion network, the attention
model uses 512 units and the GRU cells 1, 024 units, on top of which is a classifier and an upsampling layer
similar to the implementation in Section 4.1. The maximum number of inference steps is 3. ReLU is used
on top of each convolutional layer. L2-normalization is applied to the parameters of the network.
Results Table 3 shows the experimental results of our model and the baseline methods on the ReferIt
dataset. We see that our framework yields a better IoU and precision than [9]. Our hypothesis is that
the outperformance resulted from the unique attention mechanism used by our fusion network, which can
efficiently associate individual descriptive sentences with different regions of the source image. There is not
much discrepancy between the two models with T = 1 and T = 3, probably due to the fact that most textual
descriptions in this dataset are simple. Examples of semantic labels are provided in supplementary materials
11
Figure 6: First row: original images. Second row: results from the image-to-image translation model in
[11], without text input. Third row: results from our model, taking textual descriptions into account.
Figure 7: First row: original images. Second, third and fourth rows: results generated from our framework
with arbitrary text input: “The petals of the flower are red/blue/yellow”.
4.3
Experiments on Oxford-102 Flower Dataset
Dataset In this experiment, we use the Oxford-102 Flowers dataset [17], which contains 8, 189 images
from 102 flower categories. Each image has five textual descriptions [20]. Following [20], [19] and [1],
we split the dataset into 82 classes for training and 20 classes for testing. Given a grayscale image of a
flower and a description of the shapes and colors of the flower, the goal is to colorize the image following
the description.
Model Implementation A fifteen-layer convolutional network similar to [31] is used for encoding 64×64
images. Textual descriptions are encoded with an LSTM of 512 units. In the fusion network, the attention
model uses 128 units and the GRU cells 128 units. The image encoder is composed of 2 deconvolutional
layers, each followed by 2 convolutional layers, to upsample the fusion feature map to the target image space
of 64×64×2. The maximum length of the spatial RNN is 1. The discriminator is composed of five layers of
convolutional networks of stride 2, with the output dimension 256, 128, 64, 32, 31. The discriminator score
is the average of the final output. ReLU is used for nonlinearity following each convolutional layer, except
for the last one which uses the sigmoid function.
12
Results Due to the lack of available models for the task, we compare our framework with a previous model
developed for image-to-image translation as baseline, which colorizes images without the injection of texts.
The colorization results on 10 randomly-sampled images from the test set are shown in Figure 6. As we can
see, without text input, the baseline approach often colorizes images with the same color (in this dataset,
most images are painted with blue and white), while our framework can generate flowers similar to their
original colors which are specified in texts. Figure 7 also provides some examples of images generated with
arbitrary text input using the trained model.
5
Conclusion and Future Work
In this paper we introduce the problem of Language-Based Image Editing (LBIE), and propose a generic
modeling framework for two sub-tasks of LBIE: language-based image segmentation and colorization. At
the heart of the proposed framework is a fusion module that uses recurrent attentive models to dynamically
decide, for each region of an image, whether to continue the text-to-image fusion process. Our models have
demonstrated superior empirical results on three datasets, including the ReferIt dataset for image segmentation, the Oxford-102 Flower dataset for colorization, and a proposed CoSaL dataset for evaluating the
end-to-end performance of the LBIE system. In future, we will extend the framework to other image editing
subtasks. Another interesting area is to build a dialogue-based image editing system that allows users to edit
images in a more interactive and natural way.
References
[1] Z. Akata, S. Reed, D. Walter, H. Lee, and B. Schiele. Evaluation of output embeddings for fine-grained image
classification. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages
2927–2936, 2015. 12
[2] S. Antol, A. Agrawal, J. Lu, M. Mitchell, D. Batra, C. Lawrence Zitnick, and D. Parikh. Vqa: Visual question
answering. In Proceedings of the IEEE International Conference on Computer Vision, pages 2425–2433, 2015.
3
[3] Z. Cheng, Q. Yang, and B. Sheng. Deep colorization. In Proceedings of the IEEE International Conference on
Computer Vision, pages 415–423, 2015. 2
[4] B. Dhingra, H. Liu, Z. Yang, W. W. Cohen, and R. Salakhutdinov. Gated-attention readers for text comprehension. arXiv preprint arXiv:1606.01549, 2016. 4
[5] J. Donahue, L. Anne Hendricks, S. Guadarrama, M. Rohrbach, S. Venugopalan, K. Saenko, and T. Darrell.
Long-term recurrent convolutional networks for visual recognition and description. In Proceedings of the IEEE
conference on computer vision and pattern recognition, pages 2625–2634, 2015. 11
[6] I. Goodfellow, J. Pouget-Abadie, M. Mirza, B. Xu, D. Warde-Farley, S. Ozair, A. Courville, and Y. Bengio.
Generative adversarial nets. In Advances in neural information processing systems, pages 2672–2680, 2014. 5
[7] K. Gregor, I. Danihelka, A. Graves, D. J. Rezende, and D. Wierstra. Draw: A recurrent neural network for image
generation. arXiv preprint arXiv:1502.04623, 2015. 4
[8] K. M. Hermann, T. Kocisky, E. Grefenstette, L. Espeholt, W. Kay, M. Suleyman, and P. Blunsom. Teaching
machines to read and comprehend. In Advances in Neural Information Processing Systems, pages 1693–1701,
2015. 3
[9] R. Hu, M. Rohrbach, and T. Darrell. Segmentation from natural language expressions. In European Conference
on Computer Vision, pages 108–124. Springer, 2016. 2, 3, 5, 11
[10] R. Hu, H. Xu, M. Rohrbach, J. Feng, K. Saenko, and T. Darrell. Natural language object retrieval. In Proceedings
of the IEEE Conference on Computer Vision and Pattern Recognition, pages 4555–4564, 2016. 3, 11
13
[11] P. Isola, J.-Y. Zhu, T. Zhou, and A. A. Efros. Image-to-image translation with conditional adversarial networks.
arXiv preprint arXiv:1611.07004, 2016. 2, 3, 5, 8, 12
[12] E. Jang, S. Gu, and B. Poole. Categorical reparameterization with gumbel-softmax. arXiv preprint
arXiv:1611.01144, 2016. 3, 8
[13] A. Karpathy and L. Fei-Fei. Deep visual-semantic alignments for generating image descriptions. In Proceedings
of the IEEE Conference on Computer Vision and Pattern Recognition, pages 3128–3137, 2015. 3, 5
[14] S. Kazemzadeh, V. Ordonez, M. Matten, and T. L. Berg. Referit game: Referring to objects in photographs of
natural scenes. In EMNLP, 2014. 9, 10
[15] D. P. Kingma and M. Welling. Auto-encoding variational bayes. arXiv preprint arXiv:1312.6114, 2013. 4
[16] M. Mirza and S. Osindero. Conditional generative adversarial nets. arXiv preprint arXiv:1411.1784, 2014. 5
[17] M.-E. Nilsback and A. Zisserman. Automated flower classification over a large number of classes. In Proceedings
of the Indian Conference on Computer Vision, Graphics and Image Processing, Dec 2008. 3, 9, 12
[18] A. Radford, L. Metz, and S. Chintala. Unsupervised representation learning with deep convolutional generative
adversarial networks. arXiv preprint arXiv:1511.06434, 2015. 5
[19] S. Reed, Z. Akata, H. Lee, and B. Schiele. Learning deep representations of fine-grained visual descriptions. In
Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 49–58, 2016. 3, 12
[20] S. Reed, Z. Akata, X. Yan, L. Logeswaran, B. Schiele, and H. Lee. Generative adversarial text to image synthesis.
In International Conference on Machine Learning, pages 1060–1069, 2016. 2, 5, 7, 12
[21] Y. Shen, P.-S. Huang, M.-W. Chang, and J. Gao. Implicit reasonet: Modeling large-scale structured relationships
with shared memory. arXiv preprint arXiv:1611.04642, 2016. 4
[22] Y. Shen, P.-S. Huang, J. Gao, and W. Chen. Reasonet: Learning to stop reading in machine comprehension. In
Proceedings of the 23rd ACM SIGKDD International Conference on Knowledge Discovery and Data Mining,
pages 1047–1055. ACM, 2017. 4
[23] K. Simonyan and A. Zisserman. Very deep convolutional networks for large-scale image recognition. arXiv
preprint arXiv:1409.1556, 2014. 11
[24] O. Vinyals, A. Toshev, S. Bengio, and D. Erhan. Show and tell: A neural image caption generator. In Proceedings
of the IEEE conference on computer vision and pattern recognition, pages 3156–3164, 2015. 3
[25] S. Wang and J. Jiang. Machine comprehension using match-lstm and answer pointer. arXiv preprint
arXiv:1608.07905, 2016. 4
[26] J. Weston, A. Bordes, S. Chopra, A. M. Rush, B. van Merriënboer, A. Joulin, and T. Mikolov. Towards aicomplete question answering: A set of prerequisite toy tasks. arXiv preprint arXiv:1502.05698, 2015. 4
[27] C. Xiong, V. Zhong, and R. Socher. Dynamic coattention networks for question answering. arXiv preprint
arXiv:1611.01604, 2016. 4
[28] K. Xu, J. Ba, R. Kiros, K. Cho, A. Courville, R. Salakhudinov, R. Zemel, and Y. Bengio. Show, attend and tell:
Neural image caption generation with visual attention. In International Conference on Machine Learning, pages
2048–2057, 2015. 3
[29] Z. Yang, X. He, J. Gao, L. Deng, and A. Smola. Stacked attention networks for image question answering. In
Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 21–29, 2016. 3, 4
[30] H. Zhang, T. Xu, H. Li, S. Zhang, X. Huang, X. Wang, and D. Metaxas. Stackgan: Text to photo-realistic image
synthesis with stacked generative adversarial networks. arXiv preprint arXiv:1612.03242, 2016. 2, 3, 5
[31] R. Zhang, P. Isola, and A. A. Efros. Colorful image colorization. In European Conference on Computer Vision,
pages 649–666. Springer, 2016. 2, 12
14
| 1cs.CV
|
Simple average-case lower bounds for approximate near-neighbor
from isoperimetric inequalities
arXiv:1602.05391v2 [cs.DS] 6 Jan 2017
Yitong Yin∗
Abstract
We prove an Ω(d/ log sw
nd ) lower bound for the average-case cell-probe complexity of deterministic or Las Vegas randomized algorithms solving approximate near-neighbor (ANN) problem in
d-dimensional Hamming space in the cell-probe model with w-bit cells, using a table of size s.
This lower bound matches the highest known worst-case cell-probe lower bounds for any static
data structure problems.
This average-case cell-probe lower bound is proved in a general framework which relates the
cell-probe complexity of ANN to isoperimetric inequalities in the underlying metric space. A
tighter connection between ANN lower bounds and isoperimetric inequalities is established by
a stronger richness lemma proved by cell-sampling techniques.
1
Introduction
The nearest neighbor search (NNS) problem is a fundamental problem in Computer Science. In this
problem, a database y = (y1 , y2 , . . . , yn ) of n points from a metric space (X, dist) is preprocessed
to a data structure, and at the query time given a query point x from the same metric space, we
are asked to find the point yi in the database which is closest to x according to the metric.
In this paper, we consider a decision and approximate version of NNS, the approximate nearneighbor (ANN) problem, where the algorithm is asked to distinguish between the two cases: (1)
there is a point in the databases that is λ-close to the query point for some radius λ, or (2) all
points in the database are γλ-far away from the query point, where γ ≥ 1 is the approximation
ratio.
The complexity of nearest neighbor search has been extensively studied in the cell-probe model,
a classic model for data structures. In this model, the database is encoded to a table consisting of
memory cells. Upon each query, a cell-probing algorithm answers the query by making adaptive
cell-probes to the table. The complexity of the problem is measured by the tradeoff between the
time cost (in terms of number of cell-probes to answer a query) and the space cost (in terms of
sizes of the table and cells). There is a substantial body of work on the cell-probe complexity of
NNS for various metric space [2, 3, 5–8, 11, 12, 14, 16, 17, 20].
It is widely believed that NNS suffers from the “curse of dimensionality” [10]: The problem may
become intractable to solve when the dimension of the metric space becomes very high. Consider the
most important example, d-dimensional Hamming space {0, 1}d with d ≥ C log n for a sufficiently
∗
State Key Lab for Novel Software Technology, Nanjing University, China. [email protected]. This work was
supported by NSFC grants 61272081 and 61321491.
1
large constant C. The conjecture is that NNS in this metric remains hard to solve when either
approximation or randomization is allowed individually.
In a series of pioneering works [3, 5, 6, 11, 14], by a rectangle-based technique of asymmetric
communication complexity known as the richness lemma [15], cell-probe lower bounds in form of
Ω(d/ log s), where s stands for the number of cells in the table, were proved for deterministic approximate near-neighbor (due to Liu [14]) and randomized exact near-neighbor (due to Barkol and
Rabani [5]). Such lower bound is the highest possible lower bound one can prove in the communication model. This fundamental barrier was overcome by an elegant self-reduction technique
introduced in the seminal work of Pǎtraşcu and Thorup [18], in which the cell-probe lower bounds
for deterministic ANN and randomized exact near-neighbor were improved to Ω(d/ log sw
n ), where
w represents the number of bits in a cell. More recently, in a previous work of us [20], by applying
the technique of Pǎtraşcu and Thorup to the certificates in data structures, the lower bound for
deterministic ANN was further improved to Ω(d/ log sw
nd ). This last lower bound behaves differently
for the polynomial space where sw = poly(n), near-linear space where sw = n · polylog(n), and
linear space where sw = O(nd). In particular, the bound becomes Ω(d) when the space cost is
strictly linear in the entropy of the database, i.e. when sw = O(nd).
When both randomization and approximation are allowed, the complexity of NNS is substantially reduced. With polynomial-size tables, a Θ(log log d/ log log log d) tight bound was proved
for randomized approximate NNS in d-dimensional Hamming space [7, 8]. If we only consider the
decision version, the randomized ANN can be solved with O(1) cell-probes on a table of polynomial
size [8]. For tables of near-linear size, a technique called cell-sampling was introduced by Panigrahy et al. [16, 17] to prove Ω(log n/ log sw
n ) lower bounds for randomized ANN. This was later
extended to general asymmetric metrics [1].
Among these lower bounds, the randomized ANN lower bounds of Panigrahy et al. [16, 17] were
proved explicitly for average-case cell-probe complexity. The significance of average-case complexity
for NNS was discussed in their papers. A recent breakthrough in upper bounds [4] also attributes to
solving the problem on a random database. Retrospectively, the randomized exact near-neighbor
lower bounds due to the density version of richness lemma [5,6,11] also hold for random inputs. All
these average-case lower bounds hold for Monte Carlo randomized algorithms with fixed worst-case
cell-probe complexity. This leaves open an important case: the average-case cell-probe complexity
for the deterministic or Las Vegas randomized algorithms for ANN, where the number of cell-probes
may vary for different inputs.
1.1
Our contributions
We study the average-case cell-probe complexity of deterministic or Las Vegas randomized algorithms for the approximate near-neighbor (ANN) problem, where the number of cell-probes to
answer a query may vary for different query-database pairs and the average is taken with respect
to the distribution over input queries and databases.
For ANN in Hamming space {0, 1}n , the hard distribution over inputs is very natural: Every
point yi in the database y = (y1 , y2 , . . . , yn ) is sampled uniformly and independently from the
Hamming space {0, 1}d , and the query point x is also a point sampled uniformly and independently
from {0, 1}d . According to earlier average-case lower bounds [16,17] and the recent data-dependent
LSH algorthm [4], this input distribution seems to capture the hardest case for nearest neighbor
search and is also a central obstacle to overcome for efficient algorithms.
2
By a simple proof, we show the following lower bound for the average-case cell-probe complexity
of ANN in Hamming space with this very natural input distribution.
Theorem 1.1. For d ≥ 32 log n and d < no(1) , any deterministic or Las Vegas randomized algorithm solving (γ, λ)-approximate near-neighbor problem in d-dimensional Hamming space in the
cell-probe model with w-bitcells for w< no(1) , using a table of size s < 2d , must have expected
cell-probe complexity t = Ω
d
2
γ 2 log swγ
nd
, where the expectation is taken over both the uniform and
independent input database and query and the random bits of the algorithm.
This lower bound matches the highest known worst-case cell-probe lower bounds for any static
data structure problems. Such lower bound was only known for polynomial evaluation [13, 19] and
also worst-case deterministic ANN due to our previous work [20].
We also prove an average-case cell-probe lower bound for ANN under ℓ∞ -distance. The lower
bound matches the highest known worst-case lower bound for the problem [2].
In fact, we prove these lower bounds in a unified framework that relates the average-case cellprobe complexity of ANN to isoperimetric inequalities regarding an expansion property of the metric
space.
Inspired by the notions of metric expansion defined in [17], we define the following notion of
expansion for metric space. Let (X, dist) be a metric space. The λ-neighborhood of a point x ∈ X,
denoted as Nλ (x) is the set of all points in X within distance λ from x. Consider a distribution
µ over X. We say the λ-neighborhoods are weakly independent under distribution µ, if for any
point x ∈ X, the measure of the λ-neighborhood µ(Nλ (x)) < nβ for a constant β < 1. We say
the λ-neighborhoods are (Φ, Ψ)-expanding under distribution µ, if for any point set A ⊆ X with
1
, where Nλ (A) denotes the set of all points within distance
µ(A) ≥ Φ1 , we have µ(Nλ (A)) ≥ 1 − Ψ
λ from some point in A.
Consider the database y = (y1 , y2 , . . . , yn ) ∈ X n with every point yi sampled independently
from µ, and the query x ∈ X sampled independently from µ. We denote this input distribution as
µ × µn . We prove the following lower bound.
Theorem 1.2. For a metric space (X, dist), assume the followings:
• the γλ-neighborhoods are weakly independent under distribution µ;
• the λ-neighborhoods are (Φ, Ψ)-expanding under distribution µ.
Then any deterministic or Las Vegas randomized algorithm solving (γ, λ)-approximate near-neighbor
problem in (X, dist) in the cell-probe model with w-bit cells, using a table of size s, must have
expected cell-probe complexity
!
n log Ψ
log Φ
or t = Ω
t=Ω
sw
log n log
w + log s
Ψ
under input distribution µ × µn .
The key step to prove such a theorem is a stronger version of the richness lemma that we
prove in Section 3. The proof of this stronger richness lemma uses an idea called “cell-sampling”
introduced by Panigrahy et al. [17] and later refined by Larsen [13]. This new richness lemma as
well as this connection between the rectangle-based techniques (such as the richness lemma) and
information-theory-based techniques (such as cell-sampling) are of interests by themselves.
3
2
Preliminary
Let (X, dist) be a metric space. Let γ ≥ 1 and λ ≥ 0. The (γ, λ)-approximate near-neighbor
problem (γ, λ)-ANNnX is defined as follows: A database y = (y1 , y2 , . . . , yn ) ∈ X n of n points from
X is preprocessed and stored as a data structure. Upon each query x ∈ X, by accessing the data
structure we want to distinguish between the following two cases: (1) there is a point yi in the
database such that dist(x, z) ≤ λ; (2) for all points yi in the database we have dist(x, z) > γλ. For
all other cases the answer can be arbitrary.
More abstractly, given a universe X of queries and a universe Y of all databases, a data
structure problem is a function f : X × Y → Z that maps every pair of query x ∈ X and
database y ∈ Y to an answer f (x, y) ∈ Z. In our example of (γ, λ)-ANNnX , the query universe
is the metric space X, the database universe is the set Y = X n of all tuples of n points from X,
and f maps each query x ∈ X and database y ∈ Y to an Boolean answer: f (x, y) = 0 if there is
a λ-near neighbor of x in the database y; f (x, y) = 1 if no points in the database y is a γλ-near
neighbor of x; and f (x, y) can be arbitrary if otherwise. Note that due to a technical reason, we
usually use 1 to indicate the “no near-neighbor” case.
Given a data structure problem f : X × Y → Z, a code T : Y → Σs with alphabet Σ = {0, 1}w
encodes every database y ∈ Y to a table Ty of s cells with each cell storing a word of w bits. We
use [s] = {1, 2, . . . , s} to denote the set of indices of cells. For each i ∈ [s], we use Ty [i] to denote
the content of the i-th cell of table Ty ; and for S ⊆ [s], we write Ty [S] = (Ty [i])i∈S for the tuple
of the contents of the cells in S. Upon each query x ∈ X, a cell-probing algorithm adaptive
retrieves the contents of the cells in the table Ty (which is called cell-probes) and outputs the
answer f (x, y) at last. Being adaptive means that the cell-probing algorithm is actually a decision
tree: In each round of cell-probing the address of the cell to probe next is determined by the query
x as well as the contents of the cells probed in previous rounds. Together, this pair of code and
decision tree is called a cell-probing scheme.
For randomized cell-probing schemes, the cell-probing algorithm takes a sequence of random bits
as its internal random coin. In this paper we consider only deterministic or Las Vegas randomized
cell-probing algorithms, therefore the algorithm is guaranteed to output a correct answer when it
terminates.
When a cell-probing scheme is fixed, the size s of the table as well as the length w of each cell
are fixed. These two parameters together give the space complexity. And the number of cell-probes
may vary for each pair of inputs (x, y) or may be a random variable if the algorithm is randomized.
Given a distribution D over X × Y , the average-case cell-probe complexity for the cell-probing
scheme is given by the expected number of cell-probes to answer f (x, y) for a (x, y) sampled from
D, where the expectation is taken over both the input distribution D and the internal random bits
of the cell-probing algorithm.
3
A richness lemma for average-case cell-probe complexity
The richness lemma (or the rectangle method) introduced in [15] is a classic tool for proving cellprobe lower bounds. A data structure problem f : X × Y → {0, 1} is a natural communication
problem, and a cell-probing scheme can be interpreted as a communication protocol between the
cell-probing algorithm and the table, with cell-probes as communications.
Given a distribution D over X × Y , a data structure problem f : X × Y → {0, 1} is α-dense
4
under distribution D if ED [f (x, y)] ≥ α. A combinatorial rectangle A × B for A ⊆ X and B ⊆ Y
is a monochromatic 1-rectangle in f if f (x, y) = 1 for all (x, y) ∈ A × B.
The richness lemma states that if a problem f is dense enough (i.e. being rich in 1’s) and is
easy to solve by communication, then f contains large monochromatic 1-rectangles. Specifically, if
an α-dense problem f can be solved by Alice sending a bits and Bob sending b bits in total, then
f contains a monochromatic 1-rectangle of size α · 2−O(a) × α · 2−O(a+b) in the uniform measure.
In the cell-probe model with w-bit cells, tables of size s and cell-probe complexity t, it means the
monochromatic 1-rectangle is of size α · 2−O(t log s) × α · 2−O(t log s+tw) . The cell-probe lower bounds
can then be proved by refuting such large 1-rectangles for specific data structure problems f .
We prove the following richness lemma for average-case cell-probe complexity.
Lemma 3.1. Let µ, ν be distributions over X and Y respectively, and let f : X × Y → {0, 1}
be α-dense under the product distribution µ × ν. If there is a deterministic or randomized Las
Vegas cell-probing scheme solving f on a table of s cells, eachcell containing
w bits, with expected t
cell-probes under input distribution µ × ν, then for any ∆ ∈ 32t/α2 , s , there is a monochromatic
O(t/α2 )
s
and ν(B) ≥ α · 2−O(∆ ln ∆ +∆w) .
1-rectangle A × B ⊆ X × Y in f such that µ(A) ≥ α · ∆
s
Compared to the classic richness lemma, this new lemma has the following advantages:
• It holds for average-case cell-probe complexity.
• It gives stronger result even restricted to worst-case complexity. The newly introduced parameter ∆ should not be confused as an overhead caused by the average-case complexity
argument, rather, it strengthens the result even for the worst-case lower bounds. When
∆ = t it gives the bound in the classic richness lemma.
• The lemma claims the existence of a family of rectangles parameterized by ∆, therefore to
prove a cell-probe lower bound it is enough to refute any one rectangle from this family. As
we will see, this gives us a power to prove the highest lower bounds (even for the worst case)
known to any static data structure problems.
The proof of this lemma uses an argument called “cell-sampling” introduced by Panigrahy et
al. [16, 17] for approximate nearest neighbor search and later refined by Larsen [13] for polynomial
evaluation. Our proof is greatly influenced by Larsen’s approach.
The rest of this section is dedicated to the proof of this lemma.
3.1
Proof of the average-case richness lemma (Lemma 3.1)
By fixing random bits, it is sufficient to consider only deterministic cell-probing algorithms.
The high level idea of the proof is simple. Fix a table Ty . A procedure called the “cell-sampling
procedure” chooses the subset Γ of ∆ many cells that resolve the maximum amount of positive
queries. This associates each database y to a string ω = (Γ, Ty [Γ]), which we call a certificate,
where Ty [Γ] = (Ty [i])i∈Γ represent the contents of the cells in Γ. Due to the nature of the cellprobing algorithm, once the certificate is fixed, the set of queries it can resolve is fixed. We also
observe that if the density of 1’s in the problem f is Ω(1), then there is a Ω(1)-fraction of good
databases y such that amount of positive queries resolved by the certificate ω constructed by the
cell-sampling procedure is at least an ( ∆s )O(t) -fraction of all queries. On the other hand, since
s
+∆w)
∆w there are at most s 2∆w = 2O(∆ ln ∆
many certificates ω. Therefore,
ω ∈ [s]
∆
∆ × {0, 1}
5
s
s
at least 2−O(∆ ln ∆ +∆w) -fraction of good databases (which is at least 2−O(∆ ln ∆ +∆w) -fraction of all
databases) are associated with the same ω. Pick this popular certificate ω, the positive queries that
ω resolves together with the good databases that ω is associated with form the large monochromatic
1-rectangle.
Now we proceed to the formal parts of the proof. Given a database y ∈ Y , let Xy+ = {x ∈ X |
f (x, y) = 1} denote the set of positive queries on y. We use µ+
y = µXy+ to denote the distribution
+
induced by µ on Xy .
Let Pxy ⊆ [s] denote the set of cells probed by the algorithm to resolve query x on database y.
Fix a database y ∈ Y . Let Γ ⊆ [s] be a subset of cells. We say a query x ∈ X is resolved by Γ if
x can be resolved by probing only cells in Γ on the table storing database y, i.e. if Pxy ⊆ Γ. We
denote by
Xy+ (Γ) = {x ∈ Xy+ | Pxy ⊆ Γ}
the set of positive queries resolved by Γ on database y. Assume two databases y and y ′ are
indistinguishable over Γ: meaning that for the tables Ty and Ty′ storing y and y ′ respectively, the
cell contents Ty [i] = Ty′ [i] for all i ∈ Γ. Then due to the determinism of the cell-probing algorithm,
we have Xy+ (Γ) = Xy+′ (Γ), i.e. Γ resolve the same set of positive queries on both databases.
The cell-sampling procedure: Fix a database y ∈ Y and any ∆ ∈ 32t/α2 , s . Suppose we
have a cell-sampling procedure which does the following: The procedure deterministically1 chooses
a unique Γ ⊆ [s] such that |Γ| = ∆ and the measure µ(Xy+ (Γ)) of positive queries resolved by Γ
is maximized (and if there are more than one such Γ, the procedure chooses an arbitrary one of
them). We use Γ∗y to denote this set of cells chosen by the cell-sampling procedure. We also denote
by Xy∗ = Xy+ (Γ∗y ) the set of positive queries resolved by this chosen set of cells.
On each database y, the cell-sampling procedure chooses for us the most informative set Γ of
cells of size |Γ| = ∆ that resolve the maximum amount of positive queries. We use ωy = (Γ∗y , Ty [Γ∗y ])
to denote the contents (along with addresses) of the cells chosen by the cell-sampling procedure for
database y. We call such ωy a certificate chosen by the cell-sampling procedure for y.
Let y and y ′ be two databases. A simple observation is that if two databases y and y ′ have the
same certificate ωy = ωy′ chosen by the cell-sampling procedure, then the respective sets Xy∗ , Xy∗′
of positive queries resolved on the certificate are going to be the same as well.
Proposition 3.2. For any databases y, y ′ ∈ Y , if ωy = ωy′ then Xy∗ = Xy∗′ .
Let τ (x, y) = |P (x, y)| denote the number of cell-probes to resolve query x on database y. By
the assumption of the lemma, Eµ×ν [τ (x, y)] ≤ t for the inputs (x, y) sampled from the product
distribution µ × ν. We claim that there are many “good” columns (databases) with high density
of 1’s and low average cell-probe costs.
Claim 3.3. There is a collection Ygood ⊆ Y of substantial amount of good databases, such that
ν(Ygood ) ≥ α4 and for every y ∈ Ygood , the followings are true:
• the amount of positive queries is large: µ(Xy+ ) ≥ α2 ;
• the average cell-probe complexity among positive queries is bounded:
Ex∼µ+
[τ (x, y)] ≤
y
1
8t
.
α2
Being deterministic here means that the chosen set Γ∗y is a function of y.
6
Proof. The claim is proved by a series of averaging principles. First consider Ydense = {y ∈ Y |
µ(Xy+ ) ≥ α2 } the set of databases with at least α2 -density of positive queries. By the averaging
principle, we have ν(Ydense ) ≥ α/2. Since E[τ (x, y)] ≥ ν(Ydense )E[τ (x, y) | y ∈ Ydense ], we have
Eµ×νdense [τ (x, y)] ≤ 2t
α , where νdense = νYdense is the distribution induced by ν on Ydense . We then
construct Ygood ⊆ Ydense as the set of y ∈ Ydense with average cell-probe complexity bounded as
1
α
Ex∼µ [τ (x, y)] ≤ 4t
α . By Markov inequality νdense (Ygood ) ≥ 2 and hence ν(Ygood ) ≥ 4 . Note that
Ex∼µ [τ (x, y)] ≥ Ex∼µ+
[τ (x, y)]µ(Xy+ ). We have Ex∼µ+
[τ (x, y)] ≤ Ex∼µ [τ (x, y)]/µ(Xy+ ) ≤ α8t2 for
y
y
all y ∈ Ygood .
For the rest, we consider only these good databases. Fix any ∆ ∈ 32t/α2 , s . We claim that
for every good database y ∈ Ygood , the cell-sampling procedure always picks a subset Γ∗y ⊆ [s] of ∆
many cells, which can resolve a substantial amount of positive queries:
Claim 3.4. For every y ∈ Ygood , it holds that µ(Xy∗ ) ≥
α
4
2
∆ 8t/α
.
2s
Proof. Fix any good database y ∈ Ygood . We only need to prove there exists a Γ ⊆ [s] with |Γ| = ∆
2
∆ 8t/α
. The claims follows immediately.
that resolve positive queries µ(Xy+ (Γ)) ≥ α4 2s
[s]
We construct a hypergraph H ⊆ 2 with vertex set [s] as H = {Pxy | x ∈ Xy+ }, so that each
positive queries x ∈ Xy+ on database y is associated (many-to-one) to a hyperedge e ∈ H such
that e = Pxy is precisely the set of cells probed by the cell-probing algorithm to resolve query x on
database y.
We also define a measure µ̃ over hyperedges e ∈ H as the total measure (in µ+
y ) of the positive
queries x associated to e. Formally, for every e ∈ H,
X
µ̃(e) =
µ+
y (x).
x∈Xy+ :Pxy =e
P
P
Since e∈H µ̃(e) = x∈Xy+ µ+
y (x) = 1, this µ̃ is a well-defined probability distribution over hyperedges in H. Moreover, recalling that τ (x, y) = |Pxy |, the the average size of hyperedges
Ee∼µ̃ [|e|] = Ex∼µ+
[τ (x, y)] ≤
y
8t
.
α2
By the probabilistic method (whose proof is in the full paper [21]), there must exist a Γ ⊆ [s] of
size |Γ| = ∆, such that the sub-hypergraph HΓ induced by Γ has
1
µ̃(HΓ ) ≥
2
∆
2s
8t/α2
.
By our construction of H, the positive queries associated (many-to-one) to the hyperedges in the
induced sub-hypergraph HΓ = {Pxy | x ∈ Xy+ ∧ Pxy ⊆ Γ} are precisely those positive queries in
Xy+ (Γ) = {x ∈ Xy+ | Pxy ⊆ Γ}. Therefore,
+
µ+
y (Xy (Γ))
=
X
µ+
y (x)
x∈Xy+ ,Pxy ⊆Γ
7
1
= µ̃(HΓ ) ≥
2
∆
2s
8t/α2
.
Recall that µ(Xy+ ) ≥
α
2
for every y ∈ Ygood . And since Xy+ (Γ) ⊆ Xy+ , we have
µ(Xy+ (Γ))
=
+
+
µ+
y (Xy (Γ))µ(Xy )
α
≥
4
∆
2s
8t/α2
.
The claim is proved.
Recall that the certificate ωy = (Γ∗y , Ty [Γ∗y ]) is constructed by the cell-sampling procedure for
∆w of certificate, let Y denote the set of
database y. For every possible assignment ω ∈ [s]
ω
∆ × {0, 1}
good databases y ∈ Ygood with this certificate ωy = ω. Due to the determinism of the cell-sampling
s
procedure, this classifies the Ygood into at most ∆
2∆w many disjointed subclasses Yω . Recall that
α
ν(Ygood ) ≥ 4 . By the averaging principle, the following proposition is natural.
∆w , denoted as ω ∗ , such that
Proposition 3.5. There exists a certificate ω ∈ [s]
∆ × {0, 1}
ν(Yω∗ ) ≥
α
4
s ∆w .
∆ 2
On the other hand, fixed any ω, since all databases y ∈ Yω have the same ωy∗ , by Proposition 3.2
they must have the same Xy∗ . We can abuse the notation and write Xω = Xy∗ for all y ∈ Yω .
Now we let A = Xω∗ and B = Yω∗ , where ω ∗ satisfies Proposition 3.5. Due to Claim 3.4 and
Proposition 3.5, we have
α
µ(A) ≥
4
∆
2s
8t/α2
=α·
∆
s
O(t/α2 )
and
ν(B) ≥
4
s
α
= α · 2−O(∆ ln ∆ +∆w) .
∆w
2
s
∆
Note for every y ∈ B = Yω∗ , the A = Xω∗ = Xy+ (Γ∗y ) is a set of positive queries on database y, thus
A × B is a monochromatic 1-rectangle in f . This finishes the proof of Lemma 3.1.
4
Rectangles in conjunction problems
Many natural data structure problems can be expressed as a conjunction of point-wise relations
between the query point and database points. Consider data structure problem f : X × Y → {0, 1}.
Let Y = Y n , so that each database y ∈ Y is a tuple y = (y1 , y2 , . . . , yn ) of n points from Y. A
point-wise function g : X × Y → {0, 1} is given. The data structure problem f is defined as the
conjunction of these subproblems:
∀x ∈ X, ∀y = (y1 , y2 , . . . , yn ) ∈ Y,
f (x, y) =
n
^
g(x, yi ),
i=1
Many natural data structure problems can be defined in this way, for example:
• Membership query: X = Y is a finite domain. The point-wise function g(·, ·) is 6= that
indicates whether the two points are unequal.
• (γ, λ)-approximate near-neighbor (γ, λ)-ANNnX : X = Y is a metric space with distance
dist(·, ·). The point-wise function g is defined as: for x, z ∈ X, g(x, z) = 1 if dist(x, z) > γλ,
or g(x, z) = 0 if dist(x, z) ≤ λ. The function value can arbitrary for all other cases.
8
d
d
• Partial match PMd,n
Σ : Σ is an alphabet, Y = Σ and X = (Σ ∪ {⋆}) . The point-wise function
g is defined as: for x ∈ X and z ∈ Y, g(x, z) = 1 if there is an i ∈ [d] such that xi 6∈ {⋆, zi },
or g(x, z) = 0 if otherwise.
We show that refuting the large rectangles in the point-wise function g can give us lower bounds
for the conjunction problem f .
Let µ, ν be distributions over X and Y respectively, and let ν n be the product distribution on
Y = Y n . Let g : X × Y → {0, 1} be a point-wise function and f : X × Y → {0, 1} a data structure
problem defined by the conjunction of g as above.
Lemma 4.1. For f, g, µ, ν defined as above, assume that there is a deterministic or randomized Las
Vegas cell-probing scheme solving f on a table of s cells, each cell containing w bits, with expected
t cell-probes under input distribution µ × ν n . If the followings are true:
• the density of 0’s in g is at most
β
n
under distribution µ × ν for some constant β < 1;
1
Φ
• g does not contain monochromatic 1-rectangle of measure at least
µ × ν;
then
sw
n log Ψ
O(t)
≥Φ
or
t=Ω
n log Ψ
w + log s
×
1
Ψ
under distribution
.
Proof. By union bound, the density of 0’s in f under distribution µ × ν n is:
#
"n
^
β
g(x, yi ) = 0 ≤ n · x∼µ
Pr [g(x, z) = 0] ≤ n · = β.
Pr
x∼µ
n
z∼ν
y=(y1 ,...,yn )∼ν n
i=1
By Lemma 3.1, the Ω(1)-density of 1’s in f and the assumption of existing a cell-probing scheme
with parameters s, w and t, altogether imply that for any 4t ≤ ∆ ≤ s, f has a monochromatic
1-rectangle A × B such that
c 1 t
s
∆
µ(A) ≥
and ν n (B) ≥ 2−c2 ∆(ln ∆ +w),
(1)
s
for some constants c1 , c2 > 0 depending only on β.
Let C ⊂ Y be the largest set of columns in g to form a 1-rectangle with A. Formally,
C = {z ∈ Y | ∀x ∈ A, g(x, z) = 1}.
Clearly, for any monochromatic 1-rectangle A × D in g, we must have D ⊆ C. By definition of f as
a conjunction of g, it must hold that for all y = (y1 , y2 , . . . , yn ) ∈ B, none of yi ∈ y has g(x, yi ) = 0
for any x ∈ A, which means B ⊆ C n , and hence
ν n (B) ≤ ν n (C n ) = ν(C)n .
Recall that A × C is monochromatic 1-rectangle in g. Due to the assumption of the lemma, either
µ(A) < Φ1 or ν(C) < Ψ1 . Therefore, either µ(A) < Φ1 or ν n (B) < Ψ1n .
9
We can always choose a ∆ such that ∆ = O
n log Ψ
w
s
2−c2 ∆(ln ∆ +w) >
and ∆ = Ω
1
.
Ψn
n log Ψ
w+log s
to satisfy
If such ∆ is less than 32t/(1 − β)2 , then we immediately have a lower bound
n log Ψ
t=Ω
.
w + log s
Otherwise, due to (1), A × B is monochromatic 1-rectangle in f with ν n (B) >
must hold that µ(A) < Φ1 , which by (1) gives us
1
> µ(A) ≥
Φ
∆
s
O(t)
=
n log Ψ
sw
O(t)
1
Ψn ,
therefore it
,
which gives the lower bound
5
sw
n log Ψ
O(t)
≥ Φ.
Isoperimetry and ANN lower bounds
Given a metric space X with distance dist(·, ·) and λ ≥ 0, we say that two points x, x′ ∈ X are
λ-close if dist(x, x′ ) ≤ λ, and λ-far if otherwise. The λ-neighborhood of a point x ∈ X, denoted by
Nλ (x), is S
the set of all points from X which are λ-close to x. Given a point set A ⊆ X, we define
Nλ (A) = x∈A Nλ (x) to be the set of all points which are λ-close to some point in A.
In [17], a natural notion of metric expansion was introduced.
Definition 5.1 (metric expansion [17]). Let X be a metric space and µ a probability distribution
over X. Fix any radius λ > 0. Define
Φ(δ) ,
min
A⊂X,µ(A)≤δ
µ(Nλ (A))
.
µ(A)
The expansion Φ of the λ-neighborhoods in X under distribution µ is defined as the largest k such
1
, Φ(δ) ≥ k.
that for all δ ≤ 2k
We now introduce a more refined definition of metric expansion using two parameters Φ and Ψ.
Definition 5.2 ((Φ, Ψ)-expanding). Let X be a metric space and µ a probability distribution over
X. The λ-neighborhoods in X are (Φ, Ψ)-expanding under distributions µ if we have µ(Nλ (A)) ≥
1 − 1/Ψ for any A ⊆ X that µ(A) ≥ 1/Φ.
The metric expansion defined in [17] is actually a special case of (Φ, Ψ)-expanding: The expansion of λ-neighborhoods in a metric space X is Φ means the λ-neighborhoods are (Φ, 2)-expanding.
The notion of (Φ, Ψ)-expanding allows us to describe a more extremal expanding situation in metric
space: The expanding of λ-neighborhoods does not stop at measure 1/2, rather, it can go all the way
10
to be very close to measure 1. This generality may support higher lower bounds for approximate
near-neighbor.
Given a radius λ > 0 and an approximation ratio γ > 1, recall that the V
(γ, λ)-approximate near
neighbor problem (γ, λ)-ANNnX can be defined as a conjunction f (x, y) = i g(x, yi ) of point-wise
function g : X × X → {0, 1} where g(x, z) = 0 if x is λ-close to z; g(x, z) = 1 if x is γλ-far
from z; and g(x, z) is arbitrary for all other cases. Observe that g is actually (γ, λ)-ANN1X , the
point-to-point version of the (γ, λ)-approximate near neighbor.
The following proposition gives an intrinsic connection between the expansion of metric space
and size of monochromatic rectangle in the point-wise near-neighbor relation.
Proposition 5.1. If the λ-neighborhoods in X are (Φ, Ψ)-expanding under distribution µ, then the
function g defined as above does not contain a monochromatic 1-rectangle of measure ≥ Φ1 × 1.01
Ψ
under distribution µ × µ.
Proof. Since the λ-neighborhoods in X are (Φ, Ψ)-expanding, for any A ⊆ X with µ(A) ≥ Φ1 , we
have µ(Nλ (A)) ≥ 1 − Ψ1 . And by definition of g, for any monochromatic A × B, it must hold that
B ∩ Nλ (A) = ∅, i.e. B ⊆ X \ Nλ (A). Therefore, either µ(A) < Φ1 , or µ(B) = 1 − µ(Nλ (A)) ≤ Ψ1 <
1.01
Ψ .
The above proposition together with Lemma 4.1 immediately gives us the following corollary
which reduces lower bounds for near-neighbor problems to the isoperimetric inequalities.
Corollary 5.2. Let µ be a distribution over a metric space X. Let λ > 0 and γ ≥ 1. Assume
that there is a deterministic or randomized Las Vegas cell-probing scheme solving (γ, λ)-ANNnX on
a table of s cells, each cell containing w bits, with expected t cell-probes under input distribution
µ × µn . If the followings are true:
• Ex∼µ [µ(Nγλ (x))] ≤
β
n
for a constant β < 1;
• the λ-neighborhoods in X are (Φ, Ψ)-expanding under distribution µ;
then
sw
n log Ψ
O(t)
≥Φ
or
t=Ω
n log Ψ
w + log s
.
Remark 5.1. In [17], a lower bound for (γ, λ)-ANNnX was proved with the following form:
swt
n
t
≥ Φ.
In our Corollary 5.2, unless the cell-size w is unrealistically large to be comparable to n, the corollary
always gives the first lower bound
O(t)
sw
≥ Φ.
n log Ψ
Θ(d) , 2Θ(d) when
the
metric
space
is
2
This strictly improves the lower bound in [17]. For example,
expanding, this would give us a lower bound t = Ω logdsw , which in particular, when the space is
nd
linear (sw = O(nd)), becomes t = Ω(d).
11
5.1
Lower bound for ANN in Hamming space
Let X = {0, 1}d be the Hamming space with Hamming distance dist(·, ·). Recall that Nλ (x)
represents the λ-neighborhood around x, in this case, the Hamming ball of radius λ centered at x;
and for a set A ⊂ X, the Nλ (A) is the set of all points within distance λ to any point in A. For
any 0 ≤ r ≤ d B(r) = |Nr (0̄)| denote
of Hamming ball of radius r, where 0̄ ∈ {0, 1}d is
Pthe volume
d
the zero vector. Obviously B(r) = k≤r k .
The following isoperimetric inequality of Harper is well known.
Lemma 5.3 (Harper’s theorem [9]). Let X = {0, 1}d be the d-dimensional Hamming space. For
A ⊂ X, let r be such that |A| ≥ B(r). Then for every λ > 0, |Nλ (A)| ≥ B(r + λ).
In words, Hamming balls have the worst vertex expansion.
For 0 < r < d2 , the following upper bound for the volume of Hamming ball is well known:
d
2(1−o(1))dH(r/d) ≤
≤ B(r) ≤ 2dH(r/d) ,
r
where H(x) = −x log2 x − (1 − x) log2 (1 − x) is the Boolean entropy function.
Consider the Hamming (γ, λ)-approximate near-neighbor problem (γ, λ)-ANNnX . The hard distribution for this problem is just the uniform and independent distribution: For the database
y = (y1 , y2 , . . . , yn ) ∈ X n , each database point yi is sampled uniformly and independently from
X = {0, 1}n ; and the query point x is sampled uniformly and independently from X.
Theorem 5.4. Let d ≥ 32 log n. For any γ ≥ 1, there is a λ > 0 such that if (γ, λ)-ANNnX can be
solved by a deterministic or Las Vegas randomized cell-probing scheme on a table of s cells, each
cell containing
t cell-probes for uniform and independent database and query,
w bits, with expected
nd
d
.
then t = Ω 2 swγ2 or t = Ω γ 2 (w+log
s)
γ log
nd
Proof. Choose λ to satisfy γλ =
going to show:
• Ex∼µ [µ(Nγλ (x))] ≤
d
2
−
p
2d ln(2n). Let µ be uniform distribution over X. We are
1
2n ;
• the λ-neighborhoods in X are (Φ, Ψ)-expanding under distribution µ for some Φ = 2Ω(d/γ
2
and Ψ = 2Ω(d/γ ) .
2)
Then the cell-probe lower bounds follows directly from Corollary 5.2.
1
First, by the Chernoff bound, µ(Nγλ (x)) ≤ 2n
for any point x ∈ X. Thus trivially Ex∼µ [µ(Nγλ (x))] ≤
1
.
2n
d
On the other hand, for d ≥ 32 log n and n being sufficiently large, it holds that λ ≥ 4γ
. Let
d
d
−(1−H(r/d))d
dH(r/d)
r = 2 − 8γ . And consider any A ⊆ X with µ(A) ≥ 2
. We have |A| ≥ 2
≥ B(r).
Then by Harper’s theorem,
d
d
≥ 2d − B d2 − 8γ
= 2d − B(r) ≥ 2d − 2dH(r/d) ,
|Nλ (A)| ≥ B (r + λ) ≥ B d2 + 8γ
which means µ(Nλ (A)) ≥ 1 − 2−(1−H(r/d))d . In other words, the λ-neighborhoods in X are (Φ, Ψ)1
expanding under distribution µ for Φ = Ψ = 2(1−H(r/d))d , where r/d = 12 − 8γ
. Apparently
2
1 − H( 12 − x) = Θ(x2 ) for small enough x > 0. Hence, Φ = Ψ = 2Θ(d/γ ) .
12
5.2
Lower bound for ANN under L-infinity norm
Let Σ = {0, 1, . . . , m} and the metric space is X = Σd with ℓ∞ distance dist(x, y) = kx − yk∞ for
any x, y ∈ X.
Let µ be the distribution over X as defined
in [2]: First define a distribution π over Σ as
P
i
−(2ρ)
for all i > 0 and π(0) = 1 − i>0 π(i); and then µ is defined as µ(x1 , x2 , . . . , xd ) =
p(i) = 2
π(x1 )π(x2 ) . . . π(xd ).
The following isoperimetric inequality is proved in [2].
Lemma 5.5 (Lemma 9 of [2]). For any A ⊆ X, it holds that µ(N1 (A)) ≥ (µ(A))1/ρ .
Consider the (γ, λ)-approximate near-neighbor problem (γ, λ)-ANNnℓ∞ defined in the metric
space X under ℓ∞ distance. The hard distribution for this problem is µ × µn : For the database
y = (y1 , y2 , . . . , yn ) ∈ X n , each database point yi is sampled independently according to µ; and the
query point x is sampled independently from X according to µ. The following lower bound has
been proved in [2] and [12].
Fix any ǫ > 0 and 0 < δ < 21 . Assume Ω log1+ǫ n ≤ d ≤ o(n). For 3 < c ≤ O(log log d), define
ρ = 21 ( 4ǫ log d)1/c > 10. Now we choose γ = logρ log d and λ = 1.
Theorem 5.6. With d, γ, λ, ρ and the metric space X defined as above, if (γ, λ)-ANNnℓ∞ can be
solved by a deterministic or Las Vegas randomized cell-probing scheme on a table of s cells, each
cell containing w ≤ n1−2δ bits, with expected t ≤ ρ cell-probes under input distribution µ × µn , then
sw = nΩ(ρ/t) .
Proof. The followings are true
• µ(Nγλ (x)) =
e− log
1+ǫ/3 n
n
≤
1
2n
for any x ∈ X (Claim 6 in [2]);
δ
• the λ-neighborhoods in X are (nδρ , nδn−1 )-expanding under distribution µ for Φ = nδρ and
2
Ψ = 2Ω(d/γ ) .
δ
To see the expansion is true, let Φ = nδρ and Ψ = nδn−1 . By Lemma 5.5, for any set A ⊂ X
with µ(A) ≥ Φ, we have µ(Nλ (A)) ≥ n−δ ≥ 1 − Ψ1 . This means λ-neighborhoods of M are
δ
(nδρ , nδn−1 )-expanding.
1−δ
O(t)
n
≥ nδρ or t = Ω w+log
Due to Corollary 5.2, either nsw
1−δ
s . The second bound is always
higher with our ranges for w and t. The first bound gives sw = nΩ(ρ/t) .
References
[1] Amirali Abdullah and Suresh Venkatasubramanian. A directed isoperimetric inequality with
application to bregman near neighbor lower bounds. In STOC’15.
[2] Alexandr Andoni, Dorian Croitoru, and Mihai Pǎtraşcu. Hardness of nearest neighbor under
L-infinity. In FOCS’08.
[3] Alexandr Andoni, Piotr Indyk, and Mihai Pǎtraşcu. On the optimality of the dimensionality
reduction method. In FOCS’06.
13
[4] Alexandr Andoni and Ilya Razenshteyn. Optimal data-dependent hashing for approximate
near neighbors. In STOC’15.
[5] Omer Barkol and Yuval Rabani. Tighter lower bounds for nearest neighbor search and related
problems in the cell probe model. Journal of Computer and System Sciences, 64(4):873–896,
2002. Conference version in STOC’00.
[6] Allan Borodin, Rafail Ostrovsky, and Yuval Rabani. Lower bounds for high dimensional nearest
neighbor search and related problems. In Discrete and Computational Geometry, pages 253–
274, 2003. Conference version in STOC’99.
[7] Amit Chakrabarti, Bernard Chazelle, Benjamin Gum, and Alexey Lvov. A lower bound on
the complexity of approximate nearest-neighbor searching on the hamming cube. In Discrete
and Computational Geometry, pages 313–328, 2003. Conference version in STOC’99.
[8] Amit Chakrabarti and Oded Regev. An optimal randomised cell probe lower bound for approximate nearest neighbour searching. In SIAM Journal on Computing, 39(5):1919–1940,2010.
Conference version in FOCS’04.
[9] L.H. Harper. Optimal numberings and isoperimetric problems on graphs. Journal of Combinatorial Theory, 1(3):385 – 393, 1966.
[10] Piotr Indyk. Nearest neighbors in high-dimensional spaces. Handbook of Discrete and Computational Geometry, pages 877–892, 2004.
[11] T.S. Jayram, Subhash Khot, Ravi Kumar, and Yuval Rabani. Cell-probe lower bounds for
the partial match problem. In Journal of Computer and System Sciences, 69(3):435–447, 2004.
Conference version in STOC’03.
[12] Michael Kapralov and Rina Panigrahy. NNS lower bounds via metric expansion for ℓ∞ and
EMD. In ICALP’12.
[13] Kasper Green Larsen. Higher cell probe lower bounds for evaluating polynomials. In FOCS’12.
[14] Ding Liu. A strong lower bound for approximate nearest neighbor searching. Information
Processing Letters, 92(1):23–29, 2004.
[15] Peter Bro Miltersen, Noam Nisan, Shmuel Safra, and Avi Wigderson. On data structures and
asymmetric communication complexity. Journal of Computer and System Sciences, 57(1):37–
49, 1998. Conference version in STOC’95.
[16] Rina Panigrahy, Kunal Talwar, and Udi Wieder. A geometric approach to lower bounds for
approximate near-neighbor search and partial match. In FOCS’08.
[17] Rina Panigrahy, Kunal Talwar, and Udi Wieder. Lower bounds on near neighbor search via
metric expansion. In FOCS’10.
[18] Mihai Pǎtraşcu and Mikkel Thorup. Higher lower bounds for near-neighbor and further rich
problems. SIAM Journal on Computing, 39(2):730–741, 2010. Conference version in FOCS’06.
14
[19] Alan Siegel. On universal classes of fast high performance hash functions, their time-space
tradeoff, and their applications. In FOCS’89.
[20] Yaoyu Wang and Yitong Yin. Certificates in data structures. In ICALP’14.
[21] Yitong Yin. Simple average-case lower bounds for approximate near-neighbor from isoperimetric inequalities. arXiv preprint arXiv:1602.05391.
15
| 8cs.DS
|
arXiv:1710.04349v1 [math.ST] 12 Oct 2017
Asymptotic theory for maximum likelihood estimates in
reduced-rank multivariate generalised linear models
Efstathia Bura
a∗
∗a∗
, Sabrina Duarteb, Liliana Forzanib, Ezequiel Smuclerc, Mariela Suedc
Institute of Statistics and Mathematical Methods in Economics, TU Wien, A-1040
Vienna, Austria and Department of Statistics, George Washington University,
Washington, D.C. 20052, U.S.A.
b
Facultad de Ingenierı́a Quı́mica, UNL, Researcher of CONICET, 3000 Santa Fe,
Argentina
c
Instituto de Cálculo, UBA, CONICET, FCEN, 1428 Buenos Aires, Argentina
Abstract
Reduced-rank regression is a dimensionality reduction method with many applications. The asymptotic theory for reduced rank estimators of parameter matrices in multivariate linear models has been
studied extensively. In contrast, few theoretical results are available for reduced-rank multivariate generalised linear models. We develop M-estimation theory for concave criterion functions that are maximised
over parameters spaces that are neither convex nor closed. These results are used to derive the consistency
and asymptotic distribution of maximum likelihood estimators in reduced-rank multivariate generalised
linear models, when the response and predictor vectors have a joint distribution. We illustrate our results
in a real data classification problem with binary covariates.
Keywords: M-estimation; exponential family; rank restriction; non-convex; parameter spaces
∗
Corresponding author. Email: [email protected]
1
1
Introduction
The multivariate multiple linear regression model for a q-dimensional response Y = (Y1, . . . , Yq )T and a pdimensional predictor vector X = (X1 , . . . , Xp )T postulates that Y = BX +ǫ, where B is a q×p matrix and ǫ =
(ǫ1 , . . . , ǫq )T is the error term, with E(ǫ) = 0 and var(ǫ) = Σ. Based on a random sample (X1 , Y1), . . . , (Xn , Yn )
satisfying the model,
P ordinary least squares estimates the parameter matrix B by minimizing the squared
error loss function ni=1 ||Yi − BXi ||2 to obtain
bols =
B
n
X
i=1
Yi XiT
!
n
X
i=1
Xi XiT
!−1
.
(1)
Reduced-rank regression introduces a rank constraint on B, so that (1) is minimized subject to the
T
bT
bT
constraint rank(B) ≤ r, where r < min(p, q). The solution is B
RRR = Bols Ur Ur , where Ur are the first r
bols X T [see, e.g., Reinsel and Velu (1998)].
singular vectors of Yb T = B
Reduced-rank regression has attracted attention as a regularization method by introducing a shrinkage
penalty on B. Moreover, it is used as a dimensionality reduction method as it constructs latent factors in the
predictor space that explain the variance of the responses.
Anderson (1951) obtained the likelihood-ratio test of the hypothesis that the rank of B is a given number
and derived the associated asymptotic theory under the assumption of normality of Y and non-stochastic
X. Under the assumption of joint normality of Y and X, Izenman (1975) obtained the asymptotic distribution of the estimated reduced-rank regression coefficient matrix and drew connections between reduced-rank
regression, principal component analysis and correlation analysis. Specifically, principal component analysis
coincides with reduced-rank regression when Y = X (Izenman (2008)). The monograph by Reinsel and Velu
(1998) contains a comprehensive survey of the theory and history of reduced rank regression, including in
time series, and its many applications.
Despite the application potential of reduced-rank regression, it has received limited attention in generalised
linear models. Yee and Hastie (2003) were the first to introduce reduced-rank regression to the class of
multivariate generalised linear models, which covers a wide range of data types for both the response and
predictor vectors, including categorical data. Multivariate or vector generalised linear models is the topic of
Yee’s (2015) book, which is accompanied by the associated R packages VGAM and VGAMdata. Yee and Hastie
(2003) proposed an alternating estimation algorithm, which was shown to result in the maximum likelihood
estimate of the parameter matrix in reduced-rank multivariate generalised linear models by Bura, Duarte and
Forzani (2016). Asymptotic theory for the restricted rank maximum likelihood estimates of the parameter
matrix in multivariate GLMs has not been developed yet.
In general, a maximum likelihood estimator is a concave M-estimator in the sense that it maximizes the
empirical mean of a concave criterion function. Asymptotic theory for M-estimators defined through a concave
function has received much attention. Huber (1967), Haberman (1989) and Niemiro (1992) are among the
classical references. More recently, Hjort and Pollard (2011) presented a unified framework for the statistical
2
theory of M-estimation for convex criterion functions that are minimized over open convex sets of a Euclidean
space. Geyer (1994) studied M-estimators restricted to a closed subset of a Euclidean space.
The rank restriction in reduced rank regression imposes constraints that have not been studied before
in M-estimation as they result in neither convex nor closed parameter spaces. In this paper we (a) develop
M-estimation theory for concave criterion functions, which are maximized over parameters spaces that are
neither convex nor closed, and (b) apply the results from (a) to obtain asymptotic theory for reduced rank
regression estimators in generalised linear models. Specifically, we derive the asymptotic distribution and
properties of maximum likelihood estimators in reduced-rank multivariate generalised linear models where
both the response and predictor vectors have a joint distribution. The asymptotic theory we develop covers
reduced-rank regression for linear models as a special case. We show the improvement in inference the
asymptotic theory offers via analysing the data set Yee and Hastie (2003) analysed.
Throughout, for a function f : Rq → R, ∇f (x) denotes the row vector ∇f (x) = (∂f (x)/∂x1 , . . . , ∂f (x)/∂xq ),
f˙(x) stands for the column vector of derivatives, while ∇2 f (x) denotes the symmetric matrix of second order
derivatives. For a vector valued function f : Rq1 → Rq2 , ∇f denotes the q2 × q1 matrix,
∂f1
∂f1
∂f1
·
·
·
∂x2
∂xq1
1
∂x
∂f2 ∂f2 · · · ∂f2
∂x
∂x2
∂xq1
∇f (x) = . 1
.
.
..
.
.
..
..
.
∂fq2
∂fq2
∂fq2
· · · ∂xq
∂x1
∂x2
1
2
M-estimators
Let Z be a random vector taking values in a measurable space Z and distributed according to the law
P . We are interested in estimating a finite dimensional parameter ξ0 = ξ0 (P ) using n independent and
identically distributed copies Z1 , Z2 , . . . , Zn of Z. In the sequel, we use P f to denote the mean of f (Z); i.e.,
P f = E[f (Z)].
Let Ξ be a subset of an Euclidean space and mΞ : Z → R be a known function. Assume that the parameter
of interest ξ0 is the maximizer of the map ξ 7→ P mξ defined on Ξ. One can estimate ξ0 by maximizing an
empirical version of the optimization criterion. Specifically, given a known function mξ : Z 7→ R, define
M(ξ) := P mξ
and Mn (ξ) := Pn mξ ,
(2)
P
where, here and throughout, Pn denotes the empirical mean operator Pn V = n−1 n1=1 Vi . Hereafter, Mn (ξ)
and Pn mξ will be used interchangeably as the criterion function, depending on which of the two is appropriate
in a given setting.
Assume that ξ0 is the unique maximizer of the deterministic function M defined in (2). An M-estimator
3
for the criterion function Mn over Ξ is defined as
n
1X
mξ (Zi ) over Ξ .
ξbn = ξbn (Z1 , . . . , Zn ) maximizing Mn (ξ) =
n i=1
(3)
If the maximum of the criterion function Mn over Ξ is not attained but the supremum of Mn over Ξ is finite,
any value ξbn that almost maximizes the criterion function, in the sense that it satisfies
Mn (ξbn ) ≥ sup Mn (ξ) − An
(4)
ξ∈Ξ
for An small, can be used instead.
Definition 2.1. An estimator ξbn that satisfies (4) with An = op (1) is called a weak M-estimator for the
criterion function Mn over Ξ. When An = op (n−1 ), ξbn is called a strong M-estimator.
Proposition 6.1 in the Appendix lists the conditions for the existence, uniqueness and strong consistency
of an M-estimator, as defined in (3), when Mn is concave and the parameter space is convex. Under regularity
conditions, as those stated in Theorem 5.23 in van der Vaart (2000), the asymptotic expansion and distribution
of a consistent strong M-estimator [see Definition 2.1] for ξ0 is given by
√
n
1 X
IFξ0 (Zi ) + op (1),
n(ξbn − ξ0 ) = √
n i=1
(5)
where IFξ0 (Zi ) = −Vξ−1
ṁξ0 (Zi ), and Vξ0 is the nonsingular symmetric second derivative matrix of M(ξ) at
0
ξ0 .
2.1
Restricted M-estimators
We now consider the optimization of Mn over Ξres ⊂ Ξ, where Ξres is the image of a function that is not
necessarily injective. Specifically, we restrict the optimization problem to the set Ξres by requiring:
Condition 2.2. There exists an open set Θ ⊂ Rq and a map g : Θ → Ξ such that ξ0 ∈ g(Θ) = Ξres .
Even when an M-estimator for the unrestricted problem as defined in (3) exists, there is no a priori
guarantee that the supremum is attained when considering the restricted problem. Nevertheless, Lemma 2.3
establishes the existence of a restricted strong M-estimator, regardless of whether the original M-estimator is
a real maximiser, weak or strong. All proofs are provided in the Appendix.
Lemma 2.3. Assume there exists a weak/strong M-estimator ξbn for the criterion function Mn over Ξ. If
Condition 2.2 holds, then there exists a strong M-estimator ξbnres for the criterion function Mn over Ξres .
4
Proposition 2.4 next establishes the existence and consistency of a weak restricted M-estimator sequence
when mξ (z) is a concave function in ξ under the same assumptions as those of the unrestricted problem,
stated in Proposition 6.1 in the Appendix.
Proposition 2.4. Assume that Condition 2.2 holds. Then, under the assumptions of Proposition 6.1 in
the Appendix, there exists a strong M-estimator of the restricted problem over Ξres . Morevoer, any strong
M-estimator of the restricted problem converges to ξ0 in probability.
We derive next the asymptotic distribution of ξbnres = g(θbn ), with θbn ∈ Θ. The constrained estimator ξbnres is
√
well defined under Lemma 2.3, even when θbn is not uniquely determined. If θbn were unique and n(θbn − θ)
had an asymptotic distribution, one could use a Taylor series expansion for g to derive asymptotic results for
ξbnres . Building on this idea, Condition 2.5 introduces a parametrization of a neighborhood of ξ0 that allows
applying standard tools in order to obtain the asymptotic distribution of the restricted M-estimator.
Condition 2.5. Given ξ0 ∈ g(Θ), there exist an open set M in Ξres with ξ0 ∈ M ⊂ g(Θ), and (S, h),
where S is an open set in Rqs , qs ≤ q, and h : S → M is one-to-one, bi-continuous and twice continuously
differentiable, with ξ0 = h(s0 ) for some s0 ∈ S.
Under the setting in Condition 2.5, we will prove that sbn = h−1 (ξbnres ) is a strong M-estimator for the
criterion function Pn mh(s) over S. Then, we can apply Theorem 5.23 of van der Vaart (2000) to obtain the
asymptotic behavior of sbn , which, combined with a Taylor expansion of h about s0 , yield a linear expansion
for ξbnres . Finally, requiring Condition 2.6, which relates the parametrizations (S, h) and (Θ, g), suffices to
derive the asymptotic distribution of ξbnres in terms of g.
Condition 2.6. Consider (Θ, g) as in Condition 2.2 and (S, h) as in Condition 2.5. For each θ0 ∈ g −1 (ξ0 ),
span∇g(θ0 ) = span∇h(s0 ).
Condition 2.6 ensures that Tξ0 = span∇g(θ0 ) is well defined regardless of the fact that g −1 (ξ0 ) may contain
multiple θ0 ’s. Moreover, Tξ0 also agrees with span∇h(s0 ). Consequently, the orthogonal projection Πξ0 (Σ) onto
Tξ0 with respect to the inner product defined by a definite positive matrix Σ satisfies
Πξ0 (Σ) = ∇g(θ0 )(∇g(θ0 )T Σ ∇g(θ0 ))† ∇g(θ0 )T Σ
= ∇h(s0 )(∇h(s0 )T Σ∇h(s0 ))−1 ∇h(s0 )T Σ,
(6)
where A† denotes a generalised inverse of the matrix A. The gradient of g is not necessarily of full rank, in
contrast to the gradient of h.
Proposition 2.7. Assume that Conditions 2.2, 2.5 and 2.6 hold. Assume also that the unrestricted problem
satisfies the regularity Conditions 6.2, 6.3 and 6.4 in the Appendix. Then, any strong M-estimator ξbnres of the
restricted problem that converges in probability to ξ0 satisfies
√
n
1 X
Πξ (−V ) IFξ0 (Zi ) + op (1),
n(ξbnres − ξ0 ) = √
n i=1 0 ξ0
5
(7)
where IFξ0 (Zi ) = −Vξ−1
ṁξ0 (Zi ) is the influence function of the unrestricted estimator defined in (5), Vξ0 is
0
the nonsingular symmetric second derivative matrix of M(ξ) at ξ0 , and Πξ0 (−Vξ0 ) is defined according to (6).
√
Moreover, n(ξbnres − ξ0 ) is asymptotically normal with mean zero and asymptotic variance
o
n
√
−1 T
T
avar{ n(ξbnres − ξ0 )} = Πξ0 (−Vξ0 ) Vξ−1
(8)
P
ṁ
ṁ
ξ 0 ξ 0 Vξ0 Πξ 0 (−Vξ ) .
0
0
As a side remark, we conjecture that for estimators that are maximizers of a criterion function under restrictions that satisfy Conditions 2.2, 2.5 and 2.6, when (5) is true, (7) also holds. This can be important since
the asymptotic distribution of restricted estimators will be derived directly from the asymptotic distribution
of the unrestricted one.
3
Asymptotic theory for the maximum likelihood estimator in
reduced rank multivariate generalised linear models
In this section we show that maximum likelihood estimators in reduced-rank multivariate generalised linear
models are restricted strong M-estimators for the conditional log-likelihood. Using results in Section 2.1, we
obtain the existence, consistency and asymptotic distribution of maximum likelihood estimators in reducedrank multivariate generalised linear models.
3.1
Exponential Family
Let Y = (Y1 , . . . , Yq )T be a q-dimensional random vector and assume that its distribution belongs to a kparameter canonical exponential family with pdf (pms)
fη (y) = exp{η T T (y) − ψ(η)}h(y),
(9)
where T (y) = (T1 (y), . . . , Tk (y))T is a vector of known real-valued functions, h(y) ≥ 0 is a nonnegative known
function and η ∈ Rk is the vector of natural parameters, taking values in
Z
k
H = {η ∈ R : exp{η T T (y)}h(y)dy < ∞},
(10)
where the integral is replaced by a sum when Y is discrete. The set H of the natural parameter space is
assumed to be open and convex in Rk , and ψ a strictly convex function defined on H. Moreover, we assume
ψ(η) is convex and infinitely differentiable in H. In particular,
∇ψ(η) = ETη (T (Y )) and ∇2 ψ(η) = varη (T (Y )),
for every η ∈ H,
(11)
where ∇2 ψ is the k ×k matrix of second derivatives of ψ. Since ψ is strictly convex, varη (T (Y )) is non-singular
for every η ∈ H.
6
3.2
Multivariate generalised linear models
Let Z = (X, Y ) be a random vector, where now Y ∈ Rq is a multivariate response and X ∈ Rp is a vector of
predictors. The multivariate generalised linear model postulates that the conditional distribution of Y given
X belongs to some fixed exponential family and hypothesizes that the k-vector of natural parameters, which
we henceforth call ηx to emphasise the dependence on x, depends linearly on the vector of predictors. Thus,
the pdf (pms) of Y | X = x is given by
fY |X=x (y) = exp{ηxT T (y) − ψ(ηx )}h(y),
(12)
where ηx ∈ Rk depends linearly on x.
Frequently, a subset of the natural parameters depends on x while its complement does not. The normal
linear model with constant variance is such an example. To accommodate this structure, we partition the
vector ηx indexing model (12) into ηx1 and ηx2 , with k1 and k2 components, and assume that H, the natural
parameter space of the exponential family, is Rk1 × H2 , where H2 is an open convex subset of Rk2 , and assume
that
!
!
ηx1
η 1 + βx
ηx =
=
(13)
ηx2
η2
T
where β ∈ Rk1 ×p , η 1 ∈ Rk1 and η 2 ∈ H2 . Let ξ = ηT1 , vecT (β), ηT2 ∈ Ξ = Rk1 × Rk1 p × H2 , denotes a generic
vector and ξ0 the true parameter. Suppose n independent and identically distributed copies of Z = (X, Y )
satisfying (12) and (13), with true parameter vector ξ0 , are available. Given a realisation zi = (xi , yi ),
i = 1, . . . , n, the conditional log-likelihood, up to a factor that does not depend on the parameter of interest,
is
n
1 X T
Ln (β; η) =
η T (yi ) − ψ(ηxi ) .
(14)
n i=1 xi
Let
m(β;η) (z) = ηxT T (y) − ψ(ηx ).
(15)
By definition (3), the maximum likelihood estimator (MLE), if exits, of the parameter indexing model (12)
subject to (13) is an M-estimator.
Theorem 3.1 next establishes the existence, consistency and asymptotic normality of ξbn , the MLE of ξ0 .
Theorem 3.1. Assume that Z = (X, Y ) satisfies model (12) subject to (13) with true parameter ξ0 . Under
regularity conditions (47), (48), (49) and (50) in the Appendix, the maximum likelihood estimate of ξ0 , ξbn ,
√
exists, is unique and converges in probability to ξ0 . Moreover, n (ξbn − ξ0 ) is asymptotically normal with
covariance matrix
−1
Wξ0 = E F (X)T ∇2 ψ(F (X)ξ0)F (X)
,
(16)
7
where
F (x) =
(1, xT ) ⊗ Ik1
0
0
Ik 2
!
,
and ∇2 ψ was defined in (11)
3.3
Partial reduced rank multivariate generalised linear models
When the number of natural parameters or the number of predictors is large, the precision of the estimation
and/or the interpretation of results can be adversely affected. A way to address this is to assume that
the parameters live in a lower dimensional space. That is, we assume that the vector of predictors can be
partitioned as x = (xT1 , xT2 )T with x1 ∈ Rr and x2 ∈ Rp−r , and that the parameter corresponding to x1 ,
β1 ∈ Rk1 ×r has rank d < min{k1 , r}. In this way, the natural parameters ηx in (12) are related to the
predictors via
!
!
ηx1
η1 + β1 x1 + β2 x2
(17)
ηx =
=
ηx2
η2
where β1 ∈ Rkd1 ×r , the set of matrices in Rk1 ×p of rank d ≤ min{k1 , r}, while β2 ∈ Rk1 ×(p−r) , and β = (β1 , β2 ).
Following Yee and Hastie (2003), we refer to the exponential conditional model (12) subject to the restrictions imposed in (17) as partial reduced rank multivariate generalised linear model. The reduced-rank
multivariate generalised linear model is the special case with β2 = 0 in the partial reduced rank multivariate
generalised linear model.
To obtain the asymptotic distribution of the M-estimators for this reduced model, we will show that
Conditions 2.2, 2.5 and 2.6 are satisfied for Ξres , (Θ, g), M and (S, h), which are defined next. To maintain
consistency with notation introduced in Section 2.1, we vectorise each matrix involved in the parametrisation
of our model and reformulate accordingly the parameter space for each vectorised object. With this understanding, we use the symbol ∼
= to indicate that a matrix space component in a product space is identified with
its image through the operator vec : Rm×n → Rmn . In the sequel, to keep the notation as simple as possible,
we concatenate column vectors without transposing them; i.e., we write (a, b) for (aT , bT )T . Moreover, we
write ξ = (η 1 , β, η2 ), with the understanding that β stands for vec(β).
For the non-restricted problem, β1 belongs to Rk1 ×r , so that the entire parameter ξ = (η 1 , β1 , β2 , η2 )
belongs to
(18)
Ξ∼
= Rk1 × Rk1 ×r × Rk1 ×(p−r) × H2 .
However, for the restricted problem, we assume that the true parameter ξ0 = (η 01 , β01 , β02 , η 02 ) belongs to
Ξres ∼
= Rk1 × Rkd1 ×r × Rk1 ×(p−r) × H2 .
Let
× Rk1 ×(p−r) × H2
Θ∼
= Rk1 × Rkd1 ×d × Rd×r
d
8
(19)
(20)
and consider g : Θ → Ξres , with (η 1 , (S, T ), β2 , η2 ) 7→ (η1 , ST, β2 , η2 ). Without loss of generality, we assume
1 ×r
, the set of matrices in Rkd1 ×r whose first d rows are linearly independent. Therefore,
that β01 ∈ Rkfirst,d
!
Id
B0 ,
with A0 ∈ R(k1 −d)×d and B0 ∈ Rd×r
(21)
β01 =
d .
A0
This is trivial since β01 ∈ Rkd1 ×r and its first d rows linearly independent. Consider
1 ×r
× Rk1 ×(p−r) × H2 ,
M∼
= Rk1 × Rkfirst,d
Finally, let
and h : S → M be the map
and so ξ0 ∈ M ⊂ Ξres .
× Rk1 ×(p−r) × H2 ,
S∼
= Rk1 × R(k1 −d)×d × Rd×r
d
(η 1 , (A, B), β2 , η2 )
7→
η1 ,
B
AB
!
, β2 , η 2
!
(22)
(23)
.
(24)
Proposition 3.2. Conditions 2.2, 2.5 and 2.6 are satisfied for ξ0 , Ξ, Ξres , (Θ, g) and (S, h) defined in (18)–
(24), respectively.
Under the rank restriction on β1 in (17), the existence of the maximum likelihood estimate cannot be
guaranteed in the sense of an M-estimator as defined in (3) with mξ = m(β;η) in (15), and Ξ replaced by Ξres
in (18). However, using Lemma 2.3 we can work with a strong M-estimator sequence for the criterion function
Pn m(β;η) over Ξres . Theorem 3.3 states our main contribution.
Theorem 3.3. Let ξ0 = (η01 , β01 , β02 , η 02 ) denote the true parameter value of ξ = (η1 , β1 , β2 , η 2 ). Assume that
Z = (X, Y ) satisfies model (12), subject to (17) with ξ0 ∈ Ξres defined in (19). Then, there exists a strong
maximising sequence for the criterion function Pn mξ over Ξres for mξ = m(β;η) defined in (15). Moreover,
any weak M-estimator sequence {ξbnres } converges to ξ0 in probability.
√
If {ξbnres } is a strong M-estimator sequence, then n(ξbnres − ξ0 ) is asymptotically normal with covariance
matrix
√
avar{ n(ξbnres − ξ0 )} = Πξ0 (W −1 ) Wξ0
ξ0
where Wξ0 is defined in (16)
Remark 3.4. Since Πξ0 (W −1 ) is a projection, Πξ0 (W −1 ) Wξ0 ≤ Wξ0 . That is, the eigenvalues of Wξ0 −
ξ0
ξ0
Πξ0 (W −1 ) Wξ0 are non-negative, so that using partial reduced-rank multivariate generalised linear models reξ0
sults in efficiency gain.
9
4
Application: Marital status in a workforce study
Yee and Hastie (2003) analyse data from a self-administered questionnaire collected in a large New Zealand
workforce observational study conducted during 1992-1993. For homogeneity, the analysis was restricted
to a subset of 4105 European males with no missing values in any of the variables used. Yee and Hastie
were interested in exploring whether certain lifestyle and psychological variables were associated with marital
status, especially separation/divorce. The response variable is Y = marital status, with levels 1 = single, 2 =
separated or divorced, 3 = widower, and 4 = married or living with a partner. The married/partnered are
the reference group. Data on 14 regressors were collected, 12 of which are binary (1/0 for presence/absence,
respectively). These have been coded so that their presence is negative healthwise. Their goal was to
investigate if and how these 12 unhealthy variables were related to Y , adjusting for age and level of education.
The variables are described in Table 1.
Variable Name
Description
marital
Marital status. 1=single, 2=separated or divorced, 3=widower, and 4=married or living with a partner
Age - 30, in years
log(1+ years of education at secondary or high school)
In the last 3 months what is the largest number of drinks that
you had on any one day? (1=20 or more, 0=less than 20)
Current smoker?
Does not usually wear a hat, shirt or suntan lotion when outside during summer
Do you suffer from ‘nerves’ ?
Would you call yourself a ‘nervous’ person?
Are your feelings easily hurt?
Would you call yourself tense or ‘highly strung’ ?
Do you feel ‘just miserable’ for no reason?
Do you often feel ‘fed-up’ ?
Do you worry about awful things that might happen?
Are you a worrier?
Does your mood often go up and down?
age30
logedu1
binge
smokenow
sun
nerves
nervous
hurt
tense
miserable
fedup
worry
worrier
mood
Table 1: Variables used in the workforce study
10
A categorical response Y taking values 1, 2, 3, 4 with probabilities pr(Y = i) = pi can be expressed as a
T
multinomial vector to fit the generalised linear
P4 model presented in this paper, Y = (Y1 , Y2 , Y3 , Y4 ) , where
Yi = 1 if Y = i and Yi = 0 otherwise, and i=1 Yi = 1. The pdf of Y can be written as
fY (y) = exp{η T T (y) − ψ(η)}h(y)
(25)
P
where y = (y1 , y2 , y3, y4 ), T (y) = (y1 , y2, y3 ), η = (η1 , η2 , η3 ) ∈ R3 , ψ(η) = 1 + 3i=1 exp(ηi ) and h(y) =
1/(y1 !y2 !y3 !(1 − y1 − y2 − y3 )!). The natural parameter η = (η1 , η2 , η3 ) ∈ R3 , is related to the pdf of Y
through the identity ηi = log(pi /p4 ), i = 1, 2, 3.
Let X be the vector of predictor variables and consider pi = pi (x) = pr(Yi = 1|X = x). The dependence of
pi on x will not be made explicit in the notation. As in Yee and Hastie (2003) we fit a multinomial regression
model to Y |X = x, as in (25), with ηx
log(p1 /p4 )
ηx = log(p2 /p4 ) = η + βx
(26)
log(p3 /p4 )
where η ∈ R3 is the intercept and β ∈ R3×14 is the coefficient matrix, so that there are 3 ×14 + 3 = 42 + 3 = 45
parameters to estimate. When a multinomial linear model is fitted to the data at level 0.05, age30 and binge
are significant for log(p1 /p4 ), smokehow and tense for log(p2 /p4 ), and only age30 is significant for log(p3 /p4 ).
We next fitted a partial reduced rank multivariate generalised linear model, where the two continuous
variables, age30 and logedu1, were not subject to restriction. That is,
ηx = η + βx = η + β1 x1 + β2 x2 ,
(27)
where x2 represent the continuous variables and x1 the 12 binary predictors. The AIC criterion estimates the
rank of β1 in (27) to be one [see Yee and Hastie (2003)]. Using the asymptotic results from the current paper,
Duarte (2016) developed a test based on Bura and Yang (2011) that also estimates the dimension to be 1.
Therefore, in our notation, q = 4, k = k1 = 3, p = 14, r = 2, d = 1, and β1 = AB, A : 3 × 1 y B : 1 × 12,
β2 : 3 × 2 and η : 3 × 1. The rank restriction results in a drastic reduction in the total number of parameters
from 45 to 24.
The reduction in the estimation burden is also reflected in how tight the confidence intervals are compared
with those in the unrestricted model, as can be seen in Tables 3, 4 and Table 2 in Yee and Hastie (2003).
As a consequence the variables nervous, hurt, which are not significant in the unrestricted generalised linear
model, are significant in the reduced (27). Furthermore, some variables, such as binge, smokenow, nervous
and tense, are now significant for all responses.
All significant coefficients are positive. These correspond to the variables binge, smokenow, nervous,
tense and hurt only for widowers. Since the positive value of the binary variables indicates poor lifestyle and
negative psychological characteristics, our analysis concludes that for men with these features, the chance of
11
being single, divorced or widowed is higher than the chance of being married, adjusting for age and education.
Also, the coefficients corresponding to the response log(p3 /p4 ) are twice as large as those of log(p1 /p4 ),
suggesting the effect of the predictors differs in each group. All computations were performed using the R
package VGAM, developed by Yee (2017).
Variable
log(p1 /p4 )
log(p2 /p4 )
log(p3 /p4 )
-1.762 *
(0.377)
-1.573*
(0.388)
-2.699 *
(0.383)
-2.921*
(0.396)
-6.711 *
(1.047)
-6.123 *
(1.106)
-0.191 *
(0.008)
-0.190*
(0.008)
0.012
(0.008)
0.012
(0.008)
0.086 *
(0.023)
0.077*
(0.024)
0.338
(0.225)
0.254
(0.228)
-0.365
(0.213)
-0.316
(0.214)
-0.089
(0.559)
-0.198
(0.566)
Intercept
GLM
PRR-GLM
age30
GLM
PRR-GLM
logedu1
GLM
PRR-GLM
Table 2: Estimators with standard errors for β1 for the generalised linear model (first two rows) and the
reduced generalised linear model (last two rows). ∗ indicates significant at 5% level.
12
Variable
log(p1 /p4 )
log(p2 /p4 )
log(p3 /p4 )
0.801*
(0.143)
0.569 *
(0.125)
0.318
(0.256)
0.786 *
(0.196)
1.127
(0.670)
1.114 *
(0.409)
0.022
(0.126)
0.222 *
(0.088)
0.501 *
(0.157)
0.306 *
(0.119)
0.654
(0.469)
0.434 *
(0.208)
-0.066
(0.122)
0.011
(0.084)
0.120
(0.161)
0.015
(0.116)
-0.088
(0.518)
0.021
(0.164)
-0.102
(0.138)
-0.054
(0.101)
0.123
(0.198)
-0.074
(0.139)
-1.456
(0.841)
-0.105
(0.197)
0.297
(0.169)
0.312 *
(0.124)
0.353
(0.228)
0.430 *
(0.168)
1.007
(0.665)
0.609 *
(0.291)
0.184
(0.126)
0.180
(0.089)
0.210
(0.167)
0.248
(0.122)
0.483
(0.501)
0.352 *
(0.199)
binge
GLM
PRR-GLM
smokenow
GLM
PRR-GLM
sun
GLM
PRR-GLM
nerves
GLM
PRR-GLM
nervous
GLM
PRR-GLM
hurt
GLM
PRR-GLM
Table 3: MLEs with their standard errors in parentheses for the full rank generalised linear model (first two
rows) and the partial reduced rank generalised linear model (last two rows). ∗ indicates significant at 5%
level.
13
Variable
log(p1 /p4 )
log(p2 /p4 )
log(p3 /p4 )
0.166
(0.176)
0.302 *
(0.122)
0.483 *
(0.214)
0.416 *
(0.163)
1.108
(0.612)
0.590 *
(0.284)
-0.050
(0.138)
0.019
(0.094)
0.128
(0.178)
0.0268
(0.129)
-0.093
(0.613)
0.038
(0.185)
0.112
(0.122)
0.117
(0.094)
0.249
(0.171)
0.161
(0.129)
-0.214
(0.548)
0.229
(0.185)
0.113
(0.145)
0.003
(0.106)
-0.102
(0.209)
0.004
(0.146)
-0.548
(0.818)
0.005
(0.207)
-0.027
(0.131)
-0.116
(0.092)
-0.243
(0.180)
-0.160
(0.128)
-0.548
(0.550)
-0.227
(0.193)
-0.111
(0.123)
-0.037
(0.087)
0.092
(0.171)
-0.052
(0.120)
-0.193
(0.553)
-0.073
(0.172)
tense
GLM
PRR-GLM
miserable
GLM
PRR-GLM
fedup
GLM
PRR-GLM
worry
GLM
PRR-GLM
worrier
GLM
PRR-GLM
mood
GLM
PRR-GLM
Table 4: MLEs with their standard errors in parentheses for the full rank generalised linear model (first two
rows) and the partial reduced rank generalised linear model (last two rows). ∗ indicates significant at 5%
level.
14
5
Discussion
With the exception of the work of Yee on vector generalised linear models (VGLMs) (Yee and Wild, (1996),
Yee and Hastie (2003), Yee (2017)) reduced-rank regression has been almost exclusively restricted to data
where the response variable is continuous. Estimation in reduced-rank multinomial logit models (RR-MLM)
was studied in Yee and Hastie (2003), but no distribution results for the estimators were obtained. In this
paper we fill this gap by developing asymptotic theory for the restricted rank maximum likelihood estimates
of the parameter matrix in multivariate GLMs.
To illustrate the potential impact of our results, we refer to the real data analysis example in Section 4.
In order to assess the significance of the predictors, Yee and Hastie (2003) calculate the standard errors
for the coefficient matrix factors, A and B, independently and can only infer about the significance of the
components of the matrix A and the components of the matrix B separately. The asymptotic distribution
for either estimator is obtained assuming that the other is fixed and known. In this way, they first analyse
ν = Bx1 to check which predictors are significant and then Aν to examine how they influence each response.
Their standard errors are ad-hoc and it is unclear what the product of standard errors measures as relates to
the significance of the product of the components of the coefficient matrix β1 = AB. Moreover, this practical
ad-hoc approach cannot readily be extended when d > 1.
Using the results of Theorem 3.3, we can obtain the errors of each component of the coefficient matrices
A, B simultaneously, and assess the statistical significance of each predictor on each response. Using the
ad-hoc approach of Yee and Hastie (2003), a predictor can only be found to be significant across all responses.
For example, Yee and Hastie (2003) find the predictor hurt to be significant for all three groups (single,
divorced/separated, widower). On the other hand, we can assess the significance of any response/predictor
combination. Thus, we find hurt to be significant only for widower but not for single or separated/divorced
men groups [see Table 3].
6
Proofs
The consistency of M-estimators has long been established [see, for instance, Theorem 5.7 in van der Vaart
(2000)]. The functions M(ξ) and Mn (ξ) are defined in (2). Typically, the proof for the consistency of Mestimators assumes that ξ0 , the parameter of interest, is a well-separated point of maximum of M, which
is ascertained by assumptions (a) and (b) of Proposition 6.1. Assumption (c) of Proposition 6.1 yields
uniform consistency of Mn as an estimator of M, a property needed in order to establish the consistency of
M-estimators.
Proposition 6.1. Assume (a) mξ (z) is a strictly concave function in ξ ∈ Ξ, where Ξ is a convex open subset
of a Euclidean space; (b) the function M(ξ) is well defined for any ξ ∈ Ξ and has a unique maximizer ξ0 ; that
15
is, M(ξ0 ) > M(ξ) for any ξ 6= ξ0 ; and (c) for any compact set K in Ξ,
E sup |mξ (Z)| < ∞.
(28)
ξ∈K
Then, for each n ∈ N, there exists a unique M-estimator ξbn for the criterion function Mn over Ξ. Moreover,
ξbn → ξ0 a.e., as n → ∞.
Proof. For each compact subset K of Ξ, {mξ : ξ ∈ K} is a collection of measurable functions which, by
assumption (c), has an integrable envelope. Moreover, for each fixed z, the map ξ 7→ mξ (z) is continuous,
since it is concave and defined on the open set Ξ. As stated in Example 19.8 of van der Vaart (2000), these
conditions guarantee that the class is Glivenko-Cantelli. That is,
pr lim sup |Pn mξ − P mξ | = 0 = 1.
(29)
n→∞ ξ∈K
We need to prove that there exist a unique maximizer of Mn (ξ) = Pn ξ, and that it converges to the maximizer
of M(ξ) = P mξ . We first consider the deterministic case ignoring for the moment that {Mn } is a sequence of
random functions.
Since ξ0 belongs to the open set Ξ, there exists ε0 > 0 such that the closed ball B[ξ0 , ε0 ] is contained in
Ξ. Uniform convergence of {Mn } to M over K = {ξ : kξ − ξ0 k = ε0 } guarantees that
lim
sup
n→∞ kξ−ξ k=ε
0
0
{Mn (ξ) − Mn (ξ0 )} =
sup
kξ−ξ0 k=ε0
{M(ξ) − M(ξ0 )} < 0
(30)
because M(ξ0 ) > M(ξ) for any ξ 6= ξ0 . Then, for n ≥ n0 (ε0 ),
sup
kξ−ξ0 k=ε0
Mn (ξ) − Mn (ξ0 ) < 0.
(31)
Let ζn (ξ) = Mn (ξ) − Mn (ξ0 ). Since Mn is concave and continuous, ζn attains its maximum over the compact
set B[ξ0 , ε0 ], which we denote by ξbn . Note that ζn (ξ0 ) = 0 and ζn is strictly smaller than zero in the boundary
of the ball, as shown in (31); therefore, we conclude that ξbn ∈ B(ξ0 , ε0 ), so that ξbn is a local maximum for ζn .
Let ξ satisfy kξ − ξ0 k > ε0 . The convexity of Ξ implies there exists t ∈ (0, 1) such that ξe = (1 − t)ξ0 + tξ
satisfies kξe − ξ0 k = ε0 , and therefore
e ≥ tζn (ξ) + (1 − t)ζn (ξ0 ) = tζn (ξ),
ζn (ξbn ) ≥ ζn (ξ0 ) = 0 > ζn (ξ)
(32)
implying that ζn (ξ) < 0 ≤ ζn (ξbn ). Therefore, the maximum ξbn ∈ B(ξ0 , ε0 ) is global. The strict concavity
of Mn guarantees that such global maximum is unique, thus ξbn is the unique solution to the optimization
16
problem in (3). By repeating this argument for any ε < ε0 , we prove the convergence of the sequence {ξbn } to
ξ0 .
Turning to the stochastic case, the uniform convergence of Mn to M over K = B[ξ0 , ε0 ] on a set Ω1 , with
pr(Ω1 ) = 1, as assumed in (29), guarantees the deterministic result can be applied to any element of Ω1 ,
which obtains the result.
Proof of Lemma 2.3. Let {ξbn } be any (weak/strong) M-estimator sequence of the unconstrained maximization
problem. Since Mn (ξbn ) ≥ supξ∈Ξ Mn (ξ) − An with An → 0, we have
(33)
1 = lim pr sup Pn mξ < ∞ ≤ lim pr sup Pn mξ < ∞ .
n→∞
n→∞
ξ∈Ξ
Define
Ωn :=
For all n, there exists ξbnres such that
ξ∈Ξres
sup Pn mξ < ∞ .
ξ∈Ξres
1
Mn (ξbnres ) ≥ sup Mn (ξ) − 2
n
ξ∈Ξres
(34)
(35)
.
on Ωn . Let ξbnres = 0 on Ωcn . Then, since pr(Ωn ) → 1, {ξbnres } is a strong M-estimator for the criterion function
Mn over Ξres .
Proof of Proposition 2.4. In Proposition 6.1 we have established the existence of a unique maximizer ξbn for
the criterion function Mn over Ξ. We can now invoke Lemma 2.3 to guarantee the existence of ξbnres , a strong
M-estimator for the criterion function Mn over Ξres . Let {ξbnres } be any strong M-estimator for the criterion
function Mn over Ξres . We start from the deterministic case:
Mn (ξbnres ) ≥ sup Mn (ξ) − An ,
(36)
ξ∈Ξres
where Mn is defined in (2) and An is a sequence of real numbers with An → 0. As in the proof of Proposition 6.1, define ζn (ξ) = Mn (ξ) − Mn (ξ0 ) to obtain that, for ǫ0 small enough,
sup
kξ−ξ0 k=ε0
ζn (ξ) ≤
1
sup {M(ξ) − M(ξ0 )} := −δ(ε0 )
2 kξ−ξ0 k=ε0
(37)
for n large enough. Under Condition 2.2, ξ0 ∈ Ξres , and therefore, by (36),
ζn (ξbnres ) = Mn (ξbnres ) − Mn (ξ0 ) ≥ Mn (ξbnres ) − sup Mn (ξ) ≥ −An .
ξ∈Ξres
17
(38)
Since An → 0, −An > −δ(ε0 ) for n large enough. Combining this with (37) and (38) obtains
sup
kξ−ξ 0 k=ε0
ζn (ξ) < ζn (ξbnres ).
We will deduce that kξbnres − ξ0 k < ε0 , once we prove that
sup
ζn (ξ) =
sup
ζn (ξ).
(39)
kξ−ξ 0 k≥ε0
kξ−ξ 0 k=ε0
Now, let prove (39). Choose ξ with kξ − ξ0 k > ε0 , and take t ∈ (0, 1) such that ξe = (1 − t)ξbn + tξ is a distance
ε0 from ξ0 , where ξbn is the maximizer of ζn over Ξ, as defined in Proposition 6.1, which is assumed to be at
distance smaller than ε0 from ξ0 . Then,
e = ζn (1 − t)ξbn + tξ ≥ (1 − t)ζn (ξbn ) + tζn (ξ) ≥ ζn (ξ).
ζn (ξ)
Thus,
sup
kξ−ξ 0 k=ε0
e ≥ ζn (ξ) ,
ζn (ξ) ≥ ζn (ξ)
for any ξ with kξ − ξ0 k > ε0
which in turn yields (39). When An = op (1), convergence in probability of {ξbnres } to ξ0 is equivalent to the
existence of an almost everywhere convergent sub-sub-sequence for any subsequence {ξbnresk }. Therefore, by
applying the deterministic result to the set of probability one, where there exists a sub-subsequence of Ank
that converges to zero a.e., we obtain the result.
Regularity conditions for Proposition 2.7 (from Theorem 5.23, p. 53 in [15])
Condition 6.2. For each ξ in an open subset Ξ of a Euclidean space, mξ (z) is a measurable function in z
such that mξ (z) is differentiable in ξ0 for almost every z with derivative ṁξ0 (z).
Condition 6.3. There exists a measurable function φ(z) with P φ2 < ∞ such that, for any ξ1 and ξ2 in a
neighborhood of ξ0 ,
|mξ1 (z) − mξ2 (z)| ≤ φ(z)kξ1 − ξ2 k
(40)
Condition 6.4. The map ξ → P mξ admits a second order Taylor expansion at a point of maximum ξ0 with
nonsingular symmetric second derivative matrix Vξ0 .
Under regularity conditions 6.2, 6.3 and 6.4, van der Vaart proved in Theorem 5.23 of his book ([15]) that
if {ξbn } is a strong M-estimator sequence for the criterion function Pn mξ over Ξ and ξbn → ξ0 in probability,
then
√
n
1 X
√
ṁξ0 (Zi ) + op (1).
n(ξbn − ξ0 ) = −Vξ−1
0
n i=1
18
(41)
Moreover,
√
n(ξbn − ξ0 ) is asymptotically normal with mean zero and
n√
o
b
n(ξn − ξ0 ) = Vξ−1
avar
P ṁξ0 ṁTξ0 Vξ−1
.
0
0
(42)
This result will be invoked in the following proofs.
Proof of Proposition 2.7. Assume that {ξbnres } is a sequence in Ξres that converges in probability to ξ0 ∈ M,
which is assumed to be open in Ξres . Then, pr(ξbnres ∈ M) → 1. Bicontinuity of h guarantees that s∗n = h−1 (ξbnres )
converges in probability to s0 = h−1 (ξ0 ). Note that
Pn mh(s∗n ) = Pn mbξres ≥ sup Pn mξ ≥ sup Pn mξ ≥ sup Pn mh(s) ,
n
ξ ∈Ξres
ξ ∈M
(43)
s∈S
except for an op (n−1 ) term that is omitted in the last three inequalities. Therefore, {s∗n } is a strong maximizing
sequence for the criterion function Pn mh(s) (z) over S.
We next verify Conditions 6.2, 6.3 and 6.4 are satisfied for {s∗n }, s0 , mh(s) (z) and P mh(s) . Specifically,
Condition 6.2 holds since mh(s) is a measurable function in z for all s ∈ S and mh(s) (z) is differentiable in s0
for almost every z. In fact, h(s0 ) = ξ0 , mξ (z) is differentiable at ξ0 and h(s) is also differentiable. Moreover,
the derivative function is ∇h(s0 )ṁξ0 .
For all s1 and s2 in a neighborhood of s0 , by the continuity of h, h(s1 ) and h(s2 ) are in a neighborhood of
ξ0 . Then
|mh(s1 ) (z) − mh(s2 ) (z)| ≤ φ(z)kh(s1 ) − h(s2 )k
≤ φ(z)k∇hk∞,Ns0 ks1 − s2 k
where k∇hk∞,Ns0 denotes the maximum of k∇h(s)k in a neighborhood Ns0 of s0 . The first inequality holds
because such condition is valid in the unconstrained problem and the second inequality follows since h is
continuously differentiable at s0 . Thus, the Lipschitz Condition 6.3 is satisfied.
For Condition 6.4, we observe that the function s 7→ P mh(s) is twice continuously differentiable in s0
because both P mξ and h(s) satisfy the required regularity properties at ξ0 and s0 , respectively. Moreover,
since P ṁξ0 = 0, the second derivative matrix of P mh(s) at s0 , is Ws0 = ∇h(s0 )T Vξ0 ∇h(s0 ), where Vξ0 is the
second derivative matrix of P mξ at ξ0 . The matrix Ws0 is nonsingular and symmetric because ∇h(s0 ) is full
rank and Vξ0 is nonsingular and symmetric.
We can now apply Theorem 5.23 in van der Vaart (2000), and obtain
√
n(s∗n − s0 ) = − ∇h(s0 )T Vξ0 ∇h(s0 )
n
−1 1 X
√
∇h(s0 )T ṁξ0 (Zi ) + op (1),
n i=1
19
(44)
so that the first order Taylor series expansion of h(s∗n ) around s0 is
√
√
n(h(s∗n ) − h(s0 )) = ∇h(s0 ) n(s∗n − s0 )) + op (1)
n
−1
1 X
∇h(s0 ) ∇h(s0 )T Vξ0 ∇h(s0 )
∇h(s0 )T ṁξ0 (Zi ) + op (1)
= −√
n i=1
n
1 X
= √
Πξ (−V IFξ0 (Zi ) + op (1),
n i=1 0 ξ0
for Πξ0 (−Vξ0 and IFξ0 (z) defined in (6) and (5), respectively, which obtains the expansion in (7).
Proof of Theorem 3.1. Write
ηx1 = η 1 + βx = (η1 , β) f (x) = (f (x)T ⊗ Ik1 )vec (η 1 , β) ,
where f (x)T = (1, xT ) ∈ R1×(p+1) . Then, in matrix form,
!
!
f (x)T ⊗ Ik1 0
η 1 + βx
=
ηx =
0
Ik 2
η2
where
f (x)T ⊗ Ik1 0
0
Ik 2
F (x) =
!
vec (η 1 , β)
η2
!
= F (x)ξ,
(45)
∈ Rk×(k1 (p+1)+k2 ) ,
T
and ξ = η T1 , vecT (β), ηT2
is the vector of parameters of model (13). Note that F (x)ξ ∈ H for any value of
ξ with H defined in (10). This notation allows to simplify the expression for the log-likelihood function in
(15), and replace it with
mξ (z) = T (y)T F (x)ξ − ψ(F (x)ξ).
(46)
The regularity conditions required to derive consistency and asymptotic distribution of the MLE are: For any
ξ and any η and for any compact K ⊂ Ξ = Rk1 × Rk1 p × H2 .
pr (F (X)ξ = η) < 1.
T
E kT (Y ) F (X)k < ∞ , E sup |ψ(F (X)ξ)| < ∞,
(47)
(48)
ξ∈K
2
E kT (Y ) F (X)k < ∞ , E sup k∇ψ(F (X)ξ)F (X)k < ∞ ,
ξ∈K
T 2
E sup kF (X) ∇ ψ(F (X)ξ)F (X)k < ∞.
T
2
ξ∈K
20
(49)
(50)
To prove the existence, uniqueness and consistency of the MLE under the present model, ξbn , we next show
that the assumptions stated in Proposition 6.1 are satisfied.
The strict convexity of ψ implies that, for each fixed z, mξ (z) is a strictly concave function in ξ ∈ Ξ =
k1 +pk1
R
× H2 . The concavity of mξ (z) is preserved under expectation, thus M(ξ) = P mξ is concave. The
identifiability condition satisfied by the exponential family in Section 3.1 allows applying Lemma 5.35 of van
der Vaart (2000, p. 62) to conclude that
E T (Y )T F (Y )ξ − ψ(F (X)ξ) | X ≤ E T (Y )T F (X)ξ0 − ψ(F (X)ξ0) | X
(51)
Taking expectation with respect to X, we conclude that P mξ ≤ P mξ0 for any ξ. Moreover, if P mξ1 = P mξ0 ,
pr(F (X)ξ1 = F (X)ξ0) = 1, which contradicts the hypothesis (47). Finally, the integrability of (28) follows
from (48).
The conditions 6.2, 6.3 and 6.4 required by van der Vaart’s Theorem 5.23 to derive the asymptotic
distribution of M-estimators are easily verifiable under the integrability assumptions stated in (49) and (50).
The second derivative matrix of P mξ at ξ0 is
Vξ0 = −E F (X)T ∇2 ψ(F (X)ξ0 )F (X) .
(52)
Finally, observe that
h
i
P ṁξ0 ṁTξ0 = E F (X)T T (Y ) − ∇T ψ(F (X)ξ0 ) T (Y )T − ∇ψ(F (X)ξ0) F (X)
= E F (X)T ∇2 ψ(F (X)ξ0)F (X)
to deduce that, according to the general formula for the asymptotic variance of an M-estimator in (42), the
√
asymptotic variance of n(ξbn − ξ0 ) is given by (16).
Lemmas 6.5– 6.10 and Corollary 6.6 are required to prove Proposition 3.2.
1 ×r
and can be written as in (21). Let (S0 , T0 ) ∈ Rkd1 ×d × Rd×q
Lemma 6.5. Assume that β01 ∈ Rkfirst,d
with
d
d×d
S0 T0 = β01 . Then, there exists U ∈ Rd so that S0 = S0 (U) and T0 = T0 (U), with
!
U
S0 (U) :=
and T0 (U) := U −1 B0 .
(53)
A0 U
Proof. Let
S0 =
S01
S02
21
!
with S01 ∈ Rd×d and S02 ∈ R(k1 −d)×d . The matrix S01 T0 is comprised of the first d rows of β01 which are
linearly independent. Then, d = rank(S01 T0 ) ≤ rank(S01 ) ≤ d, hence S01 is invertible. Take U = S01 . From
the expression (21) for β01 , we have
S01 T0 = B0
S02 T0 = A0 B0 .
Thus, T0 = U −1 B0 , and since T0 T0T ∈ Rd×d and rank d,
S02 = A0 B0 T0T (T0 T0T )−1 = A0 B0 B0T U −1 (U −1 B0 B0T U −1 )−1 = A0 U.
1 ×r
Corollary 6.6. Let β01 ∈ Rkfirst,d
and can be written as in (21) . For U in Rd×d
and S0 (U), T0 (U) as defined
d
−1
in (53), the pre-image of ξ0 through g, g (ξ0 ), satisfies
.
(54)
g −1 (ξ0) ∼
= (η 01 , S0 (U), T0 (U), β02 , η02 ) : U ∈ Rd×d
d
Lemma 6.7. Rd×m
with d ≤ m is an open set in Rd×m .
d
Proof. We will show that the complement of Rd×m
is closed. Consider (Tn )n≥1 ∈ Rd×m , each Tn of rank
d
strictly less than d, and assume that Tn converges to T ∈ Rd×m . Note that, |Tn TnT | = 0 for all n ≥ 1, so that
|T T T | is also equal to zero. Hence, rank(T ) < d.
Lemma 6.8. Θ and S are open sets.
Proof. Up to an homeomorphism, Θ and S are equivalent to
Rk1 × Rkd1 ×d × Rd×r
× Rk1 ×(p−r) × H2
d
and Rk1 × R(k1 −d)×d × Rd×r
× Rk1 ×(p−r) × H2 ,
d
respectively, which are products of open sets by Lemma 6.7.
Lemma 6.9. h : S → M is one to one bicontinuous.
22
1 ×q
Proof. It suffices to note that h1 : R(k1 −d)×d × Rd×q
7→ Rkfirst,d
with
d
B
AB
(A, B) →
!
k1 ×q
(k1 −d)×d
satisfies the required properties for h. Let h−1
× Rd×q
with
1 : Rfirst,d 7→ R
d
B
C
!
→ (CB T BB T
−1
, B)
(55)
Then, h−1
1 (h1 (A, B)) = (A, B) and
h1
h−1
1
B
AB
!!
=
B
AB
!
,
and therefore h−1
1 is the inverse of h1 . Thus, h1 is one to one and bicontinuous.
Lemma 6.10. M is open in Ξres
1 ×r
1 ×r
is a
is an open set in Rkd1 ×r , or equivalently, that the complement of Rkfirst,d
Proof. It suffices to show Rkfirst,d
k1 ×r
k1 ×r
be a sequence such that the first d rows are not linearly independent
closed set in Rd . Let {Tn } ⊂ Rd
and Tn converges to T ∈ Rkd1 ×r . Then, if we write
!
Tn1
Tn =
Tn2
with Tn1 ∈ Rd×r and Tn2 ∈ R(k1 −d)×r , we obtain that | Tn1 TnT1 |= 0 for all n. Then | T1 T1T |= 0, where T1
comprises of the first d rows of T , and therefore the first d rows of T are not linearly independent.
Proof of Proposition 3.2. We verify that Conditions 2.2, 2.5 and 2.6 are satisfied.
Condition 2.2: By Lemma 6.8, the set Θ, defined in (20), is open and ξ0 belongs to g(Θ) = Ξres .
Condition 2.5: Since the first d rows of β01 are linearly independent, ξ0 ∈ M. Then, applying Lemma 6.10
obtains that M is an open set in Ξres .
Consider (S, h) as in (23). By Lemma 6.8, S is an open set, and h is one-to-one and bicontinuous by
Lemma 6.9. Furthermore, if
s0 = (η 01 , A0 , B0 , β02 , η02 ) ∈ S,
23
where A0 y B0 are given in (21), then h(s0 ) = (η 01 , β01 , β02 , η02 ) = ξ0 .
The function h is twice continuously differentiable and its Jacobian is full rank. In
Ik 1
0
0
0
!
!
0
Id
0 BT ⊗
I
⊗
0
r
∇h(η 1 , A, B, β2 , η 2 ) =
Ik1 −d
A
0
0
Ik1 (p−r)
0
0
0
0
0
fact, the latter is
0
0
(56)
0
Ik 2
of order (k1 p + k) × (d(k1 + r − d) + k1 (p − r) + k) with full column rank d(k1 + r − d) + k1 (p − r) + k [see
[4]; also it is a direct implication of Theorem 5 in [12]].
Condition 2.6: Consider θ0 ∈ g −1(ξ0 ), associated with U, as shown in Lemma 6.5. Since
∇g(η1 , S, T, β2 , η2 ) =
Ik 1
0 TT
0
0
0
0
0
0
⊗ Ik 1 Ir ⊗ S
0
0
0
0
Ik1 (p−r) 0
0
0
0
Ik 2
then,
∇g η 10 ,
where
U
A0 U
Ik 1
0
=
0
0
G=
!
, U −1 B0 , β20 , η20
0
B0T
0
Id
A0
⊗ Ik 1 Ir ⊗
0
0
0
0
!
!
0
0
Ik1 (p−r)
0
0
0
G.
0
Ik 2
Ik 1
0
0
0
0
0 U −T ⊗ Ik1
0
0
0
0
0
(Ir ⊗ U)
0
0
0
0
0
Ik1 (p−r) 0
0
0
0
0
Ik 2
24
(57)
Since G is invertible of order (d(k1 + r) + k1 (p − r) + k) × (d(k1 + r) + k1 (p − r) + k),
Ik 1
0
0
0
0
!
I
d
T
0 B0 ⊗ Ik1 Ir ⊗
0
0
.
span∇g(θ0 ) = span
A0
0
0
Ik1 (p−r) 0
0
0
0
0
0
Ik 2
e with
Next, since ∇h(s0 ) = ∇g(θ0 )G−1 I,
Ie = diag
Ik 1 Id ⊗
0
Ik1 −d
!
Ird Ik1 (p−r) Ik2
!
we have span∇h(s0 ) ⊂ span∇g(θ0 ). Applying again Theorem 5 in [12] obtains rank(∇g(θ0 )) = d(k1 + r −
d) + k1 (p − r) + k = rank(∇h(s0 )). Therefore, span∇h(s0 ) = span∇g(θ0 ).
Proof of Theorem 3.3. In Theorem 3.1 we verified that the conditions of Theorem 5.23 in [15] are satisfied
for the maximum likelihood estimator under model (12) satisfying (13). The asymptotic variance of the MLE
estimator is given in (16). We also showed Conditions 2.2, 2.5 and 2.6 are satisfied in the proof of Proposition
3.2. The result follows from Proposition 2.7 since −Vξ0 = Wξ−1
and
0
√
avar{ n(ξbnres − ξ0 )} = Πξ0 (W −1 ) Wξ0 ΠTξ (W −1 )
ξ0
= ∇g(θ0 )(∇g(θ0 )
= Πξ0 (W −1 ) Wξ0
0
T
ξ0
Wξ−1
∇g(θ0 ))† ∇g(θ0 )T
0
ξ0
References
[1] Anderson, T. W. (1951). Estimating linear restrictions on regression coefficients for multivariate normal
distributions. The Annals of Mathematical Statistics, 22, 327–351.
[2] Bura, E., Duarte, S. and Forzani, L. (2016). Sufficient Reductions in Regressions With Exponential Family
Inverse Predictors. Journal of the American Statistical Association 111, 1313–1329.
[3] Bura, E. and Yang, J. (2011). Dimension estimation in sufficient dimension reduction: a unifying approach.
J. Multivariate Anal., 102, 130–142.
25
[4] Cook, R. D. and Ni, L. (2005). Sufficient dimension reduction via inverse regression: a minimum discrepancy approach. Journal of the American Statistical Association, 100, 410–428.
[5] Duarte, S. L. (2016). Modelos lineales generalizados: regresión de rango reducido y reducción suficiente de
dimensiones (Tesis Doctoral). Universidad Nacional del Litoral. Argentina.
[6] Geyer, C. J. (1994). On the asymptotics of constrained M-estimation. The Annals of Statistics, 22, 1993–
2010.
[7] Haberman, S. J. (1989). Concavity and estimation. The Annals of Statistics, 17, 1631–1661.
[8] Hjort, N. L. and Pollard, D. (2011). Asymptotics for minimisers of convex processes. arXiv:1107.3806.
[9] Huber, P. J. (1967). The behavior of maximum likelihood estimates under nonstandard conditions. In
Proceedings of the fifth Berkeley symposium on mathematical statistics and probability (Vol. 1, No. 1, pp.
221–233).
[10] Izenman, A. J. (1975). Reduced-rank regression for the multivariate linear model. Journal of Multivariate
Analysis, 5, 248–264.
[11] Izenman, A. J. (2008). Modern Multivariate. Statistical Techniques: Regression, Classification and Manifold Learning. New York: Springer.
[12] Puntanen, S., Styan, G. P., and Isotalo, J. (2011). Matrix tricks for linear statistical models: our personal
top twenty. Springer Science and Business Media.
[13] Reinsel, G. C. and Velu, R. P. (1998). Multivariate Reduced rank regression. Lecture Notes in Statistics.
New York: Springer.
[14] Niemiro, W. (1992). Asymptotics for M-estimators defined by convex minimization. The Annals of Statistics, 20, 1514–1533.
[15] Van der Vaart, A. W. (2000). Asymptotic statistics. Cambridge University Press.
[16] Yee, T. W. (2015). Vector generalized linear and additive models. New York, USA: Springer.
[17] Yee, T. W. (2017). VGAM: Vector Generalized Linear and Additive Models. R package version 1.0-3.
URL https://CRAN.R-project.org/package=VGAM
[18] Yee, T. W. and Hastie, T. J. (2003). Reduced-rank Vector Generalized Linear Models. Statistical Modelling, 3, 15–41.
[19] Yee, T. W. and Wild, C. J. (1996). Vector Generalized Additive Models. Journal of Royal Statistical
Society, Series B, 58(3), 481–493.
26
| 10math.ST
|
On the Scalability of the GPU EXPLORE
Explicit-State Model Checker
Nathan Cassee
Thomas Neele
Anton Wijs∗
Eindhoven University of Technology
Eindhoven, The Netherlands
[email protected] {T.S.Neele, A.J.Wijs}@tue.nl
The use of graphics processors (GPUs) is a promising approach to speed up model checking to
such an extent that it becomes feasible to instantly verify software systems during development.
GPU EXPLORE is an explicit-state model checker that runs all its computations on the GPU. Over
the years it has been extended with various techniques, and the possibilities to further improve its
performance have been continuously investigated. In this paper, we discuss how the hash table of
the tool works, which is at the heart of its functionality. We propose an alteration of the hash table
that in isolated experiments seems promising, and analyse its effect when integrated in the tool.
Furthermore, we investigate the current scalability of GPU EXPLORE, by experimenting both with
input models of varying sizes and running the tool on one of the latest GPUs of NVIDIA.
1
Introduction
Model checking [2] is a technique to systematically determine whether a concurrent system adheres to
desirable functional properties. There are numerous examples in which it has been successfully applied,
however, the fact that it is computationally very demanding means that it is not yet a commonly used procedure in software engineering. Accelerating these computations with graphics processing units (GPUs)
is one promising way to model check a system design in mere seconds or minutes as opposed to many
hours.
GPU EXPLORE [28, 29, 33] is a model checker that performs all its computations on a GPU. Initially,
it consisted of a state space exploration engine [28], which was extended to perform on-the-fly deadlock
and safety checking [29]. Checking liveness properties has also been investigated [27], with positive
results, but liveness checking has yet to be integrated in the official release of the tool. Finally, in order
to reduce the memory requirements of GPU EXPLORE, partial order reduction has been successfully
integrated [20].
Since the first results achieved with GPU EXPLORE [28], considerable progress has been made. For
instance, the original version running on an NVIDIA K20 was able to explore the state space of the
peterson7 model in approximately 72 minutes. With many improvements to GPU EXPLORE’s algorithms, reported in [33], the GPU hardware and the CUDA compiler, this has been reduced to 16 seconds.
With these levels of speed-up, it has become much more feasible to interactively check and debug large
models. Furthermore, GPU developments continue and many options can still be investigated.
Performance is very important for a tool such as GPU EXPLORE. However, so far, the scalability of
the tool has not yet been thoroughly investigated. For instance, currently, we have access to a NVIDIA
Titan X GPU, which is equipped with 12 GB global memory, but for all the models we have been using
∗ We gratefully acknowledge the support of NVIDIA Corporation with the donation of the GeForce Titan X used for this
research.
Timo Kehrer and Alice Miller (Eds.):
Workshop on Graphs as Models 2017 (GAM ’17)
EPTCS 263, 2017, pp. 38–52, doi:10.4204/EPTCS.263.4
c N.W. Cassee, T.S. Neele & A.J. Wijs
N.W. Cassee, T.S. Neele & A.J. Wijs
39
so far, 5 GB of global memory suffices as the models used for the run-time analysis of GPU EXPLORE
do not require more than 5GB for a state space exploration. In the current paper, we report on results we
have obtained when scaling up models to utilise up to 12 GB.
In addition, we also experimentally compared running GPU EXPLORE on a Titan X GPU with the
Maxwell architecture, which was released on 2015, with GPU EXPLORE running on a Titan X GPU with
the Pascal architecture, which was released a year later. This provides insights regarding the effect recent
hardware developments have on the tool.
Finally, we analyse the scalability of a critical part of the tool, namely its hash table. This structure
is used during model checking to keep track of the progress made through the system state space. Even
a small improvement of the hash table design may result in a drastic improvement in the performance of
GPU EXPLORE. Recently, we identified, by conducting isolated experiments, that there is still potential
for further improvement [9]. In the current paper, we particularly investigate whether changing the size
of the so-called buckets, i.e., data structures that act as containers in which states are stored, can have a
positive effect on the running time.
The structure of the paper is as follows. In Section 2, we discuss related work. Next, an overview of
the inner working of GPU EXPLORE is presented in Section 3. The hash table and its proposed alterations
are discussed in Section 4. After that, we present the results we obtained through experimentation in
Section 5. Finally, conclusions and pointers to future work are given in Section 6.
2
Related work
In the literature, several different designs for parallel hash tables can be found. First of all, there is the
hash table for GPUs proposed by Alcantara et al. [1], which is based on Cuckoo hashing. Secondly,
Laarman et al. [24] designed a hash table for multi-core shared memory systems. Their implementation
was later used as a basis for the hash table underlying the LTS MIN model checker. Other lock-free hash
tables for the GPU are those proposed by Moazeni & Sarrafzadeh [19], by Bordawekar [6] and by Misra
& Chaudhuri [18]. Cuckoo hashing as implemented by Alcantara et al. is publicly available as part of the
CUDPP library 1 . Unfortunately, to the best of our knowledge, there are no implementations available of
the other hash table designs.
Besides GPU EXPLORE [33], there are several other GPU model checking tools. Bartocci et al. [5]
developed an extension for the SPIN model checker that performs state-space exploration on the GPU.
They achieved significant speed-ups for large models.
A GPU extension to the parallel model checking tool DIVINE, called DIVINE-CUDA, was developed by Barnat et al. [4]. To speed-up the model checking process, they offload the cycle detection
procedure to the GPU. Their tool can even benefit from the use of multiple GPUs. DIVINE-CUDA
achieves a significant speed-up when model checking properties that are valid.
Edelkamp and Sulewski address the issues arising from the limited amount of global memory available on a GPU. In [12], they implement a hybrid approach in a tool called C U DM O C, using the GPU
for next state computation, while keeping the hash table in the main memory, to be accessed by multiple
threads running on the Central Processing Unit (CPU). In [13], they keep part of the state space in the
global memory and store the rest on disk. The record on disk can be queried through a process they
call delayed duplicate detection. Even though disk access causes overhead, they manage to achieve a
speed-up over single-threaded tools.
1 http://cudpp.github.io/
40
On the Scalability of the GPU EXPLORE Explicit-State Model Checker
consumers
producer
rec
p0
gen work
c0
send
p1
c1
work
c1
work
τ
rec
c0
τ
par using
send * rec * _ -> trans,
send * _ * rec -> trans
in
"producer.aut"
||
"consumer.aut"
||
"consumer.aut"
end par
Figure 1: Example of LTS network with one producer and two consumers. On the right, the communication between the LTSs is specified using the EXP syntax [14]. Here, producer.aut and consumer.aut
are files containing the specification of the producer and the consumer respectively.
Edelkamp et al. [7, 8] also applied GPU programming to probabilistic model checking. They solve
systems of linear equations on the GPU in order to accelerate the value iteration procedure. GPUs are
well suited for this purpose, and can enable a speed-up of 18 times over a traditional CPU implementation.
Wu et al. [35] have developed a tool called GPURC that performs the full state-space exploration
process on the GPU, similar to GPU EXPLORE. Their implementation applies dynamic parallelism, a
relatively new feature in CUDA that allows launching of new kernels from within a running kernel. Their
tool shows a good speed-up compared to traditional single-threaded tools, although the added benefit of
dynamic parallelism is minimal.
Finally, GPUs are also successfully applied to accelerate other computations related to model checking. For instance, Wu et al. [34] use the GPU to construct counter-examples, and state space decomposition and minimisation are investigated in [26, 31, 32]. For probabilistic model checking, Češka et al. [10]
implemented GPU accelerated parameter synthesis for parameterized continous time Markov chains.
3
GPUs and GPU EXPLORE
GPU EXPLORE [28, 29, 33] is an explicit-state model checker that practically runs entirely on a GPU
(only the general progress is checked on the host side, i.e. by a thread running on the CPU). It is written
in CUDA C, an extension of C offered by N VIDIA. CUDA (Compute Unified Device Architecture) provides an interface to write applications for N VIDIA’s GPUs. GPU EXPLORE takes a network of Labelled
Transition Systems (LTSs) [17] as input, and can construct the synchronous product of those LTSs using
many threads in a Breadth-First-Search-based exploration, while optionally checking on-the-fly for the
presence of deadlocks and violations of safety properties. A (negation of a) safety property can be added
as an automaton to the network.
An LTS is a directed graph in which the nodes represent states and the edges are transitions between
the states. Each transition has an action label representing an event leading from one state to another.
An example network is shown in Figure 1, where the initial states are indicated by detached incoming
arrows. One producer generates work and sends it to one of two consumers. This happens by means of
synchronisation of the ‘send’ and ‘rec’ actions. The other actions can be executed independently. How
the process LTSs should be combined using the relevant synchronisation rules is defined on the right in
Figure 1, using the syntax of the E XP. OPEN tool [17]. The state space of this network consists of 8 states
and 24 transitions.
The general approach of GPU EXPLORE to perform state space exploration is discussed in this sec-
N.W. Cassee, T.S. Neele & A.J. Wijs
41
SM 0
threads (exploration)
SM N
threads (exploration)
shared mem. (cache)
shared mem. (cache)
L1 & L2 cache
texture cache
global memory
(network input)
(global hash table)
Figure 2: Schematic overview of the GPU hardware architecture and GPU EXPLORE
tion, leaving out many of the details that are not relevant for understanding the current work. The
interested reader is referred to [28, 29, 33].
In a CUDA program, the host launches CUDA functions called kernels, that are to be executed many
times in parallel by a specified number of GPU threads. Usually, all threads run the same kernel using
different parts of the input data, although some GPUs allow multiple different kernels to be executed
simultaneously (GPU EXPLORE does not use this feature). Each thread is executed by a streaming processor (SP). Threads are grouped in blocks of a predefined size. Each block is assigned to a streaming
multiprocessor (SM). An SM consists of a fixed number of SPs (see Figure 2).
Each thread has a number of on-chip registers that allow fast access. The threads in a block together
share memory to exchange data, which is located in the (on-chip) shared memory of an SM. Finally, the
blocks can share data using the global memory of the GPU, which is relatively large, but slow, since it
is off-chip. The global memory is used to exchange data between the host and the kernel. The GTX
T ITAN X, which we used for our experiments, has 12 GB global memory and 24 SMs, each having 128
SPs (3,072 SPs in total).
Writing well-performing GPU applications is challenging, due to the execution model of GPUs,
which is Single Instruction Multiple Threads. Threads are partitioned in groups of 32 called warps. The
threads in a warp run in lock-step, sharing a program counter, so they always execute the same program
instruction. Hence, thread divergence, i.e. the phenomenon of threads being forced to execute different
instructions (e.g., due to if-then-else constructions) or to access physically distant parts of the global
memory, negatively affects performance.
Model checking tends to introduce divergences frequently, as it requires combining the behaviour
of the processes in the network, and accessing and storing state vectors of the system state space in
the global memory. In GPU EXPLORE, this is mitigated by combining relevant network information as
much as possible in 32-bit integers, and storing these as textures, that only allow read access and use a
dedicated cache to speed up random accesses.
Furthermore, in the global memory, a hash table is used to store state vectors (Figure 2). The currently
used hash table has been designed to optimise accesses of entire warps: the space is partitioned into
buckets consisting of 32 integers, precisely enough for one warp to fetch a bucket with one combined
memory access. State vectors are hashed to buckets, and placed within a bucket in an available slot. If
the bucket is full, another hash function is used to find a new bucket. Each block accesses the global hash
table to collect vectors that still require exploration.
To each state vector with n process states, a group of n threads is assigned to construct its successors
using fine-grained parallelism. Since access to the global memory is slow, each block uses a dedicated
42
On the Scalability of the GPU EXPLORE Explicit-State Model Checker
state cache (Figure 2). It serves to store and collect newly produced state vectors, that are subsequently
moved to the global hash table in batches. With the cache, block-local duplicates can be detected.
4
The GPU EXPLORE Hash Table
States discovered during the search exploration phase of GPU EXPLORE are inserted into a global memory hash table. This hash table is used to keep track of the open and closed sets maintained during the
breadth first search based exploration of the state space.
Since many accesses (reads and writes) to the hash table are performed during state-space exploration, its performance is critical for our model checker. In order to allow for efficient parallel access,
the hash table should be lock-free. To prevent corruption of state vectors, insertion should be an atomic
operation, even when a state vector spans multiple 32 bit integers.
Given these requirements, we have considered several lock-free hash table implementations. One of
them, proposed by Alcantara et al. [1], uses so-called Cuckoo hashing.
With Cuckoo hashing a key is hashed to a bucket in the hash table, and in case of a collision the key
that is already in the bucket is evicted and rehashed using another hash function to a different bucket.
Re-insertions continue until the last evicted key is hashed to an empty bucket, until all hash functions are
exhausted or until the chain of re-insertions becomes too long [22].
The other hash table we considered is the one originally designed for GPU EXPLORE [33]; we refer to
its hashing mechanism as GPU EXPLORE hashing. We experimentally compared these two hash tables,
and from these comparisons we concluded that while Cuckoo hashing on average performs better, it does
not meet all the demands needed by GPU EXPLORE. However, based on the performance evaluation a
possible performance increase has been identified for GPU EXPLORE hashing [9]. This section discusses
the proposed modification, and its implementation.
4.1
GPU EXPLORE Hashing
Figure 3: Division of threads in a warp over a bucket of size 32 in GPU EXPLORE 2.0
The GPU EXPLORE hash table consists of two parts: the storage used to store the discovered vectors and
the set of hash constants used by the hash function to determine the position of a vector in the hash table.
The memory used to store vectors is divided into buckets with a size of 32 integers. Eachjbucket is addik
tionally split into two equally sized half buckets. Therefore a single bucket can store 2 ∗ vector16length
vectors. The reason for writing vectors to buckets with half-warps (a group of 16 threads) is that in
many cases, atomic writes of half-warps are scheduled in an uninterrupted sequence [33]. This results in
vectors consisting of multiple integers to be written without other write operations corrupting them. It
should be noted that the GPU EXPLORE hash table uses closed hashing, i.e., the vectors themselves are
stored in the hash table, as opposed to pointers to those vectors.
N.W. Cassee, T.S. Neele & A.J. Wijs
43
When inserting, a warp of 32 threads inserts a single vector into a bucket. The way threads are divided
over a bucket can be observed in Figure 3, this figure visualizes a single bucket of the GPU EXPLORE
hash table for a vector length of 3. Each thread in a warp is assigned to one integer of the bucket, and to
one integer of the vector. This assignment is done from left to right per half bucket. For this example the
first 3 threads, i.e., the first vector group, is assigned to the first slot in the bucket, and the first thread in
a vector group is assigned to the first integer of the vector. By assigning every thread to a single part of
the vector and of the bucket each thread has a relatively simple task, which can be executed in parallel.
The insertion algorithm first hashes the vector to insert to determine the bucket belonging to the
vector. Each thread checks its place in the bucket and compares the integer on that position to the
corresponding integer of the vector. After comparing, the insertion algorithm then uses CUDA warp
instructions to quickly exchange data between all 32 threads in the warp to determine whether a single vector group of threads has found the vector. If the vector has been found the insertion algorithm
terminates, if the vector has not been found the algorithm continues.
Figure 4: Example of inserting the vector (8, 2, 1) into the GPU EXPLORE 2.0 hash table.
Figure 4 shows an example of the vector (8, 2, 1) being inserted into the GPU EXPLORE hash table
where the first two slots are already occupied. Because the first two slots are already occupied the first
six threads compare their element of the vector to their element of the bucket. The icons above the arrows
indicate the result of these comparisons. As can be seen there is one thread that has a match, however,
because not all elements in the slot match the insertion algorithm does not report a find.
If the vector is not found in the bucket, the insertion algorithm selects the first free slot in the bucket,
in this case the third slot. This selection procedure can be done efficiently using CUDA warp instructions.
Next, the associated threads attempt to insert the vector (8, 2, 1) into the selected slot using a compare
and swap operation. If the insertion fails because another warp had claimed that slot already for another
vector, the algorithm takes the next free slot in the bucket.
If a bucket has no more free slots the next set of hash constants is used, and the next bucket is
probed. This is repeated until all hash constants have been used, and if no insertion succeeds into any of
the buckets, the insertion algorithm reports a failure and the exploration stops.
In GPU EXPLORE 2.0 the hash table is initialized using eight hash functions. After each exploration
step, blocks that have found new vectors use the insertion algorithm to insert any new vectors they found
into the global hash table. The hash table is therefore a vital part of GPU EXPLORE.
Buckets with a length of 32 integers have been chosen because of the fact that warps in CUDA
consist of 32 threads. This way, every integer in a bucket can be assigned to a single thread. Besides, this
design choice also allows for coalesced memory access: when a warp accesses a continuous block of 32
integers, this operation can be executed in a single memory fetch operation [21]. Uncoalesced accesses,
on the other hand, have to be done individually after each other. By coalescing the accesses, the available
bandwidth to global memory is used efficiently.
44
4.2
On the Scalability of the GPU EXPLORE Explicit-State Model Checker
Configurable Bucket Size
While the current implementation of the hash table makes it possible for GPU EXPLORE to achieve a
considerable increase in performance over CPU-based explicit-state model checkers [33], it suffers from
one disadvantage. Namely, after initially scanning the bucket, only x threads, where x is the vector
length, are active at a time. The other 32 − x threads are inactive while they await the result of the atomic
insertion of the active group. If the insertion fails but there is still a free slot in the hash table, another
group of x threads becomes active to attempt an atomic insertion, while the remaining 32 − x threads
again await the result of this operation.
Figure 5: Division of threads in a warp over buckets of size 8
Therefore, for buckets with a lot of free slots where insertions fail the majority of threads in the
warp are inactive. Furthermore, the vector size generally does not exceed four integers, which means
that when attempting to atomically insert a vector, the majority of threads in a warp is inactive. Ergo,
one possible improvement over the GPU EXPLORE hash table is to reduce the bucket size, so that there
are fewer slots per bucket, and therefore, fewer threads are needed to insert a single vector. As a single
insertion still uses one thread per integer in the bucket, in turn more vectors can be inserted in parallel.
Figure 5 shows what the division of threads over a warp looks like if buckets of size 8 instead of 32
are used. As can be observed, a warp can insert four elements in parallel, as in this diagram each group
of 8 threads inserts a different vector and accesses different buckets in global memory.
The logical consequence of this improvement is that after scanning a bucket, fewer threads are inactive while the vector is being inserted into a free slot. If we suppose that the vector size for a certain
model is 3, and that the new bucket size is 8, then while inserting a vector using the GPU EXPLORE hash
table 32 − 3 = 29 threads are inactive. On the other hand, if four buckets of size 8 are simultaneously
accessed, only 32 − 3 · 4 = 20 threads are inactive, and four vectors are simultaneously processed, as
opposed to only one.
However, while more threads can be active at the same time, smaller buckets also lead to thread
divergence within a warp. First of all, of course, accessing different buckets simultaneously likely leads
to uncoalesced memory accesses. Furthermore, it is also possible that in an insertion procedure, one
group needs to do more work than another in the same warp. For instance, consider that the first group in
the warp fails to find its vector in the designated bucket, and also cannot write it to the bucket since the
latter is full. In that case, the group needs to fetch another bucket. At the same time, another group in the
warp may find its vector in the designated bucket, and therefore be able to stop the insertion procedure.
In such a situation, these two groups will diverge, and the second group will have to wait until the first
group has finished inserting. This means that the use of smaller buckets can only be advantageous if the
performance increase of the smaller buckets outweighs the performance penalty of divergence. In this
paper, we address whether this is true or not in practical explorations.
The suggested performance increase has been experimentally validated by comparing an implementation of the hash table with varying bucket size to the original GPU EXPLORE 2.0 hash table, both in
N.W. Cassee, T.S. Neele & A.J. Wijs
45
GH
G H - CBS
2,500
Runtime (ms)
2,000
1,500
1,000
500
0
0
10
20
30
40
50
60
70
80
90
100
Duplication in sequence
Figure 6: Results of inserting a sequence of 100,000,000 randomly generated integers into the hash tables
of GPU EXPLORE (G H) and GPU EXPLORE with configurable buckets (G H - CBS). For the bucketed
version a bucket size of four integers has been used.
isolation and as a part of GPU EXPLORE. The results of this comparison are presented and discussed in
section 5.
5
Experiments
In this section we discuss our experiments to evaluate the scalability of GPU EXPLORE. In Section 5.1,
we report on experiments to evaluate the effect of the bucket size on the runtimes of GPU EXPLORE. In
the next two sections, our goal is to determine whether GPU EXPLORE scales well when varying the size
of the model and the performance of the hardware, respectively. For most of the experiments, we use an
NVIDIA Titan X (Maxwell) installed on a machine running L INUX M INT 17.2. The number of blocks
is set to 6,144, with 512 threads per block. The hash table is allocated 5GB in global memory.
For our benchmarks, we selected a varied set of models from the CADP toolset [14], the mCRL2
toolset [11] and the BEEM database [23]. The models with a .1 suffix have been modified to obtain a
larger state space.
5.1
Varying the Bucket Size
To test different bucket sizes two types of experiments have been performed. First, the hash table with
varying bucket sizes has been tested in isolation, where the time taken to insert 100,000,000 elements
has been measured. Second, GPU EXPLORE has been modified such that it uses a hash table with modifiable bucket size. This bucket size can be set at compile time. The performance of this version of
GPU EXPLORE has been compared to the original GPU EXPLORE w.r.t. the exploration of several input
models.
The input data for the performance evaluation of the hash tables in isolation is a set of sequences of
46
On the Scalability of the GPU EXPLORE Explicit-State Model Checker
randomly generated integers, each sequence consisting of 100,000,000 vectors with a length of 1 integer.
The sequences vary in how often an element occurs in the sequence. A duplication of 1 means that every
unique element occurs once in the sequence, and a duplication of 100 means that each unique element
in the sequence occurs 100 times. Therefore a sequence with a duplication of 100 has 100,000,000
unique
100
elements. Note that this experiment tries to replicate state space exploration, where many duplicate
insertions are performed when the fan-in is high, i.e., many transitions in the state space lead to the same
state. So this experiment replicates the performance of the hash table for models that vary in the average
fan-in of the states. For each sequence, we measured the total time required to insert all elements.
The results of this comparison are depicted in Figure 6. We refer to the standard GPU EXPLORE
2.0 hash table as G H and the hash table with configurable bucket size as G H - CBS. In the experiment,
G H - CBS with a bucket size of four integers has been compared to G H. G H - CBS is slower for sequences
where each element only occurs a few times. However, for sequences with a higher duplication degree
G H - CBS starts to outperform G H. After all, for the sequences with more duplication, less time is spent
on atomically inserting elements, since most elements are already in the hash table. G H - CBS performs
up to three times better than G H. We will see later why the amount of duplication is relevant for model
checking.
In addition to testing the performance of the hash tables in isolation, the performance of G H - CBS
has been compared with standard GPU EXPLORE hashing as a part of GPU EXPLORE 2.0. The hash
table underlying GPU EXPLORE, as implemented by Wijs, Neele and Bošnački [33], has been modified
to allow configuration of the bucket size at compile time. We compared the time required for state
space exploration of our implementation, GPU EXPLORE with configurable bucket size, with the original
implementation of GPU EXPLORE 2.0 [33].
Four bucket sizes have been compared to GPU EXPLORE 2.0, namely bucket sizes 4, 8, 16 and 32
integers. The relative performance with these bucket sizes has been recorded with the performance of
GPU EXPLORE 2.0 as baseline. For each model the experiment has been run five times and the average
running time of these five runs has been taken.
The result of these comparisons is illustrated in Figure 7. As can be observed the total exploration
time of GPU EXPLORE with configurable bucket size is for most models larger than the runtime of GPUEXPLORE 2.0. Only szymanski5 and lann7 show a small performance increase for bucket size 4. For the
other instances, however, the new hash table causes a slow down of up to 30%.
There are three reasons why the promising performance shown in the previous experiments is not
reflected here. First, the increased complexity of the hash table negatively affects register pressure, i.e.,
the number of registers each thread requires to execute the kernel. When the register usage in a kernel
increases, the compiler may temporarily place register contents in global memory. This effect is not
observed when the hash table is tested in isolation, as in that case, far fewer registers per thread are
required. The increase in register pressure is also the reason that GPU EXPLORE with a configurable
bucket size set to 32 is slower than GPU EXPLORE with a static bucket size of 32.
Furthermore, smaller bucket sizes result in more thread divergence and more uncoalesced memory
accesses when reading from the hash table. Therefore, the available memory bandwidth is used less
efficiently, leading to a drop in performance. Apparently, the increased potential for parallel insertions
of vectors cannot overcome this drawback.
Lastly, while exploring the state-space, GPU EXPLORE only discovers duplicates if those states have
several incoming transitions. On average, the models used for the experiments have a fan-in of 4 to
6, with some exceptions that have a higher fan-in of around 8 to 11. However, from Figure 6 it can
be concluded that the hash table in isolation only starts to outperform the static hash table when each
element is duplicated 21 times. This partly explains the performance seen in Figure 7.
Runtime relative to static bucket size of 32
N.W. Cassee, T.S. Neele & A.J. Wijs
47
wafer stepper.1
odp.1
1394.1
asyn3
lamport8
des
szymanski5
lann6
lann7
asyn3.1
1.3
1.2
1.1
1
4
8
16
32
Bucket size
Figure 7: Relative runtime of GPU EXPLORE 2.0 with variable bucket size. The runtime GPU EXPLORE
2.0 with a fixed bucket size of 32 integers is used as a reference and is normalized to 1.
5.2
Varying the Model Size
In addition to experimentally comparing the effect of different bucket sizes, we also investigated how
GPU EXPLORE 2.0 behaves when exploring state spaces of different size. We performed this experiment
with two different models. The first is a version of the Gas Station model [15] where the number of
pumps is fixed to two. We varied the number of customers between two and twelve. None of these
instances requires a state vector longer than two 32-bit integers.
The other model is a simple implementation of a ring-structured network, where one token is continuously passed forward between the nodes. Each node has two transitions to communicate with its
neighbours and a further three internal transitions. Here, we varied the number of nodes in the network.
We executed the tool five times on each instance and computed the average runtime. The results are
listed in Table 1. For the smallest instances, the performance of GPU EXPLORE (measured in states/sec)
is a lot worse compared to the larger instances. This has two main reasons. First of all, the relative
overhead suffered from initialization and work scanning is higher. Second, the parallelism offered by
the GPU cannot be fully exploited, because the size of one search layer is too small to occupy all blocks
with work.
For the gas station model, peak performance is achieved for the instance with ten customers, which
has 60 million states. For larger instances, the performance decreases slightly due to the increasing
occupancy of the hash table. This leads to more hash collisions, therefore more time is lost on rehashing.
The results of the token ring model show another interesting scalability aspect. There is a performance drop between the instances with 10 nodes and 11 nodes. This is caused by the fact that the
instance with 11 nodes is the smallest for which the state vector exceeds 32 bits in length. Longer state
vectors lead to more memory accesses throughout the state-space generation algorithm.
48
On the Scalability of the GPU EXPLORE Explicit-State Model Checker
Table 1: Performance of GPU EXPLORE for the Gas Station and the Token Ring model while varying the
amount of processes.
Token Ring
Gas Station
5.3
N
states
time (s)
states/sec
2
3
4
5
6
7
8
9
10
11
12
165
1,197
7,209
38,313
186,381
849,285
3,680,721
15,333,057
61,863,669
243,104,733
934,450,425
0.016
0.023
0.035
0.062
0.209
0.718
1.235
3.093
11.229
44.534
178.817
10,294
51,004
206,812
621,989
892,039
1,183,246
2,981,535
4,957,437
5,509,307
5,458,810
5,225,726
N
states
time (s)
states/sec
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
12
54
216
810
2,916
10,206
34,992
118,098
393,660
1,299,078
4,251,528
13,817,466
44,641,044
143,489,070
459,165,024
0.027
0.047
0.067
0.086
0.113
0.180
0.275
0.488
1.087
6.394
8.345
8.138
22.060
68.233
215.889
449
1,138
3,212
9,402
25,702
56,571
127,142
242,225
362,294
203,159
509,462
1,697,864
2,023,649
2,102,934
2,126,853
Speed-up Due to the Pascal Architecture
The Titan X we used for most of the benchmarks is based on the Maxwell architecture, and was launched
in 2015. Since then, NVIDIA has released several other high-end GPUs. Most aspects have been improved: the architecture has been revised, there are more CUDA cores on the GPU and there is more
global memory available. To investigate how well GPU EXPLORE scales with faster hardware, we performed several experiments with a Titan X with the Maxwell architecture (TXM) and with a Titan X
with the Pascal architecture (TXP). The latter was released in 2016, and the one we used is installed in
the DAS -5 cluster [3] on a node running C ENT OS L INUX 7.2.
The TXM experiments were performed with 6,144 blocks, while for the TXP, GPU EXPLORE was
set to use 14,336 blocks. The improvements of the hardware allows for GPU EXPLORE to launch more
blocks. To evaluate the speed-ups compared to a single-core CPU approach, we also conducted experiments with the G ENERATOR tool of the latest version (2017-i) of the C ADP toolbox [14]. These have
been performed on nodes of the DAS -5 cluster, which are equipped with an I NTEL H ASWELL E5-2630v3 2.4 GHz CPU, 64 GB memory, and C ENT OS L INUX 7.2.
The results are listed in Table 2. The reported runtimes are averages after having run the corresponding tool ten times. For most of the larger models, we see a speed-up of about 2 times when running
GPU EXPLORE on a TXP compared to running it on a TXM. The average speed-up is 1.73. This indicates that GPU EXPLORE scales well with a higher memory bandwidth and a larger amount of CUDA
cores.
Comparing GPU EXPLORE on a TXP with single-core CPU analysis, the average speed-up is 183.91,
and if we only take state spaces into account consisting of at least 10 million states, the average speedup is 280.81. Considering that with multi-core model checking, linear speed-ups can be achieved [16],
this roughly amounts to using 180 and 280 CPU cores, respectively. This, in combination with the
N.W. Cassee, T.S. Neele & A.J. Wijs
49
Table 2: Performance comparison of single-core G ENERATOR of C ADP (G EN) and GPU EXPLORE running on a Titan X with the Maxwell architecture (TXM) and with the Pascal architecture (TXP).
runtime (seconds)
model
acs
odp
1394
acs.1
transit
wafer stepper.1
odp.1
1394.1
asyn3
lamport8
des
szymanski5
peterson7
lann6
lann7
asyn3.1
speed-ups
states
G EN
TXM
TXP
G EN-TXP
TXM-TXP
4,764
91,394
198,692
200,317
3,763,192
3,772,753
7,699,456
10,138,812
15,688,570
62,669,317
64,498,297
79,518,740
142,471,098
144,151,629
160,025,986
190,208,728
4.17
3.26
2.81
5.30
34.36
22.95
65.50
82.71
358.58
1048.13
477.43
1516.71
3741.87
2751.15
3396.19
4546.84
0.05
0.08
0.20
0.18
0.77
1.01
1.34
1.42
3.15
5.81
12.34
7.48
31.60
10.57
16.67
31.03
0.06
0.05
0.15
0.14
0.48
0.51
0.66
0.83
1.98
3.11
6.65
3.90
15.74
5.39
8.41
15.37
71.91
70.76
18.95
37.88
70.99
45.17
99.54
99.66
181.47
336.80
71.84
389.10
237.81
510.80
403.92
295.92
0.89
1.64
1.36
1.27
1.59
2.00
2.03
1.71
1.59
1.87
1.86
1.92
2.01
1.96
1.98
2.02
average
183.91
1.73
observation that frequently, speed-ups over 300 times and once even over 500 times are achieved, clearly
demonstrates the effectiveness of using GPUs for explicit-state model checking.
6
Conclusion and Future Work
In this paper, we have reported on a number of scalability experiments we conducted with the GPU
explicit-state model checker GPU EXPLORE. In earlier work, we identified potential to further improve
its hash table [9]. However, experiments in which we varied the bucket size in GPU EXPLORE provided
the insight that only for very specific input models, and only if the bucket size is set very small (4), some
speed-up becomes noticeable. In the context of the entire computation of GPU EXPLORE, the additional
register use per thread and the introduced uncoalesced memory accesses and thread divergence make it
not worthwile to make the bucket size configurable. This may be different for other applications, as our
experiments with the hash table in isolation point out that hashing can be made more efficient in this way.
Besides this, we have also conducted experiments with models of different sizes. We scaled up a Gas
Station model and a Token Ring model and obtained very encouraging results; for the second model,
GPU EXPLORE can generate up to 2.1 million states per second, and for the first model, at its peak,
GPU EXPLORE is able to generate about 5.5 million states per second, exploring a state space of 934.5
million states in under three minutes. We believe these are very impressive numbers that demonstrate the
potential of GPU model checking.
Finally, we reported on some experiments we conducted with new GPU hardware. The Titan X
50
On the Scalability of the GPU EXPLORE Explicit-State Model Checker
with the Pascal architecture from 2016 provides for our benchmark set of models an average speed-up
of 1.73 w.r.t. the Titan X with the Maxwell architecture from 2015. We also compared the runtimes of
GPU EXPLORE running on the Pascal Titan X with the CPU single-core G ENERATOR tool of the C ADP
toolbox, and measured an average speed-up of 183.91 for the entire benchmark set of models, and of
280.81 for the models yielding a state space of at least 10 million states. Often speed-ups over 300 times
have been observed, and in one case even over 500 times.
Future work For future work, we will consider various possible extensions to the tool. First of all,
the ability to write explored state spaces to disk will open up the possibility to postprocess and further
analyse the state spaces. This could be done directly, or after application of bisimulation reduction on
the GPU [26].
Second of all, we will work on making the tool more user friendly. Currently, providing an input
model is quite cumbersome, since GPU EXPLORE requires a user to express system behaviour in the low
level description formalism of networks of LTSs. Specifying systems would be much more convenient if
a higher-level modelling language would be supported. We will investigate which modelling languages
would be most suitable for integration in the current tool.
Finally, we will also consider the application of GPU EXPLORE to conduct computations similar to
model checking, such as performance analysis [30]. This requires to incorporate time into the input
models, for instance by including special actions to represent the passage of time [25].
References
[1] Dan A. Alcantara, Vasily Volkov, Shubhabrata Sengupta, Michael Mitzenmacher, John D. Owens & Nina
Amenta (2012): Building an Efficient Hash Table on the GPU. In: GPU Computing Gems Jade Edition,
Morgan Kaufmann Publishers Inc., pp. 39–53, doi:10.1016/B978-0-12-385963-1.00004-6.
[2] Christel Baier & Joost-Pieter Katoen (2008): Principles of model checking. MIT Press.
[3] Henri Bal, Dick Epema, Cees de Laat, Rob van Nieuwpoort, John Romein, Frank Seinstra, Cees Snoek &
Harry Wijshoff (2016): A Medium-Scale Distributed System for Computer Science Research: Infrastructure
for the Long Term. IEEE Computer 49(5), pp. 54–63, doi:10.1109/MC.2016.127.
[4] Jiřı́ Barnat, Petr Bauch, Luboš Brim & Milan Češka (2012): Designing Fast LTL Model Checking Algorithms
for Many-Core GPUs. JPDC 72(9), pp. 1083–1097, doi:10.1016/j.jpdc.2011.10.015.
[5] Ezio Bartocci, Richard DeFrancisco & Scott A. Smolka (2014): Towards a GPGPU-parallel SPIN Model
Checker. In: SPIN 2014, ACM, New York, NY, USA, pp. 87–96, doi:10.1145/2632362.2632379.
[6] Rajesh Bordawekar (2014): Evaluation of Parallel Hashing Techniques. Presentation at GTC’14 (last
checked on 17 February 2017). http://on-demand.gputechconf.com/gtc/2014/presentations/
S4507-evaluation-of-parallel-hashing-techniques.pdf.
[7] Dragan Bošnački, Stefan Edelkamp, Damian Sulewski & Anton Wijs (2011): Parallel Probabilistic Model
Checking on General Purpose Graphics Processors. STTT 13(1), pp. 21–35, doi:10.1007/s10009-010-01764.
[8] Dragan Bošnački, Stefan Edelkamp, Damian Sulewski & Anton Wijs (2010): GPU-PRISM: An Extension of
PRISM for General Purpose Graphics Processing Units. In: PDMC, IEEE, pp. 17–19, doi:10.1109/PDMCHiBi.2010.11.
[9] Nathan Cassee & Anton Wijs (2017): Analysing the Performance of GPU Hash Tables for State Space
Exploration. In: GaM, EPTCS, Open Publishing Association.
N.W. Cassee, T.S. Neele & A.J. Wijs
51
[10] Milan Češka, Petr Pilař, Nicola Paoletti, Luboš Brim & Marta Kwiatkowska (2016): PRISM-PSY: Precise
GPU-Accelerated Parameter Synthesis for Stochastic Systems. In: TACAS, LNCS 9636, Springer, pp. 367–
384, doi:10.1007/978-3-642-54862-8.
[11] Sjoerd Cranen, Jan Friso Groote, Jeroen J. A. Keiren, Frank P. M. Stappers, Erik P. De Vink, Wieger Wesselink & Tim A. C. Willemse (2013): An Overview of the mCRL2 Toolset and Its Recent Advances. In:
TACAS, LNCS 7795, Springer, pp. 199–213, doi:10.1007/978-3-642-36742-7 15.
[12] Stefan Edelkamp & Damian Sulewski (2010): Efficient Explicit-State Model Checking on General Purpose
Graphics Processors. In: SPIN, LNCS 6349, Springer, pp. 106–123, doi:10.1007/978-3-642-16164-3 8.
[13] Stefan Edelkamp & Damian Sulewski (2010): External memory breadth-first search with delayed duplicate
detection on the GPU. In: MoChArt, LNCS 6572, Springer, pp. 12–31, doi:10.1007/978-3-642-20674-0 2.
[14] Hubert Garavel, Frédéric Lang, Radu Mateescu & Wendelin Serwe (2013): CADP 2011: A Toolbox for
the Construction and Analysis of Distributed Processes. STTT 15(2), pp. 89–107, doi:10.1007/978-3-54073368-3 18.
[15] David Heimbold & David Luckham (1985): Debugging Ada Tasking Programs. IEEE Software 2(2), pp.
47–57, doi:10.1109/MS.1985.230351.
[16] Alfons Laarman (2014): Scalable Multi-Core Model Checking.
doi:10.3990/1.9789036536561.
Ph.D. thesis, University of Twente,
[17] Frédéric Lang (2006): Refined Interfaces for Compositional Verification. In: FORTE, LNCS 4229, Springer,
pp. 159–174, doi:10.1007/11888116 13.
[18] Prabhakar Misra & Mainak Chaudhuri (2012): Performance Evaluation of Concurrent Lock-free Data Structures on GPUs. In: ICPADS, pp. 53–60, doi:10.1109/ICPADS.2012.18.
[19] Maryam Moazeni & Majid Sarrafzadeh (2012): Lock-free Hash Table on Graphics Processors. In: SAAHPC,
pp. 133–136, doi:10.1109/SAAHPC.2012.25.
[20] Thomas Neele, Anton Wijs, Dragan Bošnački & Jaco van de Pol (2016): Partial Order Reduction for GPU
Model Checking. In: ATVA, LNCS 9938, Springer, pp. 357–374, doi:10.1007/978-3-319-46520-3 23.
[21] John Nickolls, Ian Buck, Michael Garland & Kevin Skadron (2008): Scalable Parallel Programming with
CUDA. Queue 6(2), pp. 40–53, doi:10.1145/1365490.1365500.
[22] Rasmus Pagh & Flemming Friche Rodler (2001): Cuckoo Hashing. In: ESA, LNCS 2161, Springer, pp.
121–133, doi:10.1007/3-540-44676-1 10.
[23] Radek Pelánek (2007): BEEM: Benchmarks for Explicit Model Checkers. In: SPIN 2007, LNCS 4595, pp.
263–267, doi:10.1007/978-3-540-73370-6 17.
[24] Steven van der Vegt & Alfons Laarman (2011): A Parallel Compact Hash Table. In: MEMICS, LNCS 7119,
Springer, pp. 191–204, doi:10.1007/978-3-642-25929-6 18.
[25] Anton Wijs (2007): Achieving Discrete Relative Timing with Untimed Process Algebra. In: ICECCS, IEEE,
pp. 35–44, doi:10.1109/ICECCS.2007.13.
[26] Anton Wijs (2015): GPU Accelerated Strong and Branching Bisimilarity Checking. In: TACAS, LNCS
9035, Springer, pp. 368–383, doi:10.1007/978-3-662-46681-0 29.
[27] Anton Wijs (2016): BFS-Based Model Checking of Linear-Time Properties With An Application on GPUs.
In: CAV, Part II, LNCS 9780, Springer, pp. 472–493, doi:10.1007/978-3-319-41540-6 26.
[28] Anton Wijs & Dragan Bošnački (2014): GPUexplore: Many-Core On-the-Fly State Space Exploration Using
GPUs. In: TACAS, LNCS 8413, pp. 233–247, doi:10.1007/978-3-642-54862-8 16.
[29] Anton Wijs & Dragan Bošnački (2016): Many-Core On-The-Fly Model Checking of Safety Properties Using
GPUs. STTT 18(2), pp. 169–185, doi:10.1007/s10009-015-0379-9.
[30] Anton Wijs & Wan Fokkink (2005): From χt to µCRL: Combining Performance and Functional Analysis.
In: ICECCS, IEEE, pp. 184–193, doi:10.1109/ICECCS.2005.51.
52
On the Scalability of the GPU EXPLORE Explicit-State Model Checker
[31] Anton Wijs, Joost-Pieter Katoen & Dragan Bošnački (2014): GPU-Based Graph Decomposition into
Strongly Connected and Maximal End Components. In: CAV, LNCS 8559, Springer, pp. 309–325,
doi:10.1007/978-3-319-08867-9 20.
[32] Anton Wijs, Joost-Pieter Katoen & Dragan Bošnački (2016): Efficient GPU Algorithms for Parallel Decomposition of Graphs into Strongly Connected and Maximal End Components. Formal Methods in System
Design 48(3), pp. 274–300, doi:10.1007/s10703-016-0246-7.
[33] Anton Wijs, Thomas Neele & Dragan Bošnački (2016): GPUexplore 2.0: Unleashing GPU Explicit-State
Model Checking. In: FM, LNCS 9995, Springer, pp. 694–701, doi:10.1007/978-3-319-48989-6 42.
[34] Zhimin Wu, Yang Liu, Yun Liang & Jun Sun (2014): GPU Accelerated Counterexample Generation in LTL
Model Checking. In: ICFEM, LNCS 8829, Springer, pp. 413–429, doi:10.1007/978-3-319-11737-9 27.
[35] Zhimin Wu, Yang Liu, Jun Sun, Jianqi Shi & Shengchao Qin (2015): GPU Accelerated On-the-Fly Reachability Checking. In: ICECCS 2015, pp. 100–109, doi:10.1109/ICECCS.2015.21.
| 8cs.DS
|
How to Generate Pseudorandom
Permutations Over Other Groups:
Even-Mansour and Feistel Revisited
Hector B. Hougaard∗
arXiv:1707.01699v2 [cs.CR] 4 Oct 2017
Abstract
Recent results by Alagic and Russell have given some evidence that
the Even-Mansour cipher may be secure against quantum adversaries
with quantum queries, if considered over other groups than (Z/2)n .
This prompts the question as to whether or not other classical schemes
may be generalized to arbitrary groups and whether classical results
still apply to those generalized schemes.
In this paper, we generalize the Even-Mansour cipher and the Feistel cipher. We show that Even and Mansour’s original notions of secrecy are obtained on a one-key, group variant of the Even-Mansour
cipher. We generalize the result by Kilian and Rogaway, that the
Even-Mansour cipher is pseudorandom, to super pseudorandomness,
also in the one-key, group case. Using a Slide Attack we match the
bound found above. After generalizing the Feistel cipher to arbitrary
groups we resolve an open problem of Patel, Ramzan, and Sundaram
by showing that the 3-round Feistel cipher over an arbitrary group is
not super pseudorandom.
Finally, we generalize a result by Gentry and Ramzan showing that
the Even-Mansour cipher can be implemented using the Feistel cipher
as the public permutation. In this last result, we also consider the
one-key case over a group and generalize their bound.
1
Introduction
In [EM97], Even and Mansour introduced and proved security for the DES
inspired block cipher scheme we now call the Even-Mansour (EM) scheme.
Given a public permutation, P , over n-bit strings, with two different, random,
secret, n-bit keys k1 and k2 , a message x ∈ {0, 1}n could be enciphered as
EMkP1 ,k2 (x) = P (x ⊕ k1 ) ⊕ k2 ,
∗
This article is based on work done for my Master’s Thesis at the University of Copenhagen. For more details on the thesis, contact me at [email protected].
1
with an obvious decryption using the inverse public permutation. The scheme
was minimal, in the sense that they needed to XOR a key before and after
the permutation, otherwise the remaining key could easily be found. As an
improvement, Dunkelman, Keller, and Shamir [DKS12] showed that there
was only a need for a single key and the scheme would still retain an indistinguishability from random, i.e. it was pseudorandom. As another consideration of block ciphers, the Feistel cipher construction of Luby and Rackoff
[LR88] showed how to build pseudorandom permutations from pseudorandom functions.
Eventually, Kuwakado and Morii showed that both the EM scheme [KM12]
and the Feistel scheme [KM10] could be broken by quantum adversaries with
quantum queries. Rather than discard these beautiful constructions entirely,
Alagic and Russell [AR17] considered whether it would be possible to define the two-key EM scheme over Abelian groups in order to retain security
against quantum adversaries with quantum queries. What they showed was
a security reduction to the Hidden Shift Problem, over certain groups, such
as Z/2n and Sn . This result inspires us to ask whether the EM and Feistel
schemes can be generalized over all groups, and if so, whether or not we can
get pseudorandomness in some model.
1.1
Prior Work
In extension of their simplification of the EM scheme, Dunkelman, Keller, and
Shamir [DKS12] attacked the construction using variants of slide attacks in
order to show that the security bound was optimal. They further considered
other variants of the EM scheme, such as the Addition Even-Mansour with
an Involution as the Permutation (two-keyed). Also Kilian and Rogaway
[KR01] were inspired by DESX and EM to define their F X construction, of
which the EM scheme is a special case.
As referred to above, Kuwakado and Morii were able to break the EM
scheme [KM12] and the 3-round Feistel scheme [KM10] on n-bit strings, using
Simon’s algorithm, if able to query their oracle with a superposition of states.
Kaplan et al. [KLLNP16], using Kuwakado and Morii’s results, showed how
to break many classical cipher schemes, which in turn incited Alagic and
Russell [AR17].
In their work with the Hidden Shift Problem, [AR17] posit that a Feistel
cipher construction over other groups than the bit strings might be secure
against quantum adversaries with quantum queries. Many Feistel cipher variants exist, with different relaxations on the round functions, see for example
[NR99] and [PRS02], the latter of which also considered Feistel ciphers over
other groups. Vaudenay [Vau98] also considered Feistel ciphers over other
2
groups in order to protect such ciphers against differential analysis attacks
by what he called decorrelation.
Removed from the schemes considered below and with a greater degree of
abstraction, Black and Rogaway [BR02] consider ciphers over arbitrary domains. In general, on the question of the existence of quantum pseudorandom
permutations, see [Zha16].
1.2
Summary of Results
We work in the Random Oracle Model and consider groups G in the family of
finite groups, G. We consider pseudorandom permutations, given informally
as the following.
Definition 1. [Informal] A keyed permutation P on a group G is a Pseudorandom Permutation (PRP) on G if it is indistinguishable from a
random permutation for all probabilistic distinguishers having access to only
polynomially many permutation-oracle queries.
A Super Pseudorandom Permutation (SPRP) is a permutation
where the distinguisher is given access to the inverse permutation-oracle as
well.
We define the Group Even-Mansour (EM) scheme on G to be the
encryption scheme having the encryption algorithm
Ek (m) = P (m · k) · k,
where m ∈ G is the plaintext and k ∈ G is the uniformly random key.
We define two problems for the Group Even-Mansour scheme: Existential Forgery (EFP) and Cracking (CP). In EFP, the adversary must
eventually output a plaintext-ciphertext pair which satisfies correctness. In
CP, the adversary is given a ciphertext and asked to find the corresponding
plaintext.
It holds that for our Group EM scheme, the probability that an adversary
succeeds in the EFP is polynomially bounded:
Theorem 2. [Informal] If P is a uniformly random permutation on G and
k ∈ G is chosen uniformly at random. Then, for any probabilistic adversary
A, the success probability of solving the EFP is negligible, specifically, bounded
by
st
,
O
|G|
where s and t are the amount of encryption/decryption- and permutation/inverse
permutation-oracle queries, respectively.
3
By a basic reduction, and for the latter, by an inference result, we also
get that
Theorem 3. [Informal] If P is a super pseudorandom permutation on G
and k ∈ G is chosen uniformly at random. For any probabilistic adversary
A, the success probability of solving the EFP is negligible.
Corollary 4. [Informal] If P is a super pseudorandom permutation on G
and k ∈ G is chosen uniformly at random. For any polynomial-time probabilistic adversary A, the success probability of solving the CP is negligible.
With the same bound as in Theorem 2, we find that
Theorem 5. [Informal] For any probabilistic adversary A, limited to polynomially many encryption- and decryption-oracle queries and polynomially
many permutation- and inverse permutation-oracle queries, the Group EM
scheme over a group G is a super pseudorandom permutation.
We then apply a Slide Attack, to find an attack which matches the bound
given above.
Considering the Group Feistel cipher, whose encryption algorithm consists of multiple uses of the round function
Ff (x, y) = (y, x · f (y)),
where f is a pseudorandom function on G, we show that the 3-round Feistel
cipher is pseudorandom but is not super pseudorandom, regardless of the
underlying group G. We then note that the 4-round Feistel cipher is super
pseudorandom as proven in [PRS02].
Finally, we consider the Group Even-Mansour scheme instantiated
using a 4-round Feistel cipher over G2 = G×G, which uses the encryption
algorithm
Ψf,g
k (m) = Fg,f,f,g (m · k) · k,
where f and g are modelled as random functions, m ∈ G2 the plaintext, and
k ∈ G2 is a uniformly random key. We then show one of our main results:
Theorem 6. [Informal] For any probabilistic 4-oracle adversary A with at
most
• qc queries to the Ψ- and inverse Ψ-oracles (or random oracles),
• qf queries to the f -oracle, and
4
• qg queries to the g-oracle,
we have that the success probability of A distinguishing between Ψ and a
random oracle, is bounded by
q
2
2
−1
(2qc + 4qf qc + 4qg qc + 2qc − 2qc )|G| + 2 · c (2|G|−1 + |G|−2 ).
2
We may also rewrite our main theorem as the following:
Theorem 7. [Informal] For any 4-oracle adversary A, with at most q total
queries, we have that the success probability of A distinguishing between Ψ
and a random oracle, is bounded by
2(3q 2 − 2q)|G|−1 + (q 2 − q)|G|−2 .
We note that this main result is due to [GR04], however, we consider a
one-key group version and add details to their proof sketches.
1.3
Outline of Paper
In Section 2, we state the assumptions for this paper. In Section 2, we give
definitions that hold for the paper in general, leaving specialized definitions
to the various sections. In Section 3, we introduce the generalized EM scheme
over arbitrary groups, stating and proving some results about it. In Section 4,
we define the generalized Feistel cipher over arbitrary groups and prove a few
small results about it. In Section 5, we consider an implementation of the
generalized EM scheme using the generalized Feistel cipher as the public
permutation. In Section 6, we give our concluding remarks.
2
General Definitions
In the following, we work in the Random Oracle Model such that we may
assume the existence of a random permutation oracle on group elements. We
let G be the family of all finite groups, e.g. a group G ∈ G is a pair of the set
G and operation · satisfying the group axioms. We also assume that for any
group G ∈ G, |G| ≤ 2poly(n) for some n ∈ N and some polynomial poly(·).
We will need the concept of pseudorandom, which is also called indistinguishable from random, in several forms. On notation, we write x ∈R X
for an element chosen uniformly at random from a set X. In the following,
we consider the positive integer λ to be the security parameter, specified in
unary per convention. We assume that for each λ there exists a uniquely
specified group G(λ) = Gλ ∈ G with size |Gλ | ≥ 2λ .
5
Definition 8. Let Fm,n : Gλ × Gm → Gn , for Gm , Gn ∈ G, be an efficient,
keyed function. Fm,n is a pseudorandom function (PRF) if for all probabilistic distinguishers A, limited to only polynomially many queries to the
function-oracle, there exists a negligible function negl(·), such that
Pr
k∈R Gλ
AFm,n (k,·) (λ) = 1 −
Pr
π∈R FGm →Gn
π(·)
A (λ) = 1 ≤ negl(λ),
where FGm →Gn is the set of functions from Gm to Gn .
If F : G × G → G is a pseudorandom function, we say that it is a
pseudorandom function on G.
Definition 9. Let P : Gλ ×G → G be an efficient, keyed permutation. P is a
pseudorandom permutation (PRP) if for all probabilistic distinguishers
A, limited to only polynomially many queries to the permutation-oracle, there
exists a negligible function negl(·), such that
Pr
k∈R Gλ
P (k,·)
A
(λ) = 1 −
Pr
π∈R PG→G
Aπ(·) (λ) = 1 ≤ negl(λ),
where PG→G is the set of permutations on G.
Definition 10. Let P : Gλ × G → G be an efficient, keyed permutation.
P is said to be a super pseudorandom permutation (SPRP) if for all
probabilistic distinguishers A, limited to only polynomially many queries to
the permutation- and inverse permutation-oracles, there exists a negligible
function negl(·), such that
Pr
k∈R Gλ
h
AP (k,·),P
−1 (k,·)
i
(λ) = 1 −
Pr
π∈R PG→G
h
Aπ(·),π
−1 (·)
(λ) = 1
i
≤ negl(λ),
where PG→G is the set of permutations on G.
A (super) pseudorandom permutation P : G × G → G is said to be a
(super) pseudorandom permutation on G.
6
3
Even-Mansour
We first remark that the results in this section were initially proven in a
project prior to the start of the thesis but were further worked on to complement this thesis. Thus we have chosen to include parts of it, while this
inclusion accounts for the brevity in certain results. We begin by defining
the one-key Even-Mansour scheme over arbitrary groups, which we will refer
to as the Group EM scheme.
Definition 11. We define the Group Even-Mansour scheme to be the
triple of a key generation algorithm, encryption algorithm, and decryption
algorithm. The key generation algorithm takes as input the security parameter 1λ , fixes and outputs a group G ∈R G with |G| ≥ 2λ , and outputs a key
k ∈R G. The encryption algorithm Ek (m) takes as input the key k and a
plaintext m ∈ G and outputs
Ek (m) = P (m · k) · k,
where P is the public permutation. The decryption algorithm Dk (c) takes as
input the key k and a ciphertext c ∈ G and outputs
Dk (c) = P −1 (c · k −1 ) · k −1 ,
where P −1 is the inverse public permutation. This definition satisfies correctness.
3.1
Two Forms of Security for the Group EM Scheme
In this subsection, we prove classical results about our new scheme. We do so
by considering Even and Mansour’s two notions of security: the Existential
Forgery Problem and the Cracking Problem, the Cracking Problem being the
stronger of the two.
Definition 12. In the Existential Forgery Problem (EFP), we consider
the following game:
1. A group G ∈ G and a key k ∈R G are generated.
2. The adversary A gets the security parameter, in unary, and the group
G.
3. A receives oracle access to the Ek , Dk , P, and P −1 oracles.
4. A eventually outputs a pair (m, c).
7
If Ek (m) = c, and (m, c) has not been queried before, we say that A succeeds.
In the Cracking Problem (CP), we consider the following game:
1. A group G ∈ G and a key k ∈R G are generated.
2. The adversary A gets the security parameter, in unary, and the group
G.
3. A is presented with Ek (m0 ) = c0 ∈R G.
4. A receives oracle access to the Ek , Dk , P, and P −1 oracles, but the decryption oracle outputs ⊥ if A queries c = c0 .
5. A outputs a plaintext m.
If Dk (c0 ) = m, then we say that A succeeds. The success probability is the
probability that on a uniformly random chosen encryption c0 = Ek (m0 ), A
outputs m0 .
Even and Mansour show that polynomial-time EFP security infers polynomial-time CP security. There are no limiting factors prohibiting the problems and inference result from being employed on groups. In fact, there is
nothing disallowing the use of the same proof of the EFP security for the EFP
security of the one-key EM scheme, as noted in [DKS12], which we therefore
omit. Indeed, by redefining notions in the [EM97] proof to take into account
that we are working over a not necessarily abelian group, we are able to prove
that the Group EM scheme satisfies the EFP notion of security, specifically
the following.
Theorem 13. Assume P ∈R PG→G and let the key k ∈R G. For any probabilistic adversary A, the success probability of solving the EFP is bounded
by
st
,
Succ(A) = P rk,P [EF P (A) = 1] = O
|G|
where s is the number of E/D-queries and t is the number of P/P −1 -queries,
i.e. the success probability is negligible.
By the Even and Mansour inference result, we get the corollary below.
Corollary 14. Assume P ∈R PG→G and let the key k ∈R G. For any
probabilistic polynomial-time (PPT) adversary A, the success probability of
solving the Cracking Problem is negligible.
8
As Even and Mansour also note, the above results may be extended to
instances where the permutation is a pseudorandom permutation by a simple
reduction. Hence, we get the following two results.
Theorem 15. Assume P is a pseudorandom permutation on G ∈ G and let
the key k ∈R G. For any probabilistic adversary A with only polynomially
many queries to its oracles, the success probability of solving the Existential
Forgery Problem is negligible.
Corollary 16. Assume P is a pseudorandom permutation on G ∈ G and let
the key k ∈R G. For any probabilistic polynomial-time (PPT) adversary A,
the success probability of solving the Cracking Problem is negligible.
3.2
Pseudorandomness Property of the Group EM Scheme
Although the above notions of security are strong, we are more interested
in any pseudorandomness property the Group EM scheme offers us. Kilian
and Rogaway [KR01] show that the one-key EM scheme satisfies the pseudorandom permutation property, i.e. with only an encryption oracle and the
permutation oracles, the EM scheme is indistinguishable from random to any
adversary with only polynomially many queries to its oracles. We note that
they only show the pseudorandomness property, but state in their discussion
section that their proof may be adapted to include a decryption oracle, i.e.
that the one-key EM scheme satisfies the super pseudorandom permutation
property. Having done the analysis with the decryption oracle, over an arbitrary group, we concur. However, we were also able to generalize the [KR01]
proof to a one-key construction. This not entirely remarkable as the key k
will usually be different from its group inverse, hence we were able to use
the same proof, but with adjustments to the games and their analysis. The
proof is given in the appendix for posterity. For completeness, we present
the result as the following theorem.
Theorem 17. Assume P ∈R PG→G and let the key k ∈R G. For any
probabilistic adversary A, limited to polynomially many E/D- and P/P −1 oracle queries, the adversarial advantage of A is bounded by
h
i
h
i
st
def
P,P −1
P,P −1
.
Adv(A) = P r AEk ,Dk = 1 − P r Aπ,π−1 = 1 = O
|G|
where s is the number of E/D-queries and t is the number of P/P −1 -queries,
i.e. the success probability is negligible.
Stated simply,
9
Theorem 18. For any probabilistic adversary A, limited to polynomially
many E/D- and P/P −1 -oracle queries, the Group EM scheme over a group
G is a super pseudorandom permutation.
By removing the decryption oracle, we get the following corollary:
Corollary 19. For any probabilistic adversary A, limited to polynomially
many E- and P/P −1 -oracle queries, the Group EM scheme over a group G
is a pseudorandom permutation.
Remark. We see that in the group ((Z/2Z)n , ⊕), our Group EM scheme
reduces to the one-key EM scheme given in [DKS12]. The proof given in
[DKS12] proves the security of the scheme, and the proof given in [KR01]
proves the pseudorandomness, equivalently to our claims.
It can be proven that a multiple round Group EM scheme is an SPRP
because the security only depends on the last round, which is also an SPRP.
3.3
Slide Attack
We would like to show that the security bound that we have found above
is optimal, so we slightly alter the simple optimal attack on the Single-Key
Even-Mansour cipher as constructed in [DKS12]. The original version works
for abelian groups with few adjustments and [DKS12] also present another
slide attack against a modular addition DESX construction.
Consider the one-key Group Even-Mansour cipher
E(x) = P (x · k) · k,
over a group G with binary operation ·, where P is a publicly available
permutation oracle, x ∈ G, and k ∈R G. Define the following values:
x = x, y = x · k, z = P (y), w = E(x) = P (x · k) · k.
We hereby have that w · y −1 = z · x−1 . Consider the attack which follows.
p
1. For d = |G| arbitrary values xi ∈ G, i = 1, . . . , d, and d arbitrary
values yi ∈ G, i = 1, . . . , d, query the E-oracle on the xi ’s and the
P -oracle on the yi ’s. Store the values in a hash table as
(E(xi ) · yi−1 , P (yi ) · x−1
i , i),
sorted by the first coordinate.
10
2. If there exists a match in the above step, i.e. E(xi ) · yi−1 = P (yi ) · x−1
i
for some i, check the guess that k = x−1
·
y
.
i
i
It can be seen by the Birthday Problem1 , that with non-negligible probability,
there must exist a slid pair (xi , yi ) satisfying the above property, i.e. there
2
exists 1 ≤ i ≤ d such that k = x−1
i · yi . For a random pair (x, y) ∈ G it
−1
−1
holds that E(x) = P (y) · x · y with probability |G| , so we expect few, if
any, collisions in the hash table, including the collision by the slid pair where
the correct key k is found. The data complexity of the attack is d E-oracle
queries and d P -oracle queries. Hence the attack bound d2 = |G|, which
matches the lower bound given in Theorem 13 and Theorem 39. We have
therefore found that our scheme is optimal.
2
n
Considering the approximation p(n) ≈ 2m
, where p(n) is the probability of there being
a Birthday Problem
collision
from
n
randomly
chosen elements from the set of m elements,
√ 2
p
|G|
then p( |G|) ≈ 2|G| = 1/2.
1
11
4
Feistel
We now consider the Feistel cipher over arbitrary groups, which we will call
the Group Feistel cipher. The following is a complement to [PRS02] who
treat the Group Feistel cipher construction with great detail. Our main
accomplishment in this section is the settling of an open problem posed by
them.
4.1
Definitions
We define a Feistel cipher over a group (G, ·) as a series of round functions
on elements of G × G = G2 .
Definition 20. Given an efficiently computable but not necessarily invertible
function f : G → G, called a round function, we define the 1-round
Group Feistel cipher Ff to be
Ff : G × G −→ G × G,
(x, y) 7−→ (y, x · f (y)).
In the case where we have multiple rounds, we index the round functions as
fi , and denote the r-round Group Feistel cipher by Ff1 ,...,fr . We concurrently denote the input to the i’th round by (Li−1 , Ri−1 ) and having the
output (Li , Ri ) = (Ri−1 , Li−1 · fi (Ri−1 )), where Li and Ri respectively denote
the left and right parts of the i’th output.
Note that if (Li , Ri ) is the i’th round output, we may invert the i’th round
by setting Ri−1 := Li and then computing Li−1 := Ri · (fi (Ri−1 ))−1 to get
(Li−1 , Ri−1 ). As this holds for all rounds, regardless of the invertibility of the
round functions, we get that an r-round Feistel cipher is invertible for all r.
Let F : Gλ × G → G be a pseudorandom function. We define the keyed
permutation F (r) as
(r)
def
Fk1 ,...,kr (x, y) = FFk1 ,...,Fkr (x, y).
We sometimes index the keys as 1, 2, . . . , r, or omit the key index entirely.
4.2
Results
For completeness, we show some of the preliminary results for Group Feistel
ciphers, not considered in [PRS02].
12
We first note that F (1) is not a pseudorandom permutation as
(1)
Fk1 (L0 , R0 ) = (L1 , R1 ) = (R0 , L0 · Fk1 (R0 )),
such that any distinguisher A need only compare R0 to L1 .
Also F (2) is not a pseudorandom permutation: Consider a pseudorandom
function F on G. Pick k1 , k2 ∈R Gλ . Distinguisher A sets (L0 , R0 ) = (1, g)
for some g ∈ G, where 1 is the identity element of G, then queries (L0 , R0 )
to its oracle and receives,
L2 = L0 · Fk1 (R0 ) = Fk1 (g) and R2 = R0 · Fk2 (L0 · Fk1 (R0 )) = g · Fk2 (Fk1 (g)).
On its second query, the distinguisher A lets L0 ∈ G \ {1} but R0 = g, such
that it receives
L2 = L0 · Fk1 (R0 ) = L0 · Fk1 (g) and R2 = g · Fk2 (L0 · Fk1 (g)).
As A may find the inverse to elements in G, A acquires (Fk1 (g))−1 , and
by so doing, may compute L2 · (Fk1 (g))−1 = L0 . If F (2) were random, this
would only occur negligibly many times, while A may query its permutationoracle polynomially many times such that if L0 is retrieved non-negligibly
many times out of the queries, A is able to distinguish between a random
permutation and F (2) with non-negligible probability.
As one would expect, the 3-round Group Feistel cipher (see Figure 1a) is
indeed a pseudorandom permutation.
Theorem 21. If F is a pseudorandom function on G, then F (3) is a pseudorandom permutation on G.
The proof of this proposition can be generalized from the proof given in
Katz and Lindell [KL15] of the analogous result over bit-strings with XOR,
with no difficulties. We therefore omit it here.
Among the considerations in [PRS02], they showed that the 3-round Feistel cipher over abelian groups was not super pseudorandom, but left as an
open problem a proof over non-abelian groups. We present such a proof now.
Proposition 22. The 3-round Group Feistel cipher is not super pseudorandom.
Proof. The proof is a counter-example using the following procedure:
1. Choose two oracle-query pairs in G × G: (L0 , R0 ) and (L00 , R0 ) where
L0 6= L00 .
2
TikZ figure adapted from [Jea16].
13
xL
·k L
xR
·k R
g
L0
R0
f1
f
f2
f
f3
g
·k L
L3
yL
R3
·k R
yR
(b) Group EM scheme with Feistel.2
(a) 3-round Group Feistel cipher.
Figure 2: Encryption schemes.
2. Query the encryption oracle to get (L3 , R3 ) and (L03 , R30 ).
3. Query (L003 , R300 ) = (L03 , L0 · (L00 )−1 · R30 ) to the decryption oracle.
4. If R000 = L03 · (L3 )−1 · R0 , guess that the oracle is F (3) , else guess random.
For F (3) , this algorithm succeeds with probability 1. For a random permutation, this algorithm succeeds negligibly often.
For super pseudorandomness of the 4-round Group Feistel cipher, we refer
the reader to [PRS02]. In the paper, they show a strong result using certain
hash functions as round functions, from which the following is a corollary.
Corollary 23. Let G be a group, with characteristic other than 2, and let
f, g : Gλ × G → G be pseudorandom functions. Then, for any adversary
A with polynomially many queries to its E/D-oracles, the family P of permutations on G × G consisting of permutations of the form F (4) = Fg,f,f,g
are indistinguishable from random, i.e. super pseudorandom permutations
(SPRPs).
14
5
Implementing the Group Even-Mansour Scheme
Now that we have shown that both the Even-Mansour scheme and the Feistel cipher are generalizable to arbitrary groups, we might consider how to
implement one given the other. Gentry and Ramzan [GR04] considered exactly this for the two-key EM scheme over (Z/2Z)n . However, their paper
only had sketches of proofs and refer to another edition of the paper for full
details. As we are unable to find a copy in the place that they specify it to
exist, and as we generalize their result non-trivially, we have decided to fill
in the details while generalizing their proof.
In this section, we consider a generalized version of the Gentry and
Ramzan [GR04] construction, namely, the Group Even-Mansour scheme on
G2 instantiated with a 4-round Group Feistel cipher as the public permutation:
Ψf,g
k (x) = Fg,f,f,g (x · k) · k,
where k = (k L , k R ) ∈ G2 is a key consisting of two subkeys, chosen independently and uniformly at random, and f and g are round functions on G,
modelled as random function oracles, available to all parties, including the
adversary. We consider the operation x · k for x = (xL , xR ) ∈ G2 , to be the
coordinate-wise group operation, but do not otherwise discern between it and
the group operation · on elements of G. In the following, we shall follow the
proof in [GR04] closely. However, we make quite a few modifications, mostly
due to the nature of our generalization. Note that we consider a one-key
scheme, as opposed to the two-key version in [GR04] (see Figure 1b.) Our
main theorem for this section is the following.
Theorem 24. Let f, g be modelled as random oracles and let the subkeys of
k = (k L , k R ) ∈ G2 be chosen independently and uniformly at random. Let
Ψf,g
k (x) = Fg,f,f,g (x · k) · k, and let R ∈R PG2 →G2 . Then, for any probabilistic
4-oracle adversary A with at most
• qc queries to Ψ and Ψ−1 (or R and R−1 ),
• qf queries to f , and
• qg queries to g,
we have
h
i
h
i
−1
−1
P r AΨ,Ψ ,f,g = 1 − P r AR,R ,f,g = 1
≤
(2qc2
+ 4qf qc + 4qg qc +
2qc2
−1
− 2qc )|G|
15
q
+ 2 · c (2|G|−1 + |G|−2 ).
2
5.1
Definitions
Before we can begin the proof, we will need several definitions all of which
are identical to the [GR04] definitions, up to rewording.
Definition 25. Let P denote the permutation oracle (either Ψ or R), Of
and Og the f and g oracles, respectively. We get the transcripts: TP , the set
of all P queries, Tf , the set of all f queries, and Tg , the set of all g queries,
i.e. the sets
TP = {hx1 , y1 i, hx2 , y2 i, · · · , hxqc , yqc i}P ,
Tf = {hx01 , y10 i, hx02 , y20 i, · · · , hx0qf , yq0 f i}f ,
Tg = {hx001 , y100 i, hx002 , y200 i, · · · , hx00qg , yq00g i}g .
We discern between two types of oracle queries: Cipher queries (+, x) = P (x)
and (−, y) = P −1 (y); Oracle queries (Of , x0 ) and (Og , x00 ), respectively f - and
g-oracle queries.
As we have no bounds on the computational complexity of the adversary
A, we may assume that A is deterministic, as we did in the proof of Theorem 39. Hence, we may consider an algorithm CA which, given a set of A’s
queries, can determine A’s next query.
Definition 26. For 0 ≤ i ≤ qc , 0 ≤ j ≤ qf , and 0 ≤ k ≤ qg , the i+j+k+1’st
query by A is
CA {hx1 , y1 i, . . . , hxi , yi i}P , {hx01 , y10 i, . . . , hx0j , yj0 i}f , {hx001 , y100 i, . . . , hx00k , yk00 i}g
where the upper equality case on the indexes is defined to be A’s final output.
Definition 27. Let σ = (TP , Tf , Tg ) be a tuple of transcripts with length
qc , qf , qg , respectively. We say that σ is a possible A-transcript if for every
1 ≤ i ≤ qc , 1 ≤ j ≤ qf , and 1 ≤ k ≤ qg ,
CA {hx1 , y1 i, . . . , hxi , yi i}P , {hx01 , y10 i, . . . , hx0j , yj0 i}f , {hx001 , y100 i, . . . , hx00k , yk00 i}g
∈ {(+, xi+1 ), (−, yi+1 ), (Of , x0j+1 ), (Og , x00k+1 )}.
Let us define two useful ways in which we may answer A’s queries other
than what we have already defined.
Definition 28. Let Ψ̃ be the process where the Ψ- and Ψ−1 cipher query
oracles use f and g, and Of uses f , but Og is replaced by Oh for another,
independent, random function h.
16
Definition 29. Let R̃ denote the process which answers all oracle queries
using f and g, but answers the i’th cipher query as follows.
1. If A queries (+, xi ) and there exists 1 ≤ j < i, such that the j’th
query-answer pair has xj = xi , return yi := yj .
2. If A queries (−, yi ) and there exists 1 ≤ j < i, such that the j’th
query-answer pair has yj = yi , return xi := xj .
3. Otherwise, return uniformly chosen element in G2 .
The latter definition may not be consistent with any function or permutation, so we formalize exactly this event.
Definition 30. Let TP be a possible A-cipher-transcript. TP is inconsistent
if for some 1 ≤ i < j ≤ qc there exist cipher-pairs such that either
• xi = xj but yi 6= yj , or
• xi 6= xj but yi = yj .
Any σ containing such a transcript TP is called inconsistent.
Note. Assume from now on that A never repeats any part of a query if
the answer can be determined from previous queries, i.e. every possible Atranscript σ is consistent such that if i 6= j, then xi 6= xj , yi 6= yj , x0i 6= x0j ,
and x00i 6= x00j .
Note. Let TΨ , TΨ̃ , TR̃ , TR denote the transcripts seen by A when its cipher
queries are answered by Ψ, Ψ̃, R̃, R, respectively, and oracle queries by Of
and Og (noting that in the case of Ψ̃, the function in the Og has been replaced
by another random function, h.) We also note that using this notation, we
−1
have that AΨ,Ψ ,f,g = CA (TΨ ) (and likewise for Ψ̃, R̃, and R.)
5.2
Lemmas
Now, let us begin finding results that will aid us in proving our main theorem.
First, we will compare the distributions of R̃ and R, using a result by NaorReingold3 . Afterwards, we shall consider when the distributions of Ψ and Ψ̃
are equal. Lastly, we shall consider when the distributions of Ψ̃ and R̃ are
equal. Combining these results will allow us to prove our main theorem.
We remark that whenever we write k = (k L , k R ) ∈R G2 , we mean that
the subkeys are chosen independently and uniformly at random.
3
The proof of the proposition follows the argument of Proposition 3.3 in [NR99].
17
q
Lemma 31. P r [CA (TR̃ ) = 1] − P r [CA (TR ) = 1] ≤ c · |G|−2 .
2
R
R̃
Proof. Let σ be a possible and consistent A-transcript, then
2
|G|
= P r [TR̃ = σ | TR̃ is consistent] ,
P r [TR = σ] =
qc
R
R̃
simply because the only difference between TR and TR̃ is in the cipher queries,
and when TR̃ is consistent, we have no overlap on the query-answer pairs,
hence we need only consider how to choose qc elements from |G|2 many
possible elements, without replacement. Let us now consider the probability
of TR̃ being inconsistent. If TR̃ is inconsistent for some 1 ≤ i < j ≤ qc then
either xi = xj and yi 6= yj , or xi 6= xj and yi = yj . For any given i, j,
this happens with at most probability |G|−2 , because if xi = xj is queried,
then the R̃-oracle would return the corresponding yi = yj , but if xi 6= xj is
queried, then the R̃-oracle would return a uniformly random element (and
likewise if yi = yj or yi 6= yj were queried to the inverse R̃-oracle.) Hence,
q
P r [TR̃ is inconsistent] ≤ c · |G|−2 .
2
R̃
We thereby get that,
P r [CA (TR̃ ) = 1] − P r [CA (TR ) = 1]
R
R̃
≤ P r [CA (TR̃ ) = 1|TR̃ is consistent] − P r [CA (TR ) = 1] · P r [TR̃ is consistent]
R
R̃
R̃
+ P r [CA (TR̃ ) = 1|TR̃ is inconsistent] − P r [CA (TR ) = 1] · P r [TR̃ is inconsistent]
R
R̃
R̃
≤ P r [TR̃ is inconsistent]
R̃
q
≤ c · |G|−2 ,
2
as the distribution over R is independent of the (in)consistency of TR̃ .
Let us now focus on the distributions of TΨ and TΨ̃ , to show that they
are identical unless the input to g in the cipher query to Ψ is equal to the
oracle input to h in Oh . In order to do so, we first define the event BadG(k).
Definition 32. For every specific key k = (k L , k R ) ∈R G2 , we define BadG(k)
to be the set of all possible and consistent A-transcripts σ, satisfying at least
one of the following:
R
00
BG1: ∃i, j, 1 ≤ i ≤ qc , 1 ≤ j ≤ qg , such that xR
i · k = xj , or
18
BG2: ∃i, j, 1 ≤ i ≤ qc , 1 ≤ j ≤ qg , such that yiL · (k L )−1 = x00j .
Lemma 33. Let k = (k L , k R ) ∈R G2 . For any possible and consistent Atranscript σ = (TP , Tf , Tg ), we have
P r [σ ∈ BadG(k)] ≤
k
2qg qc
.
|G|
Proof. We know that σ ∈ BadG(k) if one of BG1 or BG2 occur, hence,
using the union bound,
P r [σ ∈ BadG(k)] = P r [BG1 occurs ∨ BG2 occurs |σ]
k
k
≤ P r [BG1 occurs |σ] + P r [BG2 occurs |σ]
k
k
−1
−1
≤ qg qc · |G| + qg qc · |G|
= 2qg qc · |G|−1 .
Lemma 34. Let σ be a possible and consistent A-transcript, then
P r [TΨ = σ|σ 6∈ BadG(k)] = P r [TΨ̃ = σ] .
Ψ
Ψ̃
Proof. We want to show that the query answers in the subtranscripts of the
games Ψ and Ψ̃ are equally distributed, under the condition that neither of
the events BG1 nor BG2 occur in game Ψ. Fix the key k = (k L , k R ) ∈R G2 .
Recall that the adversary does not query an oracle if it can determine the
answer from previous queries.
In both games, for any Of -oracle query x0 ∈ G, the query answer will
be equally distributed in both games as the underlying random function f is
the same in both games.
In game Ψ, an Og -oracle query, x00 ∈ G, will have a uniformly random
answer as g is a random function. Likewise, in game Ψ̃, an Og -oracle query,
x00 ∈ G, will have a uniformly random answer as h is a random function.
Consider now the permutation oracle P = Fg,f,f,g (x · k) · k. We consider
a query-answer pair hx, yi ∈ TP for x, y ∈ G2 .
In both games, xR · k R will be the input to the first round function, which
is g. In game Ψ̃ the output is always a uniformly random element, newly
selected by g. In game Ψ, if xR ·k R has already been queried to the Og -oracle,
the output of the round function is the corresponding oracle answer, else it
is a uniformly random element, newly selected by g. As the former event in
game Ψ never occurs because the event BG1 never occurs, the distributions
are equal.
19
As both games have access to the same random function f , the second
and third round function outputs will have equal distributions.
In both games, y L · (k L )−1 will be the input to the fourth round function,
which is again g. In game Ψ̃ the output is always a uniformly random element,
newly selected by g, unless y L · (k L )−1 = xR · k R , in which case the output
is equal to the output of the first round function. In game Ψ, if xR · k R has
already been queried to the Og -oracle, but not as input to the first round
function, the output of the round function is the corresponding oracle answer.
If y L · (k L )−1 = xR · k R , then the output is equal to the output of the first
round function, else it is a uniformly random element newly selected by g.
As the former event in game Ψ never occurs because the event BG2 never
occurs, the distributions are equal.
As A does not ask a query if it can determine the answer based on previous queries, we see that the inverse permutation oracle, using P −1 , yields
analogous distributions. Thus, the distributions for the two games must be
equal.
Let us show that the distributions of TΨ̃ and TR̃ are identical, unless the
same value is input to f on two separate occasions. Here we also define
when a key is "bad" as we did above, but altered such that it pertains to our
current oracles.
Definition 35. For every specific key k = (k L , k R ) ∈R G2 and function
g ∈R FG→G , define Bad(k, g) to be the set of all possible and consistent Atranscripts σ satisfying at least one of the following events:
B1: ∃1 ≤ i < j ≤ qc , such that
R
L
L
R
R
xLi · k L · g(xR
i · k ) = xj · k · g(xj · k )
B2: ∃1 ≤ i < j ≤ qc , such that
yiR · (k R )−1 · g(yiL · (k L )−1 )
−1
−1
= yjR · (k R )−1 · g(yjL · (k L )−1 )
B3: ∃1 ≤ i, j ≤ qc , such that
−1
R
R
R −1
xLi · k L · g(xR
· g(yjL · (k L )−1 )
i · k ) = yj · (k )
B4: ∃1 ≤ i ≤ qc , 1 ≤ j ≤ qf , such that
R
0
xLi · k L · g(xR
i · k ) = xj
20
B5: ∃1 ≤ i ≤ qc , 1 ≤ j ≤ qf , such that
yiR · (k R )−1 · g(yiL · (k L )−1 )
−1
= x0j
Lemma 36. Let k = (k L , k R ) ∈R G2 . For any possible and consistent Atranscript σ, we have that
q
2
P r [σ ∈ Bad(k, g)] ≤ qc + 2qf qc + 2 · c
· |G|−1 .
2
k,g
Proof. We have that σ ∈ Bad(k, g) if it satisfies a Bi for some i = {1, . . . , 5}.
Using that k L , k R are uniform and independently chosen, and g ∈R FG→G ,
we may achieve an upper bound on the individual event probabilities, and
then use the union
bound.
q
There are c many ways of picking i, j such that 1 ≤ i < j ≤ qc , also,
2
qf qc many ways of picking i, j such that 1 ≤ i ≤ qc , 1 ≤ j ≤ qf , and qc2 many
ways of picking i, j such that 1 ≤ i, j ≤ qc . The probability that two elements
chosen from G are equal is |G|−1 , so we may bound each event accordingly
and achieve, using the union bound, that
"
P r [σ ∈ Bad(k, g)] = P r
k,g
k,g
≤
5
X
i=1
5
_
#
Bi occurs |σ
i=1
P r [Bi occurs |σ]
k,g
qc
q
−1
≤
· |G| + c · |G|−1 + qc2 · |G|−1 + qf qc · |G|−1 + qf qc · |G|−1
2
2
q
= qc2 + 2qf qc + 2 c
· |G|−1 .
2
Lemma 37. Let σ be a possible and consistent A-transcript, then
P r [TΨ̃ = σ|σ 6∈ Bad(k, g)] = P r [TR̃ = σ] .
Ψ̃
R̃
The following proof is based on the proof in [GR04] which refers to [NR99]
for the first part of their argument. We need the generalization of this argument and so also include it.
Proof. Since σ is a possible A-transcript, we have for all 1 ≤ i ≤ qc , 1 ≤ j ≤
qf , 1 ≤ k ≤ qg :
CA {hx1 , y1 i, . . . , hxi , yi i}P , {hx01 , y10 i, . . . , hx0j , yj0 i}f , {hx001 , y100 i, . . . , hx00k , yk00 i}g
∈ {(+, xi+1 ), (−, yi+1 ), (Of , x0j+1 ), (Og , x00k+1 )}.
21
Therefore, TR̃ = σ if and only if ∀1 ≤ i ≤ qc , ∀1 ≤ j ≤ qf , and ∀1 ≤
k ≤ qg , the i, j, k’th respective answers R̃ gives are yi or xi , and x0j and
x00k , respectively. As A never repeats any part of a query, we have, by the
definition of R̃, that the i’th cipher-query answer is an independent and
uniform element of G2 , and as f and g were modelled as random function
oracles, so too will their oracle outputs be independent and uniform elements
of G. Hence,
P r [TR̃ = σ] = |G|−(2qc +qf +qg ) .
R̃
For the second part of this proof, we fix k, g such that σ 6∈ Bad(k, g) and
seek to compute P r [TΨ̃ = σ]. Since σ is a possible A-transcript, we have
f,h
that TΨ̃ = σ if and only if
• yi = Fg,f,f,g (xi · k) · k for all 1 ≤ i ≤ qc ,
• yj0 = f (x0j ) for all 1 ≤ j ≤ qf , and
• yk00 = g(x00k ) for all 1 ≤ k ≤ qg (note that g = h here.)
If we define
R
Xi := xLi · k L · g(xR
i ·k )
−1
Yi := yiR · (k R )−1 · g(yiL · (k L )−1 )
,
then (yiL , yiR ) = Ψ̃(xLi , xR
i ) if and only if
−1
· Yi
k R · f (Xi ) = (xR
i )
and
Xi · f (Yi ) = yiL · (k L )−1 ,
where the second equality of the latter is equivalent to (k L )−1 · (f (Yi ))−1 =
(yiL )−1 · Xi . Observe that, for all 1 ≤ i < j ≤ qc , Xi 6= Xj (by B1 ) and
Yi 6= Yj (by B2.) Similarly, 1 ≤ i < j ≤ qc , Xi 6= Yj (by B3.) Also, for
all 1 ≤ i ≤ qc and for all 1 ≤ j ≤ qf , x0j 6= Xi (by B4 ) and x0j 6= Yi (by
B5.) Hence, σ 6∈ Bad(k, g) implies that all inputs to f are distinct. This
then implies that P rf,h [TΨ̃ = σ] = |G|−(2qc +qf +qg ) as h was also modelled as
a random function, independent from g. Thus, as we assumed that k and g
were chosen such that σ 6∈ Bad(k, g),
P r [TΨ̃ = σ|σ 6∈ Bad(k, g)] = |G|−(2qc +qf +qg ) = P r [TR̃ = σ] .
Ψ̃
R̃
22
5.3
Proof of Theorem 24
To complete the proof of Theorem 24, we combine the above lemmas into
the following probability estimation.
Proof of Theorem 24. Let Γ be the set of all possible and consistent Atranscripts σ such that A(σ) = 1. In the following, we ease notation, for
the sake of the reader. We let BadG(k) be denoted by BadG and Bad(k, g)
by Bad. Furthermore, we abbreviate inconsistency as incon.. Let us consider
the cases between Ψ, Ψ̃ and R̃.
|P rΨ [CA (TΨ ) = 1] − P rΨ̃ [CA (TΨ̃ ) = 1]|
≤
X
(P rΨ [TΨ = σ] − P rΨ̃ [TΨ̃ = σ]) + P rΨ̃ [TΨ̃ incon.]
σ∈Γ
≤
X
|P rΨ [TΨ = σ | σ 6∈ BadG] − P rΨ̃ [TΨ̃ = σ]| · P rk [σ 6∈ BadG]
σ∈Γ
X
+
(P rΨ [TΨ = σ | σ ∈ BadG] − P rΨ̃ [TΨ̃ = σ]) · P rk [σ ∈ BadG]
σ∈Γ
+ P rΨ̃ [TΨ̃ incon.]
≤
X
(P rΨ [TΨ = σ | σ ∈ BadG] − P rΨ̃ [TΨ̃ = σ]) · P rk [σ ∈ BadG] + qc (qc − 1)|G|−1 ,
σ∈Γ
where we in the last estimate used Lemma 34 and a consideration of the
maximal amount of possible inconsistent pairs.
At the same time,
|P rΨ̃ [CA (TΨ̃ ) = 1] − P rR̃ [CA (TR̃ ) = 1]|
≤
X
(P rΨ̃ [TΨ̃ = σ] − P rR̃ [TR̃ = σ]) + P rR̃ [TR̃ incon.] + P rΨ̃ [TΨ̃ incon.]
σ∈Γ
≤
X
|P rΨ̃ [TΨ̃ = σ | σ 6∈ Bad] − P rR̃ [TR̃ = σ]| · P rk [σ 6∈ Bad]
σ∈Γ
+
X
(P rΨ̃ [TΨ̃ = σ | σ ∈ Bad] − P rR̃ [TR̃ = σ]) · P rk [σ ∈ Bad]
σ∈Γ
+ P rR̃ [TR̃ incon.] + P rΨ̃ [TΨ̃ incon.]
≤
X
(P rΨ̃ [TΨ̃ = σ | σ ∈ Bad] − P rR̃ [TR̃ = σ]) · P rk [σ ∈ Bad] +
σ∈Γ
qc
q
|G|−2 + 2 c |G|−1 ,
2
2
where we in the last estimate used Lemma 37 and the proof of Lemma 31.
23
Let us use the above in a temporary estimate,
|P rΨ [CA (TΨ ) = 1] − P rR [CA (TR ) = 1]|
= |P rΨ [CA (TΨ ) = 1] − P rΨ̃ [CA (TΨ̃ ) = 1]|
+ |P rΨ̃ [CA (TΨ̃ ) = 1] − P rR̃ [CA (TR̃ ) = 1]|
+ |P rR̃ [CA (TR̃ ) = 1] − P rR [CA (TR ) = 1]|
≤
X
(P rΨ [TΨ = σ | σ ∈ BadG] − P rΨ̃ [TΨ̃ = σ]) · P rk [σ ∈ BadG] + qc (qc − 1)|G|−1
σ∈Γ
qc
q
−2
|G| + 2 c |G|−1
(P rΨ̃ [TΨ̃ = σ | σ ∈ Bad] − P rR̃ [TR̃ = σ]) · P rk [σ ∈ Bad] +
+
2
2
σ∈Γ
q
+ c |G|−2 ,
(1)
2
X
where we in the last estimate also used Lemma 31.
We may assume WLOG that
X
P rΨ [TΨ = σ | σ ∈ BadG] · P rk [σ ∈ BadG] ≤
σ∈Γ
X
P rΨ̃ [TΨ̃ = σ] · P rk [σ ∈ BadG]
σ∈Γ
and likewise,
X
P rΨ̃ [TΨ̃ = σ | σ ∈ BadG] · P rk [σ ∈ BadG] ≤
σ∈Γ
X
P rR̃ [TR̃ = σ] · P rk [σ ∈ BadG] ,
σ∈Γ
such that by Lemma 33, respectively Lemma 36, we get the following
continued estimate from (1), using the triangle inequality and that |Γ| ≤
|G|2qc +qf +qg (every combination of query elements).
|P rΨ [CA (TΨ ) = 1] − P rR [CA (TR ) = 1]|
X
≤2
P rΨ̃ [TΨ̃ = σ] · P rk [σ ∈ BadG] + 2qc (qc − 1)|G|−1
σ∈Γ
+2
X
P rR̃ [TR̃ = σ] · P rk [σ ∈ Bad]
σ∈Γ
q
+ 2 c |G|−2
2
≤ 2|Γ| · |G|−(2qc +qf +qg ) · max P rk [σ ∈ BadG] + 2qc (qc − 1)|G|−1
σ∈Γ
−(2qc +qf +qg )
+ 2|Γ| · |G|
· max P rk [σ ∈ Bad]
σ∈Γ
qc
−2
+2
|G|
2
q
q
≤ 4qg qc · |G|−1 + 2qc (qc − 1)|G|−1 + 2 qc2 + 2qf qc + 2 c
|G|−1 + 2 c |G|−2
2
2
q
= (2qc2 + 4qg qc + 4qf qc + 2qc2 − 2qc )|G|−1 + 2 c 2|G|−1 + |G|−2 .
2
24
If we denote the total amount of queries as q = qc + qf + qg , then we may
quickly estimate and reword the main theorem as:
Theorem 38. Let f, g be modelled as random oracles, let k = (k L , k R ) ∈R
G2 , let Ψf,g
k (x) = Fg,f,f,g (x · k) · k, and let R ∈R PG2 →G2 . Then, for any
4-oracle adversary A, with at most q total queries, we have
h
i
h
i
−1
−1
P r AΨ,Ψ ,f,g = 1 − P r AR,R ,f,g = 1 ≤ 2(3q 2 − 2q)|G|−1 + (q 2 − q)|G|−2 .
Proof. Given Theorem 24, we get, by using that qf , qg ≥ 0,
qc2 + 2qf qc + 2qg qc + qc2 − qc
= 2(qc2 + qf qc + qg qc ) − qc
≤ 2(qc2 + qf qc + qg qc ) + (2(qf + qg )2 + 2(qf qc + qg qc ) − qf − qg ) − qc
= 2(qc2 + 2qf qc + qf2 + 2qf qg + 2qg qc + qg2 ) − (qc + qf + qg )
= 2(qc + qf + qg )2 − (qc + qf + qg )
= 2q 2 − q.
q
As 2· c = qc2 −qc ≤ q 2 −q, we get the final estimate by some reordering.
2
6
Conclusion
We generalized the Even and Mansour scheme as well as the Feistel cipher
to work over arbitrary groups and proved that classical results pertain to
the group versions. Based on the work in [AR17], we hope that this opens
avenues to proving that classical schemes may be made quantum secure by
generalizing them to certain groups. For further work, we suggest generalizing other classical schemes and using the underlying group structures to do
Hidden Shift reductions.
The author would like to thank his thesis advisor Gorjan Alagic for the
topic, enlightening questions and answers, as well as the encouragements
along the way. The author would also like to thank the Department of
Mathematical Sciences, at the University of Copenhagen, for lending their
facilities during the writing process.
25
References
[AR17] Gorjan Alagic and Alexander Russell.
Quantum-secure
symmetric-key cryptography based on hidden shifts. EUROCRYPT 2017, 2017.
[BR02] John Black and Phillip Rogaway. Ciphers with Arbitrary Finite
Domains, pages 114–130. Springer Berlin Heidelberg, 2002.
[DKS12] Orr Dunkelman, Nathan Keller, and Adi Shamir. Minimalism
in cryptography: The Even-Mansour scheme revisited. EUROCRYPT, 2012.
[DR02] Yan Zong Ding and Michael O. Rabin. Hyper-encryption and
everlasting security. In STACS 2002, 19th Annual Symposium
on Theoretical Aspects of Computer Science, Antibes - Juan les
Pins, France, March 14-16, 2002, Proceedings, pages 1–26, 2002.
[EM97] Shimon Even and Yishay Mansour. A construction of a cipher
from a single pseudorandom permutation. Cryptology, 1997.
[GR04] Craig Gentry and Zulfikar Ramzan. Eliminating random permutation oracles in the Even-Mansour cipher. ASIACRYPT,
2004.
[Jea16] Jeremy Jean. Tik Z for cryptographers. http://www.iacr.org/
authors/tikz/, 2016.
[KL15] Jonathan Katz and Yehuda Lindell. Introduction to Modern
Cryptography. CRC Press, 2 edition, 2015.
[KLLNP16] Marc Kaplan, Gaetan Leurent, Anthony Leverrier, and Maria
Naya-Plasencia. Breaking symmetric cryptosystems using quantum period finding. ArXiv, 2016.
[KM10] Hidenori Kuwakado and Masakatu Morii. Quantum distinguisher between the 3-round Feistel cipher and the random permutation. In ISIT, pages 2682–2685. IEEE, 2010.
[KM12] Hidenori Kuwakado and Masakatu Morii. Security on the
quantum-type Even-Mansour cipher. In ISITA, pages 312–316.
IEEE, 2012.
26
[KR01] Joe Kilian and Phillip Rogaway. How to protect DES against
exhaustive key search (an analysis of DESX). J. Cryptology,
14(1):17–35, 2001.
[LR88] Michael Luby and Charles Rackoff. How to construct pseudorandom permutations from pseudorandom functions. SIAM J.
Comput., 17(2):373–386, 1988.
[NR99] Moni Naor and Omer Reingold. On the construction of pseudorandom permutations: Luby-Rackoff revisited. Journal of Cryptology, 12:29–66, 1999. Preliminary version in: Proc. STOC 97.
[PRS02] Sarvar Patel, Zulfikar Ramzan, and Ganapathy S. Sundaram.
Luby-rackoff ciphers: Why XOR is not so exclusive. In Selected Areas in Cryptography, 9th Annual International Workshop, SAC 2002, St. John’s, Newfoundland, Canada, August
15-16, 2002. Revised Papers, pages 271–290, 2002.
[Vau98] Serge Vaudenay. Provable security for block ciphers by decorrelation, pages 249–275. Springer Berlin Heidelberg, Berlin, Heidelberg, 1998.
[Zha16] Mark Zhandry. A note on quantum-secure PRPs.
abs/1611.05564, 2016.
27
CoRR,
A
Super Pseudorandomness of the Group EM
Scheme
In the following, we assume that the adversary A is unbounded computationally, but may only make polynomially many queries to the E/D- and
P/P −1 -oracles, where all oracles act as black boxes and P is a truly random
permutation. We intend to play the "pseudorandom or random permutation
game": A is given an encryption oracle E (with related decryption oracle
D) which is randomly chosen with equal probability from the following two
options:
1. A random key k ∈R G is chosen uniformly and used to encrypt as
E(m) = Ek (m) = P (m · k) · k, or
2. A random permutation π ∈R PG→G is chosen and used to encrypt as
E(m) = π(m).
The adversary wins the game if it can distinguish how E was chosen, with
probability significantly better than 1/2. More explicitly, we wish to prove
the following for the group Even-Mansour scheme.
Theorem 39. Assume P ∈R PG→G and let the key k ∈R G. For any
probabilistic adversary A, limited to polynomially many E/D- and P/P −1 oracle queries, the adversarial advantage of A is bounded by
h
i
h
i
st
def
P,P −1
P,P −1
.
(2)
Adv(A) = P r AEk ,Dk = 1 − P r Aπ,π−1 = 1 = O
|G|
where s is the total number of E/D-queries and t is the total number of
P/P −1 -queries, i.e. the success probability is negligible.
Proof. We may assume that A is deterministic (in essence, being unbounded
computationally affords A the possibility of derandomizing its strategy by
searching all its possible random choices and picking the most effective choices
after having computed the effectiveness of each choice. For an example,
see [DR02].) We may also assume that A never queries a pair in Ss or Tt
more than once, where Si and Ti are the sets of i E/D- and P/P −1 -queries,
respectively. Let us define two main games, that A could play, through oracle
interactions (see next page for the explicit game descriptions.)
Note that the steps in italics have no impact on the response to A’s
queries, we simply continue to answer the queries and only note if the key
turns bad, i.e. we say that a key k is bad w.r.t. the sets Ss and Tt if
28
there exist i, j such that either mi · k = xj or ci · k −1 = yj , and k is good
2st
bad keys.
otherwise. There are at most |G|
Game R: We consider the random game which corresponds to the latter
probability in (2), i.e.
i
h
P,P −1
PR := P r Aπ,π
−1 = 1 .
From the definition of Game R, we see that, letting P rR denote the
probability when playing Game R,
h
i
−1
P rR AP,P
=
1
= PR ,
(3)
E,D
as we are simply giving uniformly random answers to each of A’s queries.
29
Notation: We let Si1 = {m|(m, c) ∈ Si }, Si2 = {c|(m, c) ∈ Si }, Ti1 = {x|(x, y) ∈ Ti }, and
Ti2 = {y|(x, y) ∈ Ti }.
GAME R: Initially, let S0 and T0 be
empty and flag unset. Choose k ∈R G,
then answer the i+1’st query as follows:
GAME X: Initially, let S0 and T0 be
empty and flag unset. Choose k ∈R G,
then answer the i+1’st query as follows:
E-oracle query with mi+1 :
1. Choose ci+1 ∈R G \ Si2 .
2.
If P (mi+1 · k) ∈ Ti2 , or
−1
P (ci+1 · k −1 ) ∈ Ti1 , then set flag to
bad.
3. Define E(mi+1 ) = ci+1 (and thereby
also D(ci+1 ) = mi+1 ) and return ci+1 .
E-oracle query with mi+1 :
1. Choose ci+1 ∈R G \ Si2 .
2. If P (mi+1 · k) ∈ Ti2 then redefine
ci+1 := P (mi+1 · k) · k and set flag to
bad. Else if P −1 (ci+1 · k −1 ) ∈ Ti1 , then
set flag to bad and goto Step 1.
3. Define E(mi+1 ) = ci+1 (and thereby
also D(ci+1 ) = mi+1 ) and return ci+1 .
D-oracle query with ci+1 :
1. Choose mi+1 ∈R G \ Si1 .
2.
If P −1 (ci+1 · k −1 ) ∈ Ti1 , or
P (mi+1 · k) ∈ Ti2 , then set flag to bad.
3. Define D(ci+1 ) = mi+1 (and thereby
also E(mi+1 ) = ci+1 ) and return mi+1 .
D-oracle query with ci+1 :
1. Choose mi+1 ∈R G \ Si1 .
2. If P −1 (ci+1 · k −1 ) ∈ Ti1 then redefine
mi+1 := P −1 (ci+1 · k −1 ) · k −1 and set
flag to bad. Else if P (mi+1 · k) ∈ Ti2 ,
then set flag to bad and goto Step 1.
3. Define D(ci+1 ) = mi+1 (and thereby
also E(mi+1 ) = ci+1 ) and return mi+1 .
P -oracle query with xi+1 :
1. Choose yi+1 ∈R G \ Ti2 .
2.
If E(xi+1 · k −1 ) ∈ Si2 , or
D(yi+1 · k) ∈ Si1 , then set flag to
bad.
3. Define P (xi+1 ) = yi+1 (and thereby
also P −1 (yi+1 ) = xi+1 ) and return yi+1 .
P -oracle query with xi+1 :
1. Choose yi+1 ∈R G \ Ti2 .
2. If E(xi+1 · k −1 ) ∈ Si2 then redefine
yi+1 := E(xi+1 · k −1 ) · k −1 and set flag
to bad. Else if D(yi+1 · k) ∈ Si1 , then
set flag to bad and goto Step 1.
3. Define P (xi+1 ) = yi+1 (and thereby
also P −1 (yi+1 ) = xi+1 ) and return yi+1 .
P −1 -oracle query with yi+1 :
1. Choose xi+1 ∈R G \ Ti1 .
2. If D(yi+1 ·k) ∈ Si1 , or E(xi+1 ·k −1 ) ∈
Si2 , then set flag to bad.
3. Define P −1 (yi+1 ) = xi+1 (and
thereby also P (xi+1 ) = yi+1 ) and return xi+1 .
P −1 -oracle query with yi+1 :
1. Choose xi+1 ∈R G \ Ti1 .
2. If D(yi+1 · k) ∈ Si1 then redefine
xi+1 := D(yi+1 · k) · k and set flag to
bad. Else if E(xi+1 · k −1 ) ∈ Si2 , then set
flag to bad and goto Step 1.
3. Define P −1 (yi+1 ) = xi+1 (and
thereby also P (xi+1 ) = yi+1 ) and return xi+1 .
30
Game X: Consider the experiment which corresponds to the game played
in the prior probability in (2) and define this probability as
i
h
−1
=
1
.
PX := P r AP,P
Ek ,Dk
We define Game X, as outlined above. Note that again the parts in italics
have no impact on the response to A’s queries, however, this time, when a
key becomes bad, we choose a new random value repeatedly for the response
until the key is no longer bad, and then reply with this value. Intuitively,
Game X behaves like Game R except that Game X checks for consistency
as it does not want A to win on some collision. It is non-trivial to see that,
letting P rX denote the probability when playing Game X,
h
i
−1
P rX AP,P
=
1
= PX .
(4)
E,D
The proof is given in Appendix B.
We have defined both games in such a way that their outcomes differ only
in the event that a key turns bad. Thus, any circumstance which causes a
difference in the instructions carried out by the games, will also cause both
games to set the flag to bad. Let BAD denote the event that the flag gets
set to bad and the case that the flag is not set to bad by ¬BAD, then the
two following lemmas follow from the previous statement.
Lemma 40. P rR [BAD] = P rX [BAD] and P rR [¬BAD] = P rX [¬BAD].
h
i
h
i
−1
P,P −1
Lemma 41. P rR AP,P
=
1|¬BAD
=
P
r
A
=
1|¬BAD
.
X
E,D
E,D
Using these two lemmas we are able to prove the lemma:
Lemma 42. Adv(A) ≤ P rR [BAD].
This is because, using (3), (4), and lemmas 40 and 41,
Adv(A) = |PX − PR |
h
i
h
i
P,P −1
P,P −1
= P rX AE,D = 1 − P rR AE,D = 1
h
i
−1
= |P rX AP,P
=
1|¬BAD
· P rX [¬BAD]
E,D
h
i
P,P −1
+ P rX AE,D = 1|BAD · P rX [BAD]
h
i
P,P −1
− P rR AE,D
= 1|¬BAD · P rR [¬BAD]
h
i
P,P −1
− P rR AE,D
= 1|BAD · P rR [BAD] |
h
i
h
i
P,P −1
P,P −1
= P rR [BAD] · P rX AE,D = 1|BAD − P rR AE,D = 1|BAD
≤ P rR [BAD] .
31
Let us now define yet another game, Game R’.
GAME
R’:
Initially,
let
S0
and
T0
be
empty
and
flag
unset.
Answer
the
i + 1’st
query
as
follows:
E-oracle query with mi+1 :
1. Choose ci+1 ∈R G \ Si2 .
2. Define E(mi+1 ) := ci+1 (and thereby also D(ci+1 ) := mi+1 ) and return ci+1 .
D-oracle query with ci+1 :
1. Choose mi+1 ∈R G \ Si1 .
2. Define D(ci+1 ) := mi+1 (and thereby also E(mi+1 ) := ci+1 ) and return mi+1 .
P -oracle query with xi+1 :
1. Choose yi+1 ∈R G \ Ti2 .
2. Define P (xi+1 ) := yi+1 (and thereby also P −1 (yi+1 ) := xi+1 ) and return yi+1 .
P −1 -oracle query with yi+1 :
1. Choose xi+1 ∈R G \ Ti1 .
2. Define P −1 (yi+1 ) := xi+1 (and thereby also P (xi+1 ) := yi+1 ) and return
xi+1 .
After all queries have been answered, choose k ∈R G. If there exists (m, c) ∈ Ss
and (x, y) ∈ Tt such that k becomes bad then set flag to bad.
This game runs as Game R except that it does not choose a key until
all of the queries have been answered and then checks for badness of the flag
(by checking whether or not the key has become bad). It can be shown that
the flag is set to bad in Game R if and only if the flag is set to bad in
Game R’ (by a consideration of cases (see Appendix C.)) Hence, we get the
following lemma.
Lemma 43. P rR [BAD] = P rR0 [BAD].
Using the above lemma, we now only have to bound P rR0 [BAD] in order
to bound Adv(A), but as the adversary queries at most s elements to the
E/D-oracles and at most t elements to the P/P −1 -oracles, and the key k is
chosen uniformly at random from G, we have that the probability of choosing
a bad key is at most 2st/|G|, i.e.
st
.
Adv(A) ≤ P rR0 [BAD] = O
|G|
Restating the theorem, we get:
Theorem 44. For any probabilistic adversary A, limited to polynomially
many E/D- and P/P −1 -oracle queries, the generalized EM scheme over a
group G is a super pseudorandom permutation.
32
B
Proof of probability of Game X
Recall the definition of Si1 , Si2 , Ti1 and Ti2 (see p. 30.) We write Ss and Tt
to denote the final transcripts. We drop the index i if it is understood. We
begin by defining Game X’.
GAME X’: Initially, let S0 and T0 be empty. Choose k ∈R G, then answer the
i + 1’st query as follows:
E-oracle query with mi+1 :
1. If P (mi+1 · k) ∈ Ti2 return P (mi+1 · k) · k
2. Else choose yi+1 ∈R G \ Ti2 , define P (mi+1 · k) = yi+1 , and return yi+1 · k.
D-oracle query with ci+1 :
1. If P −1 (ci+1 · k −1 ) ∈ Ti1 , return P −1 (ci+1 · k −1 ) · k −1 .
2. Else choose xi+1 ∈R G\Ti1 , define P −1 (ci+1 ·k −1 ) = xi+1 , and return xi+1 ·k −1 .
P -oracle query with xi+1 :
1. If P (xi+1 ) ∈ Ti2 , return P (xi+1 ).
2. Else choose yi+1 ∈R G \ Ti2 , define P (xi+1 ) = yi+1 , and return yi+1 .
P −1 -oracle query with yi+1 :
1. If P −1 (yi+1 ) ∈ Ti1 , return P −1 (yi+1 ).
2. Else choose xi+1 ∈R G \ Ti1 , define P −1 (yi+1 ) = xi+1 , and return xi+1 .
Notice that the only difference between Game X’ and the game defining PX is that the latter has defined all values for the oracles beforehand
while the former "defines as it goes." Still, an adversary cannot tell the
difference
betweeni playing the Game X’ or the game defining PX . Thus,
h
P,P −1
P rX 0 AE,D = 1 = PX .
What we wish to show is that
h
i
h
i
−1
P,P −1
0
P rX AP,P
=
1
=
P
r
A
=
1
,
X
E,D
E,D
i.e. that no adversary A may distinguish between playing Game X and
playing Game X’, even negligibly. We will do this by showing that no
adversary A may distinguish between the outputs given by the two games.
As both games begin by choosing a uniformly random key k and as we show
that for this value the games are identical, we hereby assume such a key k
to be a fixed, but arbitrary, value for the remainder of this proof.
Considering the definitions of Game X and Game X’, we see that the
two games define their E/D- and P/P −1 -oracles differently: the former defining both, while the latter defines only the P/P −1 -oracle and computes the
E/D-oracle. We show that Game X also answers its E/D-oracle queries by
referring to P/P −1 , although not directly.
33
Given the partial functions E and P in Game X, i.e. functions having
been defined for all values up to and including the i’th query, define the
partial function Pb as the following.
if P (x) is defined,
P (x)
def
−1
−1
b
P (x) = E(x · k ) · k
if E(x · k −1 ) is defined, and
undefined
otherwise.
Using the above definition, defining a value for E or P implicitly defines
a value for Pb. The first question is, whether or not Pb is well-defined, i.e.
whether there are clashes of values (that is, differences between values differing by other than ·k (or ·k −1 )) for some x for which both P (x) and E(x · k −1 )
are defined.
Lemma 45. Let E and P be partial functions arising in Game X, then the
partial function Pb is well-defined.
Proof. Proof by induction on the number of "Define" steps in Game X (i.e.
steps E −3, D −3, P −3, and P −1 −3) as these are the steps where Pb becomes
defined. The initial case of the induction proof is trivial as S0 and T0 are
empty such that no values may clash. Suppose now that in step E − 3 we
define E(m) = c. The only possibility that Pb becomes ill-defined will occur
if the new E(m) value clashes with a prior defined P (m · k) value: If P (m · k)
was not defined, then no clashes can arise. If P (m · k) was defined, then by
step E − 2, the value is E(m) · k −1 , such that there is no clash.
For D − 3, the argument is similar as E(m) will become defined as well.
Although, for the case where P (m · k) is defined, step D − 2 forces a new
uniformly random value of m to be chosen until no clash occurs.
Analogously, for P and P −1 , no clashes will arise, hence, Pb must be
well-defined.
We may also consider Pb in Game X’, in the sense that when we define a
value for P in the game, we implicitly define a value for Pb where Pb(x) = P (x)
as E(x · k −1 ) = P (x) in Game X’.
We wish now to show that the oracle query-answers of E, D, P, and P −1
in Game X, expressed in terms of Pb, correspond exactly to those in Game
X’.
Case 1: E-oracle query. Beginning with Game X, we first note that
Game X never defines E(m) unless m has been queried to the E-oracle,
or alternately, the D-oracle has been queried with a c such that E(m) = c.
However, as A never repeats a query if it can guess the answer, i.e. never
queries any part of an already defined E/D-oracle pair, we may assume that
34
E(m) is undefined when m is queried. Therefore, we see that concurrently
with m being queried, we have that Pb(m · k) will be defined if and only if
P (m · k) is defined, and Pb(m · k) = P (m · k). Let us consider the two cases:
when Pb(m · k) is defined and when it is undefined.
Case 1a: When Pb(m · k) is defined, then Game X returns c = Pb(m · k) · k.
Setting E(m) = c leaves Pb unchanged, i.e. the value Pb(m · k) remains
the same, unlike the next case.
Case 1b: When Pb(m · k) is undefined, then Game X repeatedly chooses
c ∈R G \ S 2 uniformly until P −1 (c · k −1 ) is undefined, i.e. the set
U = {c ∈ G|P −1 (c · k −1 ) 6∈ T 1 }. It follows that y = c · k −1 is uniformly distributed over G \ Tb2 .4 This can be seen by showing that
S 2 ∪ U { = Tb2 · k, where the only non-triviality in the argument follows from the definition of Pb. In this case, setting E(m) = c also sets
Pb(m · k) = y, in contrast to the prior case as it is now defined.
We now consider the same query on Game X’.
Case 1a’ : When Pb(m · k) = P (m · k) is defined, c = P (m · k) · k is returned, and
Pb is unchanged.
Case 1b’ : When Pb(m·k) = P (m·k) is undefined, we choose y ∈R G\T 2 = G\ Tb2 ,
Pb(m · k) is set to y, and c = y · k is returned.
Thus, the behaviour of Game X and Game X’ are identical on the E-oracle
queries.
We will be briefer in our arguments for the following 3 cases as the arguments are similar.
Case 2: D-oracle query. Here we again assume that no element of
an E/D-oracle pair (m, c), such that E(m) = c, has been queried before.
Like in the above case, we see that, as Pb(m · k) = P (m · k), we also have
Pb−1 (c · k −1 ) = P −1 (c · k −1 ).
Case 2a + 2a’ : When Pb−1 (c · k −1 ) = P −1 (c · k −1 ) is defined, then m = P −1 (c · k −1 ) · k −1
is returned, leaving Pb−1 (c · k −1 ) unchanged in both games.
Case 2b + 2b’ : If Pb−1 (c · k −1 ) = P −1 (c · k −1 ) is undefined, then x ∈R G \ Tb1 is chosen
uniformly and Pb−1 (c · k −1 ) = x, in both cases.
4
Tb1 and Tb2 are the corresponding sets on the query pairs of Pb.
35
Thus, the behaviour of Game X and Game X’ are identical on the D-oracle
queries.
Case 3: P -oracle query. Here we instead assume that no element of a
P/P −1 -oracle pair (x, y) such that P (x) = y, has been queried before.
Case 3a + 3a’ : Using the definition of the E- and P -oracles in Game X and the
definition of Pb we see that P (x) is defined if and only if E(x · k −1 ) is
defined, but then this also holds if and only if Pb(x) is defined (by the
assumption in the beginning of case 3). Hence, if Pb(x) is defined, then
y = E(x · k −1 ) · k −1 = Pb(x). Indeed, both games secure this value.
Case 3b + 3b’ : If Pb(x) is undefined, then y ∈R G \ Tb2 is chosen uniformly and Pb(x) is
defined to be y, in both cases.
Thus, the behaviour of Game X and Game X’ are identical on the P -oracle
queries.
Case 4: P −1 -oracle query. Again, we assume that no element of a
P/P −1 -oracle pair (x, y) such that P (x) = y, has been queried before.
Case 4a + 4a’ : Using the definition of Game X and the definition of Pb, as well as our
case 4 assumption, we see that Pb−1 (y) is defined if and only if D(y · k)
is defined. Hence, if Pb−1 (y) is defined, then x = D(y · k) · k = Pb−1 (y).
Indeed, both games secure this value.
Case 4b + 4b’ : If Pb−1 (y) is undefined, then x ∈R G\ Tb1 is chosen uniformly and Pb−1 (y)
is defined to be x, in both cases.
Thus, the behaviour of Game X and Game X’ are identical on the P −1 oracle queries. Q.E.D.
36
C
Proof that the probability of Game R and
Game R’ match
Recall the definition of Si1 , Si2 , Ti1 and Ti2 (see p. 30). We write Ss and Tt to
denote the final transcripts. We also introduce the following definition.
Definition 46. We say that two E/D-pairs (mi , ci ) and (mj , cj ) overlap
if mi = mj or ci = cj . If mi = mj and ci = cj , we say that the pairs are
identical. Likewise for P/P −1 -pairs (xi , yi ) and (xj , yj ).
If two pairs overlap, then by the definition of the E/D- and P/P −1 oracles, they must be identical. Therefore, WLOG, we may assume that all
queries to the oracles are non-overlapping. Let us now prove the lemma.
Lemma 47. P rR [BAD] = P rR0 [BAD].
Proof. We need to prove that Game R has its flag set to bad if and only if
Game R’ has its flag set to bad.
"⇒": We want to show that there exists (m, c) ∈ Ss and (x, y) ∈ Tt
such that either m · k = x or c · k −1 = y (i.e. such that k becomes bad).
We have to consider the 8 cases where the flag is set to bad. All of the
cases use an analogous argument to the following: If P (m · k) is defined then
P (m · k) = y = P (x) for some (x, y) ∈ Tt such that, as overlapping pairs are
identical, m · k = x.
"⇐": We assume that there exists (m, c) ∈ Ss and (x, y) ∈ Tt such that
k becomes bad. i.e. such that either m · k = x or c · k −1 = y. We need to
check that in all four oracle queries, the flag in Game R is set to bad, which
needs a consideration of 8 cases.
Assume that m · k = x, then
E-oracle on m : P (m · k) = P (x) = y ∈ Tt2 ,
D-oracle on c : P (m · k) = P (x) = y ∈ Tt2 ,
P -oracle on x : E(x · k −1 ) = E(m) = c ∈ Ss2 ,
P −1 -oracle on y : E(x · k −1 ) = E(m) = c ∈ Ss2 .
Assume now that c · k −1 = y, then
E-oracle on m : P −1 (c · k −1 ) = P −1 (y) = x ∈ Tt1 ,
D-oracle on c : P −1 (c · k −1 ) = P −1 (y) = x ∈ Tt1 ,
P -oracle on x : D(y · k) = D(c) = m ∈ Ss1 ,
P −1 -oracle on y : D(y · k) = D(c) = m ∈ Ss1 .
37
| 4math.GR
|
Non-constant bounded holomorphic functions of hyperbolic
numbers – Candidates for hyperbolic activation functions
arXiv:1306.1653v1 [cs.NE] 7 Jun 2013
* Eckhard Hitzer (University of Fukui)
Abstract– The Liouville theorem states that bounded holomorphic complex functions
are necessarily constant. Holomorphic functions fulfill the socalled Cauchy-Riemann
(CR) conditions. The CR conditions mean that a complex z-derivative is independent
of the direction. Holomorphic functions are ideal for activation functions of complex
neural networks, but the Liouville theorem makes them useless. Yet recently the use
of hyperbolic numbers, lead to the construction of hyperbolic number neural networks.
We will describe the Cauchy-Riemann conditions for hyperbolic numbers and show that
there exists a new interesting type of bounded holomorphic functions of hyperbolic
numbers, which are not constant. We give examples of such functions. They therefore
substantially expand the available candidates for holomorphic activation functions for
hyperbolic number neural networks.
Keywords: Hyperbolic numbers, Liouville theorem, Cauchy-Riemann conditions,
bounded holomorphic functions
1
Introduction
For the sake of mathematical clarity, we first carefully
review the notion of holomorphic functions in the two
number systems of complex and hyperbolic numbers.
The Liouville theorem states that bounded holomorphic complex functions f : C → C are necessarily constant [1]. Holomorphic functions are functions
that fulfill the socalled Cauchy-Riemann (CR) conditions. The CR conditions mean that a complex
z-derivative
df (z)
, z = x + iy ∈ C, x, y ∈ R, ii = −1,
dz
Contrary to the complex case, the hyperbolic logistic function is bounded. This is
due to the absence of singularities. Thus,
in general terms, this seems to be a suitable
activation function. Concretely, the following facts, however, might be of disadvantage. The real and imaginary part have different squashing values. Both component
functions do only significantly differ from
zero around the lines1 x = y (x > 0) and
−x = y (x < 0).
(1)
is independent of the direction with respect to which
the incremental ratio, that defines the derivative, is
taken [5]. Holomorphic functions would be ideal for
activation functions of complex neural networks, but
the Liouville theorem means that careful measures
need to be taken in order to avoid poles (where the
function becomes infinite).
Yet recently the use of hyperbolic numbers
z = x + h y, h2 = 1, x, y ∈ R, h ∈
/ R.
boundaries (hyperplanes) of the real and the unipotent h-part of the output. But Buchholz argued in
[4], p. 114, that
(2)
lead to the construction of hyperbolic number neural
networks. We will describe the generalized CauchyRiemann conditions for hyperbolic numbers and show
that there exist bounded holomorphic functions of
hyperbolic numbers, which are not constant. We give
a new example of such a function. They are therefore
excellent candidates for holomorphic activation functions for hyperbolic number neural networks [2, 3].
In [3] it was shown, that hyperbolic number neural
networks allow to control the angle of the decision
Complex numbers are isomorphic to the Clifford
geometric algebra Cl0,1 which is generated by a single
vector e1 of negative square e1 = −1, with algebraic
basis {1, e1 }. The isomorphism C ∼
= Cl0,1 is realized
by mapping i 7→ e1 .
Hyperbolic numbers are isomorphic to the Clifford
geometric algebra Cl1,0 which is generated by a single
vector e1 of positive square e1 = +1, with algebraic
basis {1, e1 }. The isomorphism between hyperoblic
numbers and Cl1,0 is realized by mapping h 7→ e1 .
2
Complex variable functions
We follow the treatment given in [5]. We assume
a complex function given by an absolute convergent
1 Note that we slightly correct the two formulas of Buchholz,
because we think it necessary to delete e1 in Buchholz’ original
x = ye1 (x > 0), etc.
power series.
w = f (z) = f (x + iy) = u(x, y) + iv(x, y),
(3)
where u, v : R2 → R are real functions of the real
variables x, y. Since u, v are obtained in an algebraic way from the complex number z = x + iy, they
cannot be arbitrary functions but must satisfy certain conditions. There are several equivalent ways
to obtain these conditions. Following Riemann, we
state that a function w = f (z) = u(x, y) + iv(x, y)
is a function of the complex variable z if its derivative is independent of the direction (in the complex
plane) with respect to which the incremental ratio is
taken. This requirement leads to two partial differential equations, named after Cauchy and Riemann
(CR), which relate u and v.
One method for obtaining these equations is the
following. We consider the expression w = u(x, y) +
iv(x, y) only as a function of z, but not of z̄, i.e. the
derivative with respect to z̄ shall be zero. First we
perform the bijective substitution
x=
1
(z + z̄),
2
1
y = −i (z − z̄),
2
(4)
based on z = x + iy, z̄ = x − iy. For computing the
derivative w,z̄ = dw
dz̄ with the help of the chain rule
we need the derivatives of x and y of (4)
x,z̄ =
1
,
2
y,z̄ =
1
i.
2
(5)
3
Hyperbolic numbers
Hyperbolic numbers are also known as split-complex
numbers. They form a two-dimensional commutative
algebra. The canonical hyperbolic system of numbers
is defined [5] by
(11)
The hyperbolic conjugate is defined as
z̄ = x − h y.
(6)
Requiring that both the real and the imaginary part
of (6) vanish we obtain the Cauchy-Riemann conditions
u,x = v,y ,
u,y = −v,x .
(7)
Functions of a complex variable that fulfill the CR
conditions are functions of x and y, but they are only
functions of z, not of z̄.
It follows from (7), that both u and v fulfill the
Laplace equation
u,xx = v,yx = v,xy = −u,yy ⇔ u,xx + u,yy = 0, (8)
and similarly
v,xx + v,yy = 0.
g(u(x, y)+iv(x, y)) = gr (u(x, y))+igi (v(x, y)). (10)
z = x + h y, h2 = 1, x, y ∈ R, h ∈
/ R.
Using the chain rule we obtain
w,z̄ = u,x x,z̄ + u,y y,z̄ + i(v,x x,z̄ + v,y y,z̄ )
1
1
1
1
= u,x + iu,y + i( v,x + iv,y )
2
2
2
2
1
!
= [u,x − v,y + i(v,x + u,y )] = 0.
2
are called harmonic functions and are important in
many fields of science, notably the fields of electromagnetism, astronomy, and fluid dynamics, because
they can be used to accurately describe the behavior
of electric, gravitational, and fluid potentials. In the
study of heat conduction, the Laplace equation is the
steady-state heat equation [6].
Liouville’s theorem [1] states, that any bounded
holomorphic function f : C → C, which fulfills the
CR conditions is constant. Therefore for complex
neural networks it is not very meaningful to use holomorphic functions as activation functions. If they
are used, special measures need to be taken to avoid
poles in the complex plane. Instead separate componentwise (split) real scalar functions for the real
part gr : R → R, u(x, y) 7→ gr (u(x, y)), and for the
imaginary part gi : R → R, v(x, y) 7→ gi (v(x, y)), are
usually adopted. Therefore a standard split activation function in the complex domain is given by
(9)
The Laplace equation is a simple example of an elliptic partial differential equation. The general theory
of solutions to the Laplace equation is known as potential theory. The solutions of the Laplace equation
(12)
Taking the hyperbolic conjugate corresponds in the
isomorphic algebra Cl1,0 to taking the main involution (grade involution), which maps 1 7→ 1, e1 7→
−e1 .
The hyperbolic invariant (corresponding to the
Lorentz invariant in physics for y = ct), or modulus,
is defined as
z z̄ = (x + h y)(x − h y) = x2 − y 2 ,
(13)
which is not positive definite.
Hyperbolic numbers are fundamentally different
from complex numbers. Complex numbers and
quaternions are division algebras, every non-zero element has a unique inverse. Hyperbolic numbers do
not always have an inverse, but instead there are
idempotents and divisors of zero.
We can define the following idempotent basis
n1 =
1
(1 + h),
2
n2 =
1
(1 − h),
2
(14)
which fulfills
1
1
(1 + h)(1 + h) = (2 + 2h) = n1 ,
4
4
n22 = n2 ,
n1 + n2 = 1,
1
1
n1 n2 = (1 + h)(1 − h) = (1 − 1) = 0,
4
4
n̄1 = n2 , n̄2 = n1 .
(15)
n21 =
The inverse basis transformation is simply
1 = n1 + n2 ,
h = n1 − n2 .
(16)
Setting
z = x + hy = ξn1 + ηn2 ,
(17)
we get the corresponding coordinate transformation
x=
1
(ξ + η),
2
y=
1
(ξ − η),
2
(18)
as well as the inverse coordinate transformation
ξ = x + y ∈ R,
η = x − y ∈ R.
(19)
Figure 1: The hyperbolic number plane [9] with horizontal x-axis and vertical yh-axis, showing: (a) Hyperbolas with modulus z z̄ = −1 (green). (b) Straight
lines with modulus z z̄ = 0 ⇔ x2 = y 2 (red), i.e. divisors of zero. (c) Hyperbolas with modulus z z̄ = 1
(blue).
The hyperbolic conjugate becomes, due to (15), in
the idempotent basis
z̄ = ξn̄1 + ηn̄2 = ηn1 + ξn2 .
(20)
In the idempotent basis, using (20) and (15), the hyperbolic invariant becomes multiplicative
z z̄ = (ξn1 + ηn2 )(ηn1 + ξn2 )
= ξη(n1 + n2 ) = ξη = x2 − y 2 .
(21)
In the following we consider the product and quotient
of two hyperbolic numbers z, z 0 both expressed in the
idempotent basis {n1 , n2 }
zz 0 = (ξn1 +ηn2 )(ξ 0 n1 +η 0 n2 ) = ξξ 0 n1 +ηη 0 n2 , (22)
due to (15). We repeat that in (24) the product is
zero, even though the factors are non-zero. The numbers ξn1 , ηn2 along the n1 , n2 axis are therefore called
divisors of zero. The divisors of zero have no inverse.
The hyperbolic plane with the diagonal lines of divisors of zero (b), and the pairs of hyperbolas with
constant modulus z z̄ = 1 (c), and z z̄ = −1 (a) is
shown in Fig. 1.
4
Hyperbolic number functions
We assume a hyperbolic number function given by
an absolute convergent power series
and
w = f (z) = f (x + hy) = u(x, y) + hv(x, y),
z
ξn1 + ηn2
z z̄ 0
= 0
= 0 0
0
0
z
ξ n1 + η n2
z z̄
(ξn1 + ηn2 )(η 0 n1 + ξ 0 n2 )
= 0
(ξ n1 + η 0 n2 )(η 0 n1 + ξ 0 n2 )
(ξη 0 n1 + ηξ 0 n2 )(η 0 n1 + ξ 0 n2 )
=
ξ0 η0
ξ
η
= 0 n1 + 0 n2 .
ξ
η
h2 = 1,
h∈
/ R.
(25)
where u, v : R2 → R are real functions of the real
variables x, y. An example of a hyperbolic number
function is the exponential function
ez = ex+hy = ex ehy = ex (cosh y + h sinh y)
= u(x, y) + hv(x, y),
(23)
(26)
with
Because of (23) it is not possible to divide by z 0 if
ξ 0 = 0, or if η 0 = 0. Moreover, the product of a
hyperbolic number with ξ = 0 (on the n2 axis) times
a hyperbolic number with η = 0 (on the n1 axis) is
(ξn1 + 0n2 )(0n1 + ηn2 ) = ξηn1 n2 = 0,
(24)
u(x, y) = ex cosh y,
v(x, y) = ex sinh y.
(27)
Since u, v are obtained in an algebraic way from the
hyperbolic number z = x + hy, they cannot be arbitrary functions but must satisfy certain conditions.
There are several equivalent ways to obtain these conditions. A function w = f (z) = u(x, y) + hv(x, y) is
a function of the hyperbolic variable z, if its derivative is independent of the direction (in the hyperbolic
plane) with respect to which the incremental ratio is
taken. This requirement leads to two partial differential equations, so called generalized Cauchy-Riemann
(GCR) conditions, which relate u and v.
To obtain the GCR conditions we consider the expression w = u(x, y) + hv(x, y) only as a function of
z, but not of z̄ = x − hy, i.e. the derivative with
respect to z̄ shall be zero. First we perform the bijective substitution
x=
1
(z + z̄),
2
1
y = h (z − z̄),
2
(28)
based on z = x + hy, z̄ = x − hy. For computing the
derivative w,z̄ = dw
dz̄ with the help of the chain rule
we need the derivatives of x and y of (28)
x,z̄ =
1
,
2
1
y,z̄ = − h.
2
w,z̄ = u,x x,z̄ + u,y y,z̄ + h(v,x x,z̄ + v,y y,z̄ )
1
1
1
1
= u,x − hu,y + h( v,x − hv,y )
2
2
2
2
1
!
= [u,x − v,y + h(v,x − u,y )] = 0.
(30)
2
Requiring that both the real and the h-part of (30)
vanish we obtain the GCR conditions
u,y = v,x .
f (z) = u(x, y) + h v(x, y),
1
.
u(x, y) = v(x, y) =
1 + e−x e−y
(31)
Functions of a hyperbolic variable that fulfill the
GCR conditions are functions of x and y, but they
are only functions of z, not of z̄. Such functions are
called (hyperbolic) holomorphic functions.
It follows from (31), that u and v fulfill the wave
equation
−1
(−e−x e−y )
(1 + e−x e−y )2
e−x e−y
=
,
(1 + e−x e−y )2
u,x =
u,y = v,x = v,y =
0<
1
1+
u,x = ex cosh y,
e−x e−y
lim
v,y = ex cosh y = u,x . (34)
< 1.
1
= 0,
1 + e−x e−y
and
lim
x,y→∞
1
1+
(38)
e−x e−y
= 1.
(39)
(40)
The function (35) is representative for how to turn
any real neural node activation function r(x) into
holomorphic hyperbolic activation function via
f (x) = r(x + y) (1 + h).
(41)
We note that in [3, 4] another holomorphic hyperbolic activation function was studied, namely
u,y = ex sinh y,
v,x = ex sinh y = u,y ,
(37)
We especially have
and similarly
The wave equation is an important second-order
linear partial differential equation for the description
of waves – as they occur in physics – such as sound
waves, light waves and water waves. It arises in fields
like acoustics, electromagnetics, and fluid dynamics.
The wave equation is the prototype of a hyperbolic
partial differential equation [7].
Let us compute the partial derivatives u,x , u,y ,
v,x , v,y for the exponential function ez of (26):
e−x e−y
.
(1 + e−x e−y )2
The GCR conditions (31) are therefore clearly fulfilled, which means that the hyperbolic function f (z)
of (35) is holomorphic. Since the exponential function e−x has a range of (0, ∞), the product e−x e−y
also has values in the range of (0, ∞). Therefore the
function 1 + e−x e−y has values in (1, ∞), and the
components of the function f (z) of (35) have values
x,y→−∞
(33)
(36)
where we repeatedly applied the chain rule for differentiation. Similarly we obtain
u,xx = v,yx = v,xy = u,yy ⇔ u,xx − u,yy = 0, (32)
v,xx − v,yy = 0.
(35)
The function u(x, y) is pictured in Fig. 2.
Let us verify that the function f of (35) fulfills the
GCR conditions
(29)
Using the chain rule we obtain
u,x = v,y ,
We clearly see that the partial derivatives (34) fulfill
the GCR conditions (31) for the exponential function
ez , as expected by its definition (26). The exponential function ez is therefore a manifestly holomorphic
hyperpolic function, but it is not bounded.
In the case of holomorphic hyperbolic functions the
GCR conditions do not imply a Liouville type theorem like for holomorphic complex functions. This can
most easily be demonstrated with a counter example
f 0 (z) =
1
,
1 + e−z
(42)
2. x2 > y 2 , x < 0:
θ = artanh (y/x),
z = −ρehθ ,
i.e. the quadrant in Fig. 1 including the negative
x-axis (to the left).
3. x2 < y 2 , y > 0:
θ = artanh (x/y),
z = hρehθ ,
i.e. the quadrant in Fig. 1 including the positive
y-axis (top).
4. x2 < y 2 , y < 0:
θ = artanh (x/y),
Figure 2: Function u(x, y) = 1/(1 + e−x e−y ). Horizontal axis −3 ≤ x ≤ 3, from left corner into paper
plane −3 ≤ y ≤ 3. Vertical axis 0 ≤ u ≤ 1. (Figure
produced with [8].)
but compare the quote from [4], p. 114, given in the
introduction. The split activation function used in
[2]
1
1
+h
,
(43)
f 00 (x, y) =
−x
1+e
1 + e−y
is clearly not holomorphic, because the real part u =
1/(1 + e−x ) depends only on x and not on y, and the
h-part v = 1/(1 + e−y ) depends only on y and not on
x, thus the GCR conditions (31) can not be fulfilled.
5
Geometric interpretation of
multiplication of hyperbolic
numbers
In order to geometrically interpret the product of two
complex numbers, it proves useful to introduce polar
coordinates in the complex plane. Similarly, for the
geometric interpretation of the product of two hyperbolic numbers, we first introduce hyperbolic polar
coordinates for z = x + hy with radial coordinate
p
p
ρ = |z z̄| = |x2 − y 2 | .
(44)
The hyperbolic polar coordinate transformation [5] is
then given as
1. x2 > y 2 , x > 0:
θ = artanh (y/x),
z = ρehθ ,
i.e. the quadrant in the hyperbolic plane of Fig.
1 limitted by the diagonal idempotent lines, and
including the positive x-axis (to the right).
z = −hρehθ ,
i.e. the quadrant in Fig. 1 including the negative
y-axis (bottom).
The product of a constant hyperbolic number (assuming a2x > a2y , ax > 0)
a = ax + hay = ρa ehθa ,
q
θa = artanh (ay /ax ),
ρa = a2x − a2y ,
(45)
with a hyperbolic number z (assuming x2 > y 2 , x >
0) in hyperbolic polar coordinates is
az = ρa ehθa ρ ehθ = ρa ρ eh(θ+θa ) .
(46)
The geometric interpretation is a scaling of the modulus ρ → ρa ρ and a hyperbolic rotation (movement
along a hyperbola) θ → θ + θa .
In the physics of Einstein’s special relativistic
space-time [11, 12], the hyperbolic rotation θ → θ+θa
corresponds to a Lorentz transformation from one
inertial frame with constant velocity tanh θ to another inertial frame with constant velocity tanh(θ +
θa ). Neural networks based on hyperbolic numbers
(dimensionally extended to four-dimensional spacetime) should therefore be ideal to compute with electromagnetic signals, including satellite transmission.
6
Conclusion
We have compared complex numbers and hyperbolic
numbers, as well as complex functions and hyperbolic
functions. We saw that according to Liouville’s theorem bounded complex holomorphic functions are necessarily constant, but non-constant bounded hyperbolic holomorphic functions exist. One such function
has already beeng studied in [3, 4]. We have studied a promising example of a hyperbolic holomorphic
function
1+h
,
(47)
f (z) =
1 + e−x−y
in some detail. The distinct notions of idempotents
and divisors of zero, special to hyperbolic numbers,
were introduced. After further introducing hyperbolic polar coordinates, a geometric interpretation of
the hyperbolic number multiplication was given.
Hyerbolic neural networks offer, compared to complex neural networks, therefore the advantage of suitable bounded non-constant hyperbolic holomorphic
activation functions. It would certainly be of interest
to study convergence, accuracy and decision boundaries of hyperbolic neural networks with the activation function (35), similar to [3, 4].
Acknowledgment
I want to acknowledge God [13]:
In the beginning was the Word2 , and the
Word was with God, and the Word was
God. He was with God in the beginning.
Through him all things were made; without
him nothing was made that has been made.
In him was life, and that life was the light
of all mankind.
I want to thank my dear family, as well as T. Nitta
and Y. Kuroe.
References
[1] K. Guerlebeck et al, Holomorphic Functions in
the Plane and n-dimensional Space, Birkhauser,
2008, chp. 7.3.3.
[2] S. Buchholz, G. Sommer, A hyperbolic multilayer perceptron, Proceedings of the International Joint Conference on Neural Networks,
Como, Italy, vol. 2, 129/133 (2000).
[3] T. Nitta, S. Buchholz, On the Decision Boundaries of Hyperbolic Neurons, Proceedings of
the International Joint Conference on Neural Networks, IJCNN’08-HongKong, June 1-6,
2973/2979(2008).
[4] S. Buchholz, PhD Thesis, A Theory of Neural
Computation with Clifford Algebras, University
of Kiel, 2005.
[5] F. Catoni et al, The Mathematics of Minkowski
Space-Time, Birkhauser, 2008.
[6] Laplace’s equation, Wikipedia, accessed 24 August 2011, http://en.wikipedia.org/wiki/
Laplace’s_equation
2 Greek term: logos. Note: A Greek philosopher named
Heraclitus first used the term Logos around 600 B.C. to designate the divine reason or plan which coordinates a changing
universe. [14]
[7] Wave equation, Wikipedia, accessed 24 August
2011, http://en.wikipedia.org/wiki/Wave_
equation
[8] Online
3D
function
grapher,
//www.livephysics.com/ptools/
online-3d-function-grapher.php?
http:
[9] Split-complex number, Wikipedia, accessed
29 August 2011, http://en.wikipedia.org/
wiki/Split-complex_number
[10] Notes of collaboration with H. Ishi, Feb. 2011,
p. 15.
[11] C. Doran and A. Lasenby, Geometric Algebra for
Physicists, Cambridge University Press, Cambridge (UK), 2003.
[12] E. Hitzer, Relativistic Physics as Application of
Geometric Algebra, in K. Adhav (ed.), Proceedings of the International Conference on Relativity 2005 (ICR2005), University of Amravati, India, January 2005, 71/90(2005).
[13] The Bible, New International Version (NIV),
The Gospel according to John, chapter 1, verses
1-4, http://www.biblegateway.com/
[14] Strong’s Bible lexicon entry G3056 for logos, available online at Blue Letter Bible.
http://www.blueletterbible.org/lang/
lexicon/lexicon.cfm?Strongs=G3056&t=KJV
| 9cs.NE
|
On Oscillations in the Social Force Model
Tobias Kretz
PTV Group, D-76131 Karlsruhe, Germany
[email protected]
arXiv:1507.02566v1 [physics.soc-ph] 9 Jul 2015
July 10, 2015
Abstract
The Social Force Model is one of the most prominent models of pedestrian dynamics. As such naturally
much discussion and criticism has spawned around it, some of which concerns the existence of oscillations in the
movement of pedestrians. This contribution is investigating under which circumstances, parameter choices, and
model variants oscillations do occur and how this can be prevented. It is shown that oscillations can be excluded
if the model parameters fulfill certain relations. The fact that with some parameter choices oscillations occur
and with some not is exploited to verify a specific computer implementation of the model.
1
Introduction
The Social Force Model of pedestrian dynamics is a model that aims at describing the movement of pedestrians
with the predominant purpose of simulating pedestrian movement on computers. The force of pedestrian β on
pedestrian α typically has the form
f~αβ = Aα w()e(−g()) êαβ
(1)
where g() is a function which grows with increasing distance between both pedestrians and can depend on the
velocities of one or both pedestrians. The function w() suppresses forces the more pedestrian β is located outside
the current walking direction of pedestrian α.
The Social Force Model has first been introduced in 1995 [1]. This variant later was called “elliptical specification I”. A second variant (circular specification) has been proposed in 2000 [2] and a third variant (elliptical
specification II) in 2007 [3]. The difference between the three variants lies mainly in the way the velocities of
two interacting pedestrians are considered in the computation of the force between them. The 1995 variant
considers only the velocity of the pedestrian who exerts the force. The 2000 variant does not at all consider
velocities (only the distance between pedestrians) and the 2007 variant considers the relative velocity between
both pedestrians (the pedestrian who exerts the force and the pedestrian on whom the force acts). For the
analytical considerations in this paper mainly the simplest variant from 2000 will be considered. Nevertheless it
will also be discussed how results will change qualitatively if the variants as of 1995 or 2007 are applied.
Under “oscillations” in this paper unrealistic artifacts in the trajectory of a pedestrian approaching another
pedestrian, his destination or a wall is understood. The occurrence of oscillations in the sense of this contribution
has been discussed in a number of contributions [4–7] and it is often claimed that oscillations cannot be avoided
in the Social Force Model, but that they just can be made small. In this paper it will be shown that this is not
correct and exact conditions for the value of the model parameter such that oscillations occur will be derived.
In the remainder of the paper first a single pedestrian approaching a destination is investigated, then a
pedestrian approaching another pedestrian who is standing still and finally two pedestrians approaching each
other. In each case the model is reduced to one dimension and the noise term is set to zero. In the first section
on approaching a destination the problem will be shown to be most severe as with certain conditions oscillations
cannot be prevented and continue infinitely long. At the same time – as will be argued – for this case it is not
very relevant, as there are simple, pragmatic solutions. The second and third case yield restrictions to the choice
of parameters which can produce realistic, oscillation-free behavior.
2
A pedestrian approaching a destination
In this section we are interested in and discuss the equations of motion and their solution of a single pedestrian
approaching a destination coordinate (i.e. a point) where he is required to come to a standstill. We assume that
in the beginning the pedestrian is walking with his desired speed v0 straight towards the destination coordinate,
so there is no tangential component of the walking velocity. Then we can describe the pedestrian as walking
from positive x coordinate into negative x direction towards the destination which is at x = 0. Since the 1995,
1
2000, and 2007 variants of the Social Force Model only differ in the force between pedestrians and not the driving
force term all results of this section hold for all three variants.
We assume for now, that the desired velocity is always some externally given v0 and is always pointing from
the pedestrians current position towards x = 0. This assumption is the simplest one and it can be questioned –
as we will do below. With it it is obvious that there will be oscillations around x = 0. Our intention here is to
investigate the quantitative details of these oscillations.
In general the equation of motion for this pedestrian reads
ẍ(t) =
−sign(x(t))v0 − ẋ(t)
τ
(2)
where τ is an external parameter which typically has values between 0.1 and 1.0 seconds.
We require the pedestrian not only to reach x = 0, but also to stand still there as arrival condition. Because
the pedestrian has a speed larger 0 (or, considering walking direction: smaller 0) he will walk over the destination
and be on the left (negative) side of x coordinates. There the desired velocity points into the direction of positive
x coordinates. So we have for the time following the moment when the pedestrian is at x = 0:
ẍ(t) =
v0 − ẋ(t)
τ
(3)
This is solved by
ẋ(t)
t
v0 − ae− τ
=
x(t)
(4)
− τt
=
b + v0 t + aτ e
(5)
where a and b are integration constants which need to be determined by initial conditions.
We choose t = 0 at the moment when the pedestrian is at x = 0. Then ẋ(t = 0) = −v0 . However, for later
usage we want to set here more general ẋ(t = 0) = −u and remember that for our particular case u = v0 . With
the two conditions x(0) = 0 and ẋ(0) = −u we can determine the values of the integration constants:
a
=
v0 + u
(6)
b
=
−(v0 + u)τ
(7)
So we have
ẋ(t)
=
x(t)
=
t
v0 − (v0 + u)e− τ
t
v0 t − (v0 + u)τ 1 − e− τ
(8)
(9)
Now we can compute the time tturn when the pedestrian stops (and turns around) ẋ(t0 ) = 0 and the position
x(t0 ) at which this happens:
tturn
=
x(tturn )
=
v0 + u
v
0
u
u
τ v0 ln 1 +
−
v0
v0
τ ln
(10)
(11)
In the initial case, when u = v0 this simplifies to tturn = τ ln(2) and x(tturn ) = τ v0 (ln(2) − 1).
This is only half the way to go. The actual question is how fast a pedestrian returns to x = 0 when he has
passed it before with speed u and how long such a loop takes.
Now we choose t = 0 for the moment when the pedestrian is standing still at x(tturn ). The time treturned
at which the pedestrian returns to x = 0 therefore has to be understood as time after tturn and not as absolute
point in time.
We begin again by determining the value of the integration constants. From equation (4) and ẋ(0) = 0
follows a = v0 . With equations (5) and (11) we have
u
u
x(0) = b + v0 τ = v0 τ ln 1 +
−
(12)
v0
v0
u
u
b = v0 τ ln 1 +
− 1+
(13)
v0
v0
t
u
t
u
x(t) = v0 τ ln 1 +
− 1+
+ + e− τ
(14)
v0
v0
τ
Because x(treturned ) = 0 we have the following equation for treturned :
treturned
u
u
treturned
τ
− 1+
+
+ e−
0 = v0 τ ln 1 +
v0
v0
τ
0
=
φreturned − α + ln(α) + e−φreturned
2
(15)
(16)
with the substitutions φreturned = treturned /τ and α = 1 + u/v0 . We do another substitution ϕ = φreturned −
α + ln(α), transforming the last equation to
0
−α
−αe
=
=
ϕ + e−(ϕ+α−ln(α))
ϕe
ϕ
Obviously one solution is ϕ = −α, but the general solution is
ϕ = W −αe−α
(17)
(18)
(19)
where W () is the Lambert W function [8–10], which by definition is the inverse relation of f (y) = yey .
Resubstituting ϕ and φreturned (for reasons of convenience we do not resubstitute α) we have:
treturned
= W −αe−α + α − ln(α)
τ
(20)
In the interval y ∈ [−1/e..0] W (y) has two branches, denoted W−1 (y) and W0 (y); and with 1 ≤ α ≤ 2 for
−αe−α we are within this interval where W (y) has two branches. It holds
W−1 −αe−α = −α
(21)
In this case it would be treturned = −tturn . Thus the W−1 branch gives the backwards in time solution which
we are not interested in here (because we already have computed it above). Therefore we have to continue with
the W0 branch for which
W0 −αe−α 6= −α
(22)
although W is the inverse relation of f (y) = yey . In the remainder we write W for W0 .
= W (−2/e2 ) + 2 − ln(2) = 0.90047.
In case u = v0 , i.e. α = 2 the numerical value of the solution is treturned
τ
Using equation (20) in equation (4) we get the speed ẋ(treturned ) at the time when the pedestrian returns to
x = 0 in dependence from u, which is the speed at the time when the pedestrian last was at x = 0:
ẋ(treturned ) = v0 1 + W −αe−α
(23)
where we have used the defining equation of the W function y = W (y)eW (y) .
From here on the properties have an index that determines the recurrence to the offspring. For example t0
is the (absolute) time when the pedestrian is for the first time at the offspring, t1 denotes the (absolute) time
when the pedestrian returns for the first time to the offspring and so on. In the case of properties that do not
describe passage of the offspring, but the loop durance and turning position the index enumerates the loops.
This means that for these two properties there is no index 0 and that tn = tn−1 + ∆tn .
Equation (23) means that for the n + 1th passage of the offspring αn+1 depends on αn like
(24)
αn+1 = 2 + W −αn e−αn
The time ∆tn+1 it takes for the n + 1th loop depends on αn like
∆tn+1
= αn + W −αn e−αn .
τ
(25)
Rewriting equation (11) in terms of α and writing |xn | for the turn around distance of the nth loop we have
|xn+1 |
= αn − 1 − ln(αn )
τ v0
(26)
Table 1 shows the results from the first 30 passages if the process begins at t = 0 and with speed v0 . Figures
(1) to (3) visualize these data.
As typical values are τ = 0.4 s and v0 = 1.5 m/s it takes about 7 passages before the amplitude |xn | gets
smaller than 1 cm. This takes about 2.1 seconds. The 7th oscillation has a period of 0.15 seconds. To resolve
this a computer implementation would have to have a time step which is at the very maximum half that value
0.075 seconds or 14 simulation steps per second.
Because αn+1 = 1 only if W (−αn e−αn ) = −1 which in turn is only the case if αn = 1 there will be infinitely
many oscillations and in A a proof is given that the sum of the ∆tn diverges, i.e. the pedestrian will oscillate
infinitely long around the destination coordinate.
At first sight this is an unsatisfactory result with regard to the Social Force Model, as such oscillations are
unrealistic. However, we have to ask how realistic our assumptions and initial conditions were. This concerns in
particular the desired velocity ~v0 . We demanded in the beginning that the pedestrian should come to a stop at
x = 0. Nevertheless we have set the desired velocity all of the time to one particular value |~v0 | > 0. This is too
simplistic. Real persons plan ahead and adjust their desired speed: if they desire to stop at a certain position
3
Passage/Loop(n)
0
1
2
3
4
5
6
7
8
9
14
19
29
|xn |
τ v0
∆tn
τ
tn
τ
un
v0
0.307
0.128
0.071
0.045
0.031
0.023
0.017
0.014
0.011
0.005
0.003
0.001
1.594
1.018
0.754
0.600
0.499
0.428
0.374
0.332
0.299
0.199
0.150
0.100
0.000
1.594
2.611
3.365
3.966
4.465
4.893
5.266
5.599
5.898
7.062
7.898
9.088
1.000
0.594
0.424
0.330
0.270
0.229
0.199
0.175
0.157
0.142
0.096
0.073
0.049
αn
2.000
1.594
1.424
1.330
1.270
1.229
1.199
1.175
1.157
1.142
1.096
1.073
1.049
Table 1: Numerical results for the first 30 passages resp. oscillations. The value for αn is understood before an
oscillation, while ∆tn /τ is the period of that oscillation, and the total time tn /τ after that same oscillation. All
values are dimensionless. The numerical values of the Lambert W function have been computed using Wolfram
Alpha [11].
Figure 1: Left: Value of αn at the beginning of oscillation n. Right: Maximum amplitude |xn |/(v0 τ ) of oscillation
n.
Figure 2: Left: Period ∆tn /τ of oscillation n. Right: Total time tn /τ at nth passage of offspring (i.e. time after
nth oscillation; t=0 is when the pedestrian reaches the offspring for the first time and with a speed v0 ).
4
Figure 3: This shows the evolution of the amplitude xn /(v0 τ ) over time in [s].
they adapt their desired velocity beforehand to just achieve that. So we have to ask how we have to modify v0
dynamically to account for that.
More precisely we ask: at which distance db does a pedestrian have to start braking in the Social Force
Model, if he sets vb as desired speed opposing his current speed u? And how long does it take before he comes
to a stand still?
For this the equation of motion reads
vb − ẋ(t)
(27)
ẋ(t) =
τ
with the following (initial) conditions:
ẋ(t = 0)
=
−u
(28)
x(t = 0)
=
d
(29)
ẋ(t0 )
=
0
(30)
x(t0 )
=
0
(31)
where we have defined the moment where the pedestrian starts to brake as t = 0 and the moment where he stops
at x = 0 as t0 ; d is the distance before stand still at which braking has to begin which we are looking for.
Solving this results in
u
t0 = τ ln 1 +
(32)
vb
u
u
db = τ vb
− ln 1 +
(33)
vb
vb
This gives finite and positive t0 and db for all vb > 0. Thus, it is not sufficient to set the desired speed to zero
for a pedestrian who wants to stop as this would imply an infinitely large braking distance and an infinitely long
time to come to a stand still. If we assume that a pedestrian for braking can at maximum have a desired speed
vb = v0 the minimum time for braking is t0 = τ ln(2) and the minimum braking distance is db = τ v0 (1 − ln(2))
to come to standstill from an initial speed u = v0 .
A different and pragmatic solution of the problem is that in real-world planning applications usually there is
no requirement set for the speed at which a pedestrian arrives at a destination. It is furthermore usual that as
destination not a point is given but that a cross section has to be crossed or the pedestrian has to move into a
specific area. If no restriction is given for the speed at arrival the problem disappears.
A third objection is that also a real person cannot comply to the request to stand still with his or her center
exactly at some given coordinate. Real people always sway a little bit when they (try to) stand still [12]. Why
then should one require this from a simulated pedestrian? The oscillations around the destination point which
were found here in their functional form surely do not match the swaying of real people when they stand “still”,
but the amplitude of the ones of the model quickly fall below those of real people. When the real swaying is not
required to be reproduced by models implicitly a required limit to precision is set and any movement of that
order of magnitude or below should be acceptable.
5
3
A pedestrian approaching a standing pedestrian
3.1
Theory
Now we want to investigate the situation when some pedestrian M is approaching another (pedestrian S) one
who is standing at a fixed position – again in one dimension, for an empirical investigation see for example [13].
Since the pedestrian who is exerting the force is not moving, the 1995 variant (elliptical specification I) of the
Social Force Model will produce the same result as the 2000 variant (circular specification) which is the one we
are investigating, but different from the 2007 variant (elliptical specification II) where the speed of the pedestrian
on whom the force acts modifies the strength of the force.
Assume pedestrian S is standing still at x = 0, facing into positive x direction and pedestrian M is approaching
from there, i.e. pedestrian M has positive x coordinate and a speed directed towards negative x, i.e. ẋ(t = 0) < 0
as well as the desired speed v0 > 0 which is constant in time. It is assumed that pedestrian M at time t=0 is far
away from pedestrian S and walking with his desired walking speed ẋ(t = 0) = −v0 .
Then the equation of motion for pedestrian M in the 1995 and 2000 variants of the Social Force Model is
ẍ(t) =
x(t)−2R
−v0 − ẋ(t)
+ Ae− B
τ
(34)
with R as radius of a pedestrian (we assume here for simplicity that all pedestrians have the same radius) and
τ , A, and B parameters of the Social Force Model.
Pedestrian M will come to a standstill at distance
Aτ
+ 2R
(35)
ds = B ln
v0
As pedestrians should not bump one into another ds must be larger than 2R. Therefore we have a first condition
for the parameters of the Social Force Model to yield realistic results:
Aτ > v0
(36)
We assume that oscillations only occur generally if they also occur close to equilibrium. Thus, if there are
no oscillations when pedestrian M is already close to x = ds then there are no oscillations at all. Expanding the
exponential function in equation (34) into a Taylor series around x = ds gives
˙
¨ + ξ(t) + v0 ξ(t) ≈ 0
ξ(t)
τ
Bτ
(37)
Equation (37) is the equation of the damped harmonic oscillator, with the known three solutions: under damped
(pedestrian M approaches the equilibrium point oscillating around it), critically damped (pedestrian M approaches the equilibrium point as quick as possible but without oscillations) and over damped (pedestrian M
approaches the equilibrium point slower than would be possible without having oscillations. Which of the three
cases is realized depends on the relation of parameters:
v0 τ
B
v0 τ
4
B
v0 τ
4
B
4
>
1 ↔ under damped
(38)
=
1 ↔ critically damped
(39)
<
1 ↔ over damped
(40)
Thus in addition to equation (36) with
4v0 τ ≤ B
(41)
we have a second condition for the parameters of the 1995 and 2000 variants of the Social Force Model to have
a chance to yield realistic results.
3.2
Implications for parameters found in literature
In the literature one can find values for some or all of the parameters A, B, τ , and v0 gained from empirical
observations or laboratory experiments. If all four parameters are given, one can use relations (36) and (41) to
test if realistic results for the scenario discussed here can be expected. If not for all four parameters values are
given those two equations allow to set limits. Table 2 shows a compilation of such data and how they relate
to or with relations (36) and (41). It can be seen that where all four parameters were taken from experiment
relation (36) easily holds while relation (41) is clearly violated. Whereas if parameters A and B (and τ ) have
been calibrated or given in a work then relations (36) and (41) require pedestrians to walk quite slowly, while if
v0 and τ are calibrated or given it needs relatively large values for A and B that relations (36) and (41) hold.
This is an indication that with the circular specification (alone) the parameter space that yields realistic results
is small.
6
Source
[3]
A in [m/s2 ]
0.42 ± 0.26
B in [m]
1.65 ± 1.01
[2]
[14]
[15]
[15]
[16]1
26.67†
> 2.25
0.16 ± 0.01
0.45 ± 0.41
12.0 ± 0.2
0.08
≥ 3.34
4.16 ± 0.65
13.5 ± 18.4
0.16 ± 0.08
τ in [s]
τ > v0 /Amax
τ ≤ Bmax /v0 /4
0.5
0.61
0.5 ‡
0.5 ‡
1.09 ± 0.35
v0 in [m/s]
< 0.67
eq. (36)
eq. (41)
0.8
1.37
< 0.087
< 0.43
1.34 ± 0.21
OK
violated
OK
violated
Table 2: Bold face marks value from literature; normal face is computed with equations (36) and (41). Where one
parameter was calculated from another one where a range is given the range was utilized such that the parameter
space for the derived parameter is as large as possible. † : from 2000N/75kg. ‡ : Assumption and input for
calibration. 1 : The simplified Nomad model discussed in said publication is identical to the circular specification
of the Social Force Model, except that the radius of pedestrians is not stated explicitly in the exponent. If – what
is not stated explicitly in said contribution – “distance” for the Nomad model means body surface to body surface
the parameter meanings are identical, while if it means center to center distance the value of parameter A would
change by a factor e−2R/B and even equation (36) could be violated.
3.3
Elliptical specification II in comparison
For elliptical specification II a rigid analysis is much more difficult than for the circular specification since the
elliptical specification II does not only depend on the distance of the two pedestrians but also on their relative
velocity. Still one can estimate the effect of the added consideration of relative velocity on the occurrence
of oscillations. In the elliptical specification II of the Social Force Model [3] the force from pedestrian β on
pedestrian α is defined as
˙
w(θαβ (d~αβ , ~r˙α ))~g (d~αβ , d~αβ )
1 + cos(θαβ (d~αβ , ~r˙α ))
λα + (1 − λα )
2
~r˙α · d~αβ
−
|~r˙α ||d~αβ |
(43)
=
~rα − ~rβ
(45)
=
~ ~ Vαβ (d~αβ , d~˙αβ )
−∇
dαβ
(46)
f~α (~rα , ~rβ , ~r˙α , ~r˙β )
=
w(θαβ (d~αβ , ~r˙α ))
=
cos(θαβ (d~αβ , ~r˙α ))
=
d~αβ
˙
~g (d~αβ , d~αβ )
˙
Vαβ (d~αβ , d~αβ )
=
˙
bαβ (d~αβ , d~αβ )
=
~
~˙
b
(d
αβ ,dαβ )
− αβ B
α
Aα Bα e
q
1
˙
˙
(|d~αβ | + |d~αβ + d~αβ ∆tα |)2 − (d~αβ ∆tα )2
2
(42)
(44)
(47)
(48)
where ~r gives the position of pedestrians and A, B, λ, and ∆t are model parameters. Pedestrians’ positions and
derived properties are time dependent, other properties are constant. In one dimension and with pedestrians
facing each other this simplifies to
cos(θαβ )
=
1
w(θαβ )
=
1
(50)
dαβ
=
(51)
fα (xα , xβ , ẋα , ẋβ )
=
xα − xβ w.l.o.g. assumed to be > 0
d
g(dαβ , d˙αβ ) = −
Vαβ (dαβ , d˙αβ )
ddαβ
Vαβ (dαβ , d˙αβ )
=
bαβ (dαβ , d˙αβ )
=
=
=
(49)
b
(d
,ḋ
)
− αβ αβ αβ
Bα
Aα Bα e
q
1
(dαβ + |dαβ + d˙αβ ∆tα |)2 − (d˙αβ ∆tα )2
2
0 for (dαβ + d˙αβ ∆tα ) ≤ 0
q
d2αβ + dαβ d˙αβ ∆tα otherwise
7
(52)
(53)
(54)
(55)
(56)
Therefore the force is either zero if (dαβ + d˙αβ ∆tα ) ≤ 0 or it is
fα (xα , xβ , ẋα , ẋβ )
=
g(dαβ , d˙αβ )
(57)
r
=
2dαβ + d˙αβ ∆tα
−
e
Aα q
2
˙
2 dαβ + dαβ dαβ ∆tα
=
d¯a − d̄g
Aα ¯ e Bα
dg
d2 +dαβ ḋαβ ∆tα
αβ
Bα
(58)
(59)
with d¯a being the arithmetic and d¯g the geometric mean of current and projected distance (dαβ resp. dαβ +
d˙αβ ∆tα ). It can directly be seen that for large distances equation (58) reduces approximately to the circular
specification from [2] which depends only on distance.
For a pedestrian α approaching another pedestrian β from positive x values d is positive and d˙ is negative.
Therefore and because of the inequality of arithmetic and geometric means it holds dαβ + d˙αβ ∆tα < d¯g < d¯a <
˙
dαβ . Then obviously as long as d > −d∆t
the force as in equation (58) is larger compared to the circular
specification and consequently pedestrian α will not overshoot for certain parameter choices where he does so
with the circular specification.
In case pedestrian α overshoots over the equilibrium point and turns around it would be desirable that the
force is not larger but smaller than with the circular specification. However, since in this case d˙ becomes positive
it holds that dαβ < d¯g < d¯a < dαβ + d˙αβ ∆tα and therefore in equation (59) the exponential factor gives a smaller
value than with the circular specification, yet the fraction factor has a value > 1 and may for large values of
parameter B outweigh the damping effect from the modification in the exponential function. There are three
indications that also in this phase of pedestrian α’s movement and in general oscillations are suppressed with
elliptical specification II: first, for large values of parameter B already in the circular specification there are no
oscillations as equation (41) tells us. This means that where in an isolated view on the “way back” the problem
is most pronounced the system may actually not even evolve to the point that it exists.
Second, one can expand equation (58) in a series for small values of d˙αβ ∆tα /dαβ :
d
dαβ d˙αβ ∆tα
− Bαβ
α
fα (xα , xβ , ẋα , ẋβ ) ≈ Aα e
1−
(60)
Bα dαβ
In the moment pedestrian α turns around it is d˙αβ = 0 and therefore circular and elliptical specification II
yield an identical force. Starting to move backward with now positive d˙αβ equation (60) tells us that elliptical
specification II yields smaller forces with the difference to circular specification leveled for Bα → ∞.
Third, with the – admittedly arguable – additional assumptions that
d˙αβ ∆tα
d˙αβ ∆tα
<<
dαβ
(61)
<<
Bα
(62)
one cannot only expand for small ξ(t), but also reduce the complexity of equation (58) with regard to d˙αβ .
Various approximate forms of that equation can be derived in this way of which one is analytically solvable and
contains parameter ∆tα (omitting the indices α):
˙ + v0 ξ(t) = 0
¨ + 1 1 + v0 ∆t ξ(t)
(63)
ξ(t)
τ
B
Bτ
˙ has changed. This leads to a less strict requirewhere in comparison to equation (37) just the factor before ξ(t)
ment for avoiding oscillations:
2
v0 ∆t
(64)
4v0 τ ≤ B 1 +
B
than equation (41).
˙ < 0 where the force is zero: if pedestrian α starts to approach β from a
Finally a note on the case d + d∆t
˙
sufficiently large distance and with a typical pedestrian walking speed it will be d + d∆t
> 0. Since equation
+
˙
(58) diverges to positive values at d + d∆t → 0 the pedestrian will slow down and eventually be at rest or turn
˙
˙ > 0. Only when a
around before d < −d∆t.
For a departing pedestrian α it is d˙ > 0 and thus always d + d∆t
˙
simulation is initiated with d + d∆t
< 0 it may yield dynamics that do not fit into this line of argumentation;
compare [17]. Such extrinsically prepared “Garden of Eden” states can be dealt with for example by a model
extension (if one does not simply want to exclude them by construction).
3.4
Simulations
For realistic applications it makes sense to choose the parameters such that no oscillations occur. However, to
verify a computer implementation of the Social Force Model it can be interesting to use parameters just around
8
the critically damped conditions and check if oscillations do and do not occur according to expectations. Model
specific validation work can amend validation tests like those in the RiMEA test cases [18] which usually are
formulated model independently. In this subsection such a model specific verification process will be carried out
exemplarily utilizing PTV Viswalk [19].
Figure 4: Simulation scenario.
Figure 4 shows how the one-dimensional setting has been implemented in the two-dimensional simulation
model. The walking area is modeled to be 0.55 m wide. A pedestrian facing into positive x direction is stopped
by a red traffic signal at x=0. Then a second pedestrian is set into the model at x=52 m moving into negative
x direction.
τ
0.7
0.8
0.9
1.0
1.2
1.5
2.0
3.0
4.0
5.0
distance
0.4556
0.4821
0.5044
0.5269
0.5645
0.6073
0.6655
0.7483
0.8044
0.8482
expected
0.4570
0.4837
0.5072
0.5283
0.5648
0.6094
0.6669
0.7480
0.8056
0.8502
difference
0.0014
0.0016
0.0028
0.0014
0.0002
0.0021
0.0014
-0.0003
0.0011
0.0020
Table 3: Stand still distance [m] in dependence of parameter τ [s], with A = 1.6 m/s2 , B = 0.2 m and v0 = 1.5
m/s. Column “expected” shows the values according to equation (35).
Parameter settings: since A soc mean controls the contribution of elliptical II specification it has been set
to zero (in the remainder ”A refers to ”A soc iso in the software; for the pedestrian on the left it has been set to
zero as well); λ = 1.0 to minimize effects from the in fact two-dimensional nature of the simulation; stochastic
noise has been set to zero . The value of the noise parameter has been set to zero for both pedestrians to not
unnecessarily blur the results.
B
0.1
0.2
0.3
0.5
1.0
2.0
distance
0.5831
0.6524
0.7240
0.8590
1.2073
1.9005
expected
0.5847
0.6540
0.7233
0.8620
1.2085
1.9017
difference
0.0016
0.0016
-0.0007
0.0029
0.0013
0.0012
B
4.0
6.0
9.0
12.0
18.0
24.0
dist.
3.2870
4.6734
6.7530
8.8326
12.9917
17.1509
exp.
3.2880
4.6743
6.7537
8.8332
12.9920
17.1509
diff.
0.0010
0.0008
0.0007
0.0005
0.0003
0.0000
Table 4: Stand still distance [m] in dependence of parameter B [m], with A = 2.0 m/s2 , τ = 1.5 s and v0 = 1.5 m/s.
Column “expected” shows the values according to equation (35). The difference between theoretical expectation
and simulation is in all cases below 3 mm.
At first we investigate where the second pedestrian comes to rest. According to equation (35) this depends
on the values of the parameters B soc iso, A soc iso, v0 , τ , and R. The latter was set to be 0.2577 m. Keeping
9
A soc iso and v0 constant and increasing the value of τ in steps of 0.1 s the second pedestrian does not pass
through the first one for the first time at a value τ = 0.6 s and both do not overlap visually for τ ≥ 0.8 s. At
τ = 0.9375 s when Aτ = v0 the distance of the central points of both pedestrians is 0.5135 m which comes 2
mm close to 2R. Table 3 and figure 5 show values for the stand still distance in dependence of some values for
parameter τ with all other parameters kept constant. The theoretical expectation is met well in all cases.
Table 4 and figure 5 show values for the stand still distance in dependence of some values for parameter
B soc iso with all other parameters kept constant. The theoretical expectation is met well in all cases.
The stand still distances of table 4 are unrealistically high for all but the smallest values for parameter B.
We have chosen the parameters in this way not to scan for realistic parameter values, but because with these
parameters one can well demonstrate that for certain cases (small values for B) there are oscillations and in
others (large values for B) there are none. Figures 6 and 7 show the time evolution of the position of the
approaching pedestrian for various values of parameter B.
Figure 5: Left: Visualization of the data of table 3. The expectation for the regression curve would be y =
0.2000 ln(x) + 0.5283. Right: Visualization of the data of table 4. The expectation for the regression curve would
be y = 0.6931x + 0.5154.
Figure 6: Left: Position of approaching pedestrian over time for various values of parameter B. Right: As A = 2.0
m/s2 , v0 = 1.5 m/s, τ = 1.5 s the system is critically damped for B = 9.0 m. For increasing B the oscillations get
smaller and vanish for B = 9.0.
Neglecting for the under damped case the damping exponential function the approximately expectated time
distance Tr between two reversal points in the under damped cases is:
Tr = q
π
v0
Bτ
−
(65)
1
4τ 2
Table 5 shows a comparison between actual and expected Tr for various values for B which generate under
damped behavior.
3.5
Two pedestrians approaching each other
If two pedestrians approach each other in one dimension they may come to a stand still at an equilibrium distance
or they might jointly (and with an equilibrium distance) move into one of the two possible directions. If all
10
Figure 7: Left: Position of approaching pedestrian vs. time for under damped and critical cases with regard to the
value of B. Right: Zoom to the region of oscillations.
parameters (v0 , A, B, and τ ) are identical for both pedestrians one can rewrite the coupled equations of motion
of both pedestrians as one equation for the movement of the center of mass and one equation for the relative
motion. Latter one can again be approximated by an oscillator equation. Compared to the case above one has
an additional factor 2 at the friction term implying that the system is critically or over damped for
8v0 τ ≤ B
(66)
which leads to the conclusion that the case of two mutually approaching pedestrians sets a stronger restriction
on parameter choice than the case when one pedestrian is approaching another who is standing, at least if
oscillations are to be avoided. The 2007 variant of the Social Force Model in this case suppresses oscillations
even more than when one of the two pedestrians is standing still.
B
0.1
0.2
0.3
0.5
1.0
1.5
2.0
3.0
number of reverses
11
7
3
3
2
2
2
2
simulation
1.035
1.442
1.800
2.250
3.100
4.000
4.850
6.550
approximate expectation
0.999
1.421
1.750
2.286
3.332
4.215
5.038
6.664
Table 5: Comparison of simulated and approximately expected time [s] between reversals.
4
Conclusions
It could be shown that the Social Force Model as proposed in 1995 (elliptical specification I) and 2000 (circular
specification) in a special and one-dimensional case and around equilibrium reduces to the equations of the
damped harmonic oscillator. This implies that indeed one has to be careful not to choose parameters with which
the pedestrians’ trajectories yield unrealistic results. However, at the same time it means that there are parts in
parameter space in which there are not only just small oscillations, but where there are exactly no oscillations.
A look at parameter values found in literature for the circular specification shows that parameters deemed to
yield realistic results in certain calibration scenarios do or may produce oscillating behavior unless the desired
walking speed(s) are set to rather small values.
The equations of the Social Force Model as of 2007 (elliptical specification II) are not as easily treated
analytically. Still in a discussion of its equations it could be argued that – compared to the circular specification
– oscillations are clearly suppressed. Elliptical specification II from 2007 therefore appears to be superior to the
two preceding variants also for the reasons discussed in this paper (this was found already in the paper from 2007
but for different reasons). It is therefore a good idea to either simply use elliptical specification II or combine it
with one of the earlier variants (e.g. by simply adding the forces) to reduce the risk of oscillations.
The phenomenon of oscillations was used to verify a specific computer implementation of the Social Force
Model. It was possible to show that it reproduces the expected results. This comparison can be seen as an
attempt to falsify either of the two – theoretical expectations and software implementation – and the attempt
11
failed, no potential issues could be found. The method can be applied to verify any implementation of the
Social Force Model variants of 1995 or 2000 and it is generally an example of model specific verification. The
tests carried out in this contribution are just examples. The phenomenon of oscillations bears the potential to
formulate further tests.
5
Acknowledgments
I thank Mohcine Chraibi and Normen Rochau for useful discussions.
6
References
References
[1] D. Helbing and P. Molnar, “Social force model for pedestrian dynamics”, Physical review E 51 no. 5,
(1995) 4282, arXiv:cond-mat/9805244.
[2] D. Helbing, I. Farkas, and T. Vicsek, “Simulating dynamical features of escape panic”, Nature 407
no. 6803, (2000) 487–490, arXiv:cond-mat/0009448.
[3] A. Johansson, D. Helbing, and P. Shukla, “Specification of the Social Force Pedestrian Model by
Evolutionary Adjustment to Video Tracking Data”, Advances in Complex Systems 10 no. 4, (2007)
271–288, arXiv:0810.4587 [physics.soc-ph].
[4] B. Steffen and A. Seyfried, “The repulsive force in continous space models of pedestrian movement”,
arXiv preprint (2008) eprint, arXiv:0803.1319 [physics.soc-ph].
[5] M. Chraibi, A. Seyfried, and A. Schadschneider, “Generalized centrifugal-force model for pedestrian
dynamics”, Physical Review E 82 no. 4, (2010) 046111, arXiv:1008.4297 [physics.soc-ph].
[6] M. Chraibi, U. Kemloh, A. Schadschneider, and A. Seyfried, “Force-Based Models of Pedestrian
Dynamics”, Networks and Heterogeneous Media 6 (2011) 425–442.
[7] G. Köster, F. Treml, and M. Gödel, “Avoiding numerical pitfalls in social force models”, Physical Review
E 87 no. 6, (2013) 063305.
[8] R. Corless, G. Gonnet, D. Hare, D. Jeffrey, and D. Knuth, “On the LambertW function”, Advances in
Computational Mathematics 5 no. 1, (1996) 329–359.
[9] R. Corless, D. Jeffrey, and D. Knuth, “A sequence of series for the Lambert W function”, in Proceedings of
the 1997 international symposium on Symbolic and algebraic computation, pp. 197–204, ACM. 1997.
[10] F. Chapeau-Blondeau and A. Monir, “Numerical evaluation of the Lambert W function and application to
generation of generalized Gaussian noise with exponent 1/2”, IEEE Transactions on Signal Processing 50
no. 9, (2002) 2160–2165.
[11] Wolfram—Alpha, 2014. Publisher: Wolfram Alpha LLC. Retrieved April 9th 2014. https:
//www.wolframalpha.com/input/?i=ProductLog%28a%29+with+a%3D-0.270670566473225&dataset=.
[12] D. Winter, “Human balance and posture control during standing and walking”, Gait & posture 3 no. 4,
(1995) 193–214.
[13] A. Gorrini, K. Shimura, S. Bandini, K. Ohtsuka, and K. Nishinari, “Experimental Investigation of
Pedestrian Personal Space”, Transportation Research Record: Journal of the Transportation Research
Board 2421 no. 1, (2014) 57–63.
[14] T. Werner and D. Helbing, “The social force pedestrian model applied to real life scenarios”, Pedestrian
and evacuation dynamics (2003) 17–26.
[15] S. Seer, C. Rudloff, T. Matyus, and N. Brändle, “Validating social force based models with comprehensive
real world motion data”, Transportation Research Procedia 2 (2014) 724–732.
[16] S. Hoogendoorn and W. Daamen, “Microscopic calibration and validation of pedestrian models:
Cross-comparison of models using experimental data”, in Traffic and Granular Flow 05, pp. 329–340.
Springer, 2007.
12
[17] A. Schadschneider and M. Schreckenberg, “Garden of Eden states in traffic models”, Journal of Physics
A: Mathematical and General 31 no. 11, (1998) L225, arXiv:cond-mat/9801061.
[18] U. Brunner, H. Kirchberger, C. Lebeda, M. Oswald, R. Könnecke, M. Kraft, A. Thoss, L. Mülli,
A. Seyfried, C. Hartnack, S. Wader, G. Spennes, and T. Kretz, RiMEA – Richtlinie für Mikroskopische
Entfluchtungs-Analysen. Initiatoren des RiMEA-Projekts: M. Schwendimann, N. Waldau, P. Gattermann,
C. Moroge, T. Meyer-König, and M. Schreckenberg, 2.2.1 ed., 2009. (eprint from http://www.rimea.de/).
(in German).
[19] PTV AG, PTV Vissim 6.0 – User Manual. PTV Group, Haid-und-Neu-Str. 15, D-76131 Karlsruhe,
Germany, 2013.
[20] Wolfram—Alpha, 2015. Publisher: Wolfram Alpha LLC. Retrieved January 18th 2015.
https://www.wolframalpha.com/input/?i=plot+2%2BLambertW%28-x*exp%28-x%29%29+for+x%3D1+to+2.
[21] Wolfram—Alpha, 2015. Publisher: Wolfram Alpha LLC. Retrieved January 18th 2015.
https://www.wolframalpha.com/input/?i=plot+%28LambertW%28-%281%2Bx%29*exp%28-%281%2Bx%29%
29%29%2C+-%28%2B+1+-+x+%2B+2%2F3*x*x+-+4%2F9*x*x*x+%2B+44%2F135*x*x*x*x+-+104%
2F405*x*x*x*x*x%29%29+for+x+from+0+to+1.
A Proof that a pedestrian heading for a destination point never
comes to rest
In this section we show that
tn =
n
X
∆t
(67)
i=1
diverges for n → ∞.
First step: We proof that αn+1 < αn : we begin with α0 = 2 and with equation (24)
αn+1 = 2 + W0 −αn e−αn
(68)
For all 1 < z ≤ 2 it holds that
− ze−z
<
0
(69)
∂
− ze−z
∂z
>
0
(70)
and for all −1/e < z < 0
W0 (z)
∂
W0 (z)
∂z
<
0
(71)
>
0
(72)
Therefore for all 1 < z < 2 it holds that
W0 (−ze−z )
<
0
(73)
∂
W0 (−ze−z ) > 0
∂z
(74)
From that and with α0 = 2 it follows that
α1 = 2 + W0 (−α0 e−α0 ) < α0
(75)
From equation (24) it follows that
αn − αn+1 = W0 (−αn−1 e−αn−1 ) − W0 (−αn e−αn )
(76)
−z
and from it that if αn−1 > αn also αn > αn+1 . See figure 8 for a plot of 2 + W0 −ze
and the visualization
of the evolution of the αn which allows to see this first step very easily.
Second step: We verify that ∆tn > 0 (physically this is obvious, mathematically it needs to be shown; this
step in effect is a check on potential errors done before): Since
αn = −W1 −αn e−αn
(77)
and because for −1/e < z < 0 it holds that
W0 (z) > W1 (z)
13
(78)
Figure 8: Plot of 2 + W0 (−xe−x ) for 1 ≤ α ≤ 2 and a visualization of the evolution of the αn ; basically created
with [20].
it results for the ∆tn according to equation (25)
∆tn+1
= αn + W0 −αn e−αn = W0 −αn e−αn − W1 −αn e−αn > 0
τ
(79)
Third step: we proof that ∆tn+1 < ∆tn . Equations (25) and (24) read
∆tn+1
τ
αn
αn + W0 (−αn e−αn )
=
=
2 + W0 (−αn−1 e
−αn−1
(80)
).
(81)
From them it follows that
∆tn+1
= αn+1 + αn − 2
(82)
τ
thus the ∆tn show the same behavior as the αn and with αn+1 < αn also ∆tn+1 < ∆tn .
Fourth step: if ∆tn+1 /∆tn for large n would approach a value smaller than 1 the sum would converge, only
if it approaches 1 it may diverge. From (82) it follows that
∆tn+1
αn−1 − αn+1
=1−
∆tn
αn−1 + αn − 2
To compute the limit for this for large n we have to expand W0 (−αn e−αn ) into a series.
For convenience we write
βn = αn − 1
(83)
(84)
knowing that this is βn = un /v0 and that these βn will approach 0 with increasing n.
By computing the right side limit towards 0 of
W0 (−αn e−αn ) = W0 (−(1 + βn )e−(1+βn ) )
(85)
and its derivatives one can write
W0 (−(1 + βn )e−(1+βn ) ) = −1 + βn −
2 2 4 3
44 4 104 5
βn + βn −
βn +
βn + O(βn6 )
3
9
135
405
(86)
compare figure 9.
Considering only terms to second order
βn
≈
βn+1
≈
2 2
βn−1
3
4 2
βn−1 − βn−1
3
βn−1 −
(87)
(88)
Using these approximations in equation (83) gives for large n
∆tn+1
2βn−1
≈1−
∆tn
3 − βn−1
which approaches 1 with n → ∞ where βn−1 → 0.
14
(89)
Thus, like with the harmonic series hm the ratio of subsequent terms approaches 1. We note that equation
∆t
(82) implies n+1
= βn+1 + βn and therefore
τ
∞
X
tn→∞
= α0 − 1 + 2
βn
τ
n=1
(90)
the sum of the ∆tn depends trivially on the sum of the βn and especially their convergence behavior is the same.
In the last step we prove that if βn > hm then also βn+1 > hm+1 and therefore beyond some n, m the harmonic
series is a lower estimate of the βn implying that with the harmonic series also the sum of βn diverges.
Assume that for some n, βn can be written as
βn =
q
m
(91)
with m ∈ N, m >> 1 and 1 ≤ q < m/(m − 1), i.e. βn is placed between two elements of the harmonic series.
This requires just that there is some 0 < βn < 1. This is obviously the case.
So, can there be a βn+1 < 1/(m + 1)?
βn+1 −
1
q
2 q2
1
=
−
−
m+1
m
3 m2
m+1
(92)
A lower estimate of the right side is if at the positive term q is replaced by its minimum value q = 1 and at the
negative term by its upper limit q = m/m − 1:
βn+1 −
1
m+1
>
>
>
1
2
1
1
−
−
m
3 (m − 1)2
m+1
m2 − 8m + 3
3m(m − 1)2 (m + 1)
0∀(m ≥ 8)
(93)
(94)
(95)
Thus, if there is a βn0 < 1/8 – it is, as can be seen in table 1) – it will hold for all βn>n0 that if βn > hm then
also βn+1 > hm+1 . Therefore the harmonic series is a lower estimate for the series of βn . Thus the series of the
βn diverges and with it the series of the ∆tn .
Figure 9: Plot comparing W (−(1 + y) exp(−(1 + y))) and its approximation [21].
15
| 5cs.CE
|
1
Robust Computer Algebra, Theorem Proving, and Oracle AI
arXiv:1708.02553v2 [cs.AI] 31 Dec 2017
Gopal P. Sarma
School of Medicine, Emory University, Atlanta, GA USA
[email protected]
Nick J. Hay
Vicarious FPC, San Francisco, CA USA
[email protected]
Keywords: Oracle AI, AI safety, CAS, theorem proving, math oracles
Received: June 24, 2013
In the context of superintelligent AI systems, the term “oracle” has two meanings. One
refers to modular systems queried for domain-specific tasks. Another usage, referring
to a class of systems which may be useful for addressing the value alignment and AI
control problems, is a superintelligent AI system that only answers questions. The
aim of this manuscript is to survey contemporary research problems related to oracles
which align with long-term research goals of AI safety. We examine existing question
answering systems and argue that their high degree of architectural heterogeneity makes
them poor candidates for rigorous analysis as oracles. On the other hand, we identify
computer algebra systems (CASs) as being primitive examples of domain-specific
oracles for mathematics and argue that efforts to integrate computer algebra systems
with theorem provers, systems which have largely been developed independent of one
another, provide a concrete set of problems related to the notion of provable safety
that has emerged in the AI safety community. We review approaches to interfacing
CASs with theorem provers, describe well-defined architectural deficiencies that have
been identified with CASs, and suggest possible lines of research and practical software
projects for scientists interested in AI safety.
1
Introduction
Recently, significant public attention has been
drawn to the consequences of achieving humanlevel artificial intelligence. While there have
been small communities analyzing the long-term
impact of AI and related technologies for decades,
these forecasts were made before the many recent
breakthroughs that have dramatically accelerated the pace of research in areas as diverse
as robotics, computer vision, and autonomous
vehicles, to name just a few [1–3].
Most researchers and industrialists view
advances in artificial intelligence as having
the potential to be overwhelmingly beneficial
to humanity.
Medicine, transportation, and
fundamental scientific research are just some of
the areas that are actively being transformed
by advances in artificial intelligence. On the
other hand, issues of privacy and surveillance,
access and inequality, or economics and policy
are also of utmost importance and are distinct
from the specific technical challenges posed by
most cutting-edge research problems [4, 5].
In the context of AI forecasting, one set of
issues stands apart, namely, the consequences
of artificial intelligence whose capacities vastly
exceed that of human beings. Some researchers
have argued that such a “superintelligence”
poses distinct problems from the more modest
AI systems described above. In particular, the
emerging discipline of AI safety has focused on
2
issues related to the potential consequences of
mis-specifying goal structures for AI systems
which have significant capacity to exert influence
on the world. From this vantage point, the
fundamental concern is that deviations from
“human-compatible values” in a superintelligent
agent could have significantly detrimental consequences [1].
One strategy that has been advocated for
addressing safety concerns related to superintelligence is Oracle AI, that is, an AI system
that only answers questions. In other words,
an Oracle AI does not directly influence the
world in any capacity except via the user of the
system. Because an Oracle AI cannot directly
take physical action except by answering questions posed by the system’s operator, some have
argued that it may provide a way to bypass the
immediate need for solving the “value alignment
problem” and would itself be a powerful resource
in enabling the safe design of autonomous,
deliberative superintelligent agents [1, 6–9].
A weaker notion of the term oracle, what we
call a domain-specific oracle, refers to a modular component of a larger AI system that is
queried for domain-specific tasks. In this article,
we view computer algebra systems as primitive
domain-specific oracles for mathematical computation which are likely to become quite powerful
on the time horizons on which many expect superintelligent AI systems to be developed [10,11].
Under the assumption that math oracles prove to
be useful in the long-term development of AI systems, addressing well-defined architectural problems with CASs and their integration with interactive theorem provers provides a concrete set of
research problems that align with long-term issues in AI safety. In addition, such systems may
also be useful in proving the functional correctness of other aspects of an AI architecture. In
Section 2, we briefly discuss the unique challenges
in allocating resources for AI safety research. In
Section 3, we briefly summarize the motivation
for developing oracles in the context of AI safety
and give an overview of safety risks and control
strategies which have been identified for superintelligent oracle AIs. In Section 4 we analyze contemporary question answering systems and argue
Sarma and Hay
that in contrast to computer algebra systems, current consumer-oriented, NLP-based systems are
poor candidates for rigorous analysis as oracles.
In Section 5, we review the differences between
theorem provers and computer algebra systems,
efforts at integrating the two, and known architectural problems with CASs. We close with a list
of additional research projects related to mathematical computation which may be of interest to
scientists conducting research in AI safety.
2
Metascience of AI Safety
Research
From a resource allocation standpoint, AI safety
poses a unique set of challenges. Few areas of
academic research operate on such long and
potentially uncertain time horizons. This is not
to say that academia does not engage in longterm research. Research in quantum gravity, for
example, is approaching nearly a century’s worth
of effort in theoretical physics [12]. However, the
key difference between open-ended, fundamental
research in the sciences or humanities and AI
safety is the possibility of negative consequences,
indeed significant ones, of key technological
breakthroughs taking place without corresponding advances in frameworks for safety [1, 13] .
These issues have been controversial, largely
due to disagreement over the time-horizons for
achieving human-level AI and the subsequent
consequences [10, 11]. Specifically, the notion of
an “intelligence explosion,” whereby the intelligence of software systems dramatically increases
due their capacity to model and re-write their
own source code, has yet to receive adequate
scientific scrutiny and analysis [14].
We affirm the importance of AI safety research
and also agree with those who have cautioned
against proceeding down speculative lines of
thinking that lack precision. Our perspective
in this article is that it is possible to fruitfully
discuss long-term issues related to AI safety while
maintaining a connection to practical research
problems. To some extent, our goal is similar
in spirit to the widely discussed manuscript
“Concrete Problems in AI Safety” [15]. However,
we aim to be a bit more bold. While the authors
3
Robust Computer Algebra, Theorem Proving, and Oracle AI
of “Concrete Problems” state at the outset that
their analysis will set aside questions related to
superintelligence, our goal is to explicitly tackle
superintelligence related safety concerns. We
believe that there are areas of contemporary research that overlap with novel ideas and concepts
that have arisen among researchers who have
purely focused on analyzing the consequences of
AI systems whose capacities vastly exceed those
of human beings.
To be clear, we do not claim that the strategy
of searching for pre-existing research objectives
that align with the aims of superintelligence
theory is sufficient to cover the full spectrum of
issues identified by AI safety researchers. There
is no doubt that the prospect of superintelligence
raises entirely new issues that have no context
in contemporary research. However, considering
how young the field is, we believe that the perspective adopted in this article is a down-to-earth
and moderate stance to take while the field is
in a critical growth phase and a new culture is
being created.
This article focuses on one area of the AI safety
landscape, Oracle AI. We identify a set of concrete
software projects that relate to more abstract,
conceptual ideas from AI safety, to bridge the gap
between practical contemporary challenges and
longer term concerns which are of an uncertain
time horizon. In addition to providing concrete
problems for researchers and engineers to tackle,
we hope this discussion will be a useful introduction to the concept of Oracle AI for newcomers to
the subject. We state at the outset that within
the context of Oracle AI, our analysis is limited
in scope to systems which perform mathematical computation, and not to oracles in general.
Nonetheless, considering how little effort has been
directed at the superintelligence control problem,
we are confident that there is low-hanging fruit
in addressing these more general issues which are
awaiting discovery.
3
Brief Overview of Oracle AI
As described above, an Oracle AI is a system
which only answers questions. Although the term
has traditionally been used to denote fully general
AI systems, we can also imagine oracles whose
capabilities are more limited. The latter systems
are what we refer to as domain-specific oracles.
AI safety researchers have argued that superintelligent oracles have the potential to help address
safety concerns of autonomous, deliberative superintelligent AI systems, as an Oracle AI’s only
ability to influence the world is through the user
of the system. Consequently, an Oracle AI itself
may be in a position to assist researchers in solving difficult technical and conceptual problems related to the design of independent, deliberative
agents. Yet superintelligent oracle AIs pose their
own distinct set of safety concerns. The problems
identified in [6], for example, presume a superintelligent AI system, which while limited in its
ability to interact with the world, may nonetheless have the potential to negatively impact the
world by manipulating the user of the system. As
Armstrong, Bostrom, and Sandberg write:
We are all vulnerable to the OAI’s social engineering. An advanced OAI can literally
offer us the world. Genuine altruism is no
defence against this; it can offer a world free
of poverty, hunger, war, disease or death, as
well as all personal trinkets we could want.
It could guess the weaknesses of each individual, and find the right arguments to convince us that granting it power or liberty
is the moral and profitable thing to do. It
will be able to play on rivalries, institutional
weaknesses and ethical dilemmas with more
virtuosity than any human could. [6]
Given the potential dangers of human manipulation by an OAI, researchers have searched for
control strategies to mitigate such risks. A few
examples of control methods that have been proposed are:
• Physical security- physically containing
the OAI system
• Limiting or eliminating external manipulators- like the previous security measure, precluding the OAI from physically manipulating the world in any direct manner
• Epistemic capability control- the OAIs
output could be restricted even further beyond question answering, for example, restricting it to answer “yes / no ” questions
4
Sarma and Hay
• Virtual worlds- the OAI could be run in
a simulated world so that even if the system’s capabilities were expanded, its behavior could be observed and modeled.
• Resetting the OAI after each questionthis would prevent the OAI from engaging in
long-term social engineering by manipulating
the answers to the users’ questions
Although the capacities of domain-specific oracles are more limited, they can nonetheless pose
safety risks. Architectural deficiencies in such
oracles might be exploited by a larger AI system
to manipulate the human user. It could give
answers which are difficult to verify and which
allow the OAI to execute complex and intricate
plans unbeknownst to the user. Therefore, while
flaws in domain-specific oracles are not inherently
risky if used solely in their domain of applicability, they may very well be dangerous as part of
a larger system with more general capabilities.
Though not a “control strategy” in the narrowest
sense, creating “robust” domain-specific oracles
is an important objective in designing safe
OAIs. Furthermore, ensuring the robustness of
domain-specific subsystems might mitigate the
need for stronger control strategies, as the OAI
would have fewer weaknesses to exploit.
It should go without saying that the arguments
presented above are highly schematic and do not
dependent on specific technologies. To our knowledge, there is very limited work on translating
analyses of superintelligent oracle AIs into the
concrete language of modern artificial intelligence
[8, 9, 16]. Our goal in this manuscript is in this
spirit, that is, to anchor schematic, philosophical
arguments in practical, contemporary research.
To do so, we will narrow our focus to the mathematical domain. In the remainder of the article,
we will use the term oracle in the more limited
sense of a domain-specific subsystem, and in particular, oracles for performing mathematical computations. We hope that the analysis presented
here will be of intrinsic value in developing robust
math oracles, as well as provide some intuition
and context for identifying concrete problems relevant to developing safe, superintelligent oracle
AI systems.
4
Are there contemporary
systems which qualify as
oracles?
The obvious class of contemporary systems which
would seem to qualify as oracles are question
answering systems (QASs). As we stated above, a
basic criterion characterizing oracles is that their
fundamental mode of interaction is answering
questions posed by a user, or for domain-specific
queries as part of a larger AI system.
Contemporary QASs are largely aimed at using
natural language processing techniques to answer
questions pertaining to useful facts about the
world such as places, movies, historical figures,
and so on. An important point to make about
QASs is the highly variable nature of the underlying technology. For instance, IBM’s original
Watson system which competed in Jeopardy, was
developed prior to the recent advances in deep
learning which have fundamentally transformed
areas ranging from computer vision, to speech
recognition, to natural language processing [17].
In this particular task, the system was nonetheless able to perform at a level beyond that of
the most accomplished human participants. The
introduction of “info panes” into popular search
engines, on the other hand, have been based
on more recent machine learning technology,
and indeed, these advances are also what power
the latest iterations of the Watson system [18].
On the other end of the spectrum is Wolfram |
Alpha, which is also a question answering system,
but which is architecturally centered around a
large, curated repository of structured data,
rather than datasets of unstructured natural
language [19].
While these systems are currently useful for
humans in navigating the world, planning social
outings, and arriving at quick and useful answers
to ordinary questions, it is not clear that they will
remain useful in quite the same capacity many
years from now, or as standalone components
of superintelligent AI systems. Although the
underlying techniques of deep learning or NLP
are of fundamental interest in their own right,
the fact that these systems are QASs at all seems
to be more of an artifact of their utility for
Robust Computer Algebra, Theorem Proving, and Oracle AI
consumers.
Another important observation about contemporary QASs is that much of their underlying
NLP-based architecture can be replaced by taking advantage of structured data, as the example of Wolfram — Alpha demonstrates. For the
other NLP or machine learning based systems,
the underlying technology can be used as part of
larger, semi-automated pipelines to turn unstructured data from textual sources into structured
data. Once again, this fact simply underscores
that contemporary QASs are not particularly appealing model systems to analyze from the Oracle
AI safety perspective.1
4.1
Computer Algebra and
Domain-Specific Oracles for
Mathematical Computation
The question answering systems described above
all rely on natural language processing to varying
degrees. In addition, their domain of applicability has tended towards “ordinary” day-to-day
knowledge useful to a wide array of consumers.
Another type of question answering system is
a computer algebra system (CAS). Computer
algebra has traditionally referred to systems for
computing specific results to specific mathematical equations, for example, computing derivatives
and integrals, group theoretic quantities, etc. In
a sense, we can think of computer algebra as a
1
We emphasize that our argument that contemporary
QASs are not good candidates for analysis as Oracle AIs
is not an argument against the traditional formulation of
Oracle AI as a tool for AI safety. We fully expect significant breakthroughs to be made in advancing the theory
and practice of oracle-based techniques for AI safety and
we hope that this manuscript will provide some motivation
to pursue such research. Rather, our point is that when
viewing contemporary systems from the lens of superintelligence, there seems little reason to believe that current
NLP-based QASs will remain sufficiently architecturally
stable to be used as standalone components in AI systems
many years from now. On the other hand, there are certainly important present-day problems to examine when
evaluating the broader impact of QASs, such as bias in
NLP systems, overgeneralization, and privacy, to name just
a few. Some of these issues overlap with the set of problems identified in [15] as examples of concrete problems in
AI safety. In addition, we are beginning to see conferences
devoted to contemporary ethical issues raised by machine
learning. See, for example, the workshop Ethics in Natural
Language Processing.
5
set of algorithms for performing what an applied
mathematician or theoretical physicist might
work out on paper and pencil. Indeed, some of
the early work in computer algebra came from
quantum field theory—one of the first computer
algebra systems was Veltman’s Schoonschip for
performing field theoretic computations that led
to the theory of electroweak unification [20].
As computer algebra systems have grown in
popularity, their functionality has expanded
substantially to cover a wide range of standard
computations in mathematics and theoretical
physics, including differentiation, integration,
matrix operations, manipulation of symbolic
expressions, symbolic substitution, algebraic
equation solving, limit computation, and many
others. Computer algebra systems typically run
in a read, evaluate, print loop (repl), and
in the research and education context, their popularity has also grown as a result of the notebook
model pioneered by the Mathematica system,
allowing for computations in CASs to closely
mimic the sequential, paper and pencil work of
mathematicians and theoretical physicists.
In assessing the long-term utility of CASs, it
is important to note that there is little reason to
believe that computer algebra will be subsumed
by other branches of AI research such as machine
learning. Indeed, recent research has demonstrated applications of machine learning to both
computer algebra and theorem proving (which we
discuss in more detail below), via algorithm selection in the former case [21] and proof assistance
in the latter [22, 23]. While certainly not as visible as machine learning, computer algebra and
theorem proving are very much active and deep
areas of research which are also likely to profit
from advances in other fields of artificial intelligence, as opposed to being replaced by them [24].
On the time horizons on which we are likely to
see human-level artificial intelligence and beyond,
we can expect that these systems will become
quite powerful, and possess capabilities that may
be useful in the construction of more general AI
systems. Therefore, it is worth examining such
systems from the perspective of AI safety.
6
4.2
Sarma and Hay
Briefly Clarifying Nomenclature
Before proceeding, we want to explicitly describe
issues relating to nomenclature that have arisen
in the discussion thus far, and state our choices
for terminology. Given that the phrase “Oracle
AI” has become common usage in the AI safety
community, we will continue to use this phrase,
with the first word capitalized, as well as the
acronym OAI. Where clarification is needed, we
may also use the full phrase “superintelligent
oracle AI,” without capitalization.
For more modest use cases of the word oracle,
we will either refer to “domain-specific oracles,”
or state the domain of knowledge where the oracle is applicable. We can, at the very least in
the abstract, consider extending this terminology
to other domains such as “physics oracles,” “cell
biology oracles,” or “ethics oracles” and so on.
Therefore, the remainder of the article will be
concerned with safety and robustness issues in the
design of “math oracles.”
5
Robust Computer Algebra
and Integrated Theorem
Proving
Today we should consider as a standard feature much closer interaction between proof
assistance and computer algebra software.
Several areas can benefit from this, including specification of interfaces among components, certification of results and domains
of applicability, justification of optimizations
and, in the other direction, use of efficient
algebra in proofs.
- Stephen Watt in On the future
of computer algebra systems at the
threshold of 2010
As we described above, computer algebra
systems can be thought of as question answering
systems for a subset of mathematics. A related
set of systems are interactive proof assistants
or interactive theorem provers (ITPs). While
ITPs are also systems for computer-assisted
mathematics, it is for a different mathematical
context, for computations in which one wishes to
construct a proof of a general kind of statement.
In other words, rather than computing specific
answers to specific questions, ITPs are used to
show that candidate mathematical structures (or
software systems) possess certain properties.
In a sense, the distinction between theorem
proving and computer algebra should be viewed
as a historical anomaly. From the perspective
of philosophical and logical efforts in the early
20th century that led to the “mechanization of
mathematics” the distinction between computing
the nth Laguerre polynomial and constructing
a proof by induction might have been viewed
as rather artificial, although with the benefit of
hindsight we can see that the two types of tasks
are quite different in practice [25].
The role of ITPs in the research world is very
different from that of CASs. Whereas CASs allow
researchers to perform difficult computations
that would be impossible with paper and pencil,
constructing proofs using ITPs is often more
difficult than even the most rigorous methods of
pure mathematics. In broad terms, the overhead
of using ITPs to formalize theorems arises from
the fact that proofs in these systems must
proceed strictly from a set of formalized axioms
so that the system can verify each computation.
Consequently, ITPs (and related systems, such as
automatic theorem provers) are largely used for
verifying properties of mission-critical software
systems which require a high-degree of assurance,
or for hardware verification, where mistakes can
lead to costly recalls [26–30].
As the quotation above suggests, many
academic researchers view the integration of
interactive proof assistants and computer algebra
systems as desirable, and there have been numerous efforts over the years at exploring possible
avenues for achieving this objective [31–34] (a
more complete list is given below). By integrating theorem proving with computer algebra,
we would be opening up a wealth of potentially
interoperable algorithms that have to date
remained largely unintegrated. To cite one such
example, in [35], the authors have developed a
framework for exchange of information between
the Maple computer algebra system and the
Isabelle interactive theorem prover. They show
Robust Computer Algebra, Theorem Proving, and Oracle AI
a simple problem involving the proof of an
elementary polynomial identity that could be
solved with the combined system, but in neither
system alone (see Fig. 1).
We cite this example to demonstrate how
a simply stated elementary problem cannot
be solved in existing environments for either
computer algebra or proof assistance. The computer algebra system does not have the capacity
for structural induction and theorem provers
generally have rather weak expression simplifiers.
There are numerous examples such as this one in
the academic literature.
Another key difference between CASs and ITPs
is the architectural soundness of the respective
systems. As we will discuss below, computer algebra systems have well-defined architectural deficiencies, which while not a practical issue for
the vast majority of use cases, pose problems for
their integration with theorem provers, which by
their nature, are designed to be architecturally
sound. In the context of superintelligent AI systems, the architectural problems of CASs are potential points of weakness that could be exploited
for malicious purposes or simply lead to unintended and detrimental consequences. Therefore,
we use the phrase “robust computer algebra” to
refer to CASs which lack the problems that have
been identified in the research literature. In the
section below, we combine the discussion of robust computer algebra and integration with interactive theorem provers, as there is a spectrum
of approaches which address both of these issues
to varying degrees.
5.1
A Taxonomy of Approaches
There are many possible avenues to tackle the integration of theorem provers with computer algebra systems. We give 4 broad categories characterizing such integration efforts2 :
1. Theorem provers built on top of computer algebra systems: These include Analytica, Theorema, RedLog, and logical extensions to the Axiom system [34, 36–39] .
2
This classification was first described by Kaliszyk and
Wiedijk [32] in a paper arguing for an architecture which
we list as the fourth category given above.
7
2. Frameworks for mathematical exchange between the two systems:
This category includes MathML, OpenMath, OMSCS, MathScheme, and Logic
Broker [40–44].
3. “Bridges” or “ad-hoc” information exchange solutions: The pairs of systems
in this category include bridges combining
PVS, HOL, or Isabelle with Maple, NuPRL
with Weyl, Omega with Maple/GAP, Isabelle with Summit, and most recently, Lean
with Mathematica [35, 45–51]. The example
given above, bridging Isabelle and Maple, is
an example of an approach from this category.
4. Embedding a computer algebra system
inside a proof assistant: This is the approach taken by Kaliszyk and Wiedijk in the
HOLCAS system. In their system, all expressions have precise semantics, and the proof
assistant proves the correctness of each simplification made by the computer algebra system [32].
One primary aspect of integration that differentiates these approaches is the degree of trust
the theorem prover places in the computer algebra system. Computer algebra systems give
the false impression of being monolithic systems
with globally well-defined semantics. In reality,
they are large collections of algorithms which are
neatly packaged into a unified interface. Consequently, there are often corner cases where the
lack of precise semantics can lead to erroneous
solutions. Consider the following example:
The system incorrectly gives 1 as a solution,
even though the given polynomial has an indeterminate value for x = 1. However, because the
expression is treated as a fraction of polynomials,
it is first simplified before the solve operation
is applied. In other words, there is an unclear
semantics between the solver module and the
simplifier which leads to an incorrect result.
Another simple example is the following integral:
Making the substitution n = −1 gives an indeterminate result, while it is clear by inspection
that the solution to the integral for n = −1 is
simply ln(x). This belongs to a class of problems
2
3
4
INDUCT
8
Sarma and Hay
1
1 : T H ` I n5 5 n
2 : T H `I 5 5 5 5
3 : T H `I n 5
4 : T H `I 8x : [x 2 N ^ 5 x ^ x5 5x ] =) (x + 1)5 5(x+1)
5 : x 2 N `M (x + 1)5 ⌘ x5 + 5x4 + 10x3 + 10x2 + 5x + 1
6 : T H `I 8x : [x 2 N ^ 5 x ^ x5 5x ] =)
x5 + 5x4 + 10x3 + 10x2 + 5x + 1 5(x+1)
7 : x 2 N `M 5(x+1) ⌘ 5 ⇤ 5x
8 : T H `I 8x : [x 2 N ^ 5 x ^ x5 5x ] =)
x5 + 5x4 + 10x3 + 10x2 + 5x + 1 5 ⇤ 5x
Circles represent object nodes, whose labels are reported in the table; rectangles
Figure 1: Example of a polynomial identity proven by integrating the Maple computer algebra system
represent link nodes, and contain their labels. The complex series of steps corwith Isabelle. Maple’s simplifier is used for expanding polynomials—a powerful complement to the
responding
the final ofphase
ofwhich
the proof
arethe
folded
triangular REST
theorem
provingtoarchitecture
Isabelle
allows for
setup within
of a proofthe
by induction.
node. Link nodes labelled with SIMPLIFY identify the points where the systems
cooperate to the solution of the problem; namely, where Maple is invoked to exAI safety,
decision-theoretic
research
pand some polynomial power. Note that the
RESTconsider
folded the
node
hides away several
agenda
for
the
development
of
safe,
superintelliadditional Maple calls, meant to perform evaluations of disequalitites.
Figure 2: Example of an incorrect solution to a
simple polynomial equation by a computer algebra system.
gent AI systems outlined in [52–56]. If we require
formal guarantees of correctness at any point in a
sequence of computations in which computer algebra is used, current systems would be unable to
provide the necessary framework for constructing
such a proof.
5.1.1
Figure 3: A problem arising in symbolic integration due to the non-commutativity of evaluation
and substitution.
known as the specialization problem, namely that
expression evaluation and variable substitution do
not commute [31]. So while we have seen above
that theorem proving can benefit tremendously
from the wealth of algorithms for expression simplification and mathematical knowledge in computer algebra, there is the potential cost of compromising the reliability of the combined system.
As a possible application to current research in
Qualitatively Certified
Computations
In our taxonomy of approaches to bridging
theorem provers with computer algebra, we
described how a key distinction was the degree of
trust that the theorem prover places in the computer algebra system. For instance, approaches
which build theorem provers on top of computer
algebra systems do not address the architectural
issues with CASs. They are integrative, but not
more sound. On the other extreme, building a
computer algebra system on top of a theorem
prover allows for a degree of trust that is on par
with that of the theorem prover itself. However,
this approach has the distinct disadvantage
that computer algebra systems represent many
hundred man-years worth of effort.
The more intermediate approaches involving
common languages for symbolic exchange or adhoc bridges, bring to light an important notion
in the spectrum of provable safety, namely the
9
Robust Computer Algebra, Theorem Proving, and Oracle AI
ability to assign probabilities for the correctness
of computations. In [57], the authors present an
algorithm for assigning probabilities to any statement in a formal language. We might ask what
strategies might look like that have a similar goal
in mind, but are significantly weaker. Interfaces
between theorem provers and computer algebra
systems provide a concrete example where we
can ask a question along these lines. Fundamentally, in such an interface, the computer algebra
system is the weaker link and should decrease
our confidence in the final result. But by how
much? For instance, in the example given in
Figure 1, how should we revise our confidence in
the result knowing that polynomial simplification
was conducted within a computer algebra system?
It is worth asking for simple answers to this
question that do not require major theoretical
advances to be made. For instance, we might
imagine curating information from computer
algebra experts about known weaknesses, and
use this information to simply give a qualitative
degree of confidence in a given result. Or, for
example, in a repository of formal proofs generated using integrated systems, steps of the proof
that require computer algebra can be flagged and
also assigned a qualitative measure of uncertainty.
The relationship that this highly informal
method of giving qualitative certification to
computations has with the formal algorithm
developed in [57] can be compared to existing
techniques in the software industry for ensuring
correctness. On the one hand, unit testing is a
theoretically trivial, yet quite powerful practice,
something along the lines of automated checklists for software. The complexities of modern
software would be impossible to handle without
extensive software testing frameworks [58–62].
On the other hand, formal verification can
provide substantially stronger guarantees, yet
is a major undertaking, and the correctness
proofs are often significantly more demanding to
construct than the software itself. Consequently,
as discussed in Section 5, formal verification is
much less frequently used in industry, typically
only in exceptional circumstances where high
guarantees of correctness are required, or for
hardware verification [26–30].
Integrated systems for computer algebra and
theorem proving give rise to a quite interesting
(and perhaps ironic) opportunity to pursue simple
strategies for giving qualitative estimates for the
correctness of a computation.
5.1.2
Logical Failures and Error
Propagation
As the examples described above demonstrate,
errors in initial calculations may very well propagate and give rise to non-sensical results. As
AI systems capable of performing mathematical
computation become increasingly sophisticated
and embedded as part of design workflows for
science and engineering (beyond what we see
today), we could imagine such errors being quite
costly and difficult to debug. In the case of
a superintelligent AI system, more concerning
scenarios would be if systematic errors in computer algebra could be exploited for adversarial
purposes or if they led to unintentional accidents
on a large scale.
The issue of error propagation is another example of a concrete context for pursuing simple strategies for assigning qualitative measures
of certainty to computations performed by integrated theorem proving / computer algebra systems. For instance, we may be less inclined to
trust a result in which the computer algebra system was invoked early on in a computation as
opposed to later. With curated data from computer algebra experts on the reliability or failure modes of various algorithms, we might also
chain together these informal estimates to arrive
at a single global qualitative estimate. If multiple systems were to be developed independently,
or which were based on fundamentally different
architectures, we might also be significantly more
confident in a result which could be verified by
two separate systems.
5.1.3
Additional Topics
Some related ideas merit investigation in the
broader context of mathematical computation:
• Integrating SMT solvers with interactive theorem provers: Satisfiability
10
Sarma and Hay
modulo theories (SMT) solvers are an important element of automated reasoning and
there have been efforts analogous to those
described above to bridge SMT solvers with
interactive theorem provers [63, 64].
• Identifying the most important /
widely used algorithms in computer
algebra: Computer algebra systems have
grown to become massive collections of
algorithms extending into domains well
outside of the realm of mathematics. If
the purely mathematical capacities of CASs
prove to be useful in future AI systems, it
would be valuable to rank order algorithms
by their popularity or importance.
One approach would be to do basic textual
analysis of the source code from GitHub or
StackExchange. This would also allow for
more targeted efforts to directly address the
issues with soundness in core algorithms
such as expression simplification or integration. In the context of the HOLCAS system
described above, for example, it would be
valuable to have rough estimates for the
number of man-hours required to implement
a minimal CAS with the most widely used
functionality on top of a theorem prover.
• Proof checkers for integrated systems:
Proof checkers are important tools in the
landscape of formal verification and theorem
proving. Indeed, as it is often much less
computationally expensive to verify the correctness of a proof than to generate it from
scratch, the availability of proof checkers for
the widely used interactive theorem provers
is one reason we can be confident in the
correctness of formal proofs [65, 66].
As we described above, strategies for integrating computer algebra with theorem
provers can potentially result in a combined
system which is less trustworthy than
the theorem prover alone. Therefore, the
availability of proof checkers for combined
systems would be a valuable resource in
verifying proof correctness, and in certain
mathematical domains, potentially provide
an avenue for surmounting the need to
directly make the CAS itself more architecturally robust.
The development of integrated proof checkers
is likely to be a substantial undertaking and
require novel architectures for integrating
the core CAS and ITP systems distinct from
what has been described above. However,
it is a largely unexplored topic that merits
further investigation.
• Analyzing scaling properties of algorithms for computer algebra and
theorem proving as a function of
hardware resources: The premise of the
analysis presented above is that CASs (and
integrated theorem proving) are likely to
remain sufficiently architecturally stable
and useful on a several decade time-horizon
in the construction of AI systems. On
the other hand, as we argued earlier, it is
much less clear that the same will be true
of the most visible, NLP-based, consumeroriented question answering systems. To
make these arguments more rigorous, it
would be valuable to develop quantitative
predictions of what the capabilities will be
of existing algorithms for computer algebra
and theorem proving when provided with
substantially expanded hardware resources.
For instance, we might examine problems in
mathematics or theoretical physics for which
naı̈ve solutions in CASs are intractable with
current resources, but which may be feasible
with future hardware.
• The cognitive science of computer
algebra: What role has computer algebra
played in theoretical physics and mathematics? How has it influenced the thinking
process of researchers?
Has computer
algebra simply been a convenience that has
shifted the way problems are solved, or has
it fundamentally enabled new problems to
be solved that would have been completely
intractable otherwise?
Robust Computer Algebra, Theorem Proving, and Oracle AI
The cognitive science of mathematical
thought is a substantial topic which overlaps
with many established areas of research [67–
71]. However, a systematic review of research
in mathematics and theoretical physics since
the advent of computer algebra and its role
in the mathematical thought process is an
underexplored topic. It would be an interesting avenue to pursue in understanding the
role that CASs, ITPs, and integrated systems
may come to play in superintelligence, particularly in the case of neuromorphic systems
that have been modeled after human cognition. These questions also relate to understanding the scaling properties of CAS and
theorem proving algorithms as well as cataloguing the most widely used algorithms in
computer algebra.
6
Conclusion
The aim of this article has been to examine preexisting research objectives in computer science
and related disciplines which align with problems
relevant to AI safety, thereby providing concrete,
practical context for problems which are otherwise of a longer time horizon than most research.
In particular, we focused on the notion of “Oracle
AI” as used in the AI safety community, and
observed that the word oracle has two meanings
in the context of superintelligent AI systems.
One usage refers to a subsystem of a larger AI
system queried for domain-specific tasks, and the
other to superintelligent AI systems restricted to
only answer questions.
We examined contemporary question answering systems (QASs) and argued that due to their
architectural heterogeneity, consumer-oriented,
NLP-based systems do not readily lend themselves to rigorous analysis from an AI safety
perspective. On the other hand, we identified
computer algebra systems (CASs) as concrete, if
primitive, examples of domain-specific oracles.
We examined well-known architectural deficiencies with CASs identified by the theorem proving
community and argued that the integration of
interactive theorem provers (ITPs) with CASs,
an objective that has been an area of research in
the respective communities for several decades,
11
provides a set of research problems and practical
software projects related to the development of
powerful and robust math oracles on a multidecade time horizon. Independent of their role as
domain-specific oracles, such systems may also
prove to be useful tools for AI safety researchers
in proving the functional correctness of other
components of an AI architecture.
Natural
choices of systems to use would be interfaces
for the Wolfram Language, the most widely
used computer algebra system, with one of the
HOL family of theorem provers or Coq, both of
which have substantial repositories of formalized
proofs [72–75], or a more modern ITP such as
Lean [51, 76].
Rather than representing a bold and profound
new agenda, we view these projects as being
concrete and achievable goals that may pave
the way to more substantial research directions.
Because the topics we have discussed have a
long and rich academic history, there are a
number of “shovel-ready” projects appropriate
for students anywhere from undergraduates to
PhD students and beyond. Good undergraduate
research projects would probably start with some
basic data science to catalogue core computer
algebra algorithms by their usage and popularity.
From there, it would be useful to have an
estimate of what certified implementations of
these algorithms would entail, whether formally
verified implementations, or along the lines of
Kaliszyk and Wiedijk’s HOLCAS system where
the CAS is built on top of a theorem prover.
Also useful would be a systematic study of role
that computer algebra has played in mathematics
and theoretical physics. This would have some
interesting overlap with cognitive psychology, and
these three projects together would make for an
approachable undergraduate thesis, or a beginning project for a graduate student. A solid PhD
thesis devoted to the topic of Oracle AI might
involve tackling approaches to oracles stemming
from reinforcement learning (RL) [8, 16], as
well as more advanced theorem proving and
CAS related topics such as investigating the
development of a hybrid architecture that would
allow for proof-checking. A student who worked
on these projects for several years would develop
a unique skill set spanning philosophy, machine
12
learning, theorem proving, and computer algebra.
In the context of superintelligent oracle AIs
which may possess the ability to manipulate a
human user, we differentiate between addressing
architectural or algorithmic deficiencies in subsystems versus general control methods or containment strategies. Given that strong mathematical
capabilities are likely to be useful in the construction of more general AI systems, designing robust
CASs (and any other domain-specific oracle) is an
important counterpart to general control strategies, as the top-level AI system will have fewer
loopholes to exploit. Controlling OAIs poses
a distinct set of challenges for which concrete
mathematical analysis is in its infancy [8, 9, 16].
Nonetheless, considering how little attention has
been given to the superintelligence control problem in general, we are optimistic about the potential to translate the high-level analyses of OAIs
that have arisen in the AI safety community
into the mathematical and software frameworks
of modern artificial intelligence.
Acknowledgements
We would like to thank Stuart Armstrong,
David Kristoffersson, Marcello Herreshoff, Miles
Brundage, Eric Drexler, Cristian Calude, and several anonymous reviewers for insightful discussions and feedback on the manuscript. We would
also like to thank the guest editors of Informatica,
Ryan Carey, Matthijs Maas, Nell Watson, and
Roman Yampolskiy, for organizing this special issue.
References
[1] N. Bostrom, Superintelligence: Paths, Dangers, Strategies. Oxford University Press,
2014.
[2] M. Shanahan, The Technological Singularity.
MIT Press, 2015.
Sarma and Hay
Artificial Intelligence (Future of Life Institute),” 2015.
[5] S. Russell, D. Dewey, and M. Tegmark, “Research Priorities for Robust and Beneficial
Artificial Intelligence,” AI Magazine, vol. 36,
no. 4, pp. 105–114, 2015.
[6] S. Armstrong, A. Sandberg, and N. Bostrom,
“Thinking inside the box: Controlling and
Using an Oracle AI,” Minds and Machines,
vol. 22, no. 4, pp. 299–324, 2012.
[7] B. Fallenstein, J. Taylor, and P. F. Christiano, “Reflective oracles: A foundation for
game theory in artificial intelligence,” in
Logic, Rationality, and Interaction, pp. 411–
415, Springer, 2015.
[8] S. Armstrong, “Value and policy networks as
Oracle AIs.” in preparation, 2017.
[9] S. Armstrong, “Good and safe uses of AI Oracles,” ArXiv e-prints, Nov. 2017.
[10] V. C. Müller and N. Bostrom, “Future
Progress in Artificial Intelligence: A survey
of expert opinion,” in Fundamental Issues of
Artificial Intelligence, pp. 553–570, Springer,
2016.
[11] K. Grace, J. Salvatier, A. Dafoe, B. Zhang,
and O. Evans, “When Will AI Exceed Human Performance? Evidence from AI Experts,” ArXiv e-prints, May 2017.
[12] C. Rovelli, “Quantum gravity,” Scholarpedia,
vol. 3, no. 5, p. 7117, 2008.
[13] S. Russell, “Should We Fear Supersmart
Robots?,” Scientific American, vol. 314,
no. 6, pp. 58–59, 2016.
[14] A. H. Eden, J. H. Moor, J. H. Soraker,
and E. Steinhart, Singularity Hypotheses:
A Scientific and Philosophical Assessment.
Springer Verlag, 2012.
[3] D. Chalmers, “The Singularity: A Philosophical Analysis,” Journal of Consciousness
Studies, vol. 17, no. 9-10, pp. 7–65, 2010.
[15] D. Amodei, C. Olah, J. Steinhardt, P. Christiano, J. Schulman, and D. Mané, “Concrete
Problems in AI Safety,” ArXiv e-prints, June
2016.
[4] M. Tegmark et al., “An Open Letter: Research Priorities for Robust and Beneficial
[16] S. M. Armstrong and L. Orseau, “Safely Interruptible Agents.” submitted, 2016.
Robust Computer Algebra, Theorem Proving, and Oracle AI
[17] D. Ferrucci, E. Brown, J. Chu-Carroll,
J. Fan, D. Gondek, A. A. Kalyanpur,
A. Lally, J. W. Murdock, E. Nyberg,
J. Prager, et al., “Building Watson: An
overview of the DeepQA project,” AI magazine, vol. 31, no. 3, pp. 59–79, 2010.
[18] W. Knight, “IBM Pushes Deep Learning
with a Watson Upgrade,” MIT Technology
Review, 7 2015.
[19] S. Wolfram, “Jeopardy, IBM, and Wolfram
— Alpha,” Stephen Wolfram — Blog, 1 2011.
13
Conference on Computer Aided Verification,
pp. 414–429, Springer, 2009.
[28] L. Fix, “Fifteen years of formal property verification in Intel,” in 25 Years of Model Checking, pp. 139–144, Springer, 2008.
[29] C. Kern and M. R. Greenstreet, “Formal verification in hardware design: a survey,” ACM
Transactions on Design Automation of Electronic Systems, vol. 4, no. 2, pp. 123–193,
1999.
[20] S. Weinzierl, “Computer Algebra in Particle
Physics,” ArXiv High Energy Physics - Phenomenology e-prints, Sept. 2002.
[30] T. Kropf, Introduction to Formal Hardware
Verification. Springer Science & Business
Media, 2013.
[21] Z. Huang, “Machine Learning and Computer
Algebra,” tech. rep., University of Cambridge, Computer Laboratory, 2016.
[31] C. Ballarin, Computer Algebra and Theorem
Proving. PhD thesis, University of Cambridge, Computer Laboratory, 1999.
[22] G. Irving, C. Szegedy, A. A. Alemi, F. Chollet, and J. Urban, “DeepMath—Deep Sequence Models for Premise Selection,” in
Advances in Neural Information Processing
Systems, pp. 2235–2243, 2016.
[32] C. Kaliszyk and F. Wiedijk, “Certified computer algebra on top of an interactive theorem prover,” in Towards Mechanized Mathematical Assistants, pp. 94–105, Springer,
2007.
[23] E. Komendantskaya, J. Heras, and G. Grov,
“Machine Learning in Proof General: Interfacing Interfaces,” ArXiv e-prints, Dec. 2012.
[33] S. M. Watt, “On the future of Computer Algebra Systems at the Threshold of 2010,”
Proceedings ASCM-MACIS, pp. 422–430,
2009.
[24] A. Bundy, D. Hutter, C. B. Jones, and J. S.
Moore, “AI meets Formal Software Development (Dagstuhl Seminar 12271),” Dagstuhl
Reports, vol. 2, no. 7, pp. 1–29, 2012.
[25] M. J. Beeson, “The Mechanization of Mathematics,” in Alan Turing: Life and Legacy of
a Great Thinker, pp. 77–134, Springer, 2004.
[26] G. Klein, K. Elphinstone, G. Heiser, J. Andronick, D. Cock, P. Derrin, D. Elkaduwe,
K. Engelhardt, R. Kolanski, M. Norrish,
et al., “seL4: Formal verification of an
OS kernel,” in Proceedings of the ACM
SIGOPS 22nd Symposium on Operating Systems Principles, pp. 207–220, ACM, 2009.
[27] R. Kaivola, R. Ghughal, N. Narasimhan,
A. Telfer, J. Whittemore, S. Pandav, A. Slobodová, C. Taylor, V. Frolov, E. Reeber,
et al., “Replacing Testing with Formal Verification in Intel CoreTM i7 Processor Execution Engine Validation,” in International
[34] W. Windsteiger, “Theorema 2.0: a system
for mathematical theory exploration,” in International Congress on Mathematical Software, pp. 49–52, Springer, 2014.
[35] P. G. Bertoli, J. Calmet, F. Giunchiglia, and
K. Homann, “Specification and integration of
theorem provers and computer algebra systems,” in International Conference on Artificial Intelligence and Symbolic Computation,
pp. 94–106, Springer, 1998.
[36] E. Clarke and X. Zhao, “Analytica—A theorem prover in Mathematica,” in International Conference on Automated Deduction,
pp. 761–765, Springer, 1992.
[37] A. Dolzmann and T. Sturm, “Redlog: Computer algebra meets computer logic,” ACM
SIGSAM Bulletin, vol. 31, no. 2, pp. 2–9,
1997.
14
[38] R. D. Jenks and R. S. Sutor, AXIOM: The
Scientific Computation System. Springer,
2013.
[39] E. Poll and S. Thompson, “Adding the axioms to Axiom,” tech. rep., Computing Laboratory, University of Kent, 1998.
[40] R. Miner, “The importance of MathML to
mathematics communication,” Notices of the
AMS, vol. 52, no. 5, pp. 532–538, 2005.
[41] S. Buswell, O. Caprotti, D. P. Carlisle, M. C.
Dewar, M. Gaetano, and M. Kohlhase, “The
Open Math Standard,” tech. rep., The Open
Math Society, 2004.
[42] J. Calmet and V. Lefevre, “Toward the Integration of Numerical Computations into the
OMSCS Framework,” in 7th International
Workshop on Computer Algebra in Scientific
Computing-CASC, pp. 71–79, 2004.
[43] J. Carette, W. M. Farmer, and R. O’Connor,
“MathScheme: project description,” in International Conference on Intelligent Computer Mathematics, pp. 287–288, Springer,
2011.
[44] A. Armando and D. Zini, “Towards Interoperable Mechanized Reasoning Systems:
the Logic Broker Architecture,” in AI*IATABOO Workshop: From Objects to Agents:
Evolutionary Trends of Software Systems,
pp. 70–75, 2000.
[45] A. Adams, M. Dunstan, H. Gottliebsen,
T. Kelsey, U. Martin, and S. Owre, “Computer algebra meets automated theorem
proving: Integrating Maple and PVS,” in International Conference on Theorem Proving
in Higher Order Logics, pp. 27–42, Springer,
2001.
[46] J. Harrison and L. Théry, “A skeptic’s approach to combining HOL and Maple,” Journal of Automated Reasoning, vol. 21, no. 3,
pp. 279–294, 1998.
[47] C. Ballarin, K. Homann, and J. Calmet,
“Theorems and algorithms: An interface between Isabelle and Maple,” in Proceedings
of the International Symposium on Symbolic
Sarma and Hay
and Algebraic Computation, pp. 150–157,
ACM, 1995.
[48] P. Jackson, “Exploring abstract algebra
in constructive type theory,” in International Conference on Automated Deduction,
pp. 590–604, Springer, 1994.
[49] J. Siekmann, C. Benzmüller, V. Brezhnev,
L. Cheikhrouhou, A. Fiedler, A. Franke,
H. Horacek, M. Kohlhase, A. Meier, E. Melis,
et al., “Proof development with OMEGA,” in
International Conference on Automated Deduction, pp. 144–149, Springer, 2002.
[50] C. Ballarin and L. C. Paulson, “A pragmatic approach to extending provers by computer algebra—with applications to coding
theory,” Fundamenta Informaticae, vol. 39,
no. 1, 2, pp. 1–20, 1999.
[51] R. Y. Lewis, “An extensible ad hoc interface
between Lean and Mathematica.” in preparation, 2017.
[52] E. Yudkowsky and M. Herreshoff, “Tiling
agents for self-modifying AI, and the Löbian
obstacle,” tech. rep., Machine Intelligence
Research Institute, 2013.
[53] P. LaVictoire, “An Introduction to Löbs
Theorem in MIRI Research,” tech. rep., Machine Intelligence Research Institute, 2015.
[54] M. Barasz, P. Christiano, B. Fallenstein,
M. Herreshoff, P. LaVictoire, and E. Yudkowsky, “Robust Cooperation in the Prisoner’s Dilemma: Program Equilibrium via
Provability Logic,” ArXiv e-prints, Jan.
2014.
[55] B. Fallenstein and N. Soares, “Problems
of self-reference in self-improving spacetime embedded intelligence,” in International Conference on Artificial General Intelligence, pp. 21–32, Springer, 2014.
[56] N. Soares and B. Fallenstein, “Toward Idealized Decision Theory,” ArXiv e-prints, July
2015.
[57] S. Garrabrant, T. Benson-Tilsen, A. Critch,
N. Soares, and J. Taylor, “Logical Induction,” ArXiv e-prints, Sept. 2016.
Robust Computer Algebra, Theorem Proving, and Oracle AI
[58] K. Beck, Test Driven Development: By Example. Addison Wesley, 2002.
[59] R. Osherove, The Art of Unit Testing: with
examples in C#. Manning Publications,
2013.
[60] E. M. Maximilien and L. Williams, “Assessing test-driven development at IBM,” in Proceedings of the 25th International Conference
on Software Engineering, pp. 564–569, IEEE,
2003.
[61] H. Erdogmus, “On the effectiveness of
test-first approach to programming,” IEEE
Transactions on Software Engineering,
vol. 31, no. 1, 2005.
[62] G. P. Sarma, T. W. Jacobs, M. D. Watts,
S. V. Ghayoomie, S. D. Larson, and R. C.
Gerkin, “Unit testing, model validation,
and biological simulation,” F1000Research,
vol. 5, 2016.
[63] C. Keller, A Matter of Trust: Skeptical
Communication Between Coq and External
Provers. PhD thesis, École Polytechnique,
2013.
[64] M. Armand, G. Faure, B. Grégoire, C. Keller,
L. Théry, and B. Werner, “A modular integration of SAT/SMT solvers to Coq through
proof witnesses,” in International Conference on Certified Programs and Proofs,
pp. 135–150, Springer, 2011.
[65] J. Harrison, “Towards self-verification of
HOL Light,” in International Joint Conference on Automated Reasoning, pp. 177–191,
Springer, 2006.
[66] R. Pollack, “How to believe a machinechecked proof,” Twenty Five Years of Constructive Type Theory, vol. 36, p. 205, 1998.
[67] G. Hardy and J. Hadamard, “The Psychology of Invention in the Mathematical Field,”
1946.
[68] S. Dehaene, The Number Sense: How the
Mind Creates Mathematics. Oxford University Press, 2011.
15
[69] P. Drijvers and K. Gravemeijer, “Computer
Algebra as an Instrument: Examples of Algebraic Schemes,” in The Didactical Challenge of Symbolic Calculators, pp. 163–196,
Springer, 2005.
[70] P. Drijvers, “Learning mathematics in a computer algebra environment: obstacles are opportunities,” Zentralblatt für Didaktik der
Mathematik, vol. 34, no. 5, pp. 221–228,
2002.
[71] G. Lakoff and R. Núñez, Where mathematics
comes from: How the embodied mind brings
mathematics into being. Basic books, 2000.
[72] S. Wolfram, An Elementary Introduction to
the Wolfram Language. Wolfram Media,
2015.
[73] L. C. Paulson, “The foundation of a generic
theorem prover,” Journal of Automated Reasoning, vol. 5, no. 3, pp. 363–397, 1989.
[74] L. C. Paulson, Isabelle: A generic theorem
prover, vol. 828. Springer Science & Business
Media, 1994.
[75] Y. Bertot and P. Castéran, Interactive
theorem proving and program development:
Coq‘Art: The Calculus of Inductive Constructions. Springer Science & Business Media, 2013.
[76] L. de Moura, S. Kong, J. Avigad,
F. Van Doorn, and J. von Raumer,
“The Lean Theorem Prover,” in International Conference on Automated Deduction,
pp. 378–388, Springer, 2015.
| 2cs.AI
|
arXiv:1702.08050v3 [math.GR] 26 Jun 2017
Hyperbolic actions and 2nd bounded cohomology
of subgroups of Out(Fn)
Part II: Finite lamination subgroups
Michael Handel and Lee Mosher
∗
March 16, 2018
Abstract
This is the second part of a two part work in which we prove that for every
finitely generated subgroup Γ < Out(Fn ), either Γ is virtually abelian or its
second bounded cohomology Hb2 (Γ; R) contains an embedding of ℓ1 . Here in
Part II we focus on finite lamination subgroups Γ — meaning that the set of
all attracting laminations of elements of Γ is finite — and on the construction
of hyperbolic actions of those subgroups to which the general theory of Part I
is applicable.
1
Introduction
This is the second part of a two part work, the main theorem of which is an alternative
for the second bounded cohomology of finitely generated subgroups of Out(Fn ), the
outer automorphism group of a free group of finite rank n:
Theorem A. For every finitely generated subgroup Γ < Out(Fn ), either Γ is virtually
abelian or Hb2 (Γ; R) has an embedded copy of ℓ1 and so is of uncountable dimension.
For background on Hb2 (Γ; R) see Part I [HM15]. In that paper, Theorem A was
reduced to Theorem C which is the main result of this paper. Theorem C explains how
to produce useful hyperbolic actions for a certain class of finite lamination subgroups
of Out(Fn ). To fully state Theorem C we shall first define the properties that arise
in its hypotheses.
Consider a group action G y S on a hyperbolic space. Recall that γ ∈ G is
loxodromic if its action on the Gromov boundary ∂S has north–south dynamics with
∗
The first author was supported by the National Science Foundation under Grant No. DMS1308710 and by PSC-CUNY under grants in Program Years 46 and 47. The second author was
supported by the National Science Foundation under Grant No. DMS-1406376.
1
a unique repeller-attractor pair (∂− γ, ∂+ γ) ∈ ∂S × ∂S − ∆. Recall also that G y S
is nonelementary if there exist independent loxodromic elements δ, γ ∈ G, meaning
that the sets {∂− δ, ∂+ δ}, {∂− γ, ∂+ γ} are disjoint. Given a loxodromic element γ ∈ Γ,
we say that γ satisfies WWPD ([BBF15], and see Proposition 2.6 of Part I [HM15])
if the G-orbit of the ordered pair (∂− γ, ∂+ γ) is a discrete subset of the space of distinct
ordered pairs ∂S × ∂S − ∆.
Let IAn (Z/3) < Out(Fn ) denote the finite index normal subgroup consisting of
all outer automorphisms whose induced action on H1 (Fn ; Z/3) is trivial. A subgroup
Γ < IAn (Z/3) has (virtually) abelian restrictions (Definition 2.1) if for each proper
free factor A < Fn whose conjugacy class [A] is fixed by each element of Γ, the natural
restriction homomorphism Γ 7→ Out(A) (virtually) abelian image. As explained in
Part I [HM15], the property of Γ < IAn (Z/3) having (virtually) abelian restrictions
plays a role in our theory analogous to the role played by irreducible subgroups of
mapping class groups in the theory of Bestvina and Fujiwara [BF02].
The decomposition theory of Bestvina, Feighn and Handel [BFH00] associates to
each φ ∈ Out(Fn ) a finite set L(φ) of attracting laminations. Associated to a subgroup
Γ < Out(Fn ) is the set L(Γ) = ∪φ∈Γ L(Γ). If L(Γ) is finite then Γ is a finite lamination
subgroup, otherwise Γ is an infinite lamination subgroup. In Part I we proved two
theorems for application to Theorem A, namely: Theorem B of Part I which is about
infinite lamination subgroups with virtually abelian restrictions, their actions on the
free splitting complex of Fn , and their WWPD elements; and Theorem D of Part I
which is a general result for determining second bounded cohomology of a group
possessing a hyperbolic action with a sufficiently rich collection of WWPD elements.
Also in Part I, by combining Theorem B with Theorem D we reduced the proof of
Theorem A to the following result which is the main theorem of the present paper.
Theorem C. For any finitely generated, finite lamination subgroup Γ < IAn (Z/3)
such that Γ is not abelian but Γ has virtually abelian restrictions, there exists a finite
index normal subgroup N ⊳ Γ and an action N y S on a hyperbolic space, such that
the following hold:
(1) Every element of N acts either elliptically or loxodromically on S;
(2) The action N y S is nonelementary;
(3) Every loxodromic element of the commutator subgroup [N, N] satisfies WWPD
with respect to the action N y S.
Remarks: WPD versus WWPD. In lectures on this topic we stated a stronger
version of Theorem C (3), saying that in the group Image(N 7→ Isom(S)), loxodromic
elements of the commutator subgroup satisfy WPD. That requires a stronger hypothesis, saying roughly that “virtually abelian restrictions” holds not just for free factors
but for a broader class of subgroups of Fn . That also makes the proof and application
2
of Theorem C considerably more intricate. In the interests of keeping the paper from
growing ever longer, we have settled for the version of Theorem C presented here.
Methods of proof of Theorem C
The first step of Theorem C will be to reduce it to a statement about subgroups
of automorphism groups of free groups, stated in Theorem F below. The key step
of this reduction, proved in Proposition 2.3 of Section 2, is the construction of an
“automorphic lift” of each Γ satisfying the hypotheses of Theorem C: there exists a
free factor A < Fn having rank k for some 2 ≤ k ≤ n − 1, such that its conjugacy
class [A] is Γ-invariant, and such that the natural homomorphism Γ 7→ Out(A) lifts
to a homomorphism Γ 7→ Aut(A) whose image in Aut(A) is not virtually abelian. In
Section 2.4 we shall show, by minimizing the rank of the free factor A, how to reduce
Theorem C to the following statement, in which we have identified A ≈ Fk and then
rewritten k as n. Recall the canonical isomorphism Fn ≈ Inn(Fn ) which associates to
each γ ∈ Fn the inner automorphism iγ (δ) = γδγ −1 .
b < Aut(Fn ) with n ≥ 2, and denote H =
Theorem F. Consider a subgroup H
b 7→ Out(Fn )) and J = Kernel(H
b 7→ H) = H
b ∩ Inn(Fn ), giving the following
Image(H
commutative diagram of short exact sequences:
1
/
⊂
J
/
⊂
1
/
//
H
⊂
Fn ≈ Inn(Fn )
b
H
⊂
/
Aut(Fn )
/
1
⊂
//
Out(Fn )
/
1
b is finitely generated and not virtually abelian, and
Suppose that H is abelian, that H
b on the set of
that no proper, nontrivial free factor of Fn is fixed by the action of H
b and an
subgroups of Fn . Then there exists a finite index normal subgroup N < H
action N y S on a hyperbolic space such that the following properties hold:
(1) Every element of N acts either elliptically or loxodromically on S;
(2) The action N y S is nonelementary;
(3) Every loxodromic element of J ∩ N satisfies WWPD with respect to the action
N y S.
The proof of Theorem F begins in Section 2.6 by assuming H < IAn (Z/3), which
we may do by replacing H by its intersection with IAn (Z/3); note that the conclusion
of Theorem F holds for H if and only if it holds for some (any) finite index subgroup
of H. Then we consider a maximal, proper, H-invariant free factor system B in Fn .
The proof breaks into cases depending on the “co-edge number” of B, which is the
3
minimum integer k ≥ 1 such that B is represented by a subgraph H ⊂ G of a marked
graph G for which the complement G \ H has k edges. The “one-edge” case, where
b on
the co-edge number of B equals 1, is handled in Section 2.6 using an action of H
a simplicial tree that is naturally associated to the free factor system B.
The “multi-edge” case, where the co-edge number of B is ≥ 2, takes up the
majority of the paper from Section 3 to the end. For a full introduction to the
multi-edge case, see Section 3. In brief, one uses the dynamics of EG strata to
produce a certain hyperbolic suspension space S, applying the Mj-Sardar combination
theorem [MS12] to prove hyperbolicity. The construction of S, including the flaring
properties needed to apply the combination theorem, is found in Sections 4 and 5.
The construction of the action on S, based on abelian subgroup methods, is found
in Sections 6, 7 and 8.1. The pieces are put together, and the multi-edge case of
Theorem F is proved, in Section 8.2.
Prerequisites from the theory of Out(Fn ). We will assume that the reader is
familiar with certain basic concepts of Out(Fn ) that have already been reviewed in
Part I of this paper [HM15], in particular:
[HM15, Section 3.1] Marked graphs and topological representatives; free factor
systems; relative train track maps; attracting laminations.
[HM15, Section 4.1] Properties of IAn (Z/3) = Kernel Out(Fn ) 7→ GL(n, Z/3) .
[HM15, Section 4.1] The co-edge number of a free factor system A in Fn ; elements
and subgroups of Out(Fn ) which are fully irreducible relative to a free factor
system A of Out(Fn ).
Where needed in this paper, we will conduct reviews of other basic concepts.
Contents
1 Introduction
1
2 Lifting to an automorphism group
2.1 Definition of automorphic lifts. . . . . . . . . . . . . . .
2.2 A sufficient condition to be abelian. . . . . . . . . . . . .
2.3 Constructing automorphic lifts: proof of Proposition 2.3
2.4 Proof that Theorem F implies Theorem C . . . . . . . .
2.5 Automorphic extensions of free splitting actions . . . . .
2.6 Case analysis of Theorem F. Proof of the one-edge case. .
.
.
.
.
.
.
6
6
8
9
12
13
18
3 Introduction to the multi-edge case of Theorem F.
3.1 Outline of the multi-edge case. . . . . . . . . . . . . . . . . . . . . . .
3.2 Motivation: Suspension actions and combination theorems. . . . . . .
20
20
23
4
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
4 Flaring in a top EG stratum
4.1 The path functions Lu and LPF . . . . . . . . . . . .
4.2 Flaring of the path functions Lu and LPF . . . . . .
4.3 Negative flaring of Lu . . . . . . . . . . . . . . . . .
4.4 Positive flaring and other properties of Lu . . . . .
4.5 Proof of Lemma 4.11: the Special Flaring Condition
4.6 Appendix: The graph homotopy principle . . . . .
. .
. .
. .
. .
for
. .
. .
. .
. .
. .
Lu
. .
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
5 Flaring in T ∗ and hyperbolicity of S.
5.1 The free splitting Fn y T and its role in flaring. . . . . . . . . . .
e and T . A piecewise Riemannian metric on T .
5.2 Path functions on G
5.3 Constructing T ∗ by coning off Nielsen axes of T . . . . . . . . . . .
5.4 Geometry and dynamics on T ∗ . . . . . . . . . . . . . . . . . . . .
5.5 Construction of S. . . . . . . . . . . . . . . . . . . . . . . . . . .
5.6 Proof of hyperbolicity of S. . . . . . . . . . . . . . . . . . . . . .
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
25
26
31
35
37
41
45
.
.
.
.
.
.
47
47
50
52
55
60
62
6 Abelian subgroups of Out(Fn )
6.1 Background review . . . . . . . . . . . . . . . . . . . . . . . . . . . .
6.1.1 More about CTs . . . . . . . . . . . . . . . . . . . . . . . . .
6.1.2 Principal automorphisms and rotationless outer automorphisms.
6.1.3 Rotationless abelian subgroups . . . . . . . . . . . . . . . . .
6.2 Disintegration subgroups . . . . . . . . . . . . . . . . . . . . . . . . .
6.2.1 QE-paths and QE-splittings . . . . . . . . . . . . . . . . . . .
6.2.2 Almost invariant subgraphs . . . . . . . . . . . . . . . . . . .
6.2.3 Admissible S-tuples; quasi-exceptional families . . . . . . . . .
6.2.4 Xs paths. . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
6.2.5 The disintegration subgroup D(f ) . . . . . . . . . . . . . . .
6.2.6 The coordinate homomorphism of a disintegration group D(f )
66
66
66
68
71
72
72
73
73
74
75
76
7 A train track semigroup action
7.1 A “homotopy semigroup action” of D+ on G. . . . . . . . . . . . . .
b+ on T and T ∗ . . . . . . . . . . . . . . . . .
7.2 Semigroup actions of D
b0 y T ∗ . . . . . . . . . . . . . . . . .
7.3 Dynamics of the group action D
77
79
82
84
8 The suspension action.
b y S. . . . . . . . . . . . . . . . . . . . . . .
8.1 The suspension action D
8.2 Proof of Theorem F: The multi-edge case. . . . . . . . . . . . . . . .
5
86
86
88
2
Lifting to an automorphism group
In this section, the first thing we do is to study the structure of finitely generated, finite
lamination subgroups Γ < IAn (Z/3) which are not abelian but have virtually abelian
restrictions. The motivating question of that study is this: If Γ fixes the conjugacy
class [A] of a proper, nontrivial free factor A < Fn , can the natural restriction map
Γ 7→ Out(A) be lifted to a homomorphism Γ 7→ Aut(A)? Sections 2.1–2.3 are devoted
to constructions of such “automorphic lifts”. Using this construction, in Section 2.4
we prove the implication Theorem F =⇒ Theorem C.
After that, in Section 2.5, we consider any free splitting Fn y T , and we study a
natural subgroup of Aut(Fn ) to which the free splitting action extends. That study
is used in Section 2.6 to prove one of the two major cases of Theorem F.
2.1
Definition of automorphic lifts.
Recall [HM17a, Fact 1.4] that for any group and subgroup H < G which is its own
normalizer (e.g. a free factor), letting Stab[H] < Out(G) be the stabilizer of the
conjugacy class of H, the natural restriction homomorphism Stab[H] 7→ Out(H),
denoted φ 7→ φ H, is well-defined by choosing Φ ∈ Aut(G) representing φ and
preserving H and taking φ H to be the outer automophism class of Φ H ∈ Aut(H).
Throughout the paper we use the theorem that virtually abelian subgroups of
IAn (Z/3) are abelian [HM17e]. We sometimes write “(virtually) abelian” as a reminder that one may freely include or ignore the adverb “virtually” in front of the
adjective “abelian” in the context of a subgroup of IAF (Z/3) for any finite rank free
group F . One has this freedom, for example, in the following definition (see [HM15,
Corollary 4.1]):
Definition 2.1 ((Virtually) Abelian restrictions). A subgroup Γ < IAn (Z/3) has
(virtually) abelian restrictions if for any proper free factor A < Fn such that Γ <
Stab[A], the restriction homomorphism Γ 7→ Out(A) has (virtually) abelian image.
Definition 2.2. Let Γ < IAn (Z/3) be a finitely generated, finite lamination subgroup
which is not (virtually) abelian and which has (virtually) abelian restrictions. An
automorphic lift of Γ is a homomorphism ρ : Γ 7→ Aut(A), where A < Fn is a proper
b = Image(ρ) is not virtually
free factor and Γ < Stab[A], such that the group H
abelian, and such that the following triangle commutes
Aut(A)
<
②②
②
②②
②②
②②
/ Out(A)
ρ
Γ
In this diagram the horizontal arrow is the natural restriction homomorphism Stab[A] 7→
Out(A) with domain restricted to Γ, and the vertical arrow is the natural quotient
6
Γ✾
✾✾
✾✾
✾✾
✾✾
✾✾
/H
b
//H
✾✾
✾✾
⊂
✾✾ ⊂
✾
/ Aut(A)
/ / Out(A)
ρ
1
/
⊂
J
⊂
1
/
A ≈ Inn(A)
⊂
/
/
1
1
b
Figure 1: Notation associated to an automorphic lift ρ : Γ → Aut(A) with image H.
b is not virtually abelian, the quotient H is virtually abelian, and the
The group H
kernel J is free of rank ≥ 2 (possibly infinite). The horizontal rows are exact.
homomorphism. To emphasize the role of A we will sometimes refer to an automorb = Image(ρ), H =
phic lift of Γ rel A. Adopting the notation of Theorem F, we set H
b 7→ Out(A)) = Image(Γ 7→ Out(A)), and J = Kernel(H
b 7→ H) = H
b ∩ Inn(A),
Image(H
thus obtaining the commutative diagram shown in Figure 1. We note two properties
which follow from the definition:
• H is abelian;
• The free group J has rank ≥ 2.
The first holds because A < Fn is proper and Γ has virtually abelian restrictions
(see Definition 2.1). The second is a consequence of the first combined with the
b is not virtually abelian, for otherwise the free group J is
defining requirement that H
b is virtually solvable, but Aut(A) injects into Out(Fn ) and solvable
abelian and so H
subgroups of Out(Fn ) are virtually abelian by [BFH04]. We put no further conditions
on the rank of J, it could even be infinite. When referring to J, each of its elements
will be thought of ambiguously as an element of the free factor A < Fn or as the
corresponding element of the inner automorphism group Inn(A); this ambiguity should
cause little trouble, by using the canonical isomorphism A ↔ Inn(A) given by δ ↔ iδ
where iδ (γ) = δγδ −1 .
The rank of the automorphic lift Γ 7→ Aut(A) is defined to be rank(A), and
note that rank(A) ≥ 2 because otherwise Aut(A) is finite in which case each of its
subgroups is virtually abelian.
This completes Definition 2.2.
Here is the first of two main results of Section 2.
Proposition 2.3. If Γ < IAn (Z/3) is a finitely generated, finite lamination subgroup
which is not (virtually) abelian but which has (virtually) abelian restrictions, then
there exists an automorphic lift Γ 7→ Aut(A).
7
The proof is found in Section 2.3, preceded by Lemma 2.4 in Section 2.2.
When n = 2, one recovers from Proposition 2.3 the simple fact that every finite
lamination subgroup of Out(F2 ) is virtually abelian, for otherwise the intersection with
IAn (Z/3) would have an automorphic lift to Aut(A) for some proper free factor A, from
which it would follow that 2 ≤ rank(A) ≤ n − 1 = 1. Of course this fact has a simple
proof, expressed in terms the isomorphism Out(F2 ) 7→ Aut(H1 (F2 ; Z)) ≈ GL(2, Z),
which we leave to the reader.
2.2
A sufficient condition to be abelian.
In this section we prove Lemma 2.4 which gives a sufficient condition for a finitely
generated, finite lamination subgroup Γ < IAn (Z/3) to be abelian. The negation of
this condition then becomes a property that must hold when Γ is not abelian.
Lemma 2.4. Let Γ < IAn (Z/3) be a finitely generated, finite lamination subgroup.
If A = {[A1 ], . . . , [AI ]} is a maximal proper Γ-invariant free factor system, if each
restriction Γ Ai < Out(Ai ) is abelian for i = 1, . . . , I, and if A has co-edge number
≥ 2 in Fn , then Γ is abelian.
Proof. By [HM17d, Theorem C], there exists η ∈ Γ which is fully irreducible rel A.
±
By relative train track theory, there is a unique lamination pair Λ±
η ∈ L (η) for η
which is not carried by A. By [HM17c, ], the nonattracting subgroup system of Λ±
η
±
has one of two forms, either Ana (Λ±
η ) = A or Ana (Λη ) = A ∪ {[C]} where C is a
maximal infinite cyclic group not carried by A (see [HM17c, Definition 1.2]). For
every nonperiodic line ℓ that is not carried by A, evidently ℓ is not carried by [C],
and so ℓ is not carried by Ana (Λ±
η ); by applying [HM17c, Theorem H] it then follows
−1
−
that ℓ is weakly attracted to either Λ+
η by iteration of η or to Λη by iteration of η .
+
−
+
From this it follows that the two laminations Λη , Λη−1 = Λη are the unique elements
of L(Γ) not carried by A, for if there existed ψ ∈ Γ with attracting lamination
+
±
Λ+
ψ ∈ L(Γ) − {Λη } not supported by A then the generic lines of Λψ would be weakly
−
−1
attracted either to Λ+
η by iteration of η or to Λη by iteration of η , and in either
+
case the set of laminations η k (Λ+
ψ ) = Ληk ψη−k ∈ L(Γ), k ∈ Z, would form an infinite
set, contradicting that Γ is a finite lamination group.
Since L(Γ) is finite, for each φ ∈ Γ each element of L(Γ) has finite orbit under
the action of of φ. Since Γ < IAn (Z/3), it follows by [HM15, Lemma 4.8] that each
element of L(Γ) is fixed by φ. In particular, Γ < Stab(Λ+
η ). By [BFH00, Corollary
+
3.3.1], there exists a homomorphism PFΛ+η : Stab(Λη ) → R having the property that
(φ) 6= 0 holds if and only if Λ+
for each φ ∈ Stab(Λ+
η ∈ L(φ).
η ), the inequality PFΛ+
η
Consider the following homomorphism, the range of which is abelian:
Ω:Γ →
R
⊕ Γ A1 ⊕ · · · ⊕ Γ AI
Ω(φ) = P FΛ+η (φ) ⊕ φ A1 ⊕ · · · ⊕ φ AI
8
We claim that that the kernel of Ω is also abelian. This claim completes the proof of
the lemma, because every solvable subgroup of Out(Fn ) is virtually abelian [BFH04],
and every virtually abelian subgroup of IAn (Z/3) is abelian [HM17e].
To prove the claim, by Proposition 5.2 of [HM], every subgroup of Kernel P FΛ+η (φ)
consisting entirely of UPG elements is abelian, and so we need only check that each
element of Kernel(Ω) is UPG. By [BFH00, Corollary 5.7.6], every PG element of
IAn (Z/3) is UPG, and so we need only check that each φ ∈ Kernel(Ω) is PG, equivalently L(φ) = ∅. Suppose to the contrary that there exists Λ+
φ ∈ L(φ), with dual
−
−1
repelling lamination Λφ ∈ L(φ ). Since φ Ai is trivial in Out(Ai ) for each component [Ai ] of A, neither of the laminations Λ±
φ is supported by A. Since L(φ) ⊂ L(Γ),
+
−
−
(φ) 6= 0 a contrait follows as shown above that {Λφ , Λφ } = {Λ+
η , Λη }, and so P FΛ+
η
diction.
2.3
Constructing automorphic lifts: proof of Proposition 2.3
Let Γ < IAn (Z/3) be a finitely generated, finite lamination subgroup that is not
(virtually) abelian and has (virtually) abelian restrictions. Choose a maximal proper
Γ-invariant free factor system A = {[A1 ], . . . , [AI ]}, and so each restricted group
denoted Hi = Γ [Ai ] < Out(Ai ) is abelian. Since Γ < IAn (Z/3), it follows that each
component [Ai ] of A is fixed by Γ [HM17b, Lemma 4.2].
The group Γ is not abelian by [HM17e], and so applying by Lemma 2.4 the extension A ⊏ {[Fn ]} is a one-edge extension. We may therefore choose a marked graph
pair (G, H) representing A so that G \ H = E is a single edge, and we may choose H
so that its components are roses, with each endpoint of E being the rose vertex of a
component of H. The number of components of A equals the number of components
of H, that number being either one or two, and we cover those cases separately.
Case 1: A has two components, say A = {[A1 ], [A2 ]} where Fn = A1 ∗ A2 . We
construct a commutative diagram as follows:
q
/ Out(A1 ) ⊕ Out(A2 )
O
❯❯❯❯
❯❯❯❯
❯❯❯❯
ρ=ρ1 ⊕ρ2
❯❯❯❯
α=α1 ⊕α2
❯❯❯❯
r
✐✐✐ Γ
✐✐✐✐
ω✐✐✐✐✐✐
⊂
✐✐✐
t✐✐✐✐
/ Out(Fn )
Aut(Fn , A1 , A2 )
Aut(A1 ) ⊕O Aut(A
)
j 2
Aut(Fn , A1 , A2 ) is the subgroup of Aut(Fn ) that preserves both A1 and A2 . The
homomorphism r is induced by restricting Aut(Fn , A1 , A2 ) to Aut(A1 ) and to Aut(A2 ).
Evidently r is injective, since an automorphism of Fn is determined by its restrictions
to the complementary free factors A1 , A2 . The homomorphisms denoted by the top
and bottom arrows of the diagram are induced by canonical homomorphisms from
9
automorphism groups to outer automorphism groups. For i = 1, 2 the homomorphism
ρi is the composition Γ ֒→ Stab[Ai ] 7→ Out(Ai ) where the latter map is the natural
restriction homomorphism.
We must construct ω, α1 , α2 . We may choose the marked graph pair (G, H)
representing the free factor system A to have the following properties: the two rose
components H1 , H2 of the subgraph H have ranks equal to rank(A1 ), rank(A2 ) respectively; the edge E is oriented and decomposes into oriented half-edges E = E 1 E2 ;
the common initial point of E1 , E2 is denoted w; their respective terminal vertices are
the rose vertices vi ∈ Hi ; and there is an isomorphism π1 (G, w) ≈ Fn which restricts
to isomorphisms π1 (Ei ∪ Hi , w) ≈ Ai for i = 1, 2. Given φ ∈ Γ, let f : G → G
be a homotopy equivalence that represents φ, preserves H1 and H2 , and restricts
to a locally injective path on E = E 1 E2 . By Corollary 3.2.2 of [BFH00] we have
f (E) = ū1 E ±1 u2 for possibly trivial paths closed paths ui in Hi , i = 1, 2, and in
fact the plus sign occurs and so f (E) = ū1 Eu2 , because φ ∈ Γ < IAn (Z/3). After
pre-composing f with a homeomorphism of G isotopic to the identity that restricts
to the identity on H1 ∪ H2 and that moves the point w ∈ E to f −1 (w) ∈ E, we
may also assume that f (w) = w, and so f (E1 ) = E1 u1 and f (E2 ) = E2 u2 . Define
αi (φ) ∈ Aut(Ai ) ≈ Aut(π1 (Ei ∪Hi , vi )) to be the automorphism induced by f Ei ∪Hi ,
and then use the isomorphism Fn = A1 ∗ A2 to define ω(φ). Note that ω(φ) is the
unique lift of φ ∈ Out(Fn ) to Aut(Fn ) which preserves A1 and A2 , because any two
such lifts differ by an inner automorphism ic that preserves both A1 and A2 , implying
by malnormality that c ∈ A1 ∩ A2 and so is trivial. It follows from uniqueness that ω,
α1 , and α2 are homomorphisms, and hence so is α. Commutativity of the diagram is
straightforward from the construction. The homomorphism ω is injective because it
is a lift of the inclusion Γ ֒→ Out(Fn ). Since r and ω are injective, by commutativity
of the diagram α is also injective.
At least one of the two maps αi : Γ → Aut(Ai ) is an automorphic lift because
bi = Image(αi ) < Aut(Ai ) is not virtually
at least one of the corresponding images H
b1 ⊕ H
b2 would be virtually
abelian: if both were virtually abelian, then α(Γ) < H
abelian, but α is injective and so Γ would be virtually abelian, a contradiction.
Case 2: A has a single component, say A = {[A]}, and so Fn = A ∗ hbi for some
b ∈ Fn . The proof in this case is similar to Case 1, the main differences being that in
place of direct sum we use fiber sum, and the marked graph pair (G, H) representing
A will have connected subgraph H.
10
We shall construct the following commutative diagram:
Aut2 (A, Ab )
q
/ Out(A)
O
i❚❚❚❚
❚❚❚❚
ρ
❚
❚
❚❚❚❚
α
❚❚❚❚
r
H
❥❥❥❥
❥
❥
❥
❥❥
⊂
❥❥❥❥
u❥❥❥❥ ω
/ Out(Fn )
Aut(Fn , A, Ab )
O
In this diagram we use the following notations: the conjugate Ab = bAb−1 ; the
restricted inner automorphism ib : A → Ab where ib (a) = bab−1 ; the adjoint isomorphism Adb : Aut(A) → Aut(Ab ) where Adb (Φ) = ib ◦ Φ ◦ i−1
b ; the canonical epimorphism qA : Aut(A) → Out(A); and the epimorphism qAb = qA ◦ Ad−1
:
b
b
Aut(A ) → Out(A). Also, the fiber sum of qA and qAb is the following subgroup of
Aut(A) ⊕ Aut(Ab ):
Aut2 (A, Ab ) = {(Φ, Φ′ ) ∈ Aut(A) ⊕ Aut(Ab ) qA (Φ) = qAb (Φ′ )}
Define the homomorphism q by q(Φ, Φ′ ) = qA (Φ) = qAb (Φ′ ). Define Aut(Fn , A, Ab ) <
Aut(Fn ) to be the subgroup that preserves both A and Ab . The homomorphism r is
jointly induced by the two restriction homomorphisms rA , rAb from Aut(Fn , A, Ab ) to
Aut(A), Aut(Ab ) because the two compositions qA ◦ rA , qAb ◦ rAb : Aut(Fn , A, Ab ) →
Out(A) are evidently the same. Note that r is injective, for if Φ ∈ Aut(Fn , A, Ab )
restricts to the identity on each of A and Ab , then Φ(a) = a and Φ(ab ) = ab for all
a ∈ A, and it follows that b−1 Φ(b) commutes with all a ∈ A; since rank(A) ≥ 2, we
have Φ(b) = b, and hence Φ is trivial.
To construct the homomorphism ω, we may in this case choose the marked graph
pair (G, H) representing A so that: H is a rose whose rank equals rank(A) = n−1; as
before E = E 1 E2 with w the common initial point of E1 , E2 ; the terminal endpoints of
both E1 and E2 equal the rose vertex v ∈ H; and there is an isomorphism π1 (G, w) ≈
Fn which restricts to π1 (E1 ∪H, w) ≈ A and E2 E 1 ≈ b. It follows that π1 (E2 ∪H, w) ≈
Ab . For each φ ∈ H, applying Corollary 3.2.2 of [BFH00] as in the previous case,
we find Φ = ω(φ) ∈ Aut(Fn , A, Ab ) that represents φ and that is represented by a
homotopy equivalence fφ : G → G that preserves H, fixes w, and takes Ei 7→ Ei ui
for possibly trivial paths u1 , u2 in H based at v. The map ω is a homomorphism
because Φ is the unique element of Aut(Fn , A, Ab ) that represents φ, for if c ∈ Fn
and if the inner automorphism ic preserves both A and Ab then by malnormality of
A and Ab we have c ∈ A ∩ Ab , and again by malnormality we have that c is trivial.
It follows that ω is injective. Since r is injective it follows that α = r ◦ ω is injective.
Denote α(φ) = (αA (φ), αAb (φ)) ∈ Aut2 (A, Ab ). Obviously the compositions qA ◦ αA ,
qAb ◦ αAb : H → Out(A) are the same, and hence there is an induced homomorphism
ρ : H → Out(A) which is topologically represented by fφ H. This completes the
construction of the above diagram, and commutativity is evident.
11
As in the previous case, we will be done if we can show that at least one of the
two homomorphisms αA : Γ0 → Aut(A) or αAb : Γ0 → Aut(Ab ) has image that is not
virtually abelian, but if both are virtually abelian then Image(α) is contained in a
virtually abelian subgroup of Aut2 (A, Ab ), which by injectivity of α implies that Γ is
virtually abelian, a contradiction.
2.4
Proof that Theorem F implies Theorem C
Let Γ < IAn (Z/3) be a finitely generated, finite lamination subgroup which is not
(virtually) abelian and which has (virtually) abelian restrictions. By applying Proposition 2.3, there exists a free factor A < Fn such that Γ < Stab[A], and there exists an
automorphic lift ρ : Γ 7→ Aut(A); we may assume that rank(A) is minimal amongst all
choices of A and ρ. We adopt the notation of Figure 1 in Section 2.1, matching that
notation with Theorem F by choosing an isomorphism A ≈ Fk where k = rank(A).
b = Image(ρ) < Aut(A)
Most of the hypotheses of Theorem F are now immediate: H
is finitely generated, and it is not virtually abelian by definition of automorphic lifts;
b 7→ Out(A)) is abelian.
also H = Image(H
We must check the one remaining hypothesis of Theorem F, namely that no
b Assuming
proper, nontrivial free factor B < A is preserved by the action of H.
b consider the restriction homomorphism
by contradiction that B is preserved by H,
b
σ : H 7→ Aut(B). We claim that the composition σρ : Γ → Aut(B) is an automorphic lift of Γ. Since rank(B) < rank(A), once this claim is proved, it contradicts
the assumption that ρ : Γ → Aut(A) is an automorphic lift of minimal rank. The
canonical isomorphism Inn(Fk ) ↔ Fk , denoted iδ ↔ δ, restricts to an isomorphism
b ∩ Inn(Fk ) and some subgroup of Fk . If iδ ∈ J then iδ preserves
between J = H
B, and since B is malnormal in A it follows that δ ∈ B. Thus σ restricts to an
injection from J to Inn(B). Also, the group J is a free group of rank ≥ 2, for if
b would be virtually solvable and hence, by [BFH04],
it were trivial or cyclic then H
ρ
σ
b−
virtually abelian, a contradiction. Since the image of the map σρ : Γ −
→H
→ Aut(B)
b preserves
contains σ(J), it follows that Image(σρ) is not virtually abelian. Since H
the A-conjugacy class of B, and since B is malnormal in Fn , it follows Γ preserves
the Fn -conjugacy class of B. Tracing through the definitions one easily sees that the
σρ
composed homomorphism Γ −→ Aut(B) 7→ Out(B) is equal to the composition of
Γ ֒→ StabOut(Fn ) [B] 7→ Out(B), where the latter map is the natural restriction homomorphism. Thus σρ : Γ → Aut(B) is an automorphic lift of Γ, completing the proof
of the claim.
b <
Applying the conclusions of Theorem F using the free group A ≈ Fk and H
b and a hyperbolic action
Aut(Fk ), we obtain a finite index normal subgroup N < H
N y S satisfying conclusions (1), (2) and (3) of that theorem. The subgroup N =
ρ−1 (N ) < Γ is a finite index normal subgroup of Γ, and by composition we have
an action N 7→ N y S. By Theorem F (1), each element of N acts elliptically
or loxodromically on S, and so the same holds for each element of N, which is
12
Theorem C (1). By Theorem F (2), the action N y S is nonelementary, and so
the same holds for the action N y S, which is Theorem C (2). Since the image
b 7→ H is abelian, it follows that the image in
of the homomorphism N 7→ N ֒→ H
b of the commutator subgroup [N, N] is contained in J = Kernel(H
b 7→ H), and
H
hence the image of [N, N] in N is contained in J ∩ N . By Theorem F (3), each
loxodromic element of J ∩ N is a WWPD element with respect to the action N y S.
By [HM15, Corollary 2.8], which says that WWPD is preserved under pullback, it
follows that every loxodromic element of [N, N] satisfies WWPD with respect to the
action N y S, which is Theorem C (3).
2.5
Automorphic extensions of free splitting actions
From the hypothesis of Theorem F, our interest is now transferred to the context of
a finite rank free group Fn — perhaps identified isomorphically with some free factor
b < Aut(Fn ) that has the following
of a higher rank free group — and of a subgroup H
b To
irreducibility property: no proper, nontrivial free factor of Fn is preserved by H.
b on hyperbolic spaces. In this
prove Theorem F one needs actions of such groups H
section we focus on a natural situation which produces actions on trees.
Free splittings. Recall that a free splitting of Fn is a minimal, simplicial action
Fn y T on a simplicial tree T such that the stabilizer of each edge is trivial. Two
free splittings Fn y S, T are simplicially equivalent if there exists an Fn -equivariant
simplicial isomorphism S 7→ T ; we sometimes use the notation [T ] for the simplicial
equivalence class of a free splitting Fn y T . Formally the action Fn y T is given by
a homomorphism α : Fn 7→ Isom(T ), which we denote more briefly as α : Fn y T .
In this formal notation, Isom(T ) refers to the group of simplicial self-isomorphisms
of T , equivalently the self-isometry group of T using the geodesic metric given by
barycentric coordinates on simplices of T . We note that an element of Isom(T ) is
determined by its restriction to the vertex set, in fact it is determined by its restriction
to the subset of vertices of valence ≥ 3. Two free splittings are equivalent if there
is an Fn -equivariant simplicial isomorphism between them, the equivalence class of a
free splitting Fn y T is denoted [T ], and the group Out(Fn ) acts naturally on the set
of equivalence classes of free splittings.
Given a free splitting Fn y T , the set of conjugacy classes of nontrivial vertex
stabilizers of a free splitting is a free factor system of Fn called the vertex group system
of T denoted as A(T ). The function which assigns to each free splitting T its vertex
group system A(T ) induces a well-defined, Out(Fn )-equivariant function
[T ] 7→ A(T )
from the set of simplicial equivalence classes of free splittings to the set of free factor
systems.
13
Every free splitting Fn y T can be realized by some marked graph pair (G, H) in
e where each
the sense that T is the Fn -equivariant quotient of the universal cover G
e
e
component of the total lift H ⊂ G is collapsed to a point. One may also assume that
each component of H is noncontractible, in which case the same marked graph pair
(G, H) topologically represents the vertex group system of T .
Twisted equivariance (functional notation). Given two free splittings Fn y
S, T and an automorphism Φ ∈ Aut(Fn ), a map h : S → T is said to be Φ-twisted
equivariant if
h(γ · x) = Φ(γ) · h(x) for all x ∈ S, γ ∈ Fn .
The special case when Φ = Id is simply called equivariance.
A twisted equivariant map behaves well with respect to stabilizers, as shown in
the following simple fact:
Lemma 2.5. For any free splittings Fn y S, T , for each Φ ∈ Aut(Fn ), and for each
Φ-twisted equivariant map f : S → T , we have
(1) Φ(Stab(x)) < Stab(f (x)) for all x ∈ S.
(2) If in addition the map f : S → T is a simplicial isomorphism, then the inclusion
of item (1) is an equality: Φ(Stab(x)) = Stab(f (x)). Furthermore, γ ∈ Fn acts
loxodromically on S if and only if Φ(γ) acts loxodromically on T , in which case
their axes ASγ ⊂ S and ATΦ (γ) ⊂ T satisfy ATΦ(γ) = f (ASγ ).
Remark. One could approach this proof by first working out the equivariant case
(Φ = Id), and then reducing the twisted case to the equivariant case by conjugating
the action Fn y T using Φ. We instead give a direct proof.
Proof. To prove (1), for each x ∈ S and γ ∈ Fn we have
γ ∈ Φ(Stab(x)) ⇐⇒
=⇒
⇐⇒
⇐⇒
⇐⇒
Φ−1 (γ) ∈ Stab(x) ⇐⇒ Φ−1 (γ) · x = x
f (Φ−1 (γ) · x) = f (x)
Φ(Φ−1 (γ)) · f (x) = f (x) (by twisted equivariance)
γ · f (x) = f (x)
γ ∈ Stab(f (x))
and so Φ(Stab(x)) < Stab(Φ · x).
To prove (2), consider the inverse simplicial automorphism f −1 : T → S. The
implication in the second line may be inverted by applying f −1 to both sides of the
equation in the second line. For the rest of (2), it suffices to prove the “only if”
direction, because the “if” direction can then be proved using that f −1 is Φ−1 -twisted
equivariant. Assuming γ is loxodromic in S with axis ASγ , consider the line f (ASγ ) ⊂ T .
14
Calculating exactly as above one shows that the equation Φ(Stab(ASγ )) = Stab(f (ASγ ))
holds. We may assume that γ is a generator of the infinite cyclic group Stab(ASγ ),
and so hΦ(γ)i = Φhγi = Stab(f (ASγ )). Since the stabilizer of the line f (ASγ ) is the
infinite cyclic group hΦ(γ)i, it follows that Φ(γ) is loxodromic and its axis ATΦ(γ) is
equal to f (ASγ ).
Free splittings of co-edge number 1. Recall the co-edge number of a free factor
system A of Fn (see for example [HM15, Section 4.1]), which is the minimal number of
edges of G \ H amongst all marked graph pairs (G, H) such that H is a representative
of A. There is a tight relationship between free splittings with a single edge orbit and
free factor systems with co-edge number 1. To be precise:
Fact 2.6. [HM13, Section 4.1] When the Out(Fn )-equivariant function [T ] 7→ A =
A(T ) is restricted to free splittings T with one edge orbit and free factor systems A
with co-edge number 1, the result is a bijection, and hence
Stab[T ] = Stab(A)
whenever T and A correspond under this bijection.
Under the bijection in Fact 2.6, the number of components of A equals the number
of vertex orbits of T which equals 1 or 2. If A = {[A]} has a single component then
rank(A) = n − 1, there is a free factorization Fn = A ∗ B where rank(B) = 1, and
the quotient graph of groups T /Fn is a circle with one edge and one vertex. On the
other hand if A = {[A1 ], [A2 ]} has two components then rank(A1 ) + rank(A2 ) = n,
A1 , A2 can be chosen in their conjugacy classes so that there is a free factorization
Fn = A1 ∗ A2 , and the quotient graph of groups T /Fn is an arc with one edge and
two vertices.
The stabilizer of a free splitting and its automorphic extension Consider
a free splitting α : Fn y T and its stabilizer subgroup Stab[T ] < Out(Fn ). Let
g ] < Aut(Fn ) be the preimage of Stab[T ] under the standard projection homoStab[T
morphism Aut(Fn ) 7→ Out(Fn ), and so we have a short exact sequence
g ] 7→ Stab[T ] 7→ 1
1 7→ Fn ≈ Inn(Fn ) ֒→ Stab[T
g ] y T which
From this setup we shall define in a natural way an action of Stab[T
extends the given free splitting action α : Fn y T . We proceed as follows.
For each Φ ∈ Aut(Fn ) we have the composed action α ◦ Φ : Fn y T . Assuming in
g ], in other words that Φ is a representative of some element
addition that Φ ∈ Stab[T
of Stab[T ], it follows the actions α and α ◦ Φ : Fn y T are equivalent, meaning that
there exists a simplicial automorphism h ∈ Isom(T ) such that h ◦ α(γ) = α(Φ(γ)) ◦ h.
When this equation is rewritten in action notation it simply says that h satisfies Φtwisted equivariance: h(γ · x) = Φ(γ) · h(x) for all γ ∈ Fn , x ∈ S. Suppose conversely
15
that for some Φ ∈ Aut(Fn ) there exists a Φ-twisted equivariant h ∈ Isom(T ). Since
h conjugates the action α : Fn y T to the action α ◦ Φ : Fn y T , it follows that
g ]. This proves the equivalence of (1), (2) and (3) in
φ ∈ Stab[T ] and Φ ∈ Stab[T
the following lemma, which also contains some uniqueness information regarding the
conjugating maps h.
Lemma 2.7. For each free splitting α : Fn y T and each φ ∈ Out(Fn ) the following
are equivalent:
(1) φ ∈ Stab[T ]
(2) For each Φ ∈ Aut(Fn ) representing φ, there exists a Φ-twisted equivariant isomorphism h : T → T .
(3) For some Φ ∈ Aut(Fn ) representing φ, there exists a Φ-twisted isomorphism
h : T → T.
Furthermore,
g ]
(4) If φ satisfies the equivalent conditions (1), (2), (3) then for each Φ ∈ Stab[T
representing φ, the Φ-twisted equivariant isomorphism of hΦ is uniquely determined by Φ, and is denoted
hΦ : T → T
(5) For each γ ∈ Fn with corresponding inner automorphism iγ (δ) = γδγ −1 , the
two maps hiγ : T → T and α(γ) : T → T are equal.
g ].
Remark: In item (5), note that hiγ is defined because iγ ∈ Inn(Fn ) < Stab[T
The uniqueness statement (4) is a special case of a more general uniqueness statement that we will make use of later:
Lemma 2.8. For any two free splittings Fn y S, T and any Φ ∈ Aut(Fn ), there exists
at most one Φ-equivariant simplicial isomorphism h : S 7→ T . In particular, taking
Φ = Id, there exists at most one equivariant simplicial isomorphism h : S 7→ T .
Proof. Suppose that a Φ-twisted equivariant simplicial isomorphism h : S 7→ T exists.
For each γ ∈ Fn , it follows by Φ-twisted equivariance that γ acts loxodromically on S
with axis ASγ if and only if Φ(γ) acts loxodromically on T with axis ATγ = f (AS γ) (by
Lemma 2.5). The map that h induces on the set of axes of loxodromic elements is
therefore uniquely determined. It follows that the restriction of h to the set of vertices
v ∈ T of valence ≥ 3 is uniquely determined by Φ, because v may be expressed in the
form {v} = ASβ ∩ ASγ ∩ ASδ for a certain choice of β, γ, δ ∈ Fn , and hence
{h(v)} = h(ASβ ) ∩ h(ASγ ) ∩ h(ASδ )
= ATΦ(β) ∩ ATΦ(γ) ∩ ATΦ(δ)
Since h is uniquely determined by its restriction to the vertices of valence ≥ 3, it
follows that h is uniquely determined amongst simplicial isomorphisms.
16
Proof of Lemma 2.7. The uniqueness clause (4) is an immediate consequence of
Lemma 2.8. Item (5) follows from the uniqueness clause (4), because for each γ ∈ Fn
the map h = α(γ) clearly satisfies iγ twisted equivariant: α(γ)◦α(δ) = α(iγ (δ))◦α(γ)
for all δ ∈ Fn .
g ] 7→ Isom[T ] defined by α̃(Φ) = hΦ as
Consider now the function α̃ : Stab[T
g ] y T , because for each
given by Lemma 2.7. This defines an action α̃ : Stab[T
g ] both sides of the action equation hΦ ◦ hΨ = hΦΨ clearly satisfy ΦΨΦ, Ψ ∈ Stab[T
twisted equivariance, and hence the equation holds by application of the uniqueness
clause (4) of Lemma 2.7.
The following lemma summarizes this discussion together with the evident geng ], and rewrites the twisted equivariance property
eralization to subgroups of Stab[T
using action notation instead of functional notation.
Lemma 2.9. Associated to each free splitting Fk y T there is a unique isometric
g ] y T which assigns to each Φ ∈ Stab[T
g ] the unique simplicial isomoraction Stab[T
phism T 7→ T as stated in Lemma 2.7, denoted in action notation as x 7→ Φ · x,
satisfying the following:
Twisted equivariance (action notation): For all Φ ∈ Aut(Fn ), γ ∈ Fn , x ∈ T ,
Φ · (γ · x) = Φ(γ) · (Φ · x)
g ] of the action Stab[T
g ]yT
More generally, the restriction to any subgroup K < Stab[T
is the unique isometric action K y T such that satisfies twisted equivariance.
Remark: In the twisted equivariance equation Φ · (γ · x) = Φ(γ) · (Φ · x), the
action dot “·” is used ambiguously for both the action of Fn on T and the action of
Aut(Fn ) on T . The meaning of any particular action dot should be clear by context.
Furthermore, in contexts where the two meanings overlap they will always agree.
For example, Lemma 2.7 (5) says that the action dots always respect the standard
isomorphism Fn ≈ Inn(Fn ) given by δ ≈ iδ .
The next lemma will be a key step of the proof of the loxodromic and WWPD
portions of Theorem F (see remarks after the statement). In brief, given a free
splitting Fn y T and a certain subgroup J ⊂ Fn , the lemma gives a criterion for
verifying that the restriction of a free splitting action Fn y T to a certain group
J ⊂ Fn is nonelementary. To understand the statement of the lemma, the reader
may note that by applying Lemma 2.9, the entire setup in the first paragraph of the
g ].
b < Stab[T
lemma is satisfied for any free splitting Fn y T and any subgroup H
Lemma 2.10. Let Fn y T be a free splitting with vertex set V , and let V nt be the
b < Aut(Fn ) be a subgroup
subset of all v ∈ V such that Stab(v) is nontrivial. Let H
b ∩ Inn(Fn ). Let H
b y V nt be an action such that for each
with normal subgroup J = H
b the map v 7→ Φ · v is Φ-twisted equivariant.
Φ∈H
17
b and if the
If no subgroup StabFn (v) (v ∈ V nt (T )) is fixed by the whole group H,
free group J has rank ≥ 2, then the action J y T is nonelementary.
Remarks. Lemma 2.10 will be applied in Section 2.6 where we prove the oneg ] for which
b < Stab[T
edge case of Theorem F. In that first case we will have a group H
assumptions of the first paragraph of the lemma hold automatically (by Lemma 2.9).
Lemma 2.10 will also be applied in Section 8 where we prove the multi-edge case
of Theorem F. In the place where Lemma 2.10 is applied in that case, the group
g ], and H
b < Aut(Fn ) will not be contained in Stab[T
b will not act on T . Nonetheless
H
b will have a kind of “semi-action” on T which will be enough to give an action
H
b y V nt (T ) satisfying the assumptions of the first paragraph of the lemma.
H
Proof. The action J y T has trivial edge stabilizers because it is the restriction of
a free splitting action. Clearly also each point of V − V nt has trivial stabilizer. It
follows that for each nontrivial α ∈ J, either α is elliptic and its fixed point set
Fix(α) ⊂ T is a single point of V nt , or α is loxodromic with repeller–attractor pair
(∂− α, ∂+ α) ⊂ ∂T × ∂T and StabJ (∂− α) = StabJ (∂+ α) = StabJ {∂− α, ∂+ α} is an
infinite cyclic group.
We claim that there is no point v ∈ V nt (T ) which is fixed by every element
of J. Otherwise that point v is unique, and its stabilizer Bv = StabFn (v) is the
unique nontrivial vertex stabilizer fixed by the action of J, because the bijection
b also fixes Bv ,
v ↔ Bv is J-equivariant (by Lemma 2.5). It follows that each Φ ∈ H
−1
because J = ΦJΦ fixes Φ(Bv ) which is also a nontrivial vertex stabilizer, namely
the stabilizer of Φ · v; by uniqueness of Bv we therefore have Φ(Bv ) = Bv . Since this
b we have contradicted the hypothesis of the lemma, thus proving
holds for all Φ ∈ H,
the claim.
The proof that the action J y T is nonelementary now follows a standard argument. Some γ ∈ J is loxodromic, for otherwise by applying the claim it follows that
J has nontrivial elliptic elements α, β with Fix(α) 6= Fix(β) ∈ T but in that case
γ = αβ is loxodromic (see for example [CM87, Proposition 1.5]). Since rank(J) ≥ 2,
there exists δ ∈ J − StabJ {∂− γ, ∂+ γ}. It follows that γ and δγδ −1 are independent
loxodromic elements of J.
2.6
Case analysis of Theorem F. Proof of the one-edge case.
Adopting the notation of Theorem F, we may assume that H < IAn (Z/3). Choose
B to be a maximal, proper, H-invariant free factor system of Fn , and so H is fully
irreducible relative to the extension B ⊏ {[Fn ]}, meaning that there is no free factor
system A with strict nesting B ⊏ A ⊏ {[Fn } such that A is invariant under any finite
index subgroup of H, equivalently (since H < IAn (Z/3)) such that A is invariant
under H.
The proof of Theorem F breaks into two cases:
The one-edge case: The co-edge number of B equals 1.
18
The multi-edge case: The co-edge number of B is ≥ 2.
Proof of Theorem F in the one-edge case. Assuming that B ⊏ {[Fn ]} is a
one-edge extension, using that H < Stab(B), and applying Fact 2.6, there exists a
free splitting Fn y T with one edge orbit whose vertex stabilizer system forms the
free factor system B, and we have equality of stabilizer subgroups Stab(B) = Stab[T ].
It follows that H < Stab[T ]. Applying Lemma 2.9, consider the resulting action
g ] y T . We show that the restricted action H
b y T satisfies the conclusions
Stab[T
b Since every isometry of a tree is either
of Theorem F (using S = T and N = H).
elliptic or loxodromic, Conclusion (1) of Theorem F holds.
For proving Conclusion (2) of Theorem F we wish to apply Lemma 2.10 to the
b y T , so we must check its hypotheses. All the assumptions in the first
action H
g ].
b < Stab[T
paragraph of Lemma 2.10 hold by applying Lemma 2.9 to the subgroup H
b < Aut(Fn ) does not preserve
Also, by the hypothesis of Theorem F, the subgroup H
any proper, nontrivial free factor of Fn , in particular it does not preserve any nontrivial vertex stabilizer of the free splitting Fn y T . Finally, using the hypotheses of
b is not virtually abelian, it follows that
Theorem F saying that H is abelian and H
b
J has rank ≥ 2, for otherwise the group H < Aut(Fn ) < Out(Fk+1 ) would be solvable and hence virtually abelian (by [HM17c]), a contradiction. The conclusion of
Lemma 2.10 therefore holds, saying that the restricted action J y T is nonelemenb y T is therefore nonelementary, which verifies conclusion (2) of
tary. The action H
Theorem F.
Conclusion (3) of Theorem F says that each loxodromic element of J is a WWPD
b y T , and this is an immediate consequence of the next
element for the action H
lemma (will also be used in the multi-edge case of Theorem F):
Lemma 2.11. If G y T is a group action on a simplicial tree, and if J ⊳ G is a
normal subgroup such that the restricted action J y T has trivial edge stabilizers,
then each loxodromic element of J is a WWPD element of the action G y T .
Proof. We work with the simplicial metric on T that assigns length 1 to each edge.
Given an oriented line A ⊂ T , let the stabilizers of A under the actions of G and of
J be denoted
StabG (A) = {γ ∈ G γ(A) = A}
StabJ (A) = J ∩ StabG (A)
Consider a loxodromic µ ∈ J. Let Aµ ⊂ T denote its axis oriented so that µ translates in the positive direction, with repelling/attracting endpoints ∂− Aµ , ∂+ Aµ ∈ ∂T .
Since J y T has trivial edge stabilizers, the group StabJ (∂− Aµ ) = StabJ (∂+ Aµ ) =
StabJ (∂− Aµ , ∂+ Aµ ) is infinite cyclic.
To prove that µ satisfies WWPD with respect to the action G y T , it suffices
to prove that the ordered pair ∂± Aµ = (∂− Aµ , ∂+ Aµ ) is an isolated point in its orbit
G · (∂± Aµ ) ⊂ ∂T × ∂T − ∆ [HM15, Proposition 2.6]. We may assume that µ is a
19
generator for the infinite cyclic group StabJ (∂± Aµ ); if this is not already true then,
without affecting the WWPD property for µ, we may replace µ with a generator.
Letting ℓµ > 0 denote the integer valued length of a fundamental domain for the
action of µ on Aµ , it follows that any edge path in Aµ of length ℓµ is a fundamental
domain for the StabJ (∂± Aµ ) on Aµ . Choose a subsegment α ⊂ Aµ of length ℓµ + 1.
There is a corresponding neighborhood Uα ⊂ ∂T × ∂T − ∆ consisting of all endpoint
pairs of oriented lines in T containing α as an oriented subsegment. Consider γ ∈ G
such that γ(∂± Aµ ) ⊂ Uα . It follows that µ′ = γ −1 µγ ∈ J has axis Aµ′ = γ(Aµ ) and
that Aµ ∩ Aµ′ contains α, and hence Aµ ∩ Aµ′ has length ≥ ℓµ + 1. Also, the map
γ : Aµ′ → Aµ takes the restricted orientation of the subsegment Aµ ∩ Aµ′ ⊂ Aµ′ to
the restricted orientation on the subsegment γ(Aµ ∩ Aµ′ ) ⊂ Aµ . Let the edges of
the oriented segment Aµ ∩ Aµ′ be parameterized as E0 E1 E2 . . . EJ , J ≥ ℓµ . Since µ
and µ′ both have translation number ℓµ it follows that µ(E0 ) = µ′ (E0 ) = Eℓµ , and
so µ−1 µ′ ∈ StabJ (E). Since J has trivial edge stabilizers it follows that µ = µ′ and
so γ ∈ StabJ {∂− Aµ , ∂+ Aµ } = hµi. And having shown that γ preserves orientation,
it follows that γ(∂± Aµ ) = (∂± Aµ ). This shows that ∂± Aµ is isolated in its orbit
G · ∂± Aµ , being the unique element in the intersection Uα ∩ (G · ∂± Aµ ).
This completes the proof of Theorem F in the one-edge case.
3
Introduction to the multi-edge case of Theorem F.
The proof of Theorem F in the multi-edge case will take up the rest of the paper.
In this section, we give a broad outline of the methods of proof, followed by some
motivation coming from well-known constructions in geometric group theory.
3.1
Outline of the multi-edge case.
We will make heavy use of the theory of abelian subgroups of Out(Fn ) developed by
Feighn and Handel [FH09]. In very brief outline here are the main features of that
theory we will need.
Disintegration groups. ([FH09], and see Section 6.2) Any element of Out(Fn )
has a uniformly bounded power which is rotationless, meaning roughly that each of
various natural finite permutations induced by that element are trivial. Any rotationless φ ∈ Out(Fn ) has a particularly nice relative train track representative called
a CT. For any CT f : G → G there is an associated abelian subgroup D(f ) < Out(Fn )
that contains φ and is called the “disintegration subgroup” of f . The idea of the disintegration group is to first disintegrate or decompose f into pieces, one piece for
each non-fixed stratum Hr , equal to f on Hr and to the identity elsewhere. Then one
re-integrates those pieces to form generators of the group D(f ), by choosing a list
of non-negative exponents, one per non-fixed stratum, and composing the associated
powers of the pieces of f . However, in order for this composition to be continuous and
20
a homotopy equivalence, and for D(f ) to be abelian, the exponents in that list cannot
be chosen independently. Instead two constraints are imposed: the non-fixed strata
are partitioned into a collection of “almost invariant subgraphs” on each of which the
exponent must be constant; and certain linear relations are required amongst strata
that wrap around a common twist path of f .
The following key theorem about disintegration groups lets us study an abelian
subgroup of Out(Fn ) (up to finite index) by working entirely in an appropriate disintegration group:
Disintegration Theorem ([FH09, Theorem 7.2]). For every rotationless abelian
subgroup H < Out(Fn ) there exists φ ∈ H such that for every CT f : G → G
representing φ, the intersection H ∩ D(f ) has finite index in H.
The proof of Theorem F in the multi-edge case. The detailed proof is
carried out in Section 8.2, based on material whose development we soon commence.
Here is a sketch.
By a finite index argument we may assume that the abelian group H is rotationless
abelian. Let B be a maximal, proper, H-invariant free factor system of Fn . Being in
the multi-edge case means that B has co-edge number ≥ 2. From the Disintegration
Theorem we obtain φ ∈ H, and we apply the conclusions of that theorem to a CT
representative f : G → G of φ having a proper core filtration element Gt representing
B. We may assume that H < D(f ), by replacing H with its finite index subgroup
H ∩ D(f ). From the construction of the disintegration group D(f ), any core filtration
element properly contained between Gt and G would represent a free factor system
which is D(f )-invariant, hence H-invariant, contradicting maximality of B. Thus Gt
is the maximal proper core filtration element. Since B has co-edge number ≥ 2,
the top stratum Hu is an EG stratum. By maximality of Gt , every stratum strictly
between Gt and G is either an NEG-linear edge with terminal endpoint attached to
Gt or a zero stratum enveloped by Hu .
The hard work of the proof breaks into two major phases, the first of which is:
Sections 4, 5: Construction and hyperbolicity of S.
The hyperbolic metric space S needed for verifying the conclusions of Theorem F
is constructed in terms of the CT f : G → G (see Section 3.2 for further motivation
of the construction). First we describe S in the simpler case that f has no height u
indivisible Nielsen path. Starting with f : G → G, lift to the universal cover to
e → G.
e Let G
eu−1 ⊂ G
e be the total lift of Gu−1 . Collapse to a point each
obtain f˜ : G
e
e →G
e induces a
component of Gu−1 , obtaining a simplicial tree T . The map f˜ : G
map fT : T → T . Let S be the bi-infinite mapping cylinder of fT , obtained from
T × Z × [0, 1] by identifying (x, n, 1) ∼ (fT (x), n + 1, 0) for each x ∈ T and n ∈ Z.
The construction of S is more complex when Hu is geometric, meaning that f
has a unique (up to inversion) height u indivisible Nielsen path ρ, and ρ is closed,
21
e obtained by
forming a circuit c in G ([BFH00], and see [HM17a]). Each line c̃ ⊂ G
lifting c projects to a line in T called a geometric axis, and the projections to T of
the lifts of ρ in c̃ are fundamental domains for that axis. In the bi-infinite mapping
cylinder as defined above, the portion of the mapping cylinder corresponding to each
geometric axis is a quasiflat, contradicting hyperbolicity. Thus we do not take S to
be the bi-infinite mapping cylinder itself, instead S is constructed from the bi-infinite
mapping cylinder by coning off each geometric axis in T × n for each n ∈ Z, using
one cone point per geometric axis, attaching arcs that connect the cone point to the
endpoints of the fundamental domains of that axis.
In Section 5.5, we use the Mj-Sardar combination theorem [MS12] to prove hyperbolicity of S. The Mj-Sardar theorem, a generalization of the Bestvina-Feighn
Combination Theorem [BF92], requires us to verify a flaring hypothesis. To do this,
in Section 4 we study relative flaring properties of CT f : G → G, specifically: how
f flares relative to the lower filtration element Gu−1 ; and in the geometric case, how
f flares relative to the Nielsen path ρ. Then in Section 5, we study flaring properties
of the induced map fT : T → T (and, in the geometric case, flaring properties of the
induced map obtained by coning off each geometric axis of T ). The required flaring
hypothesis on S itself can then be verified in Section 5.5, allowing application of the
Mj-Sardar theorem to deduce hyperbolicity of S.
The other major phase of the proof is:
Sections 7, 8.1: Use the theory of disintegration groups to obtain an isometric acb y S with appropriate WWPD elements.
tion H
To do this, one uses that there is a number λ > 0 and a homomorphism
P FΛ : D(f ) → Z
such that for each ψ ∈ D(f ), the lamination Λ is an attracting lamination for ψ if
and only if P F (ψ) > 0, in which case the stretch factor is λPF(ψ) . One would like to
think of Hu as an EG stratum for ψ having Perron-Frobenius eigenvalue λPF(ψ) , but
this does not yet make sense because we do not yet have an appropriate topological
representative of ψ on G. We define a subgroup and sub-semigroup
D0 (f ) = Kernel(P FΛ )
D+ (f ) = P FΛ−1[0, ∞)
and we then lift the sequence of inclusions D0 (f ) ⊂ D+ (f ) ⊂ D(f ) < Out(Fn ) to a
sequence of inclusions to
b 0 (f ) ⊂ D
b + (f ) ⊂ D(f
b ) < Aut(Fn )
D
The hard work in Section 7 is to use the theory of disintegration groups to construct
b+ (f ) on T , in which each element Ψ ∈ D
b+ (f )
a natural action of the semigroup D
22
acts on T by stretching each edge by a uniform factor equal to λPF(ψ) , and such that
b+ (f ) the resulting map x 7→ Ψ · x of T is Ψ-twisted equivariant.
for each Ψ ∈ D
b0 we obtain an action on T by twisted equivariant
When restricted to the subgroup D
b0 y T with the one described in
isometries, which allows us to identify the action D
Lemma 2.9.
b+ (f )
Then what we do in Section 8.1 is to suspend the semigroup action of D
on T (and, in the geometric case, on the graph obtained by coning off geometric
b ) y S. By restriction we obtain the required
axes), obtaining an isometric action D(f
b y S.
action H
Finally,
Section 8.2: Put the pieces together and verify the conclusions of Theorem F.
The basis of the WWPD conclusions in the multi-edge case of Theorem F is Lemma 2.11,
which has already played the same role for the one-edge case.
3.2
Motivation: Suspension actions and combination theorems.
b ) y S may be motivated
The construction of the hyperbolic suspension action D(f
by looking at some familiar examples in a somewhat unfamiliar way.
Example: Mapping torus hyperbolization after Thurston. Consider a
pseudo-Anosov homeomorphism f : S → S of a closed, oriented hyperbolic surface
S of genus ≥ 2 with associated deck action π1 S y Se = H2 . The map f uniquely
determines an outer automorphism φ ∈ Out(π1 S). The associated extension group
Γ < Aut(π1 S) is the inverse image of the infinite cyclic group hφi < Out(π1 S) under the natural homomorphism Aut(π1 S) 7→ Out(π1 S). A choice of Φ ∈ Aut(π1 S)
representing φ naturally determines a semidirect product structure Γ ≈ π1 (S) ⋊Φ Z.
The deck transformation action π1 S y H2 extends naturally to an action Γ y H2 ,
whereby if Ψ ∈ Γ projects to ψ = φk then the action of Ψ on H2 , denoted FΨ : H2 →
H2 , is the lift of f k whose induced action on the circle at infinity ∂H2 agrees with the
induced action of γ. Equivalently, FΨ is the unique Ψ-twisted equivariant lift of f k .
Although the deck action π1 S y H2 is by isometries, the extended action Γ y H2 is
not by isometries.
However, there is a way to suspend the action Γ y H2 obtaining an isometric
action Γ y S, as follows. The suspension space S is the bi-infinite mapping cylinder obtained as the quotient of H2 × Z × [0, 1] under the identifications (x, n, 1) ∼
(FΦ (x), n+1, 0). One may check (and we shall do in Section 8.1 in a different context)
that there is an action Γ y S which is generated by letting π1 S act as the deck group
on H2 ≈ H2 × {0} ⊂ S and extending naturally over the rest of S, and by letting
Φ according to the formula Φ · (x, n, t) = (x, n − 1, t). One may also check that the
hyperbolic metrics on the slices H2 × n × 0 extend to a path metric on S, uniquely
up to quasi-isometry, such that the action Γ y S is by isometries.
23
Returning to the land of the familiar, the group Γ may be identified with the fundamental group of the 3-dimensional mapping torus Mf of the surface homeomorphism
f : S → S. By Thurston’s hyperbolization theorem applied to Mf , the suspension
space S may be identified up to Γ-equivariant quasi-isometry with the universal cover
ff ≈ H3 equipped with its deck transformation action by the group Γ ≈ π1 Mf .
M
Example: Mapping torus hyperbolization after [BFH97]. Consider now
an irreducible train track representative f : G → G of a nongeometric, fully irreducible outer automorphism φ ∈ Out(Fn ). Again we have a natural extension group
Γ < Aut(Fn ) of hφi ∈ Out(Fn ), with a semidirect product structure Γ = Fn ⋊Φ Z
determined by a choice of Φ ∈ Aut(Fn ) representing φ.
Unlike in the previous situation where we started with a surface homeomorphism,
here we have started with a non-homeomorphic topological representative f : G → G
e does not extend to an action Γ y G.
e But it does
of φ, and so the deck action Fn y G
extend to an action of the semigroup Γ+ which is the inverse image under Aut(Fn ) 7→
Out(Fn ) of the semigroup hφi+ = hφi i ∈ {0, 1, 2, . . .}i: for each Ψ ∈ Γ+ mapping to
e→G
e is the unique Ψ-twisted equivariant
φi with i ≥ 0, the associated map FΨ : G
e is by isometries, the semigroup action
lift of f i . Although the deck action Fn y G
e
Γ+ y G is not by isometries.
e to an isometric action
But there is a way to suspend the semigroup action Γ+ y G
e → G,
e defined exactly as
Γ y S where S is the bi-infinite mapping cyclinder of FΦ : G
e
above, namely the quotient space S of G × Z × [0, 1] where (x, n, 1) ∼ (FΦ (x), n + 1, 0)
with an appropriate metric. What is done in [BFH97, Theorem 5.1] is to apply
properties of the train track map f : G → G to prove a flaring hypothesis, allowing
one to apply Bestvina–Feighn combination theorem [BF92] to conclude that S is
Gromov hyperbolic, thus proving that Γ is a hyperbolic group.
Flaring methods. Our proof of the multi-edge case will use the combination
theorem of Mj and Sarder [MS12], a generalization of the Bestvina and Feighn combination theorem [BF92]. A common feature of the above examples that is shared
in the construction of this paper is that S is a “metric bundle over R”: there is a
Lipschitz projection map π : S 7→ R such that the minimum distance from π −1 (s) to
π −1 (t) equals s − t for all s, t ∈ R, and each point x ∈ S is contained in the image
of a geodesic section σ : R → S of the map π. In studying the large scale geometry
of S it is important to study quasigeodesic sections σ : R → S and their flaring
properties. In our context it is convenient to translate these concepts into dynamical
systems parlance: each such quasigeodesic section turns out to be a “pseudo-orbit”
of a suspension semiflow on S whose true orbits are the geodesic sections (see the
closing paragraphs of Section 8.2). The combination theorems of [MS12] and its predecessor [BF92] share key hypotheses regarding the asymptotic “flaring” behavior of
such pseudo-orbits (see Definition 4.9). We remark, though, that those combination
theorems hold in much more general settings, e.g. certain kinds of metric bundles over
more general metric base spaces.
24
In Section 4 we study flaring of pseudo-orbits in the context of a relative train
track map, which is then used in Section 5 to extend to further contexts building up
to the suspension space S, its metric bundle S 7→ R, and its pseudo-orbits, allowing
application of the Mj-Sardar combination theorem.
4
Flaring in a top EG stratum
We will assume that the reader is familiar with the basic concepts of Out(Fn ) that
are laid out in Section 3.1 of the first part of this paper [HM15], in particular the
terminology and notation of a CT having a top EG stratum occurring there under
the heading EG properties of CTs. Here is a brief review, in order to fix notations
for what follows in the rest of this paper; see [HM15, Section 3.1] for detailed citations drawn primarily from [FH11]. At present we need no more about CTs than is
described here; when needed in Section 6.1.1 we will give a more thorough review of
CTs.
Notations 4.1. We fix φ ∈ Out(Fn ) and a relative train track representative f : G → G
with associated f -invariant filtration ∅ = G0 ⊂ G1 ⊂ · · · ⊂ Gu = G satisfying the
following:
(1) The top stratum Hu is EG with Perron-Frobenius transition matrix Mu having
top eigenvalue λ > 1. The attracting lamination of φ corresponding to Hu is
denoted Λ+ or just Λ, and its dual repelling lamination is Λ− .
(2) There exists (up to reversal) at most one indivisible periodic Nielsen path ρ
of height u, meaning that ρ is not contained in Gu−1 . In this case ρ and its
inverse ρ̄ are Nielsen paths, and the following hold. It decomposes as ρ = αβ
where α, β are u-legal paths with endpoints at vertices, and (α, β) is the unique
illegal turn in Hu . At least one endpoint of ρ is disjoint from Gu−1 ; we assume
that ρ is oriented with initial point p 6∈ Gu−1 and terminal point q. The initial
and terminal directions are distinct fixed directions in Hu . The stratum Hu is
geometric if and only if ρ exists and is a closed path in which case p = q.
(3) From the matrix Mu one obtains an eigenlength function lPF (σ) defined on all
paths σ in G having endpoints at vertices, and having the following properties:
(a) for each edge E ⊂ G we have lPF (E) = 0 if and only if E ⊂ Gu−1 ;
(b) in general if σ = E1 . . . EK then lPF (σ) = lPF (E1 ) + · · · + lPF (EK );
(c) for all edges E ⊂ G we have lPF (f (E)) = λ lPF(E).
(d) If ρ = αβ exists as in (2) then lPF (α) = lPF (β) = 21 lPF (ρ).
(4) If γ is a path with endpoints at vertices of Hu or a circuit crossing an edge of
Hu then for all sufficiently large i the path f#i (γ) has a splitting with terms in
the set {edges of Hu } ∪ {ρ, ρ̄} ∪ {paths in Gu−1 with endpoints in Hu }.
25
In what follows, there are three cases to consider: the ageometric case where ρ does
not exist; the parageometric case where ρ exists but is not closed; and the geometric
case where ρ exists and is closed. Occasionally we will need to consider these cases
separately. But for the most part we will attempt to smoothly handle all three cases
simultaneously, using the following conventions:
• In the ageometric case, where ρ does not exist, the notation ρ should simply
be ignored (for an exception see the “Notational Convention” following Corollary 4.4);
• In the parageometric case, where ρ exists but is not closed, the notations ρn for
n ≥ 2 should be ignored.
This completes Notations 4.1.
The main result of this section is Proposition 4.10, a uniform flaring property for
the top EG stratum Hu of f . This result generalizes [BFH97, Theorem 5.1 and Lemma
5.6] which covers the special case that f is a nongeometric train track map, that is,
u = 1 and if ρ exists then it is not closed. There are two features which distinguish
our present general situation from that special case. First, our flaring property is
formulated relative to the penultimate filtration element Gu−1 of G. Second, if the
height u indivisible Nielsen path ρ exists then our flaring property is formulated
relative to ρ. Taken together, these two features require very careful isolation and
annihilation of the metric effects of paths in Gu−1 and of the path ρ. This task
is particularly tricky in the geometric case where ρ exists and is closed, in which
situation we must also annihilate the effects of all iterates ρn .
4.1
The path functions Lu and LPF .
Throughout this paper, and following standard terminology in the literature of relative
train track maps, a (finite) path in a graph G is by definition a locally injective
continuous function σ : [0, 1] → G. We often restrict further by requiring that the
endpoints σ(0), σ(1) be vertices in which case σ is a concatenation of edges without
backtracking. A more general concatenation of edges in which locally injectivity may
fail, hence backtracking may occur, will be called an “edge path”.
Generalizing the eigenlength function lPF of Notations 4.1 (3), a path function on
a graph is simply a function l(·) which assigns, to each finite path σ having endpoints
at vertices, a number l(σ) ∈ [0, ∞), subject to two metric-like properties:
• l(·) is symmetric, meaning that l(σ) = l(σ̄).
• l(·) assigns length zero to any trivial path.
We do not require that l(σ) > 0 for all nontrivial paths σ. We also do not generally
require any version of the triangle inequality, but see Lemma 4.7 and the preceding
26
discussion regarding coarse triangle inequalities for the path functions of most interest
to us.
Henceforth in this section we adopt adopt Notations 4.1.
The “big L” path functions Lu , LPF are defined in terms of auxiliary “little l”
path functions lu , lPF by omitting certain terms. The function lPF is already defined
in Notations 4.1 (3).
Consider a path σ in G with edge path decomposition σ = E1 . . . EK . Define
lu (σ) to be the number of terms Ek contained in Hu . Now define Lu (σ) and LPF (σ)
by summing lu (Ek ) and lPF (Ek ) (respectively) only over those terms Ek which do
not occur in any ρ or ρ̄ subpath of σ. Note that for each edge E ⊂ G the “little l”
eigenvector equation lPF (f (E)) = λ lPF (E) implies:
“Big L” eigenvector equation:
LPF (f (E)) = λ LPF (E) for all edges E ⊂ G.
To see why this holds, if E ⊂ Gu−1 both sides are zero. For E ⊂ Hu this follows from
the fact that f (E) is Hu -legal and hence has no ρ or ρ̄ subpath.
The following “quasicomparability” result is an obvious consequence of the fact
that the Perron-Frobenius eigenvector of the transition matrix Mu has positive entries:
Lemma 4.2. There is a constant K = K4.2 (f ) ≥ 1 such that for all edges E of Hu
we have K1 LPF (E) ≤ Lu (E) ≤ K · LPF (E).
The next lemma says in part that ρ and ρ̄ subpaths never overlap.
Lemma 4.3 (Isolation of ρ). For any path σ in G with endpoints, if any, at vertices,
there is a unique decomposition σ = . . . σi . . . σj . . . with the following properties:
(1) Each term σi is an edge or a copy of ρ or ρ̄.
(2) Every subpath of σ which is a copy of ρ or ρ̄ is a term in the decomposition.
Note that we do not assume σ is finite in this statement.
Leaving the proof of Lemma 4.3 for the end of Section 4.1, we give applications.
For a path σ in G with endpoints at vertices, let Kρ (σ) denote the number of ρ or ρ̄
subpaths in the Lemma 4.3 decomposition of σ.
Corollary 4.4 (Some formulas for Lu and LPF ). For any finite path σ with endpoints
at vertices, letting its Lemma 4.3 decomposition be σ = σ1 . . . σB , we have
(
lu (σ) − Kρ (σ) · lu (ρ) in general
Lu (σ) =
lu (σ)
if ρ does not exist
27
and similarly
LPF (σ) =
(
lPF (σ) − Kρ (σ) · lPF (ρ) in general
lPF (σ)
if ρ does not exist
Notational convention when ρ does not exist: Motivated by Corollary 4.4, we
extend the meaning of the notations lu (ρ) and lPF (ρ) to the case that ρ does not exist
by defining
lu (ρ) = lPF (ρ) = 0
Corollary 4.5. For any finite path σ in G with endpoints at vertices, there is a
unique decomposition σ = µ0 ν1 µ1 . . . νA µA with the following properties:
(1) If ρ does not exist then A = 0 and σ = µ0 .
(2) If ρ exists then each νa is an iterate of ρ or ρ̄ (the iteration exponent equals 1
if ρ is not closed), and each µa contains no ρ or ρ̄ subpath.
(3) If ρ exists, and if 1 ≤ a < a + 1 ≤ A − 1, then at least one of the subpaths
µa , µa+1 contains an edge of Hu ; if in addition ρ is closed then each µa subpath
contains an edge of Hu .
Note that in the context of Corollary 4.5 the subpaths µa are nondegenerate for
1 ≤ a ≤ A − 1, but the subpaths µ0 and µA are allowed to be degenerate.
Proof. Everything is an immediate consequence of Lemma 4.3 except perhaps
item (3), which follows from the fact that at least one endpoint of ρ is disjoint
from Gu−1 (Notations 4.1 (2)).
The following is an immediate consequence of Lemma 4.3 and Corollary 4.5, with
Lemma 4.2 applied term-by-term for the “Furthermore” part:
Corollary 4.6 (More formulas for Lu and LPF ). For any finite path σ with endpoints
at vertices, and using the Corollary 4.5 decomposition of σ, we have
Lu (σ) =
A
X
a=1
Lu (µa ) =
A
X
lu (µa ),
LPF (σ) =
a=1
A
X
a=1
LPF (µa ) =
A
X
lPF (µa )
a=1
Furthermore, letting K = K4.2 (f ) ≥ 1, we have
1
LPF (σ) ≤ Lu (σ) ≤ K · LPF (σ)
K
The latter inequality of Corollary 4.6 gives us the freedom to switch back and forth
between LPF and Lu , using Lu in more combinatorial situations and LPF in more
geometric situations through much of the rest of the paper.
28
Coarse triangle inequalities for Lu and LPF . The path functions lu and lPF
satisfy a version of the triangle inequality: for any two paths γ, δ with endpoints at
vertices such that the terminal vertex of γ equals the initial vertex of δ we have
lu [γδ] ≤ lu (γ) + lu (δ),
lPF [γδ] ≤ lPF (γ) + lPF (δ)
where [·] denotes the operation that straightens an edge path rel endpoints to obtain
a path. For Lu and LPF the best we can get are coarse triangle inequalities:
Lemma 4.7. There exists a constant C = C4.7 such that for any finite paths γ, δ in G
with endpoints at vertices such that the terminal point of γ coincides with the initial
point of δ we have
Lu [γδ] ≤ Lu (γ) + Lu (δ) + C
LPF [γδ] ≤ LPF (γ) + LPF (δ) + C
Proof. Consider the Lemma 4.3 decompositions of γ and δ. We may write γ = γ1 γ2 γ3 ,
δ = δ1 δ2 δ3 so that [γδ] = γ1 γ2 δ2 δ3 and so that the following hold: γ1 is a concatenation
of terms of the Lemma 4.3 decomposition of γ, and γ2 is either degenerate or an initial
subpath of a single ρ or ρ̄ term of that decomposition; δ2 is either degenerate or a
terminal subpath of a single ρ or ρ̄ term of the Lemma 4.3 decomposition of δ, and
δ3 is a concatenation of terms of that decomposition. It follows that
Lu [γδ] ≤ Lu (γ1 ) + lu (γ2 ) + lu (δ2 ) + Lu (δ3 )
≤ Lu (γ) + Lu (δ) + 2lu (ρ)
and similarly
LPF [γδ] ≤ LPF (γ) + LPF (δ) + 2lPF (ρ)
Quasi-isometry properties of f . The next lemma describes quasi-isometry properties of the relative train track map f : G → G with respect to the path functions Lu
and LPF .
Lemma 4.8. There exist constants D = D4.8 > 1, E = E4.8 > 0 such that for any
finite path γ in G with endpoints at vertices, the following hold:
Lu (f# (γ)) ≤ D · Lu (γ) + E
LPF (f# (γ)) ≤ D · LPF (γ) + E
and Lu (γ) ≤ D · Lu (f# (γ)) + E
and LPF (γ) ≤ D · LPF (f# (γ)) + E
(1)
(2)
Proof. Once D, E are found satisfying (1), by applying Corollary 4.6 we find D, E
satisfying (2), and by maximizing we obtain D, E satisfying both.
29
If ρ exists let P be its endpoint set, a single point if Hu is geometric and two
points otherwise; and if ρ does not exist let P = ∅. Each point of P is fixed by f . By
elementary homotopy theory (see e.g. Lemma 4.14 (1) in Section 4.6), the map of pairs
f : (G, P ) → (G, P ) has a homotopy inverse f¯ : (G, P ) → (G, P ) in the category of
topological pairs. By homotoping f¯ rel vertices, we may assume that f¯ takes vertices
to vertices and edges to (possibly degenerate) paths. If ρ exists then, using that
ρ = f# (ρ), and applying f¯# to both sides, we obtain f¯# (ρ) = f¯# (f# (ρ)) = ρ.
We prove the following general fact: if g : (G, P ) → (G, P ) is a homotopy equivalence of pairs which takes vertices to vertices and takes edges to paths, and if g# (ρ) = ρ
assuming ρ exists, then there exist D ≥ 1, E ≥ 0 such that for each path δ with endpoints at vertices we have
Lu (g# (δ)) ≤ D · Lu (δ) + E
(3)
This obviously suffices for the first inequality of (1) taking g = f and δ = γ. It also
suffices for the second inequality of (1) taking g = f¯ and δ = γ because in that case
we have g# (ρ) = f¯# (ρ) = f¯# (f# (ρ)) = ρ.
To prove (3), consider the Corollary 4.5 decomposition γ = µ0 ν1 µ1 . . . νA µA .
Applying Corollary 4.6 we have
Lu (γ) = lu (µ0 ) + · · · + lu (µA )
We also have
g# (γ) = [g# (µ0 ) ν1 g# (µ1 ) . . . νA g# (µA )]
By inductive application of Lemma 4.7 combined with the fact that each Lu (νa ) = 0,
it follows that
Lu (g# (γ)) ≤ Lu (g# (µ0 )) + · · · + Lu (g# (µA )) + 2 A C4.7
≤ lu (g# (µ0 )) + · · · + lu (g# (µA )) + 2 A C4.7
≤ D ′ (lu (µ0 ) + · · · + lu (µA )) + 2 A C4.7
where D ′ is the maximum number of Hu edges crossed by g(E) for any edge E ⊂ G.
If ρ does not exist then A = 0 and we are done. If ρ exists then, by Corollary 4.6 (3),
if 1 ≤ a ≤ a + 1 ≤ A − 1 then at least one of lu (µa ), lu (µa+1 ) is ≥ 1. It follows that
A ≤ 2(lu (µ0 ) + · · · + lu (µA )) + 3 and therefore
Lu (g# (γ)) ≤ (D ′ + 4 C4.7 )Lu (γ) + 6 C4.7
Proof of Lemma 4.3. The lemma is equivalent to saying that two distinct ρ or
e this is
ρ̄ subpaths of σ cannot overlap in an edge. Lifting to the universal cover G
e if µ̃ and µ̃′ are subpaths of σ̃ each of
equivalent to saying that for any path σ̃ ⊂ G,
30
which projects to either ρ or ρ̄, and if µ̃ and µ̃′ have at least one edge in common,
then µ̃ = µ̃′ .
As a first case, assume that µ̃ and µ̃′ both project to ρ or both project to ρ̄; up
to reversal we may assume the former. Consider the decomposition ρ = αβ where
α, β are legal and the turn (ᾱ, β) is illegal and contained in Hu . There are induced
decompositions µ̃ = α̃β̃ and µ̃′ = α̃′ β̃ ′ . If the intersection µ̃ ∩ µ̃′ contains the height
u illegal turn of µ̃ or µ̃′ then µ̃ = µ̃′ so assume that it does not. After interchanging
µ̃ and µ̃′ we may assume that µ̃ ∩ µ̃′ ⊂ β̃. Projecting down to ρ we see that β = β1 β2
and α = α1 α2 where β2 = α1 is the projection of µ̃ ∩ µ̃′ . This implies the initial
directions of ᾱ1 and of β2 (which are respectively equal to the terminal and initial
directions of ρ) are fixed, distinct, and in Hu , and so for all k ≥ 0 the initial directions
of f#k (ᾱ1 ) and of f#k (β2 ) are also fixed, distinct, and in Hu . Using the eigenlength
function lPF (Notations 4.1 (3)) we have lPF (α) = lPF (β) and lPF (α1 ) = lPF (β2 ) and
so lPF (α2 ) = lPF (β1 ). Thus when the path f#k (α1 )f#k (α2 )f#k (β1 )f#k (β2 ) is tightened to
form f#k (ρ) = ρ for large k, the subpaths f#k (α2 ) and f#k (β1 ) cancel each other out,
and so the concatenation f#k (α1 )f#k (β2 ) tightens to ρ. But this contradicts that the
two terms of the concatenation are u-legal and that the turn {f#k (ᾱ1 ), f#k (β2 )} taken
at the concatenation point is also u-legal as seen above.
The remaining case is that orientations on the projections of µ̃ and µ̃′ do not
agree. In this case there is either an initial or terminal subpath of ρ that is its own
inverse, which is impossible.
4.2
Flaring of the path functions Lu and LPF
In this section we continue to adopt Notations 4.1.
For each path function l on G and each η ≥ 0 we define a relation β ∼ γ on paths
β, γ in G with endpoints at vertices: this relation means that there exist paths α, ω
with endpoints at vertices such that γ = [αβω] and such that l(α), l(γ) ≤ η. Note
that this relation is symmetric because β = [ᾱγ ω̄]. When we need to emphasize the
η
dependence of the relation on l and η we will write more formally β ∼l γ.
Definition 4.9. A path function l on G is said to satisfy the flaring condition with
respect to f : G → G if for each µ > 1, η ≥ 0 there exist integers R ≥ 1 and
A ≥ 0 such that for any sequence of paths β−R , β−R+1 , . . . , β0 , . . . , βR−1 , βR in G with
endpoints at vertices, the flaring inequality
µ · l(β0 ) ≤ max{l(β−R ), l(βR )}
holds if the following two properties hold:
Threshold property: l(β0 ) ≥ A;
Pseudo-orbit property: βr
η
∼
l
f# (βr−1 ) for each −R < r ≤ R.
31
Remark. These two properties are translations into our present context of two
requirements in the original treatment of flaring by Bestvina and Feighn [BF92]: the
threshold property corresponds to the requirement of large “girth”; and the “pseudoorbit property” corresponds to the “ρ-thin” requirement.
Proposition 4.10 (General Flaring). The path functions Lu and LPF each satisfy
the flaring condition with respect to f : G → G, with constants R = R4.10 (µ, η) and
A = A4.10 (µ, η).
Much of the work is to prove the following, which is a version of Proposition 4.10
under the special situation of Definition 4.9 where l = LPF and η = 0. Note that
η = 0 implies that βr = f# (βr−1 ), and so the choice of γ = β−R determines the whole
sequence up to βR .
Lemma 4.11 (Special Flaring for Lu ). For any ν > 1 there exist positive integers
N ≥ 1 and A = A4.11 ≥ 0 so that if γ is a finite path in G with endpoints at vertices
and if Lu (f#N (γ)) ≥ A then
νLu (f#N (γ)) ≤ max{Lu (γ), Lu (f#2N (γ))}
The proof of this special flaring condition, which takes up Sections 4.3–4.5, is
similar in spirit to [BFH97, Theorem 5.1], and is based on two special cases: a
“negative flaring” result, Lemma 4.12 in Section 4.3; and a “positive flaring” result,
Lemma 4.13 (4) in Section 4.4. These two special cases are united in Section 4.5,
using the uniform splitting property given in Lemma 3.1 of [HM15], in order to prove
the Special Flaring Condition.
Before embarking on all that, we apply Lemma 4.11 to:
Proof of Proposition 4.10. Applying Corollary 4.6 it is easy to see that Lu satisfies
a flaring condition if and only if LPF satisfies a flaring condition. Thus it suffices to
assume that the special flaring condition for Lu holds, and use it to prove the general
flaring condition for Lu . The idea of the proof is standard: the exponential growth
η
given by Proposition 4.10 swamps the constant error represented by the relation L∼
u
which we will denote in shorthand as ∼. The hard work is to carefully keep track of
various constants and other notations.
η
Fixing µ > 1 and η ≥ 0, consider the relation ∼ given formally as L∼ . Fix an
u
integer R ≥ 1 whose value is to be determined — once an application of Lemma 4.11
has been arranged, R will be set equal to the constant N of that lemma.
Consider a sequence of paths
β−R , β−R+1 , . . . , β0 , . . . , βR−1 , βR
in G with endpoints at vertices, such that we have
βr ∼ f# (βr−1 ) for
32
−R < r ≤R
Choose a vertex V ∈ G which is f -periodic, say f K (V ) = V for some integer K ≥ 1.
We may assume that β0 has endpoints at V : if not then there are paths α0′ , ω0′ of
uniformly bounded length such that β0′ = [α0′ β0 ω0′ ] has endpoints at V , and replacing
β0 with β0′ reduces the flaring property as stated to the flaring property with uniform
changes to the constants µ, η, and A.
We choose several paths with endpoints at vertices, denoted αr , ωr , γr , αr′ , ωr′ , as
follows. First, using that f# (βr−1 ) ∼ βr , there exist αr , ωr so that Lu (αr ), Lu (ωr ) ≤ η
and βr = [αr f# (βr−1 ) ωr ] and hence f# (βr−1 ) = [ᾱr βr ω̄r ]. Next, anticipating an
application of Lemma 4.11, choose γ−R so that f#R (γ−R ) = β0 , and then for −R ≤
r ≤ R let γr = f#r+R (γ−R ), hence γ0 = β0 . Finally, choose αr′ , ωr′ so that γr = [αr′ βr ωr′ ]
and hence βr = [ᾱr′ γr ω̄r′ ] (in particular α0′ , ω0′ are trivial paths). We also require αr′ , ωr′
to be chosen so that if r > −R then we have
′
′
f# (βr−1 ) = f# (ᾱr−1
γr−1 ω̄r−1
)
′
′
= [f# (ᾱr−1
) γr f# (ω̄r−1
)]
′
′
′
′
= [f# (ᾱr−1 ) αr βr ωr f# (ω̄r−1
)]
′
′
)]
= [f# (ᾱr−1
) αr′ ] βr [ωr′ f# (ω̄r−1
{z
}
{z
}
|
|
= ᾱr
= ω̄r
from which we record the following identities:
(∗)
′
ᾱr = [f# (ᾱr−1
)αr′ ]
′
ω̄r = [ωr′ f# (ω̄r−1
)]
To see why these choices of γr , αr′ , ωr′ are possible, we work with a lift to the universal
e → G.
e First choose any lift β̃−R . By induction for −R < r ≤ R the lifts
cover f˜ : G
α̃r , ω̃r , β̃r are determined by the equation β̃r = [α̃r f˜# (βr−1 )ω̃r ]. Using f -periodicity
of V , the f˜R -preimages of the initial and terminal vertices of β̃0 contain vertices of
e amongst which we choose the initial and terminal vertices of γ̃−R , respectively; it
G
follows that f˜#R (γ̃−R ) = β̃0 . Then define γ̃r = f˜#r+R (γ̃−R ), define α̃r′ to be the path
from the initial vertex of γ̃r to the initial vertex of β̃r , and define ω̃r′ to be the path
e
from the terminal vertex of β̃r to the terminal vertex of γ̃r . Projecting down from G
to G we obtain the paths γr , αr′ , ωr′ . The identities (∗) follow from the evident fact
that the paths α̃r′ , α̃r may be concatenated to form a path α̃r′ α̃r , that the two paths
′
′
α̃r′ α̃r and f# (α̃r−1
) have the same initial endpoint as f# (γr−1
) = γr , and that those
two paths have the same terminal endpoint as the initial endpoint of f˜# (β̃r−1 ); and
similarly for the ω’s.
We need new upper bounds on the quantities Lu (αr′ ) and Lu (ωr′ ) which represent
the difference between Lu (βr ) and Lu (γr ). These bounds will be expressed in terms
of the known upper bound η on Lu (αr ) and Lu (ωr ), and will be derived by applying
applying Lemmas 4.7 and 4.8 inductively, starting from Lu (α0′ ) = Lu (ω0′ ) = 0 and
applying (∗) in the induction step. These new upper bounds will then be used to
derive an expression for the threshold constant A4.10 , which will be so large so that
the differences Lu (αr′ ), Lu (ωr′ ) become insignificant.
33
The new upper bounds have the form
Lu (αr′ ), Lu (ωr′ ) ≤ Fr (C, D, E, η) for −R ≤ r ≤ R
where for each r the expression Fr (C, D, E, η) represents a certain polynomial with
integer coefficients in the variables C, D, E, η and where we substitute
C = C4.7 ,
D = D4.8 ,
E = E4.8
The proofs of these inequalities are almost identical for α′ and for ω ′ ; we carry out
the proof for the latter. To start the induction, since γ0 = β0 the path ω0′ degenerates
to a point and so Lu (ω0′ ) = 0 ≡ F0 (C, D, E, η). Inducting in the forward direction on
the interval 1 ≤ r ≤ R, and using from (∗) that
′
ωr′ = [ω̄r f# (ωr−1
)]
we have
′
Lu (ωr′ ) ≤ Lu (ωr ) + Lu (f# (ωr−1
)) + C
≤ η + D · Fr−1 (C, D, E, η) + E + C
≡ Fr (C, D, E, η)
Inducting in the backward direction on the interval −R ≤ r ≤ −1, and using from
(∗) that
′
f# (ωr′ ) = [ωr+1 ωr+1
]
we have
Lu (ωr′ ) ≤ D · Lu (f# (ωr′ )) + E
′
≤ D · (Lu (ωr+1 ) + Lu (ωr+1
) + C) + E
≤ D(η + Fr+1 (C, D, E, η) + C) + E
≡ Fr (C, D, E, η)
To summarize, we have proved
′
′
Lu (α−R
), Lu (ω−R
) ≤ F−R (C, D, E, η)
′
′
Lu (αR ), Lu (ωR ) ≤ FR (C, D, E, η)
From this it follows that
′
′
Lu (γ−R ) ≤ Lu (α−R
) + Lu (β−R ) + Lu (ω−R
) + 2C
≤ Lu (β−R ) + 2F−R (C, D, E, η) + 2C
34
and similarly
Lu (γR ) ≤ Lu (βR ) + 2FR (C, D, E, η) + 2C
Let M = 2 max{F−R (C, D, E, η) + FR (C, D, E, η)} + 2C, which we use below in
setting the value of the threshold constant A4.10 .
Now apply Lemma 4.11, the special flaring condition for Lu , with the constant
ν = 2µ − 1 > 1, to obtain integers N ≥ 1 and A4.11 . Setting R = N, from the
threshold requirement Lu (β0 ) ≥ A4.11 it follows that
νLu (β0 ) = νLu (γ0 )
≤ max{Lu (γ−R ), Lu (γR )}
≤ max{Lu (β−R ), Lu (βR )} + M
With the additional threshold requirement Lu (β0 ) ≥
2M
ν−1
we have:
νLu (β0 ) ≤ max{Lu (β−R ), Lu (βR )} +
µLu (β0 ) =
ν +1
Lu (β0 ) ≤ max{Lu (β−R ), Lu (βR )}
2
ν−1
Lu (β0 )
2
Thus we have proved the (general) flaring condition for Lu given any µ ≥ 1, η ≥ 0,
2M
}.
using R4.10 = N and A4.10 = max{A4.11 , ν−1
4.3
Negative flaring of Lu
We continue to adopt Notations 4.1.
In the context of Lemma 4.11 on Special Flaring for Lu , the next lemma establishes
flaring in the “negative direction”.
For any path σ ⊂ G, define lu− (σ) to be the maximum of lu (τ ) over all paths τ in G
such that that τ is a subpath both of σ and of some leaf ℓ of Λ− realized in G. In this
definition we may always assume that ℓ is a generic leaf of Λ− , because every finite
subpath of every leaf is a subpath of every generic leaf. Notice that if σ is already a
subpath of a leaf of Λ− then lu− (σ) = lu (σ).
Lemma 4.12. There exists L4.12 ≥ 1 such that for each L ≥ L4.12 and each a > 0
there exists an integer N ≥ 1, so that the following holds: for each n ≥ N, for each
finite subpath τ of a generic leaf ℓ of Λ− realized in G such that τ has endpoints at
vertices and such that lu− (τ ) = lu (τ ) = L, and for each finite path σ in G, if f#n (σ) = τ
then lu− (σ) ≥ aL.
Intuitively, at least in the nongeometric case (see the proof of Proposition 6.0.8
of [BFH00], page 609), the point of this lemma is that in any leaf of Λ− , the illegal
turns have a uniform density with respect to the path function lu , and those turns
die off at an exponential rate under iteration of f# .
35
Proof. For the proof, a height u illegal turn in a path σ is said to be out of play if it
is the illegal turn in some ρ or ρ̄ subpath of σ, otherwise that turn is in play.
The generic leaf ℓ of Λ− , as realized in G, is birecurrent, is not periodic, and is
not weakly attracted to Λ+ under iteration of f# . It follows that ℓ does not split as
a concatenation of u-legal paths and copies of ρ and ρ̄. By applying Lemma 4.3 it
follows that ℓ has an illegal turn that is in play. In addition, ℓ is quasiperiodic with
respect to edges of Hu [BFH00, Lemma 3.1.8]: for any integer L > 0 there exists an
integer L′ > 0 such that for any finite subpaths α, β of ℓ, if lu (α) = L and lu (β) ≥ L′
then β contains a subpath which is a copy of α. It follows that there exists an integer
L4.12 > 0 so that if τ is a subpath of ℓ such that lu (τ ) ≥ L4.12 , then τ has at least
three height u illegal turns that are in play.
Arguing by contradiction, if the lemma fails using the value of L4.12 just given
then there exist L ≥ L4.12 , a > 0, positive integers ni → +∞, finite paths τi in G
with endpoints at vertices, and paths σi in G, such that τi is a subpath of ℓ, and
lu (τi ) = L, and f#ni (σi ) = τi , and lu− (σi ) < aL. We derive contradictions in separate
cases.
+
Case 1: lu (σi ) has an upper bound B. Decompose σi as σi = ǫ−
i ηi ǫi where
ǫ±
i are each either partial edges or trivial, and where ηi is a path with endpoints at
vertices. By [HM17a, Lemma 1.53] there exists a positive integer d depending only
on B such that for each i the path f#d (ηi ) has a splitting into terms each of which
is either a single edge in Hu , a copy of ρ or ρ̄, or a subpath of Gu−1 , and therefore
each height u illegal turn in the interior of f#d (ηi ) is out of play. For each i such
that ni ≥ d the paths f#ni (ǫ±
i ) are each u-legal, and each height u illegal turn of
ni
ni −d
ni +
d
n
f# (ηi ) = f# (f# (ηi )) is out of play. Since τi is obtained from f#ni (ǫ−
i )f# (ηi )f# (ǫi )
by tightening, at most two illegal turns of τi are in play, a contradiction to our choice
of L4.12 .
Case 2: lu (σi ) has no upper bound. Consider a line M in G which is a weak
limit of a subsequence σim such that M crosses at least one edge of Hu . We apply to
each such M the weak attraction results of [HM17c] as follows. If Hu is non-geometric
then there are two options for M: either the closure of M contains Λ− ; or M is weakly
attracted to Λ+ [HM17c, Lemma 2.18]. If Hu is geometric — and so a height u closed
indivisible Nielsen path ρ exists — then there is a third option, namely that M is a
bi-infinite iterate of ρ or ρ̄ [HM17c, Lemma 2.19].
Since no σi contains a subpath of a leaf of Λ− that crosses aL edges of Hu , neither
does M. This shows that the closure of M does not contain Λ− . If M were weakly
attracted to Λ+ then for any K > 0 there would exist l > 0 such that f#l (M), and
hence f#l (σim ) for all sufficiently large m, contains a u-legal subpath that crosses
2K + 1 edges of Hu . By [BFH00] Lemma 4.2.2 we can choose K so that f#l (σim )
splits at the endpoints of the middle Hu edge of that subpath, and hence the number
n −l
of edges of Hu crossed by τim = f#im (f#l (σim )) goes to infinity with m, contradicting
that lu (τi ) = L.
We have shown that each of the first two options leads to a contradiction for any
36
choice of M as above. This concludes the proof if Hu is nongeometric.
It remains to show that if Hu is geometric then the third option can also be avoided
by careful choice of M, and hence the desired contradiction is achieved. That is, we
show that there exists a weak limit M of a subsequence of σi such that M contains at
least one edge of Hu and M is not a bi-infinite iterate of the closed path ρ or ρ̄. This
may be done by setting up an application of [HM17c, Lemma 1.11], but it is just as
simple to give a direct proof. Lift σi to the universal cover of G and write it as an edge
ei1 E
ei2 . . . E
eiJ ⊂ G;
e the first and last terms are allowed to be partial edges.
path σ̃i = E
i
Let b equal twice the number of edges in ρ. Given j ∈ {1 + b, . . . , Ji − b}, we say that
eij is well covered if E
ei,j−b, E
ei,j+b are full edges and there is a periodic line ρ̃ij ⊂ G
e
E
ei,j−b . . . E
eij . . . E
ei,j+b as a subpath. Since
that projects to ρ or to ρ̄ and that contains E
the intersection of distinct periodic lines cannot contain two fundamental domains of
eij and E
ei,j+1 are well covered
both lines, ρ̃ij is unique if it exists. Moreover, if both E
eij is well covered then we can inductively move
then ρ̃ij = ρ̃i,j+1. It follows that if E
forward and backward past other well covered edges of σ̃i all in the same lift of ρ,
until either encountering an edge that is not well covered, or encountering initial and
terminal subsegments of σ̃i of uniform length. After passing to a subsequence, one of
the following is therefore satisfied:
(1) There exists a sequence of integers Ki such that 1 < Ki < Ji , and Ki → ∞,
eiK ⊂ H
e u is not well covered.
and Ji − Ki → ∞, and such that E
i
(2) σi = αi ρpi βi where the number of edges crossed by αi and βi is bounded independently of i and pi → ∞.
If subcase (1) holds then the existence of a weak limit that crosses an edge of Hu
and is not a bi-infinite iterate of ρ or ρ̄ follows immediately. If subcase (2) holds, τi
is obtained from f#ni (αi )ρpi f#ni (βi ) by tightening. Since the number of illegal turns
of height u in the first and last terms is uniformly bounded, the number of edges
that are cancelled during the tightening is uniformly bounded, and it follows that τi
contains ρqi as a subpath where |qi | → ∞, a contradiction to the fact that ρ is not a
leaf of Λ− .
4.4
Positive flaring and other properties of Lu
In this section we continue to adopt Notations 4.1.
In the context of Lemma 4.11 on Special Flaring for Lu , item (4) of Lemma 4.13
establishes flaring in the “positive direction”. Lemma 4.13 also includes several useful
metrical/dynamical properties of Lu , which will be used in Section 4.5 where we tie
together the negative and positive flaring results to prove Lemma 4.11.
The lemma and its applications make use of the path map f## and its properties as
laid out in [HM17a] Section 1.1.6, particularly the “## Lemma” [HM17a, Lemma 1.6].
Roughly speaking, the path f## (σ) is defined for any finite path σ in G as follows:
37
for any finite path τ that contains σ the straightened image f# (τ ) contains a subpath
of f# (σ) that is obtained by deleting initial and terminal subpaths of f# (σ) that are
no longer than the bounded cancellation constant of f ; the path f## (σ) is the longest
such subpath of f# (σ) which survives this deletion operation for all choices of τ .
Lemma 4.13. The following conditions hold:
(1) For each path σ with endpoints at vertices and any decomposition into subpaths
σ = σ0 σ1 . . . σK with endpoints at vertices we have
PK
PK
0 Lu (σk )
0 Lu (σk ) − K lu (ρ) ≤ Lu (σ) ≤
and similarly with LPF and lPF in place of Lu and lu respectively.
(2) For all positive integers d there exist B = B4.13 (d) > 0 and b = b4.13 (d) > 0
so that for all finite paths σ with endpoints at vertices, if Lu (σ) ≥ B then
d
Lu (f##
(σ)) ≥ b Lu (σ).
(3) There exist constants A > 0 and κ ≥ 1 so that for all subpaths τ of leaves of
Λ− in G, if lu (τ ) ≥ A then
lu (τ ) ≤ Lu (τ ) ≤ κ lu (τ )
(4) There exists a positive integer A = A4.13 and a constant 0 < R = R4.13 < 1, so
that if a path α ⊂ G splits as a concatenation of u-legal paths and copies of ρ
or ρ̄, and if Lu (α) ≥ A, then for all m ≥ 1 we have
m
Lu (f##
(α)) ≥ R λm/2 Lu (α)
where λ is the expansion factor for f and Λ+ .
m
Remark: The notation f##
is disambiguated by requiring that the exponent binds
m
more tightly than the ##-operator, hence f##
= (f m )## — this is how items (2)
and (4) are applied in what follows. Note that this makes the statements of (2)
m
and (4) weaker than if f##
were intepreted as (f## )m , because (f## )m (α) is a subpath
of (f m )## (α).
Proof. We prove (1) for Lu when K = 1; the proof for general K follows by an easy
induction, and the statement for LPF is proved in the exact same manner. If there are
non-trivial decompositions σ0 = σ0′ σ0′′ and σ1 = σ1′ σ1′′ such that σ0′′ σ1′ is a copy of ρ or
ρ̄ then Hu -edges in σ0′′ σ1′ contribute to Lu (σ0 ) + Lu (σ1 ) but not to Lu (σ), and all other
Hu -edges contribute to Lu (σ0 )+Lu (σ1 ) if and only if they contribute to Lu (σ). In this
case Lu (σ) = Lu (σ0 ) + Lu (σ1 ) − lu (ρ). If there are no such non-trivial decompositions
then concatenating the decompositions of σ0 and σ1 given by Lemma 4.3 produces
the decomposition of σ given by Lemma 4.3 and Lu (σ) = Lu (σ0 ) + Lu (σ1 ). This
completes the proof of (1).
38
d
We next prove (2). Fix the integer d ≥ 1. The path f##
(σ) is obtained from
d
f# (σ) by removing initial and terminal segments that cross a number of edges that is
bounded above, independently of σ, by the bounded cancellation constant of f d . It fold
lows that Lu (f##
(σ))−Lu (f#d (σ)) is bounded above independently of σ, so it suffices
to find B, b > 0 depending only on d so that if Lu (σ) > B then Lu (f#d (σ)) > bLu (σ).
Applying Lemma 4.8 (1) we obtain D > 1, E > 0 so that Lu (γ) ≤ D Lu (f# (γ)) + E
for all finite paths γ with endpoints at vertices, from which it follows by induction
that
1
1
1
d
Lu (f# (γ)) ≥ d Lu (γ) − E
+···+ d
D
D
D
E
1
≥ d Lu (γ) −
D
D−1
and so if Lu (γ) ≥
2D d E
D−1
then
Lu (f#d (γ)) ≥
1
1
1
Lu (γ) −
Lu (γ) =
Lu (γ)
d
d
D
2D
2D d
The first inequality of (3) follows immediately from Corollary 4.4, for any subpath
τ of a leaf of Λ− . To prepare for proving the second inequality, given a positive integer
C let ΣC be the set of paths in G that do not contain a subpath of the form ρǫC where
ǫ = ±1. Since ρ has an endpoint that is not contained in Gu−1 [HM17a, Fact 1.43],
each maximal subpath of σ of the form ρk or ρ̄k that is neither initial nor terminal in
σ is adjacent in σ to an Hu -edge that contributes to Lu (σ). Applying Corollary 4.4
it follows that
(∗) For any fixed C, amongst those paths σ ∈ ΣC for which Lu (σ) > 0, the ratio
lu (σ)/Lu (σ) has a positive lower bound that is independent of σ.
For proving (3), the key observation is that there exists a positive integer C so that
each subpath τ of Λ− is contained in ΣC — this is equivalent to saying that ρ∞ is not
a weak limit of lines in Λ− , which is equivalent to saying that ρ∞ is not a leaf of Λ− ,
which follows from Lemma 3.1.15 of [BFH00] and the fact that Λ− 6= ρ∞ . Item (3)
therefore follows by combining this observation with (∗).
For proving (4), we focus primarily on an analogue for LPF , connecting it with Lu
version stated in (4) by applying Corollary 4.6. From the assumption on a splitting
of α we have
LPF (f#m (α)) = λm LPF (α)
m
We shall show how to replace f#m by f##
, at the expense of replacing λ by its square
root, and of requiring LPF (α) to exceed some threshold constant. To be precise, we
have:
39
Claim: There exists A′ ≥ 0 such that if LPF (α) ≥ A′ then for all m ≥ 0 we have
m
LPF (f##
(α)) ≥ λm/2 LPF (α)
This suffices to prove (4), because if Lu (α) ≥ K4.2 (f ) · A′ = A then from Corollary 4.6
m
it follows that LPF (α) ≥ A′ , from which using the Claim we obtain LPF (f##
(α)) ≥
m/2
λ LPF (α), and then by two more applications of Corollary 4.6 we obtain
m
Lu (f##
(α)) ≥
1
1
1
LPF (α) ≥
λm/2 LPF (α) ≥
λm/2 Lu (α)
K4.2 (f )
K4.2 (f )
(K4.2 (f ))2
To prove the claim, the case m = 0 is evident, so suppose by induction that
m−1
LPF (f##
(α)) ≥ λ(m−1)/2 LPF (α)
m−1
Since f##
(α) is a subpath of f#m−1 (α), and since the latter splits into terms each of
m−1
which is an edge of Hu , a copy of ρ or ρ̄, or a path in Gu−1 , it follows that f##
(α)
may be deconcatenated in the form
m−1
f##
(α) = ζ α̂ω
such that α̂ splits into terms exactly as above, and such that either ζ, ω are both
trivial, or ρ exists and ζ, ω are both proper subpaths of ρ or ρ̄; it follows that
lPF (ζ), lPF(ω) ≤ lPF (ρ). Applying item (1) it follows that
m−1
LPF (f##
(α)) ≤ LPF (α̂) + 2lPF (ρ)
LPF (α̂) ≥ λ(m−1)/2 LPF (α) − 2lPF (ρ)
Using the splitting of α̂ we obtain
LPF (f# (α̂)) = λLPF (α̂) ≥ λ(m+1)/2 LPF (α) − 2 λ lPF(ρ)
The path f## (α̂) is obtained from f# (α̂) by truncating initial and terminal segments
no longer than the bounded cancellation constant of f , and since this is a finite number
of paths their LPF -values have a finite upper bound C2 , so by applying item (1) it
follows that
LPF (f## (α̂)) ≥ λ(m+1)/2 LPF (α) − 2 λ lPF(ρ) − 2C2
m−1
Now we apply the ## Lemma: since α̂ is a subpath of f##
(α) it follows that f## (α̂)
m−1
m
is a subpath of f## (f## (α)) [HM17a, Lemma 1.6 (3)], which is a subpath of f##
(α)
m
[HM17a, Lemma 1.6 (4)]. Thus we have f## (α) = ηf## (α̂)θ for some paths η, θ, and
hence by item (1) we have
m
LPF (f##
(α)) ≥ λ(m+1)/2 LPF (α) − 2 λ lPF(ρ) − 2C2 − 2 lPF(ρ)
40
To complete the induction we show that with appropriate threshold constant the
quantity on the right is ≥ λm/2 LPF (α), equivalently
(m+1)/2
m/2
λ
LPF (α) ≥ λ
LPF (α) + 2 (λ + 1) lPF(ρ) + 2C2
{z
}
|
C1
λ1/2 ≥ 1 +
λm/2
C1
LPF (α)
Since λ > 1 is suffices to show
λ1/2 ≥ 1 +
C1
LPF (α)
⇐⇒
LPF (α) ≥
C1
λ1/2 − 1
Taking the threshold constant to be
A′ =
C1
λ1/2 − 1
the induction is complete.
4.5
Proof of Lemma 4.11: the Special Flaring Condition
for Lu
Once this proof is complete, the proof of the General Flaring Condition stated in
Proposition 4.10 will also be complete, as shown in Section 4.2.
If the Special Flaring Condition for Lu fails then there exists a sequence nk → ∞,
and there exist paths γk ⊂ G with endpoints at vertices, such that Lu (f#nk (γk )) → ∞
as k → ∞, and such that
(∗)
ν Lu (f#nk (γk )) ≥ max{Lu (γk ), Lu (f#2nk (γk ))}
Assuming this, we argue to a contradiction.
Consider the integer L4.12 ≥ 1 satisfying the conclusions of Lemma 4.12. By
Lemma 4.13 (3) there is an integer L2 so that if µ is a subpath of Λ− that crosses
≥ L2 edges of Hu then Lu (µ) ≥ 1. Let L1 = max{L4.12 , L2 }. Choose an integer d ≥ 1
satisfying the conclusion of [HM15, Lemma 3.1], the “uniform splitting lemma”, with
respect to the constant L1 . This conclusion says that for any finite path σ in G with
d
endpoints at vertices of Hu , if ℓ−
u (σ) < L1 then the path f# (σ) splits into terms each
of which is u-legal or a copy of ρ or ρ̄. In this context we shall refer to d as the
“uniform splitting exponent”.
Let {µik } be a maximal collection of subpaths of f#nk (γk ) with endpoints at vertices
that have disjoint interiors, that are subpaths of Λ− , and that cross ≥ L1 edges of Hu .
The complementary subpaths {νjk } of f#nk (γk ) all satisfy lu− (νjk ) < L1 and all have
endpoints at vertices as well.
41
Our first claim is that
P
Lu (µik )
=0
n
k→∞ Lu (f#k (γk ))
(i)
lim
i
If not, then after passing to a subsequence we may assume that
X
Lu (µik ) > ǫ1 Lu (f#nk (γk ))
i
for some ǫ1 > 0 and all k. Choose subpaths σik of γk with disjoint interiors such
that f#nk (σik ) = µik . Since lu (µik ) ≥ L1 ≥ L4.12 , and since nk → +∞, we may apply
′
“Negative Flaring”, Lemma 4.12, to obtain subpaths σik
of σik which have endpoints
−
at vertices and which are also subpaths of Λ such that for all i we have
′
lu (σik
)
=∞
k→∞ lu (µik )
lim
′
′
The ratios lu (σik
)/Lu (σik
) and lu (µik )/Lu (µik ) have positive upper and lower bounds
independent of i and k: the upper bound of 1 follows from Corollary 4.4; and the
lower bound comes from Lemma 4.13 (3). For all i we therefore obtain
′
Lu (σik
)
=∞
lim
k→∞ Lu (µik )
Using this limit, and using that Lu (µik ) ≥ 1, it follows that for all sufficiently large k
we have
X
ν X
′
Lu (γk ) ≥
(Lu (σik
)−2 ρ )>
Lu (µik ) > ν Lu (f#nk (γk ))
ǫ
1 i
i
where the first inequality follows by applying Lemma 4.13 (1) to the subdivision of γk
′
into the paths σik
and their complementary subpaths. This contradicts (∗), verifying
the first claim.
Our second claim is that for any constant A (on which constraints will be placed
below when this claim is applied) we have
X
Lu (νjk ) Lu (f#nk (γk )) = 1
(ii)
lim
k→∞
Lu (νjk )≥A
To see why, let Ik be the number of µik subpaths of f#nk (γk ), let Jk be the number
of νjk subpaths, and let Kk = Ik + Jk , and so Jk ≤ Ik + 1 and Kk ≤ 2Ik + 1. By
42
Lemma 4.13 (1) applied to f#nk (γk ) we obtain
Lu (f#nk (γk )) ≤
X
Lu (νjk ) +
j
1≤
P
Lu (νjk )≥A Lu (νjk )
Lu (f#nk (γk ))
X
Lu (µik )
i
≤ Lu (f#nk (γk )) + Kk lu (ρ)
P
Lu (µik )
Lu (νjk )<A Lu (νjk )
+
+
nk
Lu (f# (γk ))
Lu (f#nk (γk ))
{z
} |
{z
}
|
P
i
ǫk
δk
Kk lu (ρ)
≤ 1+
Lu (f#nk (γk ))
{z
}
|
ζk
From (i) it follows that δk → 0 as k → +∞. Multiplying the inequality Kk ≤ 2Ik + 1
by lu (ρ)/Lu (f#nk (γk )), and using that Lu (µik ) ≥ 1, it follows that
0 ≤ ζk ≤ 2 lu (ρ) δk +
lu (ρ)
Lu (f#nk (γk ))
and so ζk → 0 as k → +∞. Multiplying the inequality Jk ≤ Ik + 1 by A/Lu (f#nk (γk )),
it follows that
A
0 ≤ ǫk ≤ Aδk +
nk
Lu (f# (γk ))
and so ǫk → 0 as k → +∞. This proves the second claim.
In what follows we will be applying Lemma 4.13 (4), and we will use the constants
A4.13 , R4.13 involved in that statement.
By definition of L1 and by the choice of the uniform splitting exponent d, since
−
ℓu (νjk ) < L1 it follows that f#d (νjk ) splits into terms each of which is either u-legal
or a copy of ρ or ρ̄. Consider the constants B = B4.13 (d) > 0 and b = b4.13 (d) > 0
of Lemma 4.13. Constraining A ≥ B, we may combine (ii) with Lemma 4.13 (2) to
obtain
P
P
d
Lu (νjk )≥A Lu (f## (νjk ))
Lu (νjk )≥A Lu (νjk )
(iii)
≥
b
·
> 3b/4
Lu (f#nk (γk ))
Lu (f#nk (γk ))
for sufficiently large values of k.
By construction, the paths {νjk } occur in order of the subscript j as subpaths of
nk
f# (γk ) with disjoint interiors. By applying the ## Lemma [HM17a, Lemma 1.6 (5)]
nk
(νjk ) occur in order as subpaths of the
using f nk , it follows that the the paths f##
2nk
path f# (γk ) with disjoint interiors. By applying [HM17a, Lemma 1.6 (4)] it follows
nk −d d
nk
(νjk ). Putting these together we see that the
that f##
f## (νjk ) is a subpath of f##
43
nk −d d
paths f##
f## (νjk ) occur in order as subpaths of the path f#2nk (γk ) with disjoint
interiors. These subpaths being Jk in number, together with their complementary
subpaths one has a decomposition of f#2nk (γk ) into at most 2Jk + 1 paths. Ignoring
the complementary subpaths, Lemma 4.13 (1) therefore implies
X
nk −d d
Lu (f##
f## (νjk )) − 2 Jk lu (ρ)
Lu (f#2nk (γk )) ≥
≥
X
nk −d d
Lu (f##
f## (νjk ))
−
2 Jk lu (ρ)
nk −d d
Lu (f##
f## (νjk ))
−
lu (ρ) Lu (f#nk (γk ))
Lu (νjk )≥A
≥
X
Lu (νjk )≥A
where the last inequality follows for sufficiently large k by applying (i) and the inequality Lu (µik ) ≥ 1 to conclude that
X
Lu (µik ) + 2 ≥ 2Ik + 2 ≥ 2Jk
Lu (f#nk (γk )) ≥ 2
i
For sufficiently large k we therefore have
P
nk −d d
Lu (f#2nk (γk ))
Lu (νjk )≥A Lu (f## f## (νjk ))
>
− lu (ρ)
(iv)
Lu (f#nk (γk ))
Lu (f#nk (γk ))
We have already constrained A so that A ≥ B, and we now put one more constraint on A. Applying Lemma 4.13 (2) to f d it follows that if Lu (νjk ) ≥ B then
d
Lu (f##
(νjk )) ≥ b Lu (νjk ), and so if Lu (νjk ) ≥ A = max{B, 1b A4.13 } it follows that
d
Lu (f##
(νjk )) ≥ A4.13 . This allows us to apply “Positive Flaring”, Lemma 4.13 (4),
with the conclusion that, letting R = R4.13 ,
nk −d d
d
(v) Lu (f##
f## (νjk )) ≥ R λ(nk −d)/2 Lu (f##
(νjk ))
as long as Lu (νjk ) ≥ A and as long as k is sufficiently large. Combining (iv) and (v),
if k is sufficiently large we obtain
P
d
Lu (f#2nk (γk ))
Lu (νjk )≥A Lu (f# (νjk ))
(nk −d)/2
> Rλ
− lu (ρ)
Lu (f#nk (γk ))
Lu (f#nk (γk ))
and combining this with (iii) we obtain
Lu (f#2nk (γk ))
3bR (nk −d)/2
>
λ
− lu (ρ)
nk
Lu (f# (γk ))
4
>ν
where the second inequality holds for sufficiently large k. This gives us the final
contradiction to (∗), which completes the proof of Lemma 4.11.
44
4.6
Appendix: The graph homotopy principle
Lemma 4.14 in this section was used earlier in the proof of Lemma 4.8, and it will be
used later in the construction of the “homotopy semigroup action” in Section 7.1. It
is an elementary result in homotopy theory; for precision we state the result in the
language of category theory, and we give the complete proof.
Define the graph-point category, a subcategory of the standard homotopy category
of pairs, as follows. The objects are pairs (G, P ) where G is a finite graph and P ⊂ G is
a finite subset. Each morphism, denoted [f ] : (G, P ) 7→ (H, Q), is the homotopy class
rel P of a homotopy equivalence f : G → H that restricts to a bijection f : P → Q.
Define the fundamental group functor from the graph-point category to the category
of indexed groups as follows. To each pair (G, P ) we associate the indexed family of
groups π1 (G, P ) = π1 (G, p) p∈P , and to each morphism [f ] : (G, P ) → (H, Q) we
associate the indexed family of group isomorphisms
[f ]∗ : π1 (G, P ) → π1 (H, Q) = f∗ : π1 (G, p) → π1 (H, f (p)) p∈P
The category and functor axioms implicit in this discussion are easily checked.
Let Autgp (G, P ) denote the group of automorphisms of (G, P ) in the graph-point
gp
category. Let Autgp
0 (G, P ) < Aut (G, P ), which we call the pure automorphism group
of (G, P ) in the graph-point category, denote the finite index subgroup consisting of
those [f ] ∈ Autgp (G, P ) such that f : P → P is the identity.
45
Lemma 4.14 (The graph homotopy principle).
(1) The graph-point category is a groupoid: every morphism [f ] : (G, P ) → (H, Q)
has an inverse morphism [g] : (H, Q) → (G, P ), meaning that g ◦ f : (G, P ) →
(G, P ) is homotopic to the identity rel P and f ◦ g : (H, Q) → (H, Q) is homotopic to the identity rel Q.
(2) The fundamental group functor is faithful: for any pair of morphisms
[f ], [f ′ ] : (G, P ) → (H, Q), we have [f ] = [f ′ ] if and only if the restricted
maps f, f ′ : P 7→ Q are equal and the induced isomorphisms f∗ , f∗′ : π1 (G, p) →
π1 (H, f (p)) are equal for all p ∈ P .
(3) Two morphisms [f ] : (G, P ) → (H, Q) and [g] : (H, Q) → (G, P ) are inverses
if and only if their restrictions f : P → Q and g : Q → P are inverses and the
isomorphisms f∗ : π1 (G, p) → π1 (H, f (q)) and g∗ : π1 (H, f (q)) → π1 (G, p) are
inverses.
(4) The fundamental group functor restricts to an injective homomorphism defined
on the pure automorphism group Autgp
0 (π1 (G, P )) 7→ ⊕p∈P Aut(π1 (G, p)).
Proof. Once (1) and (3) are proved, (2) and (4) follow immediately.
To prove (3), the “only if” direction is obvious, and for the “if” direction it suffices
to prove this special case: for any self-morphism [f ] : (G, P ) → (G, P ), if f fixes each
p ∈ P and induces the identity π1 (G, p) for each p ∈ P , then f is homotopic to the
identity rel P . For the proof, we know that f is freely homotopic to the identity
map IdG , because G is an Eilenberg-MacClane space and f induces the identity on
its fundamental group. Choose a homotopy h : G × [0, 1] → G from f to IdG .
We claim that for each p ∈ P the closed path γp (t) = h(p, t) (0 ≤ t ≤ 1) is trivial in
π1 (G, p). Applying this claim, we alter the homotopy h as follows: using the homotopy
extension property, for each p ∈ P we may homotope the map h : G × [0, 1] → G,
keeping it stationary on G×0, stationary on G×1, and stationary outside of Up ×[0, 1]
for an arbitrarily small neighborhood Up of p, to arrange that h(p × [0, 1]) = p; note
that for a homotopy X × [0, 1] → Y to be “stationary on A ⊂ X” means that the
restricted map {a} × [0, 1] → Y is constant for each a ∈ A. Doing this independently
for each p ∈ P , we obtain a homotopy rel P from f to the identity and we are done,
subject to the claim.
To prove the claim, consider a closed path δ : [0, 1] → G based at p, representing
an arbitrary element [δ] ∈ π1 (G, p). We obtain a path homotopy Ht : [0, 1] → G from
the path H0 = f ◦ δ to the concatenated path H1 = γp ∗ δ ∗ γ̄p as follows:
if 0 ≤ s ≤ t/3
γp(3s)
Ht (s) = h δ 33s−−2tt , t
if t/3 ≤ s ≤ 1 − t/3
γ (3 − 3s)
if 1 − t/3 ≤ s ≤ 1
p
46
Since for all [δ] ∈ π1 (G, p) we have [δ] = [f ◦ δ] = [γp ] · [δ] · [γp ]−1 , it follows that [γp ]
is in the center of π1 (G, p) ≈ Fn , hence is trivial, completing the proof of (3).
To prove (1), start with any homotopy inverse g ′ : H → G of f . We may assume
that the maps f : P → Q and g ′ : Q → P are inverses, because by the homotopy
extension property we may homotope g ′ to be stationary outside of a small neighborhood of P so that for each p ∈ P the track of the homotopy on the point g ′ (f (p))
moves it back to p. Since g ′ ◦ f : G → G fixes each point in P and is homotopic to
the identity, for each p ∈ P the induced map (g ′ ◦ f )∗ : π1 (G, p) → π1 (G, p) is an
inner automorphism represented by some closed curve γp based at p, and so for each
element of π1 (G, p) having the form [δ] for some closed curve δ based at p we have
(g ′ ◦ f )∗ (δ) = [γp ∗ δ ∗ γ̄p ]. Let h : (G, p) → (G, p) be the morphism obtained from
the identity by a homotopy that is stationary outside a small neighborhood of P and
such that the track of the homotopy on each p ∈ P is the closed curve γ̄p ; again we
are applying the homotopy extension property. Letting g = h ◦ g ′ we may apply (3)
to conclude that the morphism [f ] is an isomorphism with inverse [g].
Remark. Note that the proof of (3) depends heavily on the fact that the center
of Fn is trivial. The proof breaks down, for instance, if G is replaced by a torus;
in fact the analogue of Lemma 4.14, where a graph is replaced by a torus and P is
a two-point subset of the torus, is false. On the other hand the analogue for any
K(π, 1) space whose fundamental group has trivial center is true.
5
Flaring in T ∗ and hyperbolicity of S.
Throughout this section we continue with Notations 4.1 (1)–(4) regarding an outer
automorphism φ ∈ Out(Fn ) and a relative train track representative f : G → G
having penultimate filtration element Gu−1 and top EG stratum Hu .
The main result of this section is the construction, carried out in Section 5.5, of the
Gromov hyperbolic space S that is used in later sections for proving the multi-edge
case of Theorem F. The construction of S is based on results found in Sections 5.1–
5.4, in particular Proposition 5.10 which is a re-interpretation of the flaring result
of Proposition 4.10 expressed in the context of a certain natural free splitting. The
statement of Proposition 5.10 is found in Section 5.4, after preliminary work carried
out in Sections 5.1–5.3.
5.1
The free splitting Fn y T and its role in flaring.
We begin with a description of the free splitting Fn y T associated to the marked
graph G and its subgraph Gu−1 , together with a description of some features of T
associated to height u Nielsen paths in G, after which we will explain some intuition
behind the flaring result Proposition 5.10.
47
Let F denote the free factor system corresponding to the subgraph Gu−1 , having the form F = {[A1 ], . . . , [AK ]} where Gu−1 has noncontractible components
C1 , . . . , CK and Ak < Fn is in the conjugacy class of the image of the injection
π1 (Ck ) ֒→ π1 (G) ≈ Fn (that injection determined up to inner automorphism of Fn by
appropriate choices of base points and paths between them).
Definition 5.1 (The free splitting Fn y T ). Let Fn y T denote the free splitting
corresponding to the subgraph Gu−1 ⊂ G. What this means is that, starting from
e associated to the universal covering map G
e 7→ G, the tree
the deck action Fn y G
e
eu−1
T is obtained from G by collapsing to a point each component of the total lift G
e → T denote the Fn -equivariant collapse map. Since G
eu−1 is
of Gu−1 . Let p : G
e induces via p an action Fn y T which is evidently
Fn -invariant, the action Fn y G
a free splitting, i.e. a minimal action on a simplicial tree with trivial edge stabilizers.
Note that the set of conjugacy classes of nontrivial vertex stabilizers of this action
is precisely the free factor system F — indeed the stabilizer of a vertex v ∈ T equals
e which is nontrivial if and only if p−1 (v) is a component
the stabilizer of p−1 (v) ⊂ G
eu−1 covering some noncontractible component Ck ⊂ Gu−1 , in which case the
of G
stabilizer of v is conjugate to Ak .
e and projecting down to T ). Fixing a
Definition 5.2 (Lifting f : G → G up to G
standard isomorphism between the deck transformation group of the universal covere 7→ G and the group Fn ≈ π1 (G), recall from covering space theory that the
ing map G
e→G
e of f : G → G are in bijective correspondence with the automorphisms
lifts G
Φ ∈ Aut(Fn ) that represent φ, where the bijection Φ ↔ f˜Φ is given by the following
relation:
e
Φ-twisted equivariance in G:
e
f˜Φ (γ · x) = Φ(γ) · f˜Φ (x) for all γ ∈ Fn , x ∈ G
Φ-twisted equivariance in T :
fTΦ (γ · x) = Φ(γ) · fTΦ (x) for all γ ∈ Fn , x ∈ T
eu−1 , and hence f˜Φ induces a
Since f preserves Gu−1 , any of its lifts f˜Φ preserves G
e implies a corresponding
map fTΦ : T → T . The Φ twisted equivariance property in G
property in T :
When the automorphism Φ is understood or is not important to the discussion, we
will often drop it from the notations f˜Φ and fTΦ , writing simply f˜ and fT instead.
In the next section we will impose an additional metric constraint on fT ; see under
the heading “Stretch properties of fT ”.
Definition 5.3 (Nielsen paths, the Nielsen set, and ρ∗ paths in T ). In the geometric
and parageometric cases, where ρ, ρ̄ exist and are the unique inverse pair of Nielsen
e is any lift of ρ or ρ̄, and a Nielsen path in T
paths of height u, a Nielsen path in G
e In the geometric case, where ρ is
is any projection to T of any Nielsen path in G.
e is a line
closed and has distinct initial and terminal directions, a Nielsen line in G
48
which projects to a bi-infinite iterate of ρ, and a Nielsen line in T is the projection
e
of a Nielsen line in G.
The Nielsen set N , a collection of subsets of T , is defined as follows. In the
ageometric case, N = ∅; in the geometric case, N is the set of Nielsen lines in T ;
and in the parageometric case, N is the set of Nielsen paths in T . Furthermore, for
each N ∈ H its basepoint set, denoted Z(N), is defined as follows. In the geometric
case, the Nielsen line N has a unique decomposition as a bi-infinite concatenation
of Nielsen paths, and Z(N) is defined to be the set of concatenation points. In the
parageometric case, where N is a Nielsen path, Z(N) is its endpoint set.
Note that in the geometric case, basepoint sets of distinct Nielsen lines are disjoint
— for all N 6= N ′ ∈ N we have Z(N) ∩ Z(N ′ ) = ∅. This follows from two facts about
e lying over p is an endpoint of exactly
the base point p of ρ. First, each point p̃ ∈ G
e both contained in the same Nielsen line in G.
e Second, p is
two Nielsen paths in G,
eu−1 , and
not contained in Gu−1 [FH11, Corollary 4.19], so no lift p̃ is contained in G
e 7→ T is locally injective on the complement of G
eu−1 .
the projection map G
e with endpoints at vertices, with projections σG in
Consider a finite path σGe in G
i
i
G and σT in T . If σG = ρ or ρ̄ for some integer i 6= 0 — in the parageometric case
where ρ is not closed, i must equal 1 — then we say that σT is a ρ∗ -path in T , the
superscript ∗ representing the exponent. Note that ρ∗ paths in T are precisely the
paths of the form QQ′ for which there exists N ∈ N such that Q, Q′ ∈ Z(N).
We now give an intuitive explanation of Proposition 5.10, which is a flaring result
expressed in the context of the free splitting Fn y T . Any path function l(·) on G
e which
which vanishes on paths in Gu−1 can be lifted to a path function ˜l(·) on G
eu−1 , and it can then be projected to a path function lT (·)
vanishes on paths in G
on T , which can in turn be re-interpreted as a function dl (·, ·) on pairs of vertices
e → G
e which then induces a
of T . Choose a lift of f : G → G to a map f˜ : G
map fT : T → T . We wish to intepret dl (·, ·) as a “metric” on T , in which case
Proposition 5.10 expresses flaring of the “metric” dl with respect to the action of fT .
If we attempt to carry out this process using either of the “little l” path functions
lu or lPF we do indeed obtain metrics du , dPF on vertices of T . However, in the case
that Hu is geometric, the desired flaring condition fails: for any Nielsen line N ⊂ T ,
iteration of fT = fTΦ produces a sequence of Nielsen lines Ni = (fTk )# (N) (k ∈ Z),
and furthermore the map fTk takes the base lattice Z(N) to the base lattice Z(Nk )
preserving path distance. Since the base lattice of a Nielsen line has infinite diameter
in any invariant path metric on T , this demonstrates the failure of flaring.
If instead we carry out this process starting with either of the “big L” path
functions Lu or LPF , obtaining “metrics” Du , DPF , then base lattices of Nielsen lines
have uniformly bounded diameter, and so we may hope to avert failure of flaring.
However, Du and DPF are not actually metrics, in that they satisfy only a coarse
version of the triangle inequality (c.f. Lemma 4.7).
In order to work with actual path metrics, not just some kind of fuzzy “coarse
49
metrics”, we use a relative hyperbolicity construction. In the geometric case we cone
off the base lattice of each Nielsen axis of T , embedding T naturally in a graph T ∗
that is equipped with a path metric d∗ that is closely related to DPF , and extending
fT naturally to fT ∗ : T ∗ → T ∗ . In the ageometric case we simply set T ∗ = T and
d∗ = dPF . In the parageometric case we will for consistency construct T ∗ by coning
off the endpoint pair of each Nielsen path in T . In all cases fT extends naturally
to fT ∗ . Thus the statement of Proposition 5.10 will unify all cases, expressing flaring
of d∗ with respect to fT ∗ .
Along the way we will state and prove two important properties of T ∗ , which are
almost trivial in the ageometric and parageometric cases but have some significant
content in the geometric case: Proposition 5.8 which gives quasicomparability of the
coarse metric DPF and the true metric d∗ ; and Proposition 5.9 which proves Gromov
hyperbolicity of T ∗ with the metric d∗ .
5.2
e and T . A piecewise Riemannian
Path functions on G
metric on T .
In a tree, a finite path with initial and terminal endpoints V, W is determined by
those endpoints and is denoted V W .
Both of the path functions lPF , LPF vanish on paths in Gu−1 . Both of them lift
e → G to an Fn -invariant path function on G
e that
via the universal covering map G
e
e
vanishes on paths in Gu−1 , and hence projects via q : G → T to a well-defined and
Fn -invariant path function on T . We re-use the notations lPF , LPF for these path
e and on T . For any path β e in G
e with endpoints at vertices, letting
functions on G
G
βG be its projection to G and βT its projection to T , it follows from the definitions
that l(βG ) = l(βGe ) = l(βT ) for l = lPF or LPF .
Remark. The point of “well-definedness” of LPF on T is that for any vertices
e both map to V and W1 , W2 both map to W then each of
V, W ∈ T , if V1 , V2 ∈ G
eu−1 , and hence
the paths V1 V2 and W1 W2 is either degenerate or is contained in G
LPF (V1 W1 ) = LPF (V2 W2 ) = LPF (V W ).
e and on T , we have Fn Associated to the path functions lPF (·), LPF (·) on G
e and in T ,
equivariant functions dPF (·, ·), LPF (·, ·) on pairs of vertices V, W in G
given by
dPF (V, W ) = lPF (V W ), DPF (V, W ) = LPF (V W )
Furthermore, for each edge E ⊂ T we have lPF (E) = LPF (E) ≥ η > 0, which is an
immediate consequence of the fact that as E varies over the finite set of edges of Hu ,
the finite set of positive numbers lPF (E) has a positive minimum. We record this fact
for later use:
Lemma 5.4. There exists η = η5.4 > 0 such that for all vertices V 6= W ∈ T we have
DPF (V, W ) = LPF (V W ) ≥ η.
50
Using the inequality lPF (E) > 0 we may construct an Fn -equivariant piecewise
Riemannian metricR on the tree T , denoted ds, such that for vertices V, W ∈ T we
have dPF (V, W ) = V W ds. From this it follows that the function dPF (·, ·) is a metric
on the vertex set of T .
Next we translate into the context of T several results on path functions in Section 4.1.
Proposition 5.5. For any vertices V 6= W ∈ T there is a unique decomposition of
the path V W = µ0 ν1 µ1 · · · νA µA such that the following properties hold:
(1) If ρ does not exist then A = 0.
(2) If ρ exists then the νa ’s are precisely all of the maximal ρ∗ paths of V W , and
so each µa contains no ρ∗ subpath.
(3) If ρ exists and if 1 ≤ a < a+ 1 ≤ A−1 then at least one of the subpaths µa , µa+1
is nondegenerate; in the geometric case, all µa -subpaths are nondegenerate.
(4) DPF (V, W ) = lPF (µ0 ) + · · · + lPF (µA )
Furthermore, given any path γ in T — finite, singly infinite, or bi-infinite — whose
endpoints, if any, are at vertices, and assuming that ρ exists, there is a unique decomposition of γ as an alternating concatenation of its maximal ρ∗ paths (the ν-subpaths)
and paths that contain no ρ∗ subpath (the µ-subpaths) such that for any two consecutive µ-subpaths at least one is nondegenerate, all µ-subpaths being nondegenerate in
the geometric case.
Proof. Items (1)–(3) are translations of Corollary 4.5, and item (4) is a translation
of Corollary 4.6; the proofs are immediate from those results combined with the
definitions.
The “Furthermore. . . ” clause is a quick consequence of the following observations
that hold for any nested pair of finite subpaths V W ⊂ V ′ W ′ in T . First, every ρ∗
subpath of V W is a ρ∗ subpath of V ′ W ′ . Also, every maximal ρ∗ subpath of V W
whose dPF distances from V and from W are greater than lPF (ρ) is a maximal ρ∗
subpath of V ′ W ′ .
Remark on item (3). Consider the paths µa for 1 ≤ a ≤ A−1. In the geometric
case each such µa is nondegenerate, and this can be used to improve the constants
in some applications of (3) (underlying nondegeneracy of µa is the fact that the base
point of the closed Nielsen path ρ is disjoint from Gu−1 ). In the parageometric case,
on the other hand, one of the paths µa , µa+1 may be degenerate. This happens for µa
only if, up to orientation reversal, the Nielsen path ρ has initial vertex p ∈ Gu−1 (the
terminal vertex is necessarily disjoint from Gu−1 , the fact which underlies item (3)),
e that projects
in which case µa is degenerate if and only if νa µa νa−1 lifts to a path in G
to a path in G of the form ρ̄µρ where µ is a nondegenerate closed path in Gu−1 based
at p.
51
5.3
Constructing T ∗ by coning off Nielsen axes of T .
Embed the tree T into a graph denoted T ∗ , and extend the action Fn y T to an
action Fn y T ∗ , as follows. Index the Nielsen set as N = {Nj }j∈J , letting Zj be
the basepoint set of Nj (Definition 5.1). For each j ∈ J we cone off Zj by adding
a new vertex Pj = P (Nj ) and attaching a unique edge Pj Q for each Q ∈ Zj . The
points Pj are called cone points and the edges Pj Q are called cone edges. Since the
simplicial action Fn y T takes Nielsen paths to Nielsen paths and hence induces a
basepoint preserving permutation of the Nielsen set N , this action extends uniquely
to Fn y T ∗ permuting the cone points and the cone edges.
Let V ∞ ⊂ T ∗ be the set of vertices v ∈ T ∗ whose stabilizer subgroup Stab(v) is
infinite.
Lemma 5.6. The formula v 7→ Stab(v) defines an injection from V ∞ to the set
of nontrivial subgroups of Fn . A subgroup S < Fn is equal to Stab(v) for some
v ∈ V ∞ if and only if S is conjugate to the fundamental group of some noncontractible
component of Gu−1 in π1 (G) ≈ Fn , or Hu is a geometric stratum and S is conjugate
to the infinite cyclic subgroup hρi < Fn .
Proof. This fact may be extracted from [HM17c, Theorem F]. But a direct proof is
easy; here is a sketch.
We have a partition
a
∞
∞
V ∞ = Vu−1
VN∞ where Vu−1
= V ∞ ∩ T and VN∞ = {cone points} = {Pj }
e 7→ T induces an equivariant and hence stabilizer preserving
The collapse map G
∞
eu−1 having nontrivial stabilizer.
bijection between Vu−1 and the set of components of G
Using covering space theory, the stabilizers of the latter components are precisely
those subgroups of Fn conjugate to the fundamental group of some noncontractible
eu−1 are disjoint
component of Gu−1 . Furthermore, since distinct components of G
e the intersections of their stabilizers are trivial, and so the stabilizers
subtrees of G,
are unequal if they are nontrivial. This completes the proof of Hu is nongeometric.
If Hu is geometric then we have equivariant and hence stabilizer preserving bijecej ↔ Nj ↔ Pj where N
ej is the Nielsen line in G
e mapping to Nj under the
tions N
e 7→ T . By definition the N
ej are precisely those lines in G
e that cover
collapse map G
the closed Nielsen path ρ. The element of π1 (G) represented by ρ (and denoted by ρ)
is root free in π1 (G) because ρ has a unique u-illegal turn (Notations 4.1 (2)), and so
ej are precisely the infinite cyclic
by covering space theory the stabilizers of the lines N
subgroups in the conjugacy class of the group hρi < π1 (G) ≈ Fn . And as before, two
different such lines have distinct stabilizers.
The proof is completed by noting that hρi is not conjugate in Fn to the fundamental group of a noncontractible component of Gu−1 , because ρ is a circuit not contained
in Gu−1 (Notations 4.1 (2)).
52
The piecewise Riemannian metric ds∗ on T ∗ . The piecewise Riemannian metric
ds on T extends to an Fn -equivariant piecewise Riemannian metric denoted ds∗ on
T ∗ as follows. In the nongeometric case there is nothing to do. In the geometric case
the group Fn acts freely and transitively on the set of cone edges {Pj Q j ∈ J, Q ∈
Z(Nj )}; extend ds over a single cone edge Pj Q, then extend
it over all other cone
R
∗
edges equivariantly to obtain ds ; note that the length Pj Q ds∗ is independent of j
and Q. In the parageometric case, the group Fn acts freely on the set of cone edges,
and there are two orbits of cone edges corresponding to the two endpoints of ρ; we
extend ds over a single cone edge in each of
orbits, then extend equivariantly
R the two
∗
∗
to obtain ds ; also, we require that length Pj Q ds be the same on both orbits of cone
edges. Define the cone height to be the length of any cone edge in the metric ds∗ .
The path metric d∗ on T ∗ . Define a metric d∗ (·,
·) on vertices of T ∗ by minimizing
R
path length: d∗ (V, W ) equals the minimum of γ ds∗ over all continuous paths in
the graph T ∗ having endpoints V, W . The infimum is evidently minimized by some
embedded edge path in T ∗ having endpoints V, W . Note that since T ∗ is not a tree,
embedded edge paths need not be determined by their endpoints.
d′ denote the path QP (N)∗P (N)Q′ in
Bypasses in T ∗ . For any ρ∗ path QQ′ we let QQ
∗
′
T , called the bypass of QQ . Note that a path in the graph T ∗ is a bypass if and only
if it is a two-edge path having a cone point as its midpoint, and furthermore a bypass
is completely determined by its endpoints. We thus have a one-to-one correspondence
ν ↔ νb between the set of ρ∗ paths and the set of bypasses.
Extending the map fTΦ : T → T to fTΦ∗ : T ∗ → T ∗ . Following Definition 5.2, and
e and project to T obtaining
choosing Φ ∈ Aut(Fn ) representing φ, lift f : G → G to G
Φ-twisted equivariant maps
e→G
e
f˜ = f˜Φ : G
fT = fTΦ : T → T
With respect to the inclusion T ֒→ T ∗ we extend fTΦ to a Φ-twisted equivariant map
fT ∗ = fTΦ∗ : T ∗ → T ∗
as follows. As noted, for convenience we will often suppress Φ from the notation for
these maps.
The action of fT on T induces a well-defined action on the Nielsen set N , and
so we can extend fT over the set of cone points by setting fT ∗ (P (N)) = P (fT (N))
for each N ∈ N . Furthermore, for each N ∈ N the map fT restricts to a bijection
of basepoint sets fT : Z(N) → Z(fT (N)), and so for each Q ∈ Z(N) we can extend
the endpoint map (P (N), Q) 7→ (P (fT (N)), fT (Q)) uniquely to an isometry between
cone edges fT ∗ : P (N), Q → P (fT (N)), fT (Q).
For each edge E ⊂ T we have the following equation that follows from the “big
L” eigenlength equation in Section 4.1:
Z
Z
Eigenlength equation in T :
ds = λ ds
fT (E)
E
53
It follows that by equivariantly homotoping fT relative to endpoints of edges we may
arrange that fT stretches each edge of T by a uniform factor λ. Since its extension fT ∗
is an isometry on edges of T ∗ \ T , it follows that fT ∗ is λ-Lipschitz. These conditions
constitute additional constraints on the maps fT and fT ∗ which we record here:
Stretch properties of fT and fT ∗ :
• The maps fT , fT ∗ stretch each edge E ⊂ T by a constant factor λ over the path
fT (E) = fT ∗ (E).
• The map T ∗ permutes cone edges, taking each cone edge isometrically to its
image.
These stretch properties are the first step of Lemma 5.7 to follow.
We recall here some basic definitions. Given constants k ≥ 1, c ≥ 0, a map of
metric spaces f : X → Y is a (k, c)-quasi-isometric embedding if for all p, q ∈ X we
have
1
d(p, q) − c ≤ d(f (p), f (q)) ≤ k d(p, q) + c
k
If in addition each point of Y has distance ≤ c from some point of f (X) then f is
a (k, c)-quasi-isometry. If the domain X is a subinterval of R then we say that f is
a (k, c)-quasigeodesic. Sometimes we conflate the constants k, c by setting k = c and
using terminology like “k-quasi-isometries” etc. Sometimes we ignore k, c altogether
and use terminology like “quasi-isometries” etc.
Lemma 5.7. The map fT ∗ : T ∗ → T ∗ is a quasi-isometry.
Proof. Let Φ ∈ Aut(Fn ) be the representative of φ corresponding to fT ∗ , and so fT ∗
satisfies the Φ-twisted equivariance equation (see Definition 5.2).
The map fT ∗ is Lipschitz, by the “Stretch properties” noted above. To complete
the proof it suffices to show that there is a Lipschitz map f¯T ∗ : T ∗ → T ∗ such that
fT ∗ and f¯T ∗ are coarse inverses, meaning that each of the composed maps f¯T ∗ ◦ fT ∗
and fT ∗ ◦ f¯T ∗ moves each point of T ∗ a uniformly bounded distance. We construct f¯T ∗
by taking advantage of twisted equivariance of fT ∗ combined with the fact that the
action Fn y T ∗ has finitely many vertex and edge orbits (this is a kind of “twisted
equivariant” version of the Milnor-Svarc lemma []).
`
Consider the vertex set V ∗ of T ∗ and its partition V ∗ = V 0 V ∞ into points
whose Fn -stabilizers are trivial and infinite, respectively. Using Φ-twisted equivariance it follows that for any vertex v ∈ V we have a subgroup inclusion Φ(Stab(v)) ⊂
Stab(fT ∗ (v)). It follows that fT ∗ (V ∞ ) ⊂ V ∞ . Furthermore, that subgroup inclusion
is an equation Φ(Stab(v)) = Stab(fT ∗ (v)) — this is a consequence of Lemma 5.6 combined with the fact that f : G → G restricts to a homotopy equivalence of the union
of noncontractible components of Gu−1 and with the fact that f# (ρ) = ρ. It follows
that the restricted map fT ∗ : V ∞ → V ∞ is a bijection of the set V ∞ .
54
Define the restriction f¯T ∗ V ∞ to equal the inverse of the restriction fT ∗ V ∞ ;
this map f¯T ∗ V ∞ is automatically Φ−1 -twisted equivariant. Define the restriction
f¯T ∗ V 0 as follows: choose one representative v ∈ V 0 of each orbit of the action
Fn y V 0 , choose f¯T ∗ (v) ∈ V 0 arbitrarily, and extend over all of V 0 by Φ−1 -twisted
equivariance. Having defined a Φ−1 -twisted equivariant map f¯T ∗ : V ∗ → V ∗ , extend
f¯T ∗ over each edge to stretch distance by a constant factor, and hence we obtain a
Φ−1 -twisted equivariant map f¯T ∗ : T ∗ → T ∗ . Since there are only finitely many orbits
of edges, f¯T ∗ is Lipschitz. Since fT ∗ is Φ-twisted equivariant and f¯T ∗ is Φ−1 -twisted
equivariant, it follows that the two compositions fT ∗ ◦ f¯T ∗ and f¯T ∗ ◦fT ∗ are equivariant
in the ordinary untwisted sense. Each of these compositions therefore moves each
point a uniformly bounded distance, hence fT ∗ and f¯T ∗ are coarse inverses.
5.4
Geometry and dynamics on T ∗.
In this section we prove several propositions regarding T ∗ , including Proposition 5.10
which is our interpretation of flaring in T ∗ . The proofs will follow after the statement
of all the propositions.
Proposition 5.8 (Quasicomparibility in T ∗ ). The inclusion of the vertex set of T
into the vertex set of T ∗ is a quasi-isometry from the coarse metric DPF (·, ·) to the
metric d∗ (·, ·), meaning that there exist constants K = K5.8 ≥ 1, C = C5.8 ≥ 0 such
that for all vertices V, W ∈ T we have
1 ∗
d (V, W ) − C ≤ DPF (V, W ) ≤ K d∗ (V, W ) + C
K
Proposition 5.9 (Hyperbolicity of T ∗ ). The graph T ∗ with the metric d∗ is Gromov
hyperbolic. Furthermore, geodesics in T are stable in T ∗ in the following sense: there
exists C = C5.9 ≥ 0 such that for any vertices V, W ∈ T , the path V W ⊂ T is
∗
C-Hausdorff close to any geodesic V W ⊂ T ∗ with endpoints V, W .
We now state our re-interpretation of the earlier flaring result Proposition 4.10 in
the setting of T ∗ . Given η > 0, and given a sequence of vertices Vr ∈ T ∗ defined for r
in some interval of integers a ≤ r ≤ b, we say that this sequence is an η-pseudo-orbit
of fT ∗ if d∗ (fT ∗ (Vr ), Vr+1 ) ≤ η for all a ≤ r < r + 1 ≤ b.
Proposition 5.10 (Flaring in T ∗ ). For each µ > 1, η ≥ 0 there exist integers R ≥ 1
and A ≥ 0 such that for any pair of η-pseudo-orbits Vr and Wr of fT ∗ defined on the
integer interval −R ≤ r ≤ R, if d∗ (V0 , W0 ) ≥ A then
µ · d∗ (V0 , W0 ) ≤ max{d∗ (V−R , W−R ), d∗ (VR , WR )}
55
Proof of Proposition 5.8: Quasicomparibility in T ∗ . In the ageometric case
where ρ does not exist then T = T ∗ and DPF = d∗ and we are done. Henceforth we
assume that ρ exists.
Letting h denote the cone height, it follows that each bypass has length 2h.
Given vertices V, W ∈ T , using Proposition 5.5 we obtain a decomposition of the
path V W and an accompanying formula for DPF (V, W ):
V W = µ0 ν1 µ1 . . . νA−1 µA−1 νA µA
DPF (V, W ) = lPF (µ0 ) + · · · + lPF (µA )
Each νi is a subpath of some element N ∈ N of the Nielsen set and the endpoints
of νi are distinct points in Z(N), and so the corresponding bypass νbi is defined. We
thus obtain a path in T ∗ and a length calculation as follows:
Vd
W = µ0 νb1 µ1 . . . νbA−1 µA−1 νbA µA
Length(Vd
W ) = lPF (µ0 ) + · · · + lPF (µA ) + 2hA
Applying Proposition 5.4 with its constant η = η5.4 , and applying Proposition 5.5 (3),
it follows that if 0 ≤ a < a + 1 ≤ A then at least one of lPF (µa ), lPF (µa+1 ) is ≥ η, and
hence
lPF (µ0 ) + · · · + lPF (µA ) ≥ (A − 1) η/2
We therefore have
d∗ (V, W ) ≤ Length(Vd
W)
≤ lPF (µ0 ) + · · · + lPF (µA ) + 2h
= 1 + 4h/η DPF (V, W ) + 2h
2(lPF (µ0 ) + · · · + lPF (µA ))
+1
η
This proves the first inequality using any K ≥ 1 + 4h/η and C = 2h/K.
We turn to the opposite inequality. Before starting, we shall normalize the choice
of cone height to be h = 12 lPF (ρ) and so each bypass νb has length lPF (ρ). Proving
Proposition 5.8 with h normalized in this fashion implies the proposition for any
value of h, because the normalized version of d∗ is bi-Lipschitz equivalent to any
other version.
Given vertices V, W ∈ T , choose an embedded edge path γ ∗ ⊂ T ∗ with endpoints
V, W satisfying two optimization conditions:
R
R
(i) γ ds∗ is minimal, and so d∗ (V, W ) = γ ds∗ .
R
R
(ii) Subject to (i), γ∩T ds∗ = γ∩T ds is minimal.
56
There is a unique decomposition of γ as an alternating concatenation of subpaths in
T and bypasses:
γ ∗ = µ0 νb1 µ1 . . . µA−1 νbA µA
from which we obtain the formula
d∗ (V, W ) = lPF (µ0 ) + · · · + lPF (µA ) + A lPF (ρ)
We claim that each µa has no ρ∗ -subpath. Otherwise we can decompose µa = µ′ ν ′ µ′′
where ν ′ is a Nielsen path in T . In the case that each of µ′ , µ′′ is nondegenerate, or
that a = 0 and µ′′ is nondegenerate, or that a = A and µ′ is nondegenerate, construct
a path γ ′ from γ by replacing
ν ′ Rwith the corresponding
R
R
R bypass; from the choice of
∗
∗
normalization we have γ ′ ds = γ ds but γ ′ ∩T ds < γ∩T ds, a contradiction. The
other cases all lead to a path γ ′ exhibiting the same contradiction, and are described
as follows. In the case that a ≥ 1 and µ′ is degenerate, replace the subpath νba ν ′ of γ
with the unique bypass having the same endpoints. And in the case that a ≤ A − 1
and µ′′ is degenerate, replace the subpath ν ′ νba+1 with the unique bypass having the
same endpoints. And in the last remaining case, where a = A = 0 and µ′, µ′′ are both
degenerate, we have ν ′ = γ ∗ and we let γ ′ = νb′ be the corresponding bypass.
Consider the concatenated edge path in T denoted
µ0 ν1 µ1 . . . µA−1 νA µA
which is obtained from γ ∗ by replacing each bypass νba with the corresponding ρ∗ subpath νa ⊂ T . Straightening this concatenation in T produces the path V W
to which we may inductively apply the coarse triangle inequality for LPF given in
Lemma 4.7, with the conclusion that
DPF (V, W ) = LPF (V W )
≤ LPF (µ0 ) + LPF (ν1 ) + LPF (µ1 ) + · · ·
+ LPF (µA−1) + LPF (νA ) + LPF (µA ) + (2A − 1)C4.7
Since µa has no ρ±i subpath we have LPF (µa ) = lPF (µa ), and since LPF (ρ±i ) = 0 we
have LPF (νa ) = 0, and therefore
DPF (V, W ) ≤ lPF (µ0 ) + · · · + lPF (µA ) + 2AC4.7
2C4.7
· AlPF (ρ)
≤ lPF (µ0 ) + · · · + lPF (µA ) +
lPF (ρ)
≤ Kd∗ (V, W )
for any K ≥ max{1, 2C4.7/lPF (ρ)}. This completes the proof of Proposition 5.8.
57
Proof of Proposition 5.9: Hyperbolicity of T ∗ . If ρ does not exist, i.e. in the
ageometric case, since T ∗ = T is a tree we are done. The parageometric case is
similarly easy, but it is also subsumed by the general proof when ρ does exist for
which purpose we will apply a result of Kapovich and Rafi [KR14, Proposition 2.5].
Let T ∗∗ be the graph obtained from T by attaching an edge Q Q′ for each unordered
pair Q 6= Q′ ∈ Z(N) for each element N ∈ N of the Nielsen set. We put the
simplicial metric on T ∗∗ , assigning length 1 to each edge. We have a map T ∗∗ → T ∗
extending the inclusion T ֒→ T ∗ , defined to take each attached edge Q Q′ ⊂ T ∗∗ to the
d′ . This map is evidently a quasi-isometry from the vertices
corresponding bypass QQ
∗∗
of T to the vertices of T ∗ , and this quasi-isometry commutes with the inclusions
of T into T ∗ and into T ∗∗ . The conclusions of Proposition 5.9, namely hyperbolicity
of T ∗ and stability in T ∗ of geodesics in T , will therefore follow once we demonstrate
the same conclusions with T ∗∗ in place of T ∗ . Those conclusions for T ∗∗ are identical
to the conclusions of [KR14, Proposition 2.5] applied to the inclusion map T ֒→ T ∗∗ :
the graph T ∗∗ is hyperbolic; and each arc V W ⊂ T is uniformly Hausdorff close in
T ∗∗ to each geodesic in T ∗∗ with the same endpoints V, W . So we need only verify
that the inclusion map T ֒→ T ∗∗ satisfies the hypotheses of [KR14, Proposition 2.5]
with respect to the simplicial path metrics on T and T ∗∗ that assign length 1 to each
edge.
One hypothesis [KR14, Proposition 2.5] is that T be hyperbolic, which holds for
all path metrics on trees. Another hypothesis is that the inclusion T ֒→ T ∗∗ be a
Lipschitz graph map which means that it takes each edge of T to a bounded length
edge path of T ∗ , but this is immediate since the inclusion is an isometry onto its
image.
The remaining hypotheses of [KR14, Proposition 2.5] are numbered (1), (2), (3),
the first two of which are trivially satisfied using that the inclusion map from vertices
of T to vertices T ∗∗ is surjective (this is why we use T ∗∗ instead of T ∗ ). The last
hypothesis (3) says that there exists an integer M > 0 so that for any vertices
V 6= W ∈ T , if V, W are connected by an edge in T ∗∗ then the diameter in T ∗∗
of the path V W ⊂ T is uniformly bounded. The only case that needs attention is
when V W is not a single edge in T but V, W are connected by an edge in T ∗∗ . This
happens only if V W is a ρ∗ subpath of some element N ∈ N of the Nielsen set, and
so V W is a concatenation of a sequence of Nielsen paths Qk−1 Qk in N, where the
concatenation points form a consecutive sequence in Z(N) ∩ V W of the form
V = Q0 , Q1 , . . . , QK = W
Each pair Qi , Qj is connected by an edge Qi Qj ⊂ T ∗∗ (i ≤ j = 0, . . . , K). Since each
vertex of T along V W is contained in one of the Nielsen paths Qk−1 Qk and hence its
distance to one of Qk−1 or Qk is at most lu (ρ)/2, it follows that the diameter of V W
in T ∗∗ is bounded above by M = 1 + lu (ρ).
58
Proof of Proposition 5.10: Flaring in T ∗ . We denote the constants of Proposition 5.8 in shorthand as K = K5.8 , C = C5.8 .
Fix µ > 1 and η ≥ 0. Consider R ≥ 1, to be specified, and a pair of η-pseudofr ∈ G
e projecting
orbits Vr , Wr for fT ∗ defined for −R ≤ r ≤ R. Choose vertices Ver , W
e
to Vr , Wr respectively. In G denote the path β̃r = Vr Wr , and let βr be its projection
to G. Also, for −R < r ≤ R denote α̃r = Vr , f˜(Vr−1 ) and ω̃r = f˜(Wr−1 ), Wr , which
are the unique paths such that β̃r = [α̃r f˜# (β̃r−1 )ω̃r ]. Let αr , ωr be their projections
to G. By assumption we have d∗ (fT ∗ (Vr−1 ), Vr ) ≤ η and d∗ (fT ∗ (Wr−1 ), Wr ) ≤ η, and
applying Proposition 5.8 it follows that
LPF (αr ), LPF (ωr ) ≤ K η + C ≡ η ′
and so f# (βr−1 )
η′
∼
LPF
2
βr .
Let µ′ = 2K µ. By Proposition 4.10 the path function LPF satisfies the flaring
condition of Definition 4.9 with respect to f , from which using µ′ and η ′ we obtain
constants R′ ≥ 1, A′ ≥ 1. We now specify R = R′ . Also let
A = max{KA′ + C, 2K 2 C +
2C
}
µ
Applying Proposition 5.8 it follows that if d∗ (V0 , W0 ) ≥ A then LPF (β0 ) ≥ A′ . The
flaring condition of LPF therefore applies with the conclusion that
µ′ LPF (β0 ) ≤ max{LPF (β−R ), LPF (βR )}
and so
µ d∗ (V0 , W0 ) ≤ Kµ LPF (β0 ) + KCµ
Kµ
≤ ′ max{LPF (β−R ), LPF (βR )} + KCµ
µ
Applying Proposition 5.8 again we have
∗
µ d (V0 , W0 ) ≤
≤
≤
1 ∗
µ d (V0 , W0 ) ≤
2
K 2µ
max{d∗ (V−R , W−R ), d∗ (VR , WR )} + K 2 Cµ + C
′
µ
1
1
max{d∗ (V−R , W−R ), d∗(VR , WR )} + µ A
2
2
1
1
max{d∗ (V−R , W−R ), d∗(VR , WR )} + µ d∗ (V0 , W0 )
2
2
1
max{d∗ (V−R , W−R ), d∗(VR , WR )}
2
which completes the proof.
59
5.5
Construction of S.
We continue to fix the choice of Φ ∈ Aut(Fn ) representing φ, and we consider the
corresponding Φ-twisted equivariant map fT ∗ = fTΦ∗ : T ∗ → T ∗ . Our definition of the
suspension space S will formally depend on this choice (but see remarks at the end
of the section regarding dependence on Φ of the constructions to follow).
Definition 5.11 (The suspension space S, its slices, fibers, and semiflow). We define
S to be the suspension space of T ∗ , namely the quotient of T ∗ × Z × [0, 1] modulo
the gluing identification (x, k, 1) ≈ (fT ∗ (x), k + 1, 0) for each k ∈ Z and x ∈ T ∗ .
Let [x, k, r] ∈ S denote the equivalence class of (x, k, r). We have a well-defined and
continuous projection map p : S → R given by p[x, k, r] = k + r. For each closed
connected subset J ⊂ R we denote SJ = p−1 (J) which we refer to as a slice of S. In
the special case of a singleton s ∈ R we refer to Ss = p−1 (s) as a fiber. Each fiber
may be regarded as a copy of T ∗ , in the sense that if k = ⌊s⌋ and r = s − k then we
obtain a homeomorphism js : T ∗ → Ss given by js (x) = [x, k, r]; in the special case
that s is an integer we have js (x) = [x, s, 0].
We have an action Fn y S which is induced by the action Fn y T ∗ as follows:
for each γ ∈ Fn and each [x, k, r] ∈ S we have
γ · [x, k, r] = [Φk (γ) · x, k, r]
This action is well-defined because, using Φ-twisted equivariance of fT ∗ : T ∗ → T ∗ ,
we have
γ · [x, k, 1] = [Φk (γ) · x, k, 1] = [fT ∗ (Φk (γ) · x), k + 1, 0] =
= [Φ(Φk (γ)) · fT ∗ (x), k + 1, 0]
= [Φk+1 (γ) · fT ∗ (x), k + 1, 0]
= γ · [fT ∗ (x), k + 1, 0]
Note that the homeomorphism j0 : T ∗ → S0 is equivariant with respect to Fn actions.
For generally, for each integer k the homeomorphism jk : T ∗ → Sk is Φ−k -twisted
equivariant, because
jk (γ · x) = [γ · x, k, 0] = Φ−k (γ) · [x, k, 0] = Φ−k (γ) · jk (x)
We have a semiflow S × [0, ∞) → S, which is partially defined by the formula
[x, k, s] · t = [x, k, s + t]
(x ∈ T ∗ , k ∈ Z, 0 ≤ s ≤ 1, 0 ≤ t ≤ 1 − s)
and which uniquely extends to all t ≥ 0 by requiring the semiflow equation (p · t) · u =
p · (t + u) to hold for all t, u ≥ 0. In particular Ss · t = Ss+t for all s ∈ R, t ≥ 0. For
each b ∈ R we define the first hitting map hb : S(−∞,b] 7→ Sb by letting ξ ∈ S(−∞,b] flow
forward from ξ ∈ Sp(ξ) to Sb along the flow segment ξ · [0, b − p(ξ)], thus obtaining
the formula hb (ξ) = ξ · (b − p(ξ)).
This completes Definition 5.11.
60
Definition 5.12 (Piecewise Riemannian metric and geodesic metric on S). We define
a piecewise Riemannian metric on S. Recall the piecewise Riemannian metric ds∗
on T ∗ (see the beginning of Section 5.3). For each edge E ⊂ T ∗ and each integer n
we define a Riemannian metric dE on E × n × [0, 1] (≈ E × [0, 1] and not depending
on n), in two cases. In the case E ⊂ T , use the metric d2E = λ2t (ds∗ )2 + dt2 ; note
that fT ∗ stretches the length of E by a constant factor λ, and hence the gluing map
(x, n, 1) 7→ (fT ∗ (x), n+ 1, 0) takes E ×n×1 isometrically onto fT ∗ (E) ×(n+ 1) ×0. In
the case where E ⊂ T ∗ \T , equivalently we are in the geometric or parageometric case
and E is a cone edge, use the metric d2E = (ds∗ )2 +dt2 ; note that E ′ = fT ∗ (E) is also a
cone edge and that fT ∗ maps E isometrically to E ′ , and so once again the gluing map
(x, n, 1) 7→ (fT ∗ (x), n + 1, 0) takes E × n × 1 isometrically onto fT ∗ (E) × (n + 1) × 0.
These metrics on all of the rectangles E × n × [0, 1] glue up isometrically along their
common boundaries in S, as follows. First, for any two edges E, E ′ ⊂ T ∗ having a
common endpoint x = E∩E ′ = ∂E∩∂E ′ , the restrictions to [0, 1] ≈ x×n×[0, 1] of the
metric dE on E ×[0, 1] ≈ E ×n×[0, 1] and the metric dE ′ on E ′ ×[0, 1] ≈ E ′ ×n×[0, 1]
are both equal to dt. Fixing n and letting E ⊂ T ∗ vary, this allows us to glue up all
of the rectangles E × n × [0, 1] to get a well-defined piecewise Riemannian metric on
T ∗ ×n×[0, 1]. Next, as noted above the gluing map from T ∗ ×n×1 to T ∗ ×(n+ 1) ×0
maps each E × n × 1 isometrically onto fT ∗ (E) × (n + 1) × 0; letting n vary this allows
us to glue up T ∗ × n × [0, 1] to get the desired piecewise Riemannian metric on S.
By minimizing path lengths using the above piecewise Riemannian metric on S,
we obtain a geodesic metric dS (·, ·) on S. We may similarly define a geodesic metric
dJ (·, ·) on any slice SJ , by minimizing path length with respect to the restricted
piecewise Riemannian metric on SJ . In the special case of a fiber Ss , letting n = ⌊s⌋,
for each edge E ⊂ T ∗ with image edge js (E) ⊂ Ss the metrics d∗ on E and ds in js (E)
are related so that if E ⊂ T then js stretches the metric by a factor of λs−n , whereas
if E ⊂ T ∗ \ T then js preserves the metric. In particular the map js : T ∗ → Ss is a
λs−n bilipschitz homeomorphism; it is therefore an isometry if s is an integer, and a
λ-bilipschitz homeomorphism in general.
This completes Definition 5.12.
For use at the very end of the proof of Theorem F, when verifying the WWPD
conclusions, we shall need the following metric property of S:
Lemma 5.13. For any two fibers Ss , St ⊂ S and any x ∈ Ss , y ∈ St we have
dS (x, y) ≥ s − t .
Proof. The lemma follows by noting that for each edge E ⊂ T ∗ , the projection
function p : E × [0, 1] → [0, 1] has the property that for each tangent vector v ∈
T (E × [0, 1]) we have Dp(v) ≥ v , using the Riemannian metric dE on the right
hand side of the inequality.
The following metric property will be used later to study how the inclusion map
of fibers Si ֒→ S can distort distance.
61
Lemma 5.14. For each integer m there exist constants km ≥ 1, cm ≥ 0 such that for
each integer a and each s ∈ J = [a, a + m] the inclusion map i : Ss ֒→ SJ is a km , cm
quasi-isometry.
Proof. We construct a cycle of equivariant Lipschitz maps as follows:
i
SO s
/
SJ
ha+m
hr
Sa o
h̄
Sa+m
The inclusion map i is 1-Lipschitz. The equivariant maps ha+m and hr are first hitting
maps, each of which is λm Lipschitz. The map h̄ will be an equivariant coarse inverse
to the equivariant map ha+m : Sa → Sa+m , for the construction of which we consider
the commutative diagram
4
f¯Tm∗
T∗
ja
/
Sa j
fTm∗
ha+m
T∗
/
ja+m
h̄
Sa+m
In this diagram the map fTm∗ is Φm -twisted equivariant, and the map ha+m is (untwisted) equivariant. The top and bottom maps are the instances k = a and k = a+m
of the Φ−k -twisted equivariant isometry jk : T ∗ → Sk , whose inverse jk−1 : Sk → T ∗
is Φk twisted equivariant. By Lemma 5.7 the map fT ∗ has a Φ−1 -twisted equivariant
Lipschitz coarse inverse f¯T ∗ . It follows that fTm∗ has a Φ−m -twisted equivariant Lipschitz coarse inverse f¯Tm∗ whose Lipschitz and coarse inverse constants depend on m.
−1
We may therefore fill in the diagram with the map h̄ = ja ◦ f¯Tm∗ ◦ ja+m
which makes
the diagram commute and which is an untwisted equivariant Lipschitz coarse inverse
for ha+m , with Lipschitz and coarse inverse constants depending on m.
Going round the cycle from Ss to itself and from SJ to itself we obtain two equivariant self-maps both of which move points uniformly bounded distances depending
on m. The maps i : Ss → SJ and hr ◦ h̄ ◦ ha+m : SJ → Ss are therefore Lipschitz
coarse inverses and hence quasi-isometries with constants depending on m.
Remark. The definition of the suspension space S depends ostensibly on the choice
of representative Φ ∈ Aut(Fn ) of φ, but in fact different choices of Φ produces suspension spaces which are Fn -equivariantly isometric, as can easily be checked using
the fact that distinct choices of Φ differ by an inner automorphism of Fn .
5.6
Proof of hyperbolicity of S.
We prove that S is a hyperbolic metric space by applying the Mj-Sardar combination
theorem [MS12], a descendant of the Bestvina–Feighn combination theorem [BF92].
62
The hypotheses of the Mj-Sardar theorem consist of an opening hypothesis and four
numbered hypotheses which we must check. The last of those four is the “flaring
condition” which we prove by application of Proposition 5.10.
Opening hypothesis of [MS12, Theorem 4.3]: A metric bundle. This hypothesis
says that S is a metric bundle over R with respect to the map p : S → R. We
must therefore verify that the map p : S → R satisfies the definition of a metric
bundle as given in [MS12, Definition 1.2]. First, p must be Lipschitz map, and each
fiber Ss must be a geodesic metric space with respect to the path metric induced
from S, each of which we have already verified. Also, item 2) of [MS12, Definition
1.2], when translated into our setting, requires that for each interval [a, b] ⊂ R such
that b − a ≤ 1, and for each s ∈ [a, b] and ξ ∈ Ss , there exists a path in S of uniformly
bounded length that passes through ξ and has endpoints on Sa and Sb respectively.
To obtain this path choose η ∈ Sa so that η · (s − a) = ξ and take the path t 7→ η · t
defined on a ≤ t ≤ b, whose length equals b − a ≤ 1.
The remaining verification needed for S to be a metric bundle is that the set of
inclusions Ss ֒→ S (s ∈ R) is uniformly proper, meaning that these inclusions are
uniformly Lipschitz — in fact they are all 1-Lipschitz by construction — and that
there exists a single nondecreasing “gauge” function δ : [0, ∞) → [0, ∞) for these
inclusions, having the following property:
(∗) for any s ∈ R, any x, y ∈ Ss , and any D ≥ 0, if dS (x, y) ≤ D then ds (x, y) ≤
δ(D).
To define δ consider x, y ∈ Ss connected by a geodesic path γ in S with Length(γ) =
dS (x, y) ≤ D. The projection p ◦ γ in R has length ≤ D and so, letting m =
⌊D + 2⌋, there are integers a and b = a + m such that Image(p ◦ γ) ⊂ [a, b], implying
that Image(γ) ⊂ S[a,b] which implies in turn that dS (x, y) = d[a,b] (x, y). Applying
Lemma 5.14 we have k1m ds (x, y) − cm ≤ d[a,b] (x, y) and hence ds (x, y) ≤ km dS (x, y) +
km cm . We may assume that km and cm are nondecreasing functions of m and hence
k⌊D⌋ and c⌊D⌋ are nondecreasing functions of D, and so using δ(D) = k⌊D⌋ (D + c⌊D⌋ )
we are done.
Hypotheses (1), (2) of [MS12, Theorem 4.3]: Base and fiber hyperbolicity. These
hypotheses require that the base space R is hyperbolic which is evident, and that the
fibers Ss are hyperbolic with uniform hyperbolicity constant. Proposition 5.9 gives
us a constant δ ′ ≥ 0 such that T ∗ is δ ′ -hyperbolic. Since each fiber Ss is λ-bilipschitz
homeomorphic to T ∗ , it follows that Ss is hyperbolic with a uniform hyperbolicity
constant δ depending only on δ ′ and λ.
Hypothesis (3) of [MS12, Theorem 4.3]: Barycenters. This hypothesis says that
the barycenter maps ∂ 3 Ss → Ss are uniformly coarsely surjective as s varies. We
review what this means from [MS12, Section 2 around the heading “The barycenter
map”]. Given a δ-hyperbolic geodesic metric space X with Gromov boundary ∂X,
consider the triple space
∂ 3 X = {(ξ1 , ξ2 , ξ3) ∈ (∂X)3 ξi 6= ξj if i 6= j}
63
The barycenter map ∂ 3 X → X is a coarsely well-defined map as follows. There
exists constants K ≥ 1, C ≥ 0 depending only on δ such that for any two points
ξ1 6= ξ2 ∈ ∂X there exists a K, C-quasigeodesically embedded line in X having
endpoints ξ1 , ξ2 ; we use the notation ξ1 ξ2 for any such quasigeodesic line. By [MS12,
Lemma 2.7] there exist constants D, L ≥ 0 depending only on δ such that for each
triple ξ = (ξ1 , ξ2, ξ3 ) ∈ ∂ 3 X there exists a point bξ ∈ X which comes within distance
D of any of the lines ξi ξj , i 6= j ∈ {1, 2, 3}, and for any other such point b′ξ the
distance between bξ and b′ξ is ≤ L. Once the constants K, C, D, L have been chosen,
any such point bξ is called a barycenter of ξ, and any map ∂ 3 X → X taking each
triple ξ to a barycenter bξ is called a barycenter map for X.
To say that the barycenter maps ∂ 3 Ss → Ss are uniformly coarsely surjective
means that there exists a “coboundedness constant” E ≥ 0 such that for each s ∈ R
the image of each barycenter map ∂ 3 Ss → Ss comes within distance E of each point
of Ss . For the hyperbolic space T ∗ , the action Fn y T ∗ has a fundamental domain τ ⊂
T ∗ of bounded diameter, and so E ′′ = Diam(τ ) is a coboundedness constant for any
Fn -equivariant barycenter map ∂ 3 T ∗ → T ∗ , hence there is a uniform coboundedness
constant E ′ for all barycenter maps ∂ 3 T ∗ → T ∗ . Since each of the fibers Ss comes
equipped with a λ-bilipschitz homeomorphism js : T ∗ → Ss , their barycenter maps
have a uniform coboundedness constant E = λE ′ .
Hypothesis (4) of [MS12, Theorem 4.3] aka [MS12, Definition 1.12]. Here is a
slight restatement of the hypothesis specialized to our present context.
Flaring 1: For all k1 ≥ 1, ν1 > 1, there exist integers A1 , R ≥ 0 such that for
any s ∈ R and any two k1 -quasigeodesics
γ1 , γ2 : [s − R, s + R] → S
which are sections of the projection map p — meaning that p ◦ γi is the identity
on the interval [s − R, s + R] — the following implication holds:
(F1 ) if ds (γ1 (s), γ2 (s)) ≥ A1 then
ν1 · ds γ1 (s), γ2(s) ≤ max{ds−R γ1 (s−R), γ2 (s−R) , ds+R γ1 (s+R), γ2(s+R) }
This statement tautologically implies the flaring hypothesis given in [MS12, Definition
1.12], the difference being that in the latter statement the quantifier order starts out
as “For all k ≥ 1 there exist µ > 1 and integers A, r ≥ 0 such that. . . ” with
the remainder of the statement unchanged (a simple geometric estimation argument
yields the converse implication, but we will not need this).
For proving Flaring 1 we first reduce it to a “discretized” version taking place
solely in the fibers Sr for integer values of r, as follows:
Flaring 2: For all k2 ≥ 1, ν2 > 1, there exist integers A2 , R ≥ 0 such that for any
integer m ∈ Z and any two k2 -quasigeodesic maps
δ1 , δ2 : {m − R, . . . , m, . . . , m + R} → S
64
which are sections of the projection map p, the following implication holds:
ν2 · dm
(F2 ) if dm (δ1 (m), δ2 (m)) ≥ A2 then
δ1 (m), δ2 (m) ≤ max{dm−R δ1 (m−R), δ2 (m−R) , dm+R δ1 (m+R), δ2 (m+R) }
To show that Flaring 2 implies Flaring 1, choose k1 , ν1 . Consider any integer R ≥ 0,
any s ∈ R, and any pair γ1 , γ2 : [s − R, s + R] → S of k1 -quasigeodesic sections of
the projection map p. Let m = ⌊s⌋ and let t = s − m. The semiflow restricts to
λ-bilipschitz homeomorphisms hr : Sr 7→ Sr+t defined by hr [x, r, 0] = [x, r, t] for any
integer r, having the property that the distance from each [x, r, 0] to its hr -image
[x, r, t] in S is at most 1. It follows that the functions
δj (r) = h−1
r (γj (r + t))
δj : {m − R, . . . , m + R} → S,
are k2 -quasi-isometric sections of p, where the constant k2 depends only on k1 and λ.
Applying Flaring 2, for any ν2 > 1 there exist integers A2 , R ≥ 0 such that the implication (F2 ) holds. Again using that the maps hr are λ-bilipschitz homeomorphisms, if
we take ν2 = ν1 · λ2 and A1 = A2 · λ then the hypothesis of (F1 ) implies the hypothesis
of (F2 ) and the conclusion of (F2 ) implies the conclusion of (F1 ).
It remains to verify Flaring 2 by applying Proposition 5.10 which we restate here
for convenience in a form matching that of Flaring 2:
Flaring 3: For each µ > 1, η ≥ 0 there exist integers A2 ≥ 0, R ≥ 1 such that for
any pair of η-pseudo-orbits Vr and Wr of fT ∗ : T ∗ → T ∗ defined for integers
−R ≤ r ≤ R, the following implication holds:
(F3 ) if d∗ (V0 , W0 ) ≥ A2 then
µ · d∗ (V0 , W0 ) ≤ max{d∗ (V−R , W−R ), d∗ (VR , WR )}
For the proof that Flaring 3 =⇒ Flaring 2, for each r ∈ Z we have a commutative
diagram
T∗
fT ∗
T∗
/
jr−1
jr
Sr−1
hr
/
Sr
where hr is the first hitting map. Note that in this diagram the bottom is equivariant,
the top is Φ-twisted equivariant, the left is a Φ1−r -twisted equivariant isometry, and
the right is Φ−r -twisted equivariant isometry.
Choose k2 ≥ 1, ν2 > 1, consider integers R ≥ 0 and m, and consider a pair of
k2 -quasigeodesic maps δi : {m − R, . . . , m + R} → S which are sections of p, for
i = 1, 2. It follows that if m − R ≤ r − 1 < r ≤ m + R then dS (δi (r − 1), δi (r)) ≤ 2k2
(recall that “k2 -quasigeodesic” is synonymous with “(k2 , k2 )-quasigeodesic”). For each
x ∈ Sr−1 we have dS (hr (x), x)) ≤ 1 and hence dS (hr (δi (r − 1)), δi (r)) ≤ 2k2 + 1. For
65
−1
−1
r ∈ {−R, . . . , +R} denote Vr = jm+r
(δ1 (m + r)), Wr = jm+r
(δ2 (m + r)). It follows
that d∗ (fT ∗ (Vr−1 ), Vr ), d∗ (fT ∗ (Wr−1 , Wr ) ≤ 2k2 + 1, and so the sequences (Vr ) and
(Wr ) are both η-pseudo-orbits of fT ∗ , defined for m − R ≤ r ≤ m + R and with
η = 2k2 + 1. Since jm is an isometry we have dm (δ1 (m), δ2 (m)) = d∗ (V0 , W0 ), and so
the hypothesis of (F2 ) implies the hypothesis of (F3 ). Similarly since jm−R and jm+R
are isometries, the conclusion of (F3 ) implies the conclusion of (F2 ) using µ = ν2 .
We have therefore proved that Flaring 3 =⇒ Flaring 2.
This completes the verification of the hypotheses of the Mj-Sardar combination
theorem, and so the space S is therefore hyperbolic.
6
Abelian subgroups of Out(Fn)
This section reviews some background material needed for the rest of the paper.
Section 6.1 contains basic material from [FH11] regarding automorphisms and outer
automorphisms (see also [HM17a] for a comprehensive overview). Section 6.2 reviews
elements of the theory of abelian subgroups of Out(Fn ) developed in [FH09], focussing
on disintegration subgroups.
6.1
6.1.1
Background review
More about CTs
In Notations 4.1 we reviewed features of a CT f : G → G with associated f -invariant
filtration G1 ⊂ · · · ⊂ Gu under the assumption that the top stratum Hu is EG. In
studying disintegration groups we shall need some further defining properties and
derived properties of CTs (for this section only the reader may ignore the assumption
that the top stratum is EG). We shall refer to [FH11] for material specific to CTs, to
[BFH00] for more general material, and to [HM17a] for certain “compilation” results
with multiple sources in [FH11] or [BFH00]. One may also consult [HM17a] for a
comprehensive overview.
General properties of strata. [FH11, Section 2.6] Each stratum Hi = Gi \Gi−1
is either an irreducible stratum meaning that its transition matrix Mi is an irreducible
matrix, or a zero stratum meaning that Mi is a zero matrix. Each irreducible stratum
satisfies one of the following:
Hi is an EG stratum: [FH11, Remark 3.20] The matrix Mi is a k × k PerronFrobenius matrix for some k ≥ 2, having eigenvalue λ > 1; or
Hi is an NEG stratum: [FH11, Section 4.1] Hi = Ei is a single edge with an
orientation such that f (Ei ) = Ei u where u is either trivial or a closed path in Gi−1
having distinct initial and terminal directions. An NEG stratum Hi is a fixed stratum
66
if u is trivial, a linear stratum if u is a Nielsen path (equivalently u is a periodic
Nielsen path), and a superlinear stratum otherwise.
Properties of NEG-linear strata: An NEG-linear stratum Hi = Ei will also be
referred to as a linear edge of G. The linear edges of G have the following features:
Twist path and twist coefficient: [FH11, Section 4.1] For each linear edge Ei we
have f (Ei ) = Ei widi for a unique closed Nielsen path wi which is root free meaning
that wi is not an iterate of any shorter closed path (equivalently, if p is the base point
of wi , then the element of the group π1 (G, p) represented by wi is root free). We say
that wi is the twist path of Ei and that the integer di 6= 0 is its twist coefficient. If
Ej 6= Ei is another linear edge having twist path wj , and if wi and wj determine the
same conjugacy class up to inversion, then wi = wj and di 6= dj .
NEG Nielsen paths: [FH11, Definition 4.7] For each NEG edge Ei , if there is an
indivisible Nielsen path contained in Gi but not in Gi−1 then Ei is a linear edge, and
every such Nielsen path has the form Ei wik E i for some k 6= 0.
Exceptional paths: [FH11, Definition 4.1] These are paths of the form Ei w k E j
where Ei 6= Ej are linear edges having the same twist path w = wi = wj and having
twist coefficients di , dj of the same sign.
Properties of EG strata: These properties were stated in Notations 4.1 with
respect to the top EG stratum Hu , but we go over them again here for an arbitrary
EG-stratum Hi , and with a somewhat different emphasis.
Lines: [BFH00, Section 2.2] Recall the spaces of lines in Fn and in G and the
canonical homeomorphism between them:
B(Fn ) = {2 point subsets of ∂Fn }/Fn
B(G) = {bi-infinite paths in G}/reparameterization
where the topology on B(Fn ) is induced by the Hausdorff topology on compact subsets
of ∂Fn , and the topology on B(G) has a basis element for each finite path γ consisting
of all bi-infinite paths having γ as a subpath. The homeomorphism B(Fn ) ↔ B(G)
e → G and the natural bijection ∂Fn ≈
is induced by the universal covering map G
e
∂ G. We refer to this homeomorphism by saying that a line in Fn is realized by the
corresponding line in G.
Attracting laminations: [BFH00, Section 3.1], [FH11, Remark 3.20] Associated
to Hi is its attracting lamination Λi ⊂ B(Fn ) which is the set of all lines in Fn whose
realization ℓ in G has the property that for each finite subpath γ of ℓ and each edge
E ⊂ Hi there exists k ≥ 1 such that γ is a subpath of f#k (E). For distinct EG
strata Hi , Hj (i 6= j) the corresponding laminations Λi , Λj are distinct. The set of
laminations L(φ) = {Λi } is independent of the choice of CT representative f : G → G.
EG Nielsen paths: [BFH00, Corollary 4.19], [HM17a, Fact 1.42] Up to inversion
there exists at most one indivisible periodic Nielsen path ρ contained in Gi but not
67
in Gi−1 . Its initial and terminal directions are distinct, and at least one endpoint of
ρ is not contained in Gi−1 .
Geometricity: [HM17a, Fact 2.3] Hi is a geometric stratum if ρ exists and is a
closed path.
Properties of zero strata: [FH11, Definition 2.18, Definition 4.4, Definition 4.7]
Each zero stratum Hi ⊂ G has the following properties:
Envelopment: There exist indices s < i < r such that Hs is an irreducible stratum,
Hr is an EG stratum, each component of Gr is noncontractible, and each Hj with
s < j < r is a zero stratum and a contractible component of Gr−1 . We say that the
zero strata Hj with s < j < r are enveloped by Hr , and we denote Hrz to be the union
of Hr with its enveloped zero strata. The filtration element Gs is the union of the
noncontractible components of Gr−1 , and Hrz = Gr \ Gs .
Taken paths. These are the paths µ in Hi for which there exists an edge E of some
irreducible stratum Hq with q > i, and there exists k ≥ 1, such that µ is a maximal
subpath in Hi of the path f#k (E); we say more specifically that the path µ is q-taken.
If Hr is the EG stratum that envelopes Hi then every edge E ⊂ Hi is an r-taken
path, from which it follows that the endpoints of E are vertices of Hr not contained
in Gr−1 .
Properties of complete splittings: [BFH00, Section 4], [FH11, Definition 4.4]
A splitting of a path γ in G is a concatenation expression γ = γ1 · . . . · γJ such
that for all k ≥ 1 we have f#k (γ) = f#k (γ1 ) · . . . · f#k (γJ ). The characteristic property
of a CT — short for “completely split relative train track map” — is the following:
Complete splitting: For each edge E there is a unique splitting f (E) = σ1 · . . . · σn
which is complete, meaning that each term σi is either an edge in an irreducible stratum, an EG indivisible Nielsen path, an NEG indivisible Nielsen path, an exceptional
path, or a taken path in a zero stratum.
6.1.2
Principal automorphisms and rotationless outer automorphisms.
Consider an automorphism Φ ∈ Aut(Fn ), with induced boundary homeomorphism
b : ∂Fn → ∂Fn , and with fixed subgroup denoted Fix(Φ) < Fn . Denote the
denoted Φ
b as Per(Φ)
b and Fix(Φ)
b ⊂ ∂Fn , respectively.
sets of periodic points and fixed points of Φ
b of period k ≥ 1. We say ξ is an attractor for Φ
b if it has a
Consider ξ ∈ Per(Φ)
k
ki
b (U) ⊂ U and the collection {Φ
b (U) i ≥ 0}
neighborhood U ⊂ ∂Fn such that Φ
b if it is an attractor for Φ
b −1 .
is a neighborhood basis of ξ. Also, ξ is a repeller for Φ
b and Fix(Φ)
b denote the sets of attracting, repelling, and nonrepelling
Within Per(Φ)
68
points, respectively, as
b Per− (Φ),
b PerN (Φ)
b ⊂ Per(Φ)
b
Per+ (Φ),
b Fix− (Φ),
b FixN (Φ)
b ⊂ Fix(Φ)
b
Fix+ (Φ),
For each c ∈ Fn associated inner automorphism ic (a) = cac−1 we use the following
special notations
∂c = Fix( bic ) = {∂− c, ∂+ c},
{∂− c} = Fix− ( bic ),
{∂+ c} = Fix+ ( bic )
The following equivalences are collected in [FH11, Lemma 2.1] based on results from
[GJLL98] and [BFH04].
Fact 6.1. For each Φ ∈ Aut(Fn ) and each nontrivial c ∈ Fn we have
b is invariant under îc
Φ ∈ Stab(c) ⇐⇒ ic commutes with Φ ⇐⇒ Fix(Φ)
b
⇐⇒ ∂c ⊂ Fix(Φ)
Principal automorphisms. [FH11, Definition 3.1]. We say that Φ is a principal
automorphism if FixN (Φ) ≥ 2, and furthermore if FixN (Φ) = 2 then FixN (Φ) is
neither equal to ∂c for any nontrivial c ∈ Fn , nor equal to the set of endpoints of a lift
of a generic leaf of an attracting lamination of the outer automorphism representing Φ.
For each φ ∈ Out(Fn ) let P (φ) ⊂ Aut(Fn ) denote the set of principal automorphisms
representing φ.
For each φ ∈ Out(Fn ) let P ± (φ) denote1 the set of all Φ ∈ Aut(Fn ) representing φ
such that either Φ or Φ−1 is principal, equivalently,
P ± (φ) = P (φ) ∪ (P (φ−1 ))−1
The symmetry equation P ± (φ−1 ) = (P ± (φ))−1 is useful in situations where one is
trying to prove a certain property of Φ ∈ P ± (φ) that is symmetric under inversion
of Φ: one may reduce to the case that Φ is principal by replacing φ ∈ Out(Fn ) and Φ
by their inverses in the case where Φ is not already principal. We use this reduction
argument with little comment in many places.
Rotationless outer automorphisms. [FH09, Definition 3.1 and Remark 3.2].
The concept of forward rotationless outer automorphisms is defined in [FH11], where
it is proved that the forward rotationless outer automorphisms are precisely those
outer automorphisms which have CT representatives. Here we use the stricter property of rotationless defined in [FH09], which is symmetric under inversion, and which
is better adapted to the study of abelian subgroups.
In [FH09] the definition of P ± (φ) was incorrectly stated as P (φ) ∪ P (φ−1 ). The definition given
here should replace the one in [FH09]. No further changes are required in [FH09], because the
arguments there which use P ± (φ) are all written using the current, correct definition.
1
69
We say that φ ∈ Out(Fn ) is rotationless if two conditions hold. First, for each
Φ ∈ P (φ) we have Per(Φ) = Fix(Φ) (in “forward rotationless” this condition is
replaced by the weaker PerN (Φ) = FixN (Φ)). Second, for each integer k ≥ 1 the map
Φ 7→ Φk is a bijection between P ± (φ) and P ± (φk ) — recall from [FH09, Remark 3.2]
that injective is always true, and so bijective holds if and only if surjective holds.
Expansion factor homomorphisms. [BFH00, Section 3.3] Consider φ ∈ Out(Fn )
and an attracting lamination Λ ∈ L(φ). Under the action of Out(Fn ) on the set of lines
B(Fn ), consider the subgroup Stab(Λ) < Out(Fn ) that stabilizes Λ. The expansion factor homomorphism of Λ is the unique surjective homomorphism PFΛ : Stab(Λ) → Z
such that for each ψ ∈ Stab(Λ) we have PFΛ (ψ) ≥ 1 if and only if Λ ∈ L(ψ). Furthermore, there exists µ > 1 such that if ψ ∈ Stab(Λ) and if PFΛ (ψ) > 1 then any
relative train track representative f : G → G of ψ has an EG-aperiodic stratum corresponding to Λ on which the Perron-Frobenius eigenvalue of the transition matrix
equals µPFΛ (ψ) .
Twistors (aka Axes). [FH11, Lemma 4.40 and the preceding page]. Recall that
two elements a, b ∈ Fn are said to be unoriented conjugate if a is conjugate to b or
to b−1 . The ordinary conjugacy class of a is denoted [a] and the unoriented conjugacy
class is denoted [a]u . For any marked graph G, nontrivial conjugacy classes of Fn
correspond bijectively with circuits S 1 7→ G up to orientation preserving homeomorphisms of the domain S 1 . Note that a is root free in Fn if and only if the oriented
circuit S 1 7→ G representing [a] does not factor through a nontrivial covering map of
S 1 , and a, b are unoriented conjugate if and only if the oriented circuits representing
[a], [b] differ by an arbitrary homeomorphism of S 1 .
Consider φ ∈ Out(Fn ) and a nontrivial, unoriented, root free conjugacy class
µ = [c]u , c ∈ Fn . For any two representatives Φ1 6= Φ2 ∈ Aut(Fn ) of φ, the following
are equivalent:
b 1 ) ∩ Fix(Φ
b 2)
Φ1 , Φ2 ∈ Stab(c) ⇐⇒ ∂c = Fix(Φ
d
Furthermore, if these equivalent conditions hold then Φ−1
2 Φ1 = ic for some integer
d 6= 0. If there exists a pair Φ1 6= Φ2 ∈ P (φ) such that the above equivalent conditions
hold, then µ is said to be a twistor for φ, and the number of distinct elements of the
set P (φ) ∩ Stab(c) is called the multiplicity of µ as a twistor for φ; these properties
are independent of the choice of c ∈ Fn representing µ.
The number of twistors of φ and the multiplicity of each twistor is finite, as follows.
First, for any CT f : G → G representing φ, for µ to be a twistor it is equivalent that
some NEG linear edge E twists around µ, meaning that E has twist path w such that
either w or w −1 represents µ. Furthermore, the multiplicity of µ is one more than the
number of linear edges that twist around µ [FH11, Lemma 4.40].
70
6.1.3
Rotationless abelian subgroups
Principal sets for rotationless abelian subgroups.
[FH11, Definition 3.9,
Corollary 2.13]. An abelian subgroup A < Out(Fn ) is rotationless if each of its elements is rotationless, equivalently A has a rotationless generating set. Every abelian
subgroup contains a rotationless subgroup of index bounded uniformly by a constant
depending only on n, namely the subgroup of k th powers where k ≥ 1 is an integer
such that the k th power of every element of Out(Fn ) is rotationless [FH11, Lemma
4.42].
Assuming A < Out(Fn ) is rotationless, a subset X ⊂ ∂Fn with ≥ 3 elements
is a principal set for A if each ψ ∈ A is represented by Ψ ∈ Aut(Fn ) such that
b (which determines Ψ uniquely amongst representatives of Ψ) and such
X ⊂ Fix(Ψ)
that Ψ ∈ P ± (ψ). When the principal set X is fixed, the map ψ 7→ Ψ defines a
homomorphism s : A → Aut(Fn ) that is a section of the canonical map Out(Fn ) →
Aut(Fn ) over the subgroup A. Also, the set
d
Y = ∩ψ∈A Fix(s(ψ))
is the unique principal set which is maximal up to inclusion subject to the property
that Y contains X ; this set Y defines the same section s : A → Aut(Fn ) as X [FH11,
Remark 3.10].
Comparison homomorphisms of rotationless abelian subgroups. [FH09,
Definition 4.1 and Lemma 4.3].
Consider a rotationless abelian subgroup A. Consider also two principal sets X1 , X2
for A that define lifts s1 , s2 : A → Aut(Fn ). Let Y1 , Y2 be the maximal principal
sets containing X1 , X2 respectively, and so s1 , s2 are also the lifts defined by Y1 , Y2 .
Suppose that s1 6= s2 : A → Aut(Fn ) (this if and only if Y1 6= Y2 ), and suppose also
that X1 ∩ X2 6= ∅ and hence Y1 ∩ Y2 6= ∅. It follows that the set Y1 ∩ Y2 is fixed by
distinct automorphisms representing the same element of A, and so Y1 ∩ Y2 = Tc± for
some nontrivial root free c ∈ Fn . In this situation there is an associated comparison
homomorphism ω : A → Z which is characterized by the equation
s2 (ψ) = iω(ψ)
s1 (ψ) for all ψ ∈ A.
c
The number of distinct comparison homomorphisms A → Z is finite.
The coordinate homomorphism. [FH09, Lemma 4.4, Definition 4.5, Lemma 4.6].
Every abelian subgroup A < Out(Fn ) is a finite lamination group, that is, the set
L(A) = ∪φ∈A L(φ) is finite. If in addition A is rotationless then each Λ ∈ L(A) is
ψ-invariant for all ψ ∈ A, and so
\
A<
Stab(Λ)
Λ∈L(A)
71
We thus obtain a finite collection of expansion factor homomorphisms defined on A,
namely
PFΛ : A → Z, Λ ∈ L(A)
Choosing an enumeration Ω1 , . . . , ΩN of the expansion homomorphisms and the comparison homomrophisms, we obtain the coordinate homomorphism
Ω : A → ZN
which is injective.
6.2
Disintegration subgroups
In this section we fix a rotationless φ ∈ Out(Fn ) and a CT representative f : G → G
with corresponding f -invariant filtration G1 ⊂ · · · ⊂ Gu = G (that is all we need
from Notations 4.1 for this section).
Using the CT structure of f (see Section 6.1.1) we build up to the definition of the
disintegration subgroup associated to f , a certain abelian subgroup D(f ) < Out(Fn )
constructed in [FH09, Section 6] by an explicit combinatorial procedure that we review
here, using which we also obtain a description of the coordinate homomorphism Ω :
D(f ) → ZN (see Section 6.1.3). Noting that D(f ) depends on the choice of f , all
definitions in this section depend on f .
6.2.1
QE-paths and QE-splittings
We describe structures associated to the collection of linear edges of G, augmenting the structures described under the heading “Properties of NEG-linear strata” in
Section 6.1.1.
[FH11, Def. 4.7] Two linear edges Ei , Ej ⊂ G with a common twist path w are
said to be linearly equivalent, abbreviated to LIN-equivalent, and the associated LINequivalence class is denoted LINw . Recall that distinct elements Ei 6= Ej ∈ LINw
have distinct twist coefficients di 6= dj .
[FH09, Lemma 6.1 and Def. 6.2] Given a twist path w and distinct linear
edges Ei 6= Ej ∈ LINw , a path of the form Ei w p E j is called a quasi-exceptional path
or QE-path (it is an exceptional path if and only if the twist coefficients di , dj have
the same sign). For every completely split path σ, any QE-subpath Ei w p E j of σ is a
concatenation of terms of the complete splitting of σ: either di , dj have the same sign
and Ei w p E j is an exceptional path and hence a single term; or the terms consist of
the initial Ei , the terms of the complete splitting of w p , and the terminal Ej . No two
QE-subpaths can overlap in an edge, and so there is a uniquely defined QE-splitting of
σ which is obtained from the complete splitting by conglomerating each QE-subpath
of σ into a single term.
72
Consider a twist path w. For each Ei , Ej ∈ LINw , the associated quasi-exceptional
family is the set of paths
Ei w ∗ E j = {Ei w p E j p ∈ Z}
Also, associated to w itself is its linear family, the set
[
LINw ∪
Ei w ∗ E j
Ei 6=Ej ∈LINw
Every quasi-exceptional path belongs to a unique quasi-exceptional family, and every
NEG linear edge or quasi-exceptional path belongs to a unique linear family.
6.2.2
Almost invariant subgraphs
The intuition behind the “almost invariant subgraphs” of G is that they are a partition
of the collection of non-fixed strata determined by the following requirement: if one
perturbs f : G → G by letting its restrictions f Hi to non-fixed strata be replaced
by iterates f ai Hi with varying integer exponents ai ≥ 0, and if one wishes to do
this perturbation so that the resulting outer automorphisms commute with the outer
automorphism represented by f , then the exponents ai should be constant on the
edges contained in each almost invariant subgraph.
[FH09, Def. 6.3] Define an equivalence relation on the non-fixed irreducible strata
{Hi } of G as follows. Given Hi , a path µ is a short path for Hi if µ an edge in Hi ,
or if Hi is EG and µ is a taken connecting path in a zero stratum enveloped by Hi .
Define a relation amongst the non-fixed irreducible strata denoted Hi → Hj , meaning
that there exists a short path µ for Hi such that some term of the QE-splitting of
f# (µ) is an edge of Hj ; note that if Hi → Hj → · · · → Hk then i ≥ j ≥ · · · ≥ k.
Let B be the directed graph whose vertex set is the set of non-fixed irreducible
strata, with a directed edge for each relation Hi → Hj . Let B1 , . . . , BS denote the
connected components of the graph B. For each s = 1, . . . , S define the almost
invariant subgraph Xs ⊂ G to be the union of the strata Hi comprising the vertices of
Bs , together with any zero stratum enveloped by one of these Hi . Note that the set
of almost invariant subgraphs {X1 , . . . , XS } partitions the set of the nonfixed strata.
6.2.3
Admissible S-tuples; quasi-exceptional families
[FH09, Defn. 6.6, Lemma 6.7, Defn. 6.8] For each S-tuple a = (a1 , . . . , aS ) of
non-negative integers define f a : G → G on each edge E ⊂ G as follows:2 if E ⊂ Xs
for some s then f a (E) = f#as (E); otherwise E is fixed by f and then f a (E) = E. Each
such f a is a homotopy equivalence representing an outer automorphism denoted φa .
By construction f a preserves the given filtration G1 ⊂ · · · ⊂ Gu = G.
2
In [FH09] the notation fa was used for what we here denote f a .
73
One would like to arrange (among other things) that (f a )k# (E) = f#as k (E) for all
k ≥ 1, all s = 1, . . . , S, and all edges E ⊂ Xs . This need not hold for general S-tuples,
but it does hold when a certain relation amongst linear edges and QE-splitting terms
is satisfied.
Given an almost invariant subgraph Xr and linear edges Ei , Ej , we say that the
triple (Xr , Ei , Ej ) is quasi-twist related if there exists a stratum Hk ⊂ Xr , a short
path µ for Hk , a twist path w, and an integer p ≥ 1 such that Ei , Ej ∈ LINw , and
such that Ei w p E j is a term in the QE-splitting of f# (µ).
We say that an S-tuple a is admissible if for all quasi-twist related triples (Xr , Ei , Ej ),
letting di , dj be the twist coefficients of Ei , Ej respectively, and letting Xs , Xt be
the almost invariant subgraphs containing Ei , Ej respectively, the following equation
holds:
ar (di − dj ) = as di − at dj
[FH09, Notation 6.12] Consider an almost invariant graph Xr . For each pair of
linear edges Ei , Ej such that the triple (Xr , Ei , Ej ) is quasi-twist related, the associated quasi-exceptional family is defined to be the set of all paths of the form Ei w ∗ E j .
We let Qr denote the set of all quasi-exceptional families associated to quasi-twist
related triples (Xr , Ei , Ej ).
QE-equivalence of linear edges. We say that a pair of linear edges Ei , Ej is
QE-related if there exists an almost invariant subgraph Xr such that (Xr , Ei , Ej ) is
quasi-twist related. Equivalently Ei , Ej are QE-related if and only if they are in the
same linear family and, letting w be the unique twist path for which Ei , Ej ∈ LINw ,
there exists r such that the family Ei w ∗ E j is in the set Qr .
The equivalence relation on linear edges generated by being QE-related is called
QE-equivalence and is written ∼QE . Note that the QE-equivalence relation amongst
linear edges is a refinement of the LIN-equivalence relation.
Note that if an exceptional path Ei w p E j occurs as a term in the complete splitting
of some iterate of some short path in Xr then the quasi-exceptional family Ei w ∗ E j is
an element of Qr and the linear edges Ei and Ej are QE-related.
6.2.4
Xs paths.
[FH09, Notation 6.12, Corollary 6.14] For each almost invariant subgraph Xs ,
we use the terminology Xs -paths to refer to the subpaths of G which form the elements
of the set Ps that is defined in [FH09, Notation 6.12] — namely, the completely split
paths γ such that each term of the QE-splitting of γ is one of the following:
(1) a Nielsen path; or
(2) a short path in a stratum contained in Xs , for example any edge in Hiz for any
EG stratum Hi ⊂ Xs ; or
(3) a quasi-exceptional path in the family Qs .
74
Furthermore, for any admissible S-tuple a, any almost invariant subgraph Xs , and
any Xs path γ, we have the following:
(4) Each iterate f#k (γ) is also an Xs path;
(5) f#a (γ) = f#as (γ).
6.2.5
The disintegration subgroup D(f )
Here we recall the definition of the disintegration subgroup D(f ) < Out(Fn ). We also
recall the “admissible semigroup” S(f ), which we will use as an aid to understanding
properties of D(f ).
The subgroup D(f ). [FH09, Defn. 6.10, Cor. 6.16, Lem. 6.18, Cor. 3.13]
This is the subgroup of Out(Fn ) generated by the set of elements
{φa
a is an admissible S-tuple}
The subgroup D(f ) is abelian and rotationless. The dependence of D(f ) on f rather
than just φ was suppressed in [FH09, Definition 6.10] where the notation D(φ) was
used; see [FH09, Example 6.11].
The admissible semigroup S(f ). [FH09, Corollary 7.6 and its proof]
Let S(f ) ⊂ ZS denote the set of admissible S-tuples, which forms a sub-semigroup
of ZS . Let L(f ) < ZS be the subgroup generated by S(f ). The map S(f ) 7→ D(f )
defined by a 7→ φa is an injective semigroup homomorphism, and it extends to an
isomorphism L(f ) 7→ D(f ). Every element of L(f ) can be written as the difference
a − b of two elements a, b ∈ S(f ), and so every ψ ∈ D(f ) can be written in the form
ψ = (φb )−1 φa for some a, b ∈ S(f ).
We record here a simple consequence of these definitions for later use:
Fact 6.2. The function which assigns to each a ∈ S(f ) the map f a : G → G satisfies
the following “homomorphic” property:
f#a (f#b (E)) = f#a+b (E) for each a, b ∈ S(f ) and each edge E ⊂ G
Proof. If E is a fixed edge of f then both sides equal E. Otherwise E is an edge in
some almost invariant graph Xs , hence E is an Xs -path, hence f#b (E) = (f bs )# (E) is
an Xs path, hence both sides equal (f as +bs )# (E).
We repeat here the theorem cited in Section 3:
Disintegration Theorem ([FH09, Theorem 7.2]). For every rotationless abelian
subgroup H < Out(Fn ) there exists φ ∈ H such that for every CT f : G → G
representing φ, the intersection H ∩ D(f ) has finite index in H.
75
Remark. In this statement of the Disintegration Theorem we have made explicit
the fact that [FH09, Theorem 7.2] holds for any choice of CT representing φ. This
was already implicit in the notation D(φ) used for disintegration groups in [FH09,
Definition 6.10] where any representative CT is allowed.
6.2.6
The coordinate homomorphism of a disintegration group D(f )
The disintegration group D(f ), being rotationless and abelian, has an injective coordinate homomorphism Ω : D(f ) 7→ ZN as defined at the end of Section 6.1.3, whose
coordinate functions are all expansion factor homomorphisms and comparison homomorphisms of D(f ). We review here how one may use the structure of the CT f
to pick out a subset of the coordinate functions of Ω — consisting of all the expansion factor homomorphisms but only a subset of the comparison homomorphisms —
and, after scaling the expansion factor homomorphisms, to obtain a homomorphism
denoted Ωf : D(f ) → ZI which is still injective.
[FH09, Lemmas 7.4 and 7.5 and preceding paragraphs].
Let I = {i Hi is either EG or NEG linear}. The homomorphism Ωf : D(f ) → ZI
will have coordinate functions denoted ωi : D(f ) → Z (i ∈ I), defined in two cases
depending on whether Hi is EG or NEG linear. Let Xs be the almost invariant
graph containing Hi . In the case that Hi = Ei is NEG linear with twist path w
and twist coefficient di , and so f (Ei ) = Ei w di , then we set ωi : D(f ) → Z to be the
homomorphism whose value on φa for each admissible S-tuple a is given by
ωi (φa ) = as di
In the case that Hi is EG then, letting λ be the Perron-Frobenius eigenvalue of
the transition matrix of f on Hi , the homomorphism ωi : D(f ) → Z is the unique
homomorphism such that ωi (φa ) = λas for each admissible S-tuple a. Letting Λi ∈
L(φ) be the attracting lamination of φ associated to the EG stratum Hi , we note
that ωi is a scaled copy of PFΛi , meaning that ωi (ψ) = m · PFΛ (ψ) for all ψ ∈ D(f ),
where m = PFΛi (φ).
We also define comparison homomorphisms ωi,j : D(f ) → Z, one for each LINequivalent pair of linear edges Ei , Ej , using the following formula:
ωi,j = ωi − ωj
The following properties hold, in particular the first two which relate the functions
Ωf : D(f ) → ZI and Ω : D(f ) → ZN :
• The coordinate functions ωi of Ωf for which Hi is an EG stratum correspond
one-to-one with the expansion factor coordinate functions PFΛ of Ω. Under this
correspondence, ωi is a scaled copy of PFΛ as explained above.
• The set consisting of the coordinate functions ωi of Ωf for which Hi is an NEG linear
stratum together with the functions ωi,j for all LIN-equivalent pairs of NEG linear
76
strata Hi , Hj is exactly the set of comparison homomorphism coordinate functions
of Ω.
• The homomorphism Ωf is injective.
• For each i ∈ I and each admissible S-tuple a we have ωi (φa ) = as ωi (φ).
7
A train track semigroup action
Throughout this section we continue to adopt Notations 4.1 regarding a rotationless
φ ∈ Out(Fn ) with a CT representative f : G → G whose top stratum Hu is EG with
associated attracting lamination Λ.
Consider the disintegration group D = D(f ), reviewed in Section 6.2. Recall
that each element of D stabilizes Λ. We shall focus on the subsemigroup D+ < D
consisting of all ψ ∈ D for which the value of top coordinate homomorphism ω(ψ) is
non-negative, meaning that the asymptotic stretch factor of ψ on Λ is ≥ 1, and so
either ψ does not stretch Λ or Λ is an attracting lamination of ψ.
Using the theory of disintegration groups, for each ψ ∈ D+ we construct a topological representative f ψ : G → G whose action on the edges of Hu agrees with the
ω(ψ)
action of the appropriate iterate of the CT f : G → G, namely f# . Then, by lifting
e projecting to the tree T ,
these topological representatives to the universal cover G,
∗
and extending to the coned off graph T , we obtain semigroup actions of D+ on T
and on T ∗ , detailed properties of which are described in Section 7.2. An important
feature of the action D+ y T is that it is not an isometric action, in fact the action
of each Ψ ∈ D+ will lengths in T by a uniform amount depending on the appropriate
value of the coordinate homomorphism ω (see Section 7.2 (4)). On the other hand,
by restricting the semigroup actions D+ y T, T ∗ to the subgroup D0 we obtain true
isometric actions D0 y T, T ∗ , some properties of which are studied in Section 7.3.
The actions constructed in this section will be the basis for the construction in
Section 8 of an isometric group action of D on the hyperbolic suspension complex S.
For use throughout Sections 7 and Section 8 we establish some further notations.
Notations 7.1. Given a rotationless φ ∈ Out(Fn ) with CT representative f : G → G
having associated filtration G1 ⊂ · · · ⊂ Gu with its top stratum Hu being EG, and
with all associated objects and notations as described in Notations 4.1, we have the
following additional objects and notations:
(1) [Section 6.1.1, “Properties of zero strata”] Huz is the union of Hu with all zero
strata enveloped by Hu . For some t ≤ u − 1 this set of zero strata has the form
{Hi t + 1 ≤ i ≤ u − 1} (which is empty if t = u − 1), where
(a) Ht is the highest irreducible stratum below Hu ,
(b) Gt is the union of noncontractible components of Gu−1 , the contractible
components being the zero strata Hi , t + 1 ≤ i ≤ u − 1.
77
(2) [Definitions 5.1, 5.3] Fn y T denotes the free splitting corresponding to the
marked graph pair (G, Gu−1 ), with associated Nielsen set {Nj }j∈J , each Nj
having basepoint set Zj ⊂ Nj . Also, Fn y T ∗ denotes the action obtained
from Fn y T by coning off each basepoint set Zj to a cone point Pj with edges
Pn Q (Q ∈ Zj ).
(3) [Definition 5.2] Associated to each automorphism Φ ∈ Aut(Fn ) representing φ
are unique Φ-twisted equivariant maps as follows:
e → G,
e a lift of f ;
(a) f˜Φ : G
e→
(b) fTΦ : T → T , induced by f˜Φ with respect to the collapse map G
7 T.
(c) fT ∗ : T ∗ → T ∗ , an extension of fTΦ that permutes cone points and cone
edges.
The maps fTΦ , fTΦ∗ satisfy the “Stretch properties” recorded in Section 5.3.
(4) The disintegration group of f is denoted D = D(f ). Its full pre-image in Aut(Fn )
b and is called the extended disintegration group.
is denoted D
(5) Associated to the top EG stratum Hu we have the following objects:
• The almost invariant subgraph Xs ⊂ G containing Huz ;
• The coordinate homomorphism ω : D → Z associated to Xs , which is a
scaled copy of the expansion factor homomorphism PFΛ ;
b → Z, obtained by pre-composing
• The lifted coordinate homomorphism ω
b:D
b 7→ D.
ω with the projection map D
• The kernels of these homomorphisms denoted
b0 = ker(b
D
ω)
D0 = ker(ω)
We thus have a commutative diagram with short exact rows:
1
/
b0
ker(b
ω) = D
b
D
/
O
⊂
Fn ≈ Inn(Fn )
1
/
/
ω
b
/Z
O
ω
b
D
/
D
⊂
1
/
Fn ≈ Inn(Fn )
/
Aut(Fn )
/
1
/
⊂
Out(Fn )
(6) Letting [0, ∞) = {0, 1, 2, . . .} we have subsemigroups D+
as follows:
D0 = ω −1 (0) < D+ = ω −1 [0, ∞) <
b0 = ω
b+ = ω
D
b −1 (0) < D
b −1 [0, ∞) <
78
1
/
/
1
b+ and inclusions
and D
D
b
D
7.1
A “homotopy semigroup action” of D+ on G.
b+ y T , in this section we
To prepare for the construction of the semigroup action D
work downstairs in G and construct a “homotopy semigroup action” of D+ on T .
What this means will be clear from the construction, but the intuitive idea is we
associate a topological representative f ψ : G → G to each ψ ∈ D+ so that although
′
′
the action equation f ψ ◦f ψ = f ψψ does not hold exactly, it does hold “up to homotopy
relative to Gt ”. The values of f ψ on edges of Hnz are determined by appropriate
iterates of f itself, and the values on Gt are determined, up to homotopy, by the
“graph homotopy principle” of Section 4.6.
Letting Vert(G) denote the set of vertices of G, we define subsets
P ⊂ F ix ⊂ V ⊂ Vert(G)
as follows. First, V = Vert(G) ∩ Hu = Vert(G) ∩ Huz , the latter equation holding
because each edge E ⊂ Huz \ Hu has both endpoints in Hu [FH11, Definition 4.10
(Zero Strata)]. Next, P = Huz ∩ Gt = Hu ∩ Gt ; note that P ⊂ G because Huz and Gt
are subgraphs having no edges in common; also P is the frontier both of Huz and of
Gt in G, G = Huz ∪ Gt . Finally, F ix denotes the set of points in V fixed by f , and
the inclusion P ⊂ F ix follows from [FH11, Remark 4.9].
Here is a summary of the construction of the “homotopy semigroup action”:
(1) For each ψ ∈ D+ we define a topological representative f ψ : G → G such that
the following hold:
(a) We have f IdD+ = IdG (where IdD+ ∈ D+ denotes the identity group element, and IdG : G → G the identity map).
(b) f ψ (Gu−1 ) ⊂ Gu−1 and f ψ (Gt ) ⊂ Gt .
ω(ψ)
(c) For each edge E ⊂ Huz we have f ψ (E) = f#
then f ψ Huz is the identity.
(E). In particular, if ψ ∈ D0
(d) f ψ V = f ω(ψ) V . In particular f ψ (V ) ⊂ V and f ψ fixes each point of
F ix and of its subset P .
(e) If a height u indivisible Nielsen path ρ exists then f ψ fixes each endpoint
of ρ and f#ψ (ρ) = ρ.
′
(2) For each ψ, ψ ′ ∈ D+ we define a homotopy hψ,ψ : G × [0, 1] → G between
′
′
f ψ ◦ f ψ and f ψψ such that the following hold:
′
(a) For each edge E ⊂ Hu the homotopy hψ,ψ straightens the edge path
′
′
f ψ ◦f ψ E relative to its endpoints to form the path f ψψ E by cancelling
only edges of Gu−1 , without ever cancelling any edges of Hu .
′
(b) hψ,ψ is stationary on V .
79
′
(c) hψ,ψ preserves Gt .
Recall that a homotopy h : A × [0, 1] → A “preserves” a subset B ⊂ A means that
h(B × [0, 1]) ⊂ B, and for h to be “stationary” on B means that h(x × [0, 1]) is
constant for each x ∈ B.
′
First we will construct the maps f ψ and then the homotopies hψ,ψ , along the way
proving the various requirements of (1) and (2).
Constructing f ψ . We being with the construction of ftψ = f ψ Gt : Gt → Gt
by applying the graph homotopy principle Lemma 4.14. Recall from Section 6.2.5
the abelian semigroup of admissible S-tuples S(f ). Recall also the injection S(f ) ≈
L(f ) < D defined by a 7→ φa with topological representative f a : G → G; this injection is a semigroup homomorphism whose image L(f ) generates D. Since the
restricted map fta = f a
Gt : Gt → Gt is a homotopy equivalence that fixes
each point of P , by Lemma 4.14 (1) the map of pairs fta : (Gt , P ) → (Gt , P )
is a homotopy equivalence in the category of pairs, and hence its homotopy class
rel P is an element [fta ] of Autgp
0 (Gt , P ), the pure automorphism group of (Gt , P )
in the graph point category (Lemma 4.14 (4)). By Fact 6.2, for each a, b ∈ S(f )
the maps fta ◦ ftb and fta+b : (Gt , P ) → (Gt , P ) are homotopic rel P , and so the
a
a
map A : L(f ) 7→ Autgp
0 (Gt , P ) defined by A(φ ) = [ft ] is a semigroup homomorphism. Since the commutative group D is generated by its subsemigroup L(f ), a
simple semigroup argument shows that A extends uniquely to a group homomorψ
phism A : D → Autgp
0 (Gt , P ). For each ψ ∈ D choose ft : (Gt , P ) → (Gt , P ) to be
a representative of A(ψ); if ψ = φa is already in L(f ) then we choose ftψ = fta , and
so if ψ is the identity then ftψ is the identity. Notice that no straightening is carried
out when ftψ is applied, and so there is no need for the “#” operator in the definition
of ftψ .
(φa )−1
: (Gt , P ) → (Gt , P ), which
For later use, for each a ∈ S(f ) we denote f¯ta = ft
is the homotopy inverse of fta that represents A((φa )−1 ) = A(φa )−1 ∈ Autgp
0 (Gt , P ).
ψ
For each ψ ∈ D+ we may now define f : G → G as follows:
ω(ψ)
f ψ (E) = f#
(E) for each E ⊂ Huz , and
f ψ Gt = ftψ
If ψ ∈ D is the identity then clearly f ψ : G → G is the identity, verifying (1a). By
construction f ψ satisfies (1b) and (1c). For item (1d), note first that for each x ∈ V
there exists an oriented edge E ⊂ Hu with initial vertex x, and for each such E the
ω(ψ)
initial vertex of the path f ψ (E) = f# (E) equals f ω(ψ) (x); and if in addition x ∈ P
then both of f ω(ψ) and ftψ fix x. This shows that f ψ is well-defined on V , and that it
restricts to f ω(ψ) on V , completing the proof of item (1d); it also follows that f ψ is
continuous.
The proof that f ψ is a homotopy equivalence and is a topological representative
of ψ will be delayed to the very end.
80
Xs -paths under f ψ . Item (1e) is encompassed in the following generalization
of (1c), which will also be used for item (2):
ω(ψ)
(3) For each ψ ∈ D+ and each Xs -path γ with endpoints in V we have f#ψ (γ) = f#
(γ).
To see why this holds, the general Xs -path γ with endpoints in V is a concatenation
of three special types, and it suffices to check (3) only for these types:
Type (a): edges in Huz ;
Type (b): Xs -paths in Gt having endpoints in the set P .
Type (c): a height u indivisible Nielsen path of f ;
Type (a) is the special case handled already in item (1c).
If γ is of type (b), first note that for a ∈ S(f ) and ψ = φa ∈ L(f ) we have
ω(ψ)
f#ψ (γ) = (ftψ )# (γ) = (fta )# (γ) = f#a (γ) = f#as (γ) = f#
(γ)
where the second-to-last equation follows from Section 6.2.4 (5). For more general
ψ ∈ D+ , choose a, b ∈ S(f ) so that ψ = (φb )−1 φa , and note that as − bs = ω(φa ) −
ω(φb) = ω(ψ) ≥ 0. In the group Autgp (Gt , P ) we have the equation [ftψ ] = [f¯tb ] [fta ]
and so we may calculate
f#ψ (γ) = (ftψ )# (γ) = (f¯tb ◦ fta )# (γ) = (f¯tb )# ((fta )# (γ))
= (f¯tb )# (f#a (γ)) = (f¯tb )# (f#as (γ))
ω(ψ)
(γ)))
= (f¯b )# (f bs (f as −bs (γ))) = (f¯b )# (f b (f
t
#
t
#
#
#
ω(ψ)
ω(ψ)
= (f¯tb )# ((ftb)# (f# (γ))) = (f¯tb ◦ ftb )# (f# (γ))
ω(ψ)
= f#
(γ)
where the second equations of the second and third lines follow from Section 6.2.4 (4)
and (5).
For type (c), let γ be a height u indivisible Nielsen path of f . We may write γ as
an alternating concatenation of nontrivial paths of the form
γ = η0 µ1 η1 · · · ηK−1 µK ηK
where each ηk is a path in Huz and each µk is a Nielsen path of f in Gt with endpoints
in P [FH11, Lemma 4.24]. By definition of f ψ we have
ω(ψ)
f#ψ (γ) = [f#
ω(ψ)
(E1 ) (ftψ )# (µ1 ) f#
ω(ψ)
(η1 ) · · · f#
ω(ψ)
(ηK−1 ) (ftψ )# (µK ) f#
(ηK )]
We claim that each µk is a Nielsen path of ftψ for each ψ ∈ D, and so (ftψ )# (µk ) = µk .
To prove this claim, it holds if ψ = φa ∈ L(f ) for some a ∈ S(f ), because in that case
81
the left hand side equals (f a )# (µk ) = µk . Using that [f¯tψ ] = [ftψ ]−1 ∈ Autgp (G, P ),
we have µk = (f¯tψ )# (ftψ (µk )) = (f¯tψ )# (µk ), and so the claim holds if ψ = (φa )−1 for
some a ∈ S(f ). The general case holds because ψ = (φb )−1 φa for some a, b ∈ S(f ).
Applying the claim we have:
ω(ψ)
(E1 ) µ1 f#
ω(ψ)
(E1 ) f#
f#ψ (γ) = [f#
= [f#
=
ω(ψ)
ω(ψ)
ω(ψ)
(η1 ) · · · f#
ω(ψ)
(µ1 ) f#
ω(ψ)
(ηK−1) µK f#
ω(ψ)
(η1 ) · · · f#
(ηK )]
ω(ψ)
(ηK−1 ) f#
ω(ψ)
(µK ) f#
(ηK )]
ω(ψ)
f# (γ)
This complete the proof of (3).
This also completes the construction of f ψ : G → G and the proof of (1) for each
ψ ∈ D+ , except that we will delay further the proof that f ψ is a homotopy equivalence
and a topological representative of ψ.
′
Constructing hψ,ψ . Given ψ, ψ ′ ∈ D+ we turn to the construction of the ho′
′
′
motopy hψ,ψ : G × [0, 1] → G from f ψ ◦ f ψ to f ψψ . Let θ = ψψ ′ ∈ D+ . First,
gp
using the homomorphism A : D → Aut0 (G, P ) we have A(ψψ ′ ) = A(ψ)A(ψ ′) which
′
′
′
translates to [ftψψ ] = [ftψ ] [ftψ ] and so there exists a homotopy rel P from ftψ ◦ ftψ to
′
′
ftψψ denoted hψ,ψ
: Gt × [0, 1] → Gt . Second, for each E ⊂ Huz we have
t
′
ω(ψ′ )
f#ψ (f#ψ (E)) = f#ψ (f#
ω(ψ)
(E)) = f#
ω(ψ′ )
(f#
ω(ψψ′ )
(E)) = f#
′
(E) = f#ψψ (E)
ω(ψ′ )
where the second equation holds by applying (3) together with the fact that f# (E)
′
is an Xs -path (Section 6.2.4 (4)). Putting these together, we may define hψ,ψ so that
for each edge E ⊂ Huz its restriction to E × [0, 1] is a homotopy rel endpoints from
′
′
′
f ψ ◦ f ψ E to f ψψ E, and its restriction to Gt × [0, 1] is equal to hψ,ψ
. Items
t
(2b) and (2c) are evident from the construction. Item (2a) follows the definition of a
relative train track map, which tells us that for each edge E ⊂ Hu and each integer
i ≥ 0 the path f#i (E) is u-legal, and that for each u-legal path γ the homotopy that
straightenes f i (γ) to produce f#i (γ) cancels no edges of Hu .
Topological representatives. For each ψ ∈ D+ , since L(f ) generates D we
may choose a, b ∈ S(f ) so that ψ = (φb )−1 φa , and hence φb ψ = φa . Since all three
b
of φb , ψ, φa are in D+ , we have a homotopy hφ ,ψ from f b ◦ f ψ to f a . Since f a , f b
are homotopy equivalences it follows that f ψ is also, and since f a , f b are topological
representatives of φa , φb respectively it follows that f ψ is a topological representative
of (φb )−1 φa = ψ.
7.2
b+ on T and T ∗
Semigroup actions of D
b+ y T , deriving various properties
We turn now to the construction of the action D
b+ y T ∗ .
of the construction, and then we extend the action to D
82
b+ we define a map f Ψ : T → T by carrying out the
Associated to each Ψ ∈ D
following steps. First let ψ ∈ D+ be the image of Ψ under the homomorphism
Aut(Fn ) 7→ Out(Fn ), and consider the map f ψ : (G, Gu−1 ) → (G, Gu−1 ), part of the
e G
eu−1 ) →
“homotopy semigroup action” constructed in Section 7.1. Let f˜Ψ : (G,
e G
eu−1 ) be the unique Ψ-twisted equivariant lift of f ψ to G.
e Let f Ψ : T → T be the
(G,
eu−1
Ψ-twisted equivariant map induced from f˜Ψ by collapsing each component of G
Ψ
to a point and then straightening on each edge E ⊂ T so that f
E stretches length
by a constant factor.
We record several properties of this semigroup action:
(1) Twisted equivariance: The map f Ψ is Ψ-twisted equivariant.
′
′
b+ .
(2) Semigroup action property: f Ψ ◦ f Ψ = f ΨΨ for all Ψ, Ψ′ ∈ D
′
′
Property (2) follows from the fact that f Ψ (f Ψ (E)) = f ΨΨ (E) for each edge E ⊂ T ,
which is a consequence of Section 7.1 item (2a) applied to edges of Hu .
(3) Vertex Action Property: f Ψ takes vertices to vertices, and it restricts to a
bijection of the set of vertices having nontrivial stabilizer.
For the proof, denote vertex sets by V (G) ⊂ G and V (T ) ⊂ T , and let V nt (T ) denote
the subset of all v ∈ V (T ) such that StabFn (v) is nontrivial. By Section 7.1 items (1d)
and (1b), the map f ψ takes the set Gu−1 ∪ V (G) to itself, and since f ψ is a topological
representative it follows that f ψ restricts further to a homotopy equivalence amongst
eu−1 ⊂ G
e and V (G)
e ⊂G
e
the noncontractible components of Gu−1 ∪ V (G). Letting G
e→G
e restricts to a self-map
be the total lifts of Gu−1 and V (G), it follows that f˜Ψ : G
eu−1 ∪ V (G),
e and it restricts further to a bijection amongst
of the components of G
e → T , the set
the components having nontrivial stabilizer. Under the quotient map G
e
e
of components of Gu−1 ∪ V (G) corresponds bijectively and Fn -equivariantly to the
vertex set V (T ). It follows that f Ψ : T → T restricts to a self map of V (T ), and that
it restricts further to a bijection of V nt (T ), proving property (3).
(4) Stretch Property: f Ψ maps each edge E ⊂ T to an edge path f Ψ (E) ⊂ T ,
stretching length by a uniform factor of λωb(Ψ) .
This follows from the definition of the piecewise Riemannian metric on T in Section 5.2, the eigenlength equation in T in Section 5.3, and Section 7.1 item (1c).
b and each edge E of T , the restriction
(5) Train Track Property: For each Ψ, Ψ′ ∈ D
Ψ
Ψ′
of f to the edge path f (E) is injective.
This follows from property (2) together with property (4) as applied to ΨΨ′ .
For the statement of property (6), recall from Section 2.5 the subgroup Stab[T ] <
g ] < Aut(Fn ). Recall particularly Lemma 2.5 of that
Out(Fn ) and its pre-image Stab[T
g ] containing Inn(Fn ) has a unique
section, which says that each subgroup of Stab[T
action on T extending the given action of the free group Fn ≈ Inn(Fn ) such that each
element of the action satisfies a twisted equivariance property.
83
g ].
b0 : We have D0 < Stab[T ] and hence D
b0 < Stab[T
(6) Restricted action of D
b0 of the semi-action D
b+ y T is identical to
Furthermore, the restriction to D
b0 y T given by Lemma 2.5: the unique isometric action of D
b0 on T
the action D
b
such that each Ψ ∈ D0 satisfies Ψ-twisted equivariance.
b0 with projected image ψ ∈ D0 . The map f ψ : G → G
For the proof, consider Ψ ∈ D
restricts to the identity on each edge E ⊂ Hu , by Section 7.1 item (1c). The map
e→G
e therefore permutes the edges of H
e u , mapping each isometrically to its
f˜Ψ : G
image. It follows that the map f Ψ : T → T is an isometry. Since f Ψ satisfies Ψtwisted equivariance (property (1)), it follows that ψ ∈ Stab[T ] (Lemma 2.7). Since
g ]. Applying,
b0 < Stab[T
ψ ∈ D0 is arbitrary, we have proved D0 < Stab[T ] and that D
Ψ
Lemma 2.9 we have also proved that the map Ψ 7→ f is the same as the restriction
g ] y T given in that lemma, namely the unique action
to D0 of the action Stab[T
assigning to Ψ the unique Ψ-twisted equivariant isometry of T .
b and each j ∈ J there exists j ′ ∈ J
(7) Invariance of Nielsen data: For each Ψ ∈ D
such that Nj ′ = (f Ψ )# (Nj ), and f Ψ restricts to an order preserving bijection of
fΨ
basepoint sets Zj −→ Zj ′ . In particular f#Ψ induces a bijection of the Nielsen
paths in T (Definition 5.3).
This follows immediately from Section 7.1 item (1e).
Finally, as an immediate consequence of property (7) we have:
b+ y T ∗ : The semigroup action D
b+ y T extends uniquely to
(8) Extension to D
b+ ,
b+ y T ∗ , denoted f Ψ = f Ψ∗ : T ∗ → T ∗ for each Ψ ∈ D
a semigroup action D
∗
T
Ψ
Ψ
such that f∗ permutes the cone points Pj , and f∗ permutes the cone edges Pj Q
by isometries. In particular, for each j ∈ J and Q ∈ Zj , letting Nj ′ = f#Ψ (Nj )
and Q′ = f Ψ (Q), we have f∗Ψ (Pj ) = Pj ′ and Ψ Pj Q = Pj′Q′ .
7.3
b0 y T ∗
Dynamics of the group action D
Isometries of a simplicial tree equipped with a geodesic metric satisfy a dichotomy:
each isometry is either elliptic or loxodromic. This dichotomy does not hold for all
isometries of Gromov hyperbolic spaces — in general, there is a third category of
isometries, namely the parabolics. In this section we prove Lemma 7.2 which says in
b0 y T ∗ .
part that the dichotomy does hold for the action D
b0 y T ∗ , each element ∆ ∈ D
b0
Lemma 7.2 (Dynamics on T ∗ ). Under the action D
b0 acts loxodromically
acts either loxodromically or elliptically on T ∗ . More precisely, D
b0 acts loxodromically on T and its axis A ⊂ T is not a Nielsen
on T ∗ if and only if D
line (Definition 5.3), in which case A is a quasi-axis for ∆ in T ∗ . Furthermore, the
b0 have “uniform uniformly positive stable translation length”
loxodromic elements of D
84
b0 acting
in the following sense: there exist constants η, κ > 0 such that for each ∆ ∈ D
∗
∗
loxodromically on T , and for each x ∈ T and each integer k ≥ 0, we have
d∗ (∆k (x), x) ≥ k η − κ
Remarks. In this lemma, the terminology “uniform uniformly positive stable
translation length” refers to the corollary that the stable translation length
1
d∗ (∆k (x), x))
k→∞ k
lim
has a uniform positive lower bound η > 0 independent of the choice of a loxodromic
b0 , and the rate at which this positive lower bound is approached is
element ∆ ∈ D
uniform. This property will be applied in Section 8.2, for purposes of characterizing
b y S that is described in Section 8.1.
the loxodromic elements of the action D
Lemma 7.2 and proof could already have been presented almost word-for-word
back in Section 5.4 for the restricted action Inn(Fn ) ≈ Fn y T ∗ . Other than the
methods available in Section 5.4, the additional things needed to prove the lemma
b0 y T ∗ , like its restriction, satisfies twisted equivariance
are that the larger action D
and preserves Nielsen paths and associated objects (Section 7.2 (7), (8)).
Proof. Throughout the proof we will apply Section 7.2 (7), (8) regarding the action
b0 on Nielsen paths, elements of the Nielsen set {Ni }, the basepoint sets Zi ⊂ Ni ,
of D
the cone points Pi , and the cone edges Pi Q, Q ∈ Zi . In particular, because the action
b0 on T preserves Nielsen paths, it takes maximal ρ∗ -subpaths of any
of each ∆ ∈ D
path α to maximal ρ∗ -subpaths of the path ∆(α).
b0 . If ∆ acts elliptically on T then it fixes a vertex of T ; and if in
Consider ∆ ∈ D
the geometric case ∆ acts loxodromically on T and its axis is a Nielsen line Nj , then
∆ fixes the corresponding cone point Pj ; in either case ∆ acts elliptically on T ∗ .
Suppose that ∆ acts loxodromically on T and its axis A ⊂ T is not a Nielsen
line (this always happens in the nongeometric case, but the terminology of our proof
applies to all cases). For each vertex x ∈ T and each integer k ≥ 1 consider the Proposition 5.5 decomposition of the path [x, ∆k (x)] into “ν-subpaths” meaning maximal
ρ∗ subpaths, alternating with “µ-subpaths” each having no ρ∗ subpath. Since the
intersection [x, ∆k (x)] ∩ A contains a concatenation of k fundamental domains for the
action of ∆ on A, it follows that if [x, ∆k (x)] ∩ A contains a ν-subpath of [x, ∆k (x)]
then it must contain k − 1 distinct ν-subpaths of [x, ∆k (x)], between which there are
k − 2 distinct µ-subpaths of [x, ∆k (x)]; here we are applying the fact that ∆ takes
maximal ρ∗ paths to maximal ρ∗ paths. As a consequence, the collection of µ-subpaths
of [x, ∆k (x)] contains at least k − 2 distinct edges of T . Applying Proposition 5.5 (4)
we obtain DPF (x, ∆k (x)) ≥ (k − 2)η ′ where η ′ = min{lPF (E) E is an edge of T }.
85
Applying Proposition 5.8, in T ∗ we have
′
2η
η′
−
+C
d∗ (x, ∆ (x)) ≥ k
K
|{z}
|K {z }
k
=η
= κ′
It immediately follows that ∆ acts loxodromically on T ∗ . This estimate is for vertices
of T , but a similar estimate holds for arbitrary points x ∈ T ∗ , replacing κ′ by κ =
κ′ + 2δ where δ is an upper bound for the distance in T ∗ from an arbitrary point of
T ∗ to the vertex set of T .
8
The suspension action.
In this section we complete the proof of the multi-edge case of Theorem F.
Throughout this section we adopt Notations 4.1 and 7.1 concerning a rotationless
φ ∈ Out(Fn ) and a CT representative f : G → G with top EG stratum Hu , with
disintegration group D = D(f ) < Out(Fn ), and with free splitting Fn y T associated
to the pair (G, Gu ). In particular, in Section 5 we used the coned off free splitting
Fn y T ∗ to construct the suspension space S, proving that S is Gromov hyperbolic by
using flaring properties of the action Fn y T ∗ . In Section 7 we studied the extended
b < Aut(Fn ) and its subsemigroup D
b+ < D,
b and we constructed
disintegration group D
b+ y T ∗ .
the train track semigroup action D
b+ y T ∗ to
In Section 8.1 we shall show how to suspend the semigroup action D
b y S (this is a completely general construction that applies to
an isometric action D
the disintegration group of any CT having top EG stratum).
In Section 8.2 we put the pieces together to complete the proof.
8.1
b y S.
The suspension action D
Recall from Notations 7.1 the short exact sequence
ω
bu
b0 7→ D
b−
1 7→ D
→Z→1
b representing φ ∈ D, and so ω
Choose an automorphism Φ ∈ D
bu (Φ) = 1. It follows
that Φ determines a splitting of the above short exact sequence, and hence a semidirect
b≈D
b0 ⋊Φ Z expressed in terms of the inner automorphism on D
b
product structure D
b may be
determined by Φ, namely I(Ψ) = IΦ (Ψ) = ΦΨΦ−1 . Thus each element of D
b0 and m ∈ Z, and the group operation in
written in the form ∆Φm for unique ∆ ∈ D
this notation is defined by
(∆1 Φm1 )(∆2 Φm2 ) = (∆1 I m1 (∆2 ))Φm1 +m2
Recall from Definition 5.11 the construction of S: using the chosen automorphism
Φ representing φ, the corresponding Φ-twisted equivariant map of T ∗ is denoted in
86
abbreviated form as f∗ = fTΦ∗ , and S is the suspension space of f∗ , namely the quotient
of T × Z × [0, 1] with identifications [x, k, 1] ∼ [f∗ (x), k + 1, 0]. Recall also various
other notations associated to S that are introduced in Definitions 5.11 and 5.12.
b it suffices to carry
To define the action on S of each element Ψ = ∆Φk ∈ D
b
out the following steps: first we define the group action D0 y S; next we define
the action of the element Φ y S; then we define the action of each element ∆Φk
by composition; and finally we verify that the two sides of the semidirect product
relation Φ∆ = I(∆)Φ act identically.
b0 act on [x, k, t] ∈ S by the equation
Let each ∆ ∈ D
∆ · [x, k, t] = [f∗I
k (∆)
(x), k, t] = [f∗Φ
k ∆Φ−k
(x), k, t]
and note that this formula is well defined because, using the properties of the semib+ y T from Section 7.2, we have
group action D
∆ · [x, k, 1] = [f∗Φ
k ∆Φ−k
(x), k, 1] = [f∗Φ ◦ f∗Φ
Φk+1 ∆Φ−(k+1)
= [f∗
k ∆Φ−k
(x), k + 1, 0]
◦ f∗Φ (x), k + 1, 0]
= ∆ · [f∗Φ (x), k + 1, 0]
= ∆ · [f∗ (x), k + 1, 0]
b0 evidently acts by an isometry. Again using the semigroup action
Each ∆ ∈ D
b0 :
properties, the action equation is satisfied for each ∆′ , ∆ ∈ D
∆′ · (∆ · [x, k, t]) = [f∗I
=
k (∆′ )
◦ f∗I
k (∆)
(x), k, t]
k
′
k
[f∗I (∆ )◦I (∆) (x), k, t]
k
′
= [f∗I (∆ ∆) (x), k, t]
= ∆′ ∆ · [x, k, t]
b0 y S.
This completes the definition of the isometric action D
b0 y S agrees with
We note that the restriction to Fn ≈ Inn(Fn ) of the action D
the action Fn y S given in Definition 5.11, defined in general by γ · [x, k, r] =
[Φk (γ) · x, k, r] for each γ ∈ Fn . In the special case k = 0, r = 0, the equation
i
γ · x = f∗γ (x) holds by Lemma 2.7 (5). The extension to the general case follows
easily.
Next, let Φ act by shifting downward,
Φ · [x, k, t] = [x, k − 1, t]
which is evidently a well-defined isometry.
87
b0 the two sides of the semidirect product equation act
As required, for each ∆ ∈ D
in the same way:
Φ · ∆ · [x, k, t] = Φ · [f∗I
k (∆)
(x), k, t]
I k (∆)
= [f∗
(x), k − 1, t]
I(∆) · Φ · [x, k, t] = I(∆) · [x, k − 1, t]
= [f∗I
k−1 (I(∆))
I k (∆)
= [f∗
(x), k − 1, t]
(x), k − 1, t]
b0 preserves each “horizontal level set” Ss ,
Notice that since the action of each ∆ ∈ D
and since the action of Φ has “vertical translation” equal to −1 = −b
ω (Φ) meaning
b
that Φ(Ss ) = Ss−bω(Φ) , it follows that for each Ψ ∈ D the integer −b
ω (Ψ) is the “vertical
translation length” for the action of Ψ in S. We record this for later use as:
b and each fiber Ss (s ∈ R), we have
Lemma 8.1. For each Ψ ∈ D
Ψ(Ss ) = Ss−bω(Ψ)
8.2
for any s ∈ R
Proof of Theorem F: The multi-edge case.
First we set up the notation, based on the formulation of the multi-edge case in
Section 2.6 and the outline of the proof in Section 3.1, and we apply the results of
b y S. After that, Conclusions (1), (2)
Section 8.1 to obtain a hyperbolic action H
and (3) of Theorem F are proved in order.
b < Aut(Fn ) with image H < Out(Fn )
Setup of the proof. We have a subgroup H
b 7→
under the quotient homomorphism Aut(Fn ) 7→ Out(Fn ), and with J = Kernel(H
b is finitely generated
H), satisfying the hypotheses of Theorem F: H is abelian; H
b
and not virtually abelian; and no proper, nontrivial free factor of Fn is H-invariant.
The conclusion of Theorem F is about the existence of a certain kind of hyperbolic
b and so we are free to replace H
b with
action of a finite index normal subgroup of H,
b ′ < H,
b because once the conclusion is proved using some
any finite index subgroup H
′
b ′ , the same
hyperbolic action N y S of some finite index normal subgroup N ′ ⊳ H
conclusion follows for the restriction of N ′ y S to the action N y S where N
b conjugates of N ′ ; one need only observe that the
is the intersection of all of its H
conclusions of Theorem F for the action N ′ y S imply the same conclusions when
restricted to any finite index subgroup of N ′ .
We may therefore assume that H is a rotationless abelian subgroup, using the
existence of an integer constant k such that the k th power of each element of Out(Fn )
88
1
/
b∩D
b0
K=H
O
b
H
/
⊂
1
/
b ∩ Inn(Fn )
J =H
⊂
ZO
/
H
⊂
/
/D
⊂
1
/
Fn ≈ Inn(Fn )
/
Aut(Fn )
/
1
⊂
b
D
1
/
ωu
Fn ≈ Inn(Fn )
/
/
b
H
/
1
ω
bu
1
/
/
⊂
Out(Fn )
/
1
b and associated objects in a commutative diagram with short
Figure 2: The group H
exact rows.
is rotationless, and replace the abelian group H by its finite index subgroup of k th
powers.
We have a maximal, proper, H-invariant free factor system B of co-edge number
≥ 2 in Fn . Applying the Disintegration Theorem, we obtain φ ∈ H such that for
any CT representative f : G → G of φ with disintegration group D = D(f ), the
subgroup H ∩ D has finite index in H. We choose f so that the penultimate filtration
element Gu−1 represents the free factor system B. Since the co-edge number of B
is ≥ 2, the top stratum Hu is EG. Replacing H with H ∩ D, we may thus assume
H < D.
We now adopt Notations 4.1 and 7.1, regarding the free splitting Fn y T associated to the marked graph pair (G, Gu−1 ), the action Fn y T ∗ obtained by coning off
the elements of the Nielsen set, the disintegration group D = D(f ), and its associated
b Using that H < D it follows that H
b < D.
b Setting
extended disintegration group D.
ω
b
u
b 7→ H) and K = Kernel(H
b −→
J = Kernel(H
Z) we may augment the commutative
diagram of Notations 4.1 (4) to obtain the commutative diagram with short exact
rows shown in Figure 8.2.
From Notations 7.1 (6) we also have the subgroups and subsemigroups D0 < D+ <
b0 < D
b+ < D.
b
D and D
∗
b+ y T be the semigroup action described in Section 7.2, associating to each
Let D
b<D
b to be any pre-image of φ ∈ H < D.
b
Ψ ∈ D+ a map fTΨ∗ : T ∗ → T ∗ . Pick Φ ∈ H
Let S be the suspension space of fTΦ∗ : T ∗ → T ∗ as constructed in Definition 5.11. Let
b y S be the isometric action constructed in Section 8.1. We will make heavy use
D
b0 -equivariant identification S0 ↔ T ∗ .
of the integer sections Sj (j ∈ Z) and of the D
We turn to the proof of the three conclusions of Theorem F for the action of
b on S.
N =H
89
b acts either elliptically or
Proof of Conclusion (1): Every element of H
b acts either
loxodromically on S. We show more generally that every element of D
b has the
elliptically or loxodromically on S. From Section 8.1 the general element of D
b0 , and Ψ(Sj ) = Sj−m for any j ∈ Z. If m 6= 0 then,
form Ψ = ∆Φm for some ∆ ∈ D
by Lemma 5.13, for each k ∈ Z and each x ∈ S we have d(x, Ψk (x)) ≥ km , and so
Ψ is loxodromic. Suppose then that m = 0 and so Ψ = ∆ preserves Sj for each j ∈ Z.
Consider in particular the restriction ∆ S0 ≈ T ∗ and the further restriction ∆ T .
If ∆ is elliptic on T then it fixes a point of T , hence ∆ fixes a point of T ∗ ≈ S0 ⊂ S,
and so ∆ is elliptic on T ∗ and on S. If ∆ is loxodromic on T with axis L∆ ⊂ T , and if
L∆ is a Nielsen line Nj then ∆ fixes the corresponding cone point Pj ∈ T ∗ ∈ S0 ⊂ S,
and so ∆ is elliptic on T ∗ and on S.
b0 which act loxodromically on T and whose
It remains to consider those ∆ ∈ D
axes in T are not Nielsen lines. Applying Lemma 7.2, each such ∆ acts loxodromically
on T ∗ and for each x ∈ T ∗ and each integer k ≥ 1 we have
(∗)
d∗ (x, ∆k (x)) ≥ k η − κ
where the constants η, κ > 0 are independent of ∆ and x.
Consider the function which assigns to each integer k the minimum translation
distance of vertices v ∈ S under the action of ∆k :
σ(k) = inf dS (v, ∆k (v))
v∈S
To prove that ∆ acts loxodromically on S we apply the classification of isometries
of Gromov hyperbolic spaces [Gro87], which says that every isometry is elliptic,
parabolic, or loxodromic. But if ∆ is elliptic or parabolic then the function σ(k)
is bounded. Thus it suffices to prove that limk→∞ σ(k) = ∞.
b0 . Consider also the Φ−i -twisted
For each integer i, consider ∆′i = Φi ∆Φ−i ∈ D
∗
equivariant map ji : T = S0 → Si given by ji [x, 0, 0] = [x, i, 0], which is an isometry
from the metric d∗ = d0 to the metric di . This map ji conjugates the action of ∆k on
b0 on S0 , because
Si to the action of (∆′i )k = Φi ∆k Φ−i ∈ D
i
∆k ji [x, 0, 0] = ∆k [x, i, 0] = [f∗Φ ∆
k Φ−i
(x), i, 0]
∆′i
∆′i
= [f∗ (x), i, 0] = ji [f∗ (x), 0, 0]
= ji ∆′i [x, 0, 0]
Applying the inequality (∗) to ∆′i , and then applying the twisted equivariant isometric
conjugacy ji between ∆k and (∆′i )k , for each vertex p ∈ Si we have
(∗∗)
di (p, ∆k (p)) ≥ k η − κ
As seen earlier, the uniformly proper maps Si ֒→ S have a uniform gauge function
δ : [0, ∞) → [0, ∞) (see (∗) in Section 5.6). If σ(k) does not diverge to ∞ then there
90
is a constant M and arbitrarily large values of k such that dS (p, ∆k (p)) ≤ M holds for
some i and some vertex p ∈ Si , implying by that di (p, ∆k (p)) ≤ δ(M), contradicting
(∗∗) for sufficiently large k.
This completes the proof of Conclusion (1).
Furthermore, we have proved the following which will be useful later:
b0 the following are equivalent: ∆ acts loxodromically on S;
• For each ∆ ∈ D
∆ acts loxodromically on T ∗ ; ∆ acts loxodromically on T and its axis is not a
Nielsen line.
b y S is nonelementary. We shall focus
Proof of Conclusion (2): The action H
on the restricted actions of the subgroup J, proving first that J has nonelementary
action on T , then on T ∗ , and then on S; from the latter it follows that the whole
b y S is nonelementary.
action H
We shall apply Lemma 2.10 and so we must check the hypotheses of that lemma.
Let V nt be the set of vertices v ∈ T having nontrivial stabilizer under the action
b+ y T restricts to a
of Fn . As shown in Section 7.2 (3), the semigroup action D
nt
b+ y V having the property that each element of Ψ ∈ D
b+
semigroup action D
nt
acts by a Ψ-twisted equivariant bijection of V , and it follows immediately that
b y V nt having the same
this semigroup action extends uniquely to a group action D
b we obtain an action H
b y V nt satisfying the hypotheses in
property. Restricting to H
the first paragraph of Lemma 2.10. By hypothesis of Theorem F, no proper nontrivial
b and in particular no subgroup StabFn (v) (v ∈ V nt )
free factor of Fn is fixed by H,
b Finally, the free group J has rank ≥ 2 because otherwise, since H
is fixed by H.
b is a solvable subgroup of Aut(Fn ) < Out(Fn+1 )
is abelian, it would follow that H
b is
and hence is virtually abelian by [HM17c], contradicting the hypothesis that H
not virtually abelian. Having verified all the hypotheses of Lemma 2.10, from its
conclusion we have that the action J y T is nonelementary.
Since the action J y T is nonelementary, its minimal subtree T J is not a line, and
so T J contains a finite path α which is not contained in any Nielsen line. Furthermore,
α is contained in the axis of some loxodromic element γ ∈ J whose axis Lγ is therefore
not a Nielsen line. Applying Lemma 7.2, the action of γ on T ∗ is loxodromic and Lγ
is a quasi-axis for γ. Choosing δ ∈ J − StabFn (Lγ ) and letting γ ′ = δγδ −1 , it follows
that γ ′ also acts loxodromically on T and on T ∗ , and that its axis Lγ ′ in T is also a
quasi-axis in T ∗ . By choice of δ the intersection Lγ ∩ Lγ ′ is either empty or finite.
Since neither of the lines Lγ , Lγ ′ is a Nielsen axis, each ray in each line has infinite
DPF diameter and so goes infinitely far away from the other line in T ∗ -distance. It
follows that γ, γ ′ are independent loxodromic elements on T ∗ , proving that J y T ∗
is nonelementary.
Finally, using the same γ, γ ′ as in the previous paragraph whose axes Lγ , L′γ in
T are not Nielsen lines, we showed in the proof of Conclusion (1) that γ, γ ′ act
loxodromically on S. Furthermore, since the lines Lγ , L′γ have infinite Hausdorff
91
distance in T ∗ , it follows by they also have infinite Hausdorff distance in S, as shown
in item (∗) of Section 5.6. This proves that γ, γ ′ are independent loxodromics on S
and hence the action J y S is nonelementary.
Proof of Conclusion (3): Each element of J acting loxodromically on S is
b y S.
a WWPD element of H
b∩D
b0 acting loxodromically on S. As shown earlier
Consider γ ∈ J < K = H
under the proof of Condition (1), this is equivalent to γ acting loxodromically on T ∗ ,
and to γ acting loxodromically on T with an axis Lγ ⊂ T that is not a Nielsen line.
We shall show first that γ is a WWPD element of the action K y T , then of the
b y S. Let Stab(Lj ) denote the subgroup
action K y T ∗ , and finally of the action H
of K that stabilizes Lj ; we may say that Lj is the axis of Stab(Lj ), meaning that Lj
is the common axis for all the elements of Stab(Lj ) that act loxodromically on T .
b∩D
b0 y T as shown in SecThe action J y T extends to the action K = H
tion 7.2 (6). Since J is normal in K, and since J y T is the restriction of the free
splitting action Inn(Fn ) ≈ Fn y T , we may apply Lemma 2.11 with the conclusion
that γ is a WWPD element of the action K y T . To express this conclusion, let
LK = {Ψ(Lγ ) Ψ ∈ K}; equivalently, LK is the set of axes for all of the conjugate
subgroups Ψ Stab(Lγ )Ψ−1 , Ψ ∈ K; and more explicitly, letting {Ψi }i∈I be a bijectively indexed set of left coset representatives of Stab(Lγ ) in the group K, we may
index LK bijectively as LK = {Li }i∈I where Li = Ψi · Lγ . Using the geodesic metric
du on T , the WWPD property for γ as an element of the action K y T says that the
lengths of the segments Li ∩ Lj for i 6= j in the metric du are uniformly bounded by
some constant B independent of i, j. For all x, y ∈ Li ∩ Lj ∩ V we obtain lu (x, y) ≤ B
(by definition of lu ). By applying Corollary 4.4 we obtain Lu (x, y) ≤ B, proving that
(∗)
Diamu (Li ∩ Lj ) ≤ B
for all i 6= j
where Diamu (X) denotes the Du -diameter of X ∩ V .
Next we use (∗) to show that γ is a WWPD element of the action K y T ∗ . For
this purpose we may work entirely in T using the coarse metric Du on the set V . To
say what this means, first recall from Lemma 4.7 that Du satisfies the metric axioms
except that the triangle inequality is replaced by the “coarse triangle inequality” with
respect to a constant C = C4.7 ,
Du (x, z) ≤ Du (x, y) + Du (y, z) + C
for all x, y, z ∈ V .
For any subtree τ ⊂ T and any R > 0 let
NR (τ ) = ∪{xy
x ∈ τ ∩ V, y ∈ V, Du (x, y) ≤ R}
By Lemma 4.2 and Corollary 4.4, the Du and DPF coarse metrics on V are quasicomparable, and by Proposition 5.8 the inclusion V ֒→ T ∗ is a quasi-isometry from
92
the DPF coarse metric to the d∗ metric. Also, for any x, y ∈ V the path xy is
a uniform quasigeodesic in the d∗ metric, and in particular the lines Li ∈ LK are
uniform quasigeodesics in d∗ . To prove that γ is a WWPD element of K y T ∗ it
suffices to prove that for each R > 0 there exists M > 0 such that
(∗∗)
Diamu (NR (Li ) ∩ NR (Lj )) ≤ M
for all i 6= j
This inequality (∗∗) is actually all we shall need below when we prove that γ is a
b y S. But just to complete the train of ideas,
WWPD element of the action H
the reason that (∗∗) implies γ is a WWPD element of K y T ∗ is that if not then,
using that the quasi-axes in LK are uniformly quasigeodesic in T ∗ , there exists a
constant R > 0 and a sequence Li1 , Li2 , . . . ∈ LK such that as k → ∞ the intersection
NR (Lγ ) ∩ NR (Lik ) contains a subpath αk ⊂ Lγ such that the αk ’s are nested and their
union equals Lγ , and hence Diamu (NR (Li ) ∩ NR (Lj )) → ∞.
We prove (∗∗). Letting p = pi (x) ∈ Li ∩ V be a point that minimizes Du (x, p), we
assume that Du (x, pi (x)) ≤ R; let pj (x), pi (y), pj (y) by similarly defined with similar
distance bounds. Let xi ∈ Li ∩ V be the projection of x through T to Li , meaning
the unique point xi ∈ Li ∩ V such that [x, xi ] = {xi }, and let xj , yi , yj be similarly
defined. Since Du (x, xi ) ≥ Du (x, pi (x)), and since [x, pi (x)] is the concatenation of
[x, xi ] and [xi , pi (x)], it follows from Lemma 4.13 (1) that
Du (xi , pi (x)) ≤ Du (x, pi (x)) − Du (x, xi ) + lu (ρ)
≤ lu (ρ)
and combining with the coarse triangle inequality we obtain
Du (x, xi ) ≤ Du (x, pi (x)) + Du (pi (x), xi ) + C
≤ R + lu (ρ) + C
and similarly for Du (x, xj ), Du (y, yi ), and Du (y, yj ). Of the two points xi , xj , at least
one of them, denoted xk , is contained in Li ∩Lj ; similarly at least one of yi , yj , denoted
yk , is in Li ∩ Lj . It follows from the coarse triangle inequality that
Du (x, y) ≤ Du (x, xk ) + Du (xk , yk ) + Du (yk , y) + 2C
≤ (R + lu (ρ) + C) + Diamu (Li ∩ Lj ) + (R + lu (ρ) + C) + 2C
≤ B + 6C + 2R + 2lu (ρ) = M
which completes the proof that γ is a WWPD element of the action K y T ∗ .
b y S. Let
Finally we show that γ is a WWPD element of the action H
b =
LS = {Ψ(Lγ ) Ψ ∈ H}
93
[
{Φ−j (L) L ∈ LK }
|
{z
}
j∈Z
Lj
which is a set of uniform quasi-axes for all subgroups of H conjugate in H to hγi; and
note that Lj = {L ∈ LS L ⊂ Sj }. If γ fails to be WWPD then there is a sequence
of lines
(#)
Li1 , Li2 , Li3 , . . . ∈ LS
and there exists a constant R > 0 and sequences wk , xk ∈ Lγ , yk , zk ∈ Lik defined
for k ≥ 1, such that [wk , xk ] is nested sequence of subpaths of Lγ whose union equals
Lγ , and such that dS (wk , yk ), dS (xk , zk ) ≤ R for all k. The lines in LS are uniform
quasigeodesics in the hyperbolic space S, and so there exist constants ǫ > 0, M > 0
such that for any lines L, L′ ∈ LS and points w, x ∈ L, y, z ∈ L′ , if dS (w, y), dS (x, z) ≤
R, and if dS (w, x) ≥ M, then the minimum distance in S between the subsegments
[w, x] and [y, z] is ≤ ǫ.
Throwing away finitely many terms of the sequence (#), we may assume that
dS (wk , xk ) ≥ M for all k = 1, 2, . . . Applying Lemma 5.13 it immediately follows
that if j > ǫ then no line in the sequence (#) is in the set Lj . Thus there exists
some integer j with j ≤ ǫ such that the set Lj contains infinitely many lines in the
sequence (#), and so passing to a subsequence we may assume that all of the lines
Li1 , Li2 , Li3 , . . . are in Lj .
Under the semiflow between S0 and Sj (flowing from S0 to Sj if 0 ≤ j and from
Sj to S0 of j ≤ 0), which is a quasi-isometry between S0 and Sj , let L′γ ⊂ Sj be
the line corresponding to Lγ ⊂ S0 , and so Lγ and L′γ have finite Hausdorff distance
in S. Choose sequences wk′ , x′k ∈ L′γ so that the distances dS (wk , wk′ ), dS (xk , x′k )
are uniformly bounded. Passing to a further subsequence we may assume that the
subsegments [wk′ , x′k ] are nested; their union is L′γ . The distances dS (wk′ , yk ), dS (x′k , zk )
are also uniformly bounded, and hence by uniform properness of the inclusion Sj ֒→ S
the distances dj (wk′ , yk ), dj (x′k , zk ) are uniformly bounded. It follows that for each M
there exists N such that if i1 , i2 ≥ N then the lines L′i1 , L′i2 ∈ Lj are uniformly close
on subsegments whose dj length is > M, and hence there exists a constant R such
that
Diamdj (NR (L′i1 ) ∩ NR (L′i2 )) > M for all i1 , i2 ≥ N
But this contradicts (∗∗), because under the isometry Φj : Sj → S0 ≈ T ∗ the set of
lines Lj corresponds to the set of lines L0 = LK .
References
[BBF15] M. Bestvina, K. Bromberg, and K. Fujiwara, Constructing group actions
on quasi-trees and applications to mapping class groups, Publ. Math. Inst.
Hautes Études Sci. 122 (2015), 1–64.
[BF92]
M. Bestvina and M. Feighn, A combination theorem for negatively curved
groups, J. Diff. Geom. 35 (1992), no. 1, 85–101.
94
[BF02]
M. Bestvina and K. Fujiwara, Bounded cohomology of subgroups of mapping
class groups, Geom. Topol. 6 (2002), 69–89 (electronic).
[BFH97] M. Bestvina, M. Feighn, and M. Handel, Laminations, trees, and irreducible automorphisms of free groups, Geom. Funct. Anal. 7 (1997), 215–
244.
[BFH00]
, The Tits alternative for Out(Fn ). I. Dynamics of exponentiallygrowing automorphisms., Ann. of Math. 151 (2000), no. 2, 517–623.
[BFH04]
, Solvable subgroups of Out(Fn ) are virtually Abelian, Geometriae
Dedicata 104 (2004), 71–96.
[CM87]
M. Culler and J. W. Morgan, Group actions on R-trees, Proc. London
Math. Soc. (3) 55 (1987), no. 3, 571–604.
[FH09]
M. Feighn and H. Handel, Abelian subgroups of Out(Fn ), Geometry and
Topology 13 (2009), 1657–1727.
[FH11]
, The recognition theorem for Out(Fn ), Groups Geom. Dyn. 5
(2011), 39–106.
[GJLL98] D. Gaboriau, A. Jaeger, G. Levitt, and M. Lustig, An index for counting
fixed points of automorphisms of free groups, Duke Math. J. 93 (1998),
no. 3, 425–452.
[Gro87]
M. Gromov, Hyperbolic groups, Essays in group theory (S. Gersten, ed.),
MSRI Publications, vol. 8, Springer, 1987.
[HM]
M. Handel and L. Mosher, Relative free splitting and free factor complexes
II: Loxodromic outer automorphisms, In preparation.
[HM13]
, Lipschitz retraction and distortion for subgroups of Out(Fn ),
Geom. Topol. 17 (2013), no. 3, 1535–1579.
[HM15]
, Hyperbolic actions and 2nd bounded cohomology of subgroups of
Out(Fn ) Part I: Infinite lamination subgroups, arXiv:1511.06913, Nov.
2015.
[HM17a]
, Subgroup decomposition in Out(Fn ), Part I: Geometric models,
Memoirs AMS (2017), to appear. arXiv:1302.2378.
[HM17b]
, Subgroup decomposition in Out(Fn ), Part II: A relative Kolchin
theorem, Memoirs AMS (2017), to appear. arXiv:1302.2379.
[HM17c]
, Subgroup decomposition in Out(Fn ), Part III: Weak attraction theory, Memoirs AMS (2017), to appear. arXiv:1306.4712.
95
[HM17d]
, Subgroup decomposition in Out(Fn ), Part IV: Relatively irreducible
subgroups, Memoirs AMS (2017), to appear. arXiv:1306.4711.
[HM17e]
, Virtually abelian subgroups of IAn (Z/3) are abelian, In preparation., 2017.
[KR14]
I. Kapovich and K. Rafi, On hyperbolicity of free splitting and free factor
complexes, Groups Geom. Dyn. 8 (2014), no. 2, 391–414.
[MS12]
M. Mj and P. Sardar, A combination theorem for metric bundles, Geom.
Funct. Anal. 22 (2012), no. 6, 1636–1707.
96
| 4math.GR
|
A PROBABILISTIC DEMAND SIDE MANAGEMENT APPROACH BY CONSUMPTION
ADMISSION CONTROL
Lorant Kovacs, Rajmund Drenyovszki, Andras Olah, Janos Levendovszky, Kalman Tornai, Istvan Pinter
New generation electricity network called Smart Grid is a recently conceived vision for a cleaner, more efficient and cheaper electricity system. One of the major
challenges of electricity network is that generation and consumption should be balanced at every moment. This paper introduces a new concept for controlling the demand
side by the means of automatically enabling/disabling electric appliances to make sure that the demand is in match with the available supplies, based on the statistical
characterization of the need. In our new approach instead of using hard limits we estimate the tail probability of the demand distribution and control system by using the
principles and the results of statistical resource management.
Keywords: Smart grid, Demand Side Management, Admission Control
1
Introduction
The main issue in electricity networks is keeping an
almost perfect balance between electricity generation and
consumption all the time. Balance between demand and
supply is crucial since oversupply means waste of energy,
while undersupply causes performance degradation of the
grid parameters (e.g. phase, voltage level, etc.).
Unfortunately the control of the supply side is almost
impossible because of the large time constants of the
fossil and nuclear plants; the only possibility is applying
cost ineffective auxiliary generators. Additionally, in
smart grids the percentage of renewable resources should
be increased which gives rise to uncertainty in the
generation side. Hence, the best way to keep the balance
is to manage the demand side. Demand Side Management
(DSM) means a new kind of challenge: system operators
should control the power grid in local scale, which is
possible by installing intelligent measurement devices
(smart meters). However, as a new perspective,
households can be controlled with the intelligent devices.
The residential sector accounts for about 30% of total
energy consumption [1] and contains time shiftable
appliances in high number. The amount of consumption
involved in direct control can eliminate the error between
daily prediction based generation and actual demand. The
spread of electric vehicles could mean an additional
opportunity. In average cars are parked in Europe for
more than 90% of the time [2]; hence, batteries of electric
vehicles can serve as an extra storage capacity for the
power grid.
In this paper we propose a new approach for shorttime demand side management. The introduced method
takes into account the probabilistic nature of the load by
the aid of a consumption admission control. The
algorithm enables/disables shiftable appliances and
reshapes the probability density function (pdf) of the
aggregate consumption.
1.1 Related work
Influencing the demand side in the context of Smart
Grid electricity networks is usually referred to as Demand
Side Management (DSM) or Demand Response (DR).
Demand Response is a mechanism managing customer
consumption in response to supply side conditions while
Demand Side Management covers all the activities or
programs undertaken by service providers to influence the
amount or timing of electricity use. There are many
solutions proposed to DR and DSM like direct control of
smart appliances, pricing and load scheduling. Good
references can be found about different DSM approaches
in [3], [4] and [5]. With direct control system operators
can remove the extreme values in electricity consumption
(peak shaving) and encourage additional energy use
during periods of lowest system demand (valley filling).
The load control as a demand response strategy is
presented in [6], where simulating (summer period, airconditioning units) is conducted with two control
algorithms. It takes into account users’ comfort (via
heuristic consumer utility metric) and uses binary on-off
policies. Fairness is maintained by two scheduling
algorithms: priority based and round robin. Results show
that significant energy and cost savings can be achieved
with the proposed algorithms.
To minimize the operating cost of a residential
microgrid, a MILP model is proposed in [7]. Decision
variables are used to model demand and also supply of
both electrical and thermal energy. It covers solar energy,
distributed generators, energy storages, and loads. A
model predictive control scheme is proposed to iteratively
produce a control sequence for the microgrid. The case
study reveals the performance of minimum cost control
by comparison with benchmark control policies. Results
show savings in annual operating cost.
A DSM model with three layers is introduced in [8].
This model consists of admission controller, load
balancer, and demand response with load forecaster
modules. Whenever a user turns on an appliance, a
request is sent to the admission controller. If the capacity
is available and we are not in peak hours, then it accepts
the request and initiates the operation of the appliance. If
the appliance operation exceeds the capacity, then the
request is rejected and forwarded to the load balancer.
The task of the load balancer is to solve an optimization
problem and to assign a future timeslot for the appliance
to start later. A game-theoretic approach for residential
energy consumption scheduling is proposed in [9]. It
introduces a pricing mechanism which is based on a
convex and increasing cost function. The authors present
a distributed algorithm for optimization problem. Most of
the proposed techniques in the literature consider fixed
load curves. However most of the papers do not deal with
randomness on the load side, an exception is [10], where
uncertainty is considered as well. The authors propose a
MILP optimization model, which performs scheduling.
The adopted DSM model forecasts the load curve of the
user from the previous knowledge of their energy usage.
Additionally, real time pricing and inclining block rates
are combined in the model for effective pricing. The
optimization is multi-stage, as the information of the
appliances is revealed over time, the schedule of the
appliances is updated accordingly. Simulation results
show efficiency by reducing total peak-to-average ratio
and energy expenses of users.
The rest of this paper is organized as follows. The
problem formulation and system model is described in
Section 3. The concept of our proposed Consumption
Admission Control algorithm is introduced in Section 4.
The results are presented with discussion in Section 5.
Finally the paper is concluded in Section 6.
Figure 1 The applied model
2
Problem formulation and system model
A large number of appliances can tolerate some delay
(e.g. executing the program of a washing machine at a
later time). Additionally in the near future the spread of
electric vehicles will mean a huge amount of elastic
demand in the power system. As a result, there are (time)
shiftable and non-shiftable demands in the system. On
the other hand most of the devices show stochastic
consumption behaviour (neither the start of use nor time
of operation is known a priori), hence, only a statistical
approach can efficiently solve the control task. The
foundation of our approach is that, the new Consumption
Admission Control algorithm can modify the pdf of a
consumption unit (e.g. a household, street, city etc.) by
the temporary enabling/disabling of shiftable appliances.
From the system operators’ perspective pdf close to Dirac
delta function (meaning constant load) is ideal. However,
we cannot reach the optimum, as a more realistic goal, we
can keep the probability density function as narrow as
possible, i.e. the mass of the pdf lies between a lower and
an upper limit. For the sake of an even more realistic
model, we allow the tail probabilities to be non-zero but
smaller than a predefined probability.
In this paper, we assume that the service provider
calculates and communicates these parameters governing
the behaviour of a customer. Using these parameters, the
subscriber’s Smart Meter (SM) can enable/disable the
appliances at a local level, resulting in a fully distributed
solution to the problem. The parameters coming from the
service provider that govern the CAC algorithm are as
follows: a capacity upper limit (𝐶𝑚𝑎𝑥 ) in every time slot,
which is allowed to be exceeded by a small probability 𝑝.
(In this paper we will concentrate on the upper limit 𝐶𝑚𝑎𝑥
and oversonsumption probability, however we plan to
extend our approach by a lower limit 𝐶𝑚𝑖𝑛 and an
underconsumption probality 𝑟 in our future work). The
tail probability will be referred to as Quality of Service
(QoS) parameter, because it can satisfy the overload of
the grid to keep under a certain limit, and hence, the
stability of the grid parameters such as frequency, voltage
level, etc. The task of the SM is the admission control
(enabling/disabling) of the appliances by such a way, that
the probability distribution function (pdf) of the aggregate
consumption satisfies the prescription of the service
provider. (The cooperative attitude of the subscriber can
be motivated by rewards.) The underlying model is
depicted in Figure 1.
In the model all subscribers are assumed to have a
Smart Meter. The SM has the following properties:
the SM can communicate with the service provider
and with the smart appliances;
the SM can register the consumption statistics of the
appliances (both smart and traditional ones);
the SM can temporarily enable/disable appliances.
In the model stochastic and deterministic, shiftable
and non-shiftable appliances are taken into consideration.
(The devices executing a fixed program can be seen as
deterministic). The defined categories and some examples
of appliances are listed in Table 1.
Table 1 Device categories used in the model
shiftable
nonshiftable
stochastic
electric heating,
air conditioner,
refrigerator
lighting,
vacuum cleaner
deterministic
washing machine,
dishwasher
circulation pump
In Figure 2 the measures used in the model are
depicted. The admission control algorithm uses discrete
time slots (denoted by k in Figure 3), in which the
enabled/disabled status of the appliances and the system
parameters (capacity limits and QoS) are supposed to be
unchanged. (New consumption requests are supposed to
be handled instantaneously). In all time slots there is a
deterministic component of the consumption and as well
as a stochastic one.
Figure 2 Illustration of the original and modified pdf of the aggregate
consumption and the free parameters that govern the algorithm
The stochastic part is described by its estimated (or
calculated) probability distribution function. The
maximum (Cmax ) and minimum (Cmin ) capacity limits can
be changed in every time instant by the service provider.
Csys builds a natural upper limit (i.e. lines and fuses) on
Cmax .
𝑛𝑖
𝑀
𝑓𝑋 (𝑥) = 𝑃𝑟 (∑ ∑ 𝑋𝑖𝑗 = 𝑥 ) =
𝑖=1 𝑗=1
(3)
= 𝑓𝑋11 (𝑥) ∗ 𝑓𝑋12 (𝑥) ∗ 𝑓𝑋13 (𝑥) ∗ … ∗ 𝑓𝑋𝑀𝑛 (𝑥)
𝑖
where 𝑀 is the number of appliance classes, and 𝑛𝑖 is
the number of appliances in class 𝑗, and ∑𝑀
𝑖=1 𝑛𝑖 is the
total number of enabled appliances (An appliance class
means a set of appliances that have the same statistical
descriptors). Considering deterministic (𝑋𝑑𝑒𝑡 ) and
stochastic (𝑋𝑠𝑡𝑜𝑐ℎ ) appliances in the model, we can write
the inequality:
𝑃𝑟(𝑋 𝑠𝑡𝑜𝑐ℎ + 𝑋 𝑑𝑒𝑡 ≥ 𝐶𝑚𝑎𝑥 ) ≤ 𝑝,
(4)
Figure 3 Measures used in the model
3
𝑋𝑑𝑒𝑡 is a constant value so the probability can be
expressed as
Consumption Admission Control Algorithm
The decision to enable or disable an appliance in the
system is carried out by the Consumption Admission
Control (CAC) algorithm. As mentioned in Section 2 the
aim of the algorithm is to sharpen the shape of the pdf of
the aggregate consumption of a customer resulting near
constant load in the time domain. The Smart Meter
calculates the aggregate pdf from the individual pdf-s of
the appliances. The individual pdf can be communicated
to the SM by smart appliances, or it can be measured in
the case of traditional ones. This concept was originally
applied for Call Admission Control for ATM
communication networks [11].
In this paper the following mathematical model will
be used: Let 𝑋𝑗 denote the random variable of the
consumption of the jth appliance, while
𝑁
𝑋 = ∑ 𝑋𝑗
(1)
𝑗=1
is the aggregate consumption random variable and 𝑁 is
the number of enabled appliances.
In the case of a new incoming consumption demand,
the CAC checks whether the inequality (2) holds for the
enabled plus the incoming appliance
𝑃𝑟(𝑋 ≥ 𝐶𝑚𝑎𝑥 ) ≤ 𝑝
(2)
where 𝑃𝑟 denotes probability of an event, and 𝑝 is the
probability limit of overconsumption. Therefore, CAC
keeps the upper tail probability of the aggregate
consumption under the limit 𝑝.
The probability of overconsumption 𝑃𝑟(𝑋 ≥ 𝐶𝑚𝑎𝑥 )
can be calculated based on the probability density
function 𝑓𝑋 (𝑥) of the aggregate consumption. The pdf of
the aggregate consumption can be calculated analytically
by the convolution of the individual pdfs of all
appliances:
𝑃𝑟(𝑋 𝑠𝑡𝑜𝑐ℎ ≥ 𝐶𝑚𝑎𝑥 − 𝑋 𝑑𝑒𝑡 ) ≤ 𝑝;
(5)
The lower limit can be checked by the same manner
as the upper limit. If the probability of underload is higher
than 𝑟, the goal can be expressed as
𝑃𝑟(𝑋 < 𝐶𝑚𝑖𝑛 ) ≤ 𝑟;
(6)
3.1 Estimation of the probability of overconsumption
The convolution operation in (3) can be very time
consuming in the case of high number of appliances
and/or classes, so it is suggested to estimate the
probability in terms of inequalities [12] of Large
Deviation Theory (LDT) bounds, such as Markov,
Chebisev, Bennett, Hoeffding and Chernoff upper
bounds. The estimation of overconsumption can be
derived from the calculation of the following upper
bound:
̂(𝑋, 𝐶𝑚𝑎𝑥 ) ≤ 𝑝
𝑃𝑟(𝑋 ≥ 𝐶𝑚𝑎𝑥 ) ≤ 𝑈
(7)
̂ (𝑋, 𝐶𝑚𝑎𝑥 ) is the bounding method on the tail
where 𝑈
probability. Because of the independence of the 𝑋𝑖𝑗
random variables, the expected value can be expressed as
𝑛𝑖
𝑀
𝜇 = 𝐸{𝑋} = ∑ ∑ 𝜇𝑖𝑗
(8)
𝑖=1 𝑗=1
and variance as
𝑀
2
2}
𝜎 = 𝐸{(𝑋 − 𝜇)
𝑛𝑖
= ∑ ∑ 𝜎𝑖𝑗2
(9)
𝑖=1 𝑗=1
The most widely known, Markov's inequality needs only
expected value to give an upper bound for the probability
that the non-negative X random variable is greater than or
equal to some positive constant (𝐶𝑚𝑎𝑥 in our case) :
𝑃𝑟(𝑋 ≥ 𝐶𝑚𝑎𝑥 ) ≤
𝜇
𝐶𝑚𝑎𝑥
(10)
Inevitably the advantage of Markov’s inequality is its
simplicity, but it is not a tight upper bound.
Chebysev’s inequality
𝑃𝑟(𝑋 ≥ 𝐶𝑚𝑎𝑥 ) ≤
𝜎2
(𝐶𝑚𝑎𝑥 − 𝜇)2
(11)
is also simple, but it is also not a tight upper bound.
Hoeffding’s inequality is an exponentially decreasing
upper bound, which results in a tighter estimation even far
from the expected value compared to Markov’s and
Chebysev’s inequalities. It is also based on the
expectation that 𝑋𝑖𝑗 random variables are independent and
additionally 𝑋𝑖𝑗 variables have upper and lower bounds:
𝑥𝑖𝑗𝑚𝑖𝑛 ≤ 𝑋𝑖𝑗 ≤ 𝑥𝑖𝑗𝑚𝑎𝑥 . Hoeffding’s inequality [13] can be
expressed as:
𝑃𝑟(𝑋 ≥ 𝐶𝑚𝑎𝑥 ) ≤ 𝑒𝑥𝑝 (
−2(𝐶𝑚𝑎𝑥 − 𝜇)2
2)
∑𝑖 ∑𝑗(𝑥𝑖𝑗𝑚𝑎𝑥 − 𝑥𝑖𝑗𝑚𝑖𝑛 )
(12)
From (12) it is clear that with the increase of Cmax , the
upper bound decreases in an exponential rate.
Bennett’s inequality gives exponentially decreasing upper
bound like Hoeffding’s, which assumes bounded input
random variables |𝑋𝑖𝑗 | ≤ 𝑥𝑚𝑎𝑥 , and it is formulated in
the following form [14]:
𝑃𝑟(𝑋 ≥ 𝐶𝑚𝑎𝑥 ) ≤
𝑒𝑥𝑝 (−
𝜎2
(𝐶𝑚𝑎𝑥 − 𝜇) ∙ 𝑥𝑚𝑎𝑥
∙ℎ(
))
2
𝑥𝑚𝑎𝑥
𝜎2
(13)
where ℎ(𝑢) = (1 + 𝑢) 𝑙𝑜𝑔(1 + 𝑢) − 𝑢. Bennett’s
inequality needs additional statistical information
compared to Hoeffding’s, the standard deviation of
appliances (σij ) and maximum value (𝑥𝑖𝑗𝑚𝑎𝑥 ).
Chernoff’s inequality is also an exponentially decreasing
upper bound [15]:
here X is the aggregate consumption random variable
containing the consumption of all enabled (both shiftable
and non-shiftable) appliances plus the incoming one.
Another approach for estimating an aggregate pdf is
based on the Central Limit Theorem (CLT)
𝑃𝑟(𝑋 > 𝐶𝑚𝑎𝑥 ) ≤ 1 − 𝐹𝑋 (𝐶𝑚𝑎𝑥 )
(17a)
𝐶𝑚𝑎𝑥 − 𝜇
𝐹𝑋 (𝐶𝑚𝑎𝑥 ) → 𝛷 (
)
√𝜎 2
(17b)
where 𝐹(𝑥) denotes the cdf of 𝑥 and 𝛷(. ) is the
standard normal cdf. We must emphasize that CLT is not
an upper bound on the tail probability. The speed of
convergence of F(x) → Φ(x) is the main question
regarding the estimations based on the Central Limit
Theorem. The absolute error of the CLT estimation
|F(x) − Φ(x)| is decreasing towards the tails, but the
relative error |F(x) − Φ(x)|/Φ(x) is increasing [16].
4
Results and discussion
In order to have a clear picture about the performance
of the Consumption Admission Control algorithm a
simulation environment was established in MATLAB.
We investigated the following aspects of the CAC
algorithm:
Relation of QoS (𝑝) and empirical probability of
overconsumption (𝑝̃) in the case of different LDT
bounds;
Model complexity of load time series;
Load shape modification made by CAC;
Number of enabled appliances in the case of different
LDT bounds and CLT;
Throughout our simulations we used stationary load time
series to explore the statistical behaviour of the CAC
algorithm. It is clear that the real benefit of the new
algorithm comes to the fore in a nonstationary
environment such a day or longer consumption period.
4.1 Relation of QoS and empirical probability of
overconsumption
𝑁
𝑃𝑟(𝑋 ≥ 𝐶𝑚𝑎𝑥 ) ≤ 𝑒𝑥𝑝 (∑ 𝜇𝑗 (𝑠 ∗ ) − 𝑠 ∗ Cmax )
(14)
𝑗=1
where μj (s) = lgE{esXj } are the so called logarithmic
momentum generating functions and s∗ is the parameter
that satisfies the possibly tightest bound:
𝐽
∗
𝑠 : 𝑖𝑛𝑓 ∑ 𝜇𝑗 (𝑠) − 𝑠𝐶𝑚𝑎𝑥
𝑠>0
(15)
𝑗=1
When a new demand of a shiftable appliance appears,
enabling or disabling will be calculated using one of the
upper-bounds:
−1, 0 𝐴𝑐𝑐𝑒𝑝𝑡
̂ (𝑋, 𝐶𝑚𝑎𝑥 )} = {
𝑠𝑔𝑛{𝑝 − 𝑈
+1 𝑅𝑒𝑗𝑒𝑐𝑡
(16)
In this section we present our investigation regarding
the relation of predefined QoS and empirical probability
of overconsumption. The ratio of predefined QoS and
empirical probability of overconsumption will be denoted
by
𝑘=
𝑝̃
𝑝
(18)
Using an upper bound on the tail probability leads to
underestimation of the number of appliances to be
enabled which results in 𝑝̃ < 𝑝, i.e. 𝑘 < 1; and vice-versa
a lower bound results in 𝑘 > 1. From the point of view of
the service provider, 𝑘 < 1 means guaranteed QoS, but
causes spare capacities.
The following assumptions were made in the
simulations:
Load of appliances were modelled by two-state
Bernoulli iid series of 50000 time instants;
-
There is only one appliance class. (All appliances
have the same statistical descriptors.)
Number of appliances in the class is 400;
The consumption demand of the temporarily disabled
appliances are deleted.
The aim of the investigation was to measure the
performance of different tail probability estimation
methods plugged into the CAC in the case of different
probability of ON state of the appliances (p ON). Figure 4
and 6 depict the results in the case of pON=0,1 and
pON=0,5, respectively.
The results in Figure 4 and 5 show that the empirical
probability can almost meet QoS (𝑘 = 1) when the tail
probability is exactly calculated from the analytical
aggregate pdf (see (3)). There is only a small deviation,
𝑘 = 0,4 … 0,6 in the case of small probabilities
(10−5 … 10−4 ) due to the difficulty of measuring rare
events in the case of Monte Carlo simulations.
p~ vs p N= 400 len= 50000 pON = 0; 1 h = 1 Cmax = mean
0
10
analytic
clt
chernoff
-1
10
bennett
hoeffding
chebisev
-2
p
~
10
-3
10
-4
10
4.2 Model complexity of load time series
The CAC algorithm needs appliance level statistical
information, therefore, load time series in our simulations
are generated with the bottom-up approach, i.e. the
aggregate time series are built up from appliance level
consumption time series. We used different appliancelevel models in the simulations:
Bernoulli iid;
First Order Markovian;
Higher Order Markovian.
In all the tree cases two-state (ON/OFF) models were
used. Bernoulli iid is not a realistic consumption model,
its aim is to prove the CAC concept. It requires only
measuring the probability of the ON state and the
maximum value of the consumption. A more realistic,
widely used model is the First Order Markovian model
[17]. This model can be described by a transition
probability matrix. As the most realistic model among the
three approaches we applied the distributions of the
holding times for ON and OFF states separately which
leads generally to a Higher Order Markovian (HOM)
model. The benefit of HOM models is the capability to
model long range dependence between samples, which is
a usual property of real load time series. In Figure 6
examples of iid and HOM time series can be seen. In all
the cases our models were fitted to measured data coming
from the REDD DataSet [18]. The DataSet contains
appliance level power data for 6 homes for several weeks
with sampling time of 3 seconds.
-5
10
-5
-4
10
10
-3
-2
10
10
-1
10
0
10
200
p
150
Figure 4 𝑝̃ vs 𝑝 for different bounds (pON=0,1)
Using Chernoff’s and Bennett’s inequalities the CAC
algorithm sets with one order of magnitude lower the ratio
𝑘 regardless of 𝑝𝑂𝑁 , which results only in an acceptable
decrease of the number of accepted appliances (for details
see Section 4.4).
Applying Hoeffding bound leads to results which
highly depend on the pON value. Applying Chebisev and
Markov bound lead to poor results regardless of the p ON
values. The performance of CLT based CAC is close to
the analytic calculation (𝑘 = 1 … 3). Note that CLT is not
an upper bound on the tail probability. As a consequence
𝑘 > 1 values can occur.
100
50
0
0
50
100
150
200
250
0
50
100
150
200
250
0
50
100
150
200
250
200
150
100
50
0
250
200
150
100
p~ vs p N= 400 len= 50000 pON = 0; 5 h = 1 Cmax = mean
0
10
50
analytic
0
clt
chernoff
-1
10
Figure 6 iid Bernoulli (top) and HOM (middle) model and original
measurement of a refrigerator (bottom)
bennett
hoeffding
chebisev
-2
p
~
10
-3
10
-4
10
-5
10
-5
10
-4
10
-3
-2
10
10
-1
10
p
Figure 5 𝑝̃ vs 𝑝 for different bounds (pON=0,5)
0
10
The CAC algorithm descripted by equations (3), (10)(14) assumes iid appliance load time series. It is an
important question, how complex time series models
affect CAC. Figure 7 demonstrates that there is only a
slight performance degradation even with the HOM
model (400 pieces of microwave ovens with ON
probability of 0,0160; simulation length is 50000 time
instances).
p~ vs p N= 400 len= 50000 pON = 0; 01608 h = 1 Cmax = mean
0
10
6500
original time series
modified time series
6000
analytic
clt
chernoff
bennett
chebisev
markov
hoeffding
-1
10
-2
5000
p
~
Load [W]
10
5500
-3
10
4500
4000
3500
3000
-4
10
2500
2000
-5
10
-5
10
-4
10
-3
-2
10
10
-1
10
1500
0
10
0
50
However, in the case of small probabilities (10−5 … 10−4)
CLT results in a higher ratio (𝑘 = 10 … 20), Chernoff and
Bennett remains almost in the same range (𝑘 = 0,1) like
in Figure 5 and 6.
4.3 Load shape modification made by CAC
However the basic mathematical idea of our CAC is
to limit the over- and underconsumption probability, the
direct objective of DSM methods is expressed as load
shape modification in the time domain (for instance by
valley filling and peak clipping). The CAC algorithm, as
stated before, forms the pdf of the aggregate consumption
towards the Dirac-delta function, which is equivalent to
constant load in the time domain. In this section we
demonstrate the effectiveness of the CAC algorithm
regarding load shaping. Assumptions are:
The consumption demand of the temporarily disabled
appliances is deleted;
Selection of the appliances to be temporarily disabled
is based on random selection which guarantees
fairness;
One appliance class;
All appliances are of shiftable stochastic type.
In Figure 8 the original aggregate consumption time series
and the modified one can be seen. From the figure one
can see, that this form of the algorithm does not yield
almost any load shaping. Our hypothesis was that the
treatment of consumption demand of the temporarily
disabled appliances (which is referred to as scheduling
strategy) plays key role in the algorithm to perform load
curve modification. To prove this, we changed the
scheduling strategy in the CAC to a so-called one-step
strategy.
150
200
250
Figure 8 Load shape modification ability of the CAC algorithm
The one-step scheduler (Figure 9) is an alternative
method to handle the disabled appliances. In this case our
assumptions are:
The one-step scheduler shifts the consumption of
temporarily disabled appliance with one time instant;
It guarantees that the sum of the consumed energy
remains the same after the modification of the load
curve;
Selection of the appliances to be temporarily disabled
is based on random selection which guarantees
fairness;
One appliance class;
All appliances are of schiftable stochastic type.
9000
original time series
modified time series
8000
7000
Load [W]
Figure 7 𝑝̃ vs 𝑝 with microwave oven HOM model
100
Time instances
p
6000
5000
4000
3000
0
50
100
150
200
250
Time instances
Figure 9 Load shape modification with one-step scheduling
It is clear that the CAC with one-step scheduler is
able to modify the load shape (red curve on Figure 9,
which is closer to constant). The Load Factor (LF) is
increased from 0,6718 to 0,8463 (LF is a widely used
measure of the efficiency of electric energy usage, and
𝑎𝑣𝑒𝑟𝑎𝑔𝑒 𝑙𝑜𝑎𝑑
calculated as 𝐿𝐹 =
). Based on the results,
𝑚𝑎𝑥𝑖𝑚𝑢𝑚 𝑙𝑜𝑎𝑑
we are planning to investigate more sophisticated
scheduling methods in our future work.
4.4 Number of enabled appliances in the case of
different LDT bounds and CLT
In the CAC algorithm the scheduler disables shiftable
appliances if the aggregate consumption exceeds the
Cmax upper limit with a higher probability than it is
allowed by 𝑝. The task in this step is to determine the
number of appliances to be enabled in each appliance
class so that the QoS must be satisfied. Figure 10 depicts
the number of enabled appliances in the case of different
LDT bounds and CLT, assuming:
One appliance class modelled with Bernoulli iid
model (𝑝𝑂𝑁 = 0,1);
400 appliances.
N= 400 len= 50000 pON = 0; 1 h = 1 Cmax = mean
400
analytic
clt
chernoff
bennett
hoeffding
chebisev
markov
Number of enabled appliances
350
300
250
200
Figure 11 Number of enabled appliances, one class (ℎ2 = 5)
150
Figure 11 shows that the decision curve is slightly
nonlinear and convex, but with other parameters (Figure
12) it can be highly nonlinear and even non-convex. We
can state that the two decision regions are generally not
linearly separable.
100
50
0
-5
10
-4
10
-3
-2
10
10
-1
0
10
10
p
Figure 10 Number of enabled appliances vs 𝑝
The number of enabled appliances is a monotonously
increasing function of 𝑝 in the case of one appliance class
(Figure 10). Estimation of the probability of
overconsumption applying LDT bounds lead to lower
number of enabled appliances compared to the
analytically calculated value for all 𝑝 values. Applying
LDT bounds, as stated before, causes spare capacities in
the system. CLT is not a bound, so it can lead to values
higher than 100%, which means breach of contract. The
exact percentages of enabled appliances (with analytically
calculated value as the reference) are collected in Table 2.
Table 2 Percentage of enabled appliances
Analytic
CLT
Chernoff
Bennett
Hoeffding
Chebisev
Markov
𝒑 < 𝟏𝟎−𝟑
𝒑 > 𝟏𝟎−𝟐
100%
(reference)
105%
92%
91%
80%
0%
0%
100%
(reference)
101%
88%
88%
75%
50-80%
10-50%
QoS
guaranteed
no
yes
yes
yes
yes
yes
In the case of two or more appliance classes, the CAC
algorithm can decide to enable different combinations of
appliances (Figure 11 and 12, where green colour
indicates the allowable set of appliances, red colour
indicates the combinations when the QoS is not satisfied).
Assumptions are:
Two appliance classes modelled with Bernoulli iid
model (100 appliances in each classes);
The tail probability is exactly calculated from the
analytical aggregate pdf;
Figure 12 Number of enabled appliances, one class (ℎ2 = 10)
The separator curve depends on the different LDT bounds
and CLT applied in CAC. The next two figures (Figure
13, 14) depict the investigations regarding the number of
enabled appliances in the case of different tail probability
estimation methods. Assumptions are:
Two appliance classes modelled with Bernoulli iid
model (100 appliances in each classes);
As a reference, the tail probability is exactly
calculated from the analytical aggregate pdf.
In the first experiment (Figure 13) pON1=0,2 and
pON2=0,001; and ON values h1=1W and h2=5W. In the
second experiment (Figure 14) the difference is only
ℎ2 = 10𝑊. The performance degradation is smaller in the
first case when the difference between ON values ratio
ℎ2 /ℎ1 is not too large. In the case of higher ℎ2 /ℎ1 ratio
(Figure 14) the separator curves lie far to each other
causing severe performance degradation.
100
Num. of en. apps to satisfy QoS p = 0; 0001 Cmax = 1; 5 $ mean
90
analytic
clt
N 2 (pON 2 = 0; 001 h2 = 5)
80
chernoff
bennett
70
hoeffding
60
50
40
30
20
10
10
20
30
40
50
60
70
80
90
100
N 1 (pON 1 = 0; 2 h1 = 1)
Figure 13 Number of enabled appliances, two classes (𝒉𝟐 = 𝟓)
In the latter case (Figure 14) the exact separator is
non-linear and non-convex but this fact is not reflected by
the estimation methods.
100
Num. of en. apps to satisfy QoS p = 0; 0001 Cmax = 1; 5 $ mean
90
analytic
clt
chernoff
bennett
N 2 (pON 2 = 0; 001 h2 = 10)
80
70
60
50
40
30
20
10
10
20
30
40
50
60
70
80
90
100
N 1 (pON 1 = 0; 2 h1 = 1)
Figure 14 Number of enabled appliances, two classes (𝒉𝟐 = 𝟏𝟎)
In Table 3 and 4 the number of enabled appliances
can be seen for certain parameters.
Table 3 Percentage of enabled appliances ℎ2 = 5
N2=100
N2=50
N2=0
Analytic
64*
100%
(reference)
71
100%
(reference)
80
100%
(reference)
CLT
80
110%
Chernoff
50
91%
Bennett
35
82%
82
109%
58
89%
38
72%
85
106%
72
90%
42
53%
*N1
Table 4 Percentage of enabled appliances ℎ2 = 10
N2=100
N2=50
N2=0
Analytic
11*
CLT
69
reference
38
reference
76
reference
152%
77
144%
88
116%
Chernoff
0
(N2=60)
54%
10
68%
60
79%
Bennett
0
(N2=60)
54%
6
64%
21
28%
* N1
The performance decrease caused by the different
LDT bounds is the smallest in the case of Chernoff bound
but it is highly sensitive to the ℎ2 /ℎ1 ratio. In the case of
ℎ2 /ℎ1 = 5 the utilization loss caused by Chernoff bound
is 9-11%. In the case of ℎ2 /ℎ1 = 10 it is 21-46%. CLT
has near the same performance but in the experiments the
number of enabled appliances is higher than the reference
which causes breach of contract regarding the QoS
criterion 𝑝. At the same time the computational
complexity of CLT is substantially lower than of Chernoff
bound and analytical convolution. As a result we
recommend using analytical computation when it is
possible. In the case of lack of time and importance of
satisfying QoS, Chernoff bound comes to the fore. CLT
has the lowest computational need and has quite good
performance but cannot guarantee QoS criterion.
5
Conclusions and future work
In this paper a new statistical approach was proposed
for managing the balance between demand and available
supplies in smart grids. The smart meter of the subscriber
performs the task of enabling/disabling of shiftable
appliances based on two parameters, obtained from the
supplier: upper capacity limit and allowable probability of
overconsumption (QoS). The smart meter influences the
probability distribution function of the aggregate
consumption in order to keep the tail probabilities under a
given threshold 𝑝. The new approach takes the
uncertainty of the consumption into account, and
furthermore it can work in a fully distributed manner,
since the calculations can be performed in the smart
meter. We conducted several simulations to evaluate the
performance of the CAC. As a result the introduced
Consumption Admission Control method is a promising
candidate for demand side management in smart grid
environment.
6
Acknowledgement
Our research was supported by the European Union
and the Hungarian Republic through the project TÁMOP4.2.2.A-11/1/KONV-2012-0072
–
Design
and
optimization of modernization and efficient operation of
energy supply and utilization systems using renewable
energy sources and ICTs.
7
References
[1] Lukas G. Swan, V. Ismet Ugursal, Modeling of end-use
energy consumption in the residential sector: A review of
modeling techniques, Renewable and Sustainable Energy
Reviews, Volume 13, Issue 8, October 2009, Pages 18191835, ISSN 1364-0321
[2] Pasaoglu G.; Fiorello D.; Martino A.; Scarcella G.;
Alemanno A.; Zubaryeva A.; Thiel C.; "Driving and
parking patterns of European car drivers – a mobility
survey", European Commission, DG JRC, Institute for
Energy and Transport, Petten, the Netherlands, 2012,
http://publications.jrc.ec.europa.eu/repository/handle/JRC7
7079
[3] Khan, M.A.; Javaid, N.; Arif, M.; Saud, S.; Qasim, U.; Khan,
Z.A., "Peak Load Scheduling in Smart Grid
Communication Environment," Advanced Information
Networking and Applications (AINA), 2014 IEEE 28th
International Conference on , vol., no., pp.1025,1032, 13-16
May 2014
[4] Ullah, M. N.; Mahmood, A.; Razzaq, S.; Ilahi, M.; Khan, R.
D.; Javaid, N. (2013). A Survey of Different Residential
Energy Consumption Controlling Techniques for
Autonomous DSM in Future Smart Grid Communications.
J. Basic. Appl. Sci. Res., 3(3)1207-1214, 2013
[5] Fang, Xi; Misra, Satyajayant; Xue, Guoliang; Yang, Dejun,
"Smart Grid — The New and Improved Power Grid: A
Survey," Communications Surveys & Tutorials, IEEE,
vol.14, no.4, pp.944,980, Fourth Quarter 2012
[6] Koutitas, G., "Control of Flexible Smart Devices in the
Smart Grid," Smart Grid, IEEE Transactions on, vol.3,
no.3, pp.1333,1343, Sept. 2012
[7] Phillip Oliver Kriett, Matteo Salani, “Optimal control of a
residential microgrid”, Energy, Volume 42, Issue 1, June
2012, Pages 321-330, ISSN 0360-5442
[8] G. T. Costanzo, Zhu Guchuan, M. F. Anjos, and G.
Savard,“A System Architecture for Autonomous Demand
Side Load Management in Smart Buildings,” IEEE
Transactions on Smart Grid, vol. 3, no. 4, pp. 2157–2165,
Dec. 2012.
[9] A. H. Mohsenian-Rad, V. W. S. Wong, J. Jatskevich, R.
Schober, and A. Leon-Garcia, “Autonomous Demand-Side
Management Based on Game-Theoretic Energy
Consumption Scheduling for the Future Smart Grid,” IEEE
Transactions on Smart Grid, vol.1, no. 3, pp. 320–331, Dec.
2010.
[10] P. Samadi, A. H. Mohsenian-Rad,V. W. S. Wong, and R.
Schober,“Tackling the Load Uncertainty Challenges for
Energy Consumption Scheduling in Smart Grid,” IEEE
Transactions on Smart Grid, vol. 4, no. 2, pp. 1007–1016,
June 2013.
[11] Levendovszky, J, E.C. van der Meulen,: "Tail Distribution
Estimation for Call Admission Control in ATM Networks",
Proccedings of IFIP, Third Workshop on Performance
Modelling and Evaluation of ATM Networks, Ilkley, West
Yorkshire, UK, 2-6th July 1995.
[12] Z. Lin; Z. Bai, Probability inequalities. Springerverlag
Berlin Heidelberg, 2010.
[13] W. Hoeffding, "Probability inequalities for sums of
bounded random variables", Journal of the American
statistical association, vol 58, sz 301, o 13–30, 1963.
[14] G. Bennett, „Probability inequalities for the sum of
independent random variables”, Journal of the American
Statistical Association, vol 57, sz 297, o 33–45, 1962.
[15] H. Chernoff, „A measure of asymptotic efficiency for tests
of a hypothesis based on the sum of observations”, The
Annals of Mathematical Statistics, vol 23, sz 4, o 493–507,
1952.
[16] DasGupta, Anirban. Asymptotic theory of statistics and
probability. Springer, 2008.
[17] O. Ardakanian, S. Keshav, and C. Rosenberg, “Markovian
models for home electricity consumption,” Proc. 2nd ACM
SIGCOMM Work. Green Netw. - GreenNets ’11, p. 31,
2011.
[18] J. Zico Kolter and Matthew J. Johnson. REDD: A public
data set for energy disaggregation research. In proceedings
of the SustKDD workshop on Data Mining Applications in
Sustainability, 2011.
Authors’ addresses
Lorant Kovacs, PhD
Kecskemet College, Faculty of Mechanical Engineering and
Automation (GAMF)
Izsaki ut 10, Kecskemet, H-6000, Hungary
[email protected]
Rajmund Drenyovszki, MSc
Kecskemet College, Faculty of Mechanical Engineering and
Automation (GAMF)
Izsaki ut 10, Kecskemet, H-6000, Hungary
[email protected]
Andras Olah, PhD
Pázmány Péter Catholic University, Faculty of Information
Technology
Práter utca 50/a, H-1083 Budapest, Hungary
[email protected]
Janos Levendovszky, PhD
Budapest University of Technology and Economics, Deptartment
of Telecommunications,
Egry József u. 18, H-1111 Budapest, Hungary
[email protected]
Kalman Tornai,PhD
Pázmány Péter Catholic University, Faculty of Information
Technology
Práter utca 50/a, H-1083 Budapest, Hungary
[email protected]
Istvan Pinter, PhD
Kecskemet College, Faculty of Mechanical Engineering and
Automation (GAMF)
Izsaki ut 10, Kecskemet, H-6000, Hungary
[email protected]
| 3cs.SY
|
Logical Methods in Computer Science
Vol. 2 (3:4) 2006, pp. 1–42
www.lmcs-online.org
Submitted
Published
Jan. 25, 2006
Sep. 13, 2006
RESOURCE USAGE ANALYSIS FOR THE π-CALCULUS ∗
NAOKI KOBAYASHI a , KOHEI SUENAGA b , AND LUCIAN WISCHIK c
a
6-3-09, Aoba, Aramaki-aza, Aoba-ku, Sendai, Miyagi, Japan
e-mail address: [email protected]
b
6-3-09, Aoba, Aramaki-aza, Aoba-ku, Sendai, Miyagi, Japan
e-mail address: [email protected]
c
1 Microsoft Way, Redmond, WA 98052, USA
e-mail address: [email protected]
Abstract. We propose a type-based resource usage analysis for the π-calculus extended
with resource creation/access primitives. The goal of the resource usage analysis is to
statically check that a program accesses resources such as files and memory in a valid
manner. Our type system is an extension of previous behavioral type systems for the πcalculus. It can guarantee the safety property that no invalid access is performed, as well as
the property that necessary accesses (such as the close operation for a file) are eventually
performed unless the program diverges. A sound type inference algorithm for the type
system is also developed to free the programmer from the burden of writing complex type
annotations. Based on our algorithm, we have implemented a prototype resource usage
analyzer for the π-calculus. To the authors’ knowledge, this is the first type-based resource
usage analysis that deals with an expressive concurrent language like the π-calculus.
1. Introduction
Computer programs access many external resources, such as files, library functions,
device drivers, etc. Such resources are often associated with certain access protocols; for
example, an opened file should be eventually closed and after the file has been closed,
no read/write access is allowed. The aim of resource usage analysis [11] is to statically
check that programs conform to such access protocols. Although a number of approaches,
including type systems and model checking, have been proposed so far for the resource
usage analysis or similar analyses [5, 6, 7, 11, 1], most of them focused on analysis of
sequential programs, and did not treat concurrent programs, especially those involving
dynamic creation/passing of channels and resources.
In the present paper, we propose a type-based method of resource usage analysis for
concurrent languages. Dealing with concurrency is especially important because concurrent
2000 ACM Subject Classification: D.2.4, D.3.1, F.3.1, F.3.2.
Key words and phrases: Type System, π-Calculus, Verification of Concurrent Programs, Resource Usage
Analysis.
∗
A preliminary version appeared in 7th International Conference on Verification, Model Checking, and
Abstract Interpretation (VMCAI 2006), Charleston, SC, USA, January 8–10, 2006.
l
LOGICAL METHODS
IN COMPUTER SCIENCE
c
DOI:10.2168/LMCS-2 (3:4) 2006
CC
N. Kobayashi, K. Suenaga, and L. Wischik
Creative Commons
2
N. KOBAYASHI, K. SUENAGA, AND L. WISCHIK
programs are hard to debug, and also because actual programs accessing resources are often
concurrent. We use the π-calculus (extended with resource primitives) as a target language
so that our analysis can be applied to a wide range of concurrency primitives (including
those for dynamically creating and passing channels) in a uniform manner.
A main new difficulty in dealing with concurrent programs is that control structures are
more complex in concurrent programs than in sequential programs. For example, consider
the following process P1 :
(νc) (read(x).ch i | c( ). close(x))
Here, read(x).ch i reads x and then sends a signal on channel c, and in parallel to that,
c( ). close(x) waits for a signal on channel c and then closes x. Because of the synchronization through channel c, x is closed only after being read. To capture this kind of causal
dependency between communications and resource access, we use CCS processes as extra
type information (which are called behavioral types). For example, the above process is
given the behavioral type (νc) (xR .c | c. xC ).
Using the behavioral types introduced above, we can construct a type system for resource usage analysis in a manner similar to previous behavioral type systems for the πcalculus [10, 3]. A type judgment is of the form Γ ⊲ P : A, where Γ is the usual type
environment and A is a behavioral type approximating the behavior of P on the free channels and resources. For example, the above process P1 is typed x : res⊲P1 : (νc) (xR .c | c. xC ).
Behavioral types are also used to augment channel types. The judgment for s(x). P1 is given
by:
Γ ⊲ s(x). P1 : s
R
where Γ = s:chanh(x:res)(νc) (x .c | c. xC )i. Here, the behavioral type of s(x). P1 is simply
a single input command s: the characteristic feature of this kind of type system is that the
behavior of the input continuation is accounted for at output, not at input. The channel
s has argument type (x:res)(νc) (xR .c | c. xC ), which specifies that the resource sent along
channel s will be read first and then closed. Using the same type environment, the output
process shri is typed as:
Γ, r:res ⊲ shri : s. (νc) (r R .c | c. r C )
Here the behavioral type is an output followed by a continuation. The continuation
(νc) (r R .c | c. r C ) has been obtained by substituting r for x in the argument type of s. In this
way, the types propagate information about how resources and channels passed thorough
channels are accessed.
An important property of our type system is that types express abstract behavior of
processes, so that certain properties of processes can be verified by verifying the corresponding properties of their types, using, for example, model checking techniques. The
latter properties (of behavioral types) are more amenable to automatic verification techniques like model checking than the former ones, because the types do not have channel
mobility and also because the types typically represent only the behavior of a part of the
entire process.
The technical contributions of the present work are summarized as follows.
• Formalization of type systems for resource usage analysis for the π-calculus, and
proof of their soundness. We have augmented previous behavioral types for the πcalculus with hiding and renaming constructors, and adapted them to the problem
RESOURCE USAGE ANALYSIS FOR THE π-CALCULUS
3
of resource usage analysis. CCS-like processes have been used as types also in previous work on type systems for the π-calculus [10, 3]. Igarashi and Kobayashi [10],
however, used a fragment without hiding and renaming, and Chaki et al. [3] used a
fragment without renaming, while the present paper uses both hiding and renaming. The inclusion of hiding and renaming is important both for accuracy and for
automatic inference (see Remark 3.12).
• Realization of fully automatic verification (while making the analysis more precise
than [10]). Igarashi and Kobayashi [10] gave only an abstract type system, without
giving a concrete type inference algorithm. Chaki et al. [3] requires type annotations.
The full automation was enabled by a combination of a number of small ideas, like
inclusion of hiding and renaming as type constructors, and approximation of a CCSlike type by a Petri net (to reduce the problem of checking conformance of inferred
types to resource usage specification).
• Verification of not only the usual safety property that an invalid resource access
does not occur, but also an extended safety (which we call partial liveness) that
necessary resource accesses (e.g. closing of a file) are eventually performed unless
the whole process diverges. The partial liveness is not guaranteed by Chaki et al.’s
type system [3]. A noteworthy point about our type system for guaranteeing the
partial liveness is that it is parameterized by a mechanism that guarantees deadlockfreedom (in the sense of Kobayashi’s definition [14]). So, our type system can be
combined with any mechanism (model checking, abstract interpretation, another
type system, or whatever) to verify deadlock-freedom for deadlock- or lock-freedom
(e.g., Yoshida’s graph type system [25]).
• Implementation of a prototype resource usage analyzer based on the proposed
method. The implementation can be tested at http://www.yl.is.s.u-tokyo.ac.jp/~kohei/usage-p
The rest of this paper is structured as follows. Section 2 introduces an extension of
the π-calculus with primitives for creating and accessing resources. Section 3 introduces a
type system for resource usage analysis, which guarantees that well-typed processes never
perform an invalid resource access. Section 4 gives a type inference algorithm for the type
system. Section 5 extends the type system to guarantee that necessary resource accesses
(such as closing of opened files) are eventually performed (unless the program diverges).
Section 6 describes a prototype resource usage analyzer we have implemented based on the
present work. Section 7 discusses related work. Section 8 concludes.
2. Processes
This section introduces the syntax and the operational semantics of our target language.
2.1. Syntax.
Definition 2.1 (processes). The set of processes is defined by the following syntax.
P (processes) ::= 0 | xhv1 , . . . , vn i. P | x(y1 , . . . , yn ). P
| (P | Q) | if v then P else Q
| (νx) P | ∗P | accξ (x).P | (NΦ x)P
::= x | true | false
v (values)
4
N. KOBAYASHI, K. SUENAGA, AND L. WISCHIK
Here, x, y, and z range over a countably infinite set Var of variables. ξ ranges over a set of
labels called access labels. Φ, called a trace set, denotes a set of sequences of access labels
that is prefix-closed. The prefixes (like (νx) and (NΦ x)) bind tighter than the parallel
composition | .
An access label specifies the kind of an access operation. Typical access labels that we
are going to use in this paper are: I for initialization, R for read, W for write, and C for
close.
Process accξ (x).P accesses the resource x, and then behaves like P . We will often write
init(x).P , read(x).P , write(x).P , and close(x).P for accI (x).P , accR (x).P , accW (x).P ,
accC (x).P . Process (NΦ x)P creates a new resource with the bound name x that should be
accessed according to Φ, and then behaves like P . Φ specifies a set of acceptable sequences
∗
#
of operations that are allowed for the new resource x. For example, (N(I(R+W ) C) x)P
creates a resource that should be first initialized, read or written an arbitrary number of
times, and then closed. Here, (S)# is the prefix closure of S, i.e., {s | ss′ ∈ S}. We write ǫ
for the empty sequence.
v i. P and x(e
y ). P for
We often abbreviate a sequence v1 , . . . , vn to ve, and write xhe
xhv1 , . . . , vn i. P and x(y1 , . . . , yn ). P . We often omit trailing 0 and write xhe
v i and accξ (x)
v i. 0 and accξ (x).0 respectively.
for xhe
The bound and free variables of P are defined in a customary manner; also (NΦ x)P
binds x. We identify processes up to α-conversion, and assume that α-conversion is implicity
applied so that bound variables are always different from each other and from free variables.
2.2. Operational Semantics. We now formally define the operational semantics of our
process calculus The operational semantics is almost the same as the standard reduction
semantics for the π-calculus, except that trace sets Φ (which represent how resources should
be accessed in future) may change during reduction.
Definition 2.2. The structural preorder is the least reflexive and transitive relation
closed under the rules in Figure 1 (P ≡ Q stands for (P Q) ∧ (Q P )).
Remark 2.3. As in our previous behavioural type systems for the π-calculus [10, 14, 15],
the structural relation is asymmetric. If the standard, symmetric structural relation were
used, the type preservation property would not hold: Γ ⊲ ∗P | P : A does not necessarily
imply Γ ⊲ ∗P : A) for the type system introduced in the next section.
Definition 2.4. The set of reduction labels, ranged over by L, is {xξ | x ∈ Var} ∪ {τ }. We
define target(L) by:
target (xξ ) = {x}
target (τ ) = ∅
Definition 2.5. Let Φ be a set of sequences of access labels. Φ−ξ is defined by:
{s | ξs ∈ Φ}.
Φ−ξ =
L
Definition 2.6. The reduction relation −→ is the least relation closed under the rules in
Figure 2.
L
We write P −→ Q when P −→ Q for some L. We write −→∗ for the reflexive and
transitive closure of −→.
Notice that when an invalid access to a resource occurs (i.e. when the program accesses
ξ but the specification Φ has no ξ-prefixes), then resource specification Φ is set to ∅ by
RESOURCE USAGE ANALYSIS FOR THE π-CALCULUS
P |0 ≡ P
5
(SP-Zero)
P |Q ≡ Q|P
(SP-Commut)
P | (Q | R) ≡ (P | Q) | R
(SP-Assoc)
∗P ∗P | P
(SP-Rep)
(νx) P | Q (νx) (P | Q)(if x not free in Q)
(SP-New)
(NΦ x)P | Q (NΦ x)(P | Q)(if x not free in Q)
(SP-NewR)
P P′
Q Q′
P | Q P ′ | Q′
(SP-Par)
P Q
(νx) P (νx) Q
(SP-CNew)
P Q
(N x)P (NΦ x)Q
(SP-CNewR)
Φ
Figure 1: Structural Preorder
L
P −→ Q
τ
z /e
y]Q (R-Com)
z i. P | x(e
y ). Q −→ P | [e
xhe
xξ
(R-New)
(νx) P −→ (νx) Q
(R-Acc)
accξ (x).P −→ P
x 6∈ target(L)
L
xξ
P −→ Q
τ
−ξ
(NΦ x)P −→ (NΦ x)Q
(R-NewR1)
L
P −→ Q
(R-Par)
L
P | R −→ Q | R
L
P −→ Q
x 6∈ target (L)
L
(NΦ x)P −→ (NΦ x)Q
τ
if true then P else Q −→ P
τ
(R-NewR2)
(R-IfT)
if false then P else Q −→ Q (R-IfF)
P P′
L
P ′ −→ Q′
L
Q′ Q
(R-SP)
P −→ Q
Figure 2: Reduction Relation
(R-NewR1). On the other hand Φ ⊇ {ǫ} indicates a resource that has been correctly used
so far, and Φ = {ǫ} indicates one that has been correctly and completely used.
Definition 2.7. A process P is resource-safe if it does not contain a sub-expression of the
form (N∅ x)Q.
6
N. KOBAYASHI, K. SUENAGA, AND L. WISCHIK
We give a type system guaranteeing that any resource-safe, well-typed process cannot
be reduced to a non-safe process (in other words, any resource-safe, well-typed process never
performs an invalid access) in Section 3.
Example 2.8. The following process first creates a resource x that should be first initialized,
read an arbitrary number of times, and then closed. It then spawns four processes; they
synchronize through channels c1 and c2 , so that x is accessed in a valid order.
∗
2
#
(N(IR C) x)(νc1 ) (νc2 )
init(x).(c1 h i | c1 h i) /* initialize x, and send signals */
| c1 ( ). read(x).c2 h i /* wait for a signal on c1 ,
then read x, and signal on c2 */
| c1 ( ). read(x).c2 h i /* wait for a signal on c1 ,
then read x, and signal on c2 */
| c2 ( ). c2 ( ). close(x) /* wait on c2 , then close x */
Example 2.9. The following program is prototypical of recursive functions. There is a
replicated service which listens on channel s; it either terminates the recursion by sending a
message back on the reply channel r, or it recursively invokes a sub-instance of itself which
will reply on a private channel r ′ . In this example each recursive step does a read(x). The
following program use an integer to decide whether or not to recurse. Though our language
does not have integers and operations on them as primitives, it is trivial to extend our
language and type system with those primitives.
(νs) ∗(s(n, x, r). if n = 0 then rhi
else (νr ′ ) (shn − 1, x, r ′ i | r ′ (). read(x).rhi)
∗
#
| (N(IR C) x)(νr) (init(x).sh100, x, ri | r(). close(x))
2 The above program corresponds to the following higher-level program:
init(x); parbegin read(x); read(x) parend; close(x)
Example 2.10. Consider the following producer/consumer program:1
(νproducer ) (νconsumer )
∗(producer (b, p, c). p(). accP (b).(chi | producer hb, p, ci)) |
∗(consumer (b, p, c). c(). acc G (b).(phi | producer hb, p, ci)) |
∗ #
(N((P G) ) buf )(νx) (νy)
∗(producer hbuf , x, yi) | ∗(consumer hbuf , x, yi) | xhi
The first two processes ∗(producer (b, p, c). · · · ) and
∗(consumer (b, p, c). · · · ) define the behavior of producers and consumers. A producer repeatedly waits to receive a signal on p, performs a put on the buffer b (by accP (b)), and
then sends a signal on c. A consumer repeatedly waits to receive a signal on c, performs a
get on the buffer b (by accP (b)), and then sends a signal on p. The third process creates a
new buffer on which put and get should be applied only alternately, creates two channels x
and y used for synchronization, and runs infinitely many producers and consumers.
Remark 2.11. We treat resources as primitives in this paper, but we could alternatively
express a resource as a tuple of channels, each of which corresponds to each access operation. For example, the resource in Example 2.8 can be expressed as a tuple consisting of
1This is an example taken from an ealier version of [20] and modified.
RESOURCE USAGE ANALYSIS FOR THE π-CALCULUS
7
three channels init, read, and close. If we did so, we could directly reuse the previous type
systems [10, 3] to infer some of the properties discussed in this paper (with different precision). Treating resources as primitives, however, simplifies the type systems introduced in
later sections and clarifies the essence: if we expressed a resource as a tuple of channels, we
would need primitives for simultaneous creation of multiple channels as in [10], and need to
care about whether communications on the resource access channels succeed or not. On the
other hand, our resource access primitives are non-blocking, which simplifies in particular
the extended type system discussed in Section 5.
3. Type System
This section introduces a type system that prevents invalid access to resources. The type
system in this section does not guarantee a liveness property that all the necessary accesses
are eventually made; extensions to guarantee that property are discussed in Section 5.
3.1. Types. We first introduce the syntax of types. We use two categories of types: value
types and behavioral types. The latter describes how a process accesses resources and
communicates through channels. As mentioned in Section 1, we use CCS processes for
behavioral types.
Definition 3.1 (types). The sets of value types σ and behavioral types A are defined by:
σ ::= bool | res | chanh(x1 : σ1 , . . . , xn : σn )Ai
A ::= 0 | α | a.A | xξ .A | τ.A | (A1 | A2 ) | A1 ⊕ A2 | ∗A
| hy1 /x1 , . . . , yn /xn iA | (νx) A | µα.A | A↑S | A↓S
a (communication labels) ::= x | x
A behavioral type A, which is a CCS process, describes what kind of communication
and resource access a process may perform. 0 describes a process that performs no communication or resource access. The types x. A, x. A, xξ .A and τ.A describe processes that
first perform an action and then behave according to A; the actions are, respectively, an
input on x, an output on x, an access operation ξ on x, and the invisible action. A1 | A2
describes a process that performs communications and resource access according to A1 and
A2 , possibly in parallel. A1 ⊕A2 describes a process that behaves according to either A1 or
A2 . ∗A describes a process that behaves like A an arbitrary number of times, possibly in
parallel. hy1 /x1 , . . . , yn /xn iA, abbreviated to he
y /e
xiA, denotes simultaneous renaming of x
e
with ye in A. (νx) A describes a process that behaves like A for some hidden channel x. For
example, (νx) (x. y | x) describes a process that performs an output on y after the invisible
action on x. The type µα.A describes a process that behaves like a recursive process de△
fined by α = A.2 The type A↑S describes a process that behaves like A, except that actions
whose targets are in S are replaced by the invisible action τ , while A↓S describes a process
that behaves like A, except that actions whose targets are not in S are replaced by τ . The
formal semantics of behavioral types is defined later using labeled transition semantics.
As for value types, bool is the type of booleans. res is the type of resources. The
type chanh(x1 : σ1 , . . . , xn : σn )Ai, abbreviated to chanh(e
x:σ
e)Ai, describes channels carrying tuples consisting of values of types σ1 , . . . , σn . Here the type A approximates how a
2The replication ∗A and µα.(A | α) have the same semantics in this section, but they are differentiated
in Section 5 by the predicate disabled.
8
N. KOBAYASHI, K. SUENAGA, AND L. WISCHIK
receiver on the channel may use the elements x1 , . . . , xn of each tuple for communications
and resource access. For example, chanh(x : res, y : res)xR .y C i describes channels carrying
a pair of resources, where a party who receives the actual pair (x′ , y ′ ) will first read x′ and
then close y ′ . We sometimes omit σ
e and write chanh(e
x)Ai for chanh(e
x:σ
e)Ai. When x
e is
empty, we also write chanhi.
Note that he
y /e
xi is treated as a constructor rather than an operator for performing the
actual substitution. We write [e
y /e
x] for the latter throughout this paper. he
y /e
xiA is slightly
different from the relabeling of the standard CCS [19]: hy/xi(x | y) allows the communication
on y, but the relabeling of CCS does not. This difference calls for the introduction of a
special transition label {x, y} in Section 3.2.
Definition 3.2. The set of free variables of A, written FV(A), is defined by:
FV(0)
FV(α)
FV(x. A)
FV(x. A)
FV(xξ .A)
FV(τ.A)
FV(A1 | A2 )
FV(A1 ⊕ A2 )
FV(∗A)
FV(he
y /e
xiA)
FV((νx) A)
FV(µα.A)
FV(A↑S )
FV(A↓S )
=
=
=
=
=
=
=
=
=
=
=
=
=
=
∅
∅
{x} ∪ FV(A)
{x} ∪ FV(A)
{x} ∪ FV(A)
FV(A)
FV(A1 ) ∪ FV(A2 )
FV(A1 ) ∪ FV(A2 )
FV(A)
(FV(A)\{e
x}) ∪ {e
y}
FV(A)\{x}
FV(A)
FV(A)\S
FV(A) ∩ S
As defined above, (νx) A, he
y /e
xiA, and A↑S bind x, x
e, and the variables in S respectively.
We identify behavioral types up to renaming of bound variables. In the rest of this paper,
we require that every channel type chanh(x1 : σ1 , . . . , xn : σn )Ai must satisfy FV(A) ⊆
{x1 , . . . , xn }. For example, chanh(x:res)xR i is a valid type but chanh(x:res)y R i is not.3
l
3.2. Semantics of behavioral types. We give a labeled transition relation −→ for behavioral types. The transition labels l (distinct from the reduction labels L of Definition 2.4)
are
l ::= x | x | xξ | τ | {x, y}
The label {x, y} indicates the potential to react in the presence of a substitution that
identifies x and y. We also extend target to the function on transition labels by:
target (x) = target (x) = {x}
target({x, y}) = {x, y}
l
The transition relation −→ on behavioral types is the least relation closed under the rules
τ
l
in Figure 3. We write =⇒ for the reflexive and transitive closure of −→. We also write =⇒
l
for =⇒−→=⇒.
3This constraint can be removed if we assume that the free variables in codom(Γ) never clash with the
bound variables of P in the judgment form Γ⊲P : A given later. In particular, we need an implicit assumption
{e
y }∩FV(Γ)=∅ in Figure 4, (T-In).
RESOURCE USAGE ANALYSIS FOR THE π-CALCULUS
xξ
a
τ
xξ .A → A
a.A→A
(Tr-Act)
τ.A→A
l
A1 →A′1
A2 →A′2
l
A1 |A2 →A1 |A′2
l
y
x
(Tr-Par1)
l
A1 |A2 →A′1 |A2
y
A2 →A′2
A1 →A′1
9
x
A2 →A′2
A1 →A′1
{x,y}
(Tr-Par2)
{x,y}
A1 |A2 −→ A′1 |A′2
A1 |A2 −→ A′1 |A′2
{x,x}
A −→ A′
(Tr-Com)
τ
A −→ A′
l
l
A2 →A′2
A1 →A′1
l
(Tr-Or)
l
A1 ⊕A2 →A′1
A1 ⊕A2 →A′2
l
A | ∗A −→ A′
(Tr-Rep)
l
∗A −→ A′
l
[µα.A/α]A −→ A′
(Tr-Rec)
l
µα.A −→ A′
l
A −→ A′
(Tr-Rename)
[e
y /e
x]l
he
y /e
xiA −→ he
y /e
xiA′
l
A −→ A′
target (l)∩{x} = ∅
(Tr-Hiding)
l
(νx) A −→ (νx) A′
l
A→A′
target (l)⊆S
τ
′↑
A↑S →A
l
A→A′
target(l)∩S=∅
l
A↑S →A′ ↑S
S
target (l)⊆S
l
l
A→A′
l
A→A′
target(l)∩S=∅
τ
A↓S →A′ ↓S
′
A↓S →A ↓S
(Tr-Exclude)
(Tr-Project)
Figure 3: Transition semantics of behavioral types
Remark 3.3. (νx) A should not be confused with A↑{x} . (νx) A is the hiding operator
of CCS, while A↑{x} just replaces any actions on x with τ [10]. For example, (νx) (x. y ξ )
τ
yξ
cannot make any transition, but (x. y ξ )↑{x} −→−→ 0↑{x} .
The set tracesx (A) defined below is the set of possible access sequences on x described
by A.
Definition 3.4 (traces).
xξ1
xξn
tracesx (A) = {ξ1 . . . ξn | A↓{x} =⇒ · · · =⇒ A′ }
Note that tracesx (A) is prefix-closed (hence a trace set) by definition.
10
N. KOBAYASHI, K. SUENAGA, AND L. WISCHIK
We define the subtyping relation A1 ≤ A2 below. Intuitively, A1 ≤ A2 means that
a process behaving according to A1 can also be viewed as a process behaving according to
A2 . To put in another way, A1 ≤ A2 means that A2 simulates A1 .We define ≤ for only
closed types, i.e., those not containing free type variables.
Definition 3.5 (subtyping). The subtyping relation ≤ on closed behavioral types is the
l
l
largest relation such that A1 ≤ A2 and A1 −→ A′1 implies A2 =⇒ A′2 and A′1 ≤ A′2 for
some A′2 .
We often write A1 ≥ A2 for A2 ≤ A1 , and write A1 ≈ A2 for A1 ≤ A2 ∧ A2 ≤ A1 .
Remark 3.6. Note that the subtyping relation defined here is the converse of the one used
in Igarashi and Kobayashi’s generic type system [10]. This is due to two different, dual
views on behavioral types. Here, we think of behavioral types as describing the behavior
of processes. On the other hand, Igarashi and Kobayashi [10] think of behavioral types as
describing the assumption on the environment about what kind of process is accepted by
the environment. Because of this difference, they write behavioral types on the lefthand
side of ⊲, and write A1 &A2 for non-deterministic choice instead of A1 ⊕ A2 .
Remark 3.7. Depending on what property the type system should guarantee, a finer
subtyping relation may need to be chosen. For example, the above definition allows
(xW .0) | (xW .0) ≤ xW .xW .0. We may want to disallow this relation if we want to infer a
property like “no simultaneous writes on x can occur.”
The following properties are satisfied by ≤ . For proofs, see Appendix A.
Lemma 3.8.
(1) ≤ is a precongruence, i.e., ≤ is closed under any behavioral type constructor.
(2) If A1 ≤ A2 , then tracesx (A1 ) ⊆ tracesx (A2 ) for any x.
(3) B1 ⊕ B2 ≤ A if and only if B1 ≤ A and B2 ≤ A .
(4) If [B/α]A ≤ B, then µα.A ≤ B.
3.3. Typing. We consider two kinds of judgments, Γ ⊲ v : σ for values, and Γ ⊲ P : A for
processes. Γ is a mapping from a finite set of variables to value types. In Γ ⊲ P : A, the type
environment Γ describes the types of the variables, and A describes the possible behaviors
of P . For example, x : chanh(b : bool)0i ⊲ P : x | x implies that P may send booleans along
the channel x twice. The judgment y : chanh(x : chanh(b : bool)0i)xi ⊲ Q : y means that Q
may perform an input on y once, and then it may send a boolean on the received value.
Note that in the judgment Γ ⊲ P : A, the type A is an approximation of the behavior of P
on free channels. P may do less than what is specified by A, but must not do more; for
example, x : chanh( )0i ⊲ xh i : x | x holds but x : chanh( )0i ⊲ xh i. xh i : x does not. Because
of this invariant, if A does not perform any invalid access, neither does P .
We write dom(Γ) for the domain of Γ. We write ∅ for the empty type environment,
and write x1 : τ1 , . . . , xn : τn (where x1 , . . . , xn are distinct from each other) for the type
environment Γ such that dom(Γ) = {x1 , . . . , xn } and Γ(xi ) = τi for each i ∈ {1, . . . , n}.
When x 6∈ dom(Γ), we write Γ, x : τ for the type environment ∆ such that dom(∆) =
dom(Γ) ∪ {x}, ∆(x) = τ , and ∆(y) = Γ(y) for y ∈ dom(Γ). We define the value judgment
relation Γ ⊲ v:σ to be the least relation closed under
Γ, x:σ ⊲ x:σ
Γ ⊲ true:bool
Γ ⊲ false:bool.
RESOURCE USAGE ANALYSIS FOR THE π-CALCULUS
11
We write Γ ⊲ ve:e
σ as an abbreviation for (Γ ⊲ v1 :σ1 ) ∧ · · · ∧ (Γ ⊲ vn :σn ).
Definition 3.9. The type judgment relation Γ ⊲ P : A is the least relation closed under the
rules given in Figure 4.
We explain key rules below.
In rule (T-Out), the first premise Γ ⊲ P : A2 implies that the continuation of the output process behaves like A2 , and the second premise Γ ⊲ x : chanh(e
y :σ
e)A1 i implies that
the tuple of values ve being sent may be used by an input process according to he
v /e
y iA1 .
v /e
y iA1 | A2 ). Here,
Therefore, the whole behavior of the output process is described by x. (he
hv1 /x1 , . . . , vn /xn iA stands for hvi1 /xi1 , . . . , vik /xik iA where
{vi1 , . . . , vik } = {v1 , . . . , vn }\{true, false}. For example,htrue/x, y/ziA stands for hy/ziA.
Note that, as in previous behavioral type systems [10, 3], the resource access and communications made on ve by the receiver of ve are counted as the behavior of the output process
(see Remark 3.13).
In rule (T-In), the first premise implies that the continuation of the input process
behaves like A2 . Following previous behavioral type systems [10, 3], we split A2 into two
parts: A2 ↓{ey} and A2 ↑{ey } . The first part describes the behavior on the received values
ye and is taken into account in the channel type. The second part describes the resource
access and communications performed on other values, and is taken into account in the
behavioral type of the input process. The condition A2 ↓{ey } ≤ A1 requires that the access
and communication behavior on ye conforms to A1 , the channel arguments’ behavior.
In (T-New), the premise implies that P behaves like A, so that (νx) P behaves like
(νx) A. Here, we only require that x is a channel, unlike in the previous behavioral type
systems for the π-calculus [10, 15]. That is because we are only interested in the resource access behavior; the communication behavior is used only for accurately inferring the resource
access behavior.
In (T-NewR), we check that the process’s behavior A conforms to the resource usage
specification Φ.
Rule (T-Sub) allows the type A′ of a process to be replaced by its approximation A.
We remark that weakening of Γ can be derived (Appendix B, Lemma B.1) and so is
not needed as a rule.
The following example shows how information about the usage of resources by an input
process is propagated to an output process.
Example 3.10. Let us consider (NΦ x)P , where
Φ = (R∗ C)#
P = (νy) (yhx, xi | y(z1 , z2 ). read(z1 ).close(z2 )).
Let Γ = y : chanh(z1 , z2 )z1R .z2C i, x : res. Then, the following judgment holds for the
output and input processes.
Γ ⊲ yhx, xi : y. xR .xC
Γ ⊲ y(z1 , z2 ). read(z1 ).close(z2 ) : y. 0
Here, we have used subtyping relations hx/z1 , x/z2 iz1R .z2C ≈ xR .xC and z1R .z2C ↑{z1 ,z2 } ≈ 0.
By using (T-Par) and (T-New), we obtain
x : res ⊲ P : (νy) (y. xR .xC | y)
12
N. KOBAYASHI, K. SUENAGA, AND L. WISCHIK
Γ ⊲ 0:0
(T-Zero)
Γ ⊲ P : A2 Γ ⊲ x : chanh(e
y:σ
e )A1 i Γ ⊲ ve : σ
e
Γ ⊲ xhe
v i. P : x. (he
v /e
yiA1 | A2 )
(T-Out)
Γ, ye : σ
e ⊲ P : A2
Γ ⊲ x : chanh(e
y:σ
e)A1 i A2 ↓{ey} ≤ A1
Γ ⊲ x(e
y ). P : x. (A2 ↑{ey} )
(T-In)
Γ ⊲ P1 : A1
Γ ⊲ P2 : A2
Γ ⊲ P1 | P2 : A1 | A2
(T-Par)
Γ ⊲ P :A
Γ ⊲ ∗P : ∗A
(T-Rep)
Γ ⊲ v : bool
Γ ⊲ P :A
Γ ⊲ Q:A
Γ ⊲ if v then P else Q : A
(T-If)
Γ, x : chanh(e
y:σ
e )A1 i ⊲ P : A2
Γ ⊲ (νx) P : (νx) A2
(T-New)
Γ ⊲ P :A
Γ ⊲ x : res
Γ ⊲ accξ (x).P : xξ .A
(T-Acc)
Γ, x : res ⊲ P : A
tracesx (A) ⊆ Φ
Φ
Γ ⊲ (N x)P : A↑{x}
Γ ⊲ P : A′
A′ ≤ A
Γ ⊲ P :A
(T-NewR)
(T-Sub)
Figure 4: Typing Rules
Using (T-Sub) with (νy) (y. xR .xC | y) ≈ xR .xC we get
x : res ⊲ P : xR .xC
Since tracesx (xR .xC )) ⊆ (R∗ C)# , we obtain ∅ ⊲ (NΦ x)P : 0 by using (T-NewR) and
(T-Sub). 2
Example 3.11. Recall Example 2.9:
P = (νs) (∗s(n, x, r). P1 | (NΦ x)P2 )
P1 = if n = 0 then rhi
else (νr ′ ) (shn − 1, x, r ′ i | r ′ (). read(x).rhi)
P2 = (νr) (init(x).sh100, x, ri | r(). close(x))
Φ = (IR∗ C)#
Let A1 = µα.(r ⊕ (νr ′ ) (hr ′ /riα|r ′ . xR .r) and
let Γ = s:chanh(n:int, x:res, r:chanhi) A1 i. Then
Γ, n:int, x:res, r:chanhi ⊲ P1 : A1
Γ ⊲ ∗s(n, x, r). P1 : ∗s. (A1 ↑{n,x,r} ) ≈ ∗s
Γ ⊲ P2 : (νr) (xI .A1 |r. xC )
RESOURCE USAGE ANALYSIS FOR THE π-CALCULUS
13
So long as tracesx ((νr) (xI .A1 |r. xC )) ⊆ Φ, we obtain ∅ ⊲ P : 0. See Section 4.3 for the
algorithm that establishes tracesx (·) ⊆ Φ. 2
Remark 3.12. The type A1 in the example above demonstrates how recursion, hiding, and
renaming are used together. In general, in order to type a recursive process of the form
∗s(x). (νy) (· · · shyi · · · ), we need to find a type that satisfies (νy) (· · · hy/xiA · · · ) ≤ A.
Moreover, for the type inference (in Section 4), we must find the least such A. Thanks to
the type constructors for recursion, hiding, and renaming, we can always do that: A can be
expressed by µα.(νy) (· · · hy/xiα · · · ) (recall Lemma 3.8.4).
Remark 3.13. A reader may wonder why the rules (T-Out) and (T-In) are asymmetric,
in the sense that information about the continuation of a receiver process is transferred
to a sender process but not vice versa. That design choice comes from the observation
that a channel or resource exchanged between a sender and a receiver are, in general,
statically known only to the sender, so that we have to put information about the behavior
on the channel or resource into the type of the sender. For example, consider the process
((νy) (xhyi | · · · ) | x(z). z h i. Since the receiver x(z). z h i is not in the scope of y, we have
to put the information that y will be used for output into the type of the sender xhyi (as
x. y). It is still useful and possible to recover the symmetry in the treatment of senders and
receivers to some extent: see Section 8 of our previous paper [10].
The following theorem states that no well-typed process performs an invalid access to
a resource.
Theorem 3.14 (type soundness (safety)). Suppose that P is safe. If Γ⊲P : A and P −→∗ Q,
then Q is safe.
Proof. We make use of the following lemma:
L
L
• Subject-reduction. If P −→ P ′ and Γ ⊲ P : A then A =⇒ A′ and Γ ⊲ P ′ : A′ . Proof:
see Appendix B.
For the proof of the theorem, we focus on just a single reduction step. By the Lemma we
know that judgements are preserved by reduction; we must show that safety is also preserved, by induction on the derivation of reduction. The only interesting case is (R-NewR1),
−ξ
τ
(NΦ x)P → (NΦ x)P ′ , since the other rules do not alter trace-sets Φ. In this case, we are
xξ
xξ
given Γ ⊲ P : A, tracesx (A) ⊆ Φ, and P → P ′ . By the Lemma, A =⇒ A′ for some Γ ⊲ P ′ : A′ .
Assume (NΦ x)P is safe; hence so is P ; by the induction hypothesis so is P ′ . From the
xξ
conditions tracesx (A) ⊆ Φ and A =⇒ A′ , we get ξ ∈ tracesx (A) ⊆ Φ, so that ǫ ∈ Φ−ξ 6= ∅.
−ξ
So, (NΦ x)P ′ is safe.
4. Type Inference Algorithm
This section discusses an algorithm which takes a closed process P as an input and
checks whether ∅ ⊲ P : 0 holds. As in similar type systems [11, 15], the algorithm consists
of the following steps.
(1) Extract constraints on type variables based on the (syntax-directed version of) typing rules.
(2) Reduce constraints to trace inclusion constraints of the form
{tracesx1 (A1 ) ⊆ Φ1 , . . . , tracesxn (An ) ⊆ Φn }
14
N. KOBAYASHI, K. SUENAGA, AND L. WISCHIK
∅ ⊲sd 0 : 0
Γi ⊲ vi : σi (for each i ∈ {1, . . . , n})
Γ0 ⊲sd P : A2
e ∪ (x : chanh(e
Γ0 ∪ Γ
y:σ
e)A1 i) ⊲sd xhe
v i. P : x. (he
v /e
yiA1 | A2 )
Γ ⊲sd P : A2
A2 ↓{ey} ≤ A1
wd(Γ ∪ ye : σ
e)
(Γ\{e
y}) ∪ x : chanh(e
y :σ
e)A1 i ⊲sd x(e
y ). P : x. A2 ↑{ey}
Γ1 ⊲sd P1 : A1
Γ2 ⊲sd P2 : A2
Γ1 ∪ Γ2 ⊲sd P1 | P2 : A1 | A2
Γ ⊲sd P : A
Γ ⊲sd ∗P : ∗A
Γ0 ⊲ v : bool
Γ1 ⊲sd P : A1
Γ2 ⊲sd Q : A2
A1 ≤ A
A2 ≤ A
Γ0 ∪ Γ1 ∪ Γ2 ⊲sd if v then P else Q : A
Γ ⊲sd P : A2
wd(Γ ∪ (x : chanh(e
x : τe)A1 i))
Γ\{x} ⊲sd (νx) P : (νx) A2
Γ ⊲sd P : A
Γ ∪ (x : res) ⊲sd accξ (x).P : xξ .A
Γ ⊲sd P : A
tracesx (A) ⊆ Φ
wd(Γ ∪ (x : res))
Γ\{x} ⊲sd (NΦ x)P : A′ ↑{x}
(T-SD-Zero)
(T-SD-Out)
(T-SD-In)
(T-SD-Par)
(T-SD-Rep)
(T-SD-If)
(T-SD-New)
(T-SD-Acc)
(T-SD-NewR)
Figure 5: Syntax Directed Typing Rules
(3) Decide whether the constraints are satisfied.
The algorithm for Step 3 is sound but not complete.
We give an overview of each step below. The first two steps are almost the same as
those in the previous work.
4.1. Step 1: Extracting Constraints. The typing rules presented in Section 3 can be
transformed to the syntax-directed typing rules shown in Figure 5. In the figure, Γ1 ∪ Γ2 is
the type environment obtained by merging both bindings, and defined only if Γ1 (x) = Γ2 (x)
for every x ∈ dom(Γ1 )∩dom(Γ2 ). Type equality here is syntactic equality up to α-renaming.
And wd(Γ1 ∪ Γ2 ) means that Γ1 ∪ Γ2 is well-defined. The two sets of typing rules are
equivalent in the following sense: If Γ ⊲ P : A is derivable, then there exists A′ such that
A′ ≤ A holds and Γ ⊲sd P : A′ is derivable. Conversely, if Γ ⊲sd P : A is derivable, so is
Γ ⊲ P : A.
Based on the syntax-directed rules, we obtain the algorithm in Figure 6, which takes
a process P and outputs a triple consisting of a type environment Γ, a behavioral type A,
and a set C of constraints. In Figure 6, Γ1 ⊗ · · · ⊗ Γn is defined to be (Γ, C) where Γ and
C are given by:
dom(Γ) = dom(Γ1 ) ∪ · · · ∪ dom(Γn )
Γ(x) = Γi (x) where x ∈ dom(Γi )\(dom(Γ1 ) ∪ · · · ∪ dom(Γi−1 ))
C = {Γi (x) = Γj (x) | x ∈ dom(Γi ) ∩ dom(Γj )}
The triple (Γ, A, C) output by PT satisfies the following properties:
• θΓ ⊲ P : θA holds for any substitution θ such that |= θC.
RESOURCE USAGE ANALYSIS FOR THE π-CALCULUS
15
• If Γ′ ⊲ P : A′ , then there exists a substitution θ such that θΓ ⊆ Γ′ and θA ≤ A′ .
Here, Γ and A may contain variables representing unknown behavioral types and value
types. C is a set of constraints on them, and the substitution θ above replaces them with
closed behavioral types and value types. Intuitively, the triple (Γ, A, C expresses a set of
type judgments for P . The first property above says that the triple contains only valid
judgments, while the second property says that every valid judgment is subsumed by the
triple.
We do not give a formal proof of the above properties; As usual, they can be proved by
induction on the structure of P .
4.2. Step 2: Reducing Constraints. Given a closed process P , PT(P ) produces a triple
(∅, A, C). The set C of constraints consists of unification constraints on value types (where
all the behavioral types occurring in them are variables), constraints of the form isChan(σ)
(which means that σ is a channel type), subtype constraints on behavioral types of the form
α ≥ A, and constraints of the form tracesx (A) ⊆ Φ. We can remove the first two kinds of
constraints (unification constraints on value types and isChan(σ)) by applying the standard
unification algorithm. Thus, we obtain the following constraints:
{α1 ≥ A1 , . . . , αn ≥ An ,
tracesx1 (B1 ) ⊆ Φ1 , . . . , tracesxm (Bm ) ⊆ Φm }
Here, we can assume that α1 , . . . , αn are different from each other, since α ≥ A1 and
α ≥ A2 can be replaced with α ≥ A1 ⊕ A2 by Lemma 3.8. We can also assume that
{α1 , . . . , αn } contains all the type variables in the constraint, since otherwise we can always
add the tautology α ≥ α. Each subtype constraint α ≥ A can also be replaced by
α ≥ µα.A, by using Lemma 3.8. Therefore, the above constraints can be further reduced,
by Lemma 3.8, to:
e′ /e
e′ /e
{tracesx1 ([A
α]B1 ) ⊆ Φ1 , . . . , tracesxm ([A
α]Bm ) ⊆ Φm }
Here, A′1 , . . . , A′n are the least solutions for the subtype constraints.
Thus, we have reduced type checking to the validity of trace inclusion constraints of
the form tracesx (A) ⊆ Φ.
Example 4.1. Recall Example 2.9. By applying the algorithm PT and the first part of
Step 2, we obtain the following constraints:
tracesx ((νr) (xI .s. α1 | r. xC )) ⊆ (IR∗ C)#
α1 ≥ r. α2 ⊕ (νr ′ ) (s. hr ′ /riα1 | r ′ . xR .r. α2 )↓{n,x,r}
α2 ≥ α2
By applying the second part of Step 2, we obtain tracesx (A1 ) ⊆ (IR∗ C)# where
A1 = (νr) (xI .s. A2 | r. xC )
A2 = µα1 .r. A3 ⊕ (νr ′ ) (s. hr ′ /riα1 | r ′ . xR .r. A3 )↓{n,x,r}
A3 = µα2 .α2 .
16
N. KOBAYASHI, K. SUENAGA, AND L. WISCHIK
PTv(x) = (x : ρ, ρ) (where ρ fresh)
PTv(b) = (∅, bool) if b ∈ {true, false}
PT(0) =(∅, 0, ∅)
v i. P0 ) =
PT(xhe
let (Γi , σi ) = PTv(vi )
(Γ0 , A0 , C0 ) = PT(P0 )
(Γ, C) = Γ0 ⊗ (x : chanh(e
y :σ
e)αi) ⊗ Γ1 ⊗ · · · ⊗ Γn
v /e
y]α | A0 ), C) (where α fresh)
in (Γ, x. ([e
PT(x(e
y ). P0 ) =
let (Γ0 , A0 , C0 ) = PT(P0 )
(Γ1 , C1 ) = Γ0 ⊗ (x : chanh(e
y : ρe)αi) ⊗ (e
y : ρe)
in (Γ\e
y, x. A0 ↑{ey} , C0 ∪ C1 ∪ {α ≥ A0 ↓{ey} })
(where α, ρe fresh)
PT(P0 | P1 ) =
let (Γ0 , A0 , C0 ) = PT(P0 )
(Γ1 , A1 , C1 ) = PT(P1 )
(Γ2 , C2 ) = Γ0 ⊗ Γ1
in (Γ2 , A0 | A1 , C0 ∪ C1 ∪ C2 )
P T (if v then P0 else P1 ) =
let (Γ0 , A0 , C0 ) = P T (P0 )
(Γ1 , A1 , C1 ) = PT(P1 )
(Γ2 , σ) = PTv(v)
(Γ, C2 ) = Γ0 ⊗ Γ1 ⊗ Γ2
in (Γ, A0 ⊕ A1 , C0 ∪ C1 ∪ C2 ∪ {σ = bool})
P T ((νx) P0 ) =
let (Γ0 , A0 , C0 ) = P T (P0 )
C1 = if x ∈ dom(Γ0 )then {isChan(Γ0 (x))}else ∅
in (Γ0 \{x}, (νx) A0 , C0 ∪ C1 )
P T (∗P0 ) =
let (Γ0 , A0 , C0 ) = P T (P0 )
in (Γ0 , ∗A0 , C0 )
P T (accξ (x).P0 ) =
let (Γ0 , A0 , C0 ) = P T (P0 )
(Γ1 , C1 ) = Γ0 ⊗ (x : res)
in (Γ1 , xξ .A0 , C0 ∪ C1 )
P T ((NΦ x)P0 ) =
let (Γ0 , A0 , C0 ) = P T (P0 )
(Γ1 , C1 ) = Γ0 ⊗ (x : res)
in (Γ1 \{x}, A0 ↑{x} , C0 ∪ C1 ∪ {tracesx (A0 ) ⊆ Φ})
Figure 6: A Type Inference Algorithm
4.3. Step 3: Constraint Solving. We present an approximation algorithm for checking
a trace inclusion constraint tracesx (A) ⊆ Φ when the trace set Φ is a regular language.
(Actually, we can extend the algorithm to deal with the case where Φ is a deterministic
Petri net language: see Remark 4.6.)
We first describe the algorithm with an example. In Example 4.1 above, we have
reduced the typability of the process to the equivalent constraint tracesx (A1 ) ⊆ Φ where
RESOURCE USAGE ANALYSIS FOR THE π-CALCULUS
17
Φ = (IR∗ C)# and
A1 ↓{x} ≈ (νr) (xI .A′′2 | r. xC )
A′′2 = r ⊕ (νr ′ ) (hr ′ /riA′′2 | r ′ . xR .r)
Here, we have removed A3 = µα.α since A3 ≈ 0.
Step 3-1. Approximate the behavior of A1 ↓{x} by a Petri net [22] NA1 ,x . This part
is similar to the translation of usage expressions into Petri nets in Kobayashi’s previous
work [16, 15, 12]. Since the behavioral types are more expressive (having recursion, hiding,
and renaming), however, we need to approximate the behavior of a behavioral type unlike
in the previous work. In this case A1 ↓{x} is infinite. To make it tractable we make a sound
approximation A′1 by pushing (ν) to top level, and we eliminate hr ′ /ri:
A′1 = (νr, r ′ ) (xI .A′2 | r. xC )
A′2 = r ⊕ (A′3 | r ′ . xR .r)
A′3 = r ′ ⊕ (A′3 | r ′ . xR .r ′ )
Then NA′1 ,x is as pictured. (Here we treat A1 ⊕ A2 as τ.A1 ⊕ τ.A2 for clarity. We also use
a version of Petri nets with labeled transitions.)
B10
B1
xI :A02
r:xC
xI
B2
¿
B5
¿
¿:r © ¿:(A03jr0:xR:r)
¿
xC
r
B8
¿:r 0 © ¿:(A03jr0 :xR:r0)
B9
¿
B4
¿
¿
r0
B6
r0 :xR :r 0
xC
xR
r0 :xR :r
B3
B11
xR:r
xR
¿
B7
xR:r 0
The rectangles are the places of the net, and the dots labeled by τ, xR , etc. are the
transitions of the net. Write ix for the number of tokens at node Bx . The behavior
e together
A′1 corresponds to the initial marking {i1 =1, i10 =1}. We say that the nodes B
′
′
with the restricted names (r, r ) constitute a basis for A1 . Note here that tracesx (A1 ) ⊆
tracesx (A′1 ) = ptraces(NA′1 ,x ) where ptraces(NA′1 ,x ) is the set of traces of the Petri net.
Thus, ptraces(NA′1 ,x ) ⊆ Φ is a sufficient condition for tracesx (A1 ) ⊆ Φ . The key point
here is that A′1 still has infinite states, but all its reachable states can be expressed in the
form (νr, r ′ ) (i1 B1 | · · · | i11 B11 ) (where ik Bk is the parallel composition of ik copies of Bk ),
e That is why we could express A′ by the
a linear combination of finitely many processes B.
1
Petri net as above.
18
N. KOBAYASHI, K. SUENAGA, AND L. WISCHIK
Step 3-2. Construct a deterministic, minimized automaton MΦ that accepts the language Φ. Here the initial marking is {i12 =1}.
B11
I.R∗.C
xR
B12
xI
R∗ .C
xC
Step 3-3. Construct another Petri net NA′1 ,x k MΦ from NA′1 ,x and MΦ , which simulates the behavior of PA and MΦ simultaneously, so that the problem of tracesx (A′1 )(=
ptraces(NA′1 ,x )) ⊆ Φ is equivalent to a reachability problem of NA′1 ,x k MΦ . In the example, NA′1 ,x k MΦ has the initial marking {i1 =1, i10 =1, i12 =1} and transitions such as
I
B1 |B12 −→ B2 |B13 . ptraces(NA′1 ,x ) ⊆ Φ if and only if the following unsafe state is unreachable.
(i1 >0 ∧ i12 =0) ∨ (i7 >0 ∧ i13 =0) ∨ (i9 >0 ∧ i13 =0) ∨ (i11 >0 ∧ i13 =0)
To explain, if i1 > 0 ∧ i12 =0 then the behavior is able to make an R transition but the
specification automaton MΦ is not able.
Step 3-4. Use an approximation algorithm to decide the reachability problem of
′
NA1 ,x k MΦ , in a manner similar to Kobayashi’s type-based analyzer TyPiCal [12] for the
π-calculus.
The above steps 3-1, 3-2, and 3-3 are described in more detail below. See Section 6 for
Step 3-4.
4.3.1. Step 3-1: Construction of NA,x . We first introduce the notion of a basis. The basis
is analogous to that of a vector space; Each state is expressed as a linear combination of
elements of the basis.
Definition 4.2. A pair ({y1 , . . . , ym }, {B1 , . . . , Bn }) is a basis of A if all of the following
conditions are satisfied:
• A ≈ (νy1 ) · · · (νym ) (i1 B1 | · · · | in Bn ) for some i1 , . . . , in ∈ Nat.
l
• If Bj −→ C, then there exist i1 , . . . , in ∈ Nat such that C ≈ i1 B1 | · · · | in Bn .
l
• For each Bj , there are only finitely many C (up to ≈) such that Bj −→ C.
Note that if ({e
y }, {B1 , . . . , Bn }) is a basis of A, then whenever A =⇒ A′ , there exist
i1 , . . . , in such that A′ ≈ (ν ye) (i1 B1 | · · · | in Bn ). Let us write Index(C) for (i1 , . . . , in ) such
that C ≈ i1 B1 | · · · in Bn . (If there are more than one such tuple, Index(C) picks one among
them.) Therefore, if A↓{x} has a basis, the behavior of A↓{x} is simulated by the (labeled)
Petri net NA,x,({ey},{B})
given below. Here, we use a process-like syntax to represent the
e
elements of a Petri net rather than the standard tuple notation (P, T, F, W, M0 ). A marking
state m which has ik tokens for each place pk (k ∈ {1, . . . , n}) is written i1 p1 | · · · | in pn . A
γ
transition that consumes a marking m1 and produces m2 is expressed by m1 −→ m2 , where
γ is the label of the transition.
• The set P of places is {pB1 , . . . , pBn }.
RESOURCE USAGE ANALYSIS FOR THE π-CALCULUS
19
• The initial marking mI is i1 pB1 | · · · | in pBn
where A↓{x} ≈ (ν ye) (i1 B1 | · · · | in Bn ).
• The set of transitions consists of:
τ
– pBj −→ i1 pB1 | · · · | in pBn
τ
where Index(C) = (i1 , . . . , in ), for each Bj −→ C.
ξ
– pBj −→ i1 pB1 | · · · | in pBn
xξ
where Index(C) = (i1 , . . . , in ), for each Bj −→ C.
τ
– pBj | pBj′ −→ (i1 + i′1 )pB1 | · · · | (in + i′n )pBn where Index(C) = (i1 , . . . , in ) and
z
z
Index(C ′ ) = (i′1 , . . . , i′n ), for each pair of transitions Bj −→ C and Bj ′ −→ C ′
such that z ∈ {e
y }.
From now on we omit the basis and just write NA,x for NA,x,({ey},{B})
e . Let us write
ptraces(NA,x ) for the set:
ξ1
ξ
k
{ξ1 · · · ξk | mI =⇒ · · · =⇒
m′ }
ξ
τ ∗ ξ
τ ∗
where =⇒ means −→ −→−→ . By the construction of NA,x , ptraces(NA,x ) = tracesx (A).
The construction of NA,x outlined above can be applied only when a basis of A↓x
can be found (by some heuristic algorithm). If A↓x has no basis or cannot be found,
we approximate A↓x by moving all the ν-prefixes to the top-level; for example, y. (νx) A,
∗(νx) A and µα.(νx) A are replaced by (νx) (y. A), (νx) ∗A, and (νx) µα.A respectively. Let
A′ be the approximation of A↓{x} . It is easy to prove that A′ is a sound approximation of
A↓{x} , in the sense that tracesx (A) ⊆ tracesx (A′ ).
We can compute a basis of A′ as follows (see Appendix D for more details). Since
ν-prefixes do not appear inside recursion, we can first eliminate the constructors ·↑S , ·↓S ,
and he
y /e
xi. Let (ν ye) A′′ be the resulting expression, where A′′ does not contain ·↑S , he
y /e
xi, or
(νx) . Let B be the set of behavioral types that are subexpressions of the behavioral types
obtained from A′′ by expanding recursive types and do not contain “unnecessary” unfolding
[µα.A/α]A. Then, B is a finite set, and ({e
y }, B) is a basis of A′ . We can therefore construct
a Petri net NA′ ,x . By the construction, ptraces(NA′ ,x ) = tracesx (A′ ) ⊇ tracesx (A), so
that ptraces(NA′ ,x ) ⊆ Φ is a sufficient condition for tracesx (A) ⊆ Φ.
4.3.2. Steps 3-2 and 3-3: Construction of NA,x k MΦ and reduction of tracesx (A) to a
reachability problem. Let PNA,x and TNA,x be the sets of places and transitions of NA,x
respectively. Let MΦ be a minimized deterministic automaton4 that accepts Φ, and let QΦ
be its set of states and δΦ be its transition function.
Definition 4.3. The composition of NA,x and MΦ , written NA,x k MΦ , is defined as follows:
• The set of places is PNA,x ∪ QΦ
• The set of transitions is:
ξ
ξ
{(m|q) −→ (m′ |q ′ ) | (m→m′ ) ∈ TNA,x ∧ δΦ (q, ξ) = q ′ }
τ
τ
∪{m −→ m′ | (m→m′ ) ∈ TNA,x }
• Initial state is mI | qI where mI is the initial state of NA,x and qI is the initial state
of MΦ .
4Note that since Φ is prefix-closed, all the states of the minimized automaton are accepting states.
20
N. KOBAYASHI, K. SUENAGA, AND L. WISCHIK
Now, ptraces(NA,x ) ⊆ Φ can be reduced to the reachability problems of NA,x k MΦ .
Theorem 4.4. ptraces(NA,x ) ⊆ Φ if and only if no marking m | q that satisfies the following conditions is reachable:
ξ
• m −→ m′ for some m′ and ξ in NA,x .
• δΦ (q, ξ) is undefined.
Thus, we can reduce ptraces(NA,x ) ⊆ Φ to a finite set of reachability problems of
NA,x k MΦ . Hence ptraces(NA,x ) ⊆ Φ is decidable [18].
Corollary 4.5. ptraces(NA,x ) ⊆ Φ if and only if for every transition rule of the form
ξ
m1 −→ m2 of NA,x and q such that δΦ (q, ξ) is undefined, no marking m such that m ≥ m1 | q
is reachable by NA,x k MΦ .
Remark 4.6. We can actually extend the above algorithm for checking tracesx (A) ⊆ Φ to
deal with the case where Φ belongs to the class of deterministic Petri net languages (more
precisely, the class of P-type languages of λ-free, deterministic Petri nets [22, 21]). If Φ is
the P-type language of a λ-free, deterministic Petri net, then its complement Φ is a Petri net
language [21]. Therefore, we can construct a Petri net that accepts the intersection of the
language of NA,x and Φ [22]), so that ptraces(NA,x ) ⊆ Φ can be reduced to the emptiness
problem of the Petri net, which is decidable due to the decidability of the reachability
problem.
Some of the useful resource usage specifications are not regular languages but are deterministic Petri net language. For example, consider a stack-like resource on which, at any
point of program execution, the number of times the operation pop has been performed is
less than the number of times push has been performed. Such specification is expressible
as a deterministic Petri net language.
5. Extensions
The type system given so far guarantees that no invalid resource access is performed,
but not that any necessary access is performed eventually; for example, the type system
does not guarantee that a file is eventually closed. We discuss extensions of the type system
to guarantee such properties.
We are interested in type systems that satisfy either partial liveness 5 or the stronger
liveness property:
• partial liveness: If P −→∗ Q and Q 6−→, then Q does not contain any resource to
which some access must be performed.
• liveness: In any fair reduction sequence P −→ P1 −→ P2 −→ · · ·, P eventually
performs all the necessary resource access. (Here, a reduction sequence is fair if an
input or output action that is infinitely enabled will eventually succeed. Without
the fairness assumption, no process can satisfy the liveness property in the presence
of a divergent process (νx) (xh i | ∗x( ). xh i, which is too restrictive.)
5This is not a standard term; actually, the partial liveness here can be viewed as the safety property that
no ‘bad’ state is reachable such that the necessary accesses have not yet been performed but the system
cannot make any move.
RESOURCE USAGE ANALYSIS FOR THE π-CALCULUS
21
Our idea is to take the resource type system from the previous sections, and combine it
with some existing system that annotates those communications that eventually succeed.
Specifically, this existing system might be (1) deadlock-freedom [16, 15], which guarantees
that the annotated communications eventually succeed unless the process diverges; the
combination would then guarantee partial liveness. Or the existing system could be (2) lockfreedom [14, 15], which guarantees that the annotated communications eventually succeed
even in the presence of divergence (assuming a strongly fair scheduler); the combination
would then guarantee full liveness.
To formally state which resource access must be performed, we extend the trace sets.
Definition 5.1. An extended trace set is a set of sequences of access labels, possibly ending
with a special label ↓, that is closed under the prefix operation.
Intuitively, the special label ↓ means that no further resource access need to be performed. For example, the trace set ({C ↓, RC ↓})# means that the close operation needs
to be performed, while ({↓, R ↓, C ↓, RC ↓})# means that the close operation need not be
performed.
e for a
Now we can state the partial liveness property more formally. We write (e
ν N)
(possibly empty) sequence of ν- and N-binders.
Φ x)Q 6−→.
e
Definition 5.2. A process P is partially live if ↓ ∈ Φ whenever P −→∗ (e
ν N)(N
5.1. A Type System for the Partial Liveness Property. We extend the syntax of
processes to allow each input and output prefix to be annotated with information about
whether the communication is guaranteed to succeed.
Definition 5.3 ((extended) processes). The set of (extended) processes is given by:
t (attributes)
P
::= c | ∅
::= xt hy1 , . . . , yn i. P | xt (y1 , . . . , yn ). P | · · ·
The attribute c indicates that when the annotated input or output operation appears
at the top-level, the operation will succeed unless the whole process diverges, while ∅ does
not give such a guarantee. We often omit tag ∅.
We assume that there exists a type system guaranteeing that any well-typed process
is well-annotated in the sense of Definition 5.4 below. There are indeed such type systems [13, 16, 15]. Moreover, the static analysis tool TyPiCal [12] can automatically infer
the annotations.
Definition 5.4. P is active, written active(P ), if
e c (e
e c he
v i. Q | R) or P (e
ν N)(x
y ). Q | R). Additionally, P is well-annotated, writP (e
ν N)(x
′
ten well annotated (P ), if for any P such that P −→∗ P ′ and active(P ′ ), there exists P ′′
such that P ′ −→ P ′′ .
For example, xc h i. 0 | xc ( ). y ∅ h i. 0 is well-annotated, but
xc h i. 0 | xc ( ). y c h i. 0 is not. Note that x∅ ( ). xc h i. 0 is well-annotated since, although the
output never succeeds, it does not appear at the top-level.
Now we introduce the type system that guarantees the partial liveness. We extend the
behavioral types by extending each input, output, or τ -action with an attribute to indicate
whether the action is guaranteed to succeed.
A ::= xt . A | xt . A | τt .A | · · ·
22
N. KOBAYASHI, K. SUENAGA, AND L. WISCHIK
disabled(0, S)
disabled(xξ .A, S)
if disabled(A, S) and x 6∈ S
disabled(ac .A, S)
disabled(a∅ .A, S)
if disabled(A, S)
disabled(τc .A, S)
disabled(τ∅ .A, S)
if disabled(A, S)
disabled(A1 | A2 , S) if disabled(A1 , S) and disabled(A2 , S)
disabled(A1 ⊕ A2 , S) if disabled(A1 , S) or disabled(A2 , S)
disabled(∗A, S)
if disabled(A, S)
disabled((νx) A, S) if disabled(A, S\{x})
disabled(A↑S ′ , S) if disabled(A, S\S ′ )
disabled(A↓S ′ , S) if disabled(A, S ∩ S ′ )
disabled(he
y /e
xiA, S) if disabled(A, {z | [e
y /e
x]z ∈ S})
disabled(µα.A, S)
if disabled([µα.A/α]A, S)
Figure 7: The definition of disabled(A, S)
For example, a process having type xc . x∅ . 0 implies that the process may send values on x
twice, and that the first send is guaranteed to succeed (i.e., the sent value will be received
by some process), while there is no such guarantee for the second send.
The transition semantics of behavioral types is unchanged; The attribute t is just
ignored.
We revise the definitions of the subtype relation and the traces by using the following
predicate disabled(A, S). Intuitively, this means that A describes a process that may get
blocked without accessing any resources in S.
Definition 5.5. disabled(A, S) is the least binary relation between extended behavioral
types and sets of variables closed under the rules in Figure 7.
Definition 5.6. The set etracesx (A) of extended traces is:
xξ1
xξn
{ξ1 · · · ξn ↓ |∃B.A↓{x} =⇒ · · · =⇒ B ∧ disabled(B, {x})}
xξ1
xξn
∪{ξ1 · · · ξn |∃B.A↓{x} =⇒ · · · =⇒ B}
xξ1
xξn
Here, A↓{x} =⇒ · · · =⇒ B ∧ disabled(B, {x}) means that ξn may be the last access to x,
so that ↓ is attached to the sequence ξ1 · · · ξn . By definition, etracesx (A) is prefix-closed.
Definition 5.7. A1 ≤ A2 is the largest relation on closed behavioral types that satisfies
the following properties:
l
l
• If A1 −→ A′1 then there exists A′2 such that A2 =⇒ A′2 and A′1 ≤ A′2 .
• disabled(A1 , S) implies disabled(A2 , S) for any set S of variables.
Note that by the definition, A1 ≤ A2 implies etracesx (A1 ) ⊆ etracesx (A2 ).
The typing rules are the same as those in Section 3, except for the rules shown in
Figure 8. The only changes are that attributes have been attached to (ET-Out) and
(ET-In), and that tracesx (A↓{x} ) has been replaced by etracesx (A↓{x} ) in (ET-NewR).
RESOURCE USAGE ANALYSIS FOR THE π-CALCULUS
23
Γ ⊲pl x : chanh(e
y :σ
e)A1 i
Γ ⊲pl ve : σ
e
Γ ⊲pl xt he
v i. P : xt . (he
v /e
y iA1 | A2 )
Γ ⊲pl P : A2
(ET-Out)
Γ, ye : σ
e ⊲pl P : A2
Γ ⊲ x : chanh(e
y:σ
e)A1 i
A2 ↓{ey} ≤ A1
(ET-In)
Γ ⊲pl xt (e
y ). P : xt . (A2 ↑{ey} )
Γ, x : res ⊲pl P : A
etracesx (A) ⊆ Φ
Φ
Γ ⊲pl (N x)P : A↑{x}
(ET-NewR)
Figure 8: Typing Rules for Partial Liveness
An important invariant maintained by the typing rules is that the type of an input/output
process is annotated with c only if the process itself is annotated with c. For example, we
cannot derive x : chanhi ⊲pl x∅ h i : xc .
The following theorem states the soundness of the extended type system.
Theorem 5.8. If well annotated (P ) and ∅ ⊲pl P : A, then P is partially live.
Proof. We make use of three lemmas. The first two show that typing and well-annotatedness
are preserved by reduction. The third means that the type of a process properly captures
the possibility of the process being blocked.
L
• Subject reduction. If Γ ⊲pl P : A and P −→ Q, then there exists some B such
L
that Γ ⊲pl Q : B and A =⇒ B. Proof: See Appendix B.
• Well-annotatedness. If well annotated (P ) and P −→∗ Q,
then well annotated (Q). Proof: trivial by definition of
well annotated (P ).
• Disabled. If well annotated (P ) and Γ ⊲pl P : A with bool 6∈ codom(Γ), then P −→
6
implies disabled(A, S) for any S. Proof: See Appendix C.
Φ x)Q −→
e
Now we are ready to prove the theorem. Suppose that P −→∗ (e
ν N)(N
6
and
well annotated (P ), ∅ ⊲pl P : A. We have to show ↓ ∈ Φ. By subject-reduction we obΦ x)Q : A′ for some A′ . By the inversion of the typing rules, we get
e
tain ∅ ⊲pl (e
ν N)(N
ye : rg
es, ze : σ
e, x : res ⊲pl Q : B and tracesx (B) ⊆ Φ for some sequence σ
e of channel types.
(Here, ye and ze are the variables bound by N.) By well-annotatedness we also have
Φ x)Q), which implies well annotated (Q). Thus, by Disabled, we get
e
ν N)(N
well annotated ((e
disabled(B, S) for any S, which implies disabled(B↓{x} , {x}). So, we have ↓ ∈ etracesx (B) ⊆
Φ as required.
Example 5.9. An annotated version of Example 3.11:
P = (νs) (∗sc (n, x, r). P1 | (NΦ x)P2 )
P1 = if n = 0 then r c hi
else (νr ′ ) (sc hn − 1, x, r ′ i. | r ′ c (). read(x).r c hi)
P2 = (νr) (init(x).sc h100, x, ri | rc (). close(x))
Φ = (IR∗ C ↓)#
24
N. KOBAYASHI, K. SUENAGA, AND L. WISCHIK
is well-annotated. Suppose
A1 = µα.(r c ⊕ (νr ′ ) (hr ′ /riα|r ′ c . xR .r c )
Γ = s:chanh(b:int, x:res, r:chanhi) A1 i.
Then
Γ ⊲ P1 : A1
Γ ⊲ ∗sc (n, x, r). P1 : ∗sc . (A1 ↑{n,x,r} ) ≈ ∗sc
Γ ⊲ P2 : (νr) (xI .A1 |rc . xC ).
So long as etracesx ((νr) (xI .A1 | rc . xC .)) ⊆ Φ, we obtain ∅ ⊲ P : 0. 2
5.2. Type Inference. The type inference algorithm for the extended type system is almost
the same as the algorithm for the basic type system discussed in Section 4. The only changes
are:
• In the constraint generaltion algorithm PT, attribute annotations for input and
ouptut processes are propagated to types. For example, the case for output processes
becomes:
v i. P0 ) =
PT(xt he
let (Γi , σi ) = PTv(vi )
(Γ0 , A0 , C0 ) = PT(P0 )
(Γ, C) = Γ0 ⊗ (x : chanh(e
y :σ
e)αi) ⊗ Γ1 ⊗ · · · ⊗ Γn
v /e
y ]α | A0 ), C) (where α fresh)
in (Γ, xt . ([e
• The constraint tracesx (A) ⊆ Φ is replaced by etracesx (A) ⊆ Φ.
The second change forces us to adjust the reduction of the constraint to the reachability problem of Petri nets (recall step 3 of the algorithm in Section 4). First, we
need to use eptraces(NA,x ) defined below, which corresponds to etracesx (A), instead
of ptraces(NA,x ) in the reduction.
Definition 5.10. eptraces(NA,x ) is the set
ξ1
ξ
ξ1
ξ
k
k
{ξ1 · · · ξk | mI =⇒ · · · =⇒
m′ } ∪ {ξ1 · · · ξk ↓| mI =⇒ · · · =⇒
m′ ∧ pdisabled(m′ , {x})}
where mI is the initial marking of NA,x . pdisabled(m, S) means that disabled(A, S) holds
for the behavioral type A expressed by m.
Second, the construction of an automaton needs to be adjusted so that it accepts
extended traces. For example, the automaton used in the explanation of Step 3-2 in Section 4
is replaced by the one that accepts IR∗ C ↓.
With these changes, the validity of a constraint etracesx (A) ⊆ Φ is reduced to the
reachability problem of a Petri net NA,x k MΦ where composition of a Petri net NA,x and
an automaton MΦ is defined in the same manner as Definition 4.3.
Theorem 5.11. eptraces(NA,x ) ⊆ Φ if and only if no marking m | q that satisfies the
following conditions is reachable:
• pdisabled(m, {x}).
• δΦ (q, ↓) is undefined.
RESOURCE USAGE ANALYSIS FOR THE π-CALCULUS
25
6. Implementation
We have implemented a prototype resource usage analyzer based on the extended type
system described in Section 5. We have tested all the examples given in the present paper.
The implementation can be tested at http://www.yl.is.s.u-tokyo.ac.jp/~kohei/usage-pi/.
The analyzer takes a pi-calculus program as an input, and uses TyPiCal[12] to annotate
each input or output action with an attribute on whether the action is guaranteed to succeed
automatically (recall the syntax of extended processes in Section 5). The annotated program
is then analyzed based on the algorithm described in Section 4.
The followings are some design decisions we made in the current implementation. We
restrict the resource usage specification (Φ) to the regular languages, although in future we
may extend it based on Remark 4.6. In Step 3-1 of the algorithm for checking etracesx (A) ⊆
Φ, we blindly approximate A by pushing all of its ν-prefixes to the top-level. In future we
might utilize an existing model checker to handle the case where A is already finite. In Step
3-4 for solving the reachability problems of Petri nets, we approximate the number of tokens
in each place by an element of the finite set {0, 1, 2, “3 or more”}. That approximation
reduces Petri nets to finite state machines, so we can use BDD to compute an approximation
of the reachable states.
Figure 9 shows a part of a successful run of the analyzer. The first process (on the second
line) of the input program runs a server, which returns a new, initialized resource. We write
! and ? for output and input actions. The resource access specification is here expressed
by the number 1 of newR 1, x, which refers to the built-in specification (I(R + W )∗ C ↓)# .
The second process runs infinitely many client processes, each of which sends a request for
a new resource, and after receiving it, reads and closes it. The third process (on the 6th
line) is a tail-recursive version of the replicated service in Example 2.9. Here, a boolean is
passed as the first argument of s instead of an integer, as the current system is not adapted
to handle integers; it does not affect the analysis, since the system ignores the value and
simply inspects both branches of the conditional. Note that the program creates infinitely
many resources and has infinitely many states. The first output is the annotated version
of the input program produced by TyPiCal, where !! and ?? are an output and an input
with the attribute c (recall Section 5).
The remaining part shows the trace inclusion constraint and the constructed Petri net.
The final line reports that the verification has succeeded, which implies that both the safety
property (in Section 3) and the partial liveness property (in Section 5) are satisfied.
7. Related Work
Resource usage analysis and similar analyses have recently been studied extensively,
and a variety of methods from type systems to model checking have been proposed [5, 6,
7, 11, 1, 17, 24]. However, only a few of them deal with concurrent languages. To our
knowledge, none of them deal with the partial liveness property (or the total liveness property) that we discussed in Section 5. Nguyen and Rathke [20] propose an effect-type system
for a kind of resource usage analysis for functional languages extended with threads and
monitors. In their language, neither resources nor monitors can be created dynamically. On
the other hand, our target language is π-calculus, so that our type system can be applied
to programs that may create infinitely many resources (due to the existence of primitives
for dynamic creation of resources: recall the example in Figure 9), and also to programs
26
N. KOBAYASHI, K. SUENAGA, AND L. WISCHIK
Input:
new create,s in
*(create?(r).newR 1,x in acc(x,init).r!(x))
| *(new r in create!(r)
| r?(y).new c in s!(false,y,c) | s!(false,y,c)
| c?().c?().acc(y,close))
| *(s?(b,x,r).if b then r!()
else acc(x,read).s!(b,x,r))
Output:
(*** The result of lock-freedom analysis ***)
new create, s in
*create??(r). newR 1,x in acc(x, I). r!!(x)
| *(new r in create!!(r)
| r??(y).new c in s!!(false,y,c) | s!!(false,y,c)
| c??().c??().acc(y,close))
...
(*** Constraints ***)
etrace(x,acc(x, init).(c!! & acc(x, read). $16 | $16 |
c??. c??. acc(x, close). O)) is included in 1
...
(*** initial marking ***)
1 * 11 | 1 * 7
(*** 14 Places ***)
0: c!!. O
...
(*** 9 Transitions ***)
(x,close): 1*12 | 1*10 -> -1*12 | 1*13 | -1*10 | 1*1
...
No error found
Figure 9: A Sample Run of the Analyzer.
that use a wide range of communication and synchronization primitives. Capability-based
type systems can deal with concurrency to a certain degree ([5], Section 4.2), by associating
each resource with a unique capability to access the resource. The type system can control
the resource access order, by ensuring the uniqueness of the capability and keeping track
of what access is currently allowed by each capability. In this approach, however, resource
accesses are completely serialized and programmers have to care about appropriately passing capabilities between threads. Capability-based type systems [5, 6] also require rather
complex type annotations. Igarashi and Kobayashi’s type system for resource usage analysis for λ-calculus [11] can be extended to deal with threads, by introducing the following
typing rule:
Γ1 ⊲ M1 : τ1
Γ2 ⊲ M2 : τ2
Γ1 ⊗ Γ2 ⊲ spawn(M1 ); M2
RESOURCE USAGE ANALYSIS FOR THE π-CALCULUS
27
Here, Γ1 ⊗ Γ2 describes resources that are used according to Γ1 and Γ2 that are used in
an interleaving manner. However, it is not obvious how to accurately capture information
about possible synchronizations between M1 and M2 .
Model checking technologies [2] can of course be applicable to concurrent languages, but
they suffer from the state explosion problem, especially for expressive concurrent languages
like π-calculus, where resources and communication channels can be dynamically created
and passed around. Appropriate abstraction must be devised for effectively performing
the resource usage analysis for the π-calculus with model checking. Actually, our typebased analysis can be considered a kind of abstract model checking. The behavioral types
extracted by (the first two steps of) the type inference algorithm are abstract concurrent
programs, each of which captures the access behavior on each resource. Then, conformance
of the abstract program with respect to the resource usage specification is checked as a model
checking problem. It would be interesting to study a relationship between the abstraction
through our behavioral type and the abstraction techniques for concurrent programs used
in the model checking community. From that perspective, an advantage of our approach is
that our type, which describes a resource-wise behavior, has much smaller state space than
the whole program. In particular, if infinitely many resources are dynamically created, the
whole program has infinite states, but it is often the case that our behavioral types are still
finite (indeed so for the example in Figure 9). The limitation of our current analysis is that
programs can be abstracted in only one way; on the other hand, the usual abstract model
checking techniques refine abstraction step by step until the verification succeeds.
Technically, closest to our type system are that of Igarashi and Kobayashi [10] and that
of Chaki, Rajamani, and Rehof [3]. Those type systems are developed for checking the
communication behavior of a process, but by viewing a set of channels as a resource, it is
possible to use those type systems directly for the resource usage analysis. We summarize
below similarities and differences between those type systems [10, 3] and the type system
in the present paper.
(1) Whether types are supplied by the programmer or inferred automatically: Types are
inferred automatically in Igarashi and Kobayashi’s generic type [10] and the type system of
the present paper, but the type of each channel must be annotated with in Chaki et al.’s
type system. The annotated type contains information about how the values (channels, in
particular) sent along the channel are used by senders and receivers, and that information
is used to make the type checking process compositional. For the purpose of the resource
usage analysis discussed here, we think that it is a burden for programmers to declare how
channels are going to be used, since their primary concern is how resources are accessed,
not channels. Ideal would be to allow the user to specify some types and infer the others,
like in ML. For that purpose, we need to develop an algorithm to check the conformance
A ≤ B of an inferred type A to a declared type B. That seems generally harder to decide
than the trace inclusion constraint tracesx (A) ⊆ Φ, but we expect to be able to develop a
sound algorithm by properly restricting the language of declared types.
(2) The languages used as behavioral types: All the three type systems use a fragment
of CCS as the language of types to check cross-channel dependency of communications. The
types in Igarashi and Kobayashi’s generic type system for the π-calculus [10], however, lacks
hiding, so that their type system cannot be applied to obtain precise information about
resource usage. In fact, their analysis would fail even for the program in Example 2.8.
Chaki et al.’s type system does use hiding, but lacks renaming as a constructor. Without
28
N. KOBAYASHI, K. SUENAGA, AND L. WISCHIK
the renaming constructor, the most general type does not necessarily exist, which hinders
automatic type inference (recall Remark 3.12).
(3) Algorithms for checking the conformance of inferred types with respect to specifications: In Igarashi and Kobayashi’s generic type system, how to check conformance of
inferred types with respect to the user-supplied specifications was left open, and only suggested that it could be solved as a model checking problem. In Chaki et al.’s type system [3],
the conformance is expressed as A |= F (for checking the global behavior, where F is an
LTL-formula) and A ≤ A′ (for checking the conformance of declared types with respect to
inferred types). In their type checker PIPER [3], those conditions are verified using SPIN,
so that A is restricted to a finite-state process. Corresponding to the conformance check of
the above work is the check of trace inclusion constraints tracesx (A) ⊆ Φ. Our algorithm
based on the reduction to Petri nets works even when A has infinite states.
(4) The guaranteed properties: Both Igarashi and Kobayashi’s generic type [10] and the
extended type system of the present paper can guarantee a certain lock-freedom property,
that necessary communications or resource accesses are eventually performed (unless the
whole process diverges), while Chaki et al.’s type system and the type system in Section 3 of
the present paper do not. The guaranteed properties depend on the choice of the language
of behavioral types and the subtyping relation. In the latter type systems, the ordinary
simulation relation is used, so that a process’s type describes only an upper-bound of the
possible behavior of the process, not a lower-bound of the behavior like a certain resource
access is eventually performed. Rajamani et al. [8, 23] recently introduced a more elaborate
notion of simulation relation called “stuck-free conformance.” Even with the stuck-free
conformance relation, however, their type system [3] still cannot guarantee the lack of
deadlock-freedom of a process. On the other hand, by relying on an external analysis to
check deadlock-freedom, the extension in Section 5 keeps the typing rules and the subtyping
relation simple, while achieving the guarantee that necessary resource accesses are eventually
performed unless the whole process diverges.
Kobayashi’s type systems for deadlock-freedom and livelock-freedom [16, 14, 15] and its
implementation [12] form the basis of the extended type systems for partial and total liveness
properties discussed in Section 5, and are used for producing well-annotated programs.
Conversely, the behavioral types introduced in this paper can be used to refine the type
systems for deadlock-freedom and livelock-freedom. Yoshida and Honda have also studied
type systems that can guarantee certain lock-freedom properties [25, 9, 26]. So, their type
systems can also be used for checking whether programs are well-annotated in the sense of
Section 5.
In Section 5, we have utilized the existing analysis for deadlock-freedom to enhance the
result of the resource usage analysis. Other type systems for concurrent languages may also
be useful. For example, the type system for atomicity [4] can be used to infer the atomicity
of a sequence of actions in a source program. By using the atomicity information, we may
be able to reduce the state space of behavioral types and check the trace inclusion relation
etracesx (A) ⊆ Φ more efficiently.
8. Conclusion
We have formalized a type system for resource usage analysis and proved its soundness.
We have also developed a sound (but incomplete because of the last phase for deciding
RESOURCE USAGE ANALYSIS FOR THE π-CALCULUS
29
the trace inclusion relation tracesx (A) ⊆ Φ) algorithm for it in order to liberate programmers from the burden of writing complex type annotations. We have also implemented a
prototype resource usage analyzer based on the algorithm.
There remains much future work. It is necessary to assess the effectiveness of our
analysis, including the design of the type system and the algorithm for deciding the trace
inclusion relation tracesx (A) ⊆ Φ, in more detail, and refine the analysis if necessary. It is
also necessary to make the analyzer more user-friendly, by devising a method for generating
comprehensive explanation of the verification result; currently, the analyzer gives only a
yes/no answer. Extensions of the type system to deal with other typical synchronization
primitives like join-patterns and internal choice is also left for future work.
References
[1] T. Ball, B. Cook, V. Levin, and S. K. Rajamani. Slam and static driver verifier: Technology transfer of
formal methods inside microsoft. In Integrated Formal Methods 2004, volume 2999 of Springer-Verlag,
pages 1–20, 2004.
[2] T. Ball and S. K. Rajamani. The SLAM project: Debugging system software via static analysis. In
Proceedings of ACM SIGPLAN/SIGACT Symposium on Principles of Programming Languages, pages
1–3, 2002.
[3] S. Chaki, S. Rajamani, and J. Rehof. Types as models: Model checking message-passing programs. In
Proceedings of ACM SIGPLAN/SIGACT Symposium on Principles of Programming Languages, pages
45–57, 2002.
[4] S. Q. Cormac Flanagan. A type and effect system for atomicity. In Proceedings of ACM SIGPLAN
Conference on Programming Language Design and Implementation, pages 338–349, 2003.
[5] R. DeLine and M. Fähndrich. Enforcing high-level protocols in low-level software. In Proceedings of ACM
SIGPLAN Conference on Programming Language Design and Implementation, pages 59–69, 2001.
[6] R. DeLine and M. Fähndrich. Adoption and focus: Practical linear types for imperative programming.
In Proceedings of ACM SIGPLAN Conference on Programming Language Design and Implementation,
2002.
[7] J. S. Foster, T. Terauchi, and A. Aiken. Flow-sensitive type qualifiers. In Proceedings of ACM SIGPLAN
Conference on Programming Language Design and Implementation, pages 1–12, 2002.
[8] C. Fournet, T. Hoare, S. K. Rajamani, and J. Rehof. Stuck-free conformance. In CAV’04, volume 3114
of Lecture Notes in Computer Science, pages 242–254. Springer-Verlag, 2004.
[9] K. Honda and N. Yoshida. A uniform type structure for secure information flow. In Proceedings of ACM
SIGPLAN/SIGACT Symposium on Principles of Programming Languages, pages 81–92, 2002.
[10] A. Igarashi and N. Kobayashi. A generic type system for the pi-calculus. Theoretical Computer Science,
311(1-3):121–163, 2004.
[11] A. Igarashi and N. Kobayashi. Resource usage analysis. ACM Transactions on Programming Languages
and Systems, 27(2):264–313, 2005. Preliminary summary appeared in Proceedings of POPL 2002.
[12] N. Kobayashi. Typical: A type-based static analyzer for the pi-calculus. Tool available at
http://www.kb.ecei.tohoku.ac.jp/~koba/typical/.
[13] N. Kobayashi. A partially deadlock-free typed process calculus. ACM Transactions on Programming
Languages and Systems, 20(2):436–482, 1998.
[14] N. Kobayashi. A type system for lock-free processes. Information and Computation, 177:122–159, 2002.
[15] N. Kobayashi. Type-based information flow analysis for the pi-calculus. Acta Informatica, 42(4-5):291–
347, 2005.
[16] N. Kobayashi, S. Saito, and E. Sumii. An implicitly-typed deadlock-free process calculus. In Proceedings
of CONCUR2000, volume 1877 of Lecture Notes in Computer Science, pages 489–503. Springer-Verlag,
August 2000.
[17] K. Marriott, P. J. Stuckey, and M. Sulzmann. Resource usage verification. In Proceedings of the First
Asian Symposium on Programming Languages and Systems (APLAS 2003), volume 2895 of Lecture
Notes in Computer Science, pages 212–229, 2003.
30
N. KOBAYASHI, K. SUENAGA, AND L. WISCHIK
[18] E. W. Mayr. An algorithm for the general petri net reachability problem. SIAM Journal on Computing,
13(3):441–461, 1984.
[19] R. Milner. Communication and Concurrency. Prentice Hall, 1989.
[20] N. Nguyen and J. Rathke. Typed static analysis for concurrent, policy-based, resource access control.
draft.
[21] E. Pelz. Closure properties of deterministic petri nets. In STACS 87: 4th Annual Symposium on Theoretical Aspects of Computer Science, volume 247 of Lecture Notes in Computer Science, pages 371–382.
Springer-Verlag, 1987.
[22] J. L. Peterson. Petri Net Theory and the Modeling of Systems. Prentice-Hall, 1981.
[23] S. K. Rajamani and J. Rehof. Models for contract conformance. In ISOLA2004, First International
Symposium on Leveraging Applications of Formal Methods, 2004.
[24] C. Skalka and S. Smith. History effects and verification. In Proceedings of the First Asian Symposium
on Programming Languages and Systems (APLAS 2004), volume 3302 of Lecture Notes in Computer
Science, pages 107–128, 2004.
[25] N. Yoshida. Graph types for monadic mobile processes. In FST/TCS’16, volume 1180 of Lecture Notes
in Computer Science, pages 371–387. Springer-Verlag, 1996.
[26] N. Yoshida. Type-based liveness guarantee in the presence of nontermination and nondeterminism.
Technical Report 2002-20, MSC Technical Report, University of Leicester, April 2002.
RESOURCE USAGE ANALYSIS FOR THE π-CALCULUS
31
Appendix
Appendix A. Properties of the Subtyping Relation
This section states and proves the properties of the subtyping relation, which are used
in the proof of type soundness (Theorems 3.14 and 5.8, in particular the proofs of the
lemmas in Appendices B and C), and in the type inference algorithm described in Section 4
(in particular, for transforming constraints on behavioral types).
Actually, there are two subtyping relations; the basic one in Definition 3.5 and the
extended one in Definition 5.7. Since the proofs are almost the same, we state and prove
the properties of the basic and extended ones simultaneously. In a few places, we have
an additional condition to check for the extended case. Such places will be marked by
“Extended case only.” When we are discussing the basic case, attributes attached to
actions should be ignored. We also omit them even for the extended case when they are
not important.
Lemma A.1 (Simulation relation).
(1) The subtyping relation is reflexive and transitive.
(2) (Simulation-up-to) Let R be a relation on behavioral types such that whenever A1 RA2
then
l
l
(i) A1 → A′1 implies A2 =⇒ A′2 and A′1 R ≤ A′2 for some A′2 and
(ii) disabled(A1 , S) implies disabled(A2 , S).
Then R ⊆ ≤ . Condition (ii) is required only for the extended case.
Proof. Part 1 is trivial by the definition. To show Part 2, suppose R is a simulation up to.
l
We show that R′ = (R ≤ ) ∪ R is a simulation, i.e., whenever A1 R′ A2 , (i) A1 −→ A′1 implies
l
A2 =⇒ A′2 and A′1 R′ A′2 for some A′2 and (Extended case only) (ii) disabled(A1 , S) implies
disabled(A2 , S). Suppose A1 R′ A2 . The case where A1 RA2 is trivial by the definition of the
simulation-up-to. To check the other case, suppose A1 RA3 ≤ A2 . To show (i), suppose
l
l
also that A1 −→ A′1 . Since R is a simulation up to, there exists A′3 such that A3 =⇒ A′3 and
l
l
A′1 R ≤ A′3 . By A3 =⇒ A′3 and A3 ≤ A2 , we have A′2 such that A2 =⇒ A′2 and A′3 ≤ A′2 .
Since ≤ is transitive, we have A′1 R ≤ A′2 , which implies A′1 R′ A′2 .
Extended case only: To show (ii), suppose disabled(A1 , S). Since R is a simulation up
to, we have disabled(A3 , S), which implies disabled(A2 , S).
Lemma A.2 (Structural congruence).
(1) A|0 ≈ A
(2) A|B ≈ B|A
(3) A|(B|C) ≈ (A|B)|C
(4) A⊕B ≈ B⊕A
(5) A⊕(B⊕C) ≈ (A⊕B)⊕C
(6) ∗A ≈ A|∗A
(7) (νx)(A|B) ≈ (νx)A | B if x ∈
/ FV(B)
(8) (νx)(A⊕B) ≈ (νx)A ⊕ B if x ∈
/ FV(B)
(9) [µα.A/α]A ≈ µα.A
Proof. These proofs are all standard.
32
N. KOBAYASHI, K. SUENAGA, AND L. WISCHIK
We next show that ≤ is a precongruence. We first show it for some basic type constructors.
Lemma A.3 (Precongruence, simple cases). If A ≤ A′ then
(1) A|B ≤ A′ |B ′ if B ≤ B ′
(2) hx/yiA ≤ hx/yiA′
(3) (νx)A ≤ (νx)A′
(4) A↑S ≤ A′ ↑S
(5) A↓S ≤ A′ ↓S
Proof. These follow from the fact that the following relations are all simulations-up-to.
R1 = {(A | B, A′ | B ′ ) | A ≤ A′ , B ≤ B ′ }
R2 = {(he
y /e
xiA, he
y /e
xiA′ ) | A ≤ A′ }
R3 = {((νx) A, (νx) A′ ) | A ≤ A′ }
R4 = {(A↑S , A′ ↑S ) | A ≤ A′ }
R5 = {(A↓S , A′ ↓S ) | A ≤ A′ }
We now show that ≤ is closed under arbitrary type constructors. FTV(B) below is
the set of free (i.e., not bound by µ) behavioral type variables.
Lemma A.4 (Precongruence, general cases). If A ≤ A′ and FTV(B) ⊆ {α}, then
[A/α]B ≤ [A′ /α]B.
l
l
Proof. Let R= {([A/α]B, [A′ /α]B)}. We will prove (i) if [A/α]B −→ B1 then [A′ /α]B =⇒
l
B1′ with B1 R ≤ B1′ , by induction on the derivation of [A′ /α]B −→ B1′ . We will also prove
(ii) disabled([A/α]B, S) implies disabled([A′ /α]B, S), by induction on the structure of B in
the extended case. In other words, R is a simulation-up-to. Hence (Lemma A.1.2) it is in
≤.
We start with (i), with case analysis on the last rule used. If B = α, then the required
condition follows immediately from A ≤ A′ . So we consider the case B 6= α below.
(1) Case (TR-Act). In this case, B = l.Bx , so
l
[A/α]B = l.[A/α]Bx −→ [A/α]Bx = B1 .
We also have
l
[A′ /α]B = l.[A/α]Bx −→ [A′ /α]Bx = B1′ .
By construction of R, we have B1 R B1′ ≤ B1′ as required.
(2) Case (Tr-Par1). We show only the left case. B = Bx |By and we assumed
l
[A/α]Bx −→ Bx1 to make
l
[A/α]B = [A/α]Bx |[A/α]By −→ Bx1 |[A/α]By = B1 .
l
′ with B
′
By the induction hypothesis, [A′ /α]Bx =⇒ Bx1
x1 R ≤ Bx1 . (Note that α is
′
not free in Bx1 or Bx1 . ) That gives
l
′
[A′ /α]B = [A′ /α]Bx |[A′ /α]By =⇒ Bx1
|[A′ /α]By = B1′ .
′ , there exists C such
It remains to prove B1 R ≤ B1′ . By the condition Bx1 R ≤ Bx1
that
′
Bx1 = [A/α]C
[A′ /α]C ≤ Bx1
RESOURCE USAGE ANALYSIS FOR THE π-CALCULUS
33
So, we get:
B1 = [A/α](C | By ) R [A′ /α](C | By )
′ | [A′ /α]B = B ′ .
= [A′ /α]C | [A′ /α]By ≤ Bx1
y
1
Here, we have used Lemma A.3, Part 1.
(3) Case (Tr-Par2). We show only the left case.
B = Bx |By and we assumed
y
x
[A/α]Bx −→ Bx1 and [A/α]By −→ By1 to make
{x,y}
[A/α]B = [A/α]Bx |[A/α]By −→ Bx1 |By1 = B1 .
x
x
′ and [A′ /α]B =⇒ B ′ with B
By the induction hypothesis, [A′ /α]Bx =⇒ Bx1
x1 R ≤
y
y1
′
′
Bx1 and By1 R ≤ By1 . That gives
{x,y}
′
′
[A′ /α]B = [A′ /α]Bx | [A′ /α]By =⇒ Bx1
| By1
= B1′ .
′
′ and B
It remains to prove B1 R ≤ B1′ . From Bx1 R ≤ Bx1
y1 R ≤ By1 , there exist
Cx and Cy such that
′
[A′ /α]Cx ≤ Bx1
′
′
[A /α]Cy ≤ By1
Bx1 = [A/α]Cx
By1 = [A/α]Cy
Hence, B1 = [A/α](Cx | Cy ) R [A′ /α](Cx | Cy ) ≤ B1′ .
(4) Cases (Tr-Com) and (Tr-Or). These cases follow immediately from the induction
hypothesis.
l
l
(5) Case (Tr-Rep). Then B = ∗Bx and [A/α]B = ∗[A/α]Bx −→. [A/α]B −→ B1
must have been derived from
l
[A/α](Bx | ∗Bx ) = [A/α]Bx | ∗[A/α]Bx −→ B1 .
By the induction hypothesis, there exists B1′ such that
B1 R ≤ B1′
and
l
[A′ /α](Bx | ∗Bx ) =⇒ B1′ .
l
Using (Tr-Rep), we get [A′ /α]B =⇒ B1′ as required.
(6) Case (Tr-Rec). Then, we have B = µβ.Bx to make
l
[A/α]B = µβ.[A/α]Bx −→ B1
l
where we assumed [µβ.[A/α]Bx /β][A/α]Bx −→ B1 . But β does not clash with A
or α so these two substitutions swap around, giving
l
[A/α][µβ.Bx /β]Bx −→ B1 .
By the induction hypothesis,
l
[A′ /α][µβ.Bx /β]Bx =⇒ B1′
with B1 R ≤ B1′ . Hence
l
[A′ /α]B = µβ.[A′ /α]Bx =⇒ B1′
as required.
34
N. KOBAYASHI, K. SUENAGA, AND L. WISCHIK
[e
y /e
x]l
(7) Case (Tr-Rename). Then, B = he
y /e
xiBx . [A/α]B −→ he
y /e
xiBx1 = B1 must have
l
been derived from [A/α]Bx −→ Bx1 . From the induction hypothesis, we get
l
′
[A′ /α]Bx −→ Bx1
′
Bx1 R ≤ Bx1
.
′
′ . It remains to prove B R ≤ B ′ . By B
y /e
xiBx1
Let B1′ = he
x1 R ≤ Bx1 , there exists
1
1
C such that
′
Bx1 = [A/α]C
[A′ /α]C ≤ Bx1
.
So, we have:
B1 = [A/α]he
y /e
xiC R [A′ /α]he
y /e
xiC
′ = B′ .
= he
y /e
xi[A′ /α]C ≤ he
y /e
xiBx1
1
Here, we used the fact that ≤ is preserved by he
y /e
xi (Lemma A.3, Part 2).
(8) Cases (Tr-Hiding), (Tr-Exclude), and (Tr-Project): Similar to (Tr-Rename).
We use the fact that ≤ is preserved by ν, ·↓S , and ·↑S (Lemma A.3).
Extended case only: In addition we need to show that disabled([A/α]B, S) implies
disabled([A′ /α]B, S). This follows by straightforward induction on the structure of B.
Lemma A.5 (Substitution).
(1) he
y /e
xi0 ≈ 0
(2) he
y /e
xi(a.A) ≈ ([e
y /e
x]a).he
y /e
xiA
(3) he
y /e
xi(z ξ .A) ≈ ([e
y /e
x]z)ξ .he
y /e
xiA
(4) he
y /e
xi(A|B) ≈ he
y /e
xiA | he
y /e
xiB
(5) he
y /e
xi(A⊕B) ≈ he
y /e
xiA ⊕ he
y /e
xiB
(6) he
y /e
xi(∗A) ≈ ∗(he
y /e
xiA)
(7) he
y /e
xihb/aiA ≈ h[e
y /e
x]b/aihe
y /e
xiA if target(a)∩{e
x, ye}=∅
(8) he
y /e
xi(νz)A ≈ (νz)(he
y /e
xiA) if {z}∩{x, y}=∅
(9) he
y /e
xi(A↑S ) ≈ (he
y /e
xiA)↑S , and
he
y /e
xi(A↓S ) ≈ he
y /e
xiA↓S ≈ A↓S ,
if S∩{x, y}=∅
(10) he
y /e
xi(A↑S ) ≈ A↑S , if {e
x} ⊆ S
Proof. Most parts are straightforward, although Part 4 is non-obvious in the case of labels
y /e
xi(A|B), he
y /e
xiA|he
y /e
xiB)} and prove
{x, y}. For Part 4, we construct a relation S = {(he
−1
S and S are simulations. The interesting case is when we infer
τ
he
y /e
xiA|he
y /e
xiB → he
y /e
xiA′ |he
y /e
xiB ′
from
z
1
A→
A′
z
2
B′
B→
[e
y /e
x]z1 = [e
y /e
x]z2 .
This gives
A|B
Hence
he
y /e
xi(A|B)
And hence as required
{z1 ,z2 }
→
A′ |B ′ .
{[e
y /e
x]z1 ,[e
y /e
x]z2 }
−→
τ
he
y /e
xi(A′ |B ′ ).
he
y /e
xi(A|B) → he
y /e
xi(A′ |B ′ ).
Part 9. Here we construct S = {( he
y /e
xi(A↑S ), (he
y /e
xiA)↑S )} where S does not clash
with {e
x, ye}, and we prove that S and S −1 are simulations. We focus on two cases.
RESOURCE USAGE ANALYSIS FOR THE π-CALCULUS
35
[e
y /e
x]l
(1) Suppose (he
y /e
xiA)↑S −→ (he
y /e
xiA′ )↑S is inferred from
l
A → A′
and
target ([e
y /e
x]l)∩S = ∅.
[e
y /e
x]l
We must infer that he
y /e
xi(A↑S ) −→ he
y /e
xi(A′ ↑S ). This requires target (l)∩S = ∅,
which we prove as follows. It is assumed that S does not clash, so {e
x, ye}∩S = ∅. We
also have target ([e
y /e
x]l)∩S = ∅, and so [e
y /e
x](target (l))∩S = ∅. Let T = target (l).
Suppose z ∈ T . Then either z ∈ x
e so z ∈
/ S, or z ∈ ye so z ∈
/ S, or z ∈
/ {e
x, ye} so
z ∈ [e
y /e
x]T so z ∈
/ S. In all cases z ∈
/ S, so T ∩S = ∅ as required.
l
τ
y /e
xiA′ )↑S is inferred from A → A′ and target ([e
y /e
x]l) ⊆ S.
(2) Suppose (he
y /e
xiA)↑S → (he
τ
We must infer he
y /e
xi(A↑S ) → he
y /e
xi(A′ ↑S ). This requires target(l) ⊆ S, which
we prove as follows. Once again let T = target (l). We have {e
x, ye}∩S = ∅ and
[e
y /e
x]T ⊆ S. Suppose z ∈ T . Then [e
y /e
x]z ∈ [e
y /e
x]T , and [e
y /e
x]z ∈ S. Either z ∈ x
e
so y ∈ S, which is a contradiction. Or z 6∈ x
e, so [e
y /e
x]z = z ∈ S. Hence T ⊆ S as
required.
Lemma A.6 (Exclusion and
(1) 0↑S ≈0
(2) (at .A)↑S ≈at .(A↑S )
(3) (at .A)↑S ≈τt .A↑S
(4) (z ξ .A)↑S ≈z ξ .(A↑S )
(5) (z ξ .A)↑S ≈τc .A↑S
(6) (A|B)↑S ≈A↑S | B↑S
(7) (A⊕B)↑S ≈A↑S ⊕B↑S
(8) (∗A)↑S ≈∗(A↑S )
(9) (A↑S )↑T ≈A↑S∪T
(10) A↑S ≈A
(11) A↑S ≤ 0
Projection).
0↓S ≈0
(at .A)↓S ≈τt .(A↓S )
(at .A)↓S ≈at .A↓S
(z ξ .A)↓S ≈τc .A↓S
(z ξ .A)↓S ≈z ξ .(A↓S )
(A|B)↓S ≈A↓S | B↓S
(A⊕B)↓S ≈A↓S ⊕B↓S
(∗A)↓S ≈∗(A↓S )
(A↓S )↓T ≈A↓S∩T
A↓S ≤ 0
A↓S ≈A
if
if
if
if
target(a)∩S=∅
target(a)⊆S
target(z ξ )∩S=∅
target(z ξ )⊆S
if FV(A)∩S=∅
if FV(A)⊆S
Proof. Straightforward.
Lemma A.7 (Simulation).
(1) If A1 ≤ A2 then tracesx (A1 ) ⊆ tracesx (A2 ) for any x.
(2)
(3)
(4)
(5)
(6)
(7)
{x,y}
x y
If A −→ A′ then A →→ A′ .
A ≤ A⊕B
A⊕A ≤ A
A ≤ A↑S | A↓S
If [B/α]A ≤ B then µα.A ≤ B
B1 ⊕B2 ≤ A if and only if B1 ≤ A and B2 ≤ A
Proof. These proofs are largely standard.
Part 1 follows immediately from the definitions of subtyping and traces.
Part 6. Suppose [B/α]A ≤ B. Let R be
{([µα.A/α]A′ , [B/α]A′ ) | FTV(A′ ) = {α}}.
By Lemma A.3.2, It suffices to prove that R is a simulation up to.
l
Suppose that [µα.A/α]A′ R [B/α]A′ and [µα.A/α]A′ −→ A′′ . We show that there
l
exists B ′ such that [B/α]A′ =⇒ B ′ and A′′ R ≤ B ′ by induction on the derivation of
36
N. KOBAYASHI, K. SUENAGA, AND L. WISCHIK
l
[µα.A/α]A′ −→ A′′ , with case analysis on the last rule used. We show main cases; the other
cases are similar or straightforward.
l
• Case (TR-Act): [µα.A/α]A′ −→ A′′ is derived from
l
l. [µα.A/α]A1 −→ [µα.A/α]A1 where A′ = l. A1 and A′′ = [µα.A/α]A1 . Thus,
l
[B/α]A′ = l. [B/α]A1 −→ [B/α]A1 .
l
• Case (TR-Par1): [µα.A/α]A′ −→ A′′ is derived from
l
[µα.A/α]A1 −→ A′1 where A′ = A1 | A2 and A′′ = A′1 | [µα.A/α]A2 . By the induction
l
hypothesis, there exists B1′ such that [B/α]A1 =⇒ B1′ and A′1 R ≤ B1′ . Thus, we
l
have [B/α]A′ =⇒ B1′ | [B/α]A2 . It remains to show A′′ = A′1 | [µα.A/α]A2 R ≤
B1′ | [B/α]A2 . From A′1 R ≤ B1′ , we get
A′1 = [µα.A/α]C
[B/α]C ≤ B1′
for some C. So,
A′′ = A′1 | [µα.A/α]A2 = [µα.A/α](C | A2 )
R [B/α](C | A2 ) = [B/α]C | [B/α]A2 ≤ B1′ | [B/α]A2
{x,y}
x
• Case (TR-Par2): [µα.A/α]A′ −→ A′′ is derived from [µα.A/α]A1 −→ A′1 and
y
[µα.A/α]A2 −→ A′2 where A′ = A1 | A2 and A′′ = A′1 | A′2 . From the induction
x
hypothesis, there exist B1′ and B2′ such that [B/α]A1 =⇒ B1′ and A′1 R ≤ B1′ and
{x,y}
y
[B/α]A2 =⇒ B2′ and A′2 R ≤ B2′ . Thus, we have [B/α]A′ =⇒ B1′ | B2′ . From
A′1 R ≤ B1′ and A′2 R ≤ B2′ , we get A′1 | A′2 R ≤ B1′ | B2′ as required.
• Case (TR-Rec):
l
– Case A′ = µβ.A1 : [µα.A/α]A′ −→ A′′ is derived from
[µα.A/α][µβ.A1 /β]A1
l
= [µβ.[µα.A/α]A1 /β][µα.A/α]A1 −→ A′′ .
Here, we assumed without loss of generality that β is not free in A and B.
Thus, by the induction hypothesis, there exists B ′ such that
l
[µβ.[B/α]A1 /β][B/α]A1 = [B/α][µβ.A1 /β]A1 =⇒ B ′
l
and A′′ R ≤ B ′ . Using (Tr-Rec), we obtain [B/α]A′ = µβ.[B/α]A1 =⇒ B ′
as required.
– Case A′ = α: [µα.A/α]A′ is equal to µα.A. From µα.A ≤ B, there exists B ′
l
such that B =⇒ B ′ and A′′ ≤ B ′ as required.
Extended case only: We also need to prove that
disabled([µα.A/α]A′ , S)
implies
disabled([B/α]A′ , S)
for any
A′ .
This is proved by induction on the derivation of disabled([µα.A/α]A′ , S). We show the only
non-trivial case, where disabled([µα.A/α]A′ , S) has been derived by using the last rule in
Figure 7. The other cases follow immediately from the induction hypothesis.
There are two cases to consider.
RESOURCE USAGE ANALYSIS FOR THE π-CALCULUS
37
• Case where A′ = α: Then, [µα.A/α]A′ = µα.A and disabled(µα.A, S) must have
been deduced from
disabled([µα.A/α]A, S). By the induction hypothesis, we have disabled([B/α]A, S).
By the assumption [B/α]A ≤ B, we have disabled(B, S) as required (note that
[B/α]A′ = B in this case).
• Case where A′ = µβ.C. Let C ′ be [µα.A/α]C. Then, [µα.A/α]A′ = µβ.C ′ , and
disabled(µβ.C ′ , S) must have been derived from disabled([µβ.C ′ /β]C ′ , S). Here, we
note
[µβ.C ′ /β]C ′ = [µα.A/α][µβ.C/β]C.
So, from the induction hypothesis, we get
disabled([B/α][µβ.C/β]C, S), i.e.,
disabled([µβ.[B/α]C/β][B/α]C, S).
By using the last rule of Figure 7, we get disabled([B/α]A′ , S) as required.
Appendix B. Proof of the Subject Reduction Property
In this section, we prove the subject reduction property used in the proofs of Theorems 3.14 and 5.8. As in Appendix A, we prove it for the basic and extended cases
simultaneously.
Lemma B.1 (Weakening).
(1) If Γ ⊲ v : σ and x 6∈ dom(Γ), then Γ, x : σ ′ ⊲ v : τ .
(2) If Γ ⊲ P : A and x ∈
/ FV(P ) and x not in dom(Γ) or FV(A) then Γ, x : σ ⊲ P : A.
Proof. Part 1 is straightforward. Part 2 is proved by straightforward induction on the
derivation of Γ ⊲ P : A.
Lemma B.2 (Judgement substitution).
(1) (For values) If Γ, x
e:σ
e ⊲ y : σ and Γ ⊲ ve : σ
e then Γ ⊲ [e
v /e
x]y : σ.
(2) (For processes) If Γ, x
e:σ
e ⊲ P : A and Γ ⊲ ve : σ
e then Γ ⊲ [e
v /e
x]P : he
v /e
xiA.
Proof. Part 1. Either y = xi for some i, in which case [e
v /e
x]y = vi and σ = σi , so that the
result follows from Γ ⊲ ve : σ
e. Or y ∈
/x
e, in which case [e
v /e
x]y = y and y : σ is in Γ. We remark
that types σ never have free names.
Part 2. By induction on the derivation of Γ⊲P : A. Most cases follow straightforwardly
on Lemma A.5. We consider four particular cases.
(1) Case (T-Sub), where Γ, x
e : τe ⊲ P : A is inferred from
Γ, x
e : τe ⊲ P : A′
A′ ≤ A
From the induction hypothesis, Γ ⊲ [e
v /e
x]P : he
v /e
xiA′ . By Lemma A.3.2 and assump′
′
tion A ≤ A we get he
v /e
xiA ≤ he
v /e
xiA, and hence as required Γ ⊲ [e
v /e
x]P : he
v /e
xiA.
(2) Case (T-NewR), where Γ, x
e : τe ⊲ (NΦ z)P : A↑{z} is inferred from
Γ, x
e : τe, z : res ⊲ P : A
tracesz (A) ⊆ Φ
Assume by alpha-renaming that z does not clash with x
e or ve. From Lemma A.5.9
we get A↓{z} ≈ (he
v /e
xiA)↓{z} , giving tracesz (A) = tracesz (he
v /e
xiA) and hence
tracesz (he
v /e
xiA) ⊆ Φ. From Γ ⊲ e
v : τe and Lemma B.1, we get Γ, z : res ⊲ ve : τe. So, by
the induction hypothesis, Γ, z : res ⊲ [e
v /e
x]P : he
v /e
xiA. These two together give
Γ ⊲ (NΦ z)[e
v /e
x]P : (he
v /e
xiA)↑{z} .
38
N. KOBAYASHI, K. SUENAGA, AND L. WISCHIK
For the process (NΦ z)[e
v /e
x]P , we can push the substitution out by definition of the
substitution operator and because z 6∈ {e
x, ve}. For the behavior (he
v /e
xiA)↑{z} we use
Lemma A.5.9 to push it out. Hence as required,
Γ ⊲ [e
v /e
x](NΦ z)P : he
v /e
xi(A↑{z} ).
Extended case only: Just replace traces with etraces in the above reasoning.
e y iA1 |A2 ) is inferred from
(3) Case (T-Out), where Γ, x
e : τe ⊲ zhwi. P : z. (hw/e
Γ, x
e : τe ⊲ P : A2
Γ, x
e : τe ⊲ w
e:σ
e
Γ, x
e : τe ⊲ z : chanh(e
y:σ
e)A1 i
Part 1 implies Γ ⊲ [e
v /e
x]w
e:σ
e and Γ ⊲ [e
v /e
x]z : chanh(e
y:σ
e)A1 i. From the induction
hypothesis, we get Γ ⊲ [e
v /e
x]P : he
v /e
xiA2 . These three give
Γ ⊲ [e
v /e
x]zh[e
v /e
x]wi.
e [e
v /e
x]P : [e
v /e
x]z. (h[e
v /e
x]w/e
e y iA1 |he
v /e
xiA2 )
For the process we push the substitution out by definition of the substitution operator. For the behavior we push it out using several parts of Lemma A.5.
(4) Case (T-In), where Γ, x
e : τe ⊲ z(e
y ). P : z. (A2 ↑{ey} ) is inferred from
Γ, ye : σ
e, x
e : τe ⊲ P : A2
Γ, x
e : τe ⊲ z : chanh(e
y:σ
e)A1 i
A2 ↓{ey } ≤ A1
We use three deductions. First from Part 1 we get Γ ⊲ [e
v /e
x]z : chanh(e
y:σ
e)A1 i.
Second, from assumption A2 ↓{ey} ≤ A1 and Lemma A.3.2 we get he
v /e
xi(A2 ↓{ey} ) ≤
he
v /e
xiA1 . The substitution on the right disappears because FV(A1 ) ⊆ {e
y } and we
can assume by alpha-renaming that ye does not clash with {e
x, ve}. The substitution on
the left can be pushed inside by Lemma A.5.9. These together give (he
v /e
xiA2 )↓{ey} ≤
A1 . And third, from the induction hypothesis we get Γ, ye : σ
e ⊲[e
v /e
x]P : he
v /e
xiA2 . These
three give
Γ ⊲ [e
v /e
x]z(e
y ). [e
v /e
x]P : [e
v /e
x]z. ((he
v /e
xiA2 )↑{ey } )
As in the previous case we push the substitution out in the process and the behavior
to get, as required,
Γ ⊲ [e
v /e
x](z(e
y ). P ) : he
v /e
xi(z. (A2 ↑{ey} )).
Lemma B.3 (Subject-reduction).
(1) If Γ ⊲ P : A and P Q then Γ ⊲ Q : A.
L
L
(2) (Subject-reduction) If P → P ′ and Γ ⊲ P : A then A =⇒ A′ and Γ ⊲ P ′ : A′ for some
A′ .
Proof. Part 1. By induction on the derivation of P Q. Most cases use Lemma A.2.
The case for (νx) P |Q (νx) (P |Q) uses Lemma B.1. The only interesting case is that for
(NΦ x)P |Q (NΦ x)(P |Q) with x ∈
/ FV(Q). The judgement Γ ⊲ (NΦ x)P |Q : A must have
been inferred from
Γ, x : res ⊲ P : A3
tracesx (A3 ) ⊆ Φ
A3 ↑{x} ≤ A1
Γ ⊲ Q : A2
A1 |A2 ≤ A
From these and Lemma B.1, we infer
Γ, x : res ⊲ P |Q : A3 |A2 .
RESOURCE USAGE ANALYSIS FOR THE π-CALCULUS
39
By alpha-renaming assume x ∈
/ FV(A2 ). By Lemmas A.6.6 and A.6.11 we get (A3 |A2 )↓{x} ≈
A3 ↓{x} |A2 ↓{x} ≤ A3 ↓{x} , and then by Lemma A.7.1 we get tracesx (A3 |A2 ) ⊆ tracesx (A3 ),
and so tracesx (A3 |A2 ) ⊆ Φ. This gives
Γ ⊲ (NΦ x)(P |Q) : (A3 |A2 )↑{x} .
Finally (A3 |A2 )↑{x} ≤ A3 ↑{x} |A2 ≤ A1 | A2 ≤ A. This gives as required
Γ ⊲ (NΦ x)(P |Q) : A.
Extended case only: Just replace traces with etraces in the above reasoning.
L
Part 2. By induction on the derivation of P −→ P ′ . We show main cases. The other
cases are straightforward.
• Case (R-Com): We are given
Γ ⊲ xhe
v i. P1 | x(e
y ). P2 : A.
This must have been deduced from
v i. P1 : A1
Γ ⊲ xhe
Γ ⊲ x(e
y ). P2 : A2
A1 |A2 ≤ A.
(B.1)
Γ ⊲ xhe
v i. P1 : A1 and Γ ⊲ x(e
y ). P2 : A2 must have been deduced from
Γ ⊲ P1 : A3
(B.2)
Γ ⊲ x : chanh(e
y:σ
e)A4 i
(B.3)
Γ ⊲ vi : σi
(B.4)
x. (he
v /e
y iA4 | A3 ) ≤ A1
(B.5)
Γ, ye : σ
e ⊲ P2 : A5
(B.6)
and
Γ ⊲ x : chanh(e
y:σ
e)A4 i
A5 ↓{ey } ≤ A4
(B.7)
x. (A5 ↑{ey } ) ≤ A2
A′
(B.8)
: A′
A′ .
respectively. We must show A ⇒
and Γ ⊲ P1 |[e
v /e
y ]P2
for some
We pick
′
′
′
some A such that A ⇒ A and A ≥ he
v /e
y iA4 |A3 |A5 ↑{ey } . The existence of such
A′ is guaranteed by A ≥ x. (he
v /e
y iA4 |A3 )|x. (A5 ↑{ey} ) −→ he
v /e
y iA4 |A3 |A5 ↑{ey } , which
follows from (B.1) and (B.5) and (B.8), and the definition of the subtyping relation
(Definition 3.5). It remains to prove Γ ⊲ P1 |[e
v /e
y ]P2 : A′ . We start with the judgment
(B.6),
Γ, ye : σ
e ⊲ P2 : A5 .
By Lemma B.2.2,
Γ ⊲ [e
v /e
y ]P2 : he
v /e
y iA5 .
Hence
Γ ⊲ P1 |[e
v /e
y ]P2 : A3 |he
v /e
y iA5 .
40
N. KOBAYASHI, K. SUENAGA, AND L. WISCHIK
Therefore, the required result Γ ⊲ P1 | [e
v /e
y ]P2 : A′ follows by (T-Sub), if we show
′
A3 | he
v /e
y iA5 ≤ A . It follows by:
A3 | he
v /e
y iA5
≤
≤
≤
≤
≤
A3 | he
v /e
y i(A5 ↓{ey } | A5 ↑{ey } )
(Lemma A.7.5)
A3 | he
v /e
y i(A5 ↓{ey } ) | he
v /e
y i(A5 ↑{ey } )
(Lemma A.5.4)
A3 | he
v /e
y i(A5 ↓{ey } ) | A5 ↑{ey }
(Lemma A.5.10)
A3 | he
v /e
y iA4 |A5 ↑{ey }
(assumption B.7 above)
A′
(the definition of A′ ).
• Case (R-Acc): We are given Γ ⊲ accξ (x).P1 : A. This must have been derived from
– Γ ⊲ P1 : A1
– Γ ⊲ x : res
– xξ .A1 ≤ A.
We have to show that
– Γ ⊲ P1 : A′
xξ
– A =⇒ A′ .
xξ
Let A′ be a behavioral type that satisfies A =⇒ A′ and A′ ≥ A1 . Such A′ is
xξ
guaranteed to exist by A ≥ xξ .A1 −→ A1 . Then, Γ ⊲ P1 : A′ follows from Γ ⊲ P1 : A1
and A′ ≥ A1 .
• Case (R-NewR1): We are given Γ ⊲ (NΦ x)P1 : A This must have been derived from
– Γ, x : res ⊲ P1 : A1
– tracesx (A1 ) ⊆ Φ
– A ≥ A1 ↑{x} .
We have to show that there exists A′ such that
−ξ
– Γ ⊲ (NΦ x)P1′ : A′
– A =⇒ A′
xξ
where P1 −→ P1′ .
By the induction hypothesis, there exists A′1 that satisfies Γ, x : res ⊲ P1′ : A′1
xξ
xξ
and A1 =⇒ A′1 . Using (Tr-Project), we get A1 ↓{x} =⇒ A′1 ↓{x} . So, from the
definition of traces and tracesx (A1 ) ⊆ Φ, we get tracesx (A′1 ) ⊆ Φ−ξ . By using
−ξ
(T-NewR), we get Γ ⊲ (NΦ x)P1′ : A′1 ↑{x} .
It remains to show there exists A′ such that A′1 ↑{x} ≤ A′ and A =⇒ A′ . That
xξ
follows from A ≥ A1 ↑{x} =⇒ A′1 ↑{x} . Here, the latter relation follows from A1 =⇒
A′1 and rule (Tr-Exclude).
Extended case only: Just replace traces with etraces in the above reasoning.
• Case (R-SP): This follows immediately from Part 1 and the induction hypothesis.
Appendix C. Proofs of the Lemma for Theorem 5.8
This section gives a proof of the lemma “Disabled” used in the proof of Theorem 5.8.
Lemma C.1 (Disabled). If well annotated (P ) and Γ ⊲pl P : A with bool 6∈ codom(Γ), then
P −→
6
implies disabled(A, S) for any S.
6
imply ¬active(P ) by the definition
Proof. We first note that well annotated (P ) and P −→
of well annotated (P ). So, it is sufficient to show (i)Γ ⊲pl P : A, (ii)P −→,
6
(iii)¬active(P ),
RESOURCE USAGE ANALYSIS FOR THE π-CALCULUS
41
and (iv) bool 6∈ codom(Γ) imply disabled(A, S) for any S. We prove this by induction on
the derivation of Γ ⊲pl P : A, with case analysis on the last rule.
• Case (T-Zero): In this case, A = 0, so we have disabled(A, S) for any S.
v i. P1 and
• Case (T-Out): In this case, P = xt he
v /e
y iA1 | A2 ). Since ¬active(P ), t = ∅. So, we have disabled(A, S) for any
A = xt . (he
S.
• Case (T-In): In this case, P = xt (e
y ). P1 and A = xt . (A2 ↑{ey } ). Since ¬active(P ),
t = ∅. So, we have disabled(A, S) for any S.
• Case (T-Par): In this case, P = P1 | P2 and A = A1 | A2 with Γ ⊲pl P1 : A1 and
Γ ⊲pl P2 : A2 . Note that P −→
6
implies P1 −→
6
and P2 −→.
6
¬active(P ) implies
¬active(P1 ) and ¬active(P2 ). So, by the induction hypothesis, we get disabled(A1 , S)
and disabled(A2 , S) for any S, which implies disabled(A, S).
• Case (T-Rep): In this case, P = ∗P1 and A = ∗A1 , with Γ ⊲pl P1 : A1 . ¬active(P )
and P −→
6
imply ¬active(P1 ) and P1 −→.
6
So, by the induction hypothesis, we get
disabled(A1 , S) for any S, which also implies disabled(A, S) as required.
• Case (T-If): This case cannot happen; by the condition (iv), P must be of the form
if true then P1 else P2
or if false then P1 else P2 , which contradicts with P −→.
6
• Case (T-New): In this case, P = (νx) P1 , A = (νx) A2 , and Γ, x : chanh(e
y:σ
e)A1 i ⊲
P1 : A2 . ¬active(P ) and P −→
6
imply ¬active(P1 ) and P1 −→.
6
So, by the induction
hypothesis, we get disabled(A2 , S) for any S. By the definition of disabled(·, S), we
get disabled(A, S).
• Case (T-Acc): This case cannot happen, since P must be of the form accξ (x).P1 ,
which contradicts with P −→.
6
• Case (T-NewR): Similar to the case for (T-New).
• Case (T-Sub): Γ⊲pl P : A must be derived from Γ⊲pl P : A′ for some A′ ≤ A. By the
induction hypothesis, for any S, we get disabled(A′ , S). By the condition A′ ≤ A,
we have disabled(A, S) for any S.
Appendix D. Computing a Basis of Behavioral Type
This section is an appendix for Section 4.3.1. Let A be a behavioral type of the form
(ν ye) B, where B does not contain any ν-prefix. Such A can be obtained by pushing all the
ν-prefixes out to the top-level, as described in Section 4.3.1. We show how to compute a
basis of A below.
42
N. KOBAYASHI, K. SUENAGA, AND L. WISCHIK
The constructor ·↑S can be eliminated by running the algorithm ElimUp∅,∅ (B, ∅) below.
ElimUpF,D (0, S) = 0
F,D
ElimUp
(α, S) =
A
if F (α, S) = A
F {(α,S)7→β},D
µβ.ElimUp
(D(α), S) if (α, S) 6∈ dom(F )
ElimUpF,D (l.A, S) = (l\S).ElimUpF,D (A, S)
ElimUpF,D (A1 | A2 , S) = ElimUpF,D (A1 , S) | ElimUpF,D (A2 , S)
ElimUpF,D (A1 ⊕ A2 , S) =
ElimUpF,D (A1 , S) ⊕ ElimUpF,D (A2 , S)
ElimUpF,D (∗A, S) = ∗ElimUpF,D (A, S)
ElimUpF,D (he
y /e
xiA, S) = ElimUpF,D (A, {z | [e
y /e
x]z ∈ S})
F,D
F {(α,S)7→α},D{α7→A}
ElimUp (µα.A, S) = µα.ElimUp
(A, S)
ElimUpF,D (A↑S1 , S) = ElimUpF,D (A, S ∪ S1 )
ElimUpF,D (A↓S1 , S) = ElimUpF,D (A, S)↓S1
Here, l\S is τ if target (l) ⊆ S and l otherwise. D keeps recursive definitions and F is a
cache for avoiding repeated computation. If A does not contain ν-prefixes, ElimUp∅,∅ (B, ∅)
always terminates since S can range over a finite set (which is the powerset of FV(B)).
The constructor ·↓S can be removed in the same manner.
We can further eliminate the renaming constructor he
y /e
xi by using the following algorithm.
ElimRenF,D (0, θ) = 0
F,D
ElimRen
(α, θ) =
A
if F (α, θ) = A
F {(α,θ)7→β},D
µβ.ElimRen
(D(α), θ) if (α, θ) 6∈ dom(F )
ElimRenF,D (l.A, θ) = θl.ElimRenF,D (A, θ)
ElimRenF,D (A1 | A2 , θ) = ElimRenF,D (A1 , θ) | ElimRenF,D (A2 , θ)
ElimRenF,D (A1 ⊕ A2 , θ) =
ElimRenF,D (A1 , θ) ⊕ ElimRenF,D (A2 , θ)
ElimRenF,D (∗A, θ) = ∗ElimRenF,D (A, θ)
ElimRenF,D (he
y /e
xiA, θ) = ElimRenF,D (A, θ ◦ [e
y /e
x])
F,D
F {(α,θ)7→α},D{α7→A}
ElimRen (µα.A, θ) = µα.ElimRen
(A, θ)
By applying the above algorithms to A = (ν ye) B, we obtain an equivalent type A′ =
(ν ye) B ′ , where B ′ does not contain any ν-prefixes, ·↓S , ·↑S , or he
y /e
xi. So, only elements
of Atoms(B ′ ) defined below (modulo folding/unfolding of recursive types) can appear in
transitions of B. So, ({e
y }, Atoms(B ′ )) forms a basis of A.
Definition D.1. Let A be a behavioral type that does not contain any ν-prefix, ·↓S , ·↑S ,
or he
y /e
xi. The set of atoms Atoms(A) is the least set that satisfies the following conditions.
Atoms(l.A) ⊇ {l.A} ∪ Atoms(A)
Atoms(A1 | A2 ) ⊇ Atoms(A1 ) ∪ Atoms(A2 )
Atoms(A1 ⊕ A2 ) ⊇ {A1 ⊕ A2 } ∪ Atoms(A1 ) ∪ Atoms(A2 )
Atoms(∗A) ⊇ {∗A} ∪ Atoms(A)
Atoms(µα.A) ⊇ {µα.A} ∪ Atoms([µα.A/α]A)
| 6cs.PL
|
F R ÉCHET A NALYSIS OF VARIANCE FOR R ANDOM O BJECTS
Paromita Dubey1 and Hans-Georg Müller2
Department of Statistics, University of California, Davis
Davis, CA 95616 USA
arXiv:1710.02761v2 [math.ST] 19 Oct 2017
October 2017
ABSTRACT
Fréchet mean and variance provide a way of obtaining mean and variance for general metric
space valued random variables and can be used for statistical analysis of data objects that lie in
abstract spaces devoid of algebraic structure and operations. Examples of such spaces include
covariance matrices, graph Laplacians of networks and univariate probability distribution functions. We derive a central limit theorem for Fréchet variance under mild regularity conditions,
utilizing empirical process theory, and also provide a consistent estimator of the asymptotic
variance. These results lead to a test to compare k populations based on Fréchet variance for
general metric space valued data objects, with emphasis on comparing means and variances.
We examine the finite sample performance of this inference procedure through simulation studies for several special cases that include probability distributions and graph Laplacians, which
leads to tests to compare populations of networks. The proposed methodology has good finite
sample performance in simulations for different kinds of random objects. We illustrate the proposed methods with data on mortality profiles of various countries and resting state Functional
Magnetic Resonance Imaging data.
KEY WORDS: Functional Data Analysis; Fréchet mean; Fréchet variance; Central Limit Theorem; Two sample test; Samples of probability distributions; Wasserstein metric; Samples of
networks; Graph Laplacians; fMRI
1 Corresponding
2 Research
author; email: [email protected]
supported by National Science Foundation grants DMS-1407852 and DMS-172864
1.
INTRODUCTION
With an increasing abundance of complex non-Euclidean data in multiple disciplines there is a
need for statisticians to develop techniques suitable for the analysis of such data. Settings where
data objects take values in a metric space that is devoid of vector space structure are common,
and occur for example when only pairwise distances between the observed data objects are
available. It is then natural to assume that observed samples of random objects are drawn from
a distribution in a metric space. The standard problem of K-sample testing, with the most
important case concerning the two sample case with K = 2, is of basic interest in statistics
and becomes more challenging when data objects lie in general metric spaces. For the case
of Euclidean data this is the classical K sample comparison problem. For Gaussian data it
corresponds to the classical analysis of variance (ANOVA), if comparing means is of primary
interest, which uses comparisons of summary variation measures between and within groups.
For general metric space valued random variables, Fréchet (1948) provided a direct generalization of the mean which implies a corresponding generalization of variance that may be used
to quantify the spread of the distribution of metric space valued random variables or random
objects around their Fréchet mean. The Fréchet mean resides in the object space and thus is
not amenable to algebraic operations, which implies that a central limit theorem cannot be directly applied to obtain limit distributions for Fréchet means. In contrast, the Fréchet variance
is always a scalar, which makes it more tractable. A key result of this paper is a central limit
theorem for the empirical Fréchet variance of data objects in general metric spaces under weak
assumptions. While this result is of interest in itself, we demonstrate how it can be applied
to derive a k-sample test based on groupwise Fréchet variances for testing the null hypothesis
of equal population distributions for random objects, with emphasis on testing the equality of
Fréchet means or of the Fréchet variances. In recent years the study of nonparametric tests for
the equality of two distributions for Euclidean data has expanded to cover non-Euclidean data,
which are increasingly encountered in frameworks that feature large and complex data. Tests
that are suitable for random objects in metric spaces are typically based on pairwise distances.
1
Major approaches have been based on nearest neighbors, graphs, energy statistics and kernels.
Nearest neighbor based tests (Henze, 1988; Henze and Penrose, 1999; Schilling, 1986)
count the number of K-nearest neighbor comparisons for which the observations and their
neighbors belong to the same sample. The choice of the tuning parameter K impacts the resulting inference. For graph-based two sample tests (Friedman and Rafsky, 1979), pairwise
distances of the pooled sample are used to form an edge-weighted graph and its minimum spanning tree. The test statistic is the number of subtrees formed by the removal of the edges for
which the defining nodes belong to different samples. Recently, modified versions of this test
were proposed (Chen and Friedman, 2017; Chen et al., 2017) also based on a similarity graph
of the pooled observations and a weighted edge count test in order to deal with the problem of
unequal sample sizes. Since it is more difficult to form within sample edges for samples with
smaller sizes, weights proportional to the reciprocal of sample sizes are assigned to the within
sample edges. A related approach is minimum distance non-bipartite pairing (Rosenbaum,
2005), where the dataset is split into disjoint pairs by minimizing the total distance within the
pairs and the test statistic is the number of cross matches.
The class of statistics popularly known as energy statistics, initially proposed for the case
of a multivariate two sample test (Baringhaus and Franz, 2004; Székely and Rizzo, 2004) and
based on the fact that for independent d-dimensional random vectors X1 , X2 ,Y1 and Y2 where X1
and X2 have the same distribution F and Y1 and Y2 have same distribution G, both with finite
expectations, the inequality
1
1
EkX1 −Y1 k − EkX1 − X2 k − EkY1 −Y2 k ≥ 0
2
2
holds, with equality if and only if F = G, where F and G are the two distributions to be compared. This framework was extended to the case of metric space valued random variables in
Lyons (2013) for metric spaces of strong negative type, leading to the distinction of measures
that have finite first moments with respect to the underlying metric. An extension of the energy
2
statistics approach for spaces admitting a manifold stratification (Patrangenaru and Ellingson,
2015) has been explored recently (Guo and Patrangenaru, 2017). A recent review is Székely
and Rizzo (2017).
Kernel based two sample testing (Gretton et al., 2012) aims at finding the the maximum
mean discrepancy (MMD) which is the largest difference in expectations over functions in the
unit ball of a reproducing kernel Hilbert space (RKHS) and corresponds to distances between
embeddings of distributions in the RKHS. Distribution free tests can be obtained based on large
deviation bounds of empirical estimates and also the asymptotic distribution of MMD, and the
energy statistics approach can be linked to the kernel method for two sample testing, as the
distance based energy statistic computed using a semi-metric of negative type is equivalent to
the MMD statistic obtained from a special type of kernel induced by the semi-metric (Sejdinovic et al., 2013). For every positive definite kernel, the MMD statistic corresponds to the energy statistic obtained from a semi-metric derived from the kernel used to compute the MMD.
Connections between kernel methods, energy statistics and Wasserstein distance between the
distributions to be compared are explored in Ramdas et al. (2017). Commonly used kernels in
the kernel framework, like the Gaussian kernel, might be sensitive to the choice of the tuning
parameter needed to scale the kernel and determining a good scaling factor can be challenging
when the goal is inference for complex data. We confirm in our simulations that this can be
problematic when adopting the kernel method to obtain inference.
Empirically, these tests have good power performance for either location type alternatives
or scale type alternatives, but usually not for both simultaneously. A major challenge for some
of these tests is the choice of the required tuning parameters, which often has a major impact on
the resulting inference. The challenges associated with existing inference procedures motivate
our proposed test, which is simple and is based on a test statistic that mimics the statistics on
which t tests and classical ANOVA are based, replacing between and within sums of squares
with the corresponding Fréchet variances for separate and combined samples, and is easy to
compute for k- sample comparisons.
3
Fréchet mean based testing and corresponding large sample theory including laws of large
numbers and central limit theorems for empirical Fréchet means have been explored previously for data objects that lie in special metric spaces, such as smooth Riemannian manifolds
(Bhattacharya and Patrangenaru, 2003, 2005; Bhattacharya and Bhattacharya, 2012) and topologically stratified spaces under certain restrictions, like phylogenetic trees (Kendall and Le,
2011; Barden et al., 2013; Bhattacharya and Lin, 2017). Virtually all of these results depend on
local linear tangent or similar approximations that are specific to these manifold spaces but do
not apply for random objects in more general metric spaces as they require local Euclidean approximation. The central limit theorem for Fréchet means was recently applied to the space of
graph Laplacians (Ginestet et al., 2017), which are of interest to obtain inference for networks.
This required choosing a high dimension for the approximating space, thus leading to problems
with small sample high dimensional data and the ensuing complications for inference.
Since Fréchet mean based testing exploits the local Euclidean approximation property of
the underlying space, it does not extend to data in general metric spaces such as the space of
univariate probability density functions, which will serve as one of our examples to illustrate the
proposed approach. Another drawback of these methods is that they lose power when the data
is high dimensional as large covariance matrices and their inverses need to be estimated. For
a network with as few as 20 nodes, the Fréchet mean of the graph Laplacians has a covariance
matrix of dimension 36100. To address this issue, strong assumptions, such as sparsity, have
been invoked for the estimation of these large covariance matrices and their inverses in order
to implement the corresponding tests. This provides additional motivation for our approach,
which is very simple and where dimension of the data enters only indirectly through properties
of the metric. To obtain asymptotic properties, we draw on empirical process theory, in a similar
spirit as a recent study of regression relationships between general metric space valued random
objects as responses and real valued predictors (Petersen and Müller, 2017).
Our goal in this paper is to develop a simple and straightforward extension of ANOVA,
which is one of the very basic tools for inference in statistics, to the case of metric space valued
4
random objects. Our starting point is a totally bounded metric space Ω equipped with a metric
d. We show that consistency of the sample Fréchet mean can be derived by using results of
Petersen and Müller (2017) concerning Fréchet regression estimators under mild assumptions
on Ω. We derive a central limit theorem for Fréchet variance under mild assumptions and
provide a consistent estimator of its asymptotic variance. Our method is applicable to a wide
class of objects including correlation matrices, univariate probability distributions, manifolds
and also the space of graph Laplacians. Making use of the central limit theorem, we derive a new
k-sample test for random objects and study the asymptotic distribution of the test statistic under
the null hypothesis of equality of the population distributions, as well as its power function.
It is customary to test for heteroscedasticity of the population groups prior to the F-test of
classical ANOVA to evaluate the assumption of equal variances across the populations that are
compared. One popular test for this purpose was proposed by Levene (1960), where the test
statistic is of the form of the usual ANOVA F-test, but applied to pseudo-observations which
could in principle be any monotonic function of the absolute deviations of the observations
from their group ‘centers’. Our proposed test unifies Levene’s test and the classical ANOVA for
testing inequality of population means for data objects in general metric spaces and therefore
aims at both location and scale type alternatives instead of only location alternatives which is the
objective of classical ANOVA. Our test statistic is a sum of two components. One component is
proportional to the squared difference of the pooled sample Fréchet variance and the weighted
average of the groupwise Fréchet variances, with weights proportional to the sample sizes of
the groups. For the special case of Euclidean data, this part of our statistic is proportional to
the squared F-ratio as in the usual ANOVA and is useful to detect differences in the Fréchet
means of the populations. A key auxiliary result is that this statistic converges to zero at rate
oP ( 1n ) under the null hypothesis of equality of Fréchet means of population distributions and
is bounded away from zero otherwise. The other component of our test statistic accounts for
differences in the Fréchet variances of the population groups and under the Euclidean setting
simplifies to a generalization of the Levene’s test applied to squared absolute deviations of
5
the observations from their group Fréchet means. When the assumptions of the central limit
theorem hold, the asymptotic distribution of this component of our test statistic is χ2(k−1) where
k is the number of populations to be compared.
The paper is organized as follows: The basic set up is defined in Section 2 and the theory
regarding the asymptotic behavior of Fréchet variance is provided in Section 3. The k-sample
test is introduced in Section 4, followed by empirical studies to study its power performance
in Section 5. Our data applications in Section 6 include samples of univariate probability distributions for which we employ the L2 Wasserstein metric and samples of graph Laplacians of
networks, for which we use the Frobenius metric. In Section 6.1 we illustrate the proposed
tests with human mortality records of 31 countries over the time period 1960-2009 to analyze
the evolution of age at death distributions of the countries over the years. In Section 6.2 we
analyze a resting state fMRI dataset where we compare the probability distribution of positive
correlations between fMRI signals in the posterior cingulate area of the brain of Alzheimer’s
Disease patients to those of similarly aged normal subjects. Finally, in Section 6.3 we analyze
brain connectivity networks of patients with dementia that are derived from fMRI signals of certain brain regions (Buckner et al., 2009) and compare the connectivity networks of demented
patients for three different age groups.
2.
PRELIMINARIES
The Fréchet mean is a generalization of centroids to metric spaces and for the special case of
Euclidean data it includes the arithmetic mean, median and geometric mean under different
choices of distance functions. The Fréchet variance is the corresponding generalized measure
of dispersion around the Fréchet mean. More formally, in all of the following, (Ω, d, P) is a
totally bounded metric space with metric d and probability measure P. Random objects in the
following are random variables Y that take values in Ω. The (population) Fréchet mean of Y is
µF = argmin E(d 2 (ω,Y )),
ω∈Ω
6
(1)
while for a random sample Y1 ,Y2 , . . . ,Yn of i.i.d. random variables with the same distribution as
Y , the corresponding sample Fréchet mean is
1 n 2
µ̂F = argmin ∑ d (ω,Yi ).
ω∈Ω n i=1
(2)
The sample Fréchet mean is an M-estimator as it is obtained by minimizing a sum of functions of the data objects. The population Fréchet variance quantifies the spread of the random
variable Y around its Fréchet mean µF ,
VF = E(d 2 (µF ,Y )),
(3)
with corresponding sample based estimator
V̂F =
1 n 2
∑ d (µ̂F ,Yi).
n i=1
(4)
Note that µF , µ̂F ∈ Ω, while VF , V̂F ∈ R .
The asymptotic consistency of the sample Fréchet mean µ̂F follows from results in Petersen
and Müller (2017), under the following assumption:
(P0) The objects µ̂F and µF exist and are unique, and for any ε > 0, infd(ω,µF )>ε E(d 2 (ω,Y )) >
E(d 2 (µF ,Y )).
Assumption (P0) is instrumental to establish the weak convergence of the empirical process Hn
to the population process H, where
1 n 2
Hn (ω) = ∑ d (ω,Yi ),
n i=1
H(ω) = E(d 2 (ω,Y )),
which in turn implies the consistency of µ̂F ,
d(µ̂F , µF ) = oP (1).
7
(5)
Consistency of µ̂F then implies the consistency of V̂F since
|V̂F −VF | ≤ 2 diam(Ω) d(µ̂F , µF ),
(6)
where diam(Ω) = sup{d(ω1 , ω2 ) : ω1 , ω2 ∈ Ω} is finite, since Ω is totally bounded. For the
central limit theorem (CLT) to hold for empirical Fréchet variance we need an assumption on
the entropy integral of the space Ω (Wellner and van der Vaart, 1996),
J(δ) =
Z 1p
0
1 + log(N(εδ/2, Bδ (µF ), d)) dε,
(7)
where Bδ (µF ) is the δ-ball in the metric d, centered at µF and N(εδ/2, Bδ (µF ), d) is the covering
number for Bδ (µF ) using open balls of radius εδ/2. Specifically, for our CLT we assume that
(P1) δ J(δ) → 0 as δ → 0.
For our results on the power of the proposed test in Section 4, we need an additional assumption
on the entropy integral of the whole space Ω,
(P2) The entropy integral of Ω is finite,
R1p
0
1 + log N(ε, Ω, d)dε < ∞.
Random objects that satisfy assumptions (P0)-(P2) include the space of univariate probability distributions on R with finite second moments equipped with the Wasserstein metric dW
and the spaces of correlation matrices and graph Laplacians of fixed dimensions equipped with
the Frobenius metric dF . For two univariate distributions F and G with finite variances, the
L2 -Wasserstein distance, also known as earth movers distance and closely related to optimal
transport (Villani, 2003),
2
dW
(F, G) =
Z 1
(F −1 (t) − G−1 (t))2 dt,
(8)
0
where F −1 and G−1 are the quantile functions corresponding to F and G, respectively. For two
8
matrices A and B of the same dimension, we consider the Frobenius metric,
dF2 (A, B) = trace((A − B)0 (A − B)).
(9)
To better characterize the space of graph Laplacians, we denote a weighted undirected graph
by G = (V, E), where V is the set of its vertices and E the set of its edges. Given an adjacency
matrix W , where wi j = w ji ≥ 0 and equality with zero holds if and only if {i, j} ∈
/ E, the graph
Laplacian is defined as L = D −W , where D is the diagonal matrix of the degrees of the vertices,
i.e. d j j = ∑i wi j . Under the assumption that the graphs are simple (i.e., there are no self loops
or multi edges), there is a one to one correspondence between the space of graphs and the graph
Laplacians and therefore the graph Laplacians can be used to characterize the space of networks
(Ginestet et al., 2017). The following results imply that the spaces described above provide
examples of spaces that satisfy assumptions (P0) and (P1). Our inference methods therefore
will apply to compare samples of univariate distributions, correlation matrices and networks.
Proposition 1. The space (Ω, dW ) satisfies assumptions (P0)-(P2) when the set Ω consists
of univariate probability distributions on R with finite second moments and dW is the L2 Wasserstein metric.
Proposition 2. The space (Ω, dF ) satisfies assumptions (P0)-(P2) when the set Ω consists of
graph Laplacians of connected, undirected and simple graphs of a fixed dimension r or correlation matrices of a fixed dimension r and dF is the Frobenius metric.
All proofs are in the Supplementary Materials.
3.
CENTRAL LIMIT THEOREM FOR FRÉCHET VARIANCE
The following Proposition lays the foundations for proving the Central Limit Theorem for the
empirical Fréchet variance V̂F .
9
Proposition 3. Suppose assumptions (P0)-(P1) hold. Then
1
1 n 2
{d (µ̂F ,Yi ) − d 2 (µF ,Yi )} = oP ( √ ).
∑
n i=1
n
Proposition 3 makes it possible to deal with the sum of dependent random variables
∑ni=1 d 2 (µ̂F ,Yi ), by replacing it with the sum of i.i.d. random variables ∑ni=1 d 2 (µF ,Yi ), a crucial
step in the derivation of the Central Limit Theorem for V̂F . Since Ω is totally bounded, the
population Fréchet variance Var(d 2 (µF ,Y )) is always finite. The Central Limit Theorem for
Fréchet variance is as follows.
Theorem 1. Under the assumptions of Proposition 3,
√
D
n(V̂F −VF ) −→ N(0, σ2F ),
where σ2F = Var(d 2 (µF ,Y )).
An intuitive sample based estimator for σ2F is
1 n
σ̂2F = ∑ d 4 (µ̂F ,Yi ) −
n i=1
and this estimator is
!2
1 n 2
∑ d (µ̂F ,Yi) ,
n i=1
(10)
√
n-consistent as the following result shows.
Proposition 4. Under the assumptions of Proposition 3,
D
√
b2F − σ2F −→ N(0, A),
n σ
where A = a0 Da with a0 = 1, −2E d 2 (µF ,Y ) and
Cov d 4 (µF ,Y ), d 2 (µF ,Y )
D=
4
2
2
Cov d (µF ,Y ), d (µF ,Y )
Var d (µF ,Y ) .
Var d 4 (µF ,Y )
10
(11)
b2F is a
Therefore σ
√
n-consistent estimator of σ2F .
Proposition 4 is a variant of Proposition 3 and follows from a simple application of the delta
method. The quantity A is finite by the boundedness of Ω. Combining Theorem 1, Proposition
3 and Slutsky’s Theorem leads to
1 √
D
n(V̂F −VF ) −→ N(0, 1).
σ̂F
(12)
A simple application of the delta method gives the asymptotic distribution of the Fréchet standard deviation which is the square root of Fréchet variance,
√
σ2
D
1/2
n(V̂F −VF 1/2 ) −→ N(0, F )
4VF
(13)
and since both σ̂F and V̂F are consistent estimators,
1/2
2V̂F √
D
1/2
n(V̂F −VF 1/2 ) −→ N(0, 1).
σ̂F
(14)
One can use (12) and (14) to construct asymptotic confidence intervals for Fréchet variance and standard deviation, which depend on the quality of the large sample approximations.
The bootstrap provides an alternative that often has better finite sample properties under weak
assumptions (Bickel and Freedman, 1981; Beran, 2003). Under fairly general assumptions,
resampling methods like bootstrapping and permutation tests work whenever a central limit
theorem holds (Janssen and Pauls, 2003). A basic criterion for bootstrap confidence sets to
have correct coverage probability asymptotically is convergence of the bootstrap distribution
√
of the root, in our case n(V̂F −V )/σ̂F . Then Monte Carlo approximations of the bootstrap
distribution of the root provide approximate quantiles for the construction of confidence sets.
For the empirical measure Pn generated by Y1 ,Y2 , . . . ,Yn which puts mass
11
1
n
to each of
Y1 , . . . ,Yn , where P is the underlying measure and any measurable set A we have
Pn (A) =
1 n
P
1A (Yi ) −→ P(A),
∑
n i=1
(15)
where 1A (·) is the indicator function for the set A, by the weak law of large numbers. Given
a sample Y1∗ ,Y2∗ , . . . ,Ym∗ of size m drawn with replacement from Y1 ,Y2 , . . . ,Yn , the bootstrap ap√
proximation of the root is R∗m,n = m(V̂m∗ − V̂F )/σ̂∗m , where V̂m∗ and σ̂∗m are the sample based
estimators of the Fréchet variance and its asymptotic variance, obtained from the bootstrap
sample Y1∗ ,Y2∗ , . . . ,Ym∗ . If the Fréchet mean µ̂∗m of the bootstrap sample Y1∗ ,Y2∗ , . . . ,Ym∗ exists and
is unique almost surely conditionally on Y1 ,Y2 , . . . ,Yn , then by applying the central limit theoD
rem in Theorem 1 conditionally on Y1 ,Y2 , . . . ,Yn , R∗m,n −→ N(0, 1) as m → ∞. Since N(0, 1)
has a continuous distribution function on the real line we have for each ε > 0, as m → ∞,
Pn sup |HR∗m,n (x) − Φ(x)| > ε → 0
(16)
x
where HR∗m,n (·) is the distribution function of R∗m,n conditional on Y1 ,Y2 , . . . ,Yn and Φ(·) is the
standard normal distribution function. From (15) for each ε > 0 we have as both m, n → ∞
P sup |HR∗m,n (x) − Φ(x)| > ε → 0,
x
which establishes asymptotic consistency of the bootstrap distribution. Therefore non-parametric
bootstrapping is a viable option for the construction of confidence intervals for Fréchet variance.
4.
COMPARING POPULATIONS OF RANDOM OBJECTS
Assume we have a sample of n Ω-valued random data objects Y1 ,Y2 , . . . ,Yn that belong to k
different groups G1 , G2 , . . . , Gk , each of size n j , j = 1, . . . , k, such that ∑kj=1 n j = n. We wish to
test the null hypothesis that the population distributions of the k groups are identical versus the
alternative that at least one of the groups has a different population distribution compared to the
12
others. Consider sample Fréchet means µ̂ j in Ω, which are random objects computed just from
the data falling into group j and the corresponding real-valued sample Fréchet variances V̂ j and
variance estimates (10) σ̂2j , j = 1, . . . , k,
µ̂ j = argmin
ω∈Ω
σ̂2j =
1
∑ d 2(ω,Yi),
n j i∈G
j
V̂ j =
1
∑ d 2(µ̂ j ,Yi),
n j i∈G
j
!2
1
∑ d 2(µ̂ j ,Yi)
n j i∈G
j
1
∑ d 4(µ̂ j ,Yi) −
n j i∈G
j
,
as well as the pooled sample Fréchet mean µ̂ p and the corresponding pooled sample Fréchet V̂p ,
µ̂ p = argmin
ω∈Ω
1
n
k
∑∑
d 2 (ω,Yi ),
V̂p =
j=1 i∈G j
1
n
k
∑ ∑ d 2(µ̂ p,Yi),
(17)
j=1 i∈G j
as well as weights
λ j,n =
nj
,
n
k
j = 1, . . . , k,
such that
∑ λ j,n = 1.
j=1
We will base our inference procedures on the auxiliary statistics
k
Fn = V̂p − ∑ λ j,nV̂ j ,
(18)
j=1
and
λ j,n λl,n
(V̂ j − V̂l )2 ,
2
2
j<l σ̂ j σ̂l
Un = ∑
(19)
where Fn is almost surely non-negative and is equal to the numerator of the F-ratio in classical
Euclidean ANOVA. Specifically, Fn corresponds to the weighted variance of the group means,
with weights proportional to the group sizes, and correspond to the between group variance in
the classical ANOVA setting. Hence Fn can be regarded as a generalization of the F-ratio in
classical ANOVA to the more general setting of metric space valued data. Under the Euclidean
13
setting, simple algebra shows that in this special case Fn is also proportional to the weighted
average of the squared pairwise distances between the group Fréchet means. Analogous to
ANOVA, the numerator of Fn is expected to be small under the null hypothesis of equality of
the population distributions, which is indeed the case as the following Proposition demonstrates.
Proposition 5. Suppose µ̂ p and µ̂ j exist and are unique almost surely for all j = 1, . . . , k. Let
0 < λ j,n < 1 for all j = 1, . . . , k and λ j,n → λ j as n → ∞, where λ j is such that 0 < λ j < 1 for
each j = 1, . . . , k, with ∑kk=1 λ j = 1. Then under the null hypothesis of equality of population
distributions and under assumptions (P0) and (P1) for each of the groups, as n → ∞,
√
nFn = o p (1).
(20)
Inference in classical ANOVA requires Gaussianity and equality of the population variances
and hence targets only differences in the group means to capture differences in the population
distributions. These assumptions are obviously too restrictive. We employ the statistics Un in
(19) to account for differences among the population variances, where in the Euclidean case, Un
turns out to be a slightly modified version of the traditional Levene’s test, substituting squared
distances of the observations from their group Fréchet means instead of just distances. The
following result provides the asymptotic distribution of Un under the null hypothesis.
Proposition 6. Under the assumptions of Proposition 5, we have under the null hypothesis, as
n → ∞,
nUn
λ j,n
k
∑ j=1 σ̂2
j
D
−→ χ2(k−1) .
(21)
We construct our test statistic Tn by combining Fn and Un in such a way that the distribution
of Tn under the null hypothesis is the same as that given in Proposition 6 while gaining power
against alternatives by ensuring that the asymptotic mean of Tn diverges when departing from
14
the null hypothesis,
Tn =
nUn
∑kj=1
λ j,n
σ̂2j
+
nFn2
.
∑kj=1 λ2j,n σ̂2j
(22)
For constructing Tn , we scale Fn by the estimated standard deviation of ∑kj=1 λ j,nV̂ j , which is
q
equal to ∑kj=1 λ2j,n σ̂2j , so that Fn is suitably scaled with respect to the variability of ∑kj=1 λ j,nV̂ j .
The second term on the r.h.s. of (22) is scaled such that both terms in Tn are of the same order
in n. Under the null hypothesis, consistency of σ̂2j for j = 1, 2, · · · , k and Proposition 5 imply
nFn2
k
∑ j=1 λ2j,n σ̂2j
= oP (1).
Theorem 2. Under the null hypothesis and the assumptions of Proposition 5,
D
Tn −→ χ2(k−1) .
(23)
For a level α test, we accordingly reject the null hypothesis of equality of population distributions if the test statistic Tn turns out to be bigger than χ2k−1,α , which is the (1 − α)th quantile
of the χ2(k−1) distribution, i.e., the rejection region that defines the test is
Rn,α = {Tn > χ2k−1,α }.
(24)
To study the consistency of the proposed test (24) we consider contiguous alternatives that capture departures from the null hypothesis of equal population distributions in terms of differences
in their Fréchet means and variances. We begin by defining the following population quantities,
k
k
µ p = argmin ∑ λ j E j (d 2 (ω,Y j )),
ω∈Ω
Vp =
j=1
∑ λ j E j (d 2(µ p,Y j ))
(25)
j=1
and
k
F = Vp − ∑ λ j V j ,
U=
j=1
λ j λl
∑ σ2σ2 (V j −Vl )2,
j<l
(26)
j l
where E j (·) denotes expectation under the probability distribution for the jth population and
15
λ j,n is as defined in Proposition 5, and the Y j are random objects distributed according to the
jth population distribution. In the Euclidean setting, µ p and Vp are analogous to the pooled
population Fréchet mean and pooled population Fréchet variance, respectively, and in the general case can be interpreted as generalizations of these quantities. While in the Euclidean case
F is proportional to the weighted sum of the differences between the groupwise population
Fréchet means and the pooled population Fréchet mean, by simple algebra it can be seen that
this property still holds in the general metric case.
Proposition 7 below states that under mild assumptions on the existence and uniqueness of
the pooled and the groupwise Fréchet means, the statistics µ̂ p and Fn are consistent estimators
of the population quantities µ p and F. By our assumptions, F is zero only under the equality of
the population Fréchet means and positive otherwise. The population quantity U is proportional
to the weighted average of the pairwise differences between the groupwise Fréchet variances,
which is nonnegative and is zero only if the population groupwise Fréchet variances are all
equal.
Proposition 7. Suppose µ̂ p , µ̂ j , µ p and µ j exist and are unique, the sample based estimators
almost surely for all j = 1, . . . , k. Assume for any ε > 0 , infd(ω,µ p )>ε ∑kj=1 λ j E j (d 2 (ω,Y j )) >
∑kj=1 λ j E j (d 2 (µ p ,Y j )) and also infd(ω,µ j )>ε E j (d 2 (ω,Y j )) > E j (d 2 (µ j ,Y j )) for all j = 1, . . . , k.
Let 0 < λ j,n < 1 for all j = 1, . . . , k and λ j,n → λ j as n → ∞, where λ j is such that 0 < λ j < 1
for each j = 1, . . . , k, with ∑kk=1 λ j = 1, as defined in Proposition 5. Then, as n → ∞,
P
µ̂ p −→ µ p
and
P
Fn −→ F.
(27)
The population quantity F is nonnegative and is zero iff the population Fréchet means µ j are all
equal.
To study the power performance of the proposed test (24), we consider sequences of alternatives Hn where Hn = {(U, F) : U ≥ an or F ≥ bn } for non negative sequences {an } or {bn }
with either an or bn strictly greater than 0. The case where either an → 0 or bn → 0, as n → ∞
16
reflects contiguous alternatives. Of interest is the asymptotic behavior of the power function
βHn , where
βHn =
inf
(U,F)∈Hn
P (Rn,α ) .
(28)
For the following result we require assumption (P2). The examples considered here satisfy this
assumptions, as shown in Propositions 1 and 2. The following result provides sufficient conditions for the consistency of the proposed test (24) under this family of contiguous alternatives,
i.e., where its asymptotic power is 1 under these alternatives, for any choice of the level α.
Theorem 3. Under the assumptions of Proposition 7 and assumption (P3), for sequences of
contiguous alternatives {Hn } for which either F ≥ an or U ≥ bn , where an → 0 and bn → 0 as
n → ∞, for all α > 0 the power function (28) satisfies
√
(A) If nan → ∞, then βHn → 1.
(B) If nbn → ∞ , then βHn → 1.
Asymptotic tests may not work very well for common situations where the group sample
sizes are modest. By arguments similar to those provided at the end of Section 3, resampling
methods like bootstrapping and permutation tests using the proposed test statistic Tn can be
adopted to obtain more accurate level α tests. In our empirical experiments, which are presented
in the next section, we found that when sample sizes of the groups are large the asymptotic test
is stable and accurate, but for small sample sizes, using bootstrap or permutation based critical
values for the test statistic Tn instead of the asymptotic critical values leads to more accurate
inference.
5.
SIMULATION STUDIES
In order to gauge the performance of the proposed test (24), we performed simulation experiments under various settings. The random objects we consider include samples of univariate
distributions equipped with the L2 -Wasserstein metric, samples of graph Laplacians of scale
17
free networks from the Barabási-Albert model (Barabási and Albert, 1999) with the Frobenius
metric and samples of multivariate data with the usual Euclidean metric. In each case we considered two groups of equal size n1 = n2 = 100 and constructed the empirical power functions
of the proposed test against departures from the null hypothesis of equality of population distributions of the two groups at level 0.05. The empirical power was computed by finding the
proportion of rejections for 1000 Monte Carlo runs. For comparing the performance of the
proposed test against existing tests we used the bootstrap version of the proposed test, where
the critical value for the test was obtained from the bootstrap distribution of Tn in 1000 Monte
Carlo simulations. The performance of the bootstrap version was found to have the correct level
for all sample sizes while the asymptotic version of the test did not always produce the correct
level of the test for very small sample sizes. We also investigated the finite sample power of
the asymptotic test for increasing sample sizes by comparing power functions for group sizes
n1 = n2 = 100, 250, 450.
In the simulations we explored not only location differences but also differences in shape
and scale of the population distributions. We compared the proposed test (24) with the graph
based test (Chen and Friedman, 2017), the energy test based on pairwise distances (Székely
and Rizzo, 2004) and a kernel based test (Gretton et al., 2012). For the graph based test, we
constructed the similarity graph of the pooled observations of the two groups by constructing
a 5−MST (minimal spanning tree) graph from the pooled pairwise distance matrix, following
the suggestion in Chen and Friedman (2017). Here a k−MST is the union of the 1st , . . . , kth
MSTs, where a kth MST is a spanning tree connecting all observations that minimizes the sum
of distances across edges subject to the constraint that this spanning tree does not contain any
edge in the 1st , . . . , (k − 1)th MST. For computing the statistic of the energy test of Székely and
Rizzo (2004), we used the pairwise distance matrix obtained from the specified metric in the
space of random objects. For the kernel based method, we chose a Gaussian kernel with the
kernel width as the median of the pairwise distances, as suggested in Gretton et al. (2012). In the
multivariate Euclidean setting, for the two sample case we additionally compared the proposed
18
test with Hotelling’s T 2 test.
0.8
0.6
power
0.0
0.2
0.4
0.6
0.4
0.0
0.2
power
0.8
1.0
Power Functions
1.0
Power Functions
−1.0
−0.5
0.0
0.5
1.0
0.5
1.0
δ
1.5
2.0
2.5
3.0
r
Figure 1: Empirical power as function of δ for N(µ, 1) probability distributions with µ from N(0, 0.5) for
group G1 and N(δ, 0.5) for group G2 (left), and empirical power as function of r for N(µ, 1) probability
distributions with µ from N(0, 0.2) for G1 and N(0, 0.2r) for G2 (right). The solid red curve corresponds
to the bootstrapped version of the proposed test with test statistic (22), the dashed blue curve to the graph
based test of Chen and Friedman (2017) , the dot-dashed black curve to the energy test of Székely and
Rizzo (2004) and the dotted green curve corresponds to the kernel test of Gretton et al. (2012). The level
of the tests is α = 0.05 and is indicated by the line parallel to the x-axis. Sample sizes of the groups are
fixed at n1 = n2 = 100.
0.8
0.6
power
0.0
0.2
0.4
0.6
0.4
0.0
0.2
power
0.8
1.0
Power Functions
1.0
Power Functions
−1.0
−0.5
0.0
0.5
1.0
0.5
δ
1.0
1.5
2.0
2.5
3.0
r
Figure 2: Empirical power as function of δ for N(µ, 1) probability distributions with µ from N(0, 0.5) for
group G1 and N(δ, 0.5) for group G2 (left), and empirical power as function of r for N(µ, 1) probability
distributions with µ from N(0, 0.2) for G1 and N(0, 0.2r) for G2 (right), for the proposed test at different
sample sizes. The tests are at level α = 0.05, indicated by the line parallel to the x-axis. The solid curve
corresponds to sample sizes n1 = n2 = 450, the dashed curve to n1 = n2 = 250 and the dotted curve to
n1 = n2 = 100.
The first type of random objects we study are random samples of univariate probability
distributions. Each datum is a N(µ, 1) distribution where µ is random. As distance between two
19
probability distributions we choose the L2 -Wasserstein metric. In the first scenario, for group
G1 , we generate µ to be distributed as N(0, 0.5) and for group G2 as N(δ, 0.5) and compute
the empirical power function of the tests for −1 ≤ δ ≤ 1. In the second scenario µ is drawn
randomly from N(0, 0.2) for group G1 and from N(0, 0.2r) for group G2 and empirical power
is evaluated for 0.125 ≤ r ≤ 3. The first scenario emphasizes location differences between the
populations and the second emphasizes scale differences. The results are presented in Figure 1.
We find that in the first scenario of mean differences, the proposed test and the graph based test
perform similarly. Both of them are outperformed by the energy test but perform better than the
kernel based test. In the second scenario of scale differences the proposed test outperforms all
other tests. Figure 2 indicates that the proposed test is consistent for large sample sizes in both
scenarios, and Figure 3 indicates that for insufficient sample sizes, the bootstrap version of the
proposed test has more stable rejection regions and overall is more reliable than the asymptotic
version of the test (24). As sample sizes increase, the asymptotic test becomes more reliable
and yields results that are similar to the bootstrap test.
−1.0
−0.5
0.0
δ
0.5
1.0
0.8
0.0
0.2
0.4
power
0.6
0.8
0.6
power
0.4
0.2
0.0
0.0
0.2
0.4
power
0.6
0.8
1.0
Power Functions
1.0
Power Functions
1.0
Power Functions
−1.0
−0.5
0.0
δ
0.5
1.0
−1.0
−0.5
0.0
0.5
1.0
δ
Figure 3: Empirical power as function of δ for N(µ, 1) probability distributions with µ from N(0, 0.5) for
group G1 and N(δ, 0.5) for group G2 . The leftmost panel corresponds to group sample sizes n1 = n2 = 10,
the middle panel corresponds n1 = n2 = 30 and the rightmost panel to n1 = n2 = 90. The solid curve
corresponds to the asymptotic version of the proposed test in (22) and the dashed curve corresponds to
the bootstrap version. The level of the test is α = 0.05 and is indicated by the line parallel to the x-axis.
Next we consider samples of graph Laplacians of scale free networks from the BarabásiAlbert model with the Frobenius metric. These popular networks have power law degree distri20
butions and are commonly used for networks related to the world wide web, social networks and
brain connectivity networks. For scale free networks the fraction P(c) of nodes in the network
having c connections to other nodes for large values of c is approximately c−γ , with γ typically
in the range 2 ≤ γ ≤ 3. Specifically, we used the Barabási-Albert algorithm to generate samples
of scale free networks with 10 nodes, as one might encounter in brain networks. For group G1 ,
we set γ = 2.5 and for group G2 we selected a fixed γ in the interval 2 ≤ γ ≤ 3, studying the
empirical power as a function of γ. The left panel in Figure 4 indicates that in this scenario
the proposed test has better power behavior than both the graph based test and the kernel based
test. The kernel based test with automatic scaling parameter choice (Gretton et al., 2012) in the
published software has very low power, while the graph based test has a high false positive rate.
The right panel in Figure 4 shows empirical evidence that the proposed test is also consistent
in this scenario as sample size increases, and Figure 5 that especially for small samples, bootstrapping the test statistic leads to the correct empirical level of the test. With increasing sample
size, the asymptotic and the bootstrap versions of the test perform similarly.
0.8
0.6
power
0.0
0.2
0.4
0.6
0.4
0.0
0.2
power
0.8
1.0
Power Functions
1.0
Power Functions
2.0
2.2
2.4
2.6
2.8
3.0
2.0
γ
2.2
2.4
2.6
2.8
3.0
γ
Figure 4: Empirical power functions of γ for scale-free networks from the Barabási-Albert model with
parameter 2.5 for G1 and γ for G2 . In the left panel, the solid red curve corresponds to the bootstrapped
version of our proposed test with test statistic (22), the blue dashed curve to the graph based test in Chen
and Friedman (2017), the dot-dashed black curve to the energy test of Székely and Rizzo (2004) and the
green dotted curve to the kernel test in Gretton et al. (2012). Sample sizes are fixed at n1 = n2 = 100. In
the right panel, the solid power function corresponds to the proposed asymptotic test (24) for n1 = n2 =
450, the dashed power function to the test for n1 = n2 = 250 and the dotted power function to the test for
n1 = n2 = 100. The level of the tests is α = 0.05 and is indicated by the line parallel to the x-axis.
21
2.0
2.2
2.4
2.6
2.8
0.8
0.0
0.2
0.4
power
0.6
0.8
0.6
power
0.4
0.2
0.0
0.0
0.2
0.4
power
0.6
0.8
1.0
Power Functions (n=90)
1.0
Power Functions (n=30)
1.0
Power Functions (n=10)
3.0
2.0
2.2
γ
2.4
2.6
2.8
3.0
2.0
2.2
2.4
γ
2.6
2.8
3.0
γ
Figure 5: Empirical power as function of γ for scale-free networks from the Barabási-Albert model, with
the model parameter 2.5 for G1 and γ for G2 . The leftmost panel corresponds to group sample sizes
n1 = n2 = 10, the middle panel to n1 = n2 = 30 and the rightmost panel to n1 = n2 = 90. The solid curve
corresponds to the proposed asymptotic test (24) and the dashed curve to the bootstrap version of the
proposed test at level α = 0.05. The level is indicated by the line parallel to the x-axis.
0.8
0.6
power
0.0
0.2
0.4
0.6
0.4
0.0
0.2
power
0.8
1.0
Power Functions
1.0
Power Functions
0.6
0.8
1.0
1.2
1.4
0
β
10
20
30
40
50
m
Figure 6: Empirical power as function of β for 5-dimensional vectors, where each component is distributed independently as Beta(1, 1) for group G1 and as Beta(β, β) with β varying between 0.5 ≤ β ≤ 1.5
for group G2 (left). Empirical power as function of degrees of freedom m for 5-dimensional vectors
which are distributed independently as truncated N(0, I5 ) for group G1 and as truncated multivariate
t-distribution tm (0, I5 ) with varying degrees of freedom between 1 ≤ m ≤ 50 for group G2 with each
component of the vectors truncated to lie between [−5, 5] (right). Sample sizes are n1 = n2 = 100 and
level α = 0.05. Solid red curves correspond to the bootstrapped version of our proposed test with test
statistic (22), dashed blue curves to the graph based test of Chen and Friedman (2017), dotted green
curves corresponds to the kernel test of Gretton et al. (2012), dot-dashed black curves to the energy test
of Székely and Rizzo (2004) and long-dashed magenta curves to Hotelling’s T 2 test.
22
In the multivariate setting we considered 5-dimensional vectors distributed as truncated multivariate normal distributions N(0, I5 ) for group G1 where each of the components was truncated
to lie between [−5, 5]. For group G2 , we chose a 5-dimensional t-distribution tm (0, I5 ), m indicating the degrees of freedom. As the degrees of freedom m increases, the shape of the
distribution of G2 becomes more similar to that of group G1 . We obtained the empirical power
as functions of m, for 1 ≤ m ≤ 50. In a second scenario, we took the five components of the
vectors to be distributed independently as Beta(1, 1) for group G1 , while for group G2 , the five
components were assumed to be distributed independently as Beta(β, β). Empirical power was
then obtained as a function of β for 0.5 ≤ β ≤ 1.5. Figure 6 illustrates that for these cases, the
proposed test overwhelmingly outperforms the comparison tests.
6.
6.1.
DATA ILLUSTRATIONS
Mortality Data
The Human Mortality Database provides data in the form of yearly lifetables differentiated by
countries and gender. Presently it includes yearly mortality data for 37 countries, available at
<www.mortality.org>. These can be converted to a density of age-at-death for each country,
gender and calendar year, by first converting the available lifetables into histograms and then
applying local least squares smoothing, for which we used the Hades package at <http://www.
stat.ucdavis.edu/hades/> with bandwidth h = 2. The random objects we consider are the
resulting densities of age-at-death. Considering the time period 1960-2009 and the 31 countries
in the database for which records are available for this time period, we obtained the densities of
age at death for the age interval [0, 80].
From these densities we obtained quantile functions to compute the Wasserstein distance,
which is the metric we choose for this distribution space. The Fréchet standard deviations were
computed as a function of calendar year, and are shown along with pointwise 95 % bootstrap
confidence bands in Figure 7, separately for males and females. One finds that there is a small
peak in variance of mortality between 1980-1985 for males followed by a larger peak between
23
Female
4.5
Standard Deviation
4
3.5
3
2.5
2
1.5
1
1960
1965
1970
1975
1980
1985
Time
1990
1995
2000
2005
2010
1990
1995
2000
2005
2010
Male
5.5
Standard Deviation
5
4.5
4
3.5
3
2.5
2
1960
1965
1970
1975
1980
1985
Time
Figure 7: Yearly Fréchet standard deviations (solid line) along with 95 % pointwise bootstrap confidence
limits (dashed lines) for females (top) and males (bottom).
Female (Group1)
Male (Group1)
5
Standard Deviation
Standard Deviation
4.5
4
3.5
3
2.5
2
1.5
1
1960
1970
1980
1990
2000
4
3
2
1
0
1960
2010
1970
Time
Female (Group 2)
2000
2010
Male (Group 2)
4
Standard Deviation
Standard Deviation
1990
Time
5
4
3
2
1
0
1960
1980
1970
1980
1990
2000
2010
Time
3
2
1
0
1960
1970
1980
1990
2000
2010
Time
Figure 8: Yearwise Fréchet standard deviations (solid line) along with 95% pointwise bootstrap confidence limits (dashed lines) for females in group 1, females in group 2, males in group 2 and males in
group 1 (counter clockwise starting from top left)
24
1993-1996. For females, this later peak is also quite prominent. These peaks might possibly
be attributed to major political upheaval in Central and Eastern Europe during that period since
several countries in the dataset belong to these regions. The countries in the dataset that experienced some turmoil associated with the end of Communist role are Belarus, Bulgaria, Czech
Republic, Estonia, Hungary, Latvia, Poland, Lithuania, Russia, Slovakia and Ukraine.
Females
0.07
0.06
Group 1
Group 2
Density
0.05
0.04
0.03
0.02
0.01
0
0
10
20
30
40
50
60
70
80
90
60
70
80
90
Age
Males
0.06
0.05
Group 1
Group 2
Density
0.04
0.03
0.02
0.01
0
0
10
20
30
40
50
Age
Figure 9: Wasserstein-Fréchet mean age-at-death densities for the years 1960-2009 for groups 1 (red)
and 2 (blue) and for females (top) and males (bottom)
To check whether it is indeed these countries that are responsible for the variance peak
around 1990-1995, we split our dataset into two groups, group 1 consisting of the above Eastern European countries and group 2 of all other countries and repeated the earlier analysis.
Figure 8 shows that for group 2 the variance of age-at-death distributions indeed has a decreasing trend over the years for both males and females, while the variance shows distinct
25
Females
p−values
1
0.5
0
1960
1970
1980
1990
2000
2010
1990
2000
2010
Years
Males
p−values
1
0.5
0
1960
1970
1980
Years
Figure 10: p-values for testing the differences in population age-at-death distributions of Groups 1 and 2
over the years for females (top) and males (bottom) with the proposed test.
fluctuations for both males and females in group 1. Figure 9 illustrates the group-wise Wasserstein Fréchet mean densities of the countries for the various calendar years. There seems to be a
clear difference between the mean densities of the two groups for both males and females, and
implementing the bootstrap version of the proposed test to compare the distributions of age-atdeath between groups 1 and 2. For obtaining the p-values we carried out the bootstrap version
of the proposed test due to the relatively small sample sizes. Figure 10 illustrates the p-values
obtained for each year. The null hypothesis that the populations are identical is rejected for both
males and females at the 5% level for most years between 1990-1995.
6.2.
Analyzing intra-hub connectivities using fMRI data for Alzheimer’s Disease
Alzheimers disease (AD) is an irreversible, progressive neuro-degenerative brain disorder that
slowly destroys memory and thinking skills, eventually leading to severe dementia. AD has
been found to have associations with abnormalities in functional integration of brain regions.
Recent studies as in Sui et al. (2015) have indicated that AD selectively targets regions of highconnectivity (so-called hubs) in the brain. The posterior midline, in particular the posterior
26
cingulate/precuneus (PCP) as described in Buckner et al. (2009) is a nexus or hub of high
cortical connectivity and functional connectivity in this region could be a potential biomarker
for AD. For each hub, a so-called seed voxel is identified as the voxel with the signal that has
the highest correlation with the signals of nearby voxels. To quantify intra-hub connectivity,
following Petersen and Müller (2016), we analyze the distribution of the correlations between
the signal at the seed voxel of the PCP hub and the signals of all other voxels within an 11 ×
11 × 11 cube of voxels that is centered at the seed voxel.
The subjects in our analysis consisted of cognitively normal elderly patients and demented
elderly patients diagnosed with AD (after removal of outliers), each of whom underwent an
fMRI scan at the UC Davis Imaging Research Center. Preprocessing of the recorded BOLD
(blood-oxygenation-level-dependent) signals was implemented by adopting the standard procedures of slice-timing correction, head motion correction and normalization, in addition to linear
detrending to account for signal drift and band-pass filtering to include only frequencies between 0.01 and 0.08 Hz. The signals for each subject were recorded over the interval [0, 470]
(in seconds), with 236 measurements available at 2 second intervals. The study actually had 171
normal subjects but since AD is a disease that is known to progress with age, for fair comparison, only 87 out of these 171 were selected, by matching their ages with that of the demented
patients. To check that the age matching worked, the age distributions of the 87 normal elderly
subjects in our analysis and the 65 AD patients were compared with the Wilcoxon rank sum
test for the null hypothesis of equal age distributions of the two groups, which yielded a p-value
of 0.84. For each subject, the target is the density function of positive correlations within the
PCP hub, where this density was estimated from the observed correlations with a kernel density
estimator, utilizing the standard Gaussian kernel and bandwidth h = 0.08. As negative correlations are commonly ignored in connectivity analyses, the densities were estimated on [0, 1].
The resulting sample of densities is then an i.i.d. sample across subjects. Figure 11 shows the
Wasserstein Fréchet mean probability distribution functions. To compare the two populations of
distributions, we applied the asymptotic and the bootstrap version of the proposed test to these
27
Wasserstein mean probability distribution functions
Probability Distribution Function
1
0.9
0.8
0.7
0.6
0.5
0.4
0.3
0.2
Demented Subjects
Normal Subjects
0.1
0
0
0.2
0.4
0.6
0.8
1
Correlation
Figure 11: Wasserstein mean probability distribution functions of positive correlations in the PCP hub for normal subjects (blue) and demented patients (red)
samples of density functions, which yielded a p-value of p = 0.002 (bootstrap p-value=0.001),
indicating that significant differences exist between in terms of intra-hub connectivity between
AD patients and age-matched normal subjects.
6.3.
Comparing brain networks of Alzheimer’s patients
Recent advances in neurological studies have revealed that brain hubs, being regions of high
connectivity in the brain, interconnect with each other for functional integration of their specialized roles. Studying interconnections between hubs can reveal important insights about
brain diseases like AD. Disorders of cognition can be associated with disrupted connectivity
between cortical hubs as discussed in Buckner et al. (2009). One question of interest is whether
the interconnections change with aging in patients having dementia. We consider connections
between 10 cortical hubs that are listed in Table 3 of Buckner et al. (2009).
In order to analyze the cognitively impaired patients we considered the 65 demented subjects
that were discussed in the preceding subsection. For each subject, a 10 × 10 connectivity matrix
was obtained whose entries are the correlations between average fMRI signals (with the same
pre-processing as described in subsection 6.2) from 3 × 3 × 3 cubes around the seed voxels of
the 10 hubs. These subject-specific connectivity matrices were thresholded at 0.25 as discussed
in Buckner et al. (2009) to obtain adjacency matrices of networks with the hubs as the nodes,
28
so that the presence of an edge indicates a correlation greater than 0.25. Subject-specific graph
Laplacians were then formed from these adjacency matrices.
These cognitively impaired patients were split into three groups based on their ages. Subjects were assigned to groups G1 , G2 or G3 based on whether they were aged 70 or below,
between age 70 and 80 or 80 and above. The left panel in Figure 12 shows the difference of the
average graph Laplacians of subjects in group G2 and subjects in group G1 and the right panel
the difference of the average graph Laplacians of subjects in group G3 and subjects in group
G1 .
Since the group sample sizes of G1 , G2 and G3 are small we applied the bootstrap version of
the proposed test to see if the differences are significant in the populations of graph Laplacians
of the three different groups of cognitively impaired patients. The null hypothesis of equality
of population distributions of the graph Laplacians was rejected with a bootstrap p-value of
0.032. The conclusion of the test indicates that there is evidence to support that the inter hub
connections do show changes with age for patients having Alzheimer’s disease.
2
2
1
4
0.5
HUBS
HUBS
4
1
6
0.5
6
0
0
8
8
−0.5
−0.5
10
10
2
4
6
HUBS
8
10
2
4
6
HUBS
8
10
Figure 12: The left panel shows the difference between the average graph Laplacians of demented patients aged between 70 and 80 and demented patients aged 70 or below. The right panel shows the
difference between the average graph Laplacians of demented patients aged 80 or above and demented
patients aged 70 or below.
REFERENCES
Barabási, A.-L. and Albert, R. (1999), “Emergence of scaling in random networks,” Science,
286, 509–512.
29
Barden, D., Le, H., and Owen, M. (2013), “Central limit theorems for Fréchet means in the
space of phylogenetic trees,” Electronic Journal of Probability, 18, 1–25.
Baringhaus, L. and Franz, C. (2004), “On a new multivariate two-sample test,” Journal of Multivariate Analysis, 88, 190–206.
Beran, R. (2003), “The impact of the bootstrap on statistical algorithms and theory,” Statistical
Science, 18, 175–184.
Bhattacharya, A. and Bhattacharya, R. (2012), Nonparametric Inference on Manifolds: With
Applications to Shape Spaces, no. 2, Cambridge University Press.
Bhattacharya, R. and Lin, L. (2017), “Omnibus CLTs for Fréchet means and nonparametric
inference on non-Euclidean spaces,” Proceedings of the American Mathematical Society,
145, 413–428.
Bhattacharya, R. and Patrangenaru, V. (2003), “Large sample theory of intrinsic and extrinsic
sample means on manifolds. I,” Annals of Statistics, 31, 1–29.
— (2005), “Large sample theory of intrinsic and extrinsic sample means on manifolds: II,”
Annals of Statistics, 33, 1225–1259.
Bickel, P. J. and Freedman, D. A. (1981), “Some asymptotic theory for the bootstrap,” Annals
of Statistics, 9, 1196–1217.
Buckner, R. L., Sepulcre, J., Talukdar, T., Krienen, F. M., Liu, H., Hedden, T., Andrews-Hanna,
J. R., Sperling, R. A., and Johnson, K. A. (2009), “Cortical hubs revealed by intrinsic functional connectivity: mapping, assessment of stability, and relation to Alzheimer’s disease,”
The Journal of Neuroscience, 29, 1860–1873.
Chen, H., Chen, X., and Su, Y. (2017), “A weighted edge-count two-sample test for multivariate
and object data,” Journal of the American Statistical Association.
Chen, H. and Friedman, J. H. (2017), “A new graph-based two-sample test for multivariate and
object data,” Journal of the American Statistical Association, 112, 397–409.
Fréchet, M. (1948), “Les éléments aléatoires de nature quelconque dans un espace distancié,”
10, 215–310.
Friedman, J. H. and Rafsky, L. C. (1979), “Multivariate generalizations of the Wald-Wolfowitz
and Smirnov two-sample tests,” Annals of Statistics, 7, 697–717.
Ginestet, C. E., Li, J., Balachandran, P., Rosenberg, S., Kolaczyk, E. D., et al. (2017), “Hypothesis testing for network data in functional neuroimaging,” The Annals of Applied Statistics,
11, 725–750.
Gretton, A., Borgwardt, K. M., Rasch, M. J., Schölkopf, B., and Smola, A. (2012), “A kernel
two-sample test,” Journal of Machine Learning Research, 13, 723–773.
Guo, R. and Patrangenaru, V. (2017), “Testing for the Equality of two Distributions on High
30
Dimensional Object Spaces,” arXiv preprint arXiv:1703.07856.
Henze, N. (1988), “A multivariate two-sample test based on the number of nearest neighbor
type coincidences,” Annals of Statistics, 16, 772–783.
Henze, N. and Penrose, M. D. (1999), “On the multivariate runs test,” Annals of Statistics, 27,
290–298.
Janssen, A. and Pauls, T. (2003), “How do bootstrap and permutation tests work?” Annals of
Statistics, 31, 768–806.
Kendall, W. S. and Le, H. (2011), “Limit theorems for empirical Fréchet means of independent and non-identically distributed manifold-valued random variables,” Brazilian Journal of
Probability and Statistics, 25, 323–352.
Levene, H. (1960), “Robust tests for equality of variances,” Contributions to Probability and
Statistics, 1, 278–292.
Lyons, R. (2013), “Distance covariance in metric spaces,” Annals of Probability, 41, 3284–
3305.
Patrangenaru, V. and Ellingson, L. (2015), Nonparametric statistics on manifolds and their
applications to object data analysis, CRC Press.
Petersen, A. and Müller, H.-G. (2016), “Functional data analysis for density functions by transformation to a Hilbert space,” Annals of Statistics, 44, 183–218.
— (2017), “Fréchet Regression for Random Objects with Euclidean predictors,” Annals of
Statistics, to appear (arXiv preprint arXiv:1608.03012), xx, xxx–xxx.
Ramdas, A., Trillos, N. G., and Cuturi, M. (2017), “On Wasserstein Two-Sample Testing and
Related Families of Nonparametric Tests,” Entropy, 19, 47.
Rosenbaum, P. R. (2005), “An exact distribution-free test for comparing two multivariate distributions based on adjacency,” Journal of the Royal Statistical Society: Series B (Statistical
Methodology), 67, 515–530.
Schilling, M. F. (1986), “Multivariate two-sample tests based on nearest neighbors,” Journal of
the American Statistical Association, 81, 799–806.
Sejdinovic, D., Sriperumbudur, B., Gretton, A., and Fukumizu, K. (2013), “Equivalence of
distance-based and RKHS-based statistics in hypothesis testing,” Annals of Statistics, 41,
2263–2291.
Sui, X., Zhu, M., Cui, Y., Yu, C., Sui, J., Zhang, X., Liu, J., Duan, Y., Zhang, Z., Wang, L.,
Zhang, X., and Jiang, T. (2015), “Functional connectivity hubs could serve as a potential
biomarker in Alzheimers disease: a reproducible study,” Current Alzheimer Research, 12,
974–983.
Szarek, S. J. (1998), “Metric Entropy of Homogeneous Spaces,” Quantum Probability, 43, 395–
31
410.
Székely, G. J. and Rizzo, M. L. (2004), “Testing for equal distributions in high dimension,”
InterStat, 5, 1–6.
— (2017), “The Energy of Data,” Annual Review of Statistics and Its Application, 4, 447–479.
Villani, C. (2003), Topics in Optimal Transportation, American Mathematical Society.
Wellner, J. A. and van der Vaart, A. W. (1996), Weak Convergence and Empirical Processes:
With Applications to Statistics, Springer Series in Statistics.
S UPPLEMENTARY M ATERIALS
S.1 Main Proofs
Proof of Proposition 1. Let QΩ be the space of quantile functions corresponding to the space
Ω of univariate distribution functions on R with finite second moments. For a random object Y
taking values in Ω, let QY denote the corresponding quantile function. By the convexity of the
space QΩ and the properties of the L2 -Wasserstein distance, the population and sample Fréchet
−1
thereby
means exist and are unique and given by µF = (E(QY (·)))−1 and µ̂F = n1 ∑ni=1 QYi (·)
proving (P0). For proving (P1) and (P2) observe that quantile functions are a part of a bigger
class M comprising of monotone functions and under L2 metric dL2 (the Wasserstein metric
corresponds to the L2 metric on the space of quantile functions) the metric entropy of the space
M is upper bounded as log N(ε, M, dL2 ) ≤
a
ε
for some constant a > 0 as per Theorem 2.7.5 in
Wellner and van der Vaart (1996). Therefore,
J(δ) ≤
Z 1r
1
1 + dε ≤
εδ
0
Z 1
0
2
1
dε = 1 + √ ,
1+ √
εδ
δ
and so δ J(δ) → 0 as δ → 0, thus establishing (P1). The entropy integral of the whole space is
Z 1p
0
1 + log N(ε, Ω, dW ) ≤
which establishes (P2).
32
Z 1r
0
1
1 + dε = 3,
ε
Proof of Proposition 2. The space of graph Laplacians for graphs considered in this Proposition
is convex (Ginestet et al., 2017) and so is the space of correlation matrices. The properties of
Frobenius distance imply that µF = E(Y ) and µ̂F = n−1 ∑ni=1 Yi , which exist and are unique by
the convexity of Ω proving (P0). (P1) and P(2) hold as both the space of graph Laplacians and
the space of correlation matrices are bounded subsets of a finite-dimensional Euclidean space
2
Rr of r × r matrices and under the Euclidean metric dE , which is equivalent to the Frobenius
metric for the matrices. The metric entropy of this space is seen to be bounded above by
2
log N(ε, Rr , dE ) ≤ ar2 log 1 + 1ε for some constant a > 0, due to a simple volume comparison
argument (Szarek, 1998), so that
J(δ) ≤
Z 1
0
s
s
!
Z 1
√
√
1
1
2
1 + ar2 log 1 +
dε ≤
1 + ar log 1 +
dε ≤ 1+ ar(1+ √ ),
εδ
εδ
0
δ
implying δ J(δ) → 0 as δ → 0 and establishing (P1). Using similar upper bounds for the entropy
integral of the whole space,
Z 1p
0
1 + log N(ε, Ω, dF ) ≤
Z 1
s
1 + ar2 log
0
√
1
dε ≤ 1 + 3 ar,
1+
ε
establishing (P2).
Proof of Proposition 3. Define
Mn (ω) =
1 n 2
∑ {d (ω,Yi) − d 2(µF ,Yi) − E(d 2(ω,Yi)) + E(d 2(µF ,Yi))}.
n i=1
In a first step we control the behavior of Mn (ω) uniformly for small d(ω, µF ). Define functions
gω : Ω → R as gω (y) = d 2 (y, ω) and the function class Mδ = {gω − gµF : d(ω, µF ) < δ}. An
envelope function for Mδ is G(δ) = 2diam(Ω)δ. Let J = J(δ) be the integral in (P1) so that
δ J(δ) → 0 as δ → 0. Theorems 2.7.11 and 2.14.2 of Wellner and van der Vaart (1996) and (P1)
33
imply that for small δ > 0,
!
E
sup
|Mn (ω)|
≤
d(ω,µF )<δ
therefore
J(δ)G(δ)
√
,
n
!
E
sup
|Mn (ω)|
≤
d(ω,µF )<δ
aδ J(δ)
√
n
(29)
for some a > 0. We want to show that for any ε > 0, γ > 0, there exists N = N(ε, γ) such that
for all n ≥ N,
P
!
1 n 2
√ ∑ {d (µ̂F ,Yi ) − d 2 (µF ,Yi )} > ε < γ.
n i=1
For any small δ > 0,
n
1
√ ∑ {d 2 (µ̂F ,Yi ) − d 2 (µF ,Yi )} > ε
n i=1
P
!
!
n
1
ε
{d 2 (µ̂F ,Yi ) − d 2 (µF ,Yi )} < − √ , d(µ̂F , µF ) ≤ δ + P (d(µ̂F , µF ) > δ)
∑
n i=1
n
!
ε
1 n 2
≤P − inf
{d (ω,Yi ) − d 2 (µF ,Yi )} > √
+ P (d(µ̂F , µF ) > δ)
∑
n
d(ω,µF )<δ n i=1
!
ε
+ P (d(µ̂F , µF ) > δ) ,
≤P
sup |Mn (ω)| > √
n
d(ω,µF )<δ
≤P
since
sup
|Mn (ω)|
d(ω,µF )<δ
=
sup
d(ω,µF )<δ
≥
inf
d(ω,µF )<δ
1 n 2
{d (ω,Yi ) − d 2 (µF ,Yi ) − E(d 2 (ω,Yi )) + E(d 2 (µF ,Yi ))}
∑
n i=1
E(d 2 (ω,Yi )) − E(d 2 (µF ,Yi )) −
1 n 2
∑ {d (ω,Yi) − d 2(µF ,Yi)}
d(ω,µF )<δ n i=1
1 n 2
{d (ω,Yi ) − d 2 (µF ,Yi )},
= − inf
∑
d(ω,µF )<δ n i=1
34
inf
(30)
using
{E(d 2 (ω,Yi )) − E(d 2 (µF ,Yi ))} = 0
inf
d(ω,µF )<δ
from assumption (P0).
By using Markov’s inequality and the bound in equation (29), for any small δ > 0 such that
δ J(δ) <
γε
2a ,
the expression in (30) can be bounded above by
!
ε
P
sup |Mn (ω)| > √
+ P (d(µ̂F , µF ) > δ)
n
d(ω,µF )<δ
!√
n
≤E
sup |Mn (ω)|
+ P (d(µ̂F , µF ) > δ)
ε
d(ω,µF )<δ
≤
aδ J(δ)
γ
+ P (d(µ̂F , µF ) > δ) < + P (d(µ̂F , µF ) > δ) .
ε
2
For any such δ, using the consistency of Fréchet mean µ̂F it is possible to choose N such that
P (d(µ̂F , µF ) > δ) <
γ
2
for all n ≥ N. This completes the proof.
Proof of Theorem 1.
√
n(V̂F −VF )
√
= n
!
√
1 n 2
{d (µ̂F ,Yi ) − d 2 (µF ,Yi )} + n
∑
n i=1
!
1 n 2
∑ {d (µF ,Yi) − E(d 2(µF ,Y1))} ,
n i=1
where the first term is oP (1) by Proposition 3 and the second term converges in distribution to
N(0, σ2F ) by applying Central Limit Theorem to i.i.d random variables d 2 (µF ,Y1 ), . . . , d 2 (µF ,Yn ).
Theorem 1 then follows directly from Slutsky’s Theorem.
Proof of Proposition 4. Observe that
!
1 n 4
1 n 4
∑ d (µ̂F ,Yi) − n ∑ d (µF ,Yi) ≤ 2diam2(Ω)
n i=1
i=1
35
!
1 n 2
1 n 2
∑ d (µ̂F ,Yi) − n ∑ d (µF ,Yi) (31)
n i=1
i=1
4 (µ ,Y )
E
d
√
F
which is o p √1n by Proposition 3. Hence n
−
1 n
2
2
E d (µF ,Y )
∑i=1 d (µ̂F ,Yi )
n
1 n
1 n
4 (µ̂ ,Y )
4 (µ ,Y )
d
d
∑
∑
√
F i
F i
i=1
− n i=1
can be decomposed into two components, n n
1 n
1 n
2
2
n ∑i=1 d (µ̂F ,Yi )
n ∑i=1 d (µF ,Yi )
1 n
4
4
E d (µF ,Y )
∑i=1 d (µF ,Yi )
√
−
which is oP (1) by equation (31) and Proposition 3 and n n
1 n
2
2
E d (µF ,Y )
n ∑i=1 d (µF ,Yi )
which converges
in
distribution to N(0, D) by applying Central Limit Theorem to i.i.d random
1 n
∑ d 4 (µ̂F ,Yi )
n i=1
d 4 (µF ,Yi )
, i = 1, 2, . . . , n, with
vectors
d 2 (µF ,Yi )
Cov d 4 (µF ,Y ), d 2 (µF ,Y )
.
D=
4
2
2
Cov d (µF ,Y ), d (µF ,Y )
Var d (µF ,Y )
Var d 4 (µF ,Y )
Observing
σ̂2F
=g
!
1 n 4
1 n 2
∑ d (µ̂F ,Yi), n ∑ d (µ̂F ,Yi) ,
n i=1
i=1
where g(x1 , x2 ) = x1 − x22 is a differentiable function with gradient function Og = (1, −2x2 ), a
simple application of the delta method establishes the asymptotic normality of σ̂2F . The asymp
4
E d (µF ,Y )
totic variance is given by a0 Da, where a is the gradient function Og evaluated at
.
2
E d (µF ,Y )
Proof of Proposition 5. Under the null hypothesis the groupwise means are all equal,
µ1 = µ2 = · · · = µk = µ.
36
Then, under assumptions (P0) and (P1),
k
√
√
nFn = n(V̂ − ∑ λ j,nV̂ j )
j=1
k √n
1 n 2
j 1
2
= √ ∑ {d (µ̂,Yi ) − d (µ,Yi )} − ∑ √ √ ∑ {d 2 (µ̂ j ,Yi ) − d 2 (µ,Yi )}
n i=1
n n j i∈G j
j=1
(32)
=o p (1),
(33)
where (33) follows from (32) by applying Proposition 3 to all observations and also for the
√
p
n
individual groups and noting that √nj → λ j for all j = 1, 2, ..., k. Slutsky’s theorem completes
the proof.
Proof of Proposition 6. Under the null hypothesis, let µ1 = µ2 = · · · = µk = µ and V1 = V2 =
· · · = Vk = V . Using Proposition 4 and since λ j,n → λ j as n → ∞ we find for the denominator
k
λ j,n P
∑ σ̂2 −→
j=1
j
k
λj
∑ σ2 .
j=1
(34)
j
Simple algebraic manipulation shows that the numerator of nUn is
λ j,n λl,n
(V̂ j − V̂l )2
2
2
j<l σ̂ j σ̂l
!2
k λ
k λ
k λ
j,n
j,n
j,n
=n ∑ 2 Ṽ j2 ∑ 2 − ∑ 2 Ṽ j
j=1 σ̂ j
j=1 σ̂ j
j=1 σ̂ j
!
k λ
˜ nλ
˜0
λ
j,n
n
=n ∑ 2 Ṽ 0 Λn −
Ṽ
λ
j,n
k
j=1 σ̂ j
∑
2
n∑
j=1 σ̂
j
37
(35)
under the null hypothesis. Here Ṽ j = V̂ j −V and
λ1,n
Ṽ1
σ̂21
Ṽ2
0
Ṽ = . , Λn = .
..
..
Ṽk
0
λ1
σ21
0
Λ= .
..
0
···
0
λ2,n
σ̂22
.
···
..
.
0
..
.
0
···
λk,n
σ̂2k
..
···
0
0
√
λ1,n
σ̂21
λ2,n
2
σ̂2
˜
, λn = . , sn = .
..
..
√
λk,n
σ̂k
λk,n
σ̂2k
λ1
σ21
λ1,n
√σ̂1
λ2,n
σ̂2
,
√
λ1
√σ1
λ
λ
22
2
0
˜ σ2
σ2
,
λ
=
,
s
=
. .
..
..
..
.
.
√
λ2
σ22
···
.. ..
.
.
λk
σ2k
···
0
0
λk
σ2k
λk
σk
Applying Theorem 1 to the individual groups we find
Zn =
Continuing from (35), we see that
√ 21
D
nΛn Ṽ −→ N(0, Ik ).
nUn
λ j,n
k
∑ j=1 2
σ̂ j
is Zn0 An Zn with An = Ik − sn (s0n sn )−1 s0n , and
P
An −→ A = Ik − s(s0 s)−1 s0 .
(36)
Here A is a symmetric idempotent matrix and is an orthogonal projection into the space orthogonal to the column space of s. The rank of A is same as its trace and equals k − 1 by the property
of orthogonal projector matrices. Applying the continuous mapping theorem, Slutsky’s theorem
and (36),
nUn
λ j,n
∑kj=1 σ̂2
j
D
−→ Z 0 AZ,
(37)
where the limiting distribution is a quadratic form of normal random variables and is therefore
distributed as a χ2 distribution with degrees of freedom equal to rank of A, which is k − 1.
38
Proof of Proposition 7. For proving consistency of the pooled Fréchet mean µ̂ p , the arguments
are essentially the same as those in the proof of Lemma 1 in Petersen and Müller (2017). Consider
1
Mn (ω) =
n
k
k
∑ ∑d
2
(ω,Yi j ),
M(ω) =
j=1 i∈G j
∑ λ j E j (d 2(ω,Y j )).
j=1
For each ω ∈ Ω,
k
Mn (ω) =
1
P
∑ λ j,n n j ∑ d 2(ω,Yi j ) −→ M(ω),
j=1
i∈G j
by the weak law of large numbers applied to the individual groups, using lim λ j,n = λ j for each
n→∞
j = 1, . . . , k. From |Mn (ω1 ) − Mn (ω2 )| ≤ 2 diam(Ω)d(ω1 , ω2 ) we find that Mn is asymptotically
equicontinuous in probability, as
sup
|Mn (ω1 ) − Mn (ω2 )| = O p (δ),
d(ω1 ,ω2 )<δ
which allows us to use Theorem 1.5.4 in Wellner and van der Vaart (1996) to conclude that
Mn converges weakly to M in l ∞ (Ω). By applying 1.3.6 of Wellner and van der Vaart (1996)
we have that supω∈Ω |Mn (ω) − M(ω)| converges to zero in probability. By our assumptions and
Corollary 3.2.3 in Wellner and van der Vaart (1996) this implies that
P
µ̂ p −→ µ p .
(38)
For proving consistency of Fn it is enough to prove the consistency of V̂p as the consistency of
the groupwise Fréchet variances follows from our earlier results. Observe that
V̂p −
1
n
k
∑ ∑ d 2(µ p,Yi j )
≤ 2 diam(Ω) d(µ̂ p , µ p ) = oP (1),
j=1 i∈G j
39
(39)
which implies
V̂p −Vp ≤ V̂p −
1
n
k
∑∑
d 2 (µ p ,Yi j ) +
j=1 i∈G j
1
n
k
∑ ∑ d 2(µ p,Yi j ) −Vp
= oP (1).
j=1 i∈G j
Here the first term is oP (1) by (39) and the second term is also oP (1) by the weak law of large
numbers applied to the individual groups, whence V̂ converges in probability to Vp . Clearly Vp −
∑kj=1 λ jV j = ∑kj=1 λ j E j (d 2 (µ p ,Y j )) − E j (d 2 (µ j ,Y j )) is nonnegative, as for each individual
group we have
E j (d 2 (µ p ,Y j )) − E j (d 2 (µ j ,Y j )) ≥ 0,
with equality with zero holding only if µ p = µ j for all j = 1, . . . , k. Therefore F is always
nonnegative and is zero if and only of µ p = µ j for all j = 1, . . . , k.
Proof of Theorem 3. This proof relies on the following auxiliary result on uniform consistency
of estimators V̂p , V̂ j and σ̂2j for all j = 1, 2, . . . , k, under the assumption of boundedness of the
entropy integral for the space Ω.
Lemma 1. Under the assumptions of Theorem 3, it holds for all ε > 0 and for all j = 1, 2, . . . , k,
where we denote any of the V̂ j and V j by V̂ and V respectively and any of the σ̂ j and σ j by σ̂
and σ respectively, that
(A) lim supP∈P P( V̂ −V > ε) = 0;
n→∞
(B) lim supP∈P P( σ̂2 − σ2 > ε) = 0;
n→∞
(C) lim supP∈P P( V̂p −Vp > ε) = 0.
n→∞
In all of the above statements the supremum is taken with respect to the underlying true probability measure P of Y1 ,Y2 , . . . ,Yn , over the class P of possible probability measures which
generate random observations from Ω.
40
The proof of Lemma 1 can be found in the following subsection titled Additional Proofs.
With similar notation as in the proof of Proposition 6, define
Ṽ j = V̂ j −V j
for j = 1, . . . , k. The statistic Un then can be represented as
λ j,n λl,n
(V̂ j − V̂l )2 = Ũn + ∆n ,
2
2
j<l σ̂ j σ̂l
Un =
∑
λ j,n λl,n
(Ṽ j − Ṽl )2
σ̂2j σ̂2l
where Ũn = ∑ j<l
and ∆n = ∑ j<l
λ j,n λl,n
(V j −Vl )2 +
σ̂2j σ̂2l
2 ∑ j<l
Ṽl ). By replicating the steps in the proof of Proposition 6 we find that
λ j,n λl,n
(V j −Vl )(Ṽ j −
σ̂2j σ̂2l
nŨn
∑kj=1
λ j,n
σ̂2j
converges in
distribution to χ2(k−1) asymptotically. Moreover as a consequence of Lemma 1 and by continuity
we have that
∆n
λ j,n
∑kj=1 2
σ̂ j
+
Fn2
k
∑ j=1 λ2j,n σ̂2j
is a uniformly consistent estimator of
U
λj
∑kj=1 2
σj
+
F2
.
∑kj=1 λ2j σ2j
For sets {An } defined as
An =
∆n
+
λ
∑kj=1 j,n2
σ̂ j
the uniform consistency of
σj
∆n
∑kj=1
Fn2
1 U
<
k
2
2
2 ∑k λ j
∑ j=1 λ j,n σ̂ j
j=1 2
λ j,n
σ̂2j
+
Fn2
k
∑ j=1 λ2j,n σ̂2j
F2
+ k
∑ j=1 λ2j σ2j
implies that as n → ∞,
sup P(An )
P∈P
≤ sup P
P∈P
∆n
∑kj=1
λ j,n
σ̂2j
+
Fn2
U
F2
F2
1 U
−
−
>
+
→ 0.
k
k
k
2
2
2
2
2
2
λ
λ
j
j
2 ∑k
λ
σ
∑ j=1 λ j,n σ̂ j ∑kj=1 2 ∑ j=1 λ j σ j
∑
j=1
j
j
2
j=1
σj
σj
Writing cα for the (1 − α)-th quantile of χ2(k−1) distribution, we can now represent the limiting
41
power function as
cα
Fn2
Ũn + ∆n
P(Rn,α ) =P
>
+
k
2
2
λ
j,n
n
∑ j=1 λ j,n σ̂ j
∑kj=1 σ̂2
j
≥P
Ũn
∑kj=1
λ j,n
σ̂2j
1 U
+
2 ∑k
λj
j=1 σ2
j
≥P
∑kj=1
λ j,n
σ̂2j
λj
j=1 σ2
j
+
nŨn
∑kj=1
n U
> cα −
2 ∑k
λ j,n
σ̂2j
F2
cα C
> , An
k
2
2
n
∑ j=1 λ j σ j
n U
> cα −
2 ∑k
nŨn
≥P
+
λj
j=1 σ2
j
+
F2
− P (An )
k
2
2
∑ j=1 λ j σ j
F2
− sup P (An ) .
k
2
2
∑ j=1 λ j σ j
P∈P
This implies for the sequence of hypotheses {Hn },
lim βHn
n→∞
= lim inf P (Rn,α )
n→∞ Hn
≥ lim P
n→∞
nŨn
∑kj=1
λ j,n
σ̂2j
n→∞
Since
that
nŨn
∑kj=1
λ j,n
σ̂2j
n bn
> cα −
2 ∑k
+ a2n − lim
λj
n→∞
j=1 σ2
j
nŨn
λ j,n
∑kj=1 σ̂2
j
sup P (An )
P∈P
= lim P
n bn
> cα −
2 ∑k
λj
j=1 σ2
j
+ a2n .
converges in distribution to a χ2(k−1) random variable, we find that if an is such
√
nan → ∞ or if bn is such that nbn → ∞, then lim βHn = 1, completing the proof.
n→∞
42
S.2 Additional Proofs
Proof of Lemma 1. (A). Since the proof is similar for all j = 1, 2, . . . , k, we ignore the index j
for the proof. Observe that
P( V̂ −V > ε) = P
!
1 n 2
1 n 2
1 n 2
∑ d (µ̂,Yi) − n ∑ d (µ,Yi) + n ∑ d (µ,Yi) − E(d 2(µ,Y )) > ε
n i=1
i=1
i=1
≤ An + Bn
with An and Bn being respectively equal to P
and P
1 n
2
2
n ∑i=1 d (µ,Yi ) − E(d (µ,Y ))
1 n
1 n
2
2
n ∑i=1 d (µ̂,Yi ) − n ∑i=1 d (µ,Yi )
> ε/2
> ε/2 .
Observing that infω∈Ω E(d 2 (ω,Y ) − d 2 (µ,Y )) = 0 we have,
n
n
n
!
n
1
1
1
1
d 2 (µ̂,Yi ) − ∑ d 2 (µ,Yi ) = inf
d 2 (ω,Yi ) − ∑ d 2 (µ,Yi )
∑
∑
ω∈Ω n i=1
n i=1
n i=1
n i=1
!
1 n 2
d (ω,Yi ) − d 2 (µ,Yi ) − inf E(d 2 (ω,Y ) − d 2 (µ,Y ))
= inf
∑
ω∈Ω
ω∈Ω n i=1
≤ sup
ω∈Ω
1 n 2
d (ω,Yi ) − d 2 (µ,Yi ) − E(d 2 (ω,Y ) + E(d 2 (µ,Y )) = sup |Mn (ω)| ,
∑
n i=1
ω∈Ω
with Mn (ω) = 1n ∑ni=1 d 2 (ω,Yi ) − d 2 (µ,Yi ) − E(d 2 (ω,Y ) + E(d 2 (µ,Y )). Replicating the
2
√ (Ω) , where J as
steps in proof of Proposition 3, one obtains E (supω∈Ω |Mn (ω)|) ≤ 2Jdiam
n
R1p
given by J = 0 1 + log N(ε, Ω, d)dε is the finite entropy integral of Ω and 2diam2 (Ω)
is the envelope for the function class {d 2 (ω, ·) − d 2 (µ, ·) : ω ∈ Ω} which indexes the
empirical process Mn (ω). By Markov’s inequality,
An ≤
4Jdiam2 (Ω)
√
.
nε
43
(40)
Next we observe that
1 n 2
1 n 2
2
d
(µ,Y
)
−
E(d
(µ,Y
))
≤
sup
d (ω,Yi ) − E(d 2 (ω,Y )) = sup |Hn (ω)| ,
i
∑
∑
n i=1
n
ω∈Ω
ω∈Ω
i=1
where Hn (ω) = 1n ∑ni=1 d 2 (ω,Yi ) − E(d 2 (ω,Y )). By similar arguments as before we can
see that E (supω∈Ω |Hn (ω)|) ≤
2
Jdiam
√ (Ω) ,
n
where J is the finite entropy integral of Ω and
diam2 (Ω) is the envelope of the function class {d 2 (ω, ·) : ω ∈ Ω}, which indexes the
empirical process Hn (ω). Again by Markov’s inequality,
B≤
2Jdiam2 (Ω)
√
.
nε
(41)
From equations (40) and (41),
sup P(|V̂ −V | > ε) ≤
P∈P
6Jdiam2 (Ω)
√
→0
nε
as n → ∞,
(42)
completing the proof of (A).
(B). Since the proof is similar for all j = 1, 2, . . . , k, we ignore the index j for the proof. For
proving the uniform consistency of σ̂2 it is enough to prove just the uniform consistency
of n1 ∑ni=1 d 4 (µ̂,Yi ) to E(d 4 (µ,Y )) and the rest follows from Lemma 1 (A) by continuity.
Similarly to the proof of (A), we find
P
=P
!
1 n 4
∑ d (µ̂,Yi) − E(d 4(µ,Y )) > ε
n i=1
!
1 n 4
1 n 4
1 n 4
∑ d (µ̂,Yi) − n ∑ d (µ,Yi) + n ∑ d (µ,Yi) − E(d 4(µ,Y )) > ε
n i=1
i=1
i=1
≤An + Bn ,
where
44
An = P
Bn = P
!
1 n 4
1 n 4
∑ d (µ̂,Yi) − n ∑ d (µ,Yi) > ε/2 ,
n i=1
i=1
!
n
1
4
4
∑ d (µ,Yi) − E(d (µ,Y )) > ε/2 .
n i=1
Observe that in analogy to the proof of (A),
1 n 4
1 n 4
1 n 2
1 n 2
2
d
(µ̂,Y
)
−
d
(µ,Y
)
≤
2diam
(Ω)
d
(µ̂,Y
)
−
i
i
i
∑
∑
∑
∑ d (µ,Yi)
n i=1
n i=1
n i=1
n i=1
implies
An ≤
8Jdiam4 (Ω)
√
.
nε
(43)
Next we observe
1 n 4
1 n 4
4
d
(µ,Y
)
−
E(d
(µ,Y
))
≤
sup
d (ω,Yi ) − E(d 4 (ω,Y )) = sup |Kn (ω)| ,
i
∑
∑
n i=1
n
ω∈Ω
ω∈Ω
i=1
(44)
where Kn (ω) = n1 ∑ni=1 d 4 (ω,Yi ) − E(d 4 (ω,Y )). Similar arguments as in proof of (A)
imply E (supω∈Ω |Kn (ω)|) ≤
4
Jdiam
√ (Ω) ,
n
where J is the finite entropy integral of Ω and
diam4 (Ω) is the envelope of the function class {d 4 (ω, ·) : ω ∈ Ω}, which indexes the
empirical process Kn (ω). By Markov’s inequality,
Bn ≤
2Jdiam4 (Ω)
√
,
nε
(45)
and from equations (43) and (45),
sup P
P∈P
!
1 n 4
10Jdiam4 (Ω)
4
√
d
(µ̂,Y
)
−
E(d
(µ,Y
))
>
ε
≤
→0
i
∑
n i=1
nε
This completes the proof.
45
as n → ∞.
(46)
(C). Note that
P V̂p −Vp > ε
1
n
=P
k
1
∑ ∑ d 2(µ̂ p,Yi) − d 2(µ p,Yi) + n
j=1 i∈G j
≤ An + Bn ,
k
!
d 2 (µ p ,Yi ) − ∑ λ j E j d 2 (µ p ,Y j ) > ε
k
∑∑
j=1 i∈G j
j=1
with
1
n
An = P
1
n
Bn = P
k
1
k
∑ ∑ d 2(µ̂ p,Yi) − n ∑ ∑ d 2(µ p,Yi)
j=1 i∈G j
>
j=1 i∈G j
k
k
ε
2
!
ε
∑ ∑ d 2(µ p,Yi) − ∑ λ j E j d 2(µ p,Y j ) > 2
j=1 i∈G j
j=1
,
!
.
Since infω∈Ω ∑kj=1 λ j E j (d 2 (ω,Y j ) − d 2 (µ p ,Y j )) = 0,
1
n
k
∑ ∑ d 2(µ̂ p,Yi) − d 2(µ p,Yi)
= inf
ω∈Ω
j=1 i∈G j
= inf
ω∈Ω
1
n
1
n
∑ ∑ d 2(ω,Yi) − d 2(µ p,Yi)
j=1 i∈G j
!
k
∑ ∑d
2
2
!
k
k
λ j E j (d 2 (ω,Y j ) − d 2 (µ p ,Y j ))
∑
ω∈Ω
(ω,Yi ) − d (µ p ,Yi ) − inf
j=1 i∈G j
j=1
≤ sup |Hn (ω)| ,
ω∈Ω
where
1
Hn (ω) =
n
k
k
∑ ∑ {d
j=1 i∈G j
2
2
(ω,Yi ) − d (µ p ,Yi )} − ∑ λ j {E j (d 2 (ω,Y j )) − E j (d 2 (µ p ,Y j ))}
j=1
k
k
= ∑ λ j,n Mn j (ω) + ∑ λ j,n − λ j {E j (d 2 (ω,Y j )) − E j (d 2 (µ p ,Y j ))}
j=1
j=1
46
and
1
∑ {d 2(ω,Yi) − d 2(µ p,Yi) − E j (d 2(ω,Y j )) + E j (d 2(µ p,Y j ))}.
n j i∈G
j
Mn j (ω) =
Using similar arguments as in the proofs of (A) and (B) for the individual groups, we
control the behavior of Hn (ω) by defining function classes {d 2 (ω, y) − d 2 (µ p , y) : ω ∈ Ω},
to obtain for some constants a1 , a2 > 0,
E
sup |Hn (ω)|
ω∈Ω
k
≤
∑ λ j,nE
ω∈Ω
j=1
k
≤
∑ λ j,n
j=1
k
sup |Mn j (ω)| + ∑ λ j,n − λ j E j (d 2 (ω,Y j )) − E j (d 2 (µ p ,Y j ))
j=1
k
k
a1
2Jdiam2 (Ω)
+ 2diam2 (Ω) ∑ λ j,n − λ j ≤ √ + a2 ∑ λ j,n − λ j .
√
nj
n
j=1
j=1
By Markov’s inequality,
2
An ≤
ε
k
a1
√ + a2 ∑ λ j,n − λ j
n
j=1
!
.
(47)
Next observe that
1
n
k
∑∑
j=1 i∈G j
k
d 2 (µ p ,Yi ) − ∑ λ j E j d 2 (µ p ,Y j )
k
k
≤
j=1
∑ λ j,nKn j (µ p) +
j=1
∑ (λ j,n − λ j )E j
d 2 (µ p ,Y j )
j=1
k
k
≤ ∑ λ j,n sup Kn j (ω) + ∑ (λ j,n − λ j ) diam2 (Ω),
j=1
where Kn j (ω) =
1
nj
ω∈Ω
j=1
∑i∈G j d 2 (ω,Yi ) − E j (d 2 (ω,Y j )).
By similar arguments as before, E supω∈Ω |Kn j (ω)| ≤
47
Jdiam2 (Ω)
√
.
nj
By Markov’s inequal-
ity, for some constants b1 , b2 > 0,
2
Bn ≤
ε
k
b
√1 + b2 ∑ λ j,n − λ j
n
j=1
!
.
(48)
From equations (47) and (48),
2
sup P V̂p −Vp > ε ≤
ε
P∈P
k
(a1 + b1 )
√
+ (a2 + b2 ) ∑ λ j,n − λ j
n
j=1
which completes the proof of (C).
48
!
→0
as n → ∞,
| 10math.ST
|
arXiv:1709.04066v1 [math.GR] 12 Sep 2017
DEHN FUNCTIONS OF SUBGROUPS OF RIGHT-ANGLED
ARTIN GROUPS
NOEL BRADY AND IGNAT SOROKO
Abstract. We show that for each positive integer k there exist right-angled
Artin groups containing free-by-cyclic subgroups whose monodromy automorphisms grow as nk . As a consequence we produce examples of right-angled
Artin groups containing finitely presented subgroups whose Dehn functions
grow as nk`2 .
Contents
1. Introduction
2. Preliminaries on growth
3. Bounding the Dehn function of the Bieri double
4. Cube complexes
4.1. Special cube complexes
4.2. Morse theory for cube complexes
5. Groups Gm,k
5.1. LOG definition
5.2. CATp0q structure for Gm,k
5.3. Free-by-cyclic structure
6. Constructing a special cover for Gm,m
6.1. The permutation representation
6.2. A p2m ` 1q–torus cover
6.3. Exploring hyperplane pathologies
7. Proof of Main Theorems
8. Growth of φ and φ´1
8.1. Upper bounds for the growth of φ, φ´1
8.2. Lower bounds for the growth of φ, φ´1
8.3. Lower bounds for the growth of φab
9. Open questions
References
1
4
8
12
12
13
14
14
15
16
19
19
23
25
28
30
31
35
36
37
38
1. Introduction
There has been intense interest in subgroups of right-angled Artin groups (RAAGs)
in recent years. This is due largely to the work of Ian Agol, Dani Wise and others
on the virtual fibering question in 3–manifold topology.
Date: September 14, 2017.
2010 Mathematics Subject Classification. Primary 20E05, 20F65, 20F67, 57M20.
1
2
NOEL BRADY AND IGNAT SOROKO
In [1] Agol showed that if M 3 is a compact, oriented, irreducible 3–manifold with
χpM q “ 0 and π1 pM q is a subgroup of a RAAG, then M virtually fibers. In [21]
Haglund and Wise showed that fundamental groups of special cubical complexes
are subgroups of RAAGs. Building on this machinery, Agol went on to solve the
virtual fibering conjecture in [2]. The fundamental result in [2] is that non-positively
curved cubical complexes with hyperbolic fundamental groups are virtually special.
In this paper we consider two questions about subgroups of RAAGs.
The first question asks which free-by-cyclic groups virtually embed in RAAGs.
In [19, 20] Hagen and Wise show that hyperbolic free-by-cyclic groups virtually
embed in RAAGs. It is also known that F2 ¸ Z groups virtually embed in RAAGs.
The hyperbolic free-by-cyclic examples all have exponentially growing monodromy
automorphisms and the F2 ¸ Z groups have exponential or linear monodromy automorphisms. In [18], Gersten gives an explicit example of an F3 ¸ Z group which
does not virtually embed in a RAAG. The group considered by Gersten is not a
CAT(0) group, and this prompts the following open question. Does every CAT(0)
free-by-cyclic group virtually embed in a RAAG? The family of the so-called Hydra
groups considered in [17] provides a test case for this question where the monodromy
automorphisms grow polynomially with arbitrary degree. While we haven’t proved
that Hydra groups virtually embed in RAAGs (this is a topic of an ongoing research
of the second-named author), we construct analogues of the Hydra groups (where
the base Z2 subgroup is replaced by a more complicated RAAG) which are CAT(0)
free-by-cyclic with polynomially growing monodromy automorphisms of arbitrary
degree and which are virtually special.
Theorem A. For each positive even integer m there exist virtually special free-bycyclic groups Gm,m – F2m ¸φ Z with growth function grφ pnq „ nm and Gm,m´1 –
F2m´1 ¸φ1 Z with growth function grφ1 pnq „ nm´1 .
Since finite index subgroups of free-by-cyclic groups are again free-by-cyclic and
since special groups embed into RAAGs we obtain the following corollary.
Corollary A. For each positive integer k there exist a right-angled Artin group
containing a free-by-cyclic subgroup whose monodromy automorphism has growth
function „ nk .
The second question asks what kinds of functions arise as Dehn functions of
finitely presented subgroups of RAAGs. Recall that Dehn functions capture the
isoperimetric behavior of Cayley complexes of groups. A lot is known about Dehn
functions of arbitrary finitely presented groups (see [8, 5, 26]). For example, a
group is hyperbolic if and only if its Dehn function is linear. CAT(0) groups and, in
particular, RAAGs, have Dehn functions which are either quadratic or linear. In [6]
there are examples of CAT(0) groups which contain finitely presented subgroups
whose Dehn functions are of the form nα for a dense set of α P r2, 8q. Restricting
to the case where the ambient groups are RAAGs it gets harder to find examples
of subgroups with a wide variety of Dehn functions. In [7] there are examples of
finitely presented Bestvina–Brady kernels of RAAGs which have polynomial Dehn
functions of degree 3 or 4. In [16] it is shown that the Dehn function of such kernels
of RAAGs are at most quartic. In [11] Bridson provides an example of a RAAG
containing a finitely presented group with exponential Dehn function. Our second
result shows that there are finitely presented subgroups of RAAGs whose Dehn
functions are polynomial of arbitrary degree.
DEHN FUNCTIONS OF SUBGROUPS OF RIGHT-ANGLED ARTIN GROUPS
3
Theorem B. For each positive integer k there exists a right-angled Artin group
which contains a finitely presented subgroup with Dehn function » nk .
This is the extent of what is currently known about the isoperimetric behavior
of subgroups of RAAGs; it would be very interesting if one could produce examples
with other types of Dehn functions.
Our paper is organized as follows.
In section 2 we introduce the growth functions of automorphisms and prove
a folklore result (Proposition 2.4) that the growth of an automorphism of a free
group is invariant under taking powers of the automorphism and under passing to
a subgroup of finite index. In doing that, we rely on the Gilbert Levitt’s Growth
Theorem [24] (whose proof uses train-track machinery). We also provide an example
due to Yves Cornulier which demonstrates that for arbitrary groups the invariance
of the growth function under passing to a subgroup of finite index does not hold.
Section 3 is devoted to providing estimates for the Dehn function of the Bieri
double of a free-by-cyclic group in terms of the growth of the monodromy automorphism. We use Bridson’s lower bound from [12] (Proposition 3.3) and adapt the
proof of the upper bound, given in [15] for the abelian-by-cyclic setting, to the case
of free-by-free groups needed for our construction (Proposition 3.4). Using these
estimates, we show later in section 7 that for the polynomially growing monodromy
automorphisms involved in our construction, the upper and lower bounds on the
Dehn function of the Bieri double actually coincide. If the monodromy automorphism has polynomial growth of order nk , then the Dehn function of the Bieri
double grows like nk`2 .
In section 4 we recollect all the relevant definitions related to the Morse theory
on groups and special cubical complexes, which will be used in sections 5 and 6.
In section 5 we introduce the free-by-cyclic groups Gm,k , which play the central
role in our construction. We define these groups through LOG notation, which is
a graphical tool to encode conjugation relations. We prove that the group Gm,k is
CAT(0) and free-by-cyclic, and exhibit explicit formulas for its monodromy automorphism (Proposition 5.4). We defer until section 8 the proof that this automorphism has growth „ nk .
The goal of section 6 is to exhibit a finite special cover for the presentation
complex Km,m of the group Gm,m , for arbitrary even m. The construction is done
in several stages. First, for arbitrary m, we engineer a certain right action of
Gm,m on a set of cardinality 22m`1 , which may be thought of as the 0–skeleton
pm Ñ
of a p2m ` 1q–dimensional torus T2m`1 . This action defines a finite cover K
Km,m , which cellularly embeds into the 2–skeleton of T2m`1 (Proposition 6.2).
p m is a subcomplex of a product of graphs, it is free from three out of four
Since K
hyperplane pathologies in the definition of a special cube complex (Proposition 6.4).
To eliminate the fourth hyperplane pathology we observe that for even values of m,
the complex Km,m is a V H-complex in the terminology of [21]. It follows that there
p m , such that K m is a special square complex
exist another finite cover K m Ñ K
(Proposition 6.6).
In section 7 we bring all the pieces together and prove Theorems A and B. For
even values of m, groups Gm,m are virtually special free-by-cyclic with the monodromy automorphism growing as nm . To obtain growth functions of odd degree,
we observe that the presentation 2–complex Km,m´1 of the free-by-cyclic group
4
NOEL BRADY AND IGNAT SOROKO
Gm,m´1 is a combinatorial subcomplex of Km,m´1 and it is obtained by deleting
the hyperplane corresponding to the last generator a2m`1 . Thus the pullback of
Km,m´1 in K m is a finite special square complex covering Km,m´1 . This makes
Gm,m´1 a virtually special free-by-cyclic group with the monodromy automorphism
growing as nm´1 .
To prove theorem B, we look at the Bieri double of the special (free-by-cyclic)
finite index subgroups H of Gm,m (Gm,m´1 ), and prove that the lower and the
upper bounds for its Dehn function coincide, and are of the order nm`2 (resp.,
nm`1 ). This Bieri double naturally embeds into a RAAG, whose underlying graph
is the join of the underlying graph for the RAAG containing H and the empty
graph on two vertices.
In section 8 we provide the computation of the growth function for the monodromy automorphism of Gm,k and its abelianization.
Finally, in section 9 we list two open questions related to the study in this paper.
2. Preliminaries on growth
In what follows we will consider functions up to the following equivalence relations.
Definition 2.1. Two functions f, g : r0, 8q Ñ r0, 8q are said to be „ equivalent if
f ĺ g and g ĺ f , where f ĺ g means that there exist constants A ą 0 and B ě 0
such that f pnq ď Agpnq ` B for all n ě 0.
Definition 2.2. Two functions f, g : r0, 8q Ñ r0, 8q are said to be » equivalent
if f ă
“ g and g ă
“ f , where f ă
“ g means that there exist constants A, B ą 0 and
C, D, E ě 0 such that f pnq ď AgpBn ` Cq ` Dn ` E for all n ě 0.
We extend these equivalence relations to functions N Ñ r0, 8q by assuming them
to be constant on each interval rn, n ` 1q.
Remark 2.3. Notice that f ĺ g implies f ă
“ g and f „ g implies f » g. However,
the relation „ is strictly finer than », as the latter identifies all single exponential
functions, i.e. k n » K n for k, K ą 1, whereas the relation „ does not. We will use
the relations „, ĺ when dealing with growth functions of automorphisms and the
relations », ă
“ when discussing Dehn functions of groups (as is done traditionally).
The term Dn in the above definition of ă
“ is essential for proving the equivalence
of Dehn functions under quasi-isometries.
Let F be a free group of finite rank k with a finite generating set A. Let dA px, yq
be the associated word metric on F . If ψ : G Ñ G is an automorphism, we define
grψ,A pnq :“ max }ψ n paq}A ,
aPA
where }g}A is equal to dA p1, gq for g P F .
The following properties of grψ,A will be used in the sequel.
Proposition 2.4. Let F be a free group of finite rank with a finite generating set
A, and let ψ be an automorphism of F . Then
(i) for each finite generating set B of F , grψ,B „ grψ,A ;
(ii) for each d P N, grψ,A „ grψd ,A ;
(iii) for each finite index subgroup H ď F invariant under ψ with a finite generating set B Ă H, we have grψ,A „ grψ|H ,B .
DEHN FUNCTIONS OF SUBGROUPS OF RIGHT-ANGLED ARTIN GROUPS
5
Proof. Let A “ ta1 , . . . , aN u, B “ tb1 , . . . , bM u. Since both sets A and B generate F , there exist words wi and vj such that ai “ wi pb1 , . . . , bM q and bj “
vj pa1 , . . . , aN q, for all 1 ď i ď N , 1 ď j ď M . Let constants K and L denote the maximal lengths of wi , vj , respectively, i.e. K “ max1ďiďN }wi }B ,
L “ max1ďjďM }vj }A . Then, obviously, for all i, j, n, one has:
}ψ n pai q}B ď K ¨ }ψ n pai q}A
}ψ n pbj q}A ď L ¨ }ψ n pbi q}B .
and
Now fix arbitrary 1 ď j ď M and assume without loss of generality that
vj pa1 , . . . , aN q “ aεi11 . . . aεiLL , for values 1 ď i` ď N and ε` “ ˘1 or 0. Then
}ψ n pbj q}B “ }ψ n paεi11 . . . aεiLL q}B ď }ψ n pai1 q}B ` ¨ ¨ ¨ ` }ψ n paiL q}B ď
K}ψ n pai1 q}A ` ¨ ¨ ¨ ` K}ψ n paiL q}A ď KL max }ψ n pai q}A “ KL grψ,A pnq.
1ďiďN
n
Therefore, grψ,B pnq “ max1ďjďM }ψ pbj q}B ď KL grψ,A pnq, and, by symmetry,
grψ,A pnq ď LK grψ,B pnq. This proves (i).
Before proving parts (ii) and (iii), we state following remarkable result of Gilbert
Levitt:
Levitt’s Growth Theorem ([24, Cor. 6.3]). Let F be a free group of finite rank
with a free generating set A. Given α P AutpF q and g P F , there exist λ ě 1, an
integer m ě 0 and constants A, B ą 0 such that the word length }αn pgq}A satisfies:
@n P N,
Aλn nm ď }αn pgq}A ď Bλn nm .
Consider a sequence of growth parameters pλi , mi q from the Levitt’s Growth
Theorem corresponding to the generators a1 , . . . , aN , so that for each 1 ď i ď N
there exist constants Ai , Bi ą 0 such that
Ai λni nmi ď }ψ n pai q}A ď Bi λni nmi
for all n P N.
Order these parameters lexicographically: pλi , mi q ă pλj , mj q if and only if λi ă λj
or λi “ λj and mi ă mj . Clearly, pλi , mi q ă pλj , mj q if and only if λnj nmj {λni nmi Ñ
8 as n Ñ 8. Pick 1 ď i0 ď N such that pλi0 , mi0 q is maximal with respect to this
order. Then for any constants C1 , C2 ą 0 and arbitrary 1 ď i ď N we have:
C1 λni nmi ! C2 λni0 nmi0 ,
which means that the left-hand side is less than or equal to the right-hand side for
all large enough n P N.
To prove (ii) in one direction, notice that for any 1 ď i ď N ,
˘
B λd dmi `
mi
}pψ d qn pai q}A ď Bi λdn
“ pBi λdi dmi qλni nmi ! i Aii
Ai0 λni0 nmi0 ď
i pdnq
0
mi
Bi λd
id
Ai0
}ψ n pai0 q}A .
Hence, there exist a constant Cbig ě 0 such that
grψd ,A pnq “ max }pψ d qn pai q}A ď D grψ,A pnq ` Cbig ,
1ďiďN
where D “ max1ďiďN Bi λdi dmi {Ai0 . Thus, grψd ,A ĺ grψ,A .
In the opposite direction, for any 1 ď i ď N we have:
`
˘
mi0
mi
}ψ n pai q}A ď Bi λni nmi ď Bi λdn
! ABii Ai0 λdn
ď
i0 pdnq
i pdnq
0
Bi
Ai0
}ψ dn pai0 q}A .
6
NOEL BRADY AND IGNAT SOROKO
By taking maximum, we get for arbitrary n P N:
grψ,A pnq “ max }ψ n pai q}A ď D}ψ dn pai0 q}A ` Cbig ď D grψd ,A pnq ` Cbig
1ďiďN
for D “ maxi Bi {Ai0 and some Cbig ě 0. This proves that grψ,A ĺ grψd ,A and
hence that grψ,A „ grψd ,A .
To prove (iii) in one direction, notice first that H, being of finite index in F ,
is quasi-convex in F (see e.g. [14, III.3.5]). Hence for any h P H, one has }h}B ď
C}h}A for some C ą 0. Writing each bj P B as a word bj “ vj pa1 , . . . , aN q and
setting L “ max1ďjďM }vj }A , we obtain for arbitrary 1 ď j ď M :
}ψ n pbj q}B ď C}ψ n pbj q}A ď CL max }ψ n pai q}A “ CL grψ,A pnq,
1ďiďN
n
so that grψ|H ,B pnq “ max1ďjďM }ψ pbj q}B ď CL grψ,A pnq, i.e. grψ|H ,B ĺ grψ,A .
In the opposite direction, notice that there exist an integer p ą 0 such that for
every ai P A, we have api P H. As above, let pλi , mi q, Ai , Bi ą 0 be a sequence
of growth parameters for the generators a1 , . . . , aN , and let pλi0 , mi0 q be maximal.
Consider a new generating set B 1 for H, B 1 “ B Y tapi0 u. Then for arbitrary 1 ď
i ď N we have:
`
˘
}ψ n pai q}A ď Bi λni nmi ! ABii Ai0 λni0 nmi0 ď ABii }ψ n pai0 q}A ď
0
0
Bi
Ai0
}ψ n papi0 q}A ď
Bi
Ai0
L}ψ n papi0 q}B1 ,
where L has a similar meaning as above. Here the fourth inequality holds since,
in general, for any automorphism α P AutpF q, any g P F and any p ą 0, one
has }αpgq}A ď }αpg p q}A . (Indeed, one can write αpgq “ uvu´1 with v cyclically
reduced. Then }αpgq} “ 2}u} ` }v}, whereas }αpg p q} “ }uv p u´1 } “ 2}u} ` p}v}.)
By taking maximum, we get for arbitrary n P N:
grψ,A pnq “ max }ψ n pai q}A ď DL}ψ n papi0 q}B1 ` Cbig ď DL grψ|H ,B1 pnq ` Cbig
1ďiďN
for D “ maxi Bi {Ai0 and some Cbig ě 0. This means that grψ,A ĺ grψ|H ,B1 . Since,
by part (i), grψ|H ,B1 „ grψ|H ,B , this proves that grψ,A ĺ grψ|H ,B and part (iii) is
proved.
Remark 2.5. The proof of property (i) of Proposition 2.4 works for automorphisms
of arbitrary finitely generated groups.
Remark 2.6. Property (iii) of Proposition 2.4 does not hold for arbitrary finitely
generated groups. We are grateful to Yves Cornulier for providing the following
example. Let G “ xa, b | aba´1 “ b´1 , a2 “ 1y be the infinite dihedral group. Then
the inner automorphism ib : x ÞÑ bxb´1 has linear growth, but its restriction to
the index 2 subgroup xby is trivial. To see that, observe that aba´1 “ b´1 implies
ab “ b´1 a and hence ba “ ab´1 . Thus bab´1 “ ab´2 and inb paq “ bn ab´n “ ab´2n .
Looking at the Cayley graph of G shows that the element g “ ab´2n is at distance
2n ` 1 from 1, so that ab´2n is a word of minimal length representing element g,
and ib indeed grows linearly on G.
In view of item (i) in Proposition 2.4, we will suppress the dependence on the
generating set and adopt the notation
grψ pnq :“ grψ,A pnq,
for an arbitrary generating set A Ă F .
DEHN FUNCTIONS OF SUBGROUPS OF RIGHT-ANGLED ARTIN GROUPS
7
For the abelianization Fab “ F {rF, F s – Zk we consider the induced automorphism ψ ab : Fab Ñ Fab and denote tēi u be the generating set of Fab corresponding
to A: ēi “ ai rF, F s, ai P A, i “ 1, . . . , k. For any v P Zk let |v|1 denote the `1 -norm
řk
řk
on Zk viewed as a subset of Ck : if v “ i“1 ci ēi , then |v|1 “ i“1 |ci |. Define
grψab pnq :“ max |pψ ab qn pēi q|1 .
i“1,...,k
Then the following is true:
Lemma 2.7.
grψ pnq ě grψab pnq.
Proof. Let ε : F Ñ Fab be the natural homomorphism. Then
pψ ab qn pēi q “ εpψ n pai qq
and hence the length of the shortest word in generators tēi uki“1 of the element
pψ ab qn pēi q P Zk is no bigger than }ψ n pai q}A . But the former is equal to |pψ ab qn pēi q|1
hence |pψ ab qn pēi q|1 ď }ψ n pai q}A for all i “ 1, . . . , k. By taking maximum, we get
the required inequality.
By embedding Zk into Ck we may consider Ck as a vector space with the basis
tēi uki“1 . Now let A be a linear operator on Ck given in basis tēi uki“1 by the matrix
řk
paij qki,j“1 , and let |v|8 denote the `8 -norm on Ck : if v “ i“1 ci ēi , then |v|8 “
maxi“1,...,k |ci |. Consider two norms on EndpCk q, one is the operator norm with
respect to `8 :
|Av|8
}A}op “ sup
“ sup |Av|8 ,
v‰0 |v|8
|v|8 “1
2
and another one is the supremum norm, which is the `8 -norm on the space Ck :
}A}sup “
max
i,j“1,...,k
|aij |.
Lemma 2.8.
max |Aēi |1 ě }A}sup .
i“1,...,k
Proof.
max |Aēi |1 ě max |Aēi |8 “ max |aij | “ }A}sup .
i
i
i,j
The following fact is well-known (see [22, Cor. 5.4.5]):
Lemma 2.9. There exist constants C1 , C2 ą 0 such that
C1 }A}op ď }A}sup ď C2 }A}op .
Corollary 2.10. The growth function
n
grsup
A : n ÞÝÑ }A }sup
is „ equivalent to the growth function
n
grop
A : n ÞÝÑ }A }op .
The following results are proved in [13, Proof of Th. 2.1]:
Lemma 2.11. The „ equivalence class of the function grop
A depends only on the
conjugacy class of A in GLpk, Cq.
8
NOEL BRADY AND IGNAT SOROKO
In view of Corollary 2.10 and Lemma 2.11, we need only to consider the growth
of the Jordan normal forms of matrices A.
Lemma 2.12 ([13, Th. 2.1]). Suppose that J is a matrix in the Jordan normal
c´1
form with all eigenvalues equal to 1. Then grsup
, where c is the maximal
J pnq „ n
size of Jordan blocks of J.
Combining all of the above, we get:
Corollary 2.13. Let ψ be an automorphism of a free group F . If the abelianization
ψ ab has all eigenvalues equal to 1, and c is the size of the largest Jordan block in
the Jordan normal form J for ψ ab , then
grψ pnq ľ nc´1
and
grψab pnq ľ nc´1 .
Proof. Indeed,
grψ pnq ě grψab pnq
(by Lemma 2.7)
ab n
(by Lemma 2.8)
ě }pψ q }sup
ab n
ě C1 }pψ q }op
(for some C1 ą 0, by Lemma 2.9)
n
(by Lemma 2.11)
c´1
(by Lemma 2.12).
„ }J }op
„n
3. Bounding the Dehn function of the Bieri double
In this section we outline what is known about the upper and the lower bounds
for the Dehn function of a Bieri double group. The lower bound was established
in [12, Lemma 1.5] (see Proposition 3.3 below). The argument for the upper bound
(see Proposition 3.4 below) follows the outline of [15, Theorem 5.1]. In the latter
paper the argument is given in the setting of abelian-by-cyclic groups; we adapt
this reasoning to the free-by-free setting.
Definition 3.1. (Bieri double) Let G be a free-by-cyclic group G “ F ¸ψ Z. The
Bieri double of G is the group ΓpGq “ G ˚F G.
If G – xA, t | tat´1 “ ψpaq for all a P Ay then ΓpGq – xA, s, t | sas´1 “
ψpaq, tat´1 “ ψpaq for all a P Ay. If one denotes F ps, tq the free group on the
generating set ts, tu and pψq : F ps, tq Ñ AutpF pAqq the homomorphism given on
the generators by s ÞÑ ψ, t ÞÑ ψ, then ΓpGq – F pAq ¸pψq F ps, tq.
Definition 3.2. (Dehn function) Let a group Γ be given by a finite presentation
P “ xA | Ry. For each word w lying in the normal closure of R in the free group
F pAq, define
Areapwq :“ min N | w “
F pAq
N
ź
(
˘
x´1
.
i ri xi with xi P F pAq, ri P R
i“1
The Dehn function of P is the function δP : N Ñ N defined by
δP pnq :“ maxtAreapwq | w “ 1, }w}A ď nu.
Γ
where }w}A denotes the length of the word w in generators A˘ .
Viewed up to » equivalence, the Dehn functions are independent of the choice
of the presentation (see [10, 1.3.3]), so we denote δP pnq as δΓ pnq.
DEHN FUNCTIONS OF SUBGROUPS OF RIGHT-ANGLED ARTIN GROUPS
9
Proposition 3.3 ([12, Lemma 1.5] and [10, Proposition 7.2.2]). Let ψ be an automorphism of F and }.} denote the word length with respect to a fixed generating set
of F . Then for the Dehn function δΓ pnq of the Bieri double Γ of F ¸ψ Z one has
n ¨ max }ψ n pbq} ă
“ δΓ pnq.
}b}ďn
bPF
Proposition 3.4. Let ψ be an automorphism of a free group F and assume that
grψ pnq ĺ nd and grψ´1 pnq ĺ nd . Then for the Dehn function δΓ pnq of the Bieri
d`2
double Γ “ ΓpF ¸ψ Zq one has δΓ pnq ă
“n .
Remark 3.5. It can be proved using train-tracks that grψ´1 pnq „ nd if and only
if grψ pnq „ nd (see e.g. [25, Th. 0.4]). However, for the reader who is unfamiliar
with the train-track machinery we make the exposition independent of this result.
Instead, in what follows we will apply Proposition 3.4 to the automorphisms φ
whose growth functions grφ pnq and grφ´1 pnq are computed in section 8 and are
shown to be „ equivalent to each other.
In order to prove Proposition 3.4, we need some preliminary results on combings
of groups. We start with some definitions from [9] and [15].
Let Γ be a group with finite generating set A and dA px, yq be the associated
word metric.
Definition 3.6. A combing (normal form) for Γ is a set of words tσg | g P Γu in
the letters A˘ such that σg “ g in Γ. We denote by |σg | or |σg |A the length of the
word σg in the free monoid on A˘ .
Definition 3.7. Let
R “ tρ : N Ñ N | ρp0q “ 0; ρpn ` 1q P tρpnq, ρpnq ` 1u @n; ρ unbounded u.
Given eventually constant paths p1 , p2 : N Ñ pΓ, dq we define
Dpp1 , p2 q “ min
1
ρ,ρ PR
(
maxtdA pp1 pρptqq, p2 pρ1 ptqqu .
tPN
Definition 3.8. Given a combing σ for Γ, the asynchronous width of σ is the
function Φσ : N Ñ N defined by
(
Φσ pnq “ max Dpσg , σh q | dA p1, gq, dA p1, hq ď n; dA pg, hq “ 1 .
Definition 3.9. A finitely generated group Γ is said to be asynchronously combable
if there exists a combing σ for Γ and a constant K ą 0 such that Φσ pnq ď K for
all n P N.
Definition 3.10. The length of a combing σ for Γ is the function L : N Ñ N given
by:
(
Lpnq “ max |σg | | dA p1, gq ď n .
The relation of combings to Dehn functions is manifested in the following result:
Proposition 3.11 ([15, Lemma 4.1]). Let Γ be a group with a finite set of semigroup
generators A˘ . If there exists a combing σ for Γ whose asynchronous width is
bounded by a constant and whose length is bounded by the function Lpnq, then the
Dehn function δΓ pnq for any presentation of Γ satisfies δΓ pnq ă
“ nLpnq.
In connection to the groups which are Bieri doubles, the following result from [9]
is useful.
10
NOEL BRADY AND IGNAT SOROKO
Theorem 3.12 ([9, Theorem B]). If G is word-hyperbolic and H is asynchronously
combable then every split extension
1 ÝÑ G ÝÑ G ¸ H ÝÑ H ÝÑ 1
of G by H is asynchronously combable.
Remark 3.13. From the proof of this result in [9] it follows that if groups G
and H have combings σ G and σ H whose asynchronous width is bounded by some
constants, then the combing for the split extension G ¸ H of G by H, whose length
is bounded by a constant, can be taken as the product (concatenation) σ H σ G of
combings σ H and σ G , meaning that we traverse path σ H first, then path σ G . Note
that the product of combings in the opposite order, σ G σ H , may not have bounded
asynchronous width, as the example of Baumslag–Solitar groups shows.
Now let again Γ “ F ¸pψq F ps, tq be the Bieri double of G “ F ¸ψ Z, where F
is a free group on the set of free generators A.
Our goal is to obtain an upper bound on the length Lpnq of the combing σ F ps,tq ¨
F pAq
σ
in terms of the growth of the automorphism ψ. (Here we treat a combing
on a free group as a unique reduced word in a fixed system of generators which
represents the given element of the group.) We prove the following proposition,
adapting the reasoning for the abelian-by-cyclic groups from [15, Theorem 5.1] to
the case of free-by-free groups.
Proposition 3.14. Let P pnq be an increasing function bounding the growth of both
ψ and ψ ´1 , i.e. dA p1, ψ n paqq ď P p|n|q for all a P A, n P Z. Then the length Lpnq
of the combing σ F ps,tq ¨ σ F pAq of the group Γ “ F pAq ¸pψq F ps, tq satisfies
Lpnq ď nP pnq ` n.
Proof. Take arbitrary γ P Γ and write it as γ “ u ¨ g, where u P F ps, tq, g P F pAq.
Let n0 “ dAYts,tu p1, γq be the length of the shortest word in generators pAYts, tuq˘
representing element γ in Γ. We would like to show that
|σγ | “ |σu ¨ σg | “ dts,tu p1, uq ` dA p1, gq ď n0 P pn0 q ` n0 .
Considering the natural homomorphism η : F pAq ¸pψq F ps, tq Ñ F ps, tq, one observes that u “ ηpγq and hence dts,tu p1, uq ď n0 . Therefore it suffices to show
that
dA p1, gq ď n0 P pn0 q.
Denote w0 the shortest word in generators pA Y ts, tuq˘ such that w0 “ γ in Γ,
so that |w0 |AYts,tu “ n0 . Then w0 can be written as
w0 “ u1 w1 u2 w2 ¨ ¨ ¨ ¨ ¨ ur wr ,
where ui P F ps, tq, wi P F pAq for all i. Then
r
r
ÿ
ÿ
n0 “ |w0 |AYts,tu “
|ui |ts,tu `
|wi |A
i“1
and u “ u1 . . . ur .
Denote
vi “
i
ź
j“1
uj ¨ wi ¨
i
´ź
j“1
i“1
¯´1
uj
,
i “ 1, . . . , r.
(*)
DEHN FUNCTIONS OF SUBGROUPS OF RIGHT-ANGLED ARTIN GROUPS
11
Then, as one easily checks,
v1 v2 . . . vr “ u1 w1 u2 w2 . . . ur wr ¨
i
´ź
¯´1
uj
j“1
so that
γ “ w0 “ v1 v2 . . . vr ¨
i
´ź
¯
uj .
j“1
Hence
g “ u´1 γ “
r
´ź
¯´1
uj
¨ v1 v2 . . . vr ¨
j“1
¯
uj “
j“1
r ”´ ź
r
ź
i“1
r
´ź
j“1
i
i
r
¯´1 ź
´ź
¯´1 ź
ı
uj
¨
uj ¨ wi ¨
uj
¨
uj “
j“1
j“1
r
ź
`
j“1
˘
`
˘
´1
´1
u´1
r ur´1 . . . ui`1 ¨ wi ¨ ui`1 . . . ur .
i“1
If we denote by ε : F ps, tq Ñ Z the homomorphism defined on the generators as:
s ÞÑ 1, t ÞÑ 1, then for any g P F pAq and any u P F ps, tq we have ugu´1 “ ψ εpuq pgq.
Therefore,
r
r
řr
ź
ź
g“
ψ ´εpui`1 ...ur q pwi q “
ψ ´ j“i`1 εpuj q pwi q.
i“1
i“1
On the other hand,
r
r
r
r
ˇ ÿ
ˇ
ÿ
ÿ
ÿ
ˇ
ˇ
εpuj qˇ ď
|εpuj q| ď
|εpuj q| ď
|uj |ts,tu “ |u|ts,tu ď n0
ˇ
j“i`1
j“i`1
j“1
(**)
j“1
by the observation above.
Moreover, since for any a P A we have dA p1, ψ n paqq ď P p|n|q, then for any
wi P F pAq we get
`
˘
dA 1, ψ n pwi q ď P p|n|q ¨ dA p1, wi q.
Finally, we get for the element g the estimate:
r
r
r
ˇ¯
´ˇ ÿ
řr
` ˘ ÿ
`
˘ ÿ
ˇ
ˇ
dA 1, g ď
dA 1, ψ ´ j“i`1 εpuj q pwi q ď
P ˇ
εpuj qˇ ¨ dA p1, wi q
i“1
r
ÿ
ď rby (**)s ď
i“1
i“1
P pn0 q ¨ dA p1, wi q “ P pn0 q ¨
r
ÿ
j“i`1
dA p1, wi q ď rby (*)s ď P pn0 qn0 .
i“1
This shows that |σγ | ď n0 P pn0 q ` n0 and finishes the proof of the Proposition.
Now we are ready to prove the upper bound for the Dehn function of the Bieri
double.
Proof of Proposition 3.4. As was noted above, Γ “ ΓpF ¸ψ Zq – F ¸pψq F ps, tq. As
a free group, F is asynchronously combable (with constant K “ 1) and F ps, tq is
also word-hyperbolic. Therefore by Theorem 3.12, Γ is asynchronously combable
and hence, by Proposition 3.11, δΓ pnq ă
“ nLpnq. But due to Proposition 3.14,
d`1
d`2
Lpnq ă
“ n , and therefore δΓ pnq ă
“n .
12
NOEL BRADY AND IGNAT SOROKO
4. Cube complexes
4.1. Special cube complexes. In their article [21] Haglund and Wise established
that the fundamental groups of the so-called special cube complexes admit embeddings into right-angled Artin groups. This gives us a natural class of subgroups of
right-angled Artin groups and suggests that we construct our examples within this
class. We summarize the relevant definitions and results from [21] about special
cube complexes in this section.
Definition 4.1. (Cube complex) An n–cube is a copy of r0, 1sn Ă Rn , viewed as
a metric space with the euclidean metric of Rn . (We will suppress ‘n–’ and call
‘n–cubes’ just ‘cubes’.) A face is a metric subspace of an n–cube obtained by
restricting some of coordinates (or all of them) to either 0 or 1. A cube complex is a
CW complex obtained by gluing n–cubes (of possibly varied dimensions n) together
along faces via isometries. If all cubes of a cube complex are 2–cubes, such cube
complex is called a square complex. A cube complex is simple if the link of every
vertex of is a simplicial complex. A simplicial complex is flag if any collection of
k`1 pairwise adjacent vertices spans a k–simplex. A cube complex is non-positively
curved if the link of each vertex is a flag simplicial complex.
Definition 4.2. (Hyperplane) A midcube of an n–cube r0, 1sn is a subset obtained
by restricting one of the coordinates to 21 . A hyperplane of a cube complex X is a
connected component of a new cube complex Y which is formed as follows:
‚ the cubes of Y are the midcubes of X;
‚ the restriction of a pk ` 1q–cell of X to a midcube of r0, 1sk defines the
attaching map of a k–cell in Y .
An edge a of X is dual to some hyperplane H if the midpoint of a is a vertex of H.
Definition 4.3. (Parallelism, Walls) Two oriented edges a, b of a cube complex X
are called elementary parallel if there is a square of X containing a and b and such
that the attaching map sends two opposite edges of r0, 1s ˆ r0, 1s with the same
orientation to a and b respectively. Define the parallelism on oriented edges of X
as the equivalence relation generated by elementary parallelism. An (oriented) wall
of X is a parallelism class of oriented edges. Note that every hyperplane H in X
defines a pair of oriented walls consisting of edges dual to H.
Now we describe four pathologies for interaction of hyperplanes in a cube complex, which are forbidden for special cube complexes.
Definition 4.4. (Self-intersection) A hyperplane H in X self-intersects, if it contains more than one midcube from the same cube of X.
Definition 4.5. (One-sided) A hyperplane H is two-sided if there exists a combinatorial map of CW complexes H ˆ r0, 1s Ñ X mapping H ˆ t 21 u identically to
H. (Recall that a cellular map f : X Ñ Y of CW complexes is combinatorial if
the restriction of f to each open cell of X is a homeomorphism onto its image.) A
hyperplane H in X is called one-sided if it is not two-sided.
Definition 4.6. (Self-osculating) A hyperplane H in X is self-osculating if there
are two edges a, b dual to H which do not belong to a common square of X but
share a common vertex. If in addition there is a consistent choice of orientation
on the edges dual to H which makes the common vertex for a, b their origin or
terminus, then the hyperplane H is called directly self-osculating.
DEHN FUNCTIONS OF SUBGROUPS OF RIGHT-ANGLED ARTIN GROUPS
13
Definition 4.7. (Inter-osculating) Two distinct hyperplanes H1 , H2 of X are interosculating if they intersect and there are edges a1 dual to H1 and a2 dual to H2
which do not belong to the same square of X but share a common vertex.
Definition 4.8. (Special cube complex) A non-positively curved cube complex is
called special if its hyperplanes are all two-sided, with no self-intersections, selfosculations or inter-osculations.
Definition 4.9. (Virtually special group) A group G is called special if there exists
a special cube complex X whose fundamental group is isomorphic to G. A group
G is virtually special if there exists a special cube complex X and a finite index
subgroup H ď G such that H is isomorphic to the fundamental group of X.
Definition 4.10. (Right-angled Artin group) Let ∆ be a simplicial graph. The
right-angled Artin group, or RAAG, associated to ∆, is a finitely presented group
Ap∆q given by the presentation:
Ap∆q “ xai P Verticesp∆q | rai , aj s “ 1 if pai , aj q P Edgesp∆qy.
Definition 4.11. (Salvetti complex) Given a right-angled Artin group Ap∆q, the
Salvetti complex associated to Ap∆q is a non-positively curved cube complex S∆
defined as follows. For each ai P Verticesp∆q let Sa1i be a circle endowed with
a structure of a CW complex havingśa single 0–cell and a single 1–cell. Let
n
1
n “ CardpVerticesp∆qq and let T “
j“1 Saj be an n–dimensional torus with
the product CW structure. For every full subgraph K Ă ∆ with VerticespKq “
tai1 , . . . , aik u define a k–dimensional torus TK as a Cartesian product of CW comśk
plexes: TK “ j“1 Sa1i and observe that TK can be identified as a combinatorial
j
subcomplex of T . Then the Salvetti complex associated with Ap∆q is
ď
(
S∆ “
TK Ă T | K a full subgraph of ∆ .
Thus S∆ has a single 0–cell and n 1–cells. Each edge pai , aj q P Edgesp∆q con´1
tributes a square 2–cell to S∆ with the attaching map ai aj a´1
i aj . And in general each full subgraph K Ă ∆ contributes a k–dimensional cell to S∆ where
k “ CardpVerticespKqq.
Theorem 4.12 ([21],Th. 4.2). A cube complex is special if and only if it admits a
local isometry into the Salvetti complex of some right-angled Artin group.
Since local isometries of CAT(0) spaces are π1 -injective, one gets the following
Corollary 4.13. The fundamental group of a special cube complex is isomorphic
to a subgroup of a right-angled Artin group.
4.2. Morse theory for cube complexes. We will use the following definitions
from [3, 4].
Definition 4.14. (Morse function) A map f : X Ñ R defined on a cube complex
X is a Morse function if
‚ for every cell e of X, with the characteristic map χe : r0, 1sm Ñ e, the
composition f χe : r0, 1sm Ñ R extends to an affine map Rm Ñ R and f χe
is constant only when dim e “ 0;
‚ the image of the 0–skeleton of X is discrete in R.
14
NOEL BRADY AND IGNAT SOROKO
Definition 4.15. (Circle-valued Morse function) A circle-valued Morse function
on a cube complex X is a cellular map f : X Ñ S 1 with the property that f lifts
to a Morse function between universal covers.
Definition 4.16. (Ascending and descending links) Suppose X is a cube complex,
f : X Ñ S 1 is a circle-valued Morse function and f˜: X̃ Ñ R is the corresponding
Morse function. Let v P X p0q and note that the link of v in X is naturally isomorphic
to the link of any lift ṽ of v in X̃. We say that a cell ẽ Ă X̃ contributes to the
ascending (respectively descending) link of ṽ if ṽ P ẽ and f˜|ẽ achieves its minimum
(respectively, maximum) value at ṽ. The ascending (respectively, descending) link
of v is then defined to be the subset of the link Lkpv, Xq naturally identified with
the ascending (respectively, descending) link of ṽ. Note that in the case when X is
a square complex, all ascending, descending and entire links are graphs.
The following characterization of free-by-cyclic groups was proven in [3] (see
also [23, Th. 10.1]).
Theorem 4.17 ([3], Proposition 2.5). If f : X Ñ S 1 is a circle-valued Morse
function on a 2–complex X all of whose ascending and descending links are trees,
then X is aspherical and π1 pXq is free-by-cyclic. This means that there is a short
exact sequence
1 ÝÑ Fm ÝÑ π1 pXq ÝÑ Z ÝÑ 1,
where the free group Fm is isomorphic to π1 pf ´1 pptqq, pt being any point on S 1 .
5. Groups Gm,k
In this section we define a sequence of groups Gm,k and study their presentation
complex. We show that it is a non-positively curved square complex and that the
groups are free-by-cyclic.
5.1. LOG definition. Recall that a labeled, oriented graph, or LOG, consists of a
finite, directed graph with labels on the vertices and edges satisfying the following:
the vertices have distinct labels, and the edge labels are chosen from the set of
vertex labels.
A LOG determines a finite presentation as follows. The set of generators is the
set of vertex labels. The set of relations is in one-to-one correspondence with the
set of edges; there is a relation of the form a´1 ua “ v for each oriented edge labeled
a from vertex u to vertex v.
Let m P N, m ě 1. For k “ 0, . . . , m, let Gm,k be a group defined by the LOG
presentation in the Figure 1:
i.e.
Gm,k “ x a1 , . . . , am`k`1 | rai , ai`1 s “ 1,
i “ 1, . . . , m,
a´1
m`j`1 aj am`j`1
“ am`j ,
j “ 1, . . . , k y.
Clearly Gm,k is an HNN extension of Gm,k´1 with the stable letter am`k`1 so
there is a natural tower of inclusions
Gm,0 Ă Gm,1 Ă Gm,2 Ă ¨ ¨ ¨ Ă Gm,m .
DEHN FUNCTIONS OF SUBGROUPS OF RIGHT-ANGLED ARTIN GROUPS
a2
...
a3
ak`1
ak`2
a1
a2
ak
am`2
am`3
am`k`1
am`1
am`2
am`k
...
15
am`1
ak`1
am
am`k`1
Figure 1. The LOG description of Gm,k .
aj
aj`1
a`
j
a`
j`1
a´
j
a´
j`1
aj`1
aj
am`j
am`j`1
a`
m`j
am`j`1
a´
m`j
a`
j
aj
a´
j
a`
m`j`1
a´
m`j`1
Figure 2. The contribution of the relations a´1
j`1 aj aj`1 “ aj ,
1 ď j ď m, and a´1
a
a
“
a
,
1
ď
j
ď k, to the link
m`j
m`j`1 j m`j`1
of the 0–cell of the presentation complex Km,k .
5.2. CATp0q structure for Gm,k . One way of producing a CATp0q structure on
groups Gm,k is to verify that the presentation 2–complex corresponding to their
LOG presentation can be metrized so that it is a non-positively curved, piecewise
euclidean (PE) complex.
Let Km,k denote the presentation 2–complex corresponding to the LOG presentation above of Gm,k . It has one 0–cell, pm`k`1q 1–cells labeled by a1 , . . . , am`k`1 ,
and pm ` kq 2–cells corresponding to the relations a´1
j`1 aj aj`1 “ aj for 1 ď j ď m
´1
and am`j`1 aj am`j`1 “ am`j for 1 ď j ď k. By construction, Km,k is a subcomplex of Km,k`1 . We endow Km,k with a PE structure by using regular euclidean
squares for the 2–cells, and using local isometric embedding attaching maps.
Proposition 5.1. The presentation complex Km,k defined above is a non-positively
curved PE complex.
16
NOEL BRADY AND IGNAT SOROKO
Proof. We need to check the Gromov link condition [14, Th. II.5.20]. For the square
2–cells, it reduces to a purely combinatorial requirement that the link of every 0–cell
has no circuits of combinatorial length less than 4. Figure 2 shows the contributions
of the relations of Gm,k to the link L of the unique 0–cell of Km,k . We adopt the
following notation: if a 1–cell a originates at 0–cell u and terminates at 0–cell v,
then it contributes a vertex denoted a´ to the link of u and a vertex denoted a`
to the link of v.
We see that the link L can be obtained as a union of a sequence of graphs:
L1 Ă L2 Ă ¨ ¨ ¨ Ă Lm`k`1 “ L,
´
where L1 is just a pair of disjoint vertices a`
1 , a1 , and Li`1 is obtained from Li by
`
´
adding a new pair of disjoint vertices ai`1 , ai`1 and connecting each one of them
´
to some pair a`
s , as with s ă i ` 1. We observe that this procedure preserves the
´
following property: “for every l, vertices a`
l , al are non-adjacent”. Indeed, the
´
`
shortest path between the “old” vertices as , as has length two, and the shortest
´
path between the newly added vertices a`
i`1 , ai`1 is at least two. This shows that
at each step we cannot create cycles of lengths two and three. Therefore the link
L has no cycles of length less than four.
Corollary 5.2. Groups Gm,k are CATp0q.
r m,k of non-positively curved square complex
Proof. Indeed, the universal cover K
Km,k is a CAT(0) complex and Gm,k acts on it by isometries, properly discontinuously and cocompactly.
5.3. Free-by-cyclic structure. Notice that all the relations of groups Gm,k have
a
the form: ai j “ al . This implies that there exists a well-defined epimorphism
Gm,k Ñ Z, sending every ai to a fixed generator of Z. This epimorphism can
be realized geometrically by a circle-valued Morse function f : Km,k Ñ S 1 , which
can be defined as follows. Consider a CW structure on S 1 consisting of one 0–cell
and one 1–cell. Then f takes the 0–cell of Km,k to the 0–cell of S 1 , maps 1–cells
of Km,k map homeomorphically onto the target 1–cell of S 1 , and extends linearly
over the 2–cells. Here by ‘extends linearly’ we mean that f lifts to a map of the
universal covers in the way depicted in the Figure 3. (Note that, by the non-positive
curvature, characteristic maps of cells lift to embeddings in the universal cover.)
R
ai`1
ai
am`j
ai`1
f˜
Bj
Ai
1
am`j`1
0
ai
am`j`1
aj
´1
Figure 3. The Morse function on each 2–cell and the preimage
set of 0.
DEHN FUNCTIONS OF SUBGROUPS OF RIGHT-ANGLED ARTIN GROUPS
17
Proposition 5.3. The (circle-valued) Morse function f : Km,k Ñ S 1 induces a
short exact sequence
1 ÝÑ Fm`k ÝÑ Gm,k ÝÑ Z ÝÑ 1,
where Fm`k is a free group of rank m ` k.
Proof. By Theorem 4.17 it suffices to show that the ascending and the descending
links of the 0–cell in Km,k are trees.
The ascending link of the 0–cell of Km,k is formed by those corners of 2–cells
of Km,k which are formed by a pair of originating edges (labeled a´
‚ in Figure 2).
Similarly, the descending link of the 0–cell of Km,k is formed by those corners of
2–cells of Km,k which are formed by a pair of terminating edges (labeled a`
‚ , in
Figure 2).
a´
1
a´
2
a´
3
a´
m`2
a´
m`3
a´
m`4
...
a´
k
a´
k`1
a`
2
a`
3
a´
m`1
a´
m`k`1
...
a`
1
a´
m
...
...
a`
m
a`
m`1
a`
m`2
a`
m`k
a`
m`k`1
Figure 4. The ascending and the descending links for the Morse
function f : Km,k Ñ R{Z.
From Figure 4 we observe that the ascending and the descending links of the 0–
cell of Km,k are indeed trees. By the definition of f , each 2–cell of Km,k contributes
its diagonal loop to f ´1 p0–cellq. Furthermore, f ´1 p0–cellq is a bouquet of these
diagonal loops. Hence f ´1 p0–cellq is a graph having a single 0–cell and pm ` kq
1–cells which are denoted in the Figure 3 by Ai , Bj .
The above Proposition implies that Gm,k – Fm`k ¸φm,k Z for some monodromy
automorphism φm,k . We shall determine explicitly the automorphism φm,k for a
particular choice of basis for Fm`k . Let Ai , Bj be the diagonals of the 2–cells of
Km,k , as shown in Figure 3. Note that they have the following expressions in the
generators of Gm,k :
Ai “ a´1
i`1 ai ,
1 ď i ď m;
Bj “ a´1
m`j`1 aj ,
1 ď j ď k.
Proposition 5.4. For 0 ď k ď m, Gm,k has the following explicit free-by-cyclic
structure:
Gm,k – Fm`k ¸φm,k Z
where
Fm`k “ xA1 , . . . , Am , B1 , . . . , Bk y;
Z “ xa1 y
18
NOEL BRADY AND IGNAT SOROKO
and the monodromy automorphism φm,k acts as follows (here overbar denotes the
inverse):
φm,k :
A1 ÞÝÑ A1
A2 ÞÝÑ A1 pA2 q A1
A3 ÞÝÑ A1 A2 pA3 q A2 A1
...
Am ÞÝÑ A1 A2 . . . Am´1 pAm q Am´1 . . . A1
B1 ÞÝÑ A1 A2 . . . Am pB1 q
B2 ÞÝÑ A1 A2 . . . Am pB1 B2 q A1
B3 ÞÝÑ A1 A2 . . . Am pB1 B2 B3 q A2 A1
...
Bk ÞÝÑ A1 A2 . . . Am pB1 B2 . . . Bk q Ak´1 Ak´2 . . . A2 A1 .
Furthermore, φm,k is the restriction of φm,m to Fm`k .
Proof. In the proof of Proposition 5.3 it was shown that Fm`k is freely generated
by all elements Ai , Bj .
As a generator of the Z factor we are free to choose any element that maps to a
generator of π1 pS 1 q; without loss of generality, we may take Z “ xa1 y.
To get the action of the monodromy automorphism φ on the generators Ai , Bj
´1
of Fm`k we need to compute the conjugations a1 Ai a´1
1 and a1 Bj a1 . That is, we
´1
need to find words in generators Ai , Bj which are equal to a1 Ai a1 and a1 Bj a´1
1
in Km,k .
Ai
ai`1
a1
a2
ai´1 ai
a3
ai
ai´1
...
A1
A2
a1
a3
a2
...
Ai´1
Ai
Ai´1
A2
A1
Figure 5. The action of the monodromy automorphism on Ai .
For a1 Ai a´1
1 , we start with the triangle having Ai on top and 1–cells ai`1 , ai
´1
forming two bottom sides. We would like to express a1 a´1
i`1 and ai a1 as products
of free generators Ai , Bj . Since the descending link of the 0–cell in Km,k is a tree,
`
there exists a unique path in it connecting a`
1 to ai`1 and a unique path connecting
`
a`
i to a1 . These paths correspond to paths A1 A2 . . . Ai´1 Ai and Ai´1 . . . A2 A1 ,
respectively, see Figure 5. Thus,
a1 Ai a´1
1 “ A1 A2 . . . Ai´1 pAi q Ai´1 . . . A2 A1 .
DEHN FUNCTIONS OF SUBGROUPS OF RIGHT-ANGLED ARTIN GROUPS
19
Bj
am`j`1
a1
a2
a3
am
am`1 am`2 am`j
...
A1
A2
aj
aj´1
...
Am
B1
a1
a3
a2
...
Bj
Aj´1
A2
A1
Figure 6. The action of the monodromy automorphism on Bj .
Similarly, for a1 Bj a´1
1 , we start with the triangle having Bj on top and am`j`1 ,
aj forming two bottom sides. Again, there are unique paths in the descend`
`
`
ing link from a`
1 to am`j`1 and from aj to a1 . They correspond to words
A1 A2 . . . Am B1 . . . Bj and Aj´1 . . . A2 A1 , respectively, see Figure 6. Hence,
a1 Bj a´1
1 “ A1 A2 . . . Am pB1 . . . Bj q Aj´1 . . . A2 A1 .
6. Constructing a special cover for Gm,m
In this section we construct a certain permutation representation for a group
Gm,m and show that it defines a finite cover for its presentation 2–complex Km,m ,
which can be embedded in an p2m ` 1q–dimensional torus. This allows us to construct a finite special cover for Km,m , for all even values of m.
6.1. The permutation representation. We now define a right transitive action
of Gm,m on a certain set H2m`1 of cardinality 22m`1 .
Action set. For any n “ 1, . . . , 2m ` 1 denote ś
Hk to be the set of all tuples
n
of length k consisting of 0’s and 1’s, i.e. Hn “ i“1 t0, 1u. There are natural
inclusions
Hn ãÑ Hn`1 , px1 , . . . , xn q ÞÑ px1 , . . . , xn , 0q,
and we identify Hn with its image in Hn`1 under these inclusions. Also denote Hn˚
a subset of Hn`1 consisting of all tuples with the last coordinate 1:
Hn˚ “ tpx1 , . . . , xn , 1qu Ă Hn`1 .
With the above identifications, we have Hn`1 “ Hn \ Hn˚ (disjoint union).
To define a right action of a group G on a set X, it suffices to associate to each
g P G a permutation πpgq of X such that
πpghq “ πphqπpgq for all g, h P G.
Equivalently, a right action of G on X is a homomorphism of the opposite group
G˝ to SympXq, the group of all permutations of X, where G˝ equals G as a set,
with the new operation ˝ defined as
a ˝ b :“ ba.
We adopt the latter approach and construct the homomorphism from G˝m,m to
SympH2m`1 q.
20
NOEL BRADY AND IGNAT SOROKO
Recall that we have a natural tower of inclusions
Gm,0 Ă Gm,1 Ă ¨ ¨ ¨ Ă Gm,m .
Since there are also inclusions
Hm`1 Ă Hm`2 Ă ¨ ¨ ¨ Ă Hpm`1q`m “ H2m`1
this allows us to define the homomorphism π : G˝m,m Ñ SympH2m`1 q inductively
by repeatedly extending the homomorphisms G˝m,k´1 Ñ SympHm`k q to G˝m,k Ñ
SympHm`k`1 q for k “ 1, . . . , m as follows.
Base of induction. Let the m ` 1 generators a1 , . . . , am`1 of Gm,0 act on Hm`1
as flips in the respective coordinates, i.e. for i “ 1, . . . , m ` 1, set
πpai q|Hm`1 :“ βi
(A0 )
where βi : H2m`1 Ñ H2m`1 given by
βi px1 , . . . , xi´1 , xi , xi`1 , . . . q “ px1 , . . . , xi´1 , 1 ´ xi , xi`1 , . . . q
is the operator that changes the i-th coordinate from 0 to 1 and vice versa, fixing
all others.
All the relations in Gm,0 (and G˝m,0 ) are commutators rai , ai`1 s “ 1, i “
1, . . . , m. Clearly, they are satisfied in SympHm`1 q since operators βi pairwise
commute. Thus we have a well-defined homomorphism π : G˝m,0 Ñ SympHm`1 q.
Inductive step. For a fixed k P t1, . . . , mu, suppose that π : G˝m,k´1 Ñ SympHm`k q
is already defined. In particular, this implies that Hm`k is invariant under πpaj q
for all j “ 1, . . . , m ` k. Also suppose that the following property holds:
for all j “ k, . . . , m, πpaj q|Hm`k “ βj |Hm`k .
(Pk )
The base of induction above guarantees that these suppositions are true for k “ 1.
Our goal is to extend the homomorphism π|G˝m,k´1 to π : G˝m,k Ñ SympHm`k`1 q.
˚
Since Hm`k`1 “ Hm`k \ Hm`k
, it will suffice to define πpaj q|H ˚
for j “
m`k
1, . . . , m ` k, and πpam`k`1 q|Hm`k \H ˚ .
m`k
To this end, we set
πpam`k`1 q|Hm`k \H ˚
m`k
:“ βm`k`1
(A1 )
and for all 1 ď j ď m ` k,
πpaj q|H ˚
m`k
:“ βm`k`1 ¨ ϕk,m`k ¨ πpaj q|Hm`k ¨ ϕk,m`k ¨ βm`k`1 ,
(A2 )
where ¨ denotes the composition and ϕk,m`k : Hm`k`1 Ñ Hm`k`1 is the involution
that interchanges k-th and pm`kq-th coordinates leaving all other coordinates fixed:
ϕk,m`k : px1 , . . . , xk , . . . , xm`k , xm`k`1 q ÞÑ px1 , . . . , xm`k , . . . , xk , xm`k`1 q.
˚
In other words, we transfer the action of Gm,k´1 from Hm`k to Hm`k
while
twisting it with ϕk,m`k . Notice that, with the above definitions, both sets Hm`k
˚
and Hm`k
are invariant under πpaj q for j “ 1, . . . , m`k. All the relations involving
˚
generators a1 , . . . , am`k are satisfied on Hm`k`1 “ Hm`k \ Hm`k
since they hold
true on Hm`k and the conjugation by βm`k`1 ¨ϕk,m`k is a homomorphism between
˚
permutation groups on Hm`k and Hm`k
.
The only relation in Gm,k involving the last generator am`k`1 is
a´1
m`k`1 ak am`k`1 “ am`k ,
DEHN FUNCTIONS OF SUBGROUPS OF RIGHT-ANGLED ARTIN GROUPS
21
which translates to
am`k`1 ˝ ak ˝ a´1
m`k`1 “ am`k
in G˝m,k .
˚
Since βm`k`1 sends Hm`k to Hm`k
, the left-hand side of this relation acts on
Hm`k as follows:
πpam`k`1 q ¨ πpak q ¨ πpa´1
m`k`1 q|Hm`k
“ πpam`k`1 q ¨ πpak q ¨ βm`k`1 |Hm`k
“ πpam`k`1 q ¨ πpak q|H ˚
m`k
¨ βm`k`1 |Hm`k
“ πpam`k`1 q ¨ pβm`k`1 ¨ ϕk,m`k ¨ πpak q|Hm`k ¨ ϕk,m`k ¨ βm`k`1 q ¨ βm`k`1 |Hm`k
“ πpam`k`1 q ¨ βm`k`1 ¨ ϕk,m`k ¨ πpak q|Hm`k ¨ ϕk,m`k |Hm`k
“ pπpam`k`1 q ¨ βm`k`1 q ¨ ϕk,m`k ¨ πpak q|Hm`k ¨ ϕk,m`k |Hm`k
“ id ¨ ϕk,m`k ¨ πpak q|Hm`k ¨ ϕk,m`k |Hm`k “ [by (Pk )]
“ ϕk,m`k ¨ βk |Hm`k ¨ ϕk,m`k |Hm`k “ βm`k |Hm`k “ πpam`k q|Hm`k ,
where the last equality holds due to the inductive definition of π. Thus, the both
sides of the above relation act the same on Hm`k .
˚
Analogously, on Hm`k
, the left-hand side acts as:
πpam`k`1 q ¨ πpak q ¨ πpa´1
m`k`1 q|H ˚
m`k
“ βm`k`1 ¨ πpak q|Hm`k ¨ βm`k`1 |H ˚
m`k
“ [by (Pk )] “ βm`k`1 ¨ βk |Hm`k ¨ βm`k`1 |H ˚
m`k
“ βk | H ˚
m`k
,
and the right-hand side:
πpam`k q|H ˚
m`k
“ βm`k`1 ¨ ϕk,m`k ¨ πpam`k q|Hm`k ¨ ϕk,m`k ¨ βm`k`1 |H ˚
m`k
“ [by the inductive definition] “ βm`k`1 ¨ ϕk,m`k ¨ βm`k ¨ ϕk,m`k ¨ βm`k`1 |H ˚
m`k
“ βm`k`1 ¨ βk ¨ βm`k`1 |H ˚
m`k
“ βk | H ˚
m`k
.
˚
Since the two sides act the same on Hm`k \ Hm`k
“ Hm`k`1 , the above relation
is satisfied in SympHm`k`1 q, which proves that π is well-defined on G˝m,k .
It remains to be proved that the auxiliary condition (Pk ) is preserved under the
inductive step, i.e. that (Pk ) implies (Pk`1 ).
Indeed, (Pk ) means that πpaj q|Hm`k “ βj |Hm`k for j “ k, . . . , m. Thus, for
j ą k,
πpaj q|H ˚
m`k
“ βm`k`1 ¨ ϕk,m`k ¨ πpaj q|Hm`k ¨ ϕk,m`k ¨ βm`k`1 |H ˚
m`k
“ βm`k`1 ¨ ϕk,m`k ¨ βj |Hm`k ¨ ϕk,m`k ¨ βm`k`1 |H ˚
m`k
“ [since j ‰ k, m ` k] “ βj |H ˚
m`k
.
This proves that πpaj q|Hm`k`1 “ βj |Hm`k`1 for j “ k ` 1, . . . , m, i.e. that (Pk`1 )
holds.
This finishes the inductive construction of the homomorphism
π : G˝m,m Ñ SympH2m`1 q
and the proof that it is well-defined. Thus one gets a right action of Gm,m on
H2m`1 which will also be denoted π.
Figure 7 shows the permutation representation for G2,2 .
22
NOEL BRADY AND IGNAT SOROKO
1
01100
11100
dh
dh
hhh 4
hhh
hhhh
h
h
@
hhhh
hhh
@
hhh 4
hhhh
3
hhhh 01110
hh
3 @ dh 1
11000
h d11110
h
d
d
h
h
h
hhhh h
hhh 4
01000
3
@
h
h hh
hhhh 1
hhh
1
h
@
h
h
h
h
h
2
hh
2
2
@ d h3hh
h d11010
h
4
2
01010
1
hh
2
d10000
hh
dthhhh
4
hhhh
@h hhh
2
h
2 2
3 00000
hhhhhhhhh
3 @
h
h
hhh
00100
1
@ dh
10100
h
h
4
dhh
d h3hh
d10010
h
h
h
hhhh
hhhh
h
hh
hh
00010
@
1h
4
hhhh 1 @
h
h
h
h
hhh
h
hh
@ d10110
h d
h
4
5 5
00110
3
5 5
5 5 5
5
5 5
3
11101
5
01101
5
5
5 5
dhh
dhh
hh
hh 2
5
hhh 1
hhhh
@
h
h
h
h
hhhh 2
h
1@
h hhhh
hhh
h
3
h
h d01111
h d11111
11001
d h
hhh
@ dh
hhhh
01001
3 1
hh2h
@ 1
h
hhhh
hhh
hh
h
h
@
h
h
h
h
4 4
h
h
4
hh
@ d h
3h
h d 11011
2
4
h
01011
d hh1
d10001
4
hhhh
hhh
4
hhh2
00001
4
4
@hhh
3
h
h
h
hhhh
3 @ hhhhh
h
hh
00101
1 h
1
hh
10011
h
@ d10101
h
h
2
dhh
d
d
h
hhh
hhhh
hhhh 3 00011
@
hhhh
hhhhh
hhh 2
3
@
hhhh
hhhh
@ d 10111
hd
h
2
00111
1
Figure 7. The case of m “ 2, k “ 2: the action of
G2,2 “ xa1 , a2 , a3 , a4 , a5 | ra1 , a2 s “ 1, ra2 , a3 s “ 1, a´1
4 a1 a4 “
´1
a3 , a5 a2 a5 “ a4 y on H5 . Elements of H5 are arranged at vertices
of the hypercube graph marked with the corresponding tuples of
0,1’s. (Thus, each edge of this graph corresponds to a pair of opposite edges in the 1–skeleton of the 5–dimensional torus T5 defined
below.) The subset H3 is represented by the upper left-hand corner
subgraph, and H4 by the upper half of the picture. If πpai q interchanges vertices u and v we mark the edge uv with the italicized
digit i.
Proposition 6.1. The right action π of Gm,m on H2m`1 , defined above, has the
following properties:
(1) Gm,k acts transitively on Hm`k`1 for all k “ 0, . . . , m.
(2) Each generator ai , i “ 1, . . . , 2m ` 1, acts as an involution on H2m`1 .
(3) For any v P H2m`1 , and any ai , i “ 1, . . . , 2m ` 1, πpai qv differs from v in
exactly one coordinate. In particular, πpai q has no fixpoints.
(4) For any i ‰ j, πpai aj q has no fixpoints.
DEHN FUNCTIONS OF SUBGROUPS OF RIGHT-ANGLED ARTIN GROUPS
23
Proof. (1) An easy induction. The case k “ 1 is obvious since Gm,0 acts on Hm`1
by coordinate flips. So one can start with any pm ` 1q–tuple of 0,1’s and obtain
any other pm ` 1q–tuple by changing one coordinate at a time. Suppose now that
˚
Gm,k´1 acts transitively on Hm`k . Then by (A2 ), Hm`k
comprises another orbit
˚
for Gm,k´1 and am`k`1 glues Hm`k and Hm`k into one orbit for Gm,k by (A1 )
˚
thus proving that Gm,k is transitive on Hm`k`1 “ Hm`k \ Hm`k
.
(2),(3) Follow by induction from formulas (A0 )–(A2 ).
(4) Again, this is obvious for ai , aj with 1 ď i, j ď m ` 1 acting on Hm`1 since
they act as different coordinate flips βi , βj . Suppose that the statement is proven
for some k P t1, . . . , m ` 1u, for all ai , aj , 1 ď i, j ď m ` k acting on Hm`k .
˚
˚
Then πpai aj q has no fixpoints on Hm`k
either, since otherwise if v P Hm`k
is such
a fixpoint, then by (A2 ), ϕk,m`k ¨ βm`k`1 pvq would be a fixpoint for πpai aj q in
Hm`k . Finally, if, say, i “ m ` k ` 1 then πpai q changes the last, pm ` k ` 1q-st,
coordinate on Hm`k`1 , whereas for j ă i, πpaj q preserves both subsets Hm`k and
˚
Hm`k
, so it doesn’t change the pm ` k ` 1q-st coordinate. Hence, the composition
of πpaj q and πpai q has no fixpoints.
e0
C2 :
0
1
e1
Figure 8. The CW complex C2 .
6.2. A p2m ` 1q–torus cover. Let C2 “ R{2Z be a 1–dimensional CW complex
with the following CW structure: its 0–cells are 0 ` 2Z and 1 ` 2Z, which we denote
by 0 and 1 respectively. The 1–cells are r0, 1s ` 2Z and r1, 2s ` 2Z, which we denote
by e0 and e1 respectively, see Figure 8.
We denote by Tn the CW complex Rn {p2Zqn – pR{2Zqn with the product CW
structure. Notice that the natural action of p2Zqn on Rn preserves the standard
unit cubulation of Rn , hence induces the structure of a cubical complex on Tn .
Observe that Tn is homeomorphic to an n–dimensional torus.
In what follows, it will be convenient to parametrize points of Tn by n–tuples
px1 , . . . , xn q of numbers from r0, 2s viewed up to the identification 0 „ 2.
The 0–skeleton of Tn is naturally identified with the set Hn of all n-tuples of
t0, 1u introduced before.
The 1–cells of Tn are formed by fixing an edge e0 or e1 in some factor of Tn “
C2 ˆ C2 ˆ ¨ ¨ ¨ ˆ C2 , say, in position i, and taking product with vertices 0 or 1 in all
other positions. (So if two 0–cells of T0 differ in only one coordinate, then there is
p1q
a unique directed edge in Tn from the first 0–cell to the second one and a unique
directed edge from the second one to the first one.) Thus, a typical 1–cell in Tn can
be identified with a product of the form
v1 ˆ v2 ˆ ¨ ¨ ¨ ˆ vi´1 ˆ eα ˆ vi`1 ˆ ¨ ¨ ¨ ˆ vn ,
where each vj P t0, 1u and α “ 0 or 1.
Similarly, an arbitrary 2–cell of Tn is a product
v1 ˆ ¨ ¨ ¨ ˆ vi´1 ˆ eα ˆ vi`1 ˆ ¨ ¨ ¨ ˆ vj´1 ˆ eβ ˆ vj`1 ˆ ¨ ¨ ¨ ˆ vn
24
NOEL BRADY AND IGNAT SOROKO
for some choice of 1 ď i, j ď n (i ‰ j), with each vk P t0, 1u, and α, β P t0, 1u.
Let Km,m be the presentation 2–complex for Gm,m . Recall that it consists of
one 0–cell, p2m ` 1q 1–cells corresponding to the generators a1 , . . . , a2m`1 of Gm,m
and 2m 2–cells corresponding to the relations of Gm,m .
Consider the right action π : G˝m,m Ñ SympH2m`1 q constructed in the previous
section and denote
S “ tg P Gm,m | πpgqp0, 0, . . . , 0q “ p0, 0, . . . , 0qu
the stabilizer of the point p0, 0, . . . , 0q in Gm,m . Subgroup S defines a finite covering
p m Ñ Km,m whose properties we now describe.
K
p m cellularly embeds into the 2–skeleton of
Proposition 6.2. The covering space K
T2m`1 .
p m are in one-to-one correspondence with the right cosets
Proof. The 0–cells of K
SzGm,m . Since the action of Gm,m is transitive on H2m`1 by Proposition 6.1(1),
p0q
p0q
pm
consists of |Gm,m : S| “ 22m`1 vertices which we can identify with T2m`1 ,
K
the 0–skeleton of T2m`1 , which was earlier identified with the set H2m`1 of all
p2m ` 1q–tuples consisting of t0, 1u.
p m are in one-to-one correspondence with pairs of right cosets
The 1–cells of K
pSg, Sgai q where ai , i “ 1, . . . , 2m ` 1 runs through all the generators of Gm,m .
Proposition 6.1(3) guarantees that each such 1–cell is not a loop, and it actually
belongs to the 1–skeleton of the p2m ` 1q–torus T2m`1 .
p m are lifts of the 2–cells in Km,m . Each such 2–cell is uniquely
The 2–cells of K
determined by the base vertex (a lift of the base vertex of Km,m ) and by the fixed
cyclic order of the relator word of Gm,m , which defines the attaching map of a
2–cell in Km,m . Indeed, since π : G˝m,m Ñ SympH2m`1 q is a homomorphism, every
relator word w of Gm,m acts as the identical permutation. There are two types of
relators in the presentation for Gm,m :
´1
ai ai`1 a´1
i ai`1 “ 1
for i “ 1, . . . , m,
´1
a´1
m`j`1 aj am`j`1 am`j
“1
and
for j “ 1, . . . , m,
p m , based at every
each of which has length 4. Thus they define length 4 loops in K
vertex, each of such loops has to be filled with a 2–cell because these loops must
p m can be given
be nullhomotopic when projected to Km,m . Thus, each 2–cell of K
´1
by a word ai aj a´1
a
for
some
values
i
‰
j,
j
‰
l,
see
Figure
9.
j
l
Sgaj
al
aj
Sg
Sgaj al “ Sgai aj
aj
ai
Sgai
pm.
Figure 9. A typical 2–cell in K
DEHN FUNCTIONS OF SUBGROUPS OF RIGHT-ANGLED ARTIN GROUPS
25
p0q
If we identify cosets SzGm,m with T2m`1 ” H2m`1 , Proposition 6.1(3) shows
that for any i “ 1, . . . , 2m ` 1, every edge pSg, Sgai q changes only one coordinate
of the p2m`1q–tuple of t0, 1u representing vertex Sg, therefore it maps to a suitable
1–cell of T2m`1 .
p0q
pm
Let’s show that each 2–cell of the above form at a vertex Sg P K
naturally
p2q
p0q
pm
embeds into T2m`1 under the embedding induced by the embeddings of K
and
p1q
p1q
p
Km to T2m`1 introduced above. Denote p, q, r, s the positions in t1, . . . , 2m ` 1u
in which the endpoints of the following edges differ: pSg, Sgai q, pSgai , Sgai aj q,
pSg, Sgaj q, pSgaj , Sgaj al q, respectively. By Proposition 6.1(4), ai aj and aj al have
no fixpoints on H2m`1 , hence p ‰ q and r ‰ s. And since the square above is
commutative, we conclude that 2–element sets tp, qu and tr, su are equal. Again,
by Proposition 6.1(2), ai and aj act as involutions, hence p ‰ r, since otherwise
a´1
i aj “ ai aj would have a fixpoint Sgai . Therefore, p “ s, q “ r, and the 2–cell
p2q
under consideration actually belongs to T2m`1 since each pair of its parallel edges
changes coordinates of vertices in the same position, one in position p “ s and
another in q “ r.
6.3. Exploring hyperplane pathologies. We will need the following description
of hyperplanes and walls in the n–torus Tn .
Lemma 6.3. The hyperplanes and oriented walls in Tn are in 1–1 correspondence
with pairs pi, eα q, where 1 ď i ď n and α P t0, 1u. More explicitly:
(1) the hyperplane corresponding to the pair pi, eα q is a subset of Tn of one of
the following two types:
(
px1 , . . . xi´1 , 21 , xi`1 , . . . , xn q | xj P r0, 2s
if eα “ e0 , and
px1 , . . . xi´1 , 32 , xi`1 , . . . , xn q | xj P r0, 2s
(
if eα “ e1 (with the identification 0 „ 2).
(2) the oriented wall through a 1–cell
v1 ˆ ¨ ¨ ¨ ˆ vi´1 ˆ eα ˆ vi`1 ˆ ¨ ¨ ¨ ˆ vn
consists of all 1–cells
u1 ˆ ¨ ¨ ¨ ˆ ui´1 ˆ eα ˆ ui`1 ˆ ¨ ¨ ¨ ˆ un
with i, eα fixed, and uk ’s taking all possible values of t0, 1u.
The oriented wall in (2) is dual to the corresponding hyperplane in (1).
Proof. (1) Recall that the structure of a cube complex on Tn – Rn {p2Zqn is induced
by the standard cubulation of Rn . Hyperplanes in Rn are subsets of the form
(
Hi,k “ px1 , . . . , xi´1 , k ` 21 , xi`1 , . . . , xn q | xj P R , 1 ď i ď n, k P Z.
Modding out by the action of p2Zqn yields the result.
(2) Recall that the oriented wall containing a 1–cell a of a cube complex X is
the class of all oriented 1–cells of X which are connected to a through a sequence
of elementary parallelisms via the 2–cells of X. Notice that an arbitrary 1–cell of
Tn is a product of the form: v1 ˆ ¨ ¨ ¨ ˆ vi´1 ˆ eα ˆ vi`1 ˆ ¨ ¨ ¨ ˆ vn , where each vertex
vk P t0, 1u and α “ 0 or 1, and an arbitrary 2–cell of Tn is a product
v1 ˆ ¨ ¨ ¨ ˆ vi´1 ˆ eα ˆ vi`1 ˆ ¨ ¨ ¨ ˆ vj´1 ˆ eβ ˆ vj`1 ˆ ¨ ¨ ¨ ˆ vn
26
NOEL BRADY AND IGNAT SOROKO
for some choice of 1 ď i, j ď n (i ‰ j) and α, β P t0, 1u. Thus, the elementary
parallelism via the above 2–cell establishes equivalence of the 1–cells
v1 ˆ ¨ ¨ ¨ ˆ vi´1 ˆ eα ˆ vi`1 ˆ ¨ ¨ ¨ ˆ vj ˆ ¨ ¨ ¨ ˆ vn
and
v1 ˆ ¨ ¨ ¨ ˆ vi´1 ˆ eα ˆ vi`1 ˆ ¨ ¨ ¨ ˆ p1 ´ vj q ˆ ¨ ¨ ¨ ˆ vn .
Since index j varies independently of i, we conclude that any 1–cell u1 ˆ ¨ ¨ ¨ ˆ
ui´1 ˆ eα ˆ ui`1 ˆ ¨ ¨ ¨ ˆ un , uk P t0, 1u, is contained in the parallelism class of
v1 ˆ ¨ ¨ ¨ ˆ vi´1 ˆ eα ˆ vi`1 ˆ ¨ ¨ ¨ ˆ vn .
p m does not have three of the four pathologies
Now we show that the complex K
in the definition of a special cube complex.
Proposition 6.4.
p m do not self-intersect.
(a) Hyperplanes of K
p m do not self-osculate.
(b) Hyperplanes of K
p m are two-sided.
(c) Hyperplanes of K
Proof. It is convenient to work with the oriented walls dual to hyperplanes. Since,
p m is a square subcomplex of T2m`1 , every wall of K
p m is a
by Proposition 6.2, K
subset of some wall of T2m`1 . By Lemma 6.3, the walls in T2m`1 consist of all
1–cells of the form u1 ˆ ¨ ¨ ¨ ˆ ui´1 ˆ eα ˆ ui`1 ˆ ¨ ¨ ¨ ˆ u2m`1 for a fixed i, eα , and
arbitrary uk P t0, 1u.
p m were self-intersecting, the corresponding wall would conIf a hyperplane of K
tain edges with eα in two different coordinate positions i and j, which is impossible.
This proves (a).
p m were self-osculating, the corresponding wall would contain
If a hyperplane of K
a pair of edges u1 ˆ u2 ˆ ¨ ¨ ¨ ˆ ui´1 ˆ eα ˆ ui`1 ˆ ¨ ¨ ¨ ˆ u2m`1 and v1 ˆ v2 ˆ ¨ ¨ ¨ ˆ
vi´1 ˆ eα ˆ vi`1 ˆ ¨ ¨ ¨ ˆ v2m`1 with common extremities: either their origins or their
termini coincide (for direct self-osculation), or the origin of one edge coincides with
the terminus of the other (for indirect self-osculation). In either case the tuples
pu1 , . . . , ui´1 , ui`1 , . . . , u2m`1 q and pv1 , . . . , vi´1 , vi`1 , . . . , v2m`1 q are equal, which
means that the original 1–cells are equal and there is actually no self-osculation
happening. This proves (b).
p m lies in a unique hyperplane
To prove (c) we observe that a hyperplane H of K
in T2m`1 . In particular, by the above lemma, in the coordinate system on T2m`1 ,
the hyperplane H has the following description:
(
H “ px1 , . . . xi´1 , t, xi`1 , . . . , x2m`1 q
for some 1 ď i ď 2m ` 1, t “ 12 or 32 , and some values from r0, 2s for the rest of the
p m is a square complex, H is the union of mid-cubes of some set
variables. Since K
p m of
of square 2–cells. Hence each point z of H belongs to a square 2–cell C of K
the form
C “ v1 ˆ ¨ ¨ ¨ ˆ vi´1 ˆ eα ˆ vi`1 ˆ ¨ ¨ ¨ ˆ vj´1 ˆ eβ ˆ vj`1 ˆ ¨ ¨ ¨ ˆ v2m`1
for some 1 ď j ď 2m ` 1, where all vk ’s are 0 or 1. (The index j may be less than
or bigger than i.)
Suppose that eα “ e0 so that t “ 21 . Then z actually has coordinates:
z “ pv1 , . . . vi´1 , 12 , vi`1 , . . . , vj´1 , s, vj`1 , . . . , v2m`1 q
DEHN FUNCTIONS OF SUBGROUPS OF RIGHT-ANGLED ARTIN GROUPS
27
where s is some value from r0, 2s. We see that the set
z ˆ r0, 1s “ pv1 , . . . vi´1 , t, vi`1 , . . . , vj´1 , s, vj`1 , . . . , v2m`1 q | t P r0, 1s
(
also belongs to the same square C above. We conclude that the product
(
H ˆ r0, 1s “ px1 , . . . xi´1 , t, xi`1 , . . . , xn q | t P r0, 1s
p m . This defines a combinatorial map H ˆ r0, 1s Ñ K
pm
is a union of 2–cells of K
1
(actually, an embedding) such that H ˆ t 2 u is identified with H itself.
A similar reasoning applies if eα “ e1 , t “ 32 (we parametrize H ˆ r0, 1s by
t P r1, 2s).
p m is two-sided.
This proves that every hyperplane of K
Unfortunately, cube subcomplexes of a Cartesian product of three or more graphs
can have inter-osculating hyperplanes, as the example in the Figure 10 shows. The
2–complex in Figure 10 is a subcomplex of the product of two segments of length
one and a segment of length two.
Figure 10. A subcomplex in a product of three graphs with interosculating hyperplanes.
However, Haglund and Wise have proved in [21, Th. 5.7] that in the case when the
square complex is a so-called VH-complex, the absence of the first three hyperplane
pathologies guarantees the existence of a finite special cover.
Definition 6.5. (VH-complex) A simple square complex is called a VH-complex
if its edges are divided into two disjoint classes: vertical and horizontal, such that
the attaching map of each square is of the form vhv 1 h1 where v, v 1 are vertical and
h, h1 are horizontal edges.
p m are VHProposition 6.6. For all even integers m, the complexes Km,m and K
pm.
complexes. Hence there exist a finite special cover K m Ñ K
Proof. From the LOG definition (see section 5.1) of groups Gm,m we observe that,
for the even integers m, the odd-indexed and the even-indexed generators form two
classes V and H (‘vertical’ and ‘horizontal’) such that all relators of Gm,m have the
form: v1h “ v2 or hv1 “ h2 for some v, v1 , v2 P V , h, h1 , h2 P H. This implies that
the complex Km,m is a VH-complex.
p m , being a finite cover of Km,m , inherits the structure of a VHThe complex K
complex from Km,m . Indeed, the preimages of the vertical and horizontal edges
28
NOEL BRADY AND IGNAT SOROKO
p m Ñ Km,m form two disjoint classes Vp “
in Km,m under the covering map p : K
´1
´1
p “ p pHq, and all edges of K
p m are contained in Vp \ H.
p The link
p pV q and H
of every vertex of Km,m is a bipartite graph corresponding to parts V and H,
and links of vertices are mapped isomorphically under covering maps. Thus all
p m are bipartite with respect to parts Vp and H.
p Therefore
links of vertices in K
p
all 2–cells of Km have boundaries of the form v1 h1 v2 h2 with v1 , v2 P Vp , h1 , h2 P
p By Proposition 6.4, hyperplanes of K
p m have no self-intersections and no selfH.
osculations. Hence, by Theorem 5.7 in [21], there exist a special cube complex K m
pm.
and a finite cover K m Ñ K
7. Proof of Main Theorems
Now we are ready to prove our main results.
Theorem A. For each positive even integer m there exist virtually special free-bycyclic groups Gm,m – F2m ¸φ Z with growth function grφ pnq „ nm and Gm,m´1 –
F2m´1 ¸φ1 Z with growth function grφ1 pnq „ nm´1 .
Proof. We have seen in Proposition 6.6 that for each even integer m ą 0, there
exist a special cover K m Ñ Km,m for the presentation complex Km,m of the group
Gm,m – F2m ¸φ Z. In Propositions 8.1 and 8.12 of section 8 we show that the
growth function for φ “ φm,m is „ nm . This proves the first part of Theorem A.
For the second part, recall that Gm,m´1 “ F2m´1 ¸φ1 Z, where φ1 “ φm,m´1
is the restriction of φ on the free subgroup on the first 2m ´ 1 generators. By
construction, the presentation complex Km,m´1 for Gm,m´1 is a subcomplex of
Km,m , and is actually obtained from Km,m by deleting the loop corresponding to
the last generator a2m`1 and also the single open 2–cell adjacent to a2m`1 (i.e.
which have a2m`1 as one of their sides).
Let p̄ : K m Ñ Km,m be the special cover of Km,m from Proposition 6.6. Consider
a square subcomplex K 1m Ă K m which is obtained by:
(1) deleting all 1–cells of K m which map under p̄ onto the loop labeled a2m`1
in Km,m ;
(2) deleting all open 2–cells of K m which have 1–cells from (1) as one of their
sides;
(3) taking a connected component of the resulting complex.
We claim that p̄ : K 1m Ñ Km,m´1 is a finite special cover of Km,m´1 .
Indeed, by construction, p̄pK 1m q lies in Km,m´1 . The hyperplanes of K 1m are twosided, do not self-intersect and do not self-osculate, since they are subcomplexes of
the corresponding hyperplanes in the special complex K m .
To see that the complex K 1m has no inter-osculating hyperplanes, observe that
in steps (1), (2) above we deleted only the hyperplanes which are dual to the 1–
cells corresponding to the last generator a2m`1 . This doesn’t change the absence of
inter-osculation of the remaining hyperplanes of K m . Therefore, the hyperplanes
in K 1m do not inter-osculate either, and K 1m is special.
Again, that the growth of φ1 is „ nm´1 is shown in Propositions 8.1 and 8.12 in
section 8, since φ1 “ φm,m´1 .
Corollary A. For each positive integer k there exist a right-angled Artin group
containing a free-by-cyclic subgroup whose monodromy automorphism has growth
function „ nk .
DEHN FUNCTIONS OF SUBGROUPS OF RIGHT-ANGLED ARTIN GROUPS
29
Proof. In Theorem A we have proved that for any positive integer k (where k “ m
or m ´ 1 for arbitrary even m) there exists a free-by-cyclic group G “ F ¸ψ Z
with grψ pnq „ nk , such that some finite index subgroup H ď G is isomorphic to a
fundamental group of a special cube complex. By D. Wise’s celebrated result (see
Corollary 4.13), there exists a right-angled Artin group Ap∆q with H isomorphic
to a subgroup of Ap∆q.
Let’s prove that H is free-by-cyclic itself. Indeed, we have a commutative diagram:
1
F
F ¸ψ Z π
Z
1
1
N
H
`Z
1
Here N “ H X F and `Z is the image of H under π. Since H has finite index in
G, ` ‰ 0. Hence the subgroup N is invariant under ψ ` and is a free group. Since
F Ÿ G, F H is a subgroup of G, and
|F : N | “ |F : H X F | “ |F H : H| ď |G : H| ă 8.
Therefore, N is a finitely generated free group, and H – N ¸ψ` Z, a free-by-cyclic
group.
Parts (ii) and (iii) of Proposition 2.4 tell us now that
grψ` |N pnq „ grψ pnq „ nk .
Theorem B. For each positive integer k there exists a right-angled Artin group
which contains a finitely presented subgroup with Dehn function » nk .
Proof. In Theorem A we proved that, for all even integers m, the free-by-cyclic
group Gm,m “ F2m ¸φ Z (resp. Gm,m´1 “ F2m´1 ¸φ1 Z) has the following properties:
grφ pnq „ nm (resp. grφ1 pnq „ nm´1 ), and it contains a finite index subgroup H
which embeds into a right-angled Artin group. In Corollary A we showed that H
is itself free-by-cyclic: H – N ¸φ` Z (resp. H – N ¸φ1 ` Z) for some finite index
subgroup N ď F2m (resp. N ď F2m´1 ) with the monodromy automorphism being
φ` (resp. pφ1 q` ) for some ` ą 0.
We claim that the Bieri double of H, ΓpHq “ H ˚F H, has Dehn function δΓ pnq
» equivalent to grφ pnq ¨ n2 (resp. grφ1 pnq ¨ n2 ), and itself embeds into a RAAG.
We now prove this claim for the case of subgroup H ď Gm,m , the monodromy
automorphism φ` , and k “ m, and notice that the case of H ď Gm,m´1 , the
`
monodromy automorphism φ1 , and k “ m ´ 1, is proved in a similar fashion.
k`2
The upper bound: δΓ pnq ă
is established as follows. By Propositions 8.1,
“n
k
and 8.12, grφ pnq „ n and grφ´1 pnq „ nk . Proposition 2.4(ii) implies that grφ` pnq „
nk and grpφ` q´1 pnq “ grpφ´1 q` pnq „ nk , and so the upper bound follows from Proposition 3.4.
The lower bound: n¨max}b}ďn, bPN }φ`n pbq} ă
“ δΓ pnq was given in Proposition 3.3.
k`1
If we show that max}b}ďn, bPN }φ`n pbq} ą
, it will follow, in view of the above,
“n
that δΓ pnq » nk`2 .
We prove in section 8 that in the group Gm,m “ F2m ¸φ Z, containing H, the
maximum in the definition of the growth functions
gr
›
› φ pnq
ˇ and ngrφab pnq
ˇ is achieved
m
k
ˇ
at the generator Bm (see Corollary 8.14): ›φn pBm q› „ ˇpφab
q
p
B̄
q
k 1 „ n “ n .
m,k
30
NOEL BRADY AND IGNAT SOROKO
(Here the bar over an element of F2m denotes its image in the abelianization of
F2m .)
Since the subgroup N is of finite index in F2m , there exists an integer p ą 0 such
p
that Bm
P N . Then we have:
›
› ›
›
ˇ
ˇ
pn ›
max ›φ`n pbq› ě ›φ`n pBm
q ě pn ¨ ˇpφab q`n pB̄m qˇ ľ pnp`nqm „ nm`1 .
1
}b}ďn
bPN
Since k “ m, this proves that δΓ pnq » nk`2 .
To prove that Γ embeds into a RAAG, consider a homomorphism µ : Γ Ñ H ˆ
F pu, vq, where F pu, vq is a free group of rank 2 on free generators u, v, from [12,
Rem. 3.7(iii)], which is described as follows. Let N “ F px1 , . . . , xq q be freely
generated by elements x1 , . . . , xq . Denote the generators of Z-factors
` of three copies
`
of H
as
s,
t
and
τ
.
Also
denote
for
brevity
ψ
“
φ
.
Then
Γ
“
F px1 , . . . , xq q ¸
˘
`
˘
`
˘ψ
xsy ˚F px1 ,...,xq q F px1 , . . . , xq q ¸ψ xty , and H ˆ F pu, vq “ F px1 , . . . , xq q ¸ψ xτ y ˆ
F pu, vq. Define µ on the generators as follows:
xi ÞÑ xi ,
s ÞÑ τ u,
t ÞÑ τ v.
We check at once that µ is a homomorphism, and it is easily proved using the
normal forms of elements in free amalgamated products, that µ is injective.
Thus, if we denote the right-angled Artin group containing H as Ap∆q, for some
graph ∆, then Γ Ă H ˆ F2 is a subgroup of Ap∆q ˆ F2 , which is itself a RAAG
(corresponding to the graph join of ∆ and the empty graph on two vertices).
Remark 7.1. Following the proof of Proposition 3.3 given in [12, Lemma 1.5]
(see also [10, Proposition 7.2.2]), we can exhibit an explicit sequence of words
wn “ rpst´1 qn , t`n Bkpn t´`n s (where k “ m or m´1, as above) which realize the lower
bound for the Dehn function δΓ pnq. To understand what van Kampen diagrams for
these words look like, the reader is referred to the proof of [12, Lemma 1.5]. In the
notation of [12], β “ Bkpn , t1 “ t, t2 “ s.
8. Growth of φ and φ´1
In Proposition 5.4 we have shown that, for all 1 ď k ď m, Gm,k “ Fm`k ¸φm,k Z,
where Fm`k is a free group with generators A1 , . . . , Am , B1 , . . . , Bk . For convenience, in what follows we adopt the notation:
φ “ φm,m
and use the fact that φm,k is the restriction of φ on the first m ` k generators.
The goal of this section is to prove that φm,k and pφm,k q´1 have growth „ nk . In
particular, grφ pnq „ grφ´1 pnq „ nm .
Throughout this section, }.} will denote the word length in F2m with respect to
the system of free generators tAi , Bj u.
DEHN FUNCTIONS OF SUBGROUPS OF RIGHT-ANGLED ARTIN GROUPS
31
Recall (see Proposition 5.4) that the automorphism φ is given by the formulas
(where the overbar denotes the inverse):
φ “ φm,m :
A1 ÞÝÑ A1
(1)
A2 ÞÝÑ A1 pA2 q A1
A3 ÞÝÑ A1 A2 pA3 q A2 A1
...
Am ÞÝÑ A1 A2 . . . Am´1 pAm q Am´1 . . . A1
B1 ÞÝÑ A1 A2 . . . Am pB1 q
B2 ÞÝÑ A1 A2 . . . Am pB1 B2 q A1
B3 ÞÝÑ A1 A2 . . . Am pB1 B2 B3 q A2 A1
...
Bm ÞÝÑ A1 A2 . . . Am pB1 B2 . . . Bm q Am´1 Am´2 . . . A2 A1 .
8.1. Upper bounds for the growth of φ, φ´1 .
Proposition 8.1. For the automorphism φm,k we have:
grφm,k pnq ĺ nk
and
grpφm,k q´1 pnq ĺ nk .
We will need few basic lemmas.
Lemma 8.2. For i “ 1, . . . , m,
φn pAi q “ An1 An2 . . . Ani´1 ¨ Ai ¨ Ani´1 . . . An2 An1 .
Proof. We prove the statement by induction on n, observing that it is true for
n “ 0, 1:
φn`1 pAi q “ φpφn pAi qq “ φpAn1 An2 . . . Ani´1 q ¨ φpAi q ¨ φpAni´1 . . . An2 An1 q
“ φpAn1 qφpAn2 qφpAn3 q . . . φpAni´1 q ¨ φpAi q ¨ φpAni´1 q . . . φpAn3 qφpAn2 qφpAn1 q
“ pAn1 qpA1 An2 A1 qpA1 A2 An3 A2 A1 q . . . pA1 . . . Ai´2 Ani´1 Ai´2 . . . A1 q
ˆ pA1 . . . Ai´1 Ai Ai´1 . . . A1 q ˆ pA1 . . . Ai´2 Ani´1 Ai´2 . . . A1 q . . . pA1 A2 An3 A2 A1 q
n`1
n`1 n`1
ˆ pA1 An2 A1 qpAn1 q “ An`1
An`1
. . . An`1
A1 .
1
2
i´1 ¨ Ai ¨ Ai´1 . . . A2
Corollary 8.3. For i “ 1, . . . , m,
}φn pAi q} “ 2pi ´ 1qn ` 1.
Lemma 8.4. φn pB1 q “ An1 An2 . . . Anm ¨ B1 .
Proof. The statement is true for n “ 0, 1. By induction,
φn`1 pB1 q “ φpφn pB1 qq “ φpAn1 An2 . . . Anm q ¨ φpB1 q
“ pAn1 qpA1 An2 A1 qpA1 A2 An3 A2 A1 q . . . pA1 . . . Am´1 Anm Am´1 . . . A1 q
ˆ pA1 A2 . . . Am ¨ B1 q “ An`1
An`1
. . . An`1
¨ B1 .
m
1
2
Corollary 8.5. }φn pB1 q} “ mn ` 1.
32
NOEL BRADY AND IGNAT SOROKO
n
n n
Lemma 8.6. φn pA1 A2 . . . Am q “ An`1
An`1
. . . An`1
1
2
m´1 ¨ Am ¨ Am´1 . . . A2 A1 .
Proof. We do induction on n, the case n “ 0 being evident:
φn`1 pA1 A2 . . . Am q “ φpφn pA1 A2 . . . Am qq “ φpAn`1
qφpAn`1
q . . . φpAn`1
1
2
m´1 q
ˆ φpAm q ¨ φpAnm´1 q . . . φpAn2 qφpAn1 q “ pAn`1
A1 qpA1 A2 An`1
A2 A1 q . . .
qpA1 An`1
3
1
2
ˆ pA1 . . . Am´2 An`1
m´1 Am´2 . . . A1 q ¨ pA1 . . . Am´1 Am Am´1 . . . A1 q
ˆ pA1 . . . Am´2 Anm´1 Am´2 . . . A1 q . . . pA1 A2 An3 A2 A1 q ¨ pA1 An2 A1 qpAn1 q
n`1
n`1 n`1
A1 .
“ An`2
An`2
. . . An`2
1
2
m´1 ¨ Am ¨ Am´1 . . . A2
Lemma 8.7. }φn pB2 q} “
m 2
2n
` pm
2 ` 2qn ` 1.
Proof. One observes that
φpB2 q “ φpB1 q ¨ B2 ¨ A1 .
This gives by induction in view of Lemma 8.4:
φn pB2 q “ φn pB1 qφn´1 pB1 q . . . φpB1 q ¨ B1 ¨ An1
n
“ pAn1 An2 . . . Anm B1 q ¨ pA1n´1 An´1
. . . An´1
m B1 q . . . pA1 A2 . . . Am B1 q ¨ B1 ¨ A1 .
2
Hence,
}φn pB2 q} “ rmn ` 1s ` rmpn ´ 1q ` 1s ` . . . ` rm ` 1s ` 1 ` n
`m
˘
2
“ m npn`1q
` 2n ` 1 “ m
2
2 n ` 2 ` 2 n ` 1.
Claim 8.8. For k “ 1, . . . , m,
}φn pBk q} ĺ nk .
Proof. From the formulas (1) we get for all k ě 2,
φpBk`1 q “ φpBk q ¨ pA1 . . . Ak´1 q ¨ Bk`1 ¨ φpAk . . . A1 q.
Therefore,
φn pBk`1 q “ φn pBk q ¨ φn´1 pA1 . . . Ak´1 q ¨ φn´1 pBk`1 q ¨ φn´1 pA1 . . . Ak q´1 .
Lemma 8.6 gives:
}φn´1 pA1 . . . Ak´1 q} “ 2pk ´ 2qpn ´ 1q ` pk ´ 1q,
}φn´1 pA1 . . . Ak q´1 } “ 2pk ´ 1qpn ´ 1q ` k,
so that the total length of φn pBk`1 q is bounded above by
}φn pBk q} ` }φn´1 pBk`1 q} ` rp4k ´ 6qn ´ p2k ´ 5qs.
Now if we denote
f pk, nq :“ }φn pBk q},
we will
‚
‚
‚
have
f pk, 0q “ 1;
f p1, nq “ mn ` 1, by Corollary 8.5;
m
2
f p2, nq “ m
2 n ` p 2 ` 2q ` 1, by Lemma 8.7;
(2)
DEHN FUNCTIONS OF SUBGROUPS OF RIGHT-ANGLED ARTIN GROUPS
33
and for k ě 2,
f pk ` 1, nq ď f pk, nq ` f pk ` 1, n ´ 1q ` rp4k ´ 6qn ´ p2k ´ 5qs.
We have an inequality here (instead of an equality) because in the formula (2) there
can be some cancellations. Let’s define another function gpk, nq as follows:
‚ gpk, 0q “ 1;
‚ gp1, nq “ f p1, nq “ mn ` 1,
m
2
‚ gp2, nq “ f p2, nq “ m
2 n ` p 2 ` 2q ` 1,
‚ gpk ` 1, nq “ gpk, nq ` gpk ` 1, n ´ 1q ` rp4k ´ 6qn ´ p2k ´ 5qs, for k ě 2.
Obviously, g is well-defined in a recurrent fashion. An easy induction shows that
f pk, nq ď gpk, nq for all k ě 1, n ě 0,
so that gpk, nq gives an upper bound for the growth of }φn pBk q}.
To estimate the order of growth of gpk, nq, let’s look at finite differences in n:
gpk ` 1, nq ´ gpk ` 1, n ´ 1q “ gpk, nq ` rp4k ´ 6qn ´ p2k ´ 5qs
so if we assume by induction that gpk, nq is a polynomial in n of degree k, then we
conclude that gpk ` 1, nq is a polynomial of degree k ` 1 in n.
Since this assumption is true for k “ 1 and 2, this proves Claim 8.8.
Now we establish a similar bound for φ´1 . One could use the train-track machinery (along the lines of [25, Th. 0.4]) to prove that the growth of φ´1 is the same
as the growth of φ, when it is polynomial, but we give here a simple direct proof.
One easily checks that the inverse automorphism φ´1 acts as follows:
φ´1 :
A1 ÞÝÑ A1
(3)
A2 ÞÝÑ A1 pA2 q A1
A3 ÞÝÑ A1 A2 pA3 q A2 A1
...
Am ÞÝÑ A1 A2 . . . Am´1 pAm q Am´1 . . . A2 A1
B1 ÞÝÑ A1 A2 . . . Am´1 Am ¨ B1
B2 ÞÝÑ pB 1 B2 qA1
B3 ÞÝÑ A1 pB 2 B3 q A2 A1
...
Bm ÞÝÑ A1 A2 . . . Am´2 pB m´1 Bm q Am´1 . . . A2 A1 .
Lemma 8.9. For i “ 1, . . . , m,
}φ´n pAi q} “ }φn pAi q} “ 2pi ´ 1qn ` 1.
Proof. Define an automorphism ι : H ÝÑ H of the subgroup H “ xA1 , . . . , Am y
given by ι : Aj ÞÑ Aj , j “ 1, . . . , m. One easily checks that
φ´1 |H “ ι ˝ pφ|H q ˝ ι´1 ,
therefore }φ´n pAi q} “ }φn pAi q} and the result follows from Corollary 8.3.
34
NOEL BRADY AND IGNAT SOROKO
Claim 8.10. For k “ 1, . . . , m,
}φ´n pBk q} ĺ nk .
Proof. Denote for any i ě 0, k ě 2:
T1,i :“ B1 ,
Tk,i :“ Bk ¨ Aik´1 Aik´2 . . . Ai2 Ai1 ,
Si :“ Ai1 Ai2 . . . Aim´1 Aim .
In this notation, the action of φ´1 can be written as follows:
φ´1 pTk,i q “ Tk´1,1 ¨ Tk,i`2 ,
φ
´1
(4)
pSi ¨ T1,1 q “ Si`1 ¨ T1,1 .
The first relation is obvious, and the second one follows by easy induction.
Lemma 8.11. }φ´n pB1 q} “ mn ` 1.
Proof. Indeed,
φ´n pB1 q “ φ´n pS0 ¨ T1,1 q “ φ´pn´1q pS1 ¨ T1,1 q “ φ´pn´2q pS2 ¨ T1,1 q “ . . .
“ φ´2 pSn´2 ¨ T1,1 q “ φ´1 pSn´1 ¨ T1,1 q “ Sn ¨ T1,1 .
Define for any k ě 1, i ě 0, n ě 0 a function f pk, i, nq as follows:
‚ f p1, i, nq “ mn ` 1;
‚ f pk, i, nq “ }φ´n pTk,i q}, for k ě 2.
Then relations (4) imply
f pk, i, n ` 1q ď f pk ´ 1, 1, nq ` f pk, i ` 1, nq,
where we have an inequality (but not an equality) because of possible cancellations
in the reduced expression for φ´n pTk,i q.
To obtain an upper bound on f pk, i, nq we introduce a function gpk, i, nq defined
recurrently as follows:
‚ gpk, i, 0q “ }Tk,i } “ pk ´ 1qi ` 1, for all k ě 1, i ě 0;
‚ gp1, i, nq “ }φ´n pB1 q} “ mn ` 1, for all i ě 0, n ě 0;
‚ gpk, i, n ` 1q “ gpk ´ 1, 1, nq ` gpk, i ` 1, nq, for all k ě 2, i ě 0, n ě 0.
These formulas define gpk, i, nq recurrently for all values of k ě 1, i ě 0, n ě 0.
Indeed, one proceeds by layers numbered by n, with the case n “ 0 given by the
first formula, and the case of arbitrary n given by the third one, which is valid for
k ě 2. The remaining case k “ 1 is given by the second formula.
Clearly, f pk, i, nq ď gpk, i, nq for all k ě 1, i ě 0, n ě 0, so that the function g
can be used to establish the upper bound for }φ´n pBk q}:
}φ´n pBk q} “ }φ´n pTk,0 q} “ f pk, 0, nq ď gpk, 0, nq.
To estimate the growth of gpk, i, nq, consider the finite difference gpk, i, n ` 1q ´
gpk, i, nq. Applying the recurrent relation several times, we get:
gpk, i, n ` 1q “ gpk ´ 1, 1, nq ` gpk, i ` 1, nq
“ gpk ´ 1, 1, nq ` gpk ´ 1, 1, n ´ 1q ` gpk, i ` 2, n ´ 1q
...
“ rgpk ´ 1, 1, nq ` gpk ´ 1, 1, n ´ 1q ` ¨ ¨ ¨ ` gpk ´ 1, 1, 0qs ` gpk, i ` n ` 1, 0q.
DEHN FUNCTIONS OF SUBGROUPS OF RIGHT-ANGLED ARTIN GROUPS
35
Similarly,
gpk, i, nq “ rgpk ´ 1, 1, n ´ 1q ` gpk ´ 1, 1, n ´ 2q ` ¨ ¨ ¨ ` gpk ´ 1, 1, 0qs ` gpk, i ` n, 0q.
Since by definition
gpk, i ` n ` 1, 0q “ pk ´ 1qpi ` n ` 1q ` 1,
gpk, i ` n, 0q “ pk ´ 1qpi ` nq ` 1,
we have for all k ě 2, i ě 0, n ě 0:
gpk, i, n ` 1q ´ gpk, i, nq “ gpk ´ 1, 1, nq ` pk ´ 1q.
(5)
In particular,
gpk, 1, n ` 1q ´ gpk, 1, nq “ gpk ´ 1, 1, nq ` pk ´ 1q.
If we assume by induction on k that gpk ´ 1, 1, nq is a polynomial function in n of
degree k ´ 1 (which is true for k “ 2 since gp1, i, nq “ mn ` 1), then we conclude
at once that gpk, 1, nq is a polynomial function in n of degree k.
Now the formula (5) similarly implies that gpk, i, nq is a polynomial in n of
degree k.
Therefore, }φ´n pBk q} ď gpk, 0, nq „ nk , which finishes the proof of Claim 8.10.
Proof of Proposition 8.1. According to Corollary 8.3 and Lemma 8.9, }φ˘n pAi q} ĺ
n for i “ 1, . . . , m, and according to Claims 8.8 and 8.10, }φ˘n pBk q} ĺ nk , for
k “ 1, . . . , m. Therefore grφm,k pnq ĺ nk and grpφm,k q´1 pnq ĺ nk .
8.2. Lower bounds for the growth of φ, φ´1 .
Proposition 8.12. For the automorphism φm,k and φ´1
m,k , we have:
grφm,k pnq ľ nk
and
grpφm,k q´1 pnq ľ nk .
Claim 8.13. The size of the largest Jordan block of the Jordan normal form for
´1 ab
both φab
is k ` 1.
m,k , pφm,k q
´1 ab
´1
Proof. It is sufficient to prove the claim just for φab
“ pφab
.
m,k , as pφm,k q
m,k q
ab
By direct inspection of formulas (1), we see that φm,k is represented by the
following pm ` kq ˆ pm ` kq matrix:
«
ff
I
D
m
mk
φab
,
m,k “
Okm Ckk
where Im is the identity m ˆ m matrix, Okm is the zero k ˆ m matrix, and Dmk
and Ckk are m ˆ k and k ˆ k matrices, respectively, given by the formulas:
»
fi
»
fi
1 1 ... 1
1 1 ... 1
—1 1 . . . 1ffi
—0 1 . . . 1ffi
ffi
ffi
Dmk “ —
Ckk “ —
–. . . . . . . . . . . . .fl ,
–. . . . . . . . . . . . .fl .
1 1 ... 1
0 0 ... 1
It is known that the number of Jordan blocks of a matrix A P GLpm ` k, Cq with
all eigenvalues 1 is given by the number
dim kerpA ´ Im`k q “ m ` k ´ rankpA ´ Im`k q,
36
NOEL BRADY AND IGNAT SOROKO
and the number of Jordan blocks of A with all eigenvalues 1 and size at least 2 is
given by
dim kerrpA ´ Im`k q2 s ´ dim kerpA ´ Im`k q “ rankpA ´ Im`k q ´ rankrpA ´ Im`k q2 s.
An easy computation shows that
»
«
φab
m,k
´ Im`k “
Omm
Dmk
Okm
1
Ckk
ff
,
and
where
«
pφab
m,k
2
´ Im`k q “
1
Ckk
fi
0 1 1 ... 1
—0 0 1 . . . 1ffi
—
ffi
ffi
“—
—. . . . . . . . . . . . . . . .ffi ,
–0 0 0 . . . 1fl
0 0 0 ... 0
Omm
1
Dmk
Okm
2
Ckk
ff
,
where
fi
0 0 1 2 ... k ´ 2
—0 0 0 1 . . . k ´ 3ffi
ffi
—
—. . . . . . . . . . . . . . . . . . . . . . .ffi
ffi
—
“—
1 ffi
ffi
—0 0 0 0 . . .
–0 0 0 0 . . .
0 fl
0 0 0 0 ...
0
»
»
1
Dmk
fi
0 1 2 ... k ´ 1
—0 1 2 . . . k ´ 1ffi
ffi
“—
–. . . . . . . . . . . . . . . . . . . .fl ,
0 1 2 ... k ´ 1
2
Ckk
1
2
Note that rank Ckk
“ k ´ 1, rank Ckk
“ k ´ 2, hence rankpφab
m,k ´ Im`k q “ k and
ab
2
rankrpφm,k ´ Im`k q s “ k ´ 1. Therefore,
the number of Jordan blocks for φab
m,k “ pm ` kq ´ k “ m,
the number of Jordan blocks of size ě 2 for φab
m,k “ k ´ pk ´ 1q “ 1.
This means that there is only one block of size bigger than 1, let’s denote this
size c, and there are m ´ 1 blocks of size 1. Hence, m ` k “ c ` pm ´ 1q ¨ 1 so that
c “ k ` 1.
Proof of Proposition 8.12. The Proposition follows now from Corollary 2.13 and
Claim 8.13.
8.3. Lower bounds for the growth of φab . In the proof of Theorem B in section 7 we needed a certificate for the growth of the abelianization of φm,k , i.e. an
element of the basis that realizes the maximum in the definition of the growth
functions. Now we can provide it:
Corollary 8.14. With the above notation, let B̄k be the image of the generator Bk
in the abelianization of F . Then
›
› ˇ
ˇ
ˇ ab n
ˇ
n
k
›pφm,k qn pBk q› „ ˇpφab
ˇ
ˇ
ˇ
m,k q pB̄k q 1 „ pφm,k q pB̄k q 8 „ n .
Proof. An elementary computation with the matrix from the proof of Claim 8.13
n
shows that pφab
m,k q is represented by the matrix with the following structure:
«
ff
P
I
m
mk
n
,
pφab
m,k q “
Okm Qkk
DEHN FUNCTIONS OF SUBGROUPS OF RIGHT-ANGLED ARTIN GROUPS
37
where Im is the identity m ˆ m matrix, Okm is the zero k ˆ m matrix, and Pmk
and Qkk are m ˆ k and k ˆ k matrices, respectively, given by the formulas:
»
fi
»
fi
1 c1n . . . ck´2,n ck´1,n
c1n c2n . . . ckn
—0 1 . . . ck´3,n ck´2,n ffi
—c1n c2n . . . ckn ffi
—
ffi
—
ffi
ffi
Pmk “ –
,
Qkk “ —
—. . . . . . . . . . . . . . . . . . . . . . . . . . . .ffi .
. . . . . . . . . . . . . . . . . . .fl
–0 0 . . .
1
c1n fl
c1n c2n . . . ckn
0 0 ...
0
1
Here Pmk has all rows equal to each other and Qkk is upper triangular with the
same number on each diagonal sequence of entries parallel to the main diagonal.
The numbers c1n , c2n , . . . , ckn satisfy the following identity, which follows from the
matrix multiplication rule:
`
ÿ
c`,n`1 “
cin ,
i“0
with the convention
that
`
˘ c0n “ 1. We now show by double induction on pairs
pi, nq that cin “ n`i´1
. Indeed, this equality is true for pairs pi, nq “ p0, nq with
i
arbitrary n, since c0n “ 1 by our convention, and for pi, nq “ pi, 1q with arbitrary
i, since ci1 “ 1 in the matrix representation for φab
(see the proof of Claim 8.13).
`n`i´1˘ m,k
Now suppose that the equality cin “
is already proved for all pairs pi, nq
i
with n fixed and i arbitrary, and for all pairs pi, n ` 1q with 0 ď i ď ` ´ 1. Then
for the pair p`, n ` 1q we get:
ˆ
˙ ˆ
˙ ˆ
˙
`´1
ÿ
n``´1
n``´1
n``
c`,n`1 “
cin ` c`,n “ c`´1,n`1 ` c`,n “
`
“
,
`´1
`
`
i“0
as needed.
n
In particular, the coefficients of the vector pφab
m,k q pB̄k q are: 1, c1n , . . . , ckn , with
`n`k´1˘
ckn “
being a polynomial „ nk .
k
Since, by Claim 8.8, }φnm,k pBk q} ĺ nk , we have:
ˇ
ˇ ab n
ˇ
› ˇ
›
k
n
ˇ
ˇ
ˇ
nk ľ ›φnm,k pBk q› ě ˇpφab
m,k q pB̄k q 1 ě pφm,k q pB̄k q 8 ľ n ,
and we conclude that
ˇ ab n
ˇ
› ˇ
› n
˘
k
n
ˇ
ˇ
›φm,k pBk q› „ ˇpφab
m,k q pB̄k q 1 „ pφm,k q pB̄k |8 „ n .
Remark 8.15. It can be proved in a similar manner that the same elements Bk and
ab ´1
,
B̄k also serve as certificates for the growth of automorphisms φ´1
m,k and pφm,k q
respectively. However the formulas involved are more complicated, and we don’t
need this result for our construction.
9. Open questions
We conclude our paper with two open questions.
Question 1. Does every CAT(0) free-by-cyclic group virtually embed into a RAAG?
Question 2. Do there exist finitely presented subgroups of RAAGs whose Dehn
functions are either super-exponential or sub-exponential but not polynomial?
38
NOEL BRADY AND IGNAT SOROKO
References
[1] Agol, I., Criteria for virtual fibering. J. Topol. 1 (2008), no. 2, 269–284.
[2] Agol, I., The virtual Haken conjecture. With an appendix by Ian Agol, Daniel Groves, and
Jason Manning. Doc. Math. 18 (2013), 1045–1087.
[3] Barnard, J., Brady, N., Distortion of surface groups in CAT(0) free-by-cyclic groups. Geom.
Dedicata, 120 (2006), 119–139.
[4] Bestvina, M., Brady, N., Morse theory and finiteness properties of groups. Invent. Math., 129
(1997), no. 3, 445–470.
[5] Birget, J., Ol’shanskii, A., Rips, E., Sapir, M., Isomperimetric functions of groups and computational complexity of the word problem. Ann. of Math. (2), 156 (2002), no. 2, 476–518.
[6] Brady, N., Forester, M., Snowflake geometry in CAT(0) groups. arXiv:1602.08379 [math.GR],
48 pages.
[7] Brady, N., Dehn functions and non-positive curvature, in The geometry of the word problem
for finitely generated groups. (Birkhäuser, Basel, 2007) 1–79.
[8] Brady, N., Bridson, M., There is only one gap in the isoperimetric spectrum. Geom. Funct.
Anal., 10 (2000), no. 5, 1053–1070.
[9] Bridson, M. R., Combings of semidirect products and 3–manifold groups. Geom. Funct. Anal.,
3 (1993), no. 3, 263–278.
[10] Bridson, M. R., The geometry of the word problem. Invitations to geometry and topology,
29–91, Oxf. Grad. Texts Math., 7, Oxford Univ. Press, Oxford, 2002.
[11] Bridson, M. R., On the subgroups of right-angled Artin groups and mapping class groups.
Math. Res. Lett., 20 (2013), no. 2, 203–212.
[12] Bridson, M. R., Polynomial Dehn functions and the length of asynchronously automatic
structures. Proc. London Math Soc., (3) 85 (2002), 441–466.
[13] Bridson, M. R., Gersten, S. M., The optimal isoperimetric inequality for torus bundles over
the circle. Quart. J. Math. Oxford Ser. (2), 47 (1996), no. 185, 1–23.
[14] Bridson, M. R., Haefliger, A., Metric Spaces of Non-Positive Curvature. Springer, 1999.
[15] Bridson, M. R., Pittet, Ch., Isoperimetric inequalities for the fundamental groups of torus
bundles over the circle. Geom. Dedicata, 49 (1994), no. 2, 203–219.
[16] Dison, W., An isoperimetric function for Bestvina–Brady groups. Bull. Lond. Math. Soc. 40
(2008), no. 3, 384–394.
[17] Dison, W., Riley, T. R., Hydra groups. Comment. Math. Helv. 88 (2013), no. 3, 507–540.
[18] Gersten, S. M., The automorphism group of a free group is not a CAT(0) group. Proc. Amer.
Math. Soc. 121 (1994), no. 4, 999–1002.
[19] Hagen, M. F., Wise, D. T., Cubulating hyperbolic free-by-cyclic groups: the irreducible case.
Duke Math. J. 165 (2016), no. 9, 1753–1813.
[20] Hagen, M. F., Wise, D. T., Cubulating hyperbolic free-by-cyclic groups: the general case.
Geom. Funct. Anal. 25 (2015), no. 1, 134–179.
[21] Haglund, F., Wise, D. T., Special cube complexes. Geom. Funct. Anal., 17 (2008), no. 5,
1551–1620.
[22] Horn, R.A., Johnson, C.R., Matrix Analysis. 2nd ed. Cambridge University Press, 2013.
[23] Howie, J., On the asphericity of ribbon disc complements. Trans. Amer. Math. Soc. 289
(1985), no. 1, 281–302.
[24] Levitt, G., Counting growth types of automorphisms of free groups. Geom. Funct. Anal. 19
(2009), no. 4, 1119–1146.
[25] Piggott, A., Detecting the growth of free group automorphisms by their action on the homology of subgroups of finite index. arXiv:math/0409319v1, 59 pages.
[26] Sapir, M., Birget, J.-C., Rips, E. Isoperimetric and isodiametric functions of groups. Ann. of
Math. (2), 156 (2002), no. 2, 345–466.
Department of Mathematics, University of Oklahoma, Norman, OK 73019, USA
E-mail address: [email protected]
Department of Mathematics, University of Oklahoma, Norman, OK 73019, USA
E-mail address: [email protected]
| 4math.GR
|
Deciding Circular-Arc Graph Isomorphism in
Parameterized Logspace
Maurice Chandoo1
1
Leibniz Universität Hannover, Theoretical Computer Science,
Appelstr. 4, 30167 Hannover, Germany
[email protected]
arXiv:1507.03348v3 [cs.DS] 28 Dec 2015
Abstract
We compute a canonical circular-arc representation for a given circular-arc (CA) graph which
implies solving the isomorphism and recognition problem for this class. To accomplish this we
split the class of CA graphs into uniform and non-uniform ones and employ a generalized version
of the argument given by Köbler et al. (2013) that has been used to show that the subclass of
Helly CA graphs can be canonized in logspace. For uniform CA graphs our approach works
in logspace and in addition to that Helly CA graphs are a strict subset of uniform CA graphs.
Thus our result is a generalization of the canonization result for Helly CA graphs. In the nonuniform case a specific set Ω of ambiguous vertices arises. By choosing the parameter k to be the
cardinality of Ω this obstacle can be solved by brute force. This leads to an O(k + log n) space
algorithm to compute a canonical representation for non-uniform and therefore all CA graphs.
1998 ACM Subject Classification G.2.2 Graph Theory
Keywords and phrases graph isomorphism, canonical representation, parameterized algorithm
1
Introduction
An arc is a connected set of points on the circle. A graph G is called a CA graph if every vertex
v can be assigned to an arc ρ(v) such that two vertices u, v are adjacent iff their respective
arcs intersect, i.e. ρ(u) ∩ ρ(v) 6= ∅. We call such a (bijective) mapping ρ a CA representation
of G and the set of arcs ρ(G) = {ρ(v) | v ∈ V (G)} a CA model. The recognition problem for
CA graphs is to decide whether a given graph G is a CA graph. The canonical representation
problem for CA graphs consists of computing a CA representation ρG for a given CA graph
G with the additional canonicity constraint that whenever two CA graphs G and H are
isomorphic the CA models ρG (G) and ρH (H) are identical.
Similarly, a graph is an interval graph if every vertex can be assigned to an interval on a
line such that two vertices share an edge iff their intervals intersect. It is easy to see that
every interval graph is a CA graph since every interval model is a CA model.
The class of CA graphs started to gain attraction after a series of papers in the 1970’s by
Alan Tucker. However, there is still no known better upper bound for deciding CA graph
isomorphism than for graph isomorphism in general even though considerable effort has been
done to this end. There have been two claimed polynomial-time algorithms in [9], [3] which
have been disproven in [2], [1] respectively. For the subclass of interval graphs a linear-time
algorithm for isomorphism has been described in [7]. A series of newer results show that
canonical representations for interval graphs, proper CA graphs and Helly CA graphs (a
superset of interval graphs) can be computed in logspace [4, 6, 5]. Furthermore, recognition
and isomorphism for interval graphs is logspace-hard[4] and these two hardness results carry
over to the class of CA graphs; for recognition the reduction requires a little additional work.
© Maurice Chandoo;
licensed under Creative Commons License CC-BY
Leibniz International Proceedings in Informatics
Schloss Dagstuhl – Leibniz-Zentrum für Informatik, Dagstuhl Publishing, Germany
2
Deciding Circular-Arc Graph Isomorphism in Parameterized Logspace
A1
2
ρ(i) = Ai
1
A2
ρ(G)
G
3
A5
5
4
A3
A4
Figure 1 A CA graph and representation
Our main contribution is that we extend the argument used in [5] to compute canonical
representations for HCA graphs to all CA graphs. We split the class of CA graphs into
uniform and non-uniform ones and show that the mentioned argument can be applied in both
cases using only O(k + log n) space. The parameter k describes the cardinality of an obstacle
set Ω that occurs only in the non-uniform case. This means k = 0 in the uniform case and
hence we also obtain a logspace algorithm for uniform CA graphs which are a superclass of
HCA graphs. To the best of our knowledge this is the first non-trivial algorithm to decide
isomorphism specifically for the class of CA graphs.
This paper is structured as follows. In section 2 we define CA graphs along with their
representations and recall the concept of normalized representations from [3]. In section 3
we explain what we mean by flip trick and formalize this with the notions of flip sets and
candidate functions. This idea has been used by [8] to compute a CA representation for CA
graphs in linear time and [5] has modified it to compute canonical (Helly) CA representations
for HCA graphs. In section 4 and 5 the flip trick is applied to uniform and non-uniform CA
graphs respectively.
2
Preliminaries
Given two sets A, B we say A and B intersect if A ∩ B 6= ∅. We say A, B overlap, in symbols
A G B, if A ∩ B, A \ B and B \ A are non-empty. Let A = (au,v )u,v∈V (A) , B = (bu,v )u,v∈V (B)
be two square matrices over vertex sets V (A), V (B). We say A and B are isomorphic, in
symbols A ∼
= B, if there exists a bijection π : V (A) → V (B) such that au,v = bπ(u),π(v) for
all u, v ∈ V (A); π is called an isomorphism. For two graphs G, H with adjacency matrices
∼ H if AG ∼
AG , AH we say that G =
= AH . We consider only undirected graphs without
self-loops. A graph class C is a subset of all graphs which is closed under isomorphism, i.e. if
G ∈ C and H ∼
= G then H ∈ C. We define the graph isomorphism problem for a graph class
∼ H}. For a graph G and a vertex v ∈ V (G) we
C as GI(C) = {(G, H) | G, H ∈ C and G =
define the open neighborhood N (v) of v as the set of vertices that are adjacent to v and the
closed neighborhood N [v] = N (v) ∪ {v}. For a subset of vertices V 0 ⊆ V (G) we define the
T
common neighborhood of V 0 as N [V 0 ] = v∈V 0 N [v]. We also write N [u, v, . . . ] instead of
N [{u, v, . . . }]. A vertex v is called universal if N [v] = V (G). For two vertices u =
6 v ∈ V (G)
we say that u and v are twins if N [u] = N [v]. A twin class is an equivalence class induced
by the twin relation. A graph is said to be without twins if every twin class has cardinality
one. For a graph G and S 0 ⊆ S ⊆ V (G) we define the exclusive neighborhood NS (S 0 ) as all
vertices v ∈ V (G) \ S such that v is adjacent to all vertices in S 0 and to none in S \ S 0 .
I Definition 1 (Label-independent). Let f be a function which maps graphs to a subset of
subsets of vertices, i.e. f (G) ⊆ P(V (G)). We say f is label-independent if for every pair of
isomorphic graphs G, H and all isomorphisms π from G to H it holds that
f (H) = {π(X) | X ∈ f (G)} with π(X) = {π(v) | v ∈ X}
M. Chandoo
2.1
Circular-Arc Graphs and Representations
A CA model is a set of arcs A = {A1 , . . . , An } on the circle. Let p 6= p0 be two points
on the circle. Then the arc A specified by [p, p0 ] is given by the part of the circle that is
traversed when starting from p going in clockwise direction until p0 is reached. We say that
p is the left and p0 the right endpoint of A and write l(·), r(·) to denote the left and right
endpoint of an arc in general. If A = [p, p0 ] then the arc obtained by swapping the endpoints
A = [p0 , p] covers the opposite part of the circle. We say A is obtained by flipping A. When
considering a CA model with respect to its intersection structure only the relative position
of the endpoints to each other matter. W.l.o.g. all endpoints can be assumed to be pairwise
different and no arc covers the full circle. Therefore a CA model A with n arcs can be
described as a unique string as follows. Pick an arbitrary arc A ∈ A and relabel the arcs with
1, . . . , n in order of appearance of their left endpoints when traversing the circle clockwise
starting from the left endpoint of A. Then write down the endpoints in order of appearance
when traversing the circle clockwise starting from the left endpoint of the chosen arc A. Do
this for every arc and pick the lexicographically smallest resulting string as representation
for A. For example, the smallest such string for the CA model in Fig. 1 would result from
choosing A1 (l(1), r(1), l(2), r(5), l(3), r(2), . . . ). In the following we identify A with its string
representation.
Let G be a graph and ρ = (A, f ) consists of a CA model A and a bijective mapping f
from the vertices of G to the arcs in A. Then ρ is called a CA representation of G if for all
u=
6 v ∈ V (G) it holds that {u, v} ∈ E(G) ⇔ f (u) ∩ f (v) 6= ∅. We write ρ(x) to mean the arc
f (x) corresponding to the vertex x, ρ(G) for the CA model A and for a subset V 0 ⊆ V (G)
let ρ[V 0 ] = {ρ(v) | v ∈ V 0 }. Given a set X ⊆ V (G) we write ρ+ (X) to denote ∪v∈X ρ(v). A
graph is a CA graph if it has a CA representation.
We say a CA model A has a hole if there exists a point on the circle which is not contained
in any arc in A. Every such CA model can be understood as interval model by straightening
the arcs. Therefore a graph is an interval graph if it admits a CA representation with a hole.
A CA graph G is called Helly (HCA graph) if it has a CA representation ρ such that for
T
all maxcliques, i.e. inclusion-maximal cliques, C in G it holds that v∈C ρ(v) 6= ∅. Every
interval model has the Helly property and therefore every interval graph is an HCA graph.
2.2
Normalized Representation
In [3] it was observed that the intersection type of two arcs A 6= B can be one of the following
five types: A and B are disjoint (di), A is contained in B (cd), A contains B (cs), A and B
jointly cover the circle (circle cover cc) or A and B overlap (ov) but do not jointly cover the
circle. Using these types we can associate a matrix with every CA model. An intersection
matrix is a square matrix with entries {di, ov, cs, cd, cc}. Given a CA model A we define its
intersection matrix µA such that (µA )a,b reflects the intersection type of the arcs a =
6 b ∈ A.
An intersection matrix µ is called a CA (interval) matrix if it is the intersection matrix of
some CA (interval) model.
When trying to construct a CA representation for a CA graph G it is clear that whenever
two vertices are non-adjacent their corresponding arcs must be disjoint. If two vertices u, v
are adjacent the intersection type of their corresponding arcs might be ambiguous. It would
be convenient if the intersection type for every pair of vertices would be uniquely determined
by G itself. This can be achieved by associating a graph G with an intersection matrix λG
3
4
Deciding Circular-Arc Graph Isomorphism in Parameterized Logspace
called the neighborhood matrix which is defined for all u 6= v ∈ V (G) as
di , if {u, v} ∈
/ E(G)
cd , if N [u] ( N [v]
cs , if N [v] ( N [u]
(λG )u,v = cc , if N [u] G N [v] and N [u] ∪ N [v] = V (G)
and ∀w ∈ N [u] \ N [v] : N [w] ⊂ N [u]
and ∀w ∈ N [v] \ N [u] : N [w] ⊂ N [v]
ov , otherwise
and the first case applies whose condition is satisfied.
Let µ be an intersection matrix over the vertex set V and ρ = (A, f ) where A is a CA
model and f is a bijective mapping from V to A. We say ρ is a CA representation of the
matrix µ if µ is isomorphic to the intersection matrix of A via f and denote the set of
such CA representations with N (µ). Then we say ρ is a normalized CA representation of a
graph G if ρ is a CA representation of the neighborhood matrix λG of G. An example of a
normalized representation can be seen in Fig. 1. Let us denote the set of all normalized CA
representations of G with N (G) = N (λG ).
I Lemma 2 ([3]). Every CA graph G without twins and universal vertices has a normalized
CA representation, that is N (G) 6= ∅.
For our purpose it suffices to consider only graphs without twins and universal vertices for
the same reasons as in [5]. The point is that a universal vertex can be removed from the
graph and later added as arc which covers the whole circle in the representation. For each
twin class an arbitrary representative vertex can be chosen and colored with the cardinality
of its twin class; the other twins are removed.
I Lemma 3. The canonical CA representation problem for CA graphs is logspace reducible to
the canonical CA representation problem for colored CA graphs without twins and universal
vertices.
Henceforth we assume every graph to be twin-free and without universal vertices and shall
only consider normalized representations. For two vertices u 6= v in a graph G we write u α v
instead of (λG )u,v = α. For example, u cd v indicates that the arc of u must be contained in
the arc of v for every (normalized) representation of G.
3
Flip Trick
Let A be a CA model
and X ⊆ A is a subset of arcs to be flipped. We define the resulting
CA model A(X) = A | A ∈ X ∪ A \ X. Consider a point x on the circle and let X be the
Table 1 Effects of flipping arcs in the intersection matrix
µA,B
µĀ,B
µA,B̄
µĀ,B̄
di
cs
cd
cc
cd
cc
di
cs
cs
di
cc
cd
cc
cd
cs
di
ov
ov
ov
ov
M. Chandoo
5
set of arcs that contain this point. Then after flipping the arcs in X no other arc contains
the point x and thus A(X) has a hole and therefore must be an interval model. Let µ and
µ(X) be the intersection matrices of A and A(X) respectively. It was observed in [8] that
the interval matrix µ(X) can be easily computed using µ and X as input via Table 1. With
this the problem of computing a canonical CA representation for colored CA graphs can be
reduced to the canonical interval representation problem for colored interval matrices, which
can be solved in logspace[5] (the colored part is not mentioned explicitly but can be easily
incorporated into the proof by adding the colors to the leaves of the colored ∆ tree).
The idea is that given a CA graph G if we can compute a set of vertices X as described
above then we can obtain a canonical CA representation for G by the following argument.
(X)
The neighborhood matrix λG of G is a CA matrix and the matrix λG must be an interval
(X)
matrix. Compute a canonical interval representation for λG and flip the arcs in X back.
This leads to a representation for λG and thus G. The required set X can be specified as
follows.
I Definition 4. Let G be a CA graph. Then a non-empty X ⊆ V (G) is a flip set iff there
exists a representation ρ ∈ N (G) and a point x on the circle such that v ∈ X ⇔ x ∈ ρ(v).
In fact, for the argument to obtain a canonical representation to hold it is only required
(X)
that X is chosen such that λG is an interval matrix. However, it can be shown that this is
equivalent to the above definition, see Appendix. Now, we can reframe the argument given
in [5] as follows.
I Definition 5 (Candidate function). Let C be a subset of all CA graphs and f is a function
which maps graphs to a subset of subsets of their vertices, i.e. f (G) ⊆ P(V (G)). We call f a
candidate function for C if the following conditions hold:
1. For every G ∈ C there exists an X ∈ f (G) such that X is a flip set
2. f is label-independent
I Theorem 6. If f is a candidate function for all CA graphs that can be computed in logspace
then the canonical representation problem for CA graphs can be solved in logspace.
Proof. Let G be a graph. To decide the recognition problem for CA graphs observe that
there is a flip set in f (G) iff G is a CA graph. To verify if a set X ⊆ V (G) is a flip set one
(X)
can check if λG is an interval matrix by trying to compute an interval representation.
For the representation problem let G be a CA graph. Let F be the subset of f (G) such
that every X ∈ F is a flip set. By the first condition it holds that F is non-empty. For every
X ∈ F a CA representation ρX of G can be computed by the previous argument. We return
a CA representation with the lexicographically smallest underlying model
argmin ρX (G)
{ρX | X∈F }
as canonical CA representation. To see that this is indeed a canonical representation consider
two isomorphic CA graphs G, H with FG , FH defined as F for G previously. Let π be an
isomorphism from G to H and MG = {ρX (G) | X ∈ FG } is the set of CA models induced
by the flip sets FG , similarly define MH . Then canonicity follows by showing MG = MH .
We show that MG ⊆ MH as the argument for the other direction is analogous. Let A be
a model in MG . Let X be a flip set in FG which induces A, i.e. ρX (G) = A. It must hold
∼ λH and
that π(X) ∈ f (H) since f is label-independent. From G ∼
= H it follows that λG =
(X)
(π(X))
therefore the interval matrices λG and λH
are isomorphic meaning that π(X) ∈ FH
6
Deciding Circular-Arc Graph Isomorphism in Parameterized Logspace
since it is a flip set as well. As the interval representations of both interval matrices have
identical underlying models due to canonicity it follows that they remain so after flipping X
resp. π(X). Therefore it holds that A ∈ MH .
This works in logspace since f and the representation ρX for every flip set X can be
computed in logspace.
J
So, we have reduced the problem of computing a canonical CA representation for CA graphs
to the problem of computing a candidate function for CA graphs.
Let C and C 0 be two graph classes that partition all CA graphs. If fC , fC 0 are candidate
functions for C, C 0 respectively then f (G) = fC (G) ∪ fC 0 (G) is a candidate function for all CA
graphs. That f is label-independent follows from label-independent functions being closed
under taking unions. The crux here is that we do not need to be able to distinguish if a
CA graph G is in C or C 0 . Hence, in the next two sections we consider two such classes that
partition all CA graphs while avoid dealing with recognition of these two classes.
We complete this section by stating the candidate function used in [5] to canonize HCA
graphs and explain why it is a candidate function for this subclass of CA graphs.
[
fHCA (G) =
N [u, v]
u,v∈V (G)
fHCA always returns at least one flip set for an HCA graph because all maxcliques in an
HCA graph are flip sets due to the Helly property and there exists at least one maxclique in
every HCA graph that can be characterized as the common neighborhood of two vertices[5].
However, neither of these two properties hold for CA graphs in general.
It remains to argue that fHCA can be computed in logspace and is label-independent.
Since the same arguments have to be made for the two candidate functions devised in the
next sections we introduce a tool that facilitates this and demonstrate it for fHCA .
I Definition 7. Let ϕ be a first-order (FO) formula over graph structures with k + 1 free
variables. Then we define the function fϕ for a graph G as:
[
fϕ (G) =
{u ∈ V (G) | G |= ϕ(v1 , . . . , vk , u)}
(v1 ,...,vk )∈V (G)k
I Lemma 8. For every FO formula ϕ over graph structures the function fϕ is computable
in logspace and label-independent.
To compute fϕ we can successively evaluate ϕ and take the union of the results, which can
be both done in logspace. To show that fϕ is label-independent it suffices to apply structural
induction to ϕ. The full argument for this can be found in the Appendix. Then fHCA is
computed by the FO formula ϕ(u, v, x) that states that x ∈ N [u, v]. By Lemma 8 it follows
that fHCA is logspace-computable and label-independent.
4
Uniform CA graphs
The difficulty when trying to compute flip sets for CA graphs in general is that for a CA
graph G there might be different normalized CA representations such that a set of vertices
shares a common point in one representation but not in an other one. We show a subset of
CA graphs, namely the uniform CA graphs, where this issue does not occur therefore making
it easy to compute flip sets.
For a CA graph G consider an arbitrary vertex u. Looking at the neighbors of u we
can try to compute the flip sets Xu,1 , Xu,2 specified in Fig. 2. Both sets contain u and all
M. Chandoo
7
vertices that contain u or form a circle cover with u. The vertices that overlap with u belong
to either Xu,1 or Xu,2 depending on the side they overlap from with u. Since we cannot
determine left and right from the neighborhood matrix we want to express an equivalence
relation ∼u which states that two vertices overlap from the same side with u. With the two
equivalence classes induced by ∼u the two flip sets can be expressed.
Given two vertices x, y that both overlap with u it is for instance easy to see that they
must overlap from different sides with u if they are disjoint. The only intersection type
between x, y for which the situation is not immediately clear is ov as further distinctions are
required. An ov-triangle is a set of three pairwise overlapping vertices. If x and y overlap
then x, y, u form such an ov-triangle. Consider the possible normalized representations for an
ov-triangle. The three vertices can either all jointly cover the circle or be a set of overlapping
intervals. In the first case they overlap pairwise but their overall intersection is empty thus
we call this a non-Helly triangle and the second case an interval triangle. For an interval
triangle there are three different possible representations up to reflection depending on which
of the three vertices is placed in-between the other two. If x, y, u is an ov-triangle then x
and y overlap from the same side with u iff x, y, u form an interval triangle and u is not
in-between x, y. We show that it is easy to derive this information in the case of a uniform
CA graph G as it does not depend on a representation of G.
I Definition 9. Let G be a graph. An ov-triangle T is in the set ∆G if the following holds:
S
1.
N [v] = V (G)
v∈T
2. For all x ∈ T it holds that if a vertex v ∈ NT (x) then v cd x
I Definition 10 (Uniform CA graph). A CA graph G is uniform if for all ov-triangle T in G
and ρ ∈ N (G) it holds that:
T ∈ ∆G ⇒ ρ[T ] is a non-Helly triangle
The idea behind the definition of ∆G is that it captures the properties that an ov-triangle
must satisfy in the graph if it can be represented as non-Helly triangle. The definition of
uniform CA graphs guarantees us that an ov-triangle T ∈ ∆G can never be represented as
interval triangle. Now, we can show that for the class of uniform CA graphs the property of
being a non-Helly triangle and the in-between predicate for interval triangles is invariant
across all normalized representations.
I Lemma 11. Let G be a uniform CA graph. Then the following statements are equivalent
for every ov-triangle T :
1. T ∈ ∆G
Xu,1
Xu,2
u
Figure 2 Uniform target flip sets
8
Deciding Circular-Arc Graph Isomorphism in Parameterized Logspace
2. ∃ρ ∈ N (G) : ρ[T ] is a non-Helly triangle
3. ∀ρ ∈ N (G) : ρ[T ] is a non-Helly triangle
Proof. This immediately follows from the definition of ∆G and uniform CA graphs.
J
As a contrasting example of a (non-uniform) CA graph for which being a non-Helly triangle
depends on the representation consider the graph G obtained by taking the complement of
the disjoint union of three K2 ’s (a triforce with a circle around the outer corners). Every edge
in G is an ov-entry in λG and a CA model for G is given by precisely two non-Helly triangles.
The possible assignments of the vertices to the arcs that follow from the automorphisms of
G yield the different representations.
I Definition 12. Let G be a graph and T = {u, v, w} is an ov-triangle with T ∈
/ ∆G . We
say v is in-between u, w if at least one of the following holds:
1. NT (u), NT (w) 6= ∅
2. NT (u, w) 6= ∅ and there exists z ∈ NT (u, w) such that {u, w, z} ∈ ∆G
I Lemma 13. Let G be a uniform CA graph and T = {u, v, w} is an ov-triangle with
T ∈
/ ∆G . Then the following statements are equivalent:
1. v in-between u, w
2. ∃ρ ∈ N (G) : ρ(v) ⊂ ρ(u) ∪ ρ(w)
3. ∀ρ ∈ N (G) : ρ(v) ⊂ ρ(u) ∪ ρ(w)
Proof. “2 ⇒ 1”: There exists ρ ∈ N (G) such that ρ(v) ⊂ ρ(u) ∪ ρ(w). For all x 6= y ∈ T
it must hold that N [x] \ N [y] 6= ∅ due to the fact that x ov y implies N [x] G N [y]. For this
to be true NT (x) or NT (x, z) must be non-empty for z ∈ T \ {x, y}. It follows that NT (u)
and NT (w) or NT (u, w) must be non-empty since NT (v) = ∅ due to ρ(v) ⊂ ρ(u) ∪ ρ(w). If
z ∈ NT (u, w) then for ρ(z) to intersect with ρ(u) and ρ(w) but not with ρ(v) implies that
ρ[u, w, z] forms a non-Helly triangle and therefore {u, w, z} ∈ ∆G by Lemma 11.
“1 ⇒ 3”: Assume there exists a ρ ∈ N (G) such that ρ(u) ⊂ ρ(v) ∪ ρ(w). This implies that
NT (u) = ∅ and therefore there exists a z ∈ NT (u, w) such that {u, w, z} ∈ ∆G . By Lemma
11 it follows that ρ[u, w, z] must be a non-Helly triangle and ρ(z) must be disjoint from ρ(v),
contradiction.
“3 ⇒ 2”: is clear.
J
Lemma 11 and 13 state a fact of the form that if a property holds in one representation then
it holds in all representations hence the name uniform CA graphs.
The α-neighborhood of a vertex u in a graph G is defined as N α (u) = {v ∈ N (u) | u α v}
for α ∈ {ov, cd, cs, cc}. Now, we can define the aforementioned equivalence relation ∼u and
state the candidate function for uniform CA graphs.
I Definition 14. Given a graph G and vertex u ∈ V (G) we define the relation ∼u on N ov (u)
such that x ∼u y holds if one of the following applies:
1. x = y
2. x contains or is contained in y
3. x and y overlap, {x, y, u} ∈
/ ∆G and u is not in-between x, y
I Lemma 15. Let G be a uniform CA graph and u ∈ V (G). Then x ∼u y holds iff x and y
overlap from the same side with u in every ρ ∈ N (G).
M. Chandoo
9
Proof. "⇒": If x, y are in a contained/contains relation this is clear. For the third condition
it holds that for all ρ ∈ N (G) ρ[x, y, u] is an interval triangle and x or y is in-between the
other two. It follows that x, y overlap from the same side with u.
"⇐": If λx,y ∈ {di, cc} this is clear. If x, y overlap then they either form a non-Helly triangle
or u is in-between x, y. In both cases x, y overlap from different sides with u.
J
I Theorem 16. The following mapping is a candidate function for uniform CA graphs and
can be computed in logspace:
fU (G) =
[
{u} ∪ N cd (u) ∪ N cc (u) ∪ {y ∈ N ov (u)|x ∼u y}
u∈V (G)
x∈N ov (u)
Proof. To show that fU (G) is a candidate function we have to prove that for every uniform
CA graph G there always exists a flip set X ∈ fU (G) and that fU is label-independent. In
fact, the even stronger claim holds that for all uniform CA graphs G every set in fU (G) is
a flip set. Let X ∈ fU (G) via some u ∈ V (G) and x ∈ N ov (u). Then the set of vertices in
X correspond to one of the two flip sets shown in Fig. 2. The correctness for the subset of
vertices in X overlapping with u follows from Lemma 15.
To show that fU (G) can be computed in logspace and is label-independent we apply
Lemma 8. We can rewrite fU as
fU (G) =
[
{z ∈ V (G) | G |= ϕ(u, x, z)}
u,x∈V (G)
with G |= ϕ(u, x, z) iff x ∈ N ov (u) and z ∈ {u} ∪ N cd (u) ∪ N cc (u) ∪ {y ∈ N ov (u)|x ∼u y}. It
remains to check that the entries in the neighborhood matrix, α-neighborhoods, exclusive
neighborhoods, ∆G , in-between and ∼u can be expressed in FO logic.
J
I Corollary 17. A canonical CA representation for uniform CA graphs can be computed in
logspace.
I Theorem 18. Helly CA graphs are a strict subset of uniform CA graphs.
Proof. First, we show "⊆" by contradiction. Assume there exists a Helly CA graph G
which is non-uniform. For G to be non-uniform there must exist an ov-triangle T ∈ ∆G
and a representation ρ ∈ N (G) such that T is represented as interval triangle in ρ. Let
T = {u, v, w} and assume w.l.o.g. that ρ(v) ⊂ ρ(u) ∪ ρ(w) (v is in-between u, w). It follows
that N [u] ∪ N [w] = V (G). Since λu,w =
6 cc there must be a u0 ∈ N [u] \ N [w] such that
N [u0 ] \ N [u] 6= ∅. This means there exists a w0 ∈ N [u0 , w] \ N [u]. As N [u0 ] G N [w0 ] it follows
that u0 and w0 overlap. For u0 it must hold that it is either in NT (u) or NT (u, v). If it
is in NT (u) then by the second condition of ∆G it follows that u0 must be contained in
u, contradiction. For the same reason w0 is in NT (v, w). It follows that u0 , v, w0 form an
ov-triangle and must be represented as non-Helly triangle in ρ. This contradicts that G is a
Helly CA graph.
To see that this inclusion is strict consider the graph G obtained by taking a triangle
T and attaching a new vertex to each vertex in T (also known as net graph). In every
representation ρ ∈ N (G) it must hold that T is represented as non-Helly triangle since
NT (v) 6= ∅ for all v ∈ T . For this reason G cannot be a Helly or a non-uniform CA graph. J
10
Deciding Circular-Arc Graph Isomorphism in Parameterized Logspace
5
Non-Uniform CA graphs
From the definition of uniform CA graphs it follows that a CA graph G is non-uniform if
there exists an ov-triangle T in ∆G and a ρ ∈ N (G) such that T is represented as interval
triangle in ρ. We call the pair (T, ρ) a witness for the non-uniformity of G and also say G
is non-uniform via (T, ρ). Additionally, for such a witness pair (T, ρ) we call T maximal if
there exists no T 0 6= T such that G is non-uniform via (T 0 , ρ) and ρ+ (T ) ⊂ ρ+ (T 0 ). Such
a maximal T for a given ρ must always exist. For the purpose of computing a candidate
function for this class we can assume that for a given non-uniform CA graph G we are
supplied with an ov-triangle T such that there exists a ρ ∈ N (G) with (T, ρ) being a witness
for G and T is maximal. This is justified by the fact that we can iterate over all ov-triangle
T ∈ ∆G trying to compute flip sets knowing that for at least one such T these conditions are
met. Additionally, we write T as ordered triple (u, v, w) to indicate that v is in-between u
and w in ρ.
Let G be non-uniform via (T, ρ). Consider for a vertex x ∈ V (G) \ T what the possible
relations between ρ(x) and ρ+ (T ) are. For instance, ρ(x) cannot be disjoint from ρ+ (T )
because this implies that x ∈
/ ∪t∈T N [t] and therefore T ∈
/ ∆G . Also, ρ(x) cannot contain
+
ρ (T ) as this would mean that x is a universal vertex.
I Definition 19. Let G be a non-uniform CA graph via (T, ρ). The set of normalized
representations that agree with ρ on T is:
NρT (G) = {ρ0 ∈ N (G) | ρ0 [T ] = ρ[T ]}
I Definition 20. Let G be a non-uniform CA graph via (T, ρ) and α ∈ {ov, cc, cd}. We say
x ∈ V (G) \ T is an α-arc in ρ0 if
ρ0 (x) α ρ+ (T )
for some ρ0 ∈ NρT (G). We call x an unambiguous α-arc if the above condition holds for all
ρ0 ∈ NρT (G).
I Definition 21. Let G be a graph and T = (u, v, w) ∈ ∆G . Then we define the following
1
u
2
3
4
15
5
25
v
w
35
14
24
13
x
l(x)
∈2
r(x)
∈5
45
55
34
44
23
33
Figure 3 Possible positions of the endpoints of x relative to T = {u, v, w}
12
22
11
M. Chandoo
11
sets w.r.t. T :
Γov,u = {x ∈ V (G) \ T | x ov u, v , x di w}
Γov,w = {x ∈ V (G) \ T | x ov v, w , x di u}
Γov
Γcc
= Γov,u ∪ Γov,w
= x ∈ V (G) \ T
x ∈ V (G) \ T
x ov u, w , x di v or
∃a ∈ T : x cc a
x ov u, w , x cs v or
∃a ∈ T : x cd a
Γcd
=
Ω
= {x ∈ V (G) \ T | x ov u, v, w}
I Lemma 22. If G is a non-uniform CA graph via (T, ρ) such that T is maximal then
T, Γov , Γcc , Γcd , Ω partition V (G).
Proof. It is not hard to see that these sets do not overlap. To show that every vertex in
G belongs to one of these sets we need to check all possible positions of the endpoints of
a vertex x not in T relative to T , consider Fig. 3. We know every vertex x ∈
/ T must be
represented as α-arc for some α ∈ {ov, cd, cc}. For α ∈ {cd, cc} it must hold that both
endpoints of x must be in one of the intervals 1–5. The exemplary x depicted in Fig. 3 is a
cd-arc in the given representation and overlaps with u, v, w. If x is a cd-arc then the number
of the interval in which its left endpoint is situated must be less than or equal that of its
right endpoint (in our example 2 ≤ 5). If it is a cc-arc then the right endpoint must come
before the left. The graph on the right encodes all possible placements of the two endpoints
and it can be verified case-by-case that an α-arc for α ∈ {cd, cc} will occur in either Γα or
Ω. It remains to argue that an ov-arc can only have the intersection structure denoted by
Γov . W.l.o.g. assume that x is an ov-arc that overlaps from u’s side with T . It holds that
the right endpoint of x is in one of the five intervals and the left endpoint must be in none of
these five intervals. If r(x) ∈ 1 then x ∈ NT (u) and by the second condition of ∆G it must
hold that x is contained in u, contradiction. If r(x) ∈ 2 then x overlaps with u, v and is
disjoint with w and therefore x ∈ Γov,u . If r(x) ∈ {3, 4} then T = {u, v, w} is not maximal
since {x, v, w} ∈ ∆G . If r(x) ∈ 5 then it can be shown that there must be a circle cover entry
in the neighborhood matrix for x, w which contradicts that they must overlap.
J
I Lemma 23. Let G be a non-uniform CA graph via (T, ρ) and T is maximal. All vertices
in Γα are unambiguous α-arcs for α ∈ {ov, cd, cc}.
Proof. The intersection structure of a vertex x ∈ Γα with T dictates the positioning of the
endpoints relative to T in every ρ0 ∈ NρT (G). It follows that this placement of the endpoints
of x must hold for all ρ0 ∈ NρT (G) and therefore x is an unambiguous α-arc.
J
For a vertex x ∈ Ω the possible placements of its endpoints to satisfy the intersection
structure with T can be one of the following four types (cd, 14), (cd, 25), (cc, 14), (cc, 25).
For example x in Fig. 3 is type (cd, 25) and flipping x would lead to type (cc, 25).
Let G be non-uniform via (T, ρ) and T = (u, v, w) is maximal. Then the two target flip
sets Xu , Xw we want to compute in the non-uniform case are immediately before the left
endpoint and immediately after the right endpoint of ρ+ (T ) and must be of the form
Xi = Γov,i ∪ Γcc ∪ Ω0
for i ∈ {u, w} and some subset Ω0 ⊆ Ω.
12
Deciding Circular-Arc Graph Isomorphism in Parameterized Logspace
I Definition 24. Let G be a non-uniform CA graph via (T, ρ) and T is maximal. We call a
subset Ω0 ⊆ Ω cc-realizable w.r.t. (T, ρ) if there exists a ρ0 ∈ NρT (G) such that for all x ∈ Ω
it holds that x is a cc-arc in ρ0 iff x ∈ Ω0 .
In other words, a cc-realizable set is a subset of Ω such that all of its vertices can be
represented as cc-arc in a normalized representation. By finding such a set we can construct
two flip sets by adding the cc-vertices in Γcc and one side of the ov-vertices as described
above. Now, the challenge consists in finding such a cc-realizable subset. A way to solve this
is to parameterize our input by the cardinality of Ω and try all possibilities. Since Ω and its
cardinality depend on the particular ov-triangle T chosen, which we do not know a priori,
we can use the following set which is a superset of every possible Ω and thus bounds the
cardinality:
KG = {x ∈ V (G) | ∃{u, v, w} ∈ ∆G s.t. x ov u, v, w}
I Theorem 25. The following mapping is a candidate function for non-uniform CA graphs
and can be computed in O(k + log n) space for k = |KG |:
[
fN (G) =
{Γov,u ∪ Γcc ∪ Ω0 }
(u,v,w)∈∆G
Ω0 ⊆Ω
where Γov,u , Γcc , Ω are taken w.r.t. T = (u, v, w).
Proof. To show that there always exists a flip set X ∈ fN (G) for a non-uniform CA graph
G we argue as follows. Let G be non-uniform via (T, ρ) and T = (u, v, w) is maximal.
Let Ω0 be a cc-realizable set w.r.t. (T, ρ), which must exist since G is non-uniform. Then
X = Γov,u ∪ Γcc ∪ Ω0 w.r.t. T is one of the target flip sets described previously. This means
there exists a ρ ∈ N (G) such that X describes the set of arcs that contain a point right
before the left endpoint of ρ+ (T ) or a point right after the right endpoint of ρ+ (T ).
To see that fN (G) is label-independent a formula ϕ(Ω0 , u, v, w, x) can be constructed that
is true iff x ∈ Γov,u ∪ Γcc ∪ Ω0 where Ω0 is a second-order set variable. Note, that |Ω| ≤ |KG |.
Therefore this works in O(k + log n) space since one can iterate over all 2k subsets of Ω using
k bits and then apply the argument in Lemma 8 via ϕ which requires additional O(log n)
space.
J
Conclusion
We showed how to canonically, or in our terms label-independently, compute flip sets for
CA graphs to acquire canonical CA representations. The properties of uniform CA graphs
enable us to do this easily in logspace. In the case of non-uniform CA graphs, however, it
seems that the cc-realizable sets pose a non-trivial obstacle when trying to compute flip sets.
The only simple remedy appears to be the proposed parameterization that enables us to use
brute force. Changing the target flip sets does not seem to improve upon this situation. As
a consequence, we suggest to investigate the space of cc-realizable sets. Given the restricted
structure of non-uniform CA graphs this could be a reasonable first step towards deciding
isomorphism for CA graphs in polynomial time.
Additionally, in [8] it was shown how to compute flip sets for CA graphs in linear time
without the canonicity constraint. Can this be done in logspace as well? This would mean
that recognition of CA graphs is logspace-complete.
Acknowledgments We thank the anonymous reviewers for their helpful comments on earlier
drafts of this paper.
M. Chandoo
13
References
1
2
3
4
5
6
7
8
9
6
Andrew Curtis, Min Chih Lin, Ross McConnell, Yahav Nussbaum, Francisco Soulignac,
Jeremy Spinrad, and Jayme Szwarcfiter. Isomorphism of graph classes related to the
circular-ones property. Discrete Mathematics and Theoretical Computer Science, 15(1),
2013.
Elaine Marie Eschen. Circular-arc Graph Recognition and Related Problems. PhD thesis,
Vanderbilt University, Nashville, TN, USA, 1998. UMI Order No. GAX98-03921.
Wen-Lian Hsu. O(M · N ) algorithms for the recognition and isomorphism problems on
circular-arc graphs. SIAM J. Comput., 24(3):411–439, June 1995.
Johannes Köbler, Sebastian Kuhnert, Bastian Laubner, and Oleg Verbitsky. Interval graphs:
Canonical representations in logspace. SIAM J. Comput., 40(5):1292–1315, 2011.
Johannes Köbler, Sebastian Kuhnert, and Oleg Verbitsky. Helly circular-arc graph isomorphism is in logspace. In MFCS 2013, volume 8087 of Lecture Notes in Computer
Science, pages 631–642. Springer Berlin Heidelberg, 2013.
Johannes Köbler, Sebastian Kuhnert, and Oleg Verbitsky. Solving the Canonical Representation and Star System Problems for Proper Circular-Arc Graphs in Logspace. In FSTTCS,
volume 18, pages 387–399. Schloss Dagstuhl, 2012.
George S. Lueker and Kellogg S. Booth. A linear time algorithm for deciding interval graph
isomorphism. J. ACM, 26(2):183–195, April 1979.
Ross M. McConnell. Linear-time recognition of circular-arc graphs. Algorithmica, 37(2):93–
147, 2003.
Tsong-Ho Wu. An O(n3 ) Isomorphism Test for Circular-Arc Graphs. PhD thesis, SUNY
Stony Brook, New York, NY, USA, 1983.
Appendix
Proof of Lemma 8. fϕ can be computed in logspace by successive evaluation of ϕ and taking
the union of the resulting sets. To prove that fϕ is label-independent we use the following
claim that can be verified by induction, see Lemma 26. For all FO formulas ψ and isomorphic
graphs G, H it holds that
G |= ψ(v1 , . . . , vk ) ⇐⇒ H |= ψ(π(v1 ), . . . , π(vk ))
for all isomorphisms π from G to H and assignments (v1 , . . . , vk ) ∈ V (G)k . Now, we
must argue that X ∈ fϕ (G) implies π(X) ∈ fϕ (H). The other direction follows from a
symmetrical argument. Let X ∈ fϕ (G) then there exist (v1 , . . . , vk ) ∈ V (G)k such that
X = {u ∈ V (G) | G |= ϕ(v1 , . . . , vk , u)}. This means
π(X) = {π(u) | u ∈ V (G) and G |= ϕ(v1 , . . . , vk , u)}
We can replace u ∈ V (G) with u0 ∈ V (H) and by the previous claim rewrite the above set as
π(X) = {u0 ∈ V (H) | H |= ϕ(π(v1 ), . . . , π(vk ), u0 )}
This concludes that π(X) ∈ fϕ (H).
J
I Lemma 26. For every FO formula ϕ with k free variables over graph structures and
isomorphic graphs G, H it holds that:
G |= ϕ(v1 , . . . , vk ) ⇐⇒ H |= ϕ(π(v1 ), . . . , π(vk ))
for all isomorphisms π from G to H and every assignment (v1 , . . . , vk ) ∈ V (G)k .
14
Deciding Circular-Arc Graph Isomorphism in Parameterized Logspace
Proof. We show this by structural induction over FO formulas. For the base case ϕ =
E(xi , xj ) the statement is clear, i.e.
G |= E(u, v) ⇐⇒ H |= E(π(u), π(v))
for all isomorphisms π and u, v ∈ V (G). For the inductive step we consider the boolean
connectives ¬, ∧, ∨ and the quantifiers ∃, ∀. Let ϕ = ¬ψ(x1 , . . . , xk ).
G |= ¬ψ(v1 , . . . , vk )
⇐⇒ G 6|= ψ(v1 , . . . , vk )
I.H.
⇐⇒ H 6|= ψ(π(v1 ), . . . , π(vk ))
⇐⇒ H |= ¬ψ(π(v1 ), . . . , π(vk ))
holds for all isomorphisms π from G to H and v1 , . . . , vk ∈ V (G). For the cases ∧ and ∨ this
is similar. Let ϕ = ∃xψ(x, x1 , . . . , xk ).
G |= ∃xψ(x, v1 , . . . , vk )
⇐⇒ there exists v ∈ V (G) such that G |= ψ(v, v1 , . . . , vk )
I.H.
⇐⇒ H |= ψ(π(v), π(v1 ), . . . , π(vk ))
⇐⇒ H |= ∃xψ(x, π(v1 ), . . . , π(vk ))
for all isomorphisms π from G to H and v1 , . . . , vk ∈ V (G). A similar argument holds for
the ∀-case.
J
(X)
I Lemma 27. Given a simple CA graph G it holds that X ⊆ V (G) is a flip set iff λG
an interval matrix.
is
Proof. If X is a flip set then there exists a ρ ∈ N (G) and a point x on the circle such that
exactly all vertices in X have the point x in common in ρ. The resulting CA representation
ρ(X) after flipping these arcs is an interval representation since ρ(X) (G) must have a hole at
(X)
x. Since λG must be isomorphic to the intersection matrix of ρ(X) (G) via ρ it follows that
it is an interval matrix.
(X)
For the other direction let X ⊆ V (G) such that λG is an interval matrix. To reach a
contradiction assume that X is not a flip set. Then for every ρ ∈ N (G) and all points x on
the circle it holds that there exists a v ∈ V (G) such that either v ∈ X or x ∈ ρ(v). Since
(X)
(X)
λG is an interval matrix it holds that there exists a ρ1 ∈ N (λG ) such that ρ1 has a hole.
(X)
By flipping the arcs in X in ρ1 we acquire the representation ρ1 ∈ N (λG ) for G. It holds
(X)
that for every point x on the circle there exists a v ∈ V (G) such that v ∈ X and x ∈
/ ρ1 (v)
(X)
(X)
or v ∈
/ X and x ∈ ρ1 (v). It follows that after flipping the set of arcs X in ρ1 back that
no hole can exists which contradicts that ρ1 has been an interval representation.
J
| 8cs.DS
|
1
Aviation Time Minimization of UAV for Data
Collection over Wireless Sensor Networks
Jie Gong, Member, IEEE, Tsung-Hui Chang, Senior Member, IEEE, Chao
arXiv:1801.02799v1 [cs.IT] 9 Jan 2018
Shen, Member, IEEE, Xiang Chen, Member, IEEE
Abstract
In this paper, we consider a scenario where an unmanned aerial vehicle (UAV) collects data from
a set of sensors on a straight line. The UAV can either cruise or hover while communicating with the
sensors. The objective is to minimize the UAV’s total aviation time from a starting point to a destination
while allowing each sensor to successfully upload a certain amount of data using a given amount of
energy. The whole trajectory is divided into non-overlapping data collection intervals, in each of which
one sensor is served by the UAV. The data collection intervals, the UAV’s navigation speed and the
sensors’ transmit powers are jointly optimized. The formulated aviation time minimization problem is
difficult to solve. We first show that when only one sensor node is present, the sensor’s transmit power
follows a water-filling policy and the UAV aviation speed can be found efficiently by bisection search.
Then we show that for the general case with multiple sensors, the aviation time minimization problem
can be equivalently reformulated as a dynamic programming (DP) problem. The subproblem involved
in each stage of the DP reduces to handle the case with only one sensor node. Numerical results present
insightful behaviors of the UAV and the sensors. Specifically, it is observed that the UAV’s optimal
speed is proportional to the given energy and the inter-sensor distance, but inversely proportional to the
data upload requirement.
J. Gong is with Guangdong Key Laboratory of Information Security Technology, School of Data and Computer Science, Sun
Yat-sen University, Guangzhou 510006, China. Email: [email protected].
T.-H. Chang is with School of Science and Engineering, The Chinese University of Hong Kong, Shenzhen, Shenzhen 518172,
China. Email: [email protected]
C. Shen is with State Key Lab of Rail Traffic Control and Safety, Beijing Jiaotong University, Beijing, China. Email:
[email protected]
X. Chen is with the School of Electronics and Information Engineering, Sun Yat-Sen University, Guangzhou 510006, China,
SYSU-CMU Shunde International Joint Research Institute and Key Lab of EDA, Research Institute of Tsinghua University in
Shenzhen, Shenzhen 518057, China. Email: [email protected].
2
I. I NTRODUCTION
Recently, wireless communication with unmanned aerial vehicles (UAVs) [1] has been considered as a promising technology to expand network coverage and enhance system throughput, by
leveraging the UAVs’ high mobility [2] and line-of-sight (LOS) dominated air-ground channels
[3]. One of the key applications is wide-area data collection in wireless sensor networks [4].
Conventionally, each sensor node delivers their monitored data to a fusion center via multi-hop
transmissions. Hence, a sensor node requires to not only transmit its own data, but also relay the
others’. As a consequence, the sensors’ battery may drain quickly and the multi-hop network
connection may be lost. By using the UAVs as mobile fusion centers, every sensor node can
directly send observed data to a UAV. In addition, the LOS channel condition results in higher data
rate for ground-to-air transmissions compared with ground-to-ground transmissions. However,
as UAVs are energy constrained due to the limited on-board battery, it is paramount to shorten
the aviation time needed for a data collection mission.
Different from the conventional communication techniques, there is a trajectory optimization issue for UAV-aided wireless communications. To improve network connectivity, UAVs’
deployment and movement were optimized to track the network topology in [5]. In [6], offline
path planning of UAVs was addressed for collision avoidance and fuel efficiency. Joint UAV
deployment and trajectory optimization problem was solved in [7] with a quantization theory
approach, and joint trajectory and communication power control for multiple UAVs was studied
in [8]. In addition, UAVs are widely used as mobile relays. Reference [9] studied the throughput
maximization problem for a UAV relay and showed that the uplink power of users should follow
a “staircase” water filling structure. In [10], joint optimization of multi-UAV beamforming and
relay positions for throughput maximization was studied based on stochastic optimization techniques. A round trip “load-carry-and-deliver” protocol was tested and evaluated by experiments
in [11]. Besides serving as relays, UAVs can also be used as mobile base stations (BSs) for
emergent communications. BS placement was optimized in the 2D space [12] and 3D space [13],
respectively, to minimize the required number of mobile BSs while maximizing their coverage.
The coverage of UAVs as mobile BSs was analytically studied in [14] considering inter-UAV
interference and beamwidth design.
In addition, there has been a growing research interest in applying UAV for data collection and
dissemination in wireless sensor networks. The aerial link characterization based on practical
3
protocols and experiments was given in [15]. Reference [16] considered data collection via uplink
transmission, and proposed to mitigate multi-sensor interference by adjusting the UAV heading
and beamforming. Adaptive modulation strategy was adopted in [17] to improve energy efficiency
of sensor nodes while guaranteeing user fairness. To avoid contention due to simultaneous
data transmissions from multiple sensor nodes, a priority-based frame selection scheme was
proposed in [18]. In [19], wake-up and sleep adaptation was applied for sensor nodes and UAV’s
trajectory optimization was jointly considered to minimize the sensors’ energy consumption.
One dimensional information dissemination problem is considered in [20], where the sensors
are served in cyclical TDMA mode, and the service regions for all sensors are optimized.
It is worthwhile to note that most of the existing works mentioned above focus on enhancing
energy efficiency or spectrum efficiency of sensor nodes, but overlook the fact that the limited
aviation energy of UAVs is one of the fundamental bottlenecks in UAV-aided wireless networks.
As a matter of fact, the dominant energy consumption of a UAV lies in the propulsion control
system that accelerates the UAV and maintains its aviation height. In [21], a UAV’s energy
consumption was modeled as a function of aviation speed and operation conditions such as
climbing, hovering, and so on. A UAV trajectory optimization problem with detailed propulsion
energy consumption considering both velocity and acceleration was studied in [22]. However,
as the UAV’s energy consumption model is quite complex, the problems are difficult to be
optimally solved. Intuitively, the energy consumption of a UAV is proportional to its aviation
time. Therefore, the energy minimization problem can be handled alternatively by formulating an
aviation time minimization problem. The UAV mission completion time minimization problem
was studied in [23] by optimizing its trajectory subject to a link quality constraint between
ground base station (GBS) and UAV in cellular networks. While in wireless sensor networks,
there is still a lack of research efforts to consider both the UAV and sensors’ energy consumption
as well as the quality of service of the sensor nodes at the same time.
In this paper, we study an aviation time minimization problem for a UAV which collects data
from a set of energy constrained ground sensors. Each of the sensors wants to upload a certain
amount of data to the UAV. The UAV can collect data either during navigation or hovering.
We assume that the sensors are located on a line and the UAV’s trajectory is divided into nonoverlapping data collection intervals, each of which is dedicated to data collection from one
sensor node. The objective is to minimize the total aviation time of the UAV for flying from
an initial point to a destination by jointly optimizing the division of intervals, the UAV speed,
4
as well as the sensors’ transmission power. The contributions of this paper are summarized as
follows.
•
The formulated UAV aviation time minimization problem is intrinsically difficult. We first
consider the single-senor scenario. While the problem is still difficult when only one sensor
node is present, we reveal some insightful structures for the optimal solution. Specifically, we
present an explicit condition on the feasibility of the problem. When the problem is feasible
and if the data collection interval is given, we show that the optimal power allocation of the
sensor follows a water-filling solution and the optimal speed can be efficiently obtained via
bisection search. The data collection interval can be numerically found via a two-dimensional
search.
•
The algorithm for solving the single-sensor case can be extended for solving the general
scenario with multiple sensor nodes. In particular, by judiciously exploiting the problem
structure, we show that the aviation time minimization problem with multiple sensors can
be equivalently formulated as a dynamic programming (DP) problem. In each stage of the
DP, the optimal data collection interval for one sensor node is searched and the algorithm for
the single-sensor case is used for finding the optimal UAV speed and sensor’s transmission
power.
•
Numerical results illustrate the optimal behaviors of the UAV and the sensor nodes under
different scenarios. In particular, the UAV’s optimal speed is proportional to the sensors’
energy budgets and the inter-sensor distance, but inversely proportional to the amount of
data to upload. For the randomly distributed sensors with random amount of data and energy,
the average minimum aviation time increases with the average amount of data and decreases
with the average amount of available energy.
The rest of the paper is organized as follows. Section II presents the system model and the
problem formulation. Section III studies the single-sensor case. Then, the multi-sensor problem
is solved in Section IV. Simulations are shown in Section V. Finally, Section VI concludes the
paper.
II. S YSTEM M ODEL
AND
P ROBLEM F ORMULATION
As shown in Fig. 1, we consider a scenario where a UAV flying over a set of N sensors for
data collection. The sensors are located on a line, labeled by S1 , S2 , · · · , SN . Each sensor n needs
to upload Bn information bits and is subject to a total energy budget En , where n = 1, 2, · · · , N.
5
A UAV flies at a fixed height H from an initial point S0 to a destination SN +1 , and applies time
division protocol to sequentially receive the uplink data from the sensors. Specifically, the whole
aviation range [S0 , SN +1 ] is divided into N non-overlapping intervals [xn , yn ], n = 1, 2, · · · , N
satisfying S0 ≤ x1 ≤ y1 ≤ x2 ≤ y2 ≤ · · · ≤ xN ≤ yN ≤ SN +1 . Each sensor node n uploads its
data when the UAV flies in the interval [xn , yn ]. If xn = yn , the UAV hovers above the location
xn and receives the data from sensor n. Otherwise, we assume the UAV flies with a constant
speed 0 < vn ≤ vmax from xn to yn and receives the data during its aviation. As no sensor
uploads data in the interval (yn , xn+1 ), the UAV flies with the maximum speed vmax in order to
minimize the total aviation time. In this paper, the UAV’s acceleration/deceleration process is
ignored for analytical tractability.
UAV
Uplink
H
...
S0
x1
S1
y1 x2
S2
y2
SN
yN
SN+1
Fig. 1. Data collection by a UAV from ground sensors along a line.
A. Data Collection Modes
Since the UAV can receive data when either navigating or hovering, we respectively consider
the data collection models for the two cases.
1) Data Collection during Aviation: If vn > 0 and xn < yn , the UAV collects data from the
ground node Sn during [xn , yn ]. The aviation time or the data collection time is tn = (yn −xn )/vn .
As the transmission distance changes during the flight, the transmit power and the data rate should
also adapt to the varying path-loss fading. The LOS ground-to-air channel model between the
UAV and the sensors with pathloss exponent α ≥ 2 is adopted [15]. With this model, the
6
instantaneous data rate in the transmission interval [xn , yn ] is given by
pn (t)β
1
,
Rn (t) = W log2 1 +
α
2
((xn + vn t − Sn )2 + H 2 ) 2
(1)
for t ∈ [0, tn ] where W is the bandwidth, β is the reference signal-to-noise ratio (SNR) at the
reference distance 1 meter, and pn (t) is the transmission power of the nth sensor, which satisfies
the total energy constraint
Z
tn
pn (t)dt ≤ En .
(2)
0
Besides, since each sensor n requires to upload Bn bits when the UAV flies over [xn , yn ], we
have the data constraint as
Z
tn
Rn (t)dt ≥ Bn .
(3)
0
Notice that there is a feasibility issue for data collection, i.e., with a given amount of sensor’s
energy En , is it feasible to upload Bn bits within the time duration tn ? Since the best channel
quality is experienced when hovering right above the sensor n, the maximum number of data
bits Bn is related to the hovering mode, which is detailed below.
2) Data Collection when Hovering: If vn = 0 and xn = yn , the UAV hovers above location xn
and sensor n uploads data with constant transmit power and data rate. Denote the transmission
time when hovering above the location xn by tn = Th,n (xn ). As the transmission link is static,
pn (t) should be a constant in this case. Thus, sensor n’s energy constraint (2) is simplified as
Th,n (xn )pn (t) ≤ En .
(4)
To fully utilize sensor n’s energy budget to minimize the aviation time, the transmission power
should be maximized, i.e., pn (t) = En /Th,n (xn ). Then the data constraint (3) is simplified as
βEn
Th,n (xn )
≥ Bn .
(5)
W log2 1 +
α
2
Th,n (xn )((xn − Sn )2 + H 2) 2
The function on the left hand side of (5) has the following property.
Lemma 1. The function f (x) = x log2 (1 + xa ), a > 0, x > 0 is an increasing function, and
f (x) <
a
.
ln 2
Proof. See Appendix A.
7
Based on Lemma 1, the left hand side of (5) is an increasing function of Th,n (xn ). Hence,
the minimum Th,n (xn ) satisfies (5) with equality, i.e.,
βEn
1
= Bn .
Th,n (xn )W log2 1 +
α
2
Th,n (xn )((xn − Sn )2 + H 2 ) 2
(6)
The above transcendental equation can be effectively solved by either line search or bisection
search. As the UAV experiences the best channel condition when hovering on top of the user
(xn = Sn ), the feasibility condition can be derived based on (6) as follows.
Proposition 1. (Feasibility) For each sensor n, the data constraint (3) is feasible if and only if
Bn <
W βEn
.
2H α ln 2
(7)
Proof. See Appendix B.
Hovering mode may be needed when the amount of information bits is large or the amount of
sensor’s energy is small. However, it may not be the most time efficient strategy when comparing
to that the UAV cruises and collects data at the same time. Therefore, navigation and hovering
modes have to be selected depending on the values of Bn and En , which will be incorporated
in our problem formulation.
B. Problem Formulation
In this paper, we aim to minimize the total aviation time while guaranteeing that all the sensors’
data are successfully collected. If (7) holds for all n = 1, 2, · · · , N, i.e., the data collection is
feasible for all the sensors, the problem can be formulated as
P
N
X
(SN +1 − S0 ) − N
n=1 (yn − xn )
+
tn
min
x,y,v,p(t)
vmax
n=1
s.t.
(8a)
(2) and (3), ∀n,
S0 ≤ x1 ≤ y1 ≤ x2 ≤ · · · ≤ yN ≤ SN +1 ,
tn =
yn − xn
Ixn 6=yn + Th,n (xn )Ixn =yn , ∀n,
vn
(8b)
(8c)
0 ≤ vn ≤ vmax , ∀n,
(8d)
pn (t) ≥ 0, ∀n, t,
(8e)
where the optimization variables are the locations x = {x1 , x2 , · · · , xN }, y = {y1 , y2 , · · · , yN },
the UAV speeds v = {v1 , v2 , · · · , vN }, and the transmission power p(t) = {p1 (t), p2 (t), · · · , pN (t)}.
8
The function Ievent is an indicator which equals 1 if the event is true and equals 0 otherwise. It
can be seen that the interval variables x, y for the sensors are coupled in the constraint (8b),
which makes (8) difficult to solve. To tackle the problem, we firstly consider a single-sensor
case, and then show how the solution of the single-sensor case can be extended to the general
multi-sensor case in (8).
III. AVIATION T IME M INIMIZATION
FOR
S INGLE - SENSOR C ASE
For the single-sensor case N = 1, without loss of generality, we set S1 = 0 (origin point in
the horizontal axis) and ignore the sensor index for all notations. Then the problem (8) reduce
to
min
x,y,v,p(t)
s.t.
(S2 − S0 ) − (y − x)
+t
vmax
Z t
p(τ )β
1
dτ ≥ B,
W log2 1 +
α
((x + vτ )2 + H 2 ) 2
0 2
Z t
p(τ )dτ ≤ E,
(9b)
S0 ≤ x ≤ y ≤ S2 ,
(9d)
(9a)
(9c)
0
t=
y−x
Ix6=y + Th (x)Ix=y ,
v
(9e)
0 ≤ v ≤ vmax ,
(9f)
p(t) ≥ 0.
(9g)
Since the hovering mode has been studied in the previous section, we mainly focus on the
navigation mode with v > 0 and x < y. The problem with only the navigation mode for the
single-sensor case is given by
min
x,y,v,p(t)
s.t.
1
1
S2 − S0
+ (y − x)
−
vmax
v vmax
Z y−x
v
p(τ )β
1
W log2 1 +
dτ ≥ B,
α
2
((x + vτ )2 + H 2) 2
0
Z y−x
v
p(τ )dτ ≤ E,
(10b)
S0 ≤ x < y ≤ S2 ,
(10d)
(10a)
(10c)
0
0 < v ≤ vmax , p(t) ≥ 0.
9
The problem (10) includes the power allocation optimization over p(t), the UAV speed
optimization over v, and the data upload interval optimization over x and y. These subproblems
are solved separately as follows.
A. Power Allocation
Suppose that the upload interval [x, y] and the UAV speed v are fixed and given. It is
obvious that to minimize the aviation time, the sensor should allocate its power to maximize the
uplink throughput on the left hand side of (10b). Thus, let us consider the following throughput
maximization problem
max
p(τ )≥0
s.t.
Z
y−x
v
p(τ )β
1
W log2 1 +
dτ
α
2
((x + vτ )2 + H 2 ) 2
(11a)
p(τ )dτ ≤ E.
(11b)
0
Z
y−x
v
0
Denote s = x + vτ , we have ds = vdτ . By changing the variable from τ to s, the throughput
maximization problem (11) can be reformulated as
Z
1 y1
p(s)β
max
W log2 1 + 2
ds
α
p(s)≥0 v x 2
(s + H 2 ) 2
Z
1 y
s.t.
p(s)ds ≤ E.
v x
(12a)
(12b)
Notice that the UAV receives the data from the sensor if and only if p(s) > 0. Otherwise, the
UAV flies with the maximum speed. Therefore, the condition for p(s) > 0 needs to be specified.
We have the following conclusion.
Theorem 1. Let p∗ (s) be an optimal solution of the problem (12). It holds that p∗ (s) > 0 for
x < s < y if and only if x, y and v satisfy
2
2
2
α
2
(y − x)(max{x , y } + H ) −
Moreover, the optimal power allocation p∗ (s) is
p∗ (s) =
Z
y
α
(s2 + H 2 ) 2 ds ≤ βEv.
(13)
x
1
1
−
,
γ0 γ(s)
(14)
where the water level is
1
vE
1
=
+
γ0
y − x (y − x)β
Z
x
y
α
(s2 + H 2 ) 2 ds,
(15)
10
and the inverse of channel gain is
α
(s2 + H 2 ) 2
1
=
.
γ(s)
β
(16)
The corresponding optimal objective value of (12a) is
W
β
αs
αH
s
Bmax (x, y, v) =
s log2
−
arctan
α +
2v
ln 2 ln 2
H
γ0 (s2 + H 2) 2
s=y
.
(17)
s=x
Proof. See Appendix C.
1
g ( s)
v
p( s )
1
g0
x
0
y
y'
s
Fig. 2. Illustration of water-filling power allocation.
The results in Theorem 1 are interpreted in Fig. 2. In this figure, the red solid curve represents
the inverse of the channel gain
1
,
γ(s)
the blue dash-dotted line represents the water level
1
,
γ0
and
the area between the two curves represents the total energy budget. If x, y and v satisfy the
condition (13), we have p(s) > 0 for x < s < y. However, if x, y ′ and v do not satisfy the
condition (13) as shown in the figure, there must be another set of x, y and v where y < y ′
so that p(s) > 0 for x < s < y. Therefore, the data upload must be within the range [x, y].
Theorem 1 explicitly gives the feasible region of x, y, v for optimal data collection.
11
For the free space LOS channel model with α = 2, the condition can be further simplified by
calculating the integration. In particular, we have
Z y
Z y
2
2 α
(s2 + H 2 )ds
(s + H ) 2 ds =
x
x
Replacing the term
Ry
x
α
y 3 − x3
=
+ (y − x)H 2
3
2
x + xy + y 2
2
+H .
= (y − x)
3
(18)
(s2 +H 2) 2 ds by the above expression, the condition (13) can be expressed
as the following two conditions:
(a) |x| ≤ |y| and 2y 3 + x3 − 3y 2x ≤ 3βEv,
(b) |x| ≥ |y| and 3x2 y − 2x3 − y 3 ≤ 3βEv,
and the water level can be expressed as
vE
x2 + xy + y 2 H 2
1
=
+
+
.
γ0
y−x
3β
β
(19)
B. UAV Speed Optimization
With the optimal power allocation, the maximum throughput Bmax (x, y, v) in (17) has the
following property.
Theorem 2. The maximum throughput Bmax (x, y, v) is a decreasing function of v.
Proof. See Appendix D.
Based on Theorem 2, the feasibility of any solution (x, y, v, p(t)) is guaranteed if the minimum
speed satisfies (13). The minimum speed can be written as a function of x and y, i.e.,
Z y
α
1
2
2
2 α
2
2
(y − x)(max{x , y } + H ) 2 −
(s + H ) 2 ds .
vm (x, y) = min vmax ,
βE
x
When α = 2, the minimum speed can be rewritten as
2y 3 +x3 −3y 2 x
, if |x| ≤ |y| and 2y 3 + x3 − 3y 2x ≤ 3βEvmax ,
3βE
3x2 y−2x3 −y 3
vm (x, y) =
, if |x| ≥ |y| and 3x2 y − 2x3 − y 3 ≤ 3βEvmax ,
3βE
vmax ,
otherwise.
(20)
(21)
12
According to Theorems 1 and 2, for given x, y and the optimal power allocation, the optimization over v can be formulated as
min
v
1
1
(S2 − S0 )
+ (y − x)
−
vmax
v vmax
s.t. Bmax (x, y, v) ≥ B,
vm (x, y) ≤ v ≤ vmax ,
(22a)
(22b)
(22c)
Problem (22) can be solved in two steps. Firstly, we check the feasibility of problem (22).
Based on Theorem 2, if Bmax (x, y, vm (x, y)) ≥ B, (22) is feasible, and we go to the second
step. As the objective function (22a) is a decreasing function of v, the optimal speed, denoted by
v ∗ (x, y), is the maximum feasible speed that satisfies (22b) in [vm (x, y), vmax]. Since Bmax (x, y, v)
is a decreasing function of v, v ∗ (x, y) can be found by bisection search algorithm. In summary,
the algorithm to obtain the optimal v and p(s) for given x and y where x < y in problem (10)
is summarized in Algorithm 1.
Algorithm 1 Calculate v and p(s) for given x, y in problem (10)
Input: β, H, W, α, B, E, δ, x and y where x < y.
Output: v ∗ , p∗ (s).
1:
Calculate vm (x, y) according to (20).
2:
if Bmax (x, y, vm (x, y)) ≥ B then
3:
Set vU = vmax , vL = vm (x, y), v = 21 (vU + vL ).
4:
while |vU − vL | > δ do
if Bmax (x, y, v) > B then
5:
Update vL = v, and reset v = 21 (vU + vL ).
6:
else
7:
Update vU = v, and reset v = 21 (vU + vL ).
8:
9:
end if
10:
end while
11:
Set v ∗ = v, and p∗ (s) is calculated according to (14)-(16).
12:
13:
14:
else
Problem (10) for given x and y is infeasible.
end if
13
In this algorithm, a precision parameter δ > 0 is introduced, which is used to control the
precision of v in the bisection search process. Line 2 examines the feasibility of the problem.
If the inequality does not hold, there is no feasible solution for the given parameters, and the
algorithm terminates. Otherwise, bisection search is launched as in lines 3-10. In line 4, the
search process will continue if |vU − vL | > δ. When the bisection search terminates, the optimal
solution is recorded in line 11.
Remark: It is interesting to remark that the maximum throughput in (12) can be re-written as
Ry 1
p(s)β
1
ds
1
+
W
log
α
2
y−x x 2
(s2 +H 2 ) 2
,
(23)
Bmax (x, y, v) = E max
v
p(s)
E
y−x
Ry
1
v
where p(s) is constrained by y−x
E which should be satisfied with equality
p(s)ds ≤ y−x
x
to achieve the maximum. The term on the right side of the operator max can be viewed as
the energy efficiency (achievable data rate per unit power) with “average power budget”
v
E.
y−x
Therefore, Theorem 2 says that the energy efficiency is a decreasing function of the power
budget in fading channels. It extends the result from the AWGN channel [24] to the UAV LOS
channel.
C. Data Collection Interval Optimization
Finally, we consider the problem of determining x and y in problem (10), which can be written
as
1
1
(S2 − S0 )
+ (y − x) ∗
−
min
x,y
vmax
v (x, y) vmax
s.t. S0 ≤ x < y ≤ S2 ,
(24a)
(24b)
where v ∗ (x, y) is the optimal solution of (22). As v ∗ (x, y) is a complex function of x and y,
there is no efficient algorithms other than two-dimensional line search to solve the problem
(24). By sampling m points in the aviation range [S0 , S2 ] with identical inter-point distance,
the total number of search pairs (x, y) where x < y is
m(m+1)
.
2
Thus, the complexity of the
two-dimensional search is O(m2 ).
IV. AVIATION T IME M INIMIZATION
FOR
M ULTI - SENSOR C ASE
In the multi-sensor case, the data upload intervals for the sensors correlates with one another.
In particular, if a sensor’s data upload interval is wide, the one next to it can only have a short
14
data upload interval. To deal with the inter-sensor correlation, we adopt the DP approach [25]
to solve the aviation time minimization problem for multiple sensors. Firstly, the basic concept
of the DP algorithm is briefly reviewed as follows.
A. Introduction to DP Algorithm
The DP algorithm deals with decision making problems in dynamic systems which can be
divided into stages. The dynamic system expresses the evolution of the system states, under the
influence of control actions taken at discrete instances of time (stage). The system has the form
sk+1 = fk (sk , uk ),
k = 0, 1, · · · , K − 1,
(25)
where
k is the index of the stage,
sk is the system state that summarizes all the information available for decision making,
uk is the control action selected in stage k,
fk describes how the system state is updated.
Once an action uk is taken under the state sk in stage k, an additive cost gk (sk , uk ) incurs. The
objective is to minimize the total cost by finding the optimal control actions for a given initial
state, which can be formulated as
min
u0 ,u1 ,··· ,uK−1
(K−1
X
gk (sk , uk ) + gK (sK ) s0
k=0
)
.
(26)
Notice that the above problem is optimized jointly over the actions for all stages u0 , u1, · · · , uK−1.
The Bellman’s equation in the following proposition tells us that the problem can be solved
stage-by-stage efficiently.
Proposition 2. (DP Algorithm) [25, Prop. 1.3.1, Vol. I] For every initial state s0 , the minimum
cost of the basic problem (26) is equal to J0 (s0 ), given by the last step of the following algorithm,
which proceeds backward in time from stage K − 1 to stage 0:
JK (sK ) = gK (sK ), ∀sK ∈ SK
Jk (sk ) = min
gk (sk , uk ) + Jk+1 (fk (sk , uk )) sk ,
uk ∈Uk (sk )
(27)
∀sk ∈ Sk , k = K − 1, K − 2, · · · , 0,
(28)
15
where Sk is the state space in stage k, Uk (sk ) is the state-dependent action space in stage k,
and (28) is called Bellman’s equation. Furthermore, if uk = µk (sk ) minimize the right side of
(28) for each sk and k, the policy π = {µ0 , µ1 , · · · , µK−1} is optimal.
The function Jk (sk ) is termed as the optimal cost-to-go, i.e., the minimum cost for the (K −k)stage problem that starts at stage k with state sk and ends at stage K. Based on Proposition
2, instead of jointly optimizing u0 , u1 , · · · , uK−1 for all stages, the DP algorithm recursively
optimizes the per-stage control action uk as in (28) based on the optimal cost-to-go Jk+1(sk+1 )
that has be calculated in the previous step.
B. DP-based Aviation Time Minimization
Now, we apply the DP algorithm to solve the aviation time minimization problem (8). Firstly,
the objective function (8a) can be rewritten as
N
yn − xn
SN +1 − S0 X
+
tn −
min
x,y,v,p(t)
vmax
vmax
n=1
N
SN +1 − S0 X
yn − xn
,
+
= min
min tn −
x,y
vn ,pn (t)
vmax
v
max
n=1
(29)
where the minimization over vn , pn (t) for a given pair xn < yn corresponds to the single-sensor
navigation case (10) and can be efficiently solved by Algorithm 1. If xn = yn , i.e., the UAV
hovers at location xn , the minimization takes the value with tn = Th,n (xn ) as the solution for
(6). Thus, according to the results in Sections II-A2), III-A, and III-B, we can define a cost
function as
yn − xn
tn −
vmax
gn (xn , yn ) = min
vn ,pn (t)
Th,n (xn ),
= (yn −xn )
+∞,
if xn = yn ,
1
1
∗ (x̃ ,ỹ ) − v
vn
n n
max
, if xn < yn and (22) is feasible,
(30)
elsewhere,
for all n = 1, 2, · · · , N, where x̃n = xn −Sn , ỹn = yn −Sn are the horizontal coordinates relative
to Sn , vn∗ (x̃n , ỹn ) is the optimal feasible solution of (22) that can be calculated via Algorithm 1,
and Th,n (xn ) is the minimum hovering time obtained by solving (6). If xn < yn while (22) is
infeasible, we set the cost as infinity.
Based on the above cost function, we formulate the aviation time minimization problem as a
DP problem. In particular, we have
16
- index of stage: n,
- system state in stage n: the end point of data upload for sensor n−1, denoted by sn = yn−1 .
The state space is Sn = [S0 , SN +1 ],
- control action in stage n: the data upload interval for sensor n, i.e., (xn , yn ). The action
space is Un (sn ) = {(xn , yn )|sn ≤ xn ≤ yn ≤ SN +1 },
- state update rule: sn+1 = fn (sn , xn , yn ) = yn ,
- per-stage cost: gn (xn , yn ), n = 1, 2, · · · , N as defined in (30), and gN +1 (sN +1 ) =
SN+1 −S0
.
vmax
As a result, the problem (29) can be rewritten as
" N
#
X
min
gn (xn , yn ) + gN +1 (yN ) ,
x,y
(31)
n=1
which can be solved by recursively calculating the cost-to-go function stage-by-stage as
JN +1 (sN +1 ) = gN +1 (sN +1 ) =
Jn (sn ) =
min
sn ≤xn ≤yn ≤SN+1
SN +1 − S0
,
vmax
∀sN +1
{gn (xn , yn ) + Jn+1 (yn )},
(32)
∀sn , n = N, N − 1, · · · , 1.
(33)
Then the minimum aviation time can be obtain in the last step, i.e.
Tmin = J1 (S0 ).
(34)
∗
In addition, if the optimal control actions for (33) are (x∗1 , y1∗), (x∗2 , y2∗ ), · · · , (x∗N , yN
), the
∗
optimal solution for the problem (29) is x∗ = {x∗1 , x∗2 , · · · , x∗N }, y ∗ = {y1∗, y2∗ , · · · , yN
}. Thus,
the optimal solution of the original problem (8) is x∗ , y ∗ joint with the optimal speeds vn∗ (x∗n −
Sn , yn∗ − Sn ), n = 1, 2, · · · , N from problem (22) and the optimal power allocation in (14).
It is remarkable that the computational complexity for the calculation of the cost-to-go functions Jn (sn ) can be reduced by exploring the property of (33).
Proposition 3. Concerning the DP algorithm (32) and (33), for any given n and sn , if the optimal
solution (x∗n , yn∗ ) for the minimization problem in (33) satisfies x∗n > sn , we have Jn (s′n ) = Jn (sn )
for all s′n ∈ [sn , x∗n ].
Proof. See Appendix E.
According to Proposition 3, to reduce the computational complexity, the calculation of Jn (sn )
for a given n can be launched from the initial point S0 to the destination SN +1 . When an optimal
solution (x∗n , yn∗ ) for a given sn is found and satisfies sn < x∗n , the calculation of Jn (s′n ) for
s′n ∈ [sn , x∗n ] can be omitted as the optimal solutions are equivalent to (x∗n , yn∗ ).
17
V. N UMERICAL R ESULTS
Some numerical results are shown in this section. In the numerical simulations, we set H = 100
m, β = 80 dB [20], and the channel bandwidth W = 20 kHz. According to the state-of-the-art
in the industry [2], we set the maximum speed vmax = 26 m/s.
A. Single-sensor Case Study
The optimal result for the single-sensor case versus different values of data upload requirement
B and sensor energy constraint E with S0 = −5000 m, S1 = 0 m, S2 = 5000 m are depicted
in Figs. 3 and 4, respectively. It can be seen that the optimal transmission interval (x, y) is
symmetric, which corresponds to the shortest average transmission distance from the sensor to
the UAV. In Fig. 3, when B > 5.7 Mb, the optimal solution is hovering above the sensor to
receive data. When 2.5 Mb < B < 5.7 Mb, both the length of data upload interval and the UAV
speed decreases as the data upload requirement increases. The decrease of data upload interval
increases the channel gain between the UAV and the sensor, and the decrease of the UAV speed
increases the transmission time. When B < 2.5 Mb, the UAV can fly with the maximum speed
while successfully receive all the uploaded data. In this range, the minimum data upload interval
is depicted, and its length decreases as the data upload requirement decreases as the time required
for data upload decreases.
In Fig. 4, the optimal result versus E is opposite to that versus B. In particular, when E < 0.3
J, the UAV also needs to hover above the sensor to receive data. When 0.3 J < E < 1.7 J, both
the length of the data upload interval and the UAV speed increases as the amount of energy
increases. While for E > 1.7 J, the UAV can fly with the maximum speed, and the minimum
length of the data upload interval decreases as the amount of energy increases.
B. Multi-sensor Case Study
Then the data collection for multiple sensors is studied by simulation. In particular, the UAV
flies from S0 = 0 m to SN +1 = 10000 m, during which N = 10 sensors are deployed. The
locations of the sensors Sn , n = 1, · · · , 10 are fixed as 500m, 2500m, 4500m, 6500m, 7000m,
7500m, 8000m, 8500m, 9000m, and 9500m, i.e., the first four sensors are 2000m apart from one
another (sparsely deployed), and the last six sensors are 500m apart from one another (densely
deployed). We study the impact of required data and energy limitation respectively.
18
1500
25
x
y 20
v
1000
15
Location (m)
5
0
0
−5
−500
Speed (m/s)
10
500
−10
−15
−1000
−20
−1500
0
1
2
3
4
5
−25
6
B (Mb)
Fig. 3. Optimal solution (x, y, v) versus B for the single-sensor case, with S0 = −5000 m, S1 = 0 m, S2 = 5000 m, and
E = 1 J.
1500
25
x
y 20
v
1000
15
Location (m)
5
0
0
−5
−500
Speed (m/s)
10
500
−10
−15
−1000
−20
−1500
0
0.5
1
1.5
2
2.5
−25
3
E (J)
Fig. 4. Optimal solution (x, y, v) versus E for the single-sensor case, with S0 = −5000 m, S1 = 0 m, S2 = 5000 m, and
B = 3 Mb.
19
30
v
n
Sn
25
x
n
y
n
Speed (m/s)
20
15
10
5
0
0
2000
4000
6000
8000
10000
Location (m)
Fig. 5. Optimal solution (xn , yn , vn ) for N = 10 sensors with En = 1.2 J for all n = 1, 2, · · · , N , B1 = · · · = B4 = B6 =
B10 = 3 Mbits, B5 = 2.5 Mbits, B7 = B9 = 3.5 Mbits, and B8 = 7 Mbits.
In Fig. 5, the amount of energy in each sensor is set identical, En = 1.2 J for all n =
1, 2, · · · , N, and the amount of data to be transmitted varies. We set B1 = · · · = B4 = B6 =
B10 = 3 Mbits, B5 = 2.5 Mbits, B7 = B9 = 3.5 Mbits, and B8 = 7 Mbits. It can be seen that
as the first four sensors are sparsely located, the upload intervals are disconnected. The reason
is that it is not energy-efficient when the transmission distance is large. In this case, the UAV
collects data from a sensor in a short range and then flies towards another with the maximum
speed. For the last six sensors, as the amount of data to be transmitted increases from sensor
S5 to sensor S8 and then decreases from S8 to S10 , the UAV’s speed firstly decreases and then
increases accordingly so that the required data can be uploaded successfully. Particularly, As the
amount of data in sensor S8 is extremely large, the UAV hovers above it to collect data with
maximum data rate, so that the overall aviation time is minimized.
In Fig. 6, we change the values of the amount of data bits as B1 = · · · = B4 = B6 = B10 = 2
Mbits, B5 = 2.5 Mbits, B7 = B9 = 3.5 Mbits, and B8 = 3.8 Mbits. Firstly, as the data bits in
sensors S1 , · · · , S4 are limited, the UAV can successfully receive all the data bits when flying
with maximum speed. In addition, as the data bits in sensor S8 are reduced compared with Fig. 5,
the hovering mode is not necessary any more. As the amount of data bits is still the largest, the
20
30
vn
S
n
25
x
n
yn
Speed (m/s)
20
15
10
5
0
0
2000
4000
6000
8000
10000
Location (m)
Fig. 6. Optimal solution (xn , yn , vn ) for N = 10 sensors with En = 1.2 J for all n = 1, 2, · · · , N , B1 = · · · = B4 = B6 =
B10 = 2 Mbits, B5 = 2.5 Mbits, B7 = B9 = 3.5 Mbits, and B8 = 3.8 Mbits.
aviation speed is quite low.
Then we set the amount of data in each sensor to be the same, i.e., Bn = 3 Mbits for all
n = 1, 2, · · · , N, while E1 = · · · = E4 = 3.6 J, E5 = 3.2 J, E6 = E10 = 1.8 J, E7 = E9 = 0.8
J, and E8 = 0.2 J to evaluate the impact of the energy constraint. It can be seen that with
sufficient amount of energy for the first four sensors, the UAV can fly with maximum speed
while successfully receiving all the data. For the last six sensors, as the amount of energy firstly
decreases and then increases from sensor S5 to sensor S10 , the optimal speed also decreases
at first and then increase. In particular, as the eighth sensor is quite energy stringent, the UAV
hovers above it with zero speed to collect its data. In addition, the transmission intervals for the
first four sensors shift towards the initial point so that more space can be reserved for the last
six sensors which has limited energy budget.
Next, we reset the amount of energy as E1 = E2 = E3 = E7 = E9 = 1.0 J, E4 = 1.2 J,
E5 = 3.2 J, E6 = E10 = 2.0 J, and E8 = 0.6 J and re-run the simulation, the result is shown
in Fig. 8. It can be found that the UAV serves the first three sensors with medium aviation
speed, as the limited amount of energy cannot support the maximum speed. Similarly, as the
energy in sensor S8 is sufficient to support data collection during aviation, the hovering mode
21
30
25
Speed (m/s)
20
15
10
v
n
Sn
5
x
n
y
n
0
0
2000
4000
6000
8000
10000
Location (m)
Fig. 7. Optimal solution (xn , yn , vn ) for N = 10 sensors with Bn = 3 Mbits for all n = 1, 2, · · · , N , and E1 = · · · = E4 = 3.6
J, E5 = 3.2 J, E6 = E10 = 1.8 J, E7 = E9 = 0.8 J, and E8 = 0.2 J.
30
v
n
Sn
25
xn
yn
Speed (m/s)
20
15
10
5
0
0
2000
4000
6000
8000
10000
Location (m)
Fig. 8. Optimal solution (xn , yn , vn ) for N = 10 sensors with Bn = 3 Mbits for all n = 1, 2, · · · , N , and E1 = E2 = E3 =
E7 = E9 = 1.0 J, E4 = 1.2 J, E5 = 3.2 J, E6 = E10 = 2.0 J, and E8 = 0.6 J.
22
40
Ē
Ē
Ē
Ē
Average Aviation Time (min)
35
=
=
=
=
0.15
0.30
0.60
1.50
J
J
J
J
30
25
20
15
10
5
0.5
1
1.5
2
2.5
3
3.5
4
4.5
5
B̄ (Mb)
Fig. 9. Average aviation time versus average amount of data with random data requirement, random energy and random locations.
is not necessary. Compared with Figs. 5-8, the energy constraint has similar impact as the data
requirement.
C. Average Performance Evaluation
We further evaluate the average performance with random data requirement, random energy
and random locations. The amount of data in each sensor follows uniform distribution with a
mean value B̄, the amount of energy in each sensor follows uniform distribution with a mean
value Ē, and each pair (Bn , En ) is set to satisfy the feasibility constraint (7). The sensors are
uniformly distributed in the range [S0 , SN +1 ] = [0, 10000] m. The results are illustrated in Figs. 9
and 10. It can be seen in Fig. 9 that when the average amount of energy is sufficient, the average
aviation time grows almost linearly with the increase of B̄. But when the amount of energy is
deficient (e.g., Ē = 0.15 J), the average aviation time grows exponentially with the increase of
B̄. This is due to the different relations between the aviation time and the amount of data in
hovering mode and aviation mode. In energy sufficient case, the UAV collects data mainly in
aviation mode. While in energy constrained case, it collects data mainly in hovering mode.
In Fig. 10, it is observed that when the amount of data is small, the average aviation time is
constant over all examined values of Ē, which means that the UAV can fly with the maximum
23
40
B̄
B̄
B̄
B̄
Average Aviation Time (min)
35
=
=
=
=
0.5
2.0
3.5
5.0
Mb
Mb
Mb
Mb
30
25
20
15
10
5
0
0.5
1
1.5
Ē (J)
Fig. 10. Average aviation time versus average amount of energy with random data requirement, random energy and random
locations.
speed and collect data during aviation. In addition, it is expected with the increase of Ē, the
curves converges to a fixed point with minimum aviation time, i.e., the UAV flies with the
maximum speed. However, the figure shows that the convergence is slow, especially for large
values of B̄. For the case with B̄ = 5.0 Mb, the curve firstly goes down exponentially, and then
linearly with close-to-zero slope.
VI. C ONCLUSION
In this paper, we have solved the aviation time minimization problem for completing the data
collection mission in a one-dimensional sensor network. The analysis on hovering mode provides
the feasibility condition for a successful data collection. The analysis on the single-sensor case
reveals the optimal solution structures. Firstly, the optimal power allocation follows the classical
water-filling policy. Secondly, the maximum amount of data bits that can be successfully uploaded
during UAV’s aviation is a decreasing function of the UAV speed, which results in a simple
bisection method to find the optimal aviation speed. For the multi-sensor case, we have shown
that the division of data collection intervals can be optimized via the DP algorithm. According to
the numerical results, is has been observed that the behavior of the UAV relies on the locations,
the data amount and the energy amount of sensors. With a sufficient amount of energy, the UAV
24
can fly with maximum speed. Otherwise, its speed is proportional to the sensors’ energy budgets
and the inter-sensor distance, but inversely proportional to the amount of data to be uploaded.
A PPENDIX A
P ROOF
OF
L EMMA 1
Since
a2
< 0,
f (x) = −
x(x + a)2
(35)
a
a
f ′ (x) = log2 1 +
−
> f ′ (+∞) = 0,
x
x+a
(36)
′′
f ′ (x) is decreasing. Therefore,
which indicates that f (x) is increasing. In addition,
a xa
lim f (x) = lim a log2 1 +
= a log2 e.
x→+∞
x→+∞
x
Hence, f (x) < a log2 e =
(37)
a
.
ln 2
A PPENDIX B
P ROOF
OF
P ROPOSITION 1
As
βEn
Th,n (xn )
W log2 1 +
α
2
Th,n (xn )((xn − Sn )2 + H 2 ) 2
W βEn
<
α
2((xn − Sn )2 + H 2 ) 2 ln 2
W βEn
,
≤
2H α ln 2
(38)
where the first inequality holds according to Lemma 1, and the gap can be arbitrarily small as
Th,n (xn ) tends to infinity. In the second inequality, the equality holds when xn = Sn . Hence, if
Bn <
W βEn
,
2H α ln 2
there is always a feasible transmission mode so that Bn bits can be successfully
transmitted.
If Bn ≥
W βEn
2H α ln 2
on the contrary, the left hand side of (5) is always less than Bn . Therefore,
the equation (6) is not feasible.
25
A PPENDIX C
P ROOF
OF
T HEOREM 1
The Lagrangian function of the problem (12) is expressed as
Z y
Z
p(s)β
1
W y
ds − λ
log2 1 + 2
p(s)ds − E .
L=
α
2v x
v x
(s + H 2 ) 2
By setting
∂L
∂p(s)
(39)
= 0, we get the optimal power allocation expressed as (14), where γ0 =
1
λ
is
the water level so that (12b) is satisfied with equality, and the channel gain is
γ(s) =
(s2
β
α ,
+ H 2) 2
(40)
which is equivalent to (16).
Based on (14) and (40), it can be found that p∗ (s) > 0 must hold in a continuous interval.
Next, we derive the necessary and sufficient condition for p∗ (s) > 0 for all x < s < y.
1) Sufficiency:
Suppose that p∗ (s) > 0 for x < s < y. As s2 ≤ max{x2 , y 2 } for any x < s < y, we have
γ(s) =
As p∗ (s) > 0, we have
1
γ0
>
β
β
α >
α .
(s2 + H 2 ) 2
(max{x2 , y 2 } + H 2 ) 2
1
γ(s)
(41)
holds for all x < s < y. To guarantee this,
than or equal to the maximum value of
1
,
γ(s)
1
γ0
must be larger
i.e.,
α
(max{x2 , y 2} + H 2 ) 2
1
≥
.
γ0
β
(42)
On the other hand, according to (12b), i.e.,
Z
Z
α
1 y ∗
(s2 + H 2 ) 2
1 y 1
−
ds
p (s)ds =
v x
v x γ0
β
Z
α
y−x 1
1 y (s2 + H 2 ) 2
=
−
ds
v γ0 v x
β
≤ E,
(43)
we have
1
vE
1
≤
+
γ0
y−x y−x
Z
x
y
α
(s2 + H 2 ) 2
ds.
β
(44)
According to (42) and (44), we have
α
vE
1
(max{x2 , y 2} + H 2 ) 2
≤
+
β
y−x y−x
Z
x
y
α
(s2 + H 2 ) 2
ds,
β
which is equivalent to (13). Therefore, the sufficiency is proved.
(45)
26
In addition, to maximize the throughput, (43) must be satisfied with equality, which results in
equality condition in (44). Hence, (15) is obtained.
2) Necessity:
1
γ0
Suppose (13) (or equivalently (45)) holds true. We let
equals to the right hand side of (44).
Then (43) is satisfied with equality, which guarantees that the power allocation is optimal as all
the energy budget is fully utilized. Based on (45) and the equality of (44), we have
α
α
1
(max{x2 , y 2} + H 2 ) 2
(s2 + H 2 ) 2
≥
>
γ0
β
β
for all x < s < y. Therefore, we have p∗ (s) =
1
γ0
− (s
α
2 +H 2 ) 2
β
(46)
> 0 for x < s < y, and hence, the
necessity is proved.
The optimal throughput can be obtained by substituting p(s) in (12a) with (14) and deducing
as follows
=
=
=
=
=
y
β
α ds
γ0
+ H 2) 2
x
Z
y
α y
β
W
2
2
−
s log2
log2 (s + H )ds
2v
γ0 x 2 x
y αZ y
β
α
W
2
2
2
2
+
s log2
sd log2 (s + H )
− s log2 (s + H )
2v
γ0
2
2 x
x
Z y
y
α
s2
β
W
+
s log2
ds
α
2v
γ0 (s2 + H 2 ) 2 x ln 2 x s2 + H 2
Z y
y
α
β
H2
W
+
s log2
1− 2
ds
α
2v
s + H2
γ0 (s2 + H 2 ) 2 x ln 2 x
y
W
αH
s
αs
β
.
(47)
s log2
−
arctan
α +
2v
ln 2 ln 2
H x
γ0 (s2 + H 2 ) 2
W
Bmax (x, y, v) =
2v
Z
log2
(s2
A PPENDIX D
P ROOF
OF
T HEOREM 2
Based on (12a) and (12b), to achieve the maximum throughput, all the energy should be fully
used, i.e.
Z
y
p(s)ds = vE.
(48)
x
Replacing p(s) in the above equation by (14), we have
Z y
vE
1
1
1
=
+
ds.
γ0
y − x y − x x γ(s)
(49)
27
According to the first line of (47), we have
Z
γ(s)
W y
ds
log2
Bmax (x, y, v) =
2v x
γ0
Z y
Z y
W
1
1
=
ds
log2 ds −
log2
2v
γ0
γ(s)
x
x
1
W
a1 log2
vE + a2
− a3 ,
=
2v
a1
(50)
where
a1 = y − x,
Z y
1
ds,
a2 =
x γ(s)
Z y
1
ds.
a3 =
log2
γ(s)
x
(51)
(52)
(53)
Define a function
Since
1 E
g(u) = u a1 log2
+ a2
− a3 .
a1 u
g ′′ (u) = −
a1 E 2
<0
u(E + a2 u)2
(54)
(55)
for all u > 0, g ′(u) is a decreasing function of u. Therefore,
1 E
a1 E
′
g (u) = a1 log2
+ a2
− a3 −
a1 u
E + a2 u
> g ′ (+∞)
a2
− a3
= a1 log2
a1
Z y
Z y
1
1
1
ds −
ds
log2
= (y − x) log2
y − x x γ(s)
γ(s)
x
≥ 0,
(56)
where the first inequality holds due to the monotonicity of g ′ (u), and the second inequality holds
due to the concavity of log function. Based on (56), we conclude that g(u) is an increasing
function of u. Since Bmax (x, y, v) = 21 W g( v1 ), it is a decreasing function of v.
28
A PPENDIX E
P ROOF
OF
P ROPOSITION 3
As (x∗n , yn∗ ) is the optimal solution for the minimization problem in (33), we have
Jn (sn ) =
min
sn ≤xn ≤yn ≤SN+1
{gn (xn , yn ) + Jn+1 (yn )}
= gn (x∗n , yn∗ ) + Jn+1 (yn∗ ).
(57)
For a given s′n ∈ [sn , x∗n ], as s′n ≥ sn , we have [s′n , SN +1 ] ⊆ [sn , SN +1 ]. Therefore,
Jn (s′n ) =
≥
min
{gn (xn , yn ) + Jn+1 (yn )}
min
{gn (xn , yn ) + Jn+1 (yn )} = Jn (sn ).
s′n ≤xn ≤yn ≤SN+1
sn ≤xn ≤yn ≤SN+1
(58)
Secondly, as s′n ≤ x∗n , we have s′n ≤ x∗n ≤ yn∗ ≤ SN +1 . Hence,
Jn (s′n ) =
min
s′n ≤xn ≤yn ≤SN+1
{gn (xn , yn ) + Jn+1 (yn )}
≤ gn (x∗n , yn∗ ) + Jn+1 (yn∗ ) = Jn (sn ).
(59)
Combining (58) and (59), we prove that Jn (s′n ) = Jn (sn ) for all s′n ∈ [sn , x∗n ].
R EFERENCES
[1] Y. Zeng, R. Zhang, and T. J. Lim, “Wireless communications with unmanned aerial vehicles: opportunities and challenges,”
IEEE Communications Magazine, vol. 54, no. 5, pp. 36–42, May 2016.
[2] Dji. INSPIRE 2: Power beyond imagination. [Online]. Available: http://www.dji.com/inspire-2
[3] R. Sun, “Dual-band non-stationary channel modeling for the air-ground channel,” Ph.D. dissertation, University of South
Carolina, Jul. 2015.
[4] M. Dong, K. Ota, M. Lin, Z. Tang, S. Du, and H. Zhu, “UAV-assisted data gathering in wireless sensor networks,” The
Journal of Supercomputing, vol. 70, no. 3, pp. 1142–1155, Dec. 2014.
[5] Z. Han, A. L. Swindlehurst, and K. J. R. Liu, “Optimization of MANET connectivity via smart deployment/movement of
unmanned air vehicles,” IEEE Transactions on Vehicular Technology, vol. 58, no. 7, pp. 3533–3546, Sept. 2009.
[6] E. I. Grøtli and T. A. Johansen, “Path planning for UAVs under communication constraints using SPLAT! and MILP,”
Journal of Intelligent & Robotic Systems, vol. 65, no. 1, pp. 265–282, Jan. 2012.
[7] E. Koyuncu, R. Khodabakhsh, N. Surya, and H. Seferoglu, “Deployment and trajectory optimization for UAVs: A
quantization theory approach,” arXiv:1708.08832, 2017.
[8] Q. Wu, Y. Zeng, and R. Zhang, “Joint trajectory and communication design for multi-UAV enabled wireless networks,”
arXiv:1705.02723, 2017.
[9] Y. Zeng, R. Zhang, and T. J. Lim, “Throughput maximization for UAV-enabled mobile relaying systems,” IEEE Transactions
on Communications, vol. 64, no. 12, pp. 4983–4996, Dec. 2016.
[10] D. S. Kalogerias and A. P. Petropulu, “Mobile beamforming & spatially controlled relay communications,” in Proceedings
of IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), 2016, pp. 6405–6409.
29
[11] C.-M. Cheng, P.-H. Hsiao, H. Kung, and D. Vlah, “Maximizing throughput of UAV-relaying networks with the load-carryand-deliver paradigm,” in IEEE Wireless Communications and Networking Conference, 2007, pp. 4417–4424.
[12] J. Lyu, Y. Zeng, R. Zhang, and T. J. Lim, “Placement optimization of UAV-mounted mobile base stations,” IEEE
Communications Letters, vol. 21, no. 3, pp. 604–607, Mar. 2017.
[13] M. Alzenad, A. El-Keyi, F. Lagum, and H. Yanikomeroglu, “3D placement of an unmanned aerial vehicle base station
(UAV-BS) for energy-efficient maximal coverage,” IEEE Wireless Communications Letters, vol. 6, no. 4, pp. 434–437,
Aug. 2017.
[14] M. Mozaffari, W. Saad, M. Bennis, and M. Debbah, “Efficient deployment of multiple unmanned aerial vehicles for optimal
wireless coverage,” IEEE Communications Letters, vol. 20, no. 8, pp. 1647–1650, Aug. 2016.
[15] N. Ahmed, S. S. Kanhere, and S. Jha, “On the importance of link characterization for aerial wireless sensor networks,”
IEEE Communications Magazine, vol. 54, no. 5, pp. 52–57, May 2016.
[16] F. Jiang and A. L. Swindlehurst, “Optimization of UAV heading for the ground-to-air uplink,” IEEE Journal on Selected
Areas in Communications, vol. 30, no. 5, pp. 993–1005, Jun. 2012.
[17] A. E. A. A. Abdulla, Z. M. Fadlullah, H. Nishiyama, N. Kato, F. Ono, and R. Miura, “An optimal data collection technique
for improved utility in UAS-aided networks,” in IEEE Conference on Computer Communications (Infocom), Apr. 2014,
pp. 736–744.
[18] S. Say, H. Inata, J. Liu, and S. Shimamoto, “Priority-based data gathering framework in UAV-assisted wireless sensor
networks,” IEEE Sensors Journal, vol. 16, no. 14, pp. 5785–5794, Jul. 2016.
[19] C. Zhan, Y. Zeng, and R. Zhang, “Energy-efficient data collection in UAV enabled wireless sensor network,”
arXiv:1708.00221, pp. 1–4, Aug. 2017.
[20] J. Lyu, Y. Zeng, and R. Zhang, “Cyclical multiple access in UAV-aided communications: A throughput-delay tradeoff,”
IEEE Wireless Communications Letters, vol. 5, no. 6, pp. 600–603, Dec. 2016.
[21] C. D. Franco and G. Buttazzo, “Energy-aware coverage path planning of UAVs,” in IEEE International Conference on
Autonomous Robot Systems and Competitions, Apr. 2015, pp. 111–117.
[22] Y. Zeng and R. Zhang, “Energy-efficient UAV communication with trajectory optimization,” IEEE Transactions on Wireless
Communications, vol. 16, no. 6, pp. 3747–3760, Jun. 2017.
[23] S. Zhang, Y. Zeng, and R. Zhang, “Cellular-enabled UAV communication: Trajectory optimization under connectivity
constraint,” arXiv:1710.11619, 2017.
[24] S. Verdu, “On channel capacity per unit cost,” IEEE Transactions on Information Theory, vol. 36, no. 5, pp. 1019–1030,
Sep. 1990.
[25] D. P. Bertsekas, Dynamic programming and optimal control.
Athena Scientific Belmont, MA, 2005.
| 7cs.IT
|
WEIGHTED GROWTH FUNCTIONS OF AUTOMATIC GROUPS
arXiv:1711.01256v1 [math.GR] 3 Nov 2017
MIKAEL VEJDEMO-JOHANSSON
Abstract. The growth function is the generating function for sizes of spheres
around the identity in Cayley graphs of groups. We present a novel method
to calculate growth functions for automatic groups with normal form recognizing automata that recognize a single normal form for each group element,
and are at most context free in complexity: context free grammars can be
translated into algebraic systems of equations, whose solutions represent generating functions of their corresponding non-terminal symbols. This approach
allows us to seamlessly introduce weightings on the growth function: assign
different or even distinct weights to each of the generators in an underlying
presentation, such that this weighting is reflected in the growth function. We
recover known growth functions for small braid groups, and calculate growth
functions that weight each generator in an automatic presentation of the braid
groups according to their lengths in braid generators.
1. Introduction
Analytic combinatorics provides tools for enumerating structures as described by
formal grammars, producing generating functions. In this paper we will approach
the enumeration of minimal length words representing group elements in finitely
presented automatic groups using generating functions generated from the formal
grammars associated to the group’s automatic structure.
For a group with presentation G = hg|ri, we define
Cayley graph: the graph with group elements as vertices, and an edge from
each vertex h for each generator in g, to the vertex gh.
geodesic word: shortest word in the generators and their inverses representing a group element; corresponds to a shortest path in the Cayley graph.
radius r sphere around the identity: the set of elements whose geodesic
words have length r. We denote this S(r).
growth function: the generating function of the sequence S(r) for r nonnegative integers.
First, in Section 2 we will introduce the route from a formal grammar to a generating function, and in Section 3 we will demonstrate how these methods apply to
automatic group, by working with the explicit example of the braid group B3 on
three strands.
Date: November 6, 2017.
1
2
MIKAEL VEJDEMO-JOHANSSON
2. Counting with grammars
Chomsky and Schützenberger proved [2] that a contextfree language can studied
using generating functions. Their article provides a construction for finding the
generating function related to a specific grammar.
Starting with a Backus-Naur form of the grammar, each rewriting rule can be translated into an algebraic equation. Each terminal symbol is assigned some expression
in the variables of the resulting generating function, and each non-terminal symbol
is assigned a generating function of its own. The rewriting assignment is replaced
by an equality, each concatenation with a multiplication and each disjunction with
an addition.
For a first and simple example, balanced two-symbol sequences have the grammar
S → ∅ | aSb
Translating this to an algebraic equation, we would get
S(x, y) = 1 + xyS(x, y)
by weighting each symbol a by x1 and each symbol b by y 1 . The resulting generating
function will count the number of strings by the number of a and b symbols in the
result, or by evaluating S(t, t) will count by length of the string.
P
This equation is solved straightforwardly to S(x, y) = 1/(1 − xy) = (xy)j , from
which we can immediately read that there is exactly one
for each combination
P string
of j each of as and bs. From S(t, t) = 1/(1 − t2 ) =
t2j follows that there is one
unique string for each even length, and no odd-length strings.
Chomsky and Schützenberger proved that as long as the grammar is at most
context-free, the corresponding generating function(s) will be rational functions.
For anything that can be described by a context-free grammar, this suggests a
concrete approach for enumeration:
(1) Find a Backus-Naur form of a grammar describing your structures
(2) Translate the grammar to a system of polynomial equations
(3) Use a Gröbner basis with an elimination order to solve the system of equations
(4) Isolating the Gröbner basis elements concentrated to the interesting nonterminal symbol and the terminal variables, solve for a rational form of the
generating function
3. Braids and Automatic Groups
Braid groups are usually introduced with a finite presentations in terms of elementary braids: for k strands, the braid group Bk has generators σj for 1 ≤ j < k,
where σj crosses strand j over strand j + 1. We give an illustration for B4 in
WEIGHTED GROWTH FUNCTIONS OF AUTOMATIC GROUPS
3
Figure 1. Generators of the Braid group B4
=
=
Figure 2. Relations of the Braid group B4
Figure 1. By inspecting the effects of Reidemeister moves, and of manipulations of
separated areas of the 3-sphere, we can derive the finite presentation
Bk = hσ1 , . . . , σk |σi σj = σj σi ; σi σi+1 σi = σi+ σi σi+1 i
where |i − j| > 1. Figure 2 shows these relations in B4 , the smallest braid group
where all relations are applicable.
One proof that the word problem is solvable for braid groups was described in [3],
demonstrating that automatic groups solve the word problem, and that braid groups
are (bi)automatic. An automatic group, here, is a finitely presented group coupled
with several automata: one to detect whether a given string in the generators is
the normal form of a group element, and one to detect right products of a normal
form by a generator.
Braid groups form an example of biautomatic groups: there are grammars both
for recognizing right products and left products. The part that really interests us
here, though, is the normal form recognizer. With a grammar for normal forms,
4
MIKAEL VEJDEMO-JOHANSSON
algebraic equations to compute generating functions for group sizes can be computed. These generating functions are also studied extensively for finitely presented
groups: they are called growth series. For braid groups, we even know grammars
that pick out exactly one normal form for each group element, such that this normal form is geodesic: has the shortest possible expression in some specific set of
generators.
Charney in [1] gives a grammar for the Braid group B3 with the following transition
rules, with each state terminal.
B3 → e|s1 v2 |s2 v3 |s1 s2 v3 |s2 s1 v2 |s1 s2 s1 v1 |s1 v5 |s2 v6 |s1 s2 v6 |s2 s1 v5 |s1 s2 s1 v4
v1 → s1 s2 s1 v1
v2 → s1 v2 |s1 s2 v3 |s2 v6 |s2 s1 v5 |s1 s2 s1 v1
v3 → s2 v3 |s2 s1 v2 |s1 v5 |s1 s2 v6 |s1 s2 s1 v1
v4 → s1 s2 s1 v4 |s1 s2 v6 |s2 v6 |s2 s1 v5 |s1 v5
v5 → s1 v5 |s1 s2 v6
v6 → s2 v6 |s2 s1 v5
The construction of this automaton generalizes to all braid groups, with exponential growth in the number of rules in the grammar. This improves on previous
constructions that needed factorial growth in the number of rules.
From this grammar we can produce a system of algebraic equations that counts
each generator in Charney’s presentation equally
B3 = (1 + t · v2 + t · v3 + t · v3 + t · v2 + t · v1 + t · v5 + t · v6 + t · v6 + t · v5 + t · v4 )
v1 = (1 + t · v1 )
v2 = (1 + t · v2 + t · v3 + t · v6 + t · v5 + t · v1 )
v3 = (1 + t · v3 + t · v2 + t · v5 + t · v6 + t · v1 )
v4 = (1 + t · v4 + t · v6 + t · v6 + t · v5 + t · v5 )
v5 = (1 + t · v5 + t · v6 )
v6 = (1 + t · v6 + t · v5 )
Solving this system for B3 (t) with term order to eliminate all the vs, using your
favorite computer algebra system recovers a Gröbner basis:
4B3 (t)t3 − 8B3 (t)t2 − 4t3 + 5B3 (t)t − 8t2 − B3 (t) + 5t + 1
− 2B3 (t)t2 + 3B3 (t)t + 2t2 − B3 (t) + 5t + v6
− 2B3 (t)t2 + 3B3 (t)t + 2t2 − B3 (t) + 5t + v5
− 20B3 (t)t2 + 28B3 (t)t + 20t2 − 9B3 (t) + 52t + 2v4 + 7
20B3 (t)t2 − 26B3 (t)t − 20t2 + 5B3 (t) − 54t + 6v3 − 11
20B3 (t)t2 − 26B3 (t)t − 20t2 + 5B3 (t) − 54t + 6v2 − 11
4B3 (t)t2 − 4B3 (t)t − 4t2 + B3 (t) − 12t + 6v1 − 7
WEIGHTED GROWTH FUNCTIONS OF AUTOMATIC GROUPS
5
The first of these terms completely avoids all the vs, and is the one generator of
the elimination ideal. This produces a functional equation for B3 (t):
4B3 (t)t3 − 8B3 (t)t2 − 4t3 + 5B3 (t)t − 8t2 − B3 (t) + 5t + 1 = 0
which we can rewrite to
B3 (t)(4t3 − 8t2 + 5t − 1) = 4t3 + 8t2 − 5t − 1
from which follows
B3 (t) =
2t(8t − 5)
4t3 + 8t2 − 5t − 1
=1+
4t3 − 8t2 + 5t − 1
(t − 1)(2t − 1)2
1+10t+34t2 +90t3 +218t4 +506t5 +1146t6 +2554t7 +5626t8 +12282t9 +26618t10 +57338t11 +
122874t12 +262138t13 +557050t14 +1179642t15 +2490362t16 +5242874t17 +11010042t18 +O(t19 )
This recovers the growth function for B3 as computed by Charney [1].
The method of going through Gröbner basis computations, however, is more flexible
than Charney’s linear algebra approach. Since we can choose weights at will, we
can – for instance – compute the growth series of the automatic presentation, as
weighted by the number of elementary braid generators used for each word. Doing
this still retains a strong focus on the automatic presentation, and as we will see no
longer calculates geodesic (ie shortest) words for the presentation with elementary
braid generators.
To achieve this, we weight each term when translating to a system of equations not
by the number of automatic generators involved, but by the number of elementary
braid generators in each term, producing the system of equations
B3 = (1 + t · v2 + t · v3 + t2 · v3 + t2 · v2 + t3 · v1 + t · v5 + t · v6 + t2 · v6 + t2 · v5 + t3 · v4 )
v1 = (1 + t3 · v1 )
v2 = (1 + t · v2 + t2 · v3 + t · v6 + t2 · v5 + t3 · v1 )
v3 = (1 + t · v3 + t2 · v2 + t · v5 + t2 · v6 + t3 · v1 )
v4 = (1 + t3 · v4 + t2 · v6 + t · v6 + t2 · v5 + t · v5 )
v5 = (1 + t · v5 + t2 · v6 )
v6 = (1 + t · v6 + t2 · v5 )
Calculating, again, an eliminating Gröbner basis produces
B3 (t)t5 + B3 (t)t4 − t5 − 3B3 (t)t3 − t4 − B3 (t)t2 − t3 + 3B3 (t)t − t2 − B3 (t) + t + 1
6
MIKAEL VEJDEMO-JOHANSSON
3
3
2
3
2
3
−4t4 B3 (t) −8B3 (t) t3 −2B3 (t) t4 +8B3 (t) t2 −4B3 (t) t3 +3B3 (t)t4 +12B3 (t) t+
2
3
2
2
20B3 (t) t2 + 6B3 (t)t3 + 3t4 − 8B3 (t) + 30B3 (t) t + 8B3 (t) v6 + 18B3 (t)t2 + 6t3 −
2
12B3 (t) + 27B3 (t)t + 12B3 (t)v6 + 6t2 − 6B3 (t) + 9t + 6v6
2
2
2
2
− 2B3 (t) t4 − 4B3 (t) t3 + 4B3 (t) t2 + B3 (t)t3 + 2t4 + 6B3 (t) t + 8B3 (t)t2 + 3t3 −
2
4B3 (t) + 10B3 (t)t + 4B3 (t)v6 + 4t2 + 2tv6 − 3B3 (t) + 4t + 2v6 + 1
2
2
2
2
−2B3 (t) t4 −4B3 (t) t3 +3B3 (t)t4 +4B3 (t) t2 +5B3 (t)t3 −t4 +6B3 (t) t+3B3 (t)t2
2
− t3 − 4B3 (t) + 6B3 (t)t + 4B3 (t)v6 − 3t2 + 2v6 2 − B3 (t) − 2t + 2v6 − 3
− B3 (t)t4 − 2B3 (t)t3 + t4 + 2B3 (t)t2 + 2t3 + 3B3 (t)t + 2t2 − 2B3 (t) + 3t + v5 + v6
− 4B3 (t)t4 −7B3 (t)t3 + 4t4 + 7B3 (t)t2 + 7t3 + 9B3 (t)t + 9t2 −6B3 (t) + 11t + 2v4 + 4
5
5
4
5
4
3
5
56B3 (t) t4 +88B3 (t) t3 +24B3 (t) t4 −80B3 (t) t2 +40B3 (t) t3 −40t4 B3 (t) −88B3 (t) t−
4
3
2
5
4
4
3
4
248B3 (t) t2 − 60B3 (t) t3 + 47B3 (t) t4 + 8B3 (t) − 280B3 (t) t + 96B3 (t) v3 −
3
2
3
256B3 (t) t2 + 109B3 (t) t3 + 14B3 (t)t4 − 88B3 (t) − 296B3 (t) t + 144B3 (t) v3 −
2
3
2
2
282B3 (t) t2 +25B3 (t)t3 −101t4 −148B3 (t) −387B3 (t) t+80B3 (t) v3 −385B3 (t)t2 −
2
202t3 +79B3 (t) −570B3 (t)t+12B3 (t)v3 −168B3 (t)v6 −205t2 +181B3 (t)−303t+6v3 −196v6 −6
4
4
3
4
3
2
−868B3 (t) t4 −1364B3 (t) t3 −694t4 B3 (t) +1240B3 (t) t2 −1126B3 (t) t3 +2011B3 (t) t4 +
4
3
2
4
3
1364B3 (t) t+4304B3 (t) t2 +3545B3 (t) t3 +1376B3 (t)t4 −124B3 (t) +4846B3 (t) t−
3
2
3
2
1488B3 (t) v3 + 2620B3 (t) t2 + 2595B3 (t)t3 − 1825t4 + 1318B3 (t) + 2321B3 (t) t−
2
2
2784B3 (t) v3 − 3915B3 (t)t2 − 3650t3 + 4935B3 (t) − 6772B3 (t)t − 1216B3 (t)v3 −
2064B3 (t)v6 − 3625t2 + 364tv3 + 4941B3 (t) − 5475t − 50v3 − 3336v6 − 132
4
4
3
4
3
2
252B3 (t) t4 + 396B3 (t) t3 − 4t4 B3 (t) − 360B3 (t) t2 + 4B3 (t) t3 − 1077B3 (t) t4 −
4
3
2
4
3
396B3 (t) t − 956B3 (t) t2 − 2039B3 (t) t3 + 170B3 (t)t4 + 36B3 (t) − 1084B3 (t) t+
3
2
3
2
432B3 (t) v3 + 1030B3 (t) t2 + 321B3 (t)t3 + 659t4 − 412B3 (t) + 1745B3 (t) t+
2
2
456B3 (t) v3 + 2939B3 (t)t2 + 1318t3 − 2149B3 (t) + 4426B3 (t)t + 36B3 (t)v3 +
1656B3 (t)v6 + 1299t2 − 1411B3 (t) + 1977t + 91v2 − 53v3 + 1356v6 − 38
REFERENCES
4
4
7
3
4
3
2
504B3 (t) t4 + 792B3 (t) t3 − 8t4 B3 (t) − 720B3 (t) t2 + 8B3 (t) t3 − 2154B3 (t) t4 −
4
3
2
4
3
792B3 (t) t − 1912B3 (t) t2 − 4078B3 (t) t3 + 158B3 (t)t4 + 72B3 (t) − 2168B3 (t) t+
3
2
3
2
864B3 (t) v3 + 2060B3 (t) t2 + 369B3 (t)t3 + 1500t4 − 824B3 (t) + 3490B3 (t) t+
2
2
912B3 (t) v3 + 6151B3 (t)t2 + 2909t3 − 4298B3 (t) + 9125B3 (t)t + 72B3 (t)v3 +
3312B3 (t)v6 + 3053t2 − 2822B3 (t) + 4409t + 182v1 − 288v3 + 2712v6 + 106
The first of these terms is the elimination order projection, producing the functional
equation
B3 (t)t5 + B3 (t)t4 − t5 − 3B3 (t)t3 − t4 − B3 (t)t2 − t3 + 3B3 (t)t − t2 − B3 (t) + t + 1 = 0
Which we can solve for B3 (t), producing
B3 (t) =
t5 + t4 + t3 + t2 − t − 1
= 1 + 4t + 10t2 + 22t3 + 44t4 + 84t5 + O(t6 )
t5 + t4 − 3t3 − t2 + 3t − 1
Comparing this to a hand-enumeration of small braids produces 12 braids using
two elementary generators, whereas this enumeration predicts 10. The reason for
this discrepancy is precisely the fact that geodesic here is measured in terms not of
elementary generators but in terms of automatic generators. Hence, while σ1−1 σ2
and σ2−1 σ1 are both length-2 words in the elementary generating set, they have
minimal representatives in the automatic presentation as
σ1−1 σ2 = σ2 σ1 σ2−1 σ1−1 = σ1−1 Dσ2
and
σ2−1 σ1 = σ1 σ2 σ1−1 σ2−1 = σ2−1 Dσ1
where D = σ1 σ2 σ1 σ2−1 σ1−1 σ2−1 , and hence shows up as length 3 instead.
4. Conclusion
The methods from analytical combinatorics producing generating functions directly
from contextfree grammars are directly applicable to the problem of computing
growth functions for automatic groups. They can be weighted, which provides some
insight into how the automatic group geodesic words relate to their presentation
in a different choice of generators – however, for, for instance, braid groups, the
automatic presentations tend to sort generators moving the elementary generators
to the front and their inverses to the end of a word, which may not produce a
geodesic in the simpler presentation.
It is unclear how to get closer to a growth function for the elementary presentation
of a braid group.
References
[1]
Ruth Charney. “Geodesic automation and growth functions for Artin groups
of finite type”. In: Mathematische Annalen 301.1 (1995), pp. 307–324.
[2] Noam Chomsky and Marcel P Schützenberger. “The algebraic theory of contextfree languages”. In: Studies in Logic and the Foundations of Mathematics 35
(1963), pp. 118–161.
[3] David Epstein et al. Word processing in groups. AK Peters, Ltd., 1992.
| 4math.GR
|
ON GROUPS WITH THE SAME CHARACTER DEGREES AS
ALMOST SIMPLE GROUPS WITH SOCLE THE MATHIEU
GROUPS
arXiv:1511.04129v3 [math.GR] 24 Jan 2016
SEYED HASSAN ALAVI, ASHRAF DANESHKHAH, AND ALI JAFARI
Abstract. Let G be a finite group and cd(G) denote the set of complex irreducible character degrees of G. In this paper, we prove that if G is a finite
group and H is an almost simple group whose socle is Mathieu group such that
cd(G) = cd(H), then there exists an Abelian subgroup A of G such that G/A
is isomorphic to H. This study is heading towards the study of an extension of
Huppert’s conjecture (2000) for almost simple groups.
1. Introduction
Let G be a finite group, and let Irr(G) be the set of all complex irreducible character degrees of G. Denote by cd(G) the set of character degrees of G, that is to
say, cd(G) = {χ(1)|χ ∈ Irr(G)}. It is well-known that the complex group algebra
CG determines character degrees and their multiplicities.
There is growing interest in a question regarding the structure of G which can
be recovered from the character degree set of G with or without multiplicity. It is
well-known that the character degree set of G can not use to completely determine
the structure of G. For example, the non-isomorphic groups D8 and Q8 not only
have the same set of character degrees, but also share the same character table. The
character degree set cannot be used to distinguish between solvable and nilpotent
groups. For example, if G is either Q8 or S3 , then cd(G) = {1, 2}. Recently, Navarro
[14] showed that the character degree set alone cannot determine the solvability of
the group. Indeed, he constructed a finite perfect group H and a finite solvable
group G such that cd(G) = cd(H). It is also discovered by Navarro and Rizo [15]
that there exists a finite perfect group and a finite nilpotent group with the same
character degree set. Notice that in both examples, these finite perfect groups are
not nonabelian simple. It remains open whether the complex group algebra can
determine the solvability of the group or not (see Brauer’s Problem 2 [5]).
However, the situation for simple groups and related groups is rather different.
It has been proved recently that all quasisimple groups are uniquely determined up
to isomorphism by their complex group algebras [3] in which a finite group G is
quasisimple if G is perfect and G/Z(G) is a nonabelian simple group. In the late
1990s, Huppert [10] posed a conjecture which asserts that the nonabelian simple
groups are essentially characterized by the set of their character degrees.
Date: January 26, 2016.
2010 Mathematics Subject Classification. Primary 20C15; Secondary 20D05.
Key words and phrases. Character degrees; Almost simple groups; Sporadic simple groups;
Mathieu groups; Huppert’s Conjecture.
Corresponding author: Ashraf Daneshkhah.
1
2
S.H. ALAVI, A. DANESHKHAH, AND A. JAFARI
Conjecture 1.1. (Huppert) Let G be a finite group and H a finite nonabelian
simple group such that the sets of character degrees of G and H are the same. Then
G∼
= H × A, where A is an abelian group.
This conjecture is verified for Alternating groups, many of the simple groups of
Lie type [16, 19, 20] and for all sporadic simple groups [1, 2, 11, 18].
In this paper, we initiate an investigation on an extension of Huppert’s conjecture
for almost simple groups whose socle is the sporadic simple groups. A group H is
called almost simple if there exists a nonabelian simple group H0 such that H0 6
H 6 Aut(H0 ). Indeed, this paper is devoted to studying finite groups with the same
character degrees as almost simple groups H with socle H0 being one of the Mathieu
groups:
Theorem 1.1. Let G be a finite group, and let H be an almost simple group whose
socle is one of the Mathieu groups. If cd(G) = cd(H), then there exists an abelian
group A such that G/A is isomorphic to H.
In order to prove Theorem 1.1, we establish the following steps which Huppert
introduced in [10]. Let H be an almost simple group with socle H0 , and let G be a
group with the same character degrees as H. Then we show that
Step 1.: G′ = G′′ ;
Step 2.: if G′ /M is a chief factor of G, then G′ /M is isomorphic to H0 ;
Step 3.: if θ ∈ Irr(M) with θ(1) = 1, then IG′ (θ) = G′ and so M = M ′ ;
Step 4.: M = 1 and G′ ∼
= H0 ;
′
Step 5.: G/CG (G ) is isomorphic to H.
Note that to prove Step 2, we determine all finite simple groups whose irreducible
character degrees divide some irreducible character degrees of almost simple groups
with socle sporadic simple groups, and by Proposition 3.1, all such simple groups
are listed in Table 1. This result somehow relates to [17, Theorem 1] for sporadic
simple groups.
2. Preliminaries
In this section, we present some useful results to prove Theorem 1.1. We first
establish some definitions and notation.
Throughout this paper all groups are finite. Recall that a group H is said to
be an almost simple group with socle H0 if H0 6 H 6 Aut(H0 ), where H0 is a
nonabelian simple group. For a positive integer n, π(n) denotes the set of all prime
divisors of n. If G is a group, we will write π(G) instead of π(|G|). If N E G and
θ ∈ Irr(N), then the inertia group of θ in G is denoted by IG (θ) and is defined by
P
IG (θ) = {g ∈ G | θg = θ}. If the character χ = ki=1 ei χi , where each χi is an
irreducible character of G and ei is a nonnegative integer, then those χi with ei > 0
are called the irreducible constituents of χ. The set of all irreducible constituents
of θG is denoted by Irr(G|θ). All further notation and definitions are standard and
could be found in [9, 12].
Lemma 2.1. [9, Theorems 19.5 and 21.3] Suppose N E G and χ ∈ Irr(G).
(a) If χN = θ1 +θ2 +· · ·+θk with θi ∈ Irr(N), then k divides |G/N|. In particular,
if χ(1) is prime to |G/N|, then χN ∈ Irr(N).
(b) (Gallagher’s Theorem) If χN ∈ Irr(N), then χψ ∈ Irr(G) for all ψ ∈ Irr(G/N).
GROUPS WITH CHARACTER DEGREES OF ALMOST SIMPLE MATHIEU GROUPS
3
Lemma 2.2. [9, Theorems 19.6 and 21.2] Suppose N E G and θ ∈ Irr(N). Let
I = IG (θ).
P
(a) If θI = ki=1 φi with φi ∈ Irr(I), then φG
i ∈ Irr(G). In particular, φi (1)|G :
I| ∈ cd(G).
(b) If θ extends to ψ ∈ Irr(I), then (ψτ )G ∈ Irr(G) for all τ ∈ Irr(I/N). In
particular, θ(1)τ (1)|G : I| ∈ cd(G).
(c) If ρ ∈ Irr(I) such that ρN = eθ, then ρ = θ0 τ0 , where θ0 is a character of an
irreducible projective representation of I of degree θ(1) and τ0 is a character
of an irreducible projective representation of I/N of degree e.
Lemma 2.3. [18, Lemma 3] Let G/N be a solvable factor group of G minimal with
respect to being nonabelian. Then two cases can occur.
(a) G/N is an r-group for some prime r. Hence there exists ψ ∈ Irr(G/N) such
that ψ(1) = r b > 1. If χ ∈ Irr(G) and r ∤ χ(1), then χτ ∈ Irr(G) for all
τ ∈ Irr(G/N).
(b) G/N is a Frobenius group with an elementary abelian Frobenius kernel F/N.
Then f = |G : F | ∈ cd(G) and |F/N| = r a for some prime r, and a is
the smallest integer such that r a ≡ 1 mod f . If ψ ∈ Irr(F ), then either
f ψ(1) ∈ cd(G) or r a divides ψ(1)2 . In the latter case, r divides ψ(1).
(1) If no proper multiple of f is in cd(G), then χ(1) divides f for all χ ∈
Irr(G) such that r ∤ χ(1), and if χ ∈ Irr(G) such that χ(1) ∤ f , then
r a | χ(1)2 .
(2) If χ ∈ Irr(G) such that no proper multiple of χ(1) is in cd(G), then either
f divides χ(1) or r a divides χ(1)2 . Moreover if χ(1) is divisible by no
nontrivial proper character degree in G, then f = χ(1) or r a | χ(1)2 .
Lemma 2.4. Let G be a finite group.
(a) If G is a nonabelian simple group, then there exists a nontrivial irreducible
character ϕ of G that extends to Aut(G).
(b) If N is a minimal normal subgroup of G so that N ∼
= S k , where S is a
nonabelian simple group, and ϕ ∈ Irr(S) extends to Aut(S), then ϕk ∈ Irr(N)
extends to G.
Proof. Part (a) follows from [4, Theorems 2-4] and to prove part (b) see [4, Lemma
5].
Lemma 2.5. [10, Lemma 6] Suppose that M E G′ = G′′ and for every λ ∈ Irr(M)
with λ(1) = 1, λg = λ for all g ∈ G′ . Then M ′ = [M, G′ ] and |M/M ′ | divides the
order of the Schur multiplier of G′ /M.
Lemma 2.6. [13, Theorem D] Let N be a normal subgroup of a finite group G and
let ϕ ∈ Irr(N) be G-invariant. Assume that χ(1)/ϕ(1) is odd, for all χ(1) ∈ Irr(G|ϕ).
Then G/N is solvable.
3. Degree properties of almost simple groups with socle sporadic
In this section, we determine all finite simple groups whose irreducible character
degrees divide some irreducible character degrees of almost simple groups whose
socles are sporadic simple groups.
Proposition 3.1. Let S be a simple group, and let H be an almost simple group
whose socle is a sporadic simple group. Then the character degrees of S divides some
character degrees of H if and only if H and S are as in Table 1.
4
S.H. ALAVI, A. DANESHKHAH, AND A. JAFARI
Proof. Suppose that H is an almost simple group with socle a sporadic simple group
H0 . Suppose also that S is a simple group whose degrees divide some degrees of
H. Then the prime divisors of the degrees of S are exactly those primes dividing
|S|. Therefore, π(S) ⊆ π(H), and hence by [21], we know all such possible simple
groups S. We only need to check if the degrees of S divide some degrees of H and
this could mainly be done by [7, 8]. Since our arguments are similar for each group
H, we only give a detailed proof for the cases where H is M12 : 2 or M22 : 2 which
will be frequently used and referred to in Sections 4 and 5.
Suppose first H = M12 : 2. Then π(S) ⊆ {2, 3, 5, 11}, and so by [21], S is
isomorphic to one of the simple groups A5 , A6 , L2 (11), U5 (2), S4 (3), M11 and M12 .
Note that U5 (2) and S4 (3) have degrees 220 and 81, respectively. Therefore, S is
isomorphic to A5 , A6 , L2 (11), M11 or M12 .
Suppose now H = M22 : 2. Then π(S) ⊆ {2, 3, 5, 7, 11}, and so by [21], S is
isomorphic to one of the simple groups in A1 ∪ A2 ∪ A3 , where
A1 := {A5 , A6 , A7 , L2 (7), L2 (8), M22 },
A2 := {A8 , L3 (4), L2 (49), U3 (3), S4 (7), M11 },
A3 := {A9 , A10 , A11 , A12 , L2 (11), U3 (5), U4 (3), U5 (2), U6 (2),
S4 (3), S6(2), O8+ (2), M12 , McL, HS, J2 }.
If S ∈ A2 , then S has a degree divisible by 25, 27, 44, 49 or 64, which is a contradiction. If S ∈ A3 , then S has a degree which is divisible by 12, which is also a
contradiction. Therefore S ∈ A1 as claimed.
Table 1: Simple groups S whose irreducible character
degrees divide some character degrees of almost simple
groups H with socle sporadic simple groups.
H
M11
M12 , M12 : 2
M22 , M22 : 2
M23
M24
S
An for n = 5, 6, M11
An for n = 5, 6, L2 (11), M11 , M12
An for n = 5, 6, 7, L2 (q) for q = 7, 8, M22
An for n = 5, 6, 7, L2 (q) for q = 7, 8, M11 , M23
An for n = 5, 6, 7, 8, L2 (q) for q = 7, 8, 11, 23, L3 (4),
U3 (3), M11 , M24
J1
A5 , L2 (q) for q = 7, 11, J1
J2
An for n = 5, 6, 7, L2 (q) for q = 7, 8, U3 (3), J2
J2 : 2
A8 , L3 (4), and all S in J2 ,
J3 , J3 : 2
An for n = 5, 6, S4 (3), L2 (q) for q = 16, 17, 19, J3
J4
An for n
=
5, 6, 7, 8, L2 (q) for q
=
7, 8, 11, 23, 29, 31, 32, 43, L3 (4), L5 (2), U3 (3), U3 (11),
M11 , M12 , M22 , J4
HS, HS : 2
An for n = 5, 6, 7, 8, L2 (q) for q = 7, 8, 11, L3 (4), M11 ,
M22 , HS
McL, McL : 2 An for n = 5, 6, 7, 8, L2 (q) for q = 7, 8, 11, L3 (4),
U3 (3), U3 (5) S4 (3), M11 , McL
Continued
GROUPS WITH CHARACTER DEGREES OF ALMOST SIMPLE MATHIEU GROUPS
H
Suz,
Suz : 2
Co3
Co2
Co1
He, He : 2
F i22
F i22 : 2
F i23
F i′24 , F i′24 : 2
Th
Ru
Table 1 – Continued
S
An for n = 5, . . . , 10, L2 (q) for q = 7, 8, 11, 13, 25,
27, 64, L3 (3), L3 (4), L3 (9), L4 (3), U3 (3), U3 (4), U4 (2),
U4 (3), U5 (2), S6 (2), 2 B2 (8), G2 (3),2 F4 (2)′ , M11 , M12 ,
M22 , Suz, J2
A11 , and all S in Suz,
An for n = 5, 6, 7, 8, 9, 11, L2 (q) for q = 7, 8, 11, 23,
L3 (4), U3 (q) for q = 3, 5 , U4 (3), S4 (3), S6 (2), M11 ,
M12 , M22 , M23 , M24 , Co3
An for n = 5, . . . , 11, L2 (q) for q = 7, 8, 11, 23, L3 (4),
U3 (q) for q = 3, 5, U4 (3), U5 (2), S4 (3), S6 (2), M11 ,
M12 , M22 , M23 , M24 J2 , Co2
An for n = 5, . . . , 16, L2 (q) for q = 7, 8, 11, 13, 23,
25, 27, 49, 64, L3 (q) for q = 3, 4, 9, L4 (3), U3 (q) for
q = 3, 4, 5, U4 (3),U5 (2), U6 (2), S4 (q) for q = 3, 5,
8, S6 (q) for q = 2, 3, 2 B2 (8), O7 (3), O8+ (2), G2 (q) for
q = 3, 4, 3 D4 (2), 2 F4 (2)′ , M11 , M12 , M22 , M23 , M24 ,
McL, J2 , HS, Co1 , Co3
An for n = 5, 6, 7, 8, L2 (q) for q = 7, 8, 16, 17, 49,
L3 (4), U3 (3), S4 (4), He
An for n = 5, . . . , 11, A13 , L2 (q) for q = 7, 8, 11,
13, 25, 27, 64, L3 (q) for q = 3, 4, 9, L4 (3), U3 (q) for
q = 3, 4, U4 (3), U5 (2), U6 (2), S4 (3), S6 (2), 2 B2 (8),
O8+ (2), G2 (q) for q = 3, 4, M11 , M12 , M22 , J2 , McL,
F i22
S6 (3), O7 (3), and all S in F i22 .
An for n = 5, . . . , 13, L2 (q) for q = 7, 8, 11, 13, 16,
17, 23, 25, 27, 64, L3 (q) for q = 3, 4, 9, 162 , L4 (q)
for q = 3, 4, U3 (q) for q = 3, 4, U4 (q) for q = 3,
4, U5 (2), S4 (3), S4 (4), S6 (2), S6 (3), 2 B2 (8), O7 (3),
O8+ (2), O8− (2), G2 (q) for q = 3, 4, 2 F4 (2)′ , M11 , M12 ,
M22 , M23 , M24 , HS, J2 , F i23
An for n = 5, . . . , 14, L2 (q) for q = 7, 8, 11, 13, 16, 17,
23, 25, 27, 29, 49, 64, L3 (q) for q = 3, 4, 9, 162 , L4 (q)
for q = 3, 4, L6 (3), U3 (q) for q = 3, 4, U4 (3), U5 (2),
S4 (q) for q = 3, 4, 8, S6 (2), S6 (3), S8 (2), 2 B2 (8),
O7 (3), O8± (2), G2 (q) for q = 3, 4, 3 D4 (2), 2 F4 (2)′ ,
M11 , M12 , M22 , M23 , M24 , He, J2 , F i′24
An for n = 5, . . . , 10, L2 (q) for q = 7, 8, 13, 19, 25,
27, 31, 49, 64, L3 (q) for q = 3, 4, 5, 9, L4 (3), L5 (2),
U3 (q) for q = 3, 4, 5, 8, S4 (3), S4 (8),S6 (2), 2 B2 (8),
3
D4 (2), G2 (q) for q = 3, 4, 2 F4 (2)′ , J2 , T h
An for n = 5, . . . , 8, L2 (q) for q = 7, 8, 13, 25, 27, 29,
64, L3 (q) for q = 3, 4, U3 (q) for q = 3, 4, 5, J2 , Ru
Continued
5
6
S.H. ALAVI, A. DANESHKHAH, AND A. JAFARI
H
Ly
HN, HN : 2
O ′ N, O ′N : 2
B
M
Table 1 – Continued
S
An for n = 5, . . . , 9, A11 , L2 (q) for q = 7, 8, 11, 31,
32, L3 (q) for q = 4, 5, U3 (q) for q = 3, 5, S4 (3), M11 ,
M12 , M22 , J2 , McL, HS, Ly
An for n = 5, . . . , 10, L2 (q) for q = 7, 8, 11, 19, L3 (q)
for q = 4. 19, L4 (7), U3 (q) for q = 3, 5, 8, U4 (3),
S4 (3),O8+ (2), M11 , M12 , M22 , J1 , J2 , HS, HN
An for n = 5, . . . , 8, L2 (q) for q = 7, 8, 11, 16, 17, 19,
31, 32, L3 (q) for q = 4, 7, U3 (q) for q = 3, 8, S4 (3),
S6 (2), M11 , M12 , M22 , O ′ N
An for n = 5, . . . , 28, L2 (q) for q = 7, 8, 11, 13, 16,
17, 19, 23, 27, 31, 32, 47, 49, 64, 125, L3 (q) for q = 3,
4, 5, 9, 16, 25, L4 (q) for q = 3, 4, 5, L5 (2), L5 (4),
L6 (2), L6 (4), U3 (q) for q = 3, 4, 5, 8, U4 (q) for q = 2,
3, 4, 5, 8, U5 (2), U6 (2), S4 (q) for q = 4, 5, 7, 8, G2 (q)
±
±
for q = 3, 4, 5, O7 (3), O8± (2), O8+ (3), O10
(2), O12
(2),
2
′ 3
F4 (2) , D4 (2), M11 , M12 , M22 , M23 , M24 , J1 , J2 ,
J3 , HS, McL, Suz, F i22 , Co3 , Co2 , T h, B
An for n = 5, . . . , 36, L2 (q) for q = 7, 8, 11, 13, 16,
17, 19, 23, 25, 27, 29, 31, 32, 41, 47, 49, 59, 64, 71, 81,
169, 1024, L3 (q) for q = 3, 4, 5, 7, 9, 16, 19, 25, L4 (q)
for q = 3, 4, 5, 7, 9, L5 (q) for q = 2, 4, L6 (q), for q =
2, 3, 4, U3 (q) for q = 3, 4, 5, 8, 27, U4 (q) for q = 2, 3,
4, 5, 8, 9, U5 (q) for q = 2, 4, U6 (2) for q = 2, 4, S4 (q)
for q = 4, 5, 7, 8, 9, S6 (q) for q = 2, 3, 4, 5, S8 (2),
S10 (2), 2 B2 (8), 2 B2 (32), G2 (q) for q = 3, 4, 5, O7 (3),
±
±
±
O7 (5), O9 (3),O8± (2), O8± (3), O10
(2), O10
(3), O12
(2),
2
G2 (27), 2 F4 (2)′ , F4 (2), 3 D4 (2), M11 , M12 , M22 , M23 ,
M24 , J1 , J2 , J3 , HS, McL, Suz, F i22 , Co3 , Co2 , T h,
He, O ′N, Ru, M
4. Groups with socle M12
In this section, we prove Theorem 1.1 for almost simple group H whose socle is
H0 := M12 . By [11], Theorem 1.1 is proved when H = H0 = M12 . Therefore, we
only need to deal with the case where H := M12 : 2. For convenience, we mention
some properties of H and H0 some of which can be obtained in [7, pp. 31-33].
Lemma 4.1. Let H0 := M12 and H := M12 : 2. Then
(a) The Schur multiplier and the group of outer automorphisms of H0 are Z2 ;
(b) The degrees of irreducible characters of H are
45 = 32 · 5
66 = 2 · 3 · 11
120 = 23 · 3 · 5
22 = 2 · 11
54 = 2 · 33
99 = 32 · 11
144 = 24 · 32
32 = 25
55 = 5 · 11
110 = 2 · 5 · 11
176 = 24 · 11
1
(c) If K is a maximal subgroup of H0 whose index in H0 divides some degrees
χ(1) of H, then one of the following occurs:
GROUPS WITH CHARACTER DEGREES OF ALMOST SIMPLE MATHIEU GROUPS
7
(i) K ∼
= M11 and χ(1)/|H0 : K| divides 2 · 5 or 22 · 3;
(ii) K ∼
= M10 : 2 and χ(1)/|H0 : K| = 1;
(iii) K ∼
= L2 (11) and χ(1)/|H0 : K| = 1.
(d) If S is a finite nonabelian simple group whose irreducible character degrees
divide some degrees of H, then S is isomorphic to A5 , A6 , L2 (11), M11 or
M12 .
Proof. Parts (a) and (b) follows from [7, pp. 31-33] and part (d) follows from
Proposition 3.1 and Table 1. Part (c) is a straightforward calculation.
4.1. Proof of Theorem 1.1 for M12 : 2. As noted before, by [11], we may assume
that H := M12 : 2. We further assume that G is a finite group with cd(G) =
cd(M12 : 2). The proof of Theorem 1.1 follows from the following lemmas.
Lemma 4.2. G′ = G′′ .
Proof. Assume the contrary. Then there is a normal subgroup N of G, where N is
maximal such that G/N is a nonabelian solvable group. Now we apply Lemma 2.3
and we have one of the following cases:
(a) G/N is a r-group with r prime. In this case, G/N has an irreducible character ψ
of degree r b > 1, and so does G. Since M12 : 2 has an irreducible character of degree
32, we conclude that r = 2. Let now χ ∈ Irr(G) with χ(1) = 99. Then Lemma 2.1(a)
implies that χN ∈ Irr(N), and so by Lemma 2.1(b), G has an irreducible character
of degree 99ψ(1), which is a contradiction.
(b) G/N is a Frobenius group with kernel F/N. Then |G : F | ∈ cd(G) divides
r a − 1, where |F/N| = r a . Let |G : F | ∈ {2 · 11, 5 · 11} and χ(1) = 24 · 32 . Then
Lemma 2.2(b) implies that r a divides χ2 (1) = 28 · 34 , and it is impossible as |G : F |
does not divide r a − 1, for every divisor r a of 28 · 34 . Let now |G : F | 6∈ {2 · 11, 5 · 11}.
Then no proper multiple of |G : F | is in cd(G). If r = 2, then by Lemma 2.2(a),
both 32 · 11 and 32 · 5 must divide |G : F |, which is a contradiction. If r = 3, then
by Lemma 2.2(a), |G : F | is divisible by 25 and 24 · 11, which is a contradiction.
Similarly, if r 6= 2, 3, then 25 and 24 · 32 divide |G : F |, which is a contradiction.
Lemma 4.3. Let G′ /M be a chief factor of G. Then G′ /M ∼
= M12 .
Proof. Suppose G′ /M ∼
= S k , where S is a nonabelian simple group for some positive
integer k. Since S is a finite nonabelian simple group whose irreducible character
degrees divide some degrees of M12 : 2, by Lemma 4.1(d), S is isomorphic to one of
the groups A5 , A6 , M11 , M12 or L2 (11). If S is isomorphic to one of the groups A5 ,
A6 , M11 , M12 , L2 (11), then k = 1 as G has no degree divisible by 52 . Assume that
S is isomorphic to A5 or A6 , then G′ /M has a character ψ of degree 5. If χ is an
irreducible constituent of ψ G/M , then χ(1) = tψ(1) = 5t, where t divides | Out(S)|.
Consequently, G has a character of degree at most 20, which is a contradiction.
Similarly, in the case where S is isomorphic to M11 or L2 (11), the factor group G/M
has a character of degree 10t with t = 1, 2, and this implies that G has a character
of degree at most 20, which is a contradiction. Therefore G′ /M is isomorphic to
M12 .
Lemma 4.4. Let θ ∈ Irr(M) with θ(1) = 1. Then IG′ (θ) = G′ and M = M ′ .
P
Proof. Suppose I = IG′ (θ) < G′ . By Lemma 2.2, we have θI = ki=1 φi where φi ∈
Irr(I) for i = 1, 2, ..., k. Let U/M be a maximal subgroup of G′ /M ∼
= M12 containing
8
S.H. ALAVI, A. DANESHKHAH, AND A. JAFARI
I/M and set t := |U : I|. It follows from Lemma 2.2(a) that φi (1)|G′ : I| ∈ cd(G′ ),
and so tφi (1)|G′ : U| divides some degrees of G. Then |G′ : U| must divide some
character degrees of G, and hence by Lemma 4.1(c) one of the following holds.
(i) Suppose U/M ∼
= M11 .
= M11 . Then tφi (1) divide 2 · 5 or 22 · 3. If t = 1, then I/M ∼
Since M11 has trivial Schur multiplier, it follows that θ extends to θ0 ∈ Irr(I), and so
′
by Lemma 2.2(b) (θ0 τ )G ∈ Irr(G′ ), for all τ ∈ Irr(I/M). For τ (1) = 55 ∈ cd(M11 ),
it turns out that 12 · 55 · θ0 (1) divide some degrees of G, which is a contradiction.
Therefore, t 6= 1, and hence the index of a maximal subgroup of U/M ∼
= M11
2
containing I/M must divide 2 · 5 or 2 · 3. This implies that tφi (1) divides 22 · 3
and I/M ∼
= L2 (11). In particular, φi (1) = 1. Thus θ extends to a φi , and so
by Lemma 2.2(b), 144τ (1) ∈ cd(G′ ), for all τ ∈ Irr(I/M). This leads us to a
contradiction by taking τ (1) = 10 ∈ cd(L2 (11)).
(ii) Suppose U/M ∼
= M10 : 2. In this case t = 1, or equivalently, I/M = U/M ∼
=
M10 : 2. Moreover, φi (1) = 1, for all i. Then θ extends to φi ∈ Irr(I), and so
by Lemma 2.2(b), 66τ (1) divides some degrees of G, for τ (1) = 10, which is a
contradiction.
(iii) Suppose U/M ∼
= L2 (11), and
= L2 (11) and t = φi (1) = 1, for all i. Then I/M ∼
′
so θ extends to φi ∈ Irr(I). Thus 144τ (1) ∈ Irr(G ), for all τ ∈ Irr(I/M). This is
impossible by taking τ (1) = 10.
Therefore, IG′ (θ) = G′ . By Lemma 2.5, we have that |M/M ′ | divides the order
of Schur multiplier of G′ /M ∼
= M12 which is 2. If |M/M ′ | = 2, then G′ /M ′ is
isomorphic to 2 · M12 which has a character of degree 32 [7, p. 33]. Therefore M12
must have a degree divisible by 32, which is a contradiction. Hence |M/M ′ | = 1, or
equivalently, M = M ′ .
Lemma 4.5. The subgroup M is trivial, and hence G′ ∼
= M12 .
Proof. By Lemmas 4.3 and 4.4, we have that G′ /M ∼
= M12 and M = M ′ . Suppose
that M is nonabelian, and let N 6 M be a normal subgroup of G′ such that M/N
is a chief factor of G′ . Then M/N ∼
= S k , for some nonabelian simple group S. It
follows from Lemma 2.4 that S possesses a nontrivial irreducible character ϕ such
that ϕk ∈ Irr(M/N) extends to G′ /N. By Lemma 2.1(b), we must have ϕ(1)k τ (1) ∈
cd(G′ /N) ⊆ cd(G′ ), for all τ ∈ Irr(G′ /M). Now we can choose τ ∈ G′ /M such that
τ (1) is the largest degree of M12 , and since ϕ is nontrivial, ϕ(1)k τ (1) divides no
degree of G, which is a contradiction. Therefore, M is abelian, and since M = M ′ ,
we conclude that M = 1.
Lemma 4.6. There exists an abelian group A such that G/A ∼
= M12 : 2.
Proof. Set A := CG (G′ ). Since G′ ∩ A = 1 and G′ A ∼
= G′ × A, it follows that
G′ ∼
= G′ A/A E G/A 6 Aut(G′ ). By Lemma 4.5, we have G′ ∼
= M12 , and so we
conclude that G/A is isomorphic to M12 or M12 : 2. In the case where G/A is
isomorphic to M12 , we must have G ∼
= A × M12 . This is impossible as 32 ∈ cd(G)
but M12 has no character of degree 32. Therefore, G/A is isomorphic to M12 : 2.
5. Groups with socle M22
In this section, we prove Theorem 1.1 for almost simple group H whose socle is
H0 := M22 . Note that Theorem 1.1 is proved for H = H0 = M22 , see [11]. Therefore,
we only need to focus on case where H := M22 : 2. For convenience, we mention
some properties of H and H0 some of which can be obtained from [7, pp. 39-41].
GROUPS WITH CHARACTER DEGREES OF ALMOST SIMPLE MATHIEU GROUPS
9
Lemma 5.1. Let H0 := M22 and H := M22 : 2. Then
(a) The Schur multiplier H0 is Z12 and the group of outer automorphisms of H0
is Z2 ;
(b) The degrees of irreducible characters of H are
1
55 = 5 · 11
210 = 2 · 3 · 5 · 7
385 = 5 · 7 · 11
21 = 3 · 7
99 = 32 · 11
231 = 3 · 7 · 11
560 = 24 · 5 · 7
45 = 32 · 5
154 = 2 · 7 · 11
(c) If K is a maximal subgroup of H0 whose index in H0 divides some degrees
χ(1) of H, then one of the following occurs:
(i) K ∼
= L3 (4) and χ(1)/|H0 : K| divides 7;
(ii) K ∼
= 24 : S5 and χ(1)/|H0 : K| = 1;
(iii) K ∼
= 24 : A6 and χ(1)/|H0 : K| divides 2, 3 or 5.
(d) If S is a finite nonabelian simple group whose irreducible character degrees
divide some degrees of H, then S is isomorphic to A5 , A6 , A7 , L2 (7), L2 (8)
or M22 .
Proof. Parts (a)-(b) follows from [7, pp. 31-33] part (d) follows from Proposition 3.1
and Table 1. Part (c) is a straightforward calculation.
5.1. Proof of Theorem 1.1 for M22 : 2. Theorem 1.1 is true for the Mathieu
group M22 by [11]. It remains to assume that H := M22 : 2. In what follows assume
that G is a finite group with cd(G) = cd(M22 : 2). The proof follows from the
following lemmas.
Lemma 5.2. G′ = G′′ .
Proof. Assume the contrary. Then there is a normal subgroup N of G where N is a
maximal such that G/N is a nonabelian solvable group. We now apply Lemma 2.3,
and so we have the following two cases:
(a) Suppose that G/N is a r-group with r prime. Then it has an irreducible character
ψ of degree r b > 1. It is impossible as the group M22 : 2 does not have a irreducible
character of prime power degree.
(b) G/N is a Frobenius group with kernel F/N. Then |G : F | ∈ cd(G) divides r a −1,
where |F/N| = r a . Let |G : F | ∈ {3·7, 5·11} and χ(1) = 2·7·11. Then Lemma 2.2(b)
implies that r a divides χ2 (1) = 22 · 72 · 112 , and it is impossible as |G : F | does not
divide r a − 1, for every divisor r a of 22 · 72 · 112 . Let now |G : F | 6∈ {3 · 7, 5 · 11}.
Then no proper multiple of |G : F | is in cd(G). If r = 3, then by Lemma 2.2(a),
both 5 · 7 · 11 and 24 · 5 · 7 must divide |G : F |, which is a contradiction. If r = 5,
then by Lemma 2.2(a), both 3 · 7 · 11 and 2 · 7 · 11 must divide |G : F |, which is a
contradiction. If r = 11, then by Lemma 2.2(a), |G : F | is divisible by 2 · 3 · 5 · 7
and 24 · 5 · 7, which is a contradiction. Similarly, if r 6= 3, 5, and 11, then 32 · 11 and
5 · 11 divide |G : F |, which is impossible. Therefore, G′ = G′′ .
Lemma 5.3. Let G′ /M be a chief factor of G. Then G′ /M ∼
= M22 .
Proof. Suppose G′ /M ∼
= S k , where S is a nonabelian simple group and for some
positive integer k. Since S is a finite nonabelian simple group whose irreducible
character degrees divide some degrees of M22 : 2, by Lemma 5.1(d), the group S
is isomorphic to A5 , A6 , A7 , L2 (7), L2 (8) or M22 . Observe that G has no degree
10
S.H. ALAVI, A. DANESHKHAH, AND A. JAFARI
Table 2. The triples (S, ψ, d) in Lemma 5.3
S
| Out(S)| ψ(1) d
A5
2
5
10
A6
4
5
20
A7 , L2 (7)
2
6
12
L2 (8)
3
8
24
divisible by 25 and 49. This implies that k = 1 in each case. Assume that S is
isomorphic to one of the simple groups as in the first column of Table 2. Then by
[7], G′ /M has a character ψ of degree as in the third column of the same Table 2. If χ
is an irreducible constituent of ψ G/M , then χ(1) = tψ(1), where t divides | Out(S)|.
Consequently, G has a character of degree at most d as in the forth column of
Table 2, which is a contradiction. For example, if S is isomorphic to A7 or L2 (7),
then G′ /M has a character ψ of degree 6, then χ(1) = tψ(1) = 6t, where t divides
| Out(S)|, and so G has a character of degree at most 12, which is a contradiction.
Therefore G′ /M ∼
= M22 .
Lemma 5.4. If θ ∈ Irr(M), then IG′ (θ) = G′ and M = M ′ .
P
Proof. Suppose I := IG′ (θ) < G′ . By Lemma 2.2 we have θI = ki=1 φi where φi ∈
Irr(I) for i = 1, 2, ..., k. Let U/M be a maximal subgroup of G′ /M ∼
= M22 containing
I/M and set t := |U : I|. It follows from Lemma 2.2(a) that φi (1)|G′ : I| ∈ cd(G′ ),
and so tφi (1)|G′ : U| divides some degrees of G. Then |G′ : U| must divide some
character degrees of G, and hence by Lemma 5.1 one of the following holds:
(i) Suppose U/M ∼
= L3 (4) does
= L3 (4). Then, for each i, tφi (1) divide 7. As U/M ∼
∼
not have any subgroup of index 7, by [7, p. 23], so t = 1 and I/M = U/M ∼
= L3 (4)
and φi (1) divide 7. If φi (1) = 1, then θ extend to φi , and so by Lemma 2.2(b),
′
(φi τ )G ∈ Irr(G′ ), for all τ ∈ Irr(I/M). If τ (1) = 64 ∈ cd(L3 (4)), then 22τ (1) = 27 ·11
divides some degrees of G, which is a contradiction. Hence φi (1) = 7, for all i. Then
φiM = ei θ, where ei 6= 1 is the degree of a projective representation of I/M ∼
= L3 (4),
and it is impossible by [7, p. 24].
(ii) Suppose U/M ∼
= 24 : S5 and φi (1) = 1, for all i. Thus
= 24 : S5 . Then I/M ∼
θ extends to φi in I. It follows from Lemma 2.2(b) that τ (1)|G′ : I| divides some
character degrees of G, for all τ ∈ I/M. This is impossible by taking τ (1) = 4.
(iii) Suppose U/M ∼
= 24 : A6 . In this case, tφi (1) divides 2, 3 or 5. It follows from
[6] that U/M has no maximal subgroup of index 2, 3 and 5, and this implies that
I = U. Therefore, I/M ∼
= 24 : A6 , that is to say, I/M has an abelian subgroup
A/M of order 24 such that I/A ∼
= A6 . P
Let now λ ∈ Irr(A|θ), and write λI = fi µi . Since AEI, the degree µi (1) divides
2, 3 or 5, for all i. Since the index of a maximal subgroup of I/A ∼
= A6 is at least
6, λ is I-invariant, and so µiA = fi λ, for all i. If fi = 1 for some i, then λ extends
to λ0 ∈ Irr(I), and so by Lemma 2.2(b), λ0 τ is an irreducible character of λI , for all
τ ∈ Irr(I/A), and so λ0 (1)τ (1) = τ (1) divides 2, 3, or 5. This is impossible as we
could take τ (1) = 8 ∈ cd(A6 ). Therefore, fi > 1, for all i. Moreover, we know from
Lemma 2.2(c) that each fi is the degree of a nontrivial proper irreducible projective
representation of A6 , by [7, p. 5], we observe that fi ∈ {3, 5}. This shows that
µ(1)/λ(1) is odd, for all µ ∈ Irr(I|A), and so by Lemma 2.6, the group I/A ∼
= A6 is
solvable, which is a contradiction.
GROUPS WITH CHARACTER DEGREES OF ALMOST SIMPLE MATHIEU GROUPS
11
Table 3. Some character degrees of G′ /M ′ in Lemma 5.4
G′ /M ′ 2 · M22 3 · M22 4 · M22 6 · M22 12 · M22
Degree
440
384
440
384
384
This show IG′ (θ) = G′ . By Lemma 2.5, we have that |M/M ′ | divides the order
6 1, then G has an
of Schur multiplier of G′ /M ∼
= M22 which is 12. If |M/M ′ | =
irreducible character of degree divisible by one the degrees in the second row of
Table 3, which is a contradiction. Therefore, M = M ′ .
Lemma 5.5. The subgroup M is trivial, and hence G′ ∼
= M22 .
Proof. It follows from Lemmas 5.3 and 5.4 that G′ /M ∼
= M22 and M = M ′ . Assume
that M is nonabelian, and let N 6 M be a normal subgroup of G′ such that M/N
is a chief factor of G′ . Then M/N ∼
= S k for some nonabelian simple group S. By
Lemma 2.4, S has a nontrivial irreducible character ϕ such that ϕk ∈ Irr(M/N)
extends to G′ /N. Now Lemma 2.1(b) implies that ϕ(1)k τ (1) ∈ cd(G′ /N) ⊆ cd(G′ ),
for all τ ∈ Irr(G′ /M). As ϕ(1) > 1, if we choose τ ∈ G′ /M such that τ (1) is the
largest degree of M22 , then ϕ(1)k τ (1) divides no degree of G, which is a contradiction.
Therefore, M is abelian, and hence we are done.
Lemma 5.6. There exists an abelian group A such that G/A ∼
= M22 : 2.
Proof. Set A := CG (G′ ). Since G′ ∩ A = 1 and G′ A ∼
= G′ × A, it follows that
G′ ∼
= M22 , we conclude that G/A is isomorphic
= G′ A/AEG/A 6 Aut(G′ ). Since G′ ∼
to M22 or M22 : 2. In the case where G/A is isomorphic to M22 , we conclude that
G∼
= A × M22 . This is impossible as 560 ∈ cd(G) but M22 has no character of degree
560. Therefore, G/A is isomorphic to M22 : 2.
References
[1] S. H. Alavi, A. Daneshkah, H. P. Tong-Viet, and T. P. Wakefield. Huppert’s conjecture for
F i23 . Rend. Semin. Mat. Univ. Padova, 126:201–211, 2011. 2
[2] S. H. Alavi, A. Daneshkhah, H. P. Tong-Viet, and T. P. Wakefield. On Huppert’s conjecture
for the Conway and Fischer families of sporadic simple groups. J. Aust. Math. Soc., 94(3):289–
303, 2013. 2
[3] C. Bessenrodt, H. N. Nguyen, J. B. Olsson, and H. P. Tong-Viet. Complex group algebras of
the double covers of the symmetric and alternating groups. Algebra Number Theory, 9(3):601–
628, 2015. 1
[4] M. Bianchi, D. Chillag, M. L. Lewis, and E. Pacifici. Character degree graphs that are complete
graphs. Proc. Amer. Math. Soc., 135(3):671–676 (electronic), 2007. 3
[5] R. Brauer. Representations of finite groups. In Lectures on Modern Mathematics, Vol. I, pages
133–175. Wiley, New York, 1963. 1
[6] T. Connor and D. Leemans. An Atlas of subgroup lattices of finite almost simple groups.
ArXiv e-prints, June 2013. 10
[7] J. H. Conway, R. T. Curtis, S. P. Norton, R. A. Parker, and R. A. Wilson. Atlas of finite
groups. Oxford University Press, Eynsham, 1985. Maximal subgroups and ordinary characters
for simple groups, With computational assistance from J. G. Thackray. 4, 6, 7, 8, 9, 10
[8] The GAP Group. GAP – Groups, Algorithms, and Programming, Version 4.7.8, 2015. 4
[9] B. Huppert. Character theory of finite groups, volume 25 of de Gruyter Expositions in Mathematics. Walter de Gruyter & Co., Berlin, 1998. 2, 3
[10] B. Huppert. Some simple groups which are determined by the set of their character degrees.
I. Illinois J. Math., 44(4):828–842, 2000. 1, 2, 3
12
S.H. ALAVI, A. DANESHKHAH, AND A. JAFARI
[11] B. Huppert. Some simple groups which are determined by the set of their character degrees i.
Preprint, 2000. 2, 6, 7, 8, 9
[12] I. M. Isaacs. Character theory of finite groups. AMS Chelsea Publishing, Providence, RI, 2006.
Corrected reprint of the 1976 original [Academic Press, New York; MR0460423]. 2
[13] A. Moretó. An answer to a question of Isaacs on character degree graphs. Adv. Math.,
201(1):90–101, 2006. 3
[14] G. Navarro. The set of character degrees of a finite group does not determine its solvability.
Proc. Amer. Math. Soc., 143(3):989–990, 2015. 1
[15] G. Navarro and N. Rizo. Nilpotent and perfect groups with the same set of character degrees.
J. Algebra Appl., 13(8):1450061, 3, 2014. 1
[16] H. N. Nguyen, H. P. Tong-Viet, and T. P. Wakefield. Projective special linear groups PSL4 (q)
are determined by the set of their character degrees. J. Algebra Appl., 11(6):1250108, 26, 2012.
2
[17] H. P. Tong-Viet. Alternating and sporadic simple groups are determined by their character
degrees. Algebr. Represent. Theory, 15(2):379–389, 2012. 2
[18] H. P. Tong-Viet and T. P. Wakefield. On Huppert’s conjecture for the Monster and Baby
Monster. Monatsh. Math., 167(3-4):589–600, 2012. 2, 3
[19] T. P. Wakefield. Verifying Huppert’s conjecture for PSL3 (q) and PSU3 (q 2 ). Comm. Algebra,
37(8):2887–2906, 2009. 2
[20] T. P. Wakefield. Verifying Huppert’s conjecture for 2 G2 (q 2 ). Algebr. Represent. Theory,
14(4):609–623, 2011. 2
[21] A. V. Zavarnitsine. Finite simple groups with narrow prime spectrum. ArXiv e-prints, Oct.
2008. 4
S.H. Alavi, Department of Mathematics, Faculty of Science, Bu-Ali Sina University, Hamedan, Iran
E-mail address: [email protected] [email protected] (Gmail is preferred)
A. Daneshkhah, Department of Mathematics, Faculty of Science, Bu-Ali Sina
University, Hamedan, Iran
E-mail address: [email protected] [email protected] (Gmail is preferred)
A. Jafari, Department of Mathematics, Faculty of Science, Bu-Ali Sina University, Hamedan, Iran
E-mail address: [email protected]
| 4math.GR
|
arXiv:1704.05596v2 [cs.LG] 23 Nov 2017
Insensitive Stochastic Gradient Twin Support Vector
Machines for Large Scale Problems
Zhen Wanga , Yuan-Hai Shaob,∗, Lan Baia , Li-Ming Liuc , Nai-Yang Dengd
a
School of Mathematical Sciences, Inner Mongolia University, Hohhot, 010021,
P.R.China
b
School of Economics and Management, Hainan University, Haikou, 570228, P.R. China
c
School of Statistics, Capital University of Economics and Business, Beijing, 100070,
P.R.China
d
College of Science China Agricultural University, Beijing, 100083, P.R.China
Abstract
Stochastic gradient descent algorithm has been successfully applied on support vector machines (called PEGASOS) for many classification problems.
In this paper, stochastic gradient descent algorithm is investigated to twin
support vector machines for classification. Compared with PEGASOS, the
proposed stochastic gradient twin support vector machines (SGTSVM) is
insensitive on stochastic sampling for stochastic gradient descent algorithm.
In theory, we prove the convergence of SGTSVM instead of almost sure convergence of PEGASOS. For uniformly sampling, the approximation between
SGTSVM and twin support vector machines is also given, while PEGASOS
only has an opportunity to obtain an approximation of support vector machines. In addition, the nonlinear SGTSVM is derived directly from its linear
case. Experimental results on both artificial datasets and large scale problems show the stable performance of SGTSVM with a fast learning speed.
Keywords: Classification, support vector machines, twin support vector
machines, stochastic gradient descent, large scale problem.
∗
Corresponding author. Tel./Fax:(+86)0571-87313551.
Email address: [email protected] (Yuan-Hai Shao )
Preprint submitted to Elsevier
November 27, 2017
1. Introduction
Support vector machines (SVM), being powerful tool for classification
[7, 20, 42], have already outperformed most other classifiers in a wide variety of applications [23, 17, 11]. Different from SVM with a pair of parallel hyperplanes, twin support vector machines (TWSVM) [12, 35] with a
pair of nonparallel hyperplanes has been proposed and developed, e.g., twin
bounded support vector machines (TBSVM) [35], twin parametric margin
support vector machines (TPMSVM) [24], and weighted Lagrangian twin
support vector machines (WLTSVM) [33]. These classifiers have been widely
applied in many practical problems [34, 39, 19, 38, 6, 36, 29, 28, 26, 27]. In
the training stage, SVM solves a quadratic programming problem (QPP),
whereas TWSVM solve two smaller QPPs by traditional solver such as interior method [20, 1, 12]. However, neither SVM nor TWSVM based on these
solvers can deal with the large scale problem, especially millions of samples.
In order to deal with the large scale problem, many improvements were
proposed, e.g., for SVM, sequential minimal optimization, coordinate decent method, trust region Newton, and stochastic gradient descent algorithm (SGD) in [25, 13, 5, 9, 32], and for TWSVM, successive overrelaxation
technique, Newton-Armijo algorithm, and dual coordinate decent method in
[35, 39, 37]. The stochastic gradient descent algorithm for SVM (PEGASOS)
[16, 43, 32, 41] attracts a great attention, because it partitions the large scale
problem into a series of subproblems by stochastic sampling with a suitable
size. It has been proved that PEGASOS is almost sure convergent, and thus
is able to find an approximation of the desired solution with high probability [2, 43, 32]. The existing experiments confirm the effectiveness of these
algorithms with an amazing learning speed.
However, for large scale problem, the stochastic sampling in SGD may
bring some difficulties to SVM due to only a small subset of the dataset is
selected for training. In fact, if the subset is not suitable, PEGASOS would be
weak. It is well known that in SVM the support vectors (SVs), a small subset
of the dataset, decides the final classifier. If the stochastic sampling does not
include the SVs sufficiently, the classifier would lose some generalizations.
Figure 1 is a toy example for PEGASOS. There are two classes in this figure,
where the positive and negative classes respectively include 6 and 4 samples,
and the circle is one of the potential SVs. The solid blue line is the separating
line obtained by PEGASOS with three different sampling: (i) strengthening
the circle sample; (ii) infrequently using the circle sample; (iii) ignoring the
2
5
4
5
Class +1
Class -1
w T x=0
4
w T x=+1
w T x=-1
5
Class +1
Class -1
w T x=0
4
w T x=+1
w T x=-1
3
3
2
2
2
1
1
1
0
-4
0
-4
-2
-1
0
1
2
3
4
-3
w T x=0
w T x=+1
w T x=-1
3
-3
Class +1
Class -1
-2
-1
(a)
0
1
2
3
4
0
-4
-3
-2
-1
(b)
0
1
2
3
4
(c)
Figure 1: PEGASOS on 10 samples from two classes. (i) Training includes all of the 10
samples with 11 iterations, and the circle sample is used twice; (ii) Training includes all
of the 10 samples with 28 iterations, and the circle sample is used once; (iii) Training
includes 9 samples with 27 iterations, where the circle sample is excluded.
5
5
Class +1
Class -1
f(x)=0
f 1 (x)=0
4
5
Class +1
Class -1
f(x)=0
f 1 (x)=0
4
f 2 (x)=0
f 2 (x)=0
f 2 (x)=0
3
3
3
2
2
2
1
1
1
0
-4
0
-4
-3
-2
-1
0
(a)
1
2
3
Class +1
Class -1
f(x)=0
f 1 (x)=0
4
4
-3
-2
-1
0
(b)
1
2
3
4
0
-4
-3
-2
-1
0
1
2
3
4
(c)
Figure 2: SGTSVM on 10 samples from two classes. (i) Training includes all of the 10
samples with 7 iterations, and the circle sample is used twice; (ii) Training includes all
of the 10 samples with 16 iterations, and the circle sample is used once; (iii) Training
includes 9 samples with 15 iterations, where the circle sample is excluded.
3
circle sample. Figure 1 shows that the circle sample plays an important role
on the separating line, and infrequently using or ignoring this sample would
lead to misclassify.
Compared with SVM, it is significant that TWSVM is more stable for
sampling and does not strongly depend on some special samples such as the
SVs [12, 35], which indicates SGD is more suitable for TWSVM. Therefore,
in this paper, we propose a stochastic gradient twin support vector machines
(SGTSVM). Different from PEGASOS, our method selects two samples from
different classes randomly in each iteration to construct a pair of nonparallel
hyperplanes. Due to TWSVM fits all of the training samples, our method is
stable for the stochastic sampling and thus gains well generalizations. Moreover, the characteristics inherited from TWSVM result in that our SGTSVM
suits for many cases, e.g., “cross planes” dataset [21] and preferential classification [12]. As the above toy example, Figure 2 shows the corresponding
results by SGTSVM. Comparing Figure 2 with Figure 1, it is clear that
SGTSVM performs better than PEGASOS.
The main contributions of this paper includes:
(i) a SGD-based TWSVM (SGTSVM) is proposed, and it is very easy to be
extended to other TWSVM-type classifiers;
(ii) we prove that the proposed SGTSVM is convergent, instead of almost
sure convergence in PEGASOS;
(iii) for the uniformly sampling, it is proved that the original objective of
the solution to SGTSVM is bounded by the optimum of TWSVM, which
indicates the solution to SGTSVM is an approximation of the optimal solution to TWSVM, while PEGASOS only has an opportunity to obtain an
approximation of the optimal solution to SVM (more information please see
Corollaries 1 and 2 in [32]);
(iv) the nonlinear case of SGTSVM is obtained directly based on its original
problem;
(v) each iteration of SGTSVM includes no more than 8n + 4 multiplications without additional storage, so it is the fastest one than other proposed
TWSVM-type classifiers.
The rest of this paper is organized as follow. Section 2 briefly reviews
SVM, PEGASOS, and TWSVM. Our linear and nonlinear SGTSVMs together with the theoretical analysis are elaborated in Section 3. Experiments
are arranged in Section 4. Finally, we give the conclusions.
4
2. Related Works
Consider a binary classification problem in the n-dimensional real space
Rn . The set of training samples is represented by X ∈ Rn×m , where x ∈ Rn is
the sample with the label y ∈ {+1, −1}. We further organize the m1 samples
of Class +1 into a matrix X1 ∈ Rn×m1 and the m2 samples of Class −1 into
a matrix X2 ∈ Rn×m2 . Below, we give a brief outlines of some related works.
2.1. SVM
Support vector machines (SVM) [7, 3] searches for a separating hyperplane
w ⊤ x + b = 0,
(1)
where w ∈ Rn and b ∈ R. By introducing the regularization term, the primal
problem of SVM can be expressed as a QPP as follow
1
||w||2
2
min
w,b
+
c ⊤
e ξ
m
D(X ⊤ w + b) ≥ e − ξ, ξ ≥ 0,
s.t.
(2)
where || · || denotes the L2 norm, c > 0 is a parameter with some quantitative
meanings [3], e is a vector of ones with an appropriate dimension, ξ ∈ Rm is
the slack vector, and D = diag(y1 , . . . , ym ). Note that the minimization of
the regularization term kwk2 is equivalent to maximize the margin between
two parallel supporting hyperplanes w ⊤ x + b = ±1. And the structural risk
minimization principle is implemented in this problem [7].
2.2. PEGASOS
PEGASOS [43, 32] considers a strongly convex problem by modifying (2)
as follow
min 12 ||w||2 + mc e⊤ ξ
w
(3)
s.t. DX ⊤ w ≥ e − ξ, ξ ≥ 0,
and recasts the above problem to
min
w
1
||w||2
2
+
c ⊤
e (e
m
− DX ⊤ w)+ ,
where (·)+ replaces negative components of a vector by zeros.
5
(4)
In the tth iteration (t ≥ 1), PEGASOS constructs a temporary function,
which is defined by a random sample xt ∈ X as
gt (w) = 21 ||w||2 + c(1 − yt w ⊤ xt )+ .
(5)
Then, starting with an initial w1 , PEGASOS iteratively updates wt+1 =
wt − ηt ∇wt gt (w) for t ≥ 1, where ηt = 1/t is the step size and ∇wt gt (w) is
the sub-gradient of gt (w) at wt ,
∇wt gt (w) = wt − cyt xt sign(1 − yt wt⊤ xt )+ .
(6)
When some terminate conditions are satisfied, the last wt is outputted as w.
And a new sample x can be predicted by
y = sign(w ⊤ x).
(7)
It has been proved that the average solution w̄ =
1
T
T
P
wt is bounded
t=1
by the optimal solution w ∗ to (4) with o(1), and thus PEGASOS has with
a probability of at least 1/2 to find a good approximation of w ∗ [32]. The
authors of [32] also pointed out that wT is often used instead of w̄ in practice.
The sample xt which is selected randomly can be replaced with a small subset
belonging to the whole dataset, and the subset only including a sample is
often used in practice [43, 32, 41]. In order to extend the generalization ability
of PEGASOS, the bias term b in SVM can be appended to PEGASOS by
replacing g(wt ) of (5) with
g(wt , b) = 12 ||wt ||2 + C(1 − yt (wt⊤ xt + b))+ .
(8)
However, this modification would lead to the function not to be strongly
convex and thus yield a slow convergence rate [32].
2.3. TWSVM
TWSVM [12, 35] seeks a pair of nonparallel hyperplanes in Rn which can
be expressed as
w1⊤ x + b1 = 0 and w2⊤ x + b2 = 0,
(9)
such that each hyperplane is close to samples of one class and has a certain
distance from the other class. To find the pair of nonparallel hyperplanes, it
is required to get the solutions to the primal problems
w1 ,b1
min
1
(||w1 ||2
2
s.t.
X2⊤ w1 + b1 − ξ1 ≤ −e, ξ1 ≥ 0,
+ b21 ) +
c1
kX1⊤ w1
2m1
6
+ b1 k2 +
c2 ⊤
e ξ1
m2
(10)
and
min
w2 ,b2
s.t.
c3
1
(||w2 ||2 + b22 ) + 2m
kX2⊤ w2 + b2 k2
2
2
X1⊤ w2 + b2 + ξ2 ≥ e, ξ2 ≥ 0,
+
c4 ⊤
e ξ2
m1
(11)
where c1 , c2 , c3 , and c4 are positive parameters, ξ1 ∈ Rm2 and ξ2 ∈ Rm1 are
slack vectors. Their geometric meaning is clear. For example, for (10), its
objective function makes the samples of Class +1 proximal to the hyperplane
w1⊤ x + b1 = 0 together with the regularization term, while the constraints
make each sample of Class −1 has a distance more than 1/||w1|| away from
the hyperplane w1⊤ x + b1 = −1.
Once the solutions (w1 , b1 ) and (w2 , b2 ) to the problems (10) and (11) are
respectively obtained, a new point x ∈ Rn is assigned to which class depends
on the distance to the two hyperplanes in (9), i.e.,
y = arg min
i
|wi⊤ x+bi |
,
kwi k
(12)
where | · | is the absolute value.
3. SGTSVM
In this section, we elaborate our SGTSVM and give its convergence analysis together with the boundedness.
3.1. Linear Formation
Following the notations in Section 2, we recast the QPPs (10) and (11)
in TWSVM to unconstrained problems
min
w1 ,b1
1
(||w1 ||2
2
+ b21 ) +
c1
||X1⊤ w1
2m1
+ b1 ||2 +
c2 ⊤
e (e
m2
+ X2⊤ w1 + b1 )+ ,
(13)
1
(||w2||2
2
+ b22 ) +
c3
||X2⊤ w2
2m2
+ b2 ||2 +
c4 ⊤
e (e
m1
− X1⊤ w2 − b2 )+ ,
(14)
and
min
w2 ,b2
respectively.
In order to solve the above two problems, we construct a series of strictly
convex functions f1,t (w1 , b1 ) and f2,t (w2 , b2 ) with t ≥ 1 as
f1,t = 21 (||w1||2 + b21 ) +
c1
||w1⊤ xt
2
+ b1 ||2 + c2 (1 + w1⊤ x̂t + b1 )+ ,
7
(15)
and
f2,t = 21 (||w2 ||2 + b22 ) +
c3
||w2⊤ x̂t
2
+ b2 ||2 + c4 (1 − w2⊤ xt − b2 )+ ,
(16)
where xt and x̂t are selected randomly from X1 and X2 , respectively.
The sub-gradients of the above functions at (w1,t , b1,t ) and (w2,t , b2,t ) can
be obtained as
⊤
⊤
∇w1,t f1,t = w1,t + c1 (w1,t
xt + b1,t )xt + c2 x̂t sign(1 + w1,t
x̂t + b1,t )+ ,
⊤
⊤
∇b1,t f1,t = b1,t + c1 (w1,t xt + b1,t ) + c2 sign(1 + w1,t x̂t + b1,t )+ ,
(17)
and
⊤
⊤
xt − b2,t )+ ,
x̂t + b2,t )x̂t − c4 xt sign(1 − w2,t
∇w2,t f2,t = w2,t + c3 (w2,t
⊤
⊤
∇b2,t f2,t = b2,t + c3 (w2,t x̂t + b2,t ) − c4 sign(1 − w2,t xt − b1,t )+ ,
(18)
respectively.
Our SGTSVM starts from the initial (w1,1 , b1,1 ) and (w2,t , b2,t ). Then, for
t ≥ 1, the updates are given by
w1,t+1 = w1,t − ηt ∇w1,t f1,t ,
b1,t+1 = b1,t − ηt ∇b1,t f1,t ,
w2,t+1 = w2,t − ηt ∇w2,t f2,t ,
b2,t+1 = b2,t − ηt ∇b2,t f2,t ,
(19)
where ηt is the step size and typically is set to 1/t. If the terminated condition
is satisfied, (w1,t , b1,t ) is assigned to (w1 , b1 ), and (w2,t , b2,t ) is assigned to
(w2 , b2 ). Then, a new sample x ∈ Rn can be predicted by (12).
The above procedures are summarized in Algorithm 1.
3.2. Nonlinear Formation
Now, we extend our SGTSVM to nonlinear case by the kernel trick [21,
12, 35, 31, 15, 18]. Suppose K(·, ·) is the predefined kernel function, then the
nonparallel hyperplanes can be expressed as
K(x, X)⊤ w1 + b1 = 0 and K(x, X)⊤ w2 + b2 = 0.
(20)
The counterparts of (13) and (14) can be formulated as
min
w1 ,b1
1
(||w1 ||2
2
+ b21 ) +
c1
||K(X1 , X)⊤ w1
2m1
+ b1 ||2 +
c2 ⊤
e (e
m2
+ K(X2 , X)⊤ w1 + b1 )+ ,
(21)
8
Algorithm 1 SGTSVM Framework.
Require:
Given the training dataset X1 ∈ Rn×m1 as positive class, X2 ∈ Rn×m2 as
negative class, select parameters c1 , c2 , c3 , c4 , and a small tolerance tol,
typically tol = 1e − 3.
Ensure:
w1 , b1 , w2 , b2 .
1: set w1,1 , b1,1 , w2,1 , and b2,1 be zeros;
For t = 1, 2, . . .
2: Choose a pair of samples xt and x̂t from X1 and X2 at random, respectively;
3: Compute the tth gradients by (17) and (18);
4: Update w1,t+1 , b1,t+1 , w2,t+1 , and b2,t+1 by (19);
5: If ||w1,t+1 − w1,t || + |b1,t+1 − b1,t | < tol, stop updating w1,t+1 and b1,t+1 ,
and set w1 = w1,t+1 , b1 = b1,t+1 ;
6: If ||w2,t+1 − w2,t || + |b2,t+1 − b2,t | < tol, stop updating w2,t+1 and b2,t+1 ,
and set w2 = w2,t+1 , b2 = b2,t+1 ;
and
min
w2 ,b2
1
(||w2 ||2
2
+ b22 ) +
c3
||K(X2 , X)⊤ w2
2m2
+ b2 ||2 +
c4 ⊤
e (e
m1
− K(X1 , X)⊤ w2 − b2 )+ .
(22)
Then, we construct a series of functions with t ≥ 1 as
h1,t = 21 (||w1 ||2 + b21 ) +
c1
||K(xt , X)⊤ w1
2
+ b1 ||2 + c2 (1 + K(x̂t , X)⊤ w1 + b1 )+ ,
(23)
and
h2,t = 21 (||w2 ||2 + b22 ) +
c3
||K(x̂t , X)⊤ w2
2
+ b2 ||2 + c4 (1 − K(xt , X)⊤ w2 − b2 )+ .
(24)
Similar to (17), (18), and (19), the sub-gradients and updates can be
obtained. The details are omitted.
For large scale problem, it is time consuming to calculate the kernel
K(·, X). However, the reduced kernel strategy, which has been successfully applied for SVM and TWSVM [18, 40, 39], can also be applied for our
SGTSVM. The reduced kernel strategy replaces K(·, X) with K(·, X̃), where
X̃ is a random sampled subset of X. In practice, X̃ just needs 0.01% ∼ 1%
samples from X to get a well performance, reducing the learning time without
loss of generalization [40].
9
3.3. Analysis
In this subsection, we discuss two issues: (i) the convergence of the solution in SGTSVM; (ii) the relation between the solution in SGTSVM and
the optimal one in TWSVM. For convenience, we just consider the first QPP
(13) of linear TWSVM together with the SGD formation of linear SGTSVM.
The conclusions on another QPP (14) and the nonlinear formations can be
obtained easily as the first one.
Let u = (w ⊤ , b)⊤ , Z1 = (X1⊤ , e)⊤ , Z2 = (X2⊤ , e)⊤ , z = (x⊤ , 1)⊤ , and the
notations with the subscripts in SGTSVM also comply with this definition.
Then, the first QPP (13) is reformulated as
min f (u) = 21 ||u||2 +
u
c1
||Z1u||2
2m1
+
c2 ⊤
e (e
m2
+ Z2 u)+ .
(25)
Next, we reformulate the tth (t ≥ 1) function in SGTSVM as
ft (u) = 21 ||u||2 +
c1
||u⊤ zt ||2
2
+ c2 (1 + u⊤ ẑt )+ ,
(26)
where zt and ẑt are the samples selected randomly from Z1 and Z2 for the
tth iteration, respectively. The sub-gradient of ft (u) at ut is denoted as
∇t = ut + c1 (u⊤ zt )zt + c2 ẑt sign(1 + u⊤ ẑt )+ .
(27)
Given u1 and the step size ηt = 1/t, ut+1 with t ≥ 1 is updated by
ut+1 = ut − ηt ∇t ,
(28)
i.e.,
ut+1 = (1 − 1t )ut −
c1
z z⊤ u
t t t t
−
c2
ẑ sign(1
t t
+ u⊤
t ẑt )+ .
(29)
Lemma 3.1. For all t ≥ 1, ||∇t || and ||ut || have the upper bounds.
Proof. The formation (29) can be rewritten as
ut+1 = At ut + 1t vt ,
(30)
where At = 1t ((t−1)I−c1 zt zt⊤ ), I is the identity matrix, and vt = −c2 ẑt sign(1+
u⊤
t ẑt )+ . Note that for sufficient t, there is a positive integer N such that for
t > N, At is positive definite, and the largest eigenvalue λt of At is smaller
than or equal to t−1
. Based on (30), we have
t
ut+1 =
t
Q
At+N +1−i uN +1 +
i=N +1
t
P
i=N +1
10
1
(
i
t
Q
j=i+1
At+i+1−j )vi .
(31)
For i ≥ N + 1, ||At+N +1−i uN +1 || ≤ λi ||uN +1|| ≤
t
Q
||
|| 1i (
t
Q
j=i+1
[10]. Therefore,
N
||uN +1||,
t
(32)
At+i+1−j )vi || ≤ 1t max ||vi ||.
(33)
At+N +1−i uN +1|| ≤
i=N +1
and
i−1
||uN +1||
i
i≤t
Thus, we have
||ut+1 || ≤
N
||uN +1 ||
t
+
t−N
t
max ||vi ||
i≤t
≤ ||uN +1 || + c2 max ||z||.
(34)
z∈Z2
Let M be the largest norm of the samples in the dataset and
G1 = max{max{||u1||, . . . , ||uN ||}, ||uN +1|| + c2 M}.
(35)
This leads to that G1 is an upper bound of ||ut ||, and G2 = G1 +c1 G1 M 2 +c2 M
is an upper bound of ||∇t ||, for t ≥ 1.
Theorem 3.1. The iterative formation (29) of our SGTSVM is convergent.
Proof. On the one hand, from (32) in the proof of Lemma 3.1, we have
t
Q
lim ||
t→∞
At+N +1−i uN +1 || = 0,
(36)
At+N +1−i uN +1 = 0.
(37)
i=N +1
which indicates
t
Q
lim
t→∞ i=N +1
On the other hand, from (33), we have
t
P
i=N +1
|| 1i (
t
Q
At+i+1−j )vi || ≤ M,
(38)
j=i+1
which indicates that the following limit exists
lim
t
P
t→∞ i=N +1
|| 1i (
t
Q
At+i+1−j )vi || < ∞.
j=i+1
11
(39)
Note that an infinite series of vectors is convergent if its norm series is convergent [30]. Therefore, the following limit exists
t
P
lim
t→∞ i=N +1
1
(
i
t
Q
At+i+1−j )vi < ∞.
(40)
j=i+1
Combine (37) with (40), we conclude that the series wt+1 is convergent
for t → ∞.
Based on the above theorem, it is reasonable to take the terminate condition to be ||ut+1 − ut || < tol. Moreover, if we reform (31) by u1 , then
ut+1 =
t
Q
At+1−i u1 +
i=1
t
P
i=1
1
(
i
t
Q
At+i+1−j )vi .
(41)
j=i+1
In order to keep ut+1 to be convergent fast, it is suggested to set u1 = 0.
In the following, we analyse the relation between the solution ut in SGTSVM
and the optimal solution u∗ = (w ∗⊤ , b∗ )⊤ in TWSVM.
Lemma 3.2. Let f1 , . . . , fT be a sequence of convex functions, and u1 , . . . , uT +1 ∈
Rn be a sequence of vectors. For t ≥ 1, ut+1 = ut − ηt ∇t , where ∇t belongs
to the sub-gradient set of ft at ut and ηt = 1/t. Suppose ||ut|| and ||∇t || have
the upper bounds G1 and G2 , respectively. Then, for all θ ∈ Rn , we have
T
T
P
P
1
ft (ut ) ≤ T1
ft (θ) + G2 (G1 + ||θ||) + 2T
G22 (1 + ln T );
(i) T1
t=1
t=1
(ii) for sufficiently large T , given any ε > 0, then
1
T
T
P
ft (ut ) ≤
t=1
1
T
T
P
ft (θ)+ε.
t=1
Proof. Since ft is convex and ∇t is the sub-gradient of ft at ut , we have that
ft (ut ) − ft (θ) ≤ (ut − θ)⊤ ∇t .
(42)
Note that
(ut − θ)⊤ ∇t =
1
(||ut
2ηt
− θ||2 − ||ut+1 − θ||2) +
12
ηt
||∇t ||2 .
2
(43)
Combine (42) and (43), we have
T
P
(ft (ut ) − ft (θ))
t=1
T
T
P
P
1
1
1
2
2
(ηt ||∇t ||2 )
(||u
−
θ||
−
||u
−
θ||
)
+
t
t+1
2
ηt
2
t=1
t=1
P
P
1
( Tt=1 ||ut − θ||2 − T ||uT +1 − θ||2 ) + 12 Tt=1 (ηt ||∇t ||2 )
2
T
P
(G1 + ||θ||) ||uT +1 − ut || + 12 G22 (1 + ln T )
t=1
T
T
P
P
(G1 + ||θ||) || 1i ∇i || + 12 G22 (1 + ln T )
t=1 i=t
T G2 (G1 + ||θ||) + 12 G22 (1 + ln T ).
≤
=
≤
=
≤
(44)
Multiplying (44) by 1/T leads to the conclusion (i).
On the other hand, suppose lim uT = ũ, we have lim ||uT || = ||ũ||.
Then,
T
P
lim T1
||ut −θ||
T →∞
t=1
T →∞
T →∞
G22 (1+lnT )
T
T →∞
= lim ||uT −θ|| = ||ũ−θ||. Note that lim
T →∞
=
0. Given any ε > 0, for sufficiently large T ,
1
T
T
P
(ft (ut ) − ft (θ))
t=1
≤
1 1
(
2 T
≤
1
ε
2
T
P
||ut − θ||2 − ||uT +1 − θ||2) +
t=1
1
2T
G22 (1 + lnT )
(45)
+ 12 ε = ε.
We are now ready to bound the average instantaneous objective (26).
Theorem 3.2. For ft (t = 1, . . . , T ) defined as (26) in SGTSVM, ut (t =
1, . . . , T ) is constructed by (29), and u∗ is the optimal solution to (25). Then,
(i) there are two constants G1 and G2 (actually, they are the upper bounds of
T
T
P
P
||wt || and ||∇t ||, respectively) such that T1
ft (ut ) ≤ T1
ft (u∗ ) + G2 (G1 +
||u∗||) +
1
2T
t=1
G22 (1 + ln T );
(ii) for sufficiently large T , given any ε > 0, then
t=1
1
T
T
P
t=1
ft (ut ) ≤
1
T
T
P
ft (u∗ )+ε.
t=1
Proof. Obviously, ft (t = 1, . . . , T ) is convex. Let G1 and G2 respectively be
the upper bounds of ||ut || and ||∇t ||, the conclusions come from Lemmas 3.1
and 3.2.
13
In the following, let us discuss the relation between the solutions to
SGTSVM and TWSVM with the uniform sampling.
Corollary 3.1. Assume the conditions stated in Theorem 3.1 and m1 = m2 ,
where m1 and m2 are the sample number of X1 and X2 , respectively. Suppose
T = km1 , where k > 0 is an integer, and each sample is selected k times at
random. Then
1
(i) f (uT ) ≤ f (u∗ ) + G2 (G1 + ||u∗|| + G2 ) + 2T
G21 (1 + ln T );
(ii) for sufficiently large T , given any ε > 0, then f (uT ) ≤ f (u∗ ) + G22 + ε.
Proof. First, we prove that for all i, j = 1, 2, . . . , T ,
|ft (ui ) − ft (uj )| ≤ G2 ||ui − uj ||, t = 1, 2, . . . , T.
(46)
From the formation of ft (u), we have
|ft (ui ) − ft (uj )| ≤ 21 |||ui||2 − ||uj ||2 |
2
⊤
2
+ c21 |(u⊤
i zt ) − (uj zt ) |
⊤
+c2 |(1 + u⊤
i ẑt )+ − (1 + uj ẑt )+ |.
(47)
Since G1 is the upper bound of ||ut|| (t ≥ 1) and M is the largest norm of
the samples in the dataset, the first part, the second part, and the third part
on the right hand of (47) are respectively
1
|||ui||2
2
=
≤
and
− ||uj ||2 | ≤ G1 ||ui − uj ||,
c1
2
⊤
2
|(u⊤
i zt ) − (uj zt ) |
2
c1
|(ui + uj )⊤ zt (ui − uj )⊤ zt |
2
c1 G1 M 2 ||ui − uj ||,
⊤
c2 |(1 + u⊤
i ẑt )+ − (1 + uj ẑt )+ |
= c2 |(ui − uj )⊤ ẑt |
≤ c2 M||ui − uj ||.
(48)
(49)
(50)
Therefore, there is a constant G2 = G1 + c1 G1 M 2 + c2 M satisfying (46).
From ut+1 = ut − 1t ∇t , it is easy to obtain
ut+1 = u1 −
t
P
i=1
1
∇,
i t
14
t = 1, 2, . . . , T.
(51)
Thus, for 1 ≤ i < j ≤ T ,
||ui − uj || = ||
j−1
P
t=i
1
∇ ||
t t
Since T = km1 = km2 , for all u ∈ Rn ,
1
T
≤
j−1
P
t=i
T
P
1
G .
t 2
(52)
ft (u) = f (u). Note that f (u)
t=1
is the objective of TWSVM. Based on (46) and (52), we have
f (uT ) −
=
≤
1
T
1
T
T
P
1
T
T
P
ft (ut )
t=1
(ft (uT ) − ft (ut ))
t=1
T
P
(53)
G2 ||uT − ut ||
t=1
G2 (T −1)
≤ 2T
≤ G22 .
Using the Theorem 3.1, we have the conclusion immediately.
If m1 6= m2 , we can modify the sampling rule to obtain the same result
as one in Corollary 3.1.
Corollary 3.2. Assume the conditions stated in Corollary 3.1, but m1 6= m2 .
Suppose T = kd(m1 , m2 ), where k > 0 is an integer and d is the least common
multiple of m1 and m2 . The sample in X1 is selected kd/m1 times at random,
and the one in X2 is kd/m2 times at random. Then
1
(i) f (uT ) ≤ f (u∗ ) + G2 (G1 + ||u∗|| + G2 ) + 2T
G21 (1 + ln T );
(ii) for sufficiently large T , given any ε > 0, then f (uT ) ≤ f (u∗ ) + G22 + ε.
Note that for all u ∈ Rn ,
1
T
T
P
ft (u) = f (u). The proof of the above
t=1
corollary is the same as Corollary 3.1.
The above corollaries provide the approximations of u∗ by uT . If the
sampling rule is not as stated in these corollaries, these upper bounds no
longer holds. However, Kakade and Tewari [14] have shown a way to obtain
a similar bounds with high probability.
15
2
2
1.5
1.5
1
1
0.5
0.5
0
0
-0.5
-0.5
-1
-1
-1.5
-1.5
-2
-1
-0.8
-0.6
-0.4
-0.2
0
0.2
0.4
0.6
0.8
-2
-1
1
(a) TWSVM
-0.8
-0.6
-0.4
-0.2
0
0.2
0.4
0.6
0.8
1
(b) SGTSVM
Figure 3: Results of TWSVM and SGTSVM on the “cross planes”, where the black solid
lines are w1⊤ x + b1 = 0 and w2⊤ x + b2 = 0, respectively.
4. Experiments
In the experiments, we compared our SGTSVM with SVM [7], PEGASOS
[32], and TWSVM [12, 35] on several artificial and large scale problems. All
of the methods were implemented on a PC with an Intel Core Duo processor
(3.4 GHz) with 4 GB RAM.
4.1. Artificial datasets
On the artificial datasets, PEGASOS, TWSVM, and our SGTSVM were
implemented by Matlab [22], and the corresponding SGTSVM Matlab codes
were uploaded upon http://www.optimal-group.org/Resource/SGTSVM.html.
First of all, we consider the similarity between TWSVM and SGTSVM.
These two methods were implemented on the “cross planes” dataset, where
TWSVM was superior on this dataset [12]. Figure 3 shows the proximal
lines on the dataset. It is obvious that the two proximal lines by SGTSVM
is similar as the ones by TWSVM, so TWSVM and SGTSVM can precisely
capture the data distribution, and thus both of them obtain the well classifier. To measure the similarity quantitatively, the optimums f1 of (10) and
f2 of (11) in TWSVM were calculated compared with the ones of each iteration in SGTSVM on the “cross planes” and some UCI datasets [4] (e.g.,
dataset Australia which includes 690 samples with 14 features, dataset Creadit which includes 690 samples with 15 features, and dataset Hypothyroid
which includes 3, 163 samples with 25 features). Linear TWSVM, SGTSVM,
16
0.12
0.1
f 1 of TBSVM
0.11
f 1 of SGTSVM
0.095
f 2 of TBSVM
f 2 of SGTSVM
Objective
Objective
f 1 of TBSVM
f 1 of SGTSVM
0.1
0.09
f 2 of TBSVM
f 2 of SGTSVM
0.09
0.085
0.08
0.08
5
10
15
20
25
20
40
60
Iteration
80
100
120
140
160
180
200
Iteration
(a) Cross planes
(b) Australia
0.14
0.12
f 1 of SGTSVM
0.11
f 2 of TBSVM
Objective
Objective
0.12
f 1 of TBSVM
f 2 of SGTSVM
0.1
f 1 of TBSVM
f 1 of SGTSVM
f 2 of TBSVM
f 2 of SGTSVM
0.1
0.09
0.08
0.07
0.08
0.06
0.06
0.05
20
40
60
80
100
120
140
160
180
Iteration
20
40
60
80
100
120
140
Iteration
(c) Creadit
(d) Hypothyroid
Figure 4: Results of linear TWSVM and SGTSVM on the four datasets, where the vertical
axis denotes the objectives of f1 and f2 .
and their nonlinear versions were implemented, where the Gaussian kernel
K(x, y) = exp{−µ||x−y||2} was used for nonlinear versions. The parameters
c1 , c2 , c3 , c4 , and µ are fixed to 0.1. Figure 4 shows the results from the two
linear classifiers, and Figure 5 corresponds to the nonlinear case. In Figures
4 and 5, the horizontal axis denotes the iteration of SGTSVM and the vertical axis denotes the objectives f1 and f2 of TWSVM and SGTSVM. Due
to the objectives of TWSVM are constant, they are denoted by two horizontal lines, while the objectives of SGTSVM for each iteration are denoted by
two broken lines in these figures. For different datasets, it can be seen that
our SGTSVM converges to TWSVM after different iterations. For instance,
linear SGTSVM converges to TWSVM after 20 iterations in Figure 4 (a),
whereas the same thing appears in Figure 4 (b) after 180 iterations. Generally, SGTSVM converges to TWSVM after 150 iterations on these datasets
either for linear or nonlinear case. Furthermore, the 10-fold cross validation
[8] was used on these datasets. We ran TWSVM and SGTSVM 10 times,
17
0.12
0.2
f 1 of TBSVM
0.1
f 1 of SGTSVM
0.15
f 2 of TBSVM
f 2 of SGTSVM
Objective
Objective
f 1 of TBSVM
f 1 of SGTSVM
0.08
0.06
f 2 of TBSVM
f 2 of SGTSVM
0.1
0.05
0.04
0
20
40
60
80
100
120
10
20
30
40
Iteration
(a) Cross planes
60
70
80
90
100
(b) Australia
0.1
0.1
f 1 of TBSVM
0.099
f 2 of TBSVM
f 1 of SGTSVM
f 2 of TBSVM
0.098
Objective
f 2 of SGTSVM
0.097
f 1 of TBSVM
0.099
f 1 of SGTSVM
0.098
Objective
50
Iteration
0.096
0.096
0.095
0.095
0.094
0.094
0.093
f 2 of SGTSVM
0.097
0.093
10
20
30
40
50
60
70
80
90
100
Iteration
20
40
60
80
100
120
140
160
Iteration
(c) Creadit
(d) Hypothyroid
Figure 5: Results of nonlinear TWSVM and SGTSVM on the four datasets, where the
vertical is the same as the one in Fig. 4.
and reported the mean accuracy and standard deviation on Table 1. The
differences of the mean accuracies are no more than 2%, which implies the
classifiers obtained by TWSVM and SGTSVM do not have significant difference.
Secondly, we test the stability of SGTSVM compared with PEGASOS.
100 datasets were generated randomly, and each dataset contain 10, 000
samples in R, where 5, 000 negative samples are from normal distribution
N(−2, 1) and 5, 000 positive ones are from N(2, 1). The best classification
point is at zero. We implemented PEGASOS and SGTSVM without any
restrictions on the 100 datasets and obtained 100 classifiers shown in Figure 6, where the upper right digit is the mean of these lines together with
their standard deviation (the parameters c in PEGASOS, c1 , c2 , c3 , and c4 in
SGTSVM were fixed to 0.1). It is clear that our SGTSVM obtains much more
compact classification lines than PEGASOS. The mean line of SGTSVM is
at −0.0016 which is closer to zero and its standard deviation is smaller than
18
Table 1: Mean accuracy (%) with standard deviation of TWSVM and SGTSVM by 10-fold
cross validation.
Data
Cross Planes
Australia
Creadit
Hypothyroid
TWSVM†
96.05±0.70
86.87±0.38
85.78±0.32
98.21±0.09
†
SGTSVM†
97.71±0.41
87.34±0.13
85.72±0.23
97.28±0.01
TWSVM♯
99.01±2.24
87.10±0.43
86.71±0.33
98.08±0.09
SGTSVM♯
98.51±2.15
85.21±0.16
85.21±0.45
98.07±0.03
linear case;♯ nonlinear case.
0.4
0.4
-0.0236±0.17
-0.0016±0.05
0.3
0.3
0.2
0.2
0.1
0.1
0
-6
-4
-2
0
2
4
6
(a) PEGASOS
0
-6
-4
-2
0
2
4
6
(b) SGTSVM
Figure 6: Results of PEGASOS and SGTSVM on 100 artificial datasets, where the 100
upright black solid lines are final classifiers.
19
0.4
0.4
-0.2038±0.24
-0.0537±0.07
0.3
0.3
0.2
0.2
0.1
0.1
0
-6
-4
-2
0
2
4
6
(a) PEGASOS
0
-6
-4
-2
0
2
4
6
(b) SGTSVM
Figure 7: Results of PEGASOS and SGTSVM on 100 artificial datasets, where the 100
upright black solid lines are final classifiers, and the samples on the dash line is invisible
for sampling.
PEGASOS. In order to investigate the effect of sampling, PEGASOS and
SGTSVM were implemented on the above 100 datasets with the restricted
sampling (i.e., some possible support vectors from negative samples in SVM
and the samples close to these support vectors are invisible for sampling).
Figure 7 shows the results of PEGASOS and SGTSVM, where the dash line
denotes that the samples in this scope are invisible for sampling. From Figure 7, it can be seen that the classification lines by PEGASOS fall into two
regions, while SGTSVM obtains a compact region. Thus, it means that the
possible support vectors significantly influence PEGASOS, while SGTSVM
relatively relies on the data distribution. From Figures 6 and 7, PEGASOS
always acquires a mean classification line further from zero with a larger
standard deviation than SGTSVM. Therefore, SGTSVM is more stable than
PEGASOS on these datasets with or without restricted sampling. To further show the classifiers’ stability, we recorded the classification accuracies
(%) of PEGASOS and SGTSVM on one of the 100 datasets. PEGASOS
and SGTSVM were implemented 100 times on this dataset, where the parameters were set as before and two methods were iterated 200 times. Every
accuracies of these methods are reported in Figure 8. From Figure 8, the
accuracies of SGTSVM belong to [99.0, 99.5] while PEGASOS is [96.5, 99.5],
which indicates SGTSVM is more stable than PEGASOS from the aspect
of classification result. Although PEGASOS obtains the highest accuracy in
this test, SGTSVM obtains higher accuracies than PEGASOS in most cases.
Finally, we test the convergence of PEGASOS and SGTSVM. A dataset
20
100
PEGASOS
Accuracy (%)
99.5
SGTSVM
99
98.5
98
97.5
97
96.5
10
20
30
40
50
60
70
80
90
100
Times
Figure 8: Accuracies of PEGASOS and SGTSVM on a normal distribution dataset, where
each method is implemented 100 times.
0.2
mean
0
-0.2
10
100
200
300
400
500
600
700
800
900
1000
PEGASOS
0.2
mean
0
-0.2
10
100
200
300
400
500
600
700
800
900
1000
SGTSVM
Figure 9: Results of PEGASOS and SGTSVM on a normal distribution dataset, where
each method is implemented 10 times. The horizontal axis is the iteration and the vertical
one is the classification location.
21
0
6
10
10
4
Iteration
10
PEGASOS (mean)
SGTSVM (mean)
PEGASOS (std)
SGTSVM (std)
Time (Sec.)
PEGASOS (mean)
SGTSVM (mean)
PEGASOS (std)
SGTSVM (std)
102
100
10-2
-1
10
-2
10
-3
10
10
-4
-5
10
10-5
-6
10
10
tol
-1
10
-2
-3
-4
10
10
10
-5
10
-6
tol
Figure 10: Iteration and time of PEGASOS and SGTSVM on a normal distribution
dataset, where each method is implemented 100 times.
contains 20, 000 samples in R was generated randomly, where 10, 000 negative
samples are from normal distribution N(−2, 1) and 10, 000 positive ones are
from N(2, 1). PEGASOS and SGTSVM were implemented 10 times and
each method was iterated 1, 000 times. The current classification locations
for different iterations were reported in Figure 9, where the horizontal axis is
the iteration and the vertical one is the classification location. From Figure
9, it can be seen that: (i) the initial selected samples do not very affect
both PEGASOS and SGTSVM after iterating 150 times; (ii) after iterating
100 times, the classification locations of two methods are centralized to zero
and the error is less than 0.1; (iii) it is important that PEGASOS gets higher
error after iterating 800 times than SGTSVM, indicates PEGASOS converges
slower than SGTSVM. To more precisely discuss the convergence, PEGASOS
and SGTSVM were implemented 100 times and each method was terminated
by the solution error parameter tol (more details about tol can be found
in Algorithm 1). tol is selected from {10i |i = −1, −2, . . . , −6}, and the
corresponding iteration and spent time are reported in Figure 10. It is clear
from Figure 10 that our SGTSVM converges faster than PEGASOS when
tol ≤ 10−3 . Moreover, if one needs smaller solution error such as tol = 10−4
or tol = 10−5 , the iterations of PEGASOS would be about 10 times more than
SGTSVM, and it would be 100 times when tol = 10−6 (thus the learning time
between PEGASOS and SGTSVM is more than a hundredfold). Therefore,
SGTSVM converges much faster than PEGASOS.
4.2. Large scale datasets
To test the feasibility of these methods on large scale datasets, we ran
SVM, PEGASOS, and SGTSVM on six large scale datasets [4]. Table 2 shows
22
Table 2: The details of the large scale datasets.
Data
(a)
(b)
(c)
(d)
(e)
(f)
Name
Skin
Gashome
Susy
Kddcup
Gas
Hepmass
No. of samples
245,057
928,990
5,000,000
4,898,432
8,386,764
10,500,000
Dimension
3
10
18
41
16
28
Ratio
0.262
0.578
0.844
0.248
0.077
1.000
the details of the large scale datasets, where Ratio in Table 2 is the sample
number of positive class than negative one. Each dataset is split into two
subsets where one (including 90% samples) is used for training and the other
(including 10% samples) is used for testing. SVM is implemented by Liblinear [9], while PEGASOS and SGTSVM are implemented by the softwares
written in C language. The corresponding softwares can be downloaded
from http://www.optimal-group.org/Resource/SGTSVM.html. For nonlinear SGTSVM, the reduced kernel [18] is used and the kernel size is fixed
to 100.
First, let us test the influence of parameter tol on PEGASOS and SGTSVM.
These methods were implemented on the large scale datasets, where tol was
respectively set to {10i |i = −1, −2, . . . , −6} and other parameters were fixed
to 0.1. The testing accuracy and learning time are reported in Figure 11.
By comparing Figure 11 (a), (c), and (e), it can bee seen that our SGTSVM
(including linear and nonlinear cases) is more stable than PEGASOS when
tol ≤ 10−4. In order to select a high accuracy with an acceptable learning
time from Figure 11, tol is set to 10−6 for PEGASOS, and it is set to 10−4
for SGTSVM.
Then, we compare SVM and PEGASOS with our SGTSVM with fixed tol
on these datasets. These methods’ accuracies are recorded in Table 3, where
validation accuracy is obtained by 5-fold cross validation on the training subset, and testing accuracy is obtained by the testing subset. The parameters
c in SVM and PEGASOS, c1 , c2 , c3 , and c4 in SGTSVM are selected from
{2i |i = −8, −7, . . . , 1}, and the Gaussian kernel parameter µ in nonlinear
SGTSVM is selected from {2i |i = −10, −9, . . . , −1}. For simplicity, we also
set c1 = c3 and c2 = c4 in SGTSVM. The optimal parameters are recorded
in Table 4. From Table 3, it is obvious that our SGTSVM owns the highest
23
1800
90
1600
80
1400
70
1200
60
1000
Time (Sec.)
Accuracy (%)
100
50
40
800
600
30
400
20
200
10
Skin
0
10-1
Gashome
10-2
Susy
Kddcup
10-3
Gas
10-4
0
Hepmass
10-5
Skin
Gashome
Susy
Kddcup
Gas
Hepmass
10-6
10-1
10-2
10-3
tol
10-4
10-5
10-6
10-5
10-6
10-5
10-6
tol
(a) PEGASOS
(b) PEGASOS
700
100
600
90
500
Skin
Gashome
Susy
Kddcup
Gas
Hepmass
400
Time (Sec.)
Accuracy (%)
80
70
60
300
200
50
100
40
0
Skin
30
10-1
Gashome
10-2
Susy
Kddcup
10-3
Gas
10-4
Hepmass
10-5
10-6
10-1
10-2
10-3
tol
(d) SGTSVM†
100
800
90
700
80
600
70
500
Time (Sec.)
Accuracy (%)
(c) SGTSVM†
60
50
300
200
30
100
20
Skin
Gashome
Susy
Kddcup
Gas
Hepmass
400
40
0
Skin
10
10-1
10-4
tol
10-2
Gashome
Susy
Kddcup
10-3
10-4
Gas
Hepmass
10-5
10-6
tol
10-1
10-2
10-3
10-4
tol
(e) SGTSVM♯
(f) SGTSVM♯
Figure 11: The accuracy and learning time of PEGASOS, linear SGTSVM († ), and nonlinear SGTSVM (♯ ) on six large scale datasets. The dashed box corresponds to the chosen
parameter tol.
24
Table 3: The results on the large scale datasets.
Data
Skin
245,057×3
Gashome
919,438×10
Susy
5,000,000×18
Kddcup
4,898,432×41
Gas
8,386,764×16
Hepmass
10,500,000×28
†
validation(%)
testing(%)
validation(%)
testing(%)
validation(%)
testing(%)
validation(%)
testing(%)
validation(%)
testing(%)
validation(%)
testing(%)
SVM
PEGASOS
SGTSVM†
SGTSVM♯
78.87
84.28
49.11
82.57
78.41
78.52
*
*
*
*
*
*
82.46
85.39
70.09
72.85
54.11
56.44
96.39
96.42
69.77
50.54
80.63
80.84
85.23
87.70
67.50
76.09
76.14
75.09
95.24
97.45
89.73
92.45
80.80
81.10
84.70
85.34
74.49
89.13
69.90
68.61
93.19
99.20
92.60
92.86
82.18
79.59
linear case;♯ nonlinear case; ∗ out of memory.
Table 4: The optimal parameters of SVM, PEGASOS, and SGTSVM.
Data
Skin
Gashome
Susy
Kddcup
Gas
Hepmass
validation
testing
validation
testing
validation
testing
validation
testing
validation
testing
validation
testing
†
SVM
c
2i
PEGASOS
c
2i
SGTSVM†
c1 = c3 , c2 = c4
2i , 2j
SGTSVM♯
c1 = c3 , c2 = c4 , µ
2i , 2j , 2k
-1
-1
0
-1
1
0
NA
NA
NA
NA
NA
NA
-6
-4
-6
-1
0
-7
-6
-2
-1
1
0
0
0,-5
1,-6
-4,-5
-8,-7
-2,-6
-1,-3
-8,-4
-8,-4
-4,0
-3,1
-1,-2
0,-2
-6,-5,-3
-1,0,-9
-3,-5,-2
-8,-1,-2
-3,-1,-4
-3,-3,-3
0,-3,-4
-6,-1,-8
-1,-1,-6
-4,-8,-6
-4,-1,-3
-4,-2,-3
linear case;♯ nonlinear case.
25
linear SGTSVM
nonlinear SGTSVM
PEGASOS
Liblinear
Time (Sec.)
2000
1500
1000
500
0
Figure 12: Learning time of SGTSVM, PEGASOS, and Liblinear on the large scale
datasets with the optimal parameters.
accuracies on 9 groups of comparisons, and performs as well as SVM or PEGASOS on the other 3 groups. However, SVM performs much worse than
SGTSVM on the dataset Gashome and cannot work on three much larger
datasets. Though PEGASOS can work on these datasets, it performs much
worse than SGTSVM on Susy and Gas. To further comparing the learning
time of these methods, we report the one-run time in Figure 12 with the optimal parameters. It is obvious that SGTSVM (including linear and nonlinear
cases) is much faster than the others. Thus, our SGTSVM is comparable to
SVM and PEGASOS on these large scale datasets. In addition, the softwares
of SGTSVM and PEGASOS need much less RAM than Liblinear (the software of SVM). In detail, Liblinear needs store the entire training set in RAM,
while PEGASOS and SGTSVM only store a subset related to the iteration.
Due to the required memory of Liblinear increases with the size of dataset, it
tends to out of memory with the increasing data size, while the same thing
does not appear in PEGASOS or SGTSVM.
5. Conclusion
The stochastic gradient twin support vector machines (SGTSVM) based
on stochastic gradient decent algorithm has been proposed. By hiring the
26
nonparallel hyperplanes, SGTSVM is more stable on stochastic sampling
than PEGASOS. In theory, we prove that SGTSVM is convergent, and it is
an approximation of TWSVM with uniform sampling. Experimental results
have confirmed the merits of SGTSVM and shown our SGTSVM has better
accuracy compared with Liblinear and PEGASOS with the fastest learning
speed. For practical convenience, the corresponding SGTSVM codes (including Matlab and C language) can be downloaded from http://www.optimal-group.org/Resource/S
For the future work, it is possible to design some special sampling for SGTSVM
to obtain more powerful performance, together with applying SGTSVM on
the bigdata problems.
Acknowledgment
This work is supported by the National Natural Science Foundation of
China (Nos. 11501310, 11201426, and 11371365), the Natural Science Foundation of Inner Mongolia Autonomous Region of China (No. 2015BS0606),
and the Zhejiang Provincial Natural Science Foundation of China (No. LY15F
030013).
References
[1] M.S. Bazarra, H.D. Sherali, and C.M. Shetty. Nonlinear ProgrammingTheory and Algorithms, second ed. Wiley, 2004.
[2] A. Bennar and J.M. Monnez. Almost sure convergence of a stochastic
approximation process in a convex set. International Journal of Applied
Mathematics, 20(5):713–722, 2007.
[3] J.B. Bi and V.N. Vapnik. Learning with rigorous support vector machines. Springer, 2003.
[4] C.L. Blake and C.J. Merz. UCI Repository for Machine Learning
Databases. http://www.ics.uci.edu/~ mlearn/MLRepository.html,
1998.
[5] C.C. Chang and C.J. Lin. LIBSVM: A library for support vector machines. http://www.csie.ntu.edu.tw/~ cjlin, 2001.
[6] W.J. Chen, Y.H. Shao, C.N. Li, and N.Y. Deng. Mltsvm: A novel twin
support vector machine to multi-label learning. Pattern Recognition,
52:61–74, 2015.
27
[7] C. Cortes and V.N. Vapnik. Support vector networks. Machine Learning,
20:273–297, 1995.
[8] R.O. Duda, P.E. Hart, and D.G. Stork. Pattern Classification, 2nd
Edition. John Wiley and Sons, 2001.
[9] R.E. Fan, K.W. Chang, C.J. Hsieh, X.R. Wang, and C.J. Lin. LIBLINEAR: a library for large linear classification. Journal of Machine
Learning Research, 9:1871–1874, 2008.
[10] G.H. Golub and L.C.F. Van. Matrix Computations. The John Hopkins
University Press, 1996.
[11] H. Ince and T.B. Trafalis. Support vector machine for regression and
applications to financial forecasting. In International Joint Conference
on Neural Networks, pages 6348–6354, Italy, 2002.
[12] Jayadeva, R. Khemchandani, and S. Chandra. Twin support vector
machines for pattern classification. IEEE Trans.PatternAnal. Machine
Intell, 29(5):905–910, 2007.
[13] T. Joachims. Making large-scale SVM learning practical. In Advances
in Kernel Methods-Support Vector Learning, pages 169–184, Cambridge,
1998.
[14] S.M. Kakade and A. Tewari. On the generalization ability of online
strongly convex programming algorithms. In Advances in Neural Information Processing Systems, pages 801–808, 2009.
[15] R. Khemchandani, Jayadeva, and S. Chandra. Optimal kernel selection
in twin support vector machines. Optimization Letters, 3:77–88, 2009.
[16] J. Kivinen, A.J. Smola, and R.C. Williamson. Online learning with kernels. Signal Processing, IEEE Transactions on, 52(8):2165–2176, 2004.
[17] T.N. Lal, M. Schröder, T. Hinterberger, J. Weston, M. Bogdan, N. Birbaumer, and B. Schölkopf. Support vector channel selection in BCI.
Data Mining and Knowledge Discovery, 51(6):1003–1010, 2004.
[18] Y.J. Lee and O.L. Mangasarian. RSVM: Reduced support vector machines. In First SIAM International Conference on Data Mining, pages
5–7, Chicago, IL, USA, 2001.
28
[19] D.W. Li, Y.J. Tian, and H.G. Xu. Deep twin support vector machine.
In Data Mining Workshop (ICDMW), 2014 IEEE International Conference on, pages 65–73. IEEE, 2014.
[20] O.L. Mangasarian. Nonlinear Programming. SIAM, 1994.
[21] O.L. Mangasarian and E.W. Wild. Multisurface proximal support vector classification via generalize eigenvalues. IEEE Trans.PatternAnal.
Machine Intell, 28(1):69–74, 2006.
[22] Matlab.
User’s
Guide,
The
http://www.mathworks.com, 1994-2010.
MathWorks,
Inc.
[23] W.S. Noble. Support vector machine applications in computational biology. In Kernel Methods in Computational Biology, Cambridge, 2004.
[24] X.J. Peng. TPMSVM: A novel twin parametric-margin support vector
machine for pattern recognition. Pattern Recognition, 44(10-11):2678–
2692, 2011.
[25] J. Platt. Fast training of support vector machines using sequential minimal optimization. In Advances in kernel methods-support vector learning, pages 185–208, Cambridge, MA: MIT Press, 1999.
[26] Z. Qi, Y. Tian, and Y. Shi. Laplacian twin support vector machine for
semi-supervised classification. Neural Networks, 35:46–53, 2012.
[27] Z. Qi, Y. Tian, and Y. Shi. Twin support vector machine with universum
data. Neural Networks, 36:112–119, 2012.
[28] Z. Qi, Y. Tian, and Y. Shi. Robust twin support vector machine for
pattern classification. Pattern Recognition, 46(1):305–316, 2013.
[29] Z. Qi, Y. Tian, and Y. Shi. Successive overrelaxation for laplacian support vector machine. IEEE transactions on neural networks and learning
systems, 26(4):674–683, 2015.
[30] W. Rudin. Principles of mathematical analysis, volume 3. McGraw-Hill
New York, 1964.
[31] B. Schölkopf and A. Smola. Learning with kernels. MA:MIT Press,
Cambridge, 2002.
29
[32] S.S. Shai, Y. Singer, N. Srebro, and A. Cotter. Pegasos: Primal
estimated sub-gradient solver for svm. Mathematical programming,
127(1):3–30, 2011.
[33] Y.H. Shao, W.L. Chen, J.J. Zhang, Z. Wang, and N.Y. Deng. An efficient
weighted lagrangian twin support vector machine for imbalanced data
classification. Pattern Recognition, 47(9):3158–3167, 2014.
[34] Y.H. Shao and N.Y. Deng. A coordinate descent margin based-twin
support vector machine for classification. Neural Networks, 25:114–121,
2012.
[35] Y.H. Shao, C.H. Zhang, X.B. Wang, and N.Y. Deng. Improvements on
twin support vector machines. IEEE Transactions on Neural Networks,
22(6):962 – 968, 2011.
[36] Y. Tian, Z. Qi, and X. Ju. Nonparallel support vector machines for
pattern classification. IEEE transactions on cybernetics, 44(7):1067–
1079, 2014.
[37] Y.J. Tian and Y. Ping. Large-scale linear nonparallel support vector
machine solver. Neural Networks, 50:166–174, 2014.
[38] Z. Wang, Y.H. Shao, L. Bai, and N.Y. Deng. Twin support vector
machine for clustering. IEEE Transactions on Neural Networks and
Learning Systems, 26(10):2583–2588, 2015.
[39] Z. Wang, Y.H. Shao, and T.R. Wu. A ga-based model selection for
smooth twin parametric-margin support vector machine. Pattern Recognition, 46(8):2267–2277, 2013.
[40] Z. Wang, Y.H. Shao, and T.R. Wu. Proximal parametric-margin support vector classifier and its applications. Neural Computing and Applications, 24(3-4):755–764, 2014.
[41] W. Xu. Towards optimal one pass large scale learning with averaged
stochastic gradient descent. arXiv preprint arXiv:1107.2490, 2011.
[42] C.H. Zhang, Y.J. Tian, and N.Y. Deng. The new interpretation of
support vector machines on statistical learning theory. Science China,
53(1):151–164, 2010.
30
[43] T. Zhang. Solving large scale linear prediction problems using stochastic
gradient descent algorithms. In Proceedings of the twenty-first international conference on Machine learning, page 116. ACM, 2004.
31
| 1cs.CV
|
arXiv:1603.08077v1 [math.GT] 26 Mar 2016
THE MOST SYMMETRIC SURFACES IN THE 3-TORUS
SHENG BAI, VANESSA ROBINS, CHAO WANG, AND SHICHENG WANG
Abstract. Suppose an orientation preserving action of a finite group
G on the closed surface Σg of genus g > 1 extends over the 3-torus T 3 for
some embedding Σg ⊂ T 3 . Then |G| ≤ 12(g − 1), and this upper bound
12(g − 1) can be achieved for g = n2 + 1, 3n2 + 1, 2n3 + 1, 4n3 + 1, 8n3 +
1, n ∈ Z+ . Those surfaces in T 3 realizing the maximum symmetries
can be either unknotted or knotted. Similar problems in non-orientable
category is also discussed.
Connection with minimal surfaces in T 3 is addressed and when the
maximum symmetric surfaces above can be realized by minimal surfaces
is identified.
Contents
1. Introduction
2. Upper and lower bounds of extendable finite groups
3. Unknotted examples of the most symmetric surfaces
4. Knotted examples of the most symmetric surfaces
5. Minimal surfaces, space groups, proof of the main result
References
1
4
6
10
16
18
1. Introduction
Let Σg be the closed orientable surface of genus g > 1, Πg be the closed
non-orientable surface of genus g > 2, and T 3 be the three dimensional
torus (3-torus for short). We consider the following question in the smooth
category:
Suppose the action of a finite group G on a closed surface S can extend
over T 3 . Then what is the maximum order of the group and what does the
maximum action look like?
A similar problem has been addressed for surfaces embedded in the 3sphere, S 3 , which is the simplest compact 3-manifold in the sense that it is
a one point compactification of our three space and the universal spherical
3-manifold covering all spherical 3-manifolds. See [WWZZ1] and [WWZZ2]
for surfaces in the orientable category. Note that only orientable surfaces Σg
2010 Mathematics Subject Classification. 57M60, 57N10, 57S25, 53A10, 20F65, 05C10;
Key words and phrases. maximum surface symmetry in 3-torus, minimal surface.
The second author is supported by grant No.11501534 of NSFC and the last author is
supported by grant No.11371034 of NSFC.
1
2
SHENG BAI, VANESSA ROBINS, CHAO WANG, AND SHICHENG WANG
can be embedded in S 3 . T 3 is another natural and significant 3-manifold for
this question. It is the universal compact Euclidean 3-manifold in the sense
that it covers all compact Euclidean 3-manifolds; moreover T 3 is covered
by our 3-space and the preimage of a closed surface S ⊂ T 3 under such a
covering can be a triply periodic surface, which is an interesting object in
the natural sciences and engineering [HBLL+].
Definition 1.1. Let G be a finite group. A G-action on a closed surface S
is extendable over T 3 with respect to an embedding e : S ֒→ T 3 if G can act
on T 3 such that h ◦ e = e ◦ h for any h ∈ G.
For an embedding of an orientable surface in T 3 we can define whether it
is unknotted in the same way as the usual definition for knotting in S 3 .
Definition 1.2. An embedding e : Σg ֒→ T 3 is unknotted if e(Σg ) splits T 3
into two handlebodies. Otherwise it is knotted.
We assume all orientable manifolds in this note are already oriented. According to whether the action preserves the orientation of the surface and
T 3 , we can define four classes of maximum orders.
Definition 1.3. Let S be either Σg or Πg . Define E(S), E + (S), E+ (Σg )
+
(Σg ) as below:
and E+
E(S): the maximum order of all extendable G-actions on S.
E + (S): the maximum order of all extendable G-actions on S which preserve the orientation of T 3 .
E+ (Σg ): the maximum order of all extendable G-actions on Σg which
preserve the orientation of Σg .
+
(Σg ): the maximum order of extendable G-actions on Σg , which preE+
serve the orientation of both T 3 and Σg .
We will prove the following Theorem 1.4 and Theorem 1.5. Theorem
1.4 gives the upper bounds of those invariants defined in Definition 1.3,
and Theorem 1.5 provides infinitely many g to realize each upper bound in
Theorem 1.4.
Theorem 1.4. Suppose the genus g > 1 for Σg and g > 2 for Πg .
+
(Σg ) ≤ 12(g − 1).
(1) E+
(2) E+ (Σg ) ≤ 24(g − 1), E + (Σg ) ≤ 24(g − 1), E(Σg ) ≤ 48(g − 1).
(3) E + (Πg ) ≤ 12(g − 2), E(Πg ) ≤ 24(g − 2).
If an extendable G-action on S realizes one of the above upper bounds,
then the corresponding surface e(S) can be thought as a most symmetric
surface in T 3 .
Theorem 1.5. Suppose n is any positive integer.
+
(Σg ) in Theorem 1.4 can be achieved by an
(1) The upper bound of E+
unknotted embedding for g = 2n3 + 1, 4n3 + 1, 8n3 + 1; and by a knotted
embedding for g = 2n3 + 1, 4n3 + 1, 8n3 + 1, where 6 does not divide n; and
g = n2 + 1, 3n2 + 1.
THE MOST SYMMETRIC SURFACES IN THE 3-TORUS
3
(2) The upper bound of E+ (Σg ), E + (Σg ) and E(Σg ) in Theorem 1.4 can
be achieved for g = 2n3 + 1, 4n3 + 1, 8n3 + 1 by an unknotted embedding.
(3) The upper bound of E + (Πg ) and E(Πg ) in Theorem 1.4 can be achieved
for g = 2n3 + 2, 8n3 + 2, where n is odd.
The proof of Theorem 1.4, using the Riemann-Hurwitz Formula and the
Equivariant Loop Theorem, is given in §2. The proof of Theorem 1.5 occupies the remaining three sections of the paper, which is outlined as below:
In §3 and §4 we will construct various examples to realize those upper
bounds in Theorem 1.5 (1), §3 for the unknotted case and §4 for the knotted case, therefore Theorem 1.5 (1) follows, up to the verification of the
knottiness of those embeddings. Those examples are explicit in the following sense: The 3-torus is obtained by standard opposite face identification of
the cube and we define those surfaces in the cube before the identification.
Indeed, for each g in Theorem 1.5 (1), we construct all embeddings Σg ⊂ T 3
+
(Σg ) we can image for the moment, and we expect those are all
realizing E+
+
(Σg ), see a conjecture below. The constructions in
embeddings realizing E+
§3 and §4 often rely on an understanding of crystallographic space groups.
The knottedness of the examples in §3 and §4 involved in Theorem 1.5
(1), as well as Theorem 1.5 (2) and (3), could be verified by arguments
such as those in §3 and §4, and by some topological reasoning; but we
present a more convenient and interesting way in §5. In §5 we first link
the verification of knottedness of the examples in §3 and §4 with minimal
surfaces in T 3 , or equivalently, triply periodic minimal surfaces in R3 , which
itself is an important topic, see [Me] and [FH] for examples. Historically,
many minimal surfaces in T 3 were constructed using the symmetry of T 3
and it is known that minimal surfaces in T 3 must be unknotted [Me]. We
identify the examples in §3 with known triply periodic minimal surfaces
(up to isotopy) [Br], therefore they are unknotted. On the other hand,
by a simple criterion from covering space theory and our constructions,
the examples in §4 are knotted, therefore can not be realized by minimal
surfaces. We also identify the group actions in §3 and §4 with known space
groups as well as their index-2 supergroups; then Theorem 1.5 (2) and (3)
are proved.
+
(g) = 12(g − 1) can
Corollary 1.6. Σg ⊂ T 3 realizing the upper bound E+
3
3
be realized by minimal surfaces for g = 2n + 1, 4n + 1, 8n3 + 1.
For each g it is easy to construct an extendable action of order 4(g − 1)
on (T 3 , Σg ), see Example 2.3. We end the introduction by the following
Conjecture 1.7. Suppose the genus g > 1.
+
(Σg ) = 12(g − 1) are listed in
(1) All g realizing the upper bound E+
Theorem 1.5 (1), and moreover, all examples realizing those g are listed in
§3 and §4.
+
(Σg ) = 4(g − 1) for g ∈ Z+ \ K, where K = {fi (n)|n ∈ Z+ } and
(2) E+
fi runs over finitely many quadratic and cubic functions.
4
SHENG BAI, VANESSA ROBINS, CHAO WANG, AND SHICHENG WANG
2. Upper and lower bounds of extendable finite groups
We need the following two important results, see [Hu], [MY], to prove
Theorem 1.4.
Riemann-Hurwitz Formula. Σg → Σg′ is a regular branched covering
with transformation group G. Let a1 , a2 , · · · , ak be the branched points in
Σg′ having indices q1 ≤ q2 ≤ · · · ≤ qk . Then
2 − 2g = |G|(2 − 2g ′ −
k
X
1
(1 − ))
qi
i=1
Equivariant Loop Theorem. Let M be a three manifold with a smooth
action of a discrete group G. Let F be an equivariant subsurface of ∂M .
If F is not π1 -injective with respect to inclusion into M , then it admits a
G-equivariant compression disk.
Here a nonempty subset X of M is G-equivariant if h(X) = X or h(X) ∩
X = ∅, for any h ∈ G. A disk D ⊂ M is a compression disk of F if ∂D ⊂ F
and in F it does not bound any disk.
Proof of (1). Suppose there is an extendable G-action on Σg which preserves
the orientation of both T 3 and Σg . Cutting T 3 along e(Σg ) we get a three
manifold M . Since both Σg and T 3 are orientable, Σg must be two-sided in
T 3 , therefore M contains two boundaries F1 and F2 , and F1 ∼
= F2 ∼
= Σg .
Now clearly G acts on M . Since the G-action preserves the orientation of
both T 3 and Σg , each of F1 and F2 is a G-equivariant surface. Since g > 1,
π1 (Σg ) is not abelian, but π1 (T 3 ) is abelian so the induced homomorphism
π1 (Σg ) → π1 (T 3 ) is not injective. Then at least one of F1 and F2 is not π1 injective with respect to inclusion into M . Suppose F1 is not π1 -injective,
then it admits a G-equivariant compression disk D of F1 by the Equivariant
Loop Theorem.
If there exists an h ∈ G such that h(D) = D and h reverses an orientation
of D, then we can choose a G-equivariant regular neighbourhood N (D)
of D such that there is a homeomorphism i : D × [−1, 1] → N (D) and
i(D × {0}) = D. Then D ′ = i(D × {1}) is also a G-equivariant compression
disk of F1 , and clearly for this D ′ , if h′ (D ′ ) = D ′ for some h′ ∈ G, then h′
preserves an orientation of D ′ . Hence we can assume each element of G that
preserves D also preserves its orientation, in particular it is fixed point free
on ∂D.
Since the G-action on T 3 preserves the orientations of both Σg and T 3 ,
the induced G-action on M preserves F1 . With the quotient topology F1 /G
is homeomorphic to some Σg′ , and p : F1 → F1 /G is a regular branched
covering. Since the G action is fixed point free on ∂D by previous discussion,
p(∂D) is a simple closed curve in F1 /G.
Let a1 , a2 , · · · , ak be the branch points having indices q1 ≤ q2 ≤ · · · ≤ qk .
Note 2 − 2g < 0. By the Riemann-Hurwitz Formula we have
THE MOST SYMMETRIC SURFACES IN THE 3-TORUS
2 − 2g′ −
5
k
X
1
2 − 2g
(1 − ) =
< 0.
qi
|G|
i=1
If g′ = 0 and k ≤ 3, then p(∂D) must bound a disk in F1 /G containing
at most one branch point, and ∂D will bound a disk in F1 , a contradiction.
Hence either g′ ≥ 1 or g ′ = 0 and k ≥ 4.
Notice that for each qi , we have
1
≥ 1/2.
qi
Then by elementary calculation we have |G| ≤ 12(g−1), and moreover the
equality holds if and only if g′ = 0, k = 4 and (q1 , q2 , q3 , q4 ) = (2, 2, 2, 3).
1−
Remark 2.1. Zimmermann first proved the order of orientation preserving
finite group action on handlebody of genus g is bounded by 12(g − 1) [Zi]
soon after the work [MY].
By the above proof, we actually have the following:
Theorem 2.2. Suppose Σg is embedded in a three manifold M and a finite group G acts on (Σg , M ). If G preserves both the two sides and the
orientation of Σg and Σg is not π1 -injective in M , then |G| ≤ 12(g − 1).
Now (2) and (3) follow from (1).
Proof of (2). Suppose there is an extendable G-action on Σg . Let Go be
the normal subgroup of G containing all elements that preserve both the
orientations of Σg and T 3 . In the case of E+ (Σg ) or E + (Σg ), the index
of Go in G is at most two, and in the case of E(Σg ), the index of Go
in G is at most four. Hence E+ (Σg ) ≤ 24(g − 1), E + (Σg ) ≤ 24(g − 1),
E(Σg ) ≤ 48(g − 1).
Proof of (3). Suppose there is an extendable G-action on Πg and the action preserves the orientation of T 3 . We can choose an equivariant regular neighbourhood N (Πg ) of Πg , then G also acts on ∂N (Πg ). Since T 3
is orientable, N (Πg ) is homeomorphic to a twisted [−1, 1]-bundle over Πg
and ∂N (Πg ) is the orientable double cover of Πg under the bundle projection. Since χ(Πg ) = 2 − g, χ(∂N (Πg )) = 4 − 2g = 2 − 2(g − 1). It
follows that ∂N (Πg ) is homeomorphic to Σg−1 . Clearly the G-action preserves the two sides of ∂N (Πg ). Since the action preserves the orientation
of T 3 , it also preserves the orientation of ∂N (Πg ). Then by the result of
(1), we have E + (Πg ) ≤ 12(g − 2). And similar to the proof of (2), we have
E(Πg ) ≤ 24(g − 2).
Example 2.3. Let ([0, 1]3 , +) be the unit cube with a cross properly embedded shown as the right side of Figure 1. For each g > 1, put g − 1 copies of
([0, 1]3 , +) to get ([0, 1]3 , +)g−1 which is a cuboid with an antennae properly
embedded, shown as Figure 1. If we identify the opposite faces of the cuboid,
6
SHENG BAI, VANESSA ROBINS, CHAO WANG, AND SHICHENG WANG
we get (T 3 , Θg ), where Θg is a graph of genus g. If we first only identify
the right and left faces, we get ([0, 1]2 × S 1 , Ag−1 ), where Ag−1 is a circular
antennae with g − 1 bars. It is easy to see that G = Dg−1 ⊕ Z2 acts on the
pair ([0, 1]2 × S 1 , Ag−1 ), where Dg−1 is the dihedral group acts in a standard way on ([0, 1]2 × S 1 , Ag−1 ), and Z2 is π-rotation of ([0, 1]2 × S 1 , Ag−1 )
around the center-circle. Clear this action preserves the each pair of faces of
([0, 1]2 × S 1 , Ag−1 ) to be identified, therefore induces an action on (T 3 , Θg ).
Let N (Θg ) be a G-invariant regular neiborhood of Θg . Then ∂N (Θg ) = Σg
and G of order 4(g − 1) acts on (T 3 , Σg ).
Figure 1.
3. Unknotted examples of the most symmetric surfaces
In this section we give three classes of examples realizing the upper bound
+
(Σg ) = 12(g − 1). In each case we first construct a triply periodic
of E+
graph Γ in the three-dimensional Euclidean space E 3 . There will be a
three-dimensional space group G preserving Γ. Then we choose a rankthree translation normal subgroup T in G. The space E 3 /T is our T 3 . The
finite group G = G/T acts on T 3 preserving the graph Θ = Γ/T . Finally, we
choose an equivariant regular neighbourhood N (Θ) of Θ, and ∂N (Θ) is our
surface Σg . After the discussion of §5 we will see that for each example the
complement of N (Θ) is also a handlebody, hence the surface is unknotted.
Definition 3.1. Let T1 = {(a, b, c) | a, b, c ∈ Z} be the group of integer
translations. Let tx = (1, 0, 0), ty = (0, 1, 0), tz = (0, 0, 1). An element
t = (a, b, c) ∈ T1 acts on E 3 as following:
t : (x, y, z) 7→ (x + a, y + b, z + c)
For n ∈ Z+ , we define three classes of subgroups Tm of T1 for those integers
m that can be presented (uniquely) in the form n3 , 2n3 and 4n3 as follows:
Tn3 = hntx , nty , ntz i,
T2n3 = hnty + ntz , ntz + ntx , ntx + nty i,
T4n3 = h−ntx + nty + ntz , ntx − nty + ntz , ntx + nty − ntz i.
Here the subscript m of Tm is equal to the volume V ol(E 3 /Tm ), m =
n3 , 2n3 , 4n3 .
THE MOST SYMMETRIC SURFACES IN THE 3-TORUS
7
Clearly T2n3 and T4n3 are subgroups of Tn3 . One can easily verify that
T(2n)3 ⊂ T2n3 , since 2ntx , 2nty , 2ntz are linear combinations of nty +ntz , ntz +
ntx , ntx + nty , hence T32n3 = T4(2n)3 ⊂ T(2n)3 = T8n3 ⊂ T2n3 . Similarly one
can verify that T(2n)3 ⊂ T4n3 , hence T16n3 = T2(2n)3 ⊂ T(2n)3 = T8n3 ⊂ T4n3 .
Definition 3.2. We define five isometries of E 3 as follows:
ry : (x, y, z) 7→ (−x, y, −z)
rz : (x, y, z) 7→ (−x, −y, z)
rxy : (x, y, z) 7→ (y, x, −z)
rxyz : (x, y, z) 7→ (z, x, y)
t1/2 : (x, y, z) 7→ (x + 1/2, y + 1/2, z + 1/2)
The isometries ry , rz and rxy are 2-fold rotations (i.e., rotation by an
angle of π) about the y-axis, z-axis and the line x = y, z = 0 respectively.
The isometry rxyz is a positive 3-fold rotation about the cube body diagonal,
x = y = z (i.e., right-hand rule rotation by 2π/3 about the direction [1, 1, 1].)
The notations of space groups used below come from [Ha].
Example 3.3. Let Γ1min beSthe one skeleton of the unit cube [0, 1]3 in E 3
as in Figure 2. Let Γ1 = t∈T1 t(Γ1min ). Then Γ1 , the 1-skeleton of the
tessellation of E 3 by the unit cube, is a triply periodic graph called the
simple or primitive cubic lattice (pcu in [RCSR]).
Figure 2. One skeleton of [0, 1]3
Let H 1 = hry , rz , rxy , rxyz i, which is the well-known orientation-preserving
isometric group of [−1, 1]3 of order 24 (with Schönflies symbol O). Let
G 1 = hT1 , H 1 i, this is the space group P 432. Then G 1 preserves Γ1 . Now in
T3 ∼
= E 3 /T1 we have a graph Θ11 = Γ1 /T1 . It has one vertex and three edges,
hence its Euler characteristic χ(Θ11 ) = −2 and its genus g = 1 − χ(Θ11 ) = 3.
Clearly G11 = G 1 /T1 ∼
= E 3 /T1 preserving Θ11 , and G11
= H 1 acts on T 3 ∼
has order 24. Hence when we choose an equivariant regular neighbourhood
N (Θ11 ) of Θ11 , we get an extendable action of order 24 on Σ3 ∼
= ∂N (Θ11 ).
8
SHENG BAI, VANESSA ROBINS, CHAO WANG, AND SHICHENG WANG
Similarly in T 3 ∼
= E 3 /Tn3 we have a graph Θ1n3 = Γ1 /Tn3 . Since Tn3 is a
normal subgroup of G 1 , G1n3 = G 1 /Tn3 acts on T 3 ∼
= E 3 /Tn3 preserving Θ1n3 .
χ(Θ1n3 ) = χ(Θ11 ) · V ol(E 3 /Tn3 ) = −2n3
|G1n3 | = |G11 | · V ol(E 3 /Tn3 ) = 24n3
Hence the genus of Θ1n3 is 2n3 + 1. Then we can get an extendable action
of order 24n3 on Σ2n3 +1 ∼
= ∂N (Θ1n3 ).
Notice that when m = 2n3 , 4n3 , n ∈ Z+ , Tm is also a normal subgroup
of G 1 . Similar to the above construction, we can get an extendable action
of order 24m on Σ2m+1 ∼
= ∂N (Θ1m ). Here Θ1m = Γ1 /Tm is a graph in
3
3
T ∼
= E /Tm .
In the above example the superscript 1 in Γ1 or Θ1m is equal to the volume
of the ‘minimal 3-torus’ E 3 /T1 . It is also equal to the number of vertices of
the graph Θ11 . In following examples we use similar notations.
Example 3.4. Let Γ2min be the graph in the unit cube [0, 1]3 in E 3 as in
Figure 3. It consists of four
(1/2, 1/2, 1/2) to (0, 0, 0), (0, 1, 1),
S edges from
2
2
(1, 0, 1), (1, 1, 0). Let Γ = t∈T2 t(Γmin ). One can check that it is connected
and is known to scientists as the bonding structure of diamond (dia in
[RCSR]). Figure 3 shows a fundamental region of T2 .
Figure 3. Γ2min in [0, 2] × [0, 1] × [0, 1]
Let H 2 = hry , rz , rxyz i. It is the orientation-preserving isometric group of
the regular tetrahedron formed by the convex hull of (1, 1, 1), (1, −1, −1),
(−1, 1, −1), and (−1, −1, 1) and has Schönflies symbol T . Let G 2 = hT2 , H 2 , t1/2 rxy i,
this is the space group F 41 32. Then G 2 preserves Γ2 .
Now in T 3 ∼
= E 3 /T2 we have a graph Θ22 = Γ2 /T2 . It has two vertices
and four edges, hence its Euler characteristic χ(Θ22 ) = −2 and its genus is
3. T2 is a normal subgroup of G 2 and G22 = G 2 /T2 has order 24. It acts on
T3 ∼
= E 3 /T2 preserving Θ22 . Hence when we choose an equivariant regular
neighbourhood N (Θ22 ) of Θ22 , we get an extendable action of order 24 on
Σ3 ∼
= ∂N (Θ22 ).
Similarly when m = n3 , 4n3 , 16n3 , n ∈ Z+ , in T 3 ∼
= E 3 /T2m we have a
2
2
graph Θ2m = Γ /T2m . One can check that T2m is a normal subgroup of G 2 ,
THE MOST SYMMETRIC SURFACES IN THE 3-TORUS
9
and the quotient G22m = G 2 /T2m acts on T 3 ∼
= E 3 /T2m preserving Θ22m .
χ(Θ22m ) = χ(Θ22 ) · V ol(E 3 /T2m )/V ol(E 3 /T2 ) = −2m
|G22m | = |G22 | · V ol(E 3 /T2m )/V ol(E 3 /T2 ) = 24m
Hence the genus of Θ22m is 2m + 1. Then we can get an extendable action
of order 24m on Σ2m+1 ∼
= ∂N (Θ22m ).
Example 3.5. Let γ be the graph in the unit cube [0, 1]3 in E 3 as in Figure 4. It has three edges from (1/4, 1/4, 1/4) to (0, 1/2, 1/4), (1/4, 0, 1/2)
and (1/2, 1/4, 0). Let Γ4min = γ ∪ t2y tz ry rz (γ) ∪ t2x ty tz ry (γ) ∪ t2x ty rz (γ).
S
Then let Γ4 = t∈T4 t(Γ4min ). One can check that it is connected, and is
the oft-rediscovered chiral vertex-transitive net of degree-3 known by many
names [HOP], including srs in [RCSR]. Figure 4 shows a fundamental region
of T4 .
Figure 4. Γ4min in [0, 2] × [0, 2] × [0, 1]
Let G 4 = hT4 , tx tz rz , ty tz ry , rxyz , tx t1/2 rxy i, it is the space group I41 32.
Then G 4 preserves Γ4 .
Now in T 3 ∼
= E 3 /T4 we have a graph Θ44 = Γ4 /T4 . It has four vertices
and six edges, hence its Euler characteristic χ(Θ44 ) = −2 and its genus is
3. T4 is a normal subgroup of G 4 and G44 = G 4 /T4 has order 24. It acts on
T3 ∼
= E 3 /T4 preserving Θ44 . Hence when we choose an equivariant regular
neighbourhood N (Θ44 ) of Θ44 , we get an extendable action of order 24 on
Σ3 ∼
= ∂N (Θ44 ).
Similarly when m = n3 , 2n3 , 4n3 , n ∈ Z+ , in T 3 ∼
= E 3 /T4m we have a
4
4
graph Θ4m = Γ /T4m . One can check that T4m is a normal subgroup of G 4 ,
and the quotient G44m = G 4 /T4m acts on T 3 ∼
= E 3 /T4m preserving Θ44m .
χ(Θ44m ) = χ(Θ44 ) · V ol(E 3 /T4m )/V ol(E 3 /T4 ) = −2m
|G44m | = |G44 | · V ol(E 3 /T4m )/V ol(E 3 /T4 ) = 24m
10
SHENG BAI, VANESSA ROBINS, CHAO WANG, AND SHICHENG WANG
Hence the genus of Θ44m is 2m + 1. Then we can get an extendable action
of order 24m on Σ2m+1 ∼
= ∂N (Θ44m ).
4. Knotted examples of the most symmetric surfaces
In the section we give a further six classes of examples realizing the upper
+
(Σg ) = 12(g − 1). In each case below the embedding is knotted.
bound of E+
The surface e(Σg ) does bound a handlebody in T 3 on one side, but the
complement of the handlebody is not a handlebody, and this will be clear
after the discussion of §5.
Constructions of the extendable actions and surfaces is similar to §4,
but here the graph Γ in E 3 can be disconnected. There is a space group G
preserving Γ and a rank three translation normal subgroup T in G so that the
graph Θ = Γ/T is connected. The finite group G = G/T acts on T 3 ∼
= E 3 /T
preserving Θ. Then we choose an equivariant regular neighbourhood N (Θ)
of Θ, and ∂N (Θ) is our surface Σg .
Example 4.1. Let Γ1,2 = Γ1 ∪ t1/2 (Γ1 ). The graph t1/2 (Γ1 ) can be thought
of as the dual of Γ1 in E 3 . Γ1,2 has two connected components and is named
pcu-c in [RCSR]. In the unit cube [0, 1]3 it is as in Figure 5.
Figure 5. Part of Γ1,2 in [0, 1]3
Let G 1,2 = hG 1 , t1/2 i, it is the space group I432, and let T1/2 = hT1 , t1/2 i =
{(a/2, b/2, c/2) | (a, b, c) ∈ T4 }. Then G 1,2 preserves Γ1,2 and T1/2 is a normal subgroup of G 1,2 . Since t1/2 changes the two components of Γ1,2 , the
3
1,2
3 ∼
graph Θ1,2
1/2 = Γ /T1/2 in T = E /T1/2 is connected. Its Euler character-
1,2
1,2 /T
istic χ(Θ1,2
1/2 has order 24. It
1/2 ) = −2 and its genus is 3. G1/2 = G
1,2
3
3
∼
acts on T = E /T1/2 preserving Θ1/2 . Hence when we choose an equivari-
1,2
ant regular neighbourhood N (Θ1,2
1/2 ) of Θ1/2 , we get an extendable action of
1,2
order 24 on Σ3 ∼
= ∂N (Θ ).
1/2
Similarly let Tn3 /2 = {(a/2, b/2, c/2) | (a, b, c) ∈ T4n3 }, n ∈ Z+ , 2 ∤ n.
Then the graph Θ1,2
= Γ1,2 /Tn3 /2 in T 3 ∼
= E 3 /Tn3 /2 is connected. Since
n3 /2
THE MOST SYMMETRIC SURFACES IN THE 3-TORUS
11
Tn3 /2 is a normal subgroup of G 1,2 , G1,2
= G 1,2 /Tn3 /2 acts on T 3 ∼
=
n3 /2
E 3 /Tn3 /2 preserving Θ1,2
.
n3 /2
3
3
3
χ(Θ1,2
) = χ(Θ1,2
1/2 ) · V ol(E /Tn3 /2 )/V ol(E /T1/2 ) = −2n
n3 /2
3
3
3
|G1,2
| = |G1,2
1/2 | · V ol(E /Tn3 /2 )/V ol(E /T1/2 ) = 24n
n3 /2
Hence the genus of Θ1,2
is 2n3 + 1. Then we can get an extendable action
n3 /2
∼ ∂N (Θ1,2
of order 24n3 on Σ2n3 +1 =
), here n ∈ Z+ , 2 ∤ n.
n3 /2
Example 4.2. Let Γ2,2 = Γ2 ∪ tx (Γ2 ). The graph tx (Γ2 ) can be thought
as the dual of Γ2 in E 3 . Γ2,2 has two connected components and is called
dia-c in [RCSR]. In the fundamental region of T2 it is as in Figure 6.
Figure 6. Part of Γ2,2 in [0, 2] × [0, 1] × [0, 1]
Let G 2,2 = hG 2 , tx i, it is the space group P 42 32. Then G 2,2 preserves Γ2,2
and contains T1 as a normal subgroup. Since tx changes the two components
3
2,2
3 ∼
of Γ2,2 , the graph Θ2,2
1 = Γ /T1 in T = E /T1 is connected. Its Euler
2,2
2,2 /T has order
characteristic χ(Θ2,2
1
1 ) = −2, and its genus is 3. G1 = G
2,2
3
3
∼
24. It acts on T = E /T1 preserving Θ1 . Hence when we choose an
2,2
equivariant regular neighbourhood N (Θ2,2
1 ) of Θ1 , we get an extendable
2,2
action of order 24 on Σ3 ∼
= ∂N (Θ1 ).
2,2
Similarly if m = n3 , 4n3 , n ∈ Z+ , 2 ∤ n, the graph Θ2,2
m = Γ /Tm in
3
3
∼
T = E /Tm is connected. One can check that Tm is a normal subgroup of
2,2 /T acts on T 3 ∼ E 3 /T preserving Θ2,2 .
G 2,2 , G2,2
=
m =G
m
m
m
2,2
3
χ(Θ2,2
m ) = χ(Θ1 ) · V ol(E /Tm ) = −2m
2,2
3
|G2,2
m | = |G1 | · V ol(E /Tm ) = 24m
Hence the genus of Θ2,2
m is 2m + 1. Then we can get an extendable action
2,2
of order 24m on Σ2m+1 ∼
= ∂N (Θm ), here m = n3 , 4n3 , n ∈ Z+ , 2 ∤ n.
Example 4.3. Let Γ4,4 = Γ4 ∪ tx (Γ4 ) ∪ ty (Γ4 ) ∪ tz (Γ4 ). Γ4,4 has four
connected components. In the unit cube [0, 1]3 in E 3 it is as in Figure 7.
12
SHENG BAI, VANESSA ROBINS, CHAO WANG, AND SHICHENG WANG
Figure 7. Part of Γ4,4 in [0, 1]3
One can check that G 2,2 defined in Example 4.2 (space group P 42 32)
3
4,4
3 ∼
preserves Γ4,4 , and the graph Θ4,4
1 = Γ /T1 in T = E /T1 is connected.
2,2
2,2 /T has
Its Euler characteristic χ(Θ4,4
1
1 ) = −2, and its genus is 3. G1 = G
4,4
3
3
∼
order 24. It acts on T = E /T1 preserving Θ1 . Hence when we choose an
4,4
equivariant regular neighbourhood N (Θ4,4
1 ) of Θ1 , we get an extendable
4,4
action of order 24 on Σ3 ∼
= ∂N (Θ1 ).
4,4
Similarly if m = n3 , 2n3 , n ∈ Z+ , 2 ∤ n, the graph Θ4,4
m = Γ /Tm in
4,4
2,2
T3 ∼
= E 3 /Tm preserving Θm .
= E 3 /Tm is connected. Gm acts on T 3 ∼
4,4
3
χ(Θ4,4
m ) = χ(Θ1 ) · V ol(E /Tm ) = −2m
2,2
3
|G2,2
m | = |G1 | · V ol(E /Tm ) = 24m
Hence the genus of Θ4,4
m is 2m + 1. Then we can get an extendable action
2,2
of order 24m on Σ2m+1 ∼
= ∂N (Θm ), here m = n3 , 2n3 , n ∈ Z+ , 2 ∤ n.
Example 4.4. Let Γ4,8 = Γ4,4 ∪ t1/2 (Γ4,4 ). It has eight connected components. In the unit cube [0, 1]3 it is as in Figure 8.
One can check that G 1,2 defined in Example 4.1 (space group I432) pre4,8
3 ∼
3
serves Γ4,8 , and the graph Θ4,8
1/2 = Γ /T1/2 in T = E /T1/2 is connected.
Its Euler characteristic is −2 and its genus is 3. The order 24 group G1,2
1/2
4,8
3
3
∼
acts on T = E /T1/2 preserving Θ1/2 . Hence when we choose an equivari-
4,8
ant regular neighbourhood N (Θ4,8
1/2 ) of Θ1/2 , we get an extendable action of
4,8
order 24 on Σ3 ∼
= ∂N (Θ ).
1/2
THE MOST SYMMETRIC SURFACES IN THE 3-TORUS
13
Figure 8. Part of Γ4,8 in [0, 1]3
Similarly for n ∈ Z+ , 2 ∤ n, the graph Θ4,8
= Γ4,8 /Tn3 /2 in T 3 ∼
=
n3 /2
1,2
4,8
E 3 /Tn3 /2 is connected. Gn3 /2 acts on T 3 ∼
= E 3 /Tn3 /2 preserving Θn3 /2 .
3
3
3
χ(Θ4,8
) = χ(Θ4,8
1/2 ) · V ol(E /Tn3 /2 )/V ol(E /T1/2 ) = −2n
n3 /2
3
3
3
|G1,2
| = |G1,2
1/2 | · V ol(E /Tn3 /2 )/V ol(E /T1/2 ) = 24n
n3 /2
is 2n3 + 1. Then we can get an extendable action
Hence the genus of Θ4,8
n3 /2
4,8
of order 24n3 on Σ2n3 +1 ∼
= ∂N (Θn3 /2 ), here n ∈ Z+ , 2 ∤ n.
Example 4.5. Let γ ′ be the graph in the unit cube [0, 1]3 in E 3 as in Figure
3. There are three edges from (1/4, 1/4, 1/4) to (0, 1/4, 1/2), (1/2, 0, 1/4),
(1/4, 1/2, 0), and three edges from (0, 3/4, 1/2) to (1/2, 3/4, 1), (1/2, 0, 3/4)
to (1, 1/2, 3/4) and (3/4, 1/2, 0) to (3/4, 1, 1/2) separately.
S
′ 2
′
2
′
2
′
′4
′4
Let Γ′4
min = γ ∪ty tz ry rz (γ )∪tx ty tz ry (γ )∪tx ty rz (γ ), and Γ =
t∈T4 t(Γmin ).
One can check that Γ′4 has 27 connected components, each is similar to a
mirror image of Γ4 . Figure 9 shows the part of Γ′4 in a fundamental region
of T4 .
By construction, G 4 (space group I41 32) preserves Γ′4 . And one can check
3
′4
′4
3 ∼
that Θ′4
4 = Γ /T4 in T = E /T4 is connected. Θ4 has four vertices and
six edges, hence it has Euler characteristic −2 and genus 3. The order24 group G44 acts on T 3 ∼
= E 3 /T4 preserving Θ′4
4 . Hence when we choose
an equivariant regular neighbourhood N (Θ′4
)
of
Θ′4
4
4 , we get an extendable
′4
action of order 24 on Σ3 ∼
).
∂N
(Θ
=
4
′4
Similarly when m = n3 , 2n3 , 4n3 , n ∈ Z+ , 3 ∤ n, the graph Θ′4
4m = Γ /T4m
4
3
4
3
∼
in T = E /T4m is connected. The quotient G4m = G /T4m acts on T 3 ∼
=
14
SHENG BAI, VANESSA ROBINS, CHAO WANG, AND SHICHENG WANG
Figure 9. Part of Γ′4 min in [0, 2] × [0, 2] × [0, 1]
E 3 /T4m preserving Θ′4
4m .
′4
3
3
χ(Θ′4
4m ) = χ(Θ4 ) · V ol(E /T4m )/V ol(E /T4 ) = −2m
|G44m | = |G44 | · V ol(E 3 /T4m )/V ol(E 3 /T4 ) = 24m
Hence the genus of Θ′4
4m is 2m + 1. Then we can get an extendable action
3
3
3
of order 24m on Σ2m+1 ∼
= ∂N (Θ′4
4m ), here m = n , 2n , 4n , n ∈ Z+ , 3 ∤ n.
√
Definition 4.6. Let tω = (−1/2, 3/2, 0). Define
Tnω2 = hntω , ntx , tz i,
ω
T3n
2 = h2ntω + ntx , ntω + 2ntx , tz i.
Define a rotation rω on E 3 as following:
√
√
1
1
3
3
rω : (x, y, z) 7→ (− x −
y,
x − y, z)
2
2
2
2
ω
T1 is the hexagonal 3D lattice and rω is a rotation by 2π/3 right-handed
with respect to the direction [0, 0, 1].
Example 4.7. Let P be a√regular hexagon in the xy-plane. It has center
a vertex, see Figure 10. Let ΓPmin be
(0, 0, 0) and contains (1/2, 3/6, 0) as S
the boundary of P in E 3 . Let ΓP = t∈T ω t(ΓPmin ). It contains infinitely
1
many connected components, and for each n ∈ Z, the horizontal plane
{(x, y, n) | x, y ∈ E 2 } contains exactly one of them. A local picture of ΓP is
as in Figure 11.
Let H ω = hrω , rx , ry i; it is the isometric group of P (Schönflies symbol
D6 ). Let G ω = hT1ω , H ω i, this is the space group P 622. Then G ω preserves
ΓP , and ΘP1 = ΓP /T1ω is a connected graph in T 3 ∼
= E 3 /T1ω . It has two
vertices and three edges, hence its Euler characteristic χ(Θω1 ) = −1 and
its genus is 2. Clearly Gω1 = G ω /T1ω acts on T 3 ∼
= E 3 /T1ω preserving ΘP1 ,
THE MOST SYMMETRIC SURFACES IN THE 3-TORUS
15
0.6
0.4
0.2
-0.4
0.2
-0.2
0.4
-0.2
-0.4
-0.6
Figure 10. A regular hexagon
Figure 11. A local picture of ΓP
and Gω1 ∼
= H ω has order 12. Hence when we choose an equivariant regular
neighbourhood N (ΘP1 ) of ΘP1 , we get an extendable action of order 12 on
Σ2 ∼
= ∂N (ΘP1 ).
ω is a normal subgroup of
Notice that when m = n2 , 3n2 , n ∈ Z+ , Tm
ω in T 3 ∼ E 3 /T ω is connected, and
G ω . Similarly the graph ΘPm = ΓP /Tm
=
m
3
ω
3
ω
ω
ω
∼
Gm = G /Tm acts on T = E /Tm preserving ΘPm .
ω
χ(ΘPm ) = χ(ΘP1 ) · V ol(E 3 /Tm
)/V ol(E 3 /T1ω ) = −m
ω
|Gωm | = |Gω1 | · V ol(E 3 /Tm
)/V ol(E 3 /T1ω ) = 12m
16
SHENG BAI, VANESSA ROBINS, CHAO WANG, AND SHICHENG WANG
Hence the genus of ΘPm is m + 1. Then we can get an extendable action of
order 12m on Σm+1 ∼
= ∂N (ΘPm ).
5. Minimal surfaces, space groups, proof of the main result
In this section we will finish the proof of Theorem 1.5. We need some
results about triply periodic minimal surfaces, space groups, and a lemma
given below.
There are three classical triply periodic minimal surfaces that admit high
symmetries: Schwarz’s P surface, Schwarz’s D [Schw] and Schoen’s gyroid
surface illustrated in Figure 12 [Br]. Denote them by SP , SD and SG . The
fundamental region for SD is a cube of volume 2, we redraw SD in the
fundamental region [0, 1] × [0, 2] × [0, 1] we used.
SD
SP
SG
Figure 12. Translational unit cells of the minimal surfaces
SP , and SD match the translational unit cells of their associated graphs, Γ1 , Γ2 , and are displayed with the same
viewing direction. The picture of SG above shows the cube
[0, 2] × [0, 2] × [0, 2], twice the region of that used to depict
Γ4 , but depicted with the same viewing direction.
The following results are known (and can be checked):
(a) Boundaries of equivariant regular neighbourhoods of Γ1 , Γ2 and Γ4 in
Example 3.3, 3.4 and 3.5 can be realized by SP , SD and SG respectively.
THE MOST SYMMETRIC SURFACES IN THE 3-TORUS
17
(b) SP , SD and SG are unknotted in T 3 . Actually, any genus g > 1
minimal orientable closed surface in T 3 must be unknotted,[Me].
The notations of space groups below come from [Ha].
(c) G 1 , G 2 , G 4 , G 1,2 and G 2,2 in Example 3.3, 3.4, 3.5, 4.1 and 4.2 are the
space groups [P 432], [F 41 32], [I41 32], [I432] and [P 42 32] respectively.
(d) SP , SD , SG are preserved by space groups [Im3̄m], [P n3̄m], [Ia3̄d]
respectively.
(e) We have the following index two subgroup sequences:
• [P 432] ⊂ [I432] ⊂ [Im3̄m]
• [F 41 32] ⊂ [P 42 32] ⊂ [P n3̄m]
• [I41 32] ⊂ [Ia3̄d]
Lemma 5.1. Let T be a lattice in E 3 , p : E 3 → T 3 = E 3 /T be the covering
map. Σg is an embedded surface in T 3 . If p−1 (Σg ) is not connected, then
Σg is knotted.
Proof. Note first, from the definition of unknotted embedding Σg ⊂ T 3 , the
induced map on the fundamental groups must be surjective. Let F be a
connected component of p−1 (Σg ). Let St(F ) be its stable subgroup in T ,
i.e., the lattice translations that preserve F . If p−1 (Σg ) has more than one
component then St(F ) 6= T , hence π1 (Σg ) → π1 (T 3 ) is not surjective. Hence
Σg is knotted in T 3 .
Proof of Theorem 1.5. By the above results (a) and (b), the surfaces in T 3
described in §3 are unknotted. Since Γ1,2 , Γ2,2 , Γ4,4 , Γ4,8 , Γ′4 and ΓP are
disconnected, by Lemma 5.1, the surfaces in T 3 described in §4 are knotted. Then by combining the Examples in §3 and §4, we finish the proof of
Theorem 1.5 (1).
Now we are going to prove Theorem 1.5 (2) and (3) based on the examples
in §3 and §4, and the results (c), (d) and (e) above.
For (2). Recall G 1 = [P 432]. Example 3.3 told us that when m =
n3 , 2n3 , 4n3 we have the action G 1 /Tm on Γ1 /Tm , or equivalently on SP /Tm
by (a), of order 24m. By (d) [Im3̄m] preserves SP . By (e), we have an order
96m extendable [Im3̄m]/Tm -action on SP /Tm ∼
= Σ2m+1 realizing the upper
bound of E(Σg ), where g = 2m + 1.
The group [Im3̄m]/Tm contains two elements h1 and h2 satisfying: h1
preserves the orientation of SP /Tm and reverses the orientation of E 3 /Tm ,
h2 reverses the orientation of SP /Tm and preserves the orientation of E 3 /Tm .
Then choosing the index two subgroups of [Im3̄m], h[P 432], h1 i and h[P 432], h2 i,
we get an extendable action on Σ2m+1 , realizing the upper bound of E+ (Σg )
and E + (Σg ) respectively. Now (2) is proved.
For (3). Recall G 2 = [F 41 32]. Example 3.4 told us that when m =
3
n , 4n3 , 16n3 we have the action G 2 /T2m on Γ2 /T2m , or equivalently on
SD /T2m by (a), of order 24m.
Similarly we also have an order 96m extendable [P n3̄m]/T2m -action on
SD /T2m ∼
= Σ2m+1 realizing the upper bound of E(Σg ), where g = 2m + 1.
18
SHENG BAI, VANESSA ROBINS, CHAO WANG, AND SHICHENG WANG
Furthermore if m = n3 , 4n3 , n ∈ Z+ , n odd, [P n3̄m]/T2m contains a translation h reversing an orientation of SD /T2m and preserving an orientation
of E 3 /T2m (see Example 4.2 for details, or better to see SD in Figure 12).
Modulo this translation we can get an order 48m extendable [P n3̄m]/Tm action on SD /Tm ∼
= Π2m+2 , realizing the upper bound 24(g − 2) of E(Πg ),
where g = 2m + 2. Then choosing the index two subgroup [P 42 32] = G 2,2 ,
we can get an extendable action on Π2m+2 , realizing the upper bound of
E + (Πg ). Now (3) is proved.
Remark 5.2. In the above discussion for (3), the regular neighbourhood
N (SD /Tm ) of SD /Tm is homeomorphic to a twisted [−1, 1]-bundle over
SD /Tm . Its complement in E 3 /Tm is essentially the regular neighbour2,2
hood N (Θ2,2
m ) of Θm in Example 4.2. Similarly for n ∈ Z+ , n odd, and
m = n3 /2, we can get an order 48m extendable [Im3̄m]/Tn3 /2 -action on
SP /Tn3 /2 ∼
= Π2n3 +2 , realizing the upper bound of E(Πg ). Choosing the
index two subgroup [I432], we can get an extendable action on Π2n3 +2 , realizing the upper bound of E + (Πg ). The regular neighbourhood N (SP /Tn3 /2 )
is homeomorphic to a twisted [−1, 1]-bundle over SP /Tn3 /2 . Its complement
in E 3 /Tn3 /2 is essentially the regular neighbourhood N (Θn1,2
3 /2 ) in Example
4.1. There is no similar construction for a regular neighbourhood of the
Gyroid minimal surface because there is no translation that reverses the
orientation of SG /T4 ; the handlebodies on each side of SG are mirror images
of one another.
References
[Br]
[FH]
[Ha]
[He]
[Hu]
[HBLL+]
[HOP]
[Me]
[MY]
[RCSR]
K. Brakke, Minimal surface page,
http://www.susqu.edu/brakke/evolver/examples/periodic/periodic.html
C. Frohman, J. Hass, Unstable minimal surfaces and Heegaard splittings. Invent. Math. 95 (1989), no. 3, 529-540.
T. Hahn (Ed.), International tables for crystallography, D. Reidel Publishing
Company, (2005).
J. Hempel, 3–manifolds, Princeton University Press, (1976).
A. Hurwitz, Über algebraische Gebilde mit eindeutigen Transformationen in
sich, Mathematische Annalen, 41(3) (1892), 403-442.
S.T. Hyde, Z. Blum, T. Landh, S. Lidin, B.W. Ninham, S. Andersson and
K. Larsson. The Language of Shape: The Role of Curvature in Condensed
Matter. New York, USA: Elsevier, (1996).
S.T. Hyde, M. O’Keeffe, D.M. Proserpio, A Short History of an Elusive Yet
Ubiquitous Structure in Chemistry, Materials, and Mathematics. Angewandte
Chemie Int. Ed. 47, (2008), 79968000.
W. H. Meeks, The conformal structure and geometry of triply periodic minimal
surfaces in R3 , Bull. Amer. Math. Soc. 83 (1977), no. 1, 134-136.
W. H. Meeks, S. T. Yau, The equivariant Dehn’s lemma and loop theorem,
Commentarii Mathematici Helvetici, 56(1) (1981), 225-239.
M. O’Keeffe, M.A. Peskov, S.J. Ramsden, O.M. Yaghi, The Reticular Chemistry Structure Resource (RCSR) Database of, and Symbols for Crystal Nets
Accts. Chem. Res. 41, (2008), 1782-1789.
THE MOST SYMMETRIC SURFACES IN THE 3-TORUS
19
[Schw]
H. A. Schwarz, Gesammelte Mathematische Abhandlungen. Bd 1, Berlin:
Springer, (1890).
[Scho]
A. H. Schoen, NASA Tech. Note No. D-5541, (1970)
[WWZZ1] C. Wang, S. C. Wang, Y. M. Zhang, B. Zimmermann, Extending finite group
actions on surfaces over S 3 . Topology Appl. 160, (2013), no. 16, 2088-2103.
[WWZZ2] C. Wang, S. C. Wang, Y. M. Zhang, B. Zimmerman, Maximal Orders of Extendable Finite Group Actions on Surfaces. Groups Geom. Dyn. 9 (2015), no.
4, 1001-1045.
[Zi]
B. Zimmermann, Uber Homomorphismen n-dimensionaler Henkelkrper und
endliche Erweiterungen von Schottky-Gruppen, Comment. Math. Helv. 56
(1981), 474-486.
School of Mathematical Sciences, Peking University, Beijing 100871, CHINA
E-mail address: [email protected]
Department of Applied Mathematics, Research School of Physics and Engineering, The Australian National University, Canberra, Australia
E-mail address: [email protected]
School of Mathematical Sciences, University of Science and Technology
of China, Hefei 230026, CHINA
E-mail address: chao wang [email protected]
School of Mathematical Sciences, Peking University, Beijing 100871, CHINA
E-mail address: [email protected]
| 4math.GR
|
HYPERGROUPS AND HYPERFIELDS IN UNIVERSAL ALGEBRA
arXiv:1604.03415v1 [math.GR] 11 Apr 2016
LOUIS ROWEN
Abstract. Hypergroups are lifted to power semigroups with negation, yielding a method of transferring
results from semigroup theory. This applies to analogous structures such as hypergroups, hyperfields,
and hypermodules, and permits us to transfer the general theory espoused in [19] to the hypertheory.
1. Introduction
This note, a companion to [19], grew out of a conversation with Matt Baker, in which we realized
that the “tropical hyperfield” of [3] and [22, §5.2] is isomorphic to the “extended” tropical arithmetic in
Izhakian’s Ph.D. dissertation (Tel-Aviv University) in 2005, also cf. [10, 11]. On the other hand, there are
many parallels between the two theories. This motivated us to see whether hyperfields in general also can
be studied by semiring theory, which fits in well with the theory of universal algebra, and which might
be more amenable for further study. Viewing a hyperfield as a group with additive structure on part of
its power set, we want to extend this structure to all of the power set, thereby making the definitions
tighter and “improving” the additive structure to make standard tools more available for developing an
algebraic theory.
Thus, the theme is to embed the category of hyperfields (and their modules) into the category of
semirings with negation (and their modules), as studied in [19], defined on power sets. The tricky part in
passing to the power set is distributivity, which must be weakened at times to a notion that we call “weak
distributivity,” and we thereby weaken “semiring” to “T -semiring,” where distributivity holds only with
respect to a special subset T . Then we can treat all hyperrings (not just hyperfields) in this context.
It turns out that the hyperrings of [3, 22] can be injected naturally into T -semirings, which are power
sets with a negation map, in the context of [19], whereby the hyperring is identified with the subset of
singletons. Then one can develop linear algebra over hyperfields, and also go through [3], making the
appropriate adjustments to view matroids over these semifields with negation.
In the other direction, Henry [9] has defined a hypergroup structure on symmetrized monoids.
1.1. Pre-semirings.
Semigroups need not have an identity element, but monoids do, and are usually written multiplicatively,
i.e. (A, ·, 1). We start with semirings, for which the standard reference is [7]. Since distributivity will
be weakened, we remove it (as well as the element 0) from the definition. We enrich a monoid with a
“hyper” operation (resembling addition).
Definition 1.1. A monoid (A, ·, 1) acts on a set S if there is a multiplication A × S → S satisfying
1s = s and (a1 a2 )s = a1 (a2 s) for all ai ∈ A and s ∈ S.
A pre-semiring (A, ·, +, 1) is a multiplicative monoid (A, ·, 1R ) also possessing the structure of an
additive Abelian semigroup (A, +), on which (A, ·) acts.
A pre-semifield is a pre-semiring (A, ·, +, 1) for which (A, ·, 1) is a group.
A premodule S over a monoid (A, ·, 1) is an Abelian group (S, +, 0) on which A acts, also satisfying
the condition:
r0M = 0M , ∀r ∈ R.
Date: April 13, 2016.
2010 Mathematics Subject Classification. Primary 16Y60, 12K10, 06F05, 14T05 Secondary 12K10, .
Key words and phrases. hyperfield, negation map, tropical algebra, tropical geometry, power set, semiring, supertropical
algebra.
1
2
L. ROWEN
A premodule S over a pre-semiring (A, ·, +, 1) is a premodule over the monoid (A, ·, 1), also satisfying
the condition:
• “Left distributivity”: (a1 + a2 )s = a1 s + a2 s, ∀ai ∈ A, ∀s ∈ S.
Then a semiring† (R, +, ·, 1R ) (without 0) would be a pre-semiring satisfying the usual distributive
laws. There are various versions of distributivity that will be relevant to us later.
Definition 1.2. Define the following notions, where (S, +) is a premodule over a pre-semiring (A, ·, +, 1),
for ai ∈ A and sj ∈ S:
(i) “Right distributivity”: a(s1 + s2 ) = as1 + as2 .
(ii) “Two-sided distributivity”: Left and right distributivity.
(iii) “Double distributivity”: (a1 + a2 )(s1 + s2 ) = a1 s1 + a2 s1 + a1 s2 + a2 s2 .
(iv) “Generalized distributivity”:
!
X
X
X
(ai sj ).
sj =
ai
i
j
i,j
Remark 1.3. The properties are in order of increasing formal strength although, by induction, double
distributivity implies generalized distributivity.
1.2. Motivation: Power sets of semigroups.
We let P(A) denote the power set of a set A, i.e., the set of subsets of A. The sets {a} for a ∈ A are
called singletons.
Theorem 1.4.
(i) Given a monoid (A, ·, 1), we can extend its operation to P(A) elementwise, by putting
S1 S2 = {s1 s2 : sj ∈ Sj }.
Then (P(A), ·) also is a monoid with the identity element {1}, on which A acts.
Given a semigroup (A, +), we can define addition elementwise on P(A) by defining
S1 + S2 = {s1 + s2 : sj ∈ Sj }.
Then (P(A), +) also is a semigroup.
Thus, when A is a pre-semiring (resp. semiring), so is P(A), and P(A) is an A-premodule via
the action
aS = {as : s ∈ S}.
(ii) More generally, the relevant concepts in universal algebra were outlined in [19, §2.3], including
signatures defined via operators and identical relations. We can lift operators from an (Ω; Id)algebra A to P(A), as follows: Given an operator ωm = ωm (x1 , . . . , xm ) on A, we define ωm on
P(A) via
ωm (S1 , . . . , Sm ) = {ωm (s1 , . . . , sm ) : sk ∈ Sk , 1 ≤ k ≤ m}.
Let us call an identical relation φ(x1 , . . . , xℓ ) = φ(x1 , . . . , xℓ ) multilinear if each xi appears
exactly once in the definition (via the operators). Then any multilinear identical relation holding
in A also holds in P(A).
Proof. (i) First we verify associativity:
(S1 + S2 ) + S3 = {a1 + a2 : aj ∈ Sj } + S3
= {(a1 + a2 ) + a3 : aj ∈ Sj }
= {a1 + (a2 + a3 ) : aj ∈ Sj }
= S1 + {a2 + a3 : aj ∈ Sj }
= S1 + (S2 + S3 ).
For generalized distributivity we have:
(1.1)
HYPERGROUPS AND HYPERFIELDS
X
Si
X
Tj =
(
mi
XX
aik : aik ∈ Si
ik k=1
=
X
)
nj
X X
3
bjℓ : bjℓ ∈ Tj
jℓ ℓ=1
aik bjℓ : aik ∈ Si , bjℓ ∈ Tj
i,j,k,ℓ
=
X
(1.2)
Si T j .
(ii) We generalize the proof in (i). By an easy induction applied to [19, Definition 2.12], any formula
φ(x1 , . . . , xℓ ) satisfies
φ(S1 , . . . , Sℓ ) = {φ(s1 , . . . , sℓ ) : sj ∈ Sj },
and thus any multilinear identical relation φ = ψ holding elementwise in A also holds set-wise in P(A).
In particular, distributivity and generalized distributivity lift from a semiring A to P(A). When the
relation is not multilinear we encounter difficulties due to repetition. For example, being a group does not
lift/ Indeed, the defining identical relation xx−1 = 1 is quadratic in x; here we are defining the inverse as
a unary operation ω1 : x 7→ x−1 , so the left side is xω1 (x). In fact the only invertible elements in P(A)
are the singletons.
1.3. Hypermonoids and semirings.
The next step is to formulate all of our extra structure in terms of addition (and possibly other
operations) on P(A) as a premodule over A. But this is not as easy when A itself is not a semiring, so
let us pause to review hypermonoids, to see just how much of the semiring structure we would need.
The “intuitive” definition: A hypermonoid should be a triple (A, ⊞, 0) where ⊞ : A × A → P(A), and
the analog of associativity holds:
(a1 ⊞ a2 ) ⊞ a3 = a1 ⊞ (a2 ⊞ a3 ),
∀a ∈ A.
There is a fundamental difficulty in this definition — a1 ⊞ a2 is a set, not an element of A, so
technically (a1 ⊞ a2 ) ⊞ a3 is not defined. This difficulty is exacerbated when considering generalized
associativity; for example, what does (a1 ⊞ a2 ) ⊞ (a3 ⊞ a4 ) mean? We rectify this by passing to P(A).
Definition 1.5. A hypermonoid is a triple (A, ⊞, 0) where
(i) ⊞ is a commutative binary operation A × A → P(A), which also is associative in the sense that
if we define
a ⊞ S = ∪s∈S a ⊞ s,
then (a1 ⊞ a2 ) ⊞ a3 = a1 ⊞ (a2 ⊞ a3 ) for all ai in A.
(ii) 0 is the neutral element.
We always think of ⊞ as a sort of addition.
We write à for {a1 ⊞ a2 : ai ∈ A}. Note that {a} = a ⊞ 0 ∈ Ã. Thus there is a natural embedding
A ֒→ Ã given by a 7→ {a}, and we can transfer the addition to P(A) by defining
{a1 } ⊞ {a2 } = a1 ⊞ a2 .
By definition, the hypermonoid is not closed under repeated addition, which makes it difficult to check
basic identical relations such as associativity.
Many hypermonoids satisfy the extra property:
Property P. a, b ∈ a ⊞ b whenever a ⊞ b is not a singleton.
A hyperinverse of an element a in a hypermonoid (A, ⊞, 0) is an element denoted as −a, for which
0 ∈ a ⊞ (−a).
A hyperzero of a hypermonoid (A, ⊞, 0) is an element of the form a ⊞ (−a) ⊆ P(A).
A hypergroup is a hypermonoid (A, ⊞, 0), satisfying the extra property:
Every element a ∈ A has a unique hyperinverse.
By [9, §2], any hypergroup satisfies the condition:
4
L. ROWEN
(Reversibility) If a ∈ b ⊞ c,
then c ∈ a ⊞ (−b).
In [22, Definition 3.1] Viro calls this a multigroup.
Definition 1.6. A hypermodule over a monoid (A, ·, 1) is a hypermonoid (M, ⊞, 0) together with an
action of A on M such that distributivity holds for M over A.
A hyperring (A, ·, ⊞, 0) is a hypermonoid (A, ⊞, 0) which also is a hypermodule over (A, ·, 1).
In other words, a hyperring is an additive hypermonoid which also is a monoid with respect to an
associative multiplication that distributes over addition; we have the two operations · on A and ⊞ : A → Ã,
with distributivity holding on the elements of A.
A hyperfield is a hyperring (A, ·, ⊞, 0), with (A, ·) a group.
[12, Definition 2.3] defines a hypermonoid morphism to be a map f : A1 → A2 of hypergroups,
satisfying f (a ⊞ b) ⊆ f (a) ⊞ f (b). This yields the category of hypergroups and their morphisms, which
matches the definition of morphism in [19].
Here are some easy instances in which associativity fails in P(A).
Example 1.7. Consider the natural max-plus algebra.
(i) Define a ⊞ b = sup{a, b}, a 6= b, and
a ⊞ a = {0, 9}.
Then each element has a unique hyperinverse, itself, and this is associative on distinct single
elements (taking their max) but (2 ⊞ 2) ⊞ 5 = {0, 9} ⊞ 5 = {5, 9} whereas 2 ⊞ (2 ⊞ 5) = 5.
(ii) Define a ⊞ b = sup{a, b}, a 6= b, and
a ⊞ a = {−a, 0, a}.
Again each element has a unique hyperinverse, itself, and now
(2 ⊞ 3) ⊞ 3 = 3 ⊞ 3 = {−3, 0, 3};
2 ⊞ (3 ⊞ 3) = 2 ⊞ {−3, 0, 3} = {2, 2, 3}.
Lemma 1.8. If (A, ·, 1) is a monoid and (A, ⊞, 0) is a hypermonoid, then A acts on à via the action
aS = {as : s ∈ S}.
(1.3)
.
Proof. (a1 a2 )S = {(a1 a2 )s : s ∈ S} = {a1 (a2 s) : s ∈ S} = a1 (a2 S).
1.4. The power set of a hyperfield.
To proceed further, we need associativity at the level of sets, and we need the following definition to
make this precise (and hopefully more manageable, since then we can do all the calculations in the power
set).
Definition 1.9. Every operator ω on A is extended to an element-compatible operator on P(A), in
the sense that
[
ωm (S1 , . . . , Sm ) = {ωm (s1 , . . . , sm ) : sk ∈ Sk , 1 ≤ k ≤ m}.
(1.4)
If (A, ⊞, 0) is a hypermonoid, then P(A) is a monoid with respect to a commutative associative binary
operation P(A) × P(A) → P(A), compatible with the operation on singletons, in the sense that
[
S1 ⊞ S2 =
{{s1 } ⊞ {s2 } : si ∈ Si }
(1.5)
à is the subset of P(A) containing all singletons and their sums. ({0} is the neutral element.)
Lemma 1.10. For any multiplicative monoid A, any invertible element of P(A) must be a singleton.
HYPERGROUPS AND HYPERFIELDS
5
Proof. If S has two elements s1 , s2 and is invertible, then S −1 contains s1 −1 and s2 −1 , implying 1 =
s1 −1 s2 , i.e., s1 = s2 .
Theorem 1.4 generalizes easily to:
Theorem 1.11. Given a hyperfield (A, ⊞, 0), we can define addition elementwise on P(A) by means of
(1.5). Then (P(A), ⊞) is a monoid, whose identity element is {0}. In this case A \ {0} (viewed as the set
of singletons) is the set of invertible elements of P(A).
Proof. We need to verify associativity, repeating the proof of Theorem 1.4, replacing + by ⊞.
[
(S1 ⊞ S2 ) ⊞ S3 = {s1 ⊞ s2 : sj ∈ Sj } ⊞ S3
[
= ((s1 ⊞ s2 ) ⊞ s3 ) : sj ∈ Sj }
[
= (s1 ⊞ (s2 ⊞ s3 )) : sj ∈ Sj }
[
= S1 ⊞ {s2 ⊞ s3 : sj ∈ Sj }
(1.6)
= S1 ⊞ (S2 ⊞ S3 ).
The set of invertible elements of P(A) must be contained in the set of singletons of P(A), which is A.
For any finite set S = {s1 , . . . , sm } ⊂ P(A), we write ⊞S for s1 ⊞ · · · ⊞ sm , which makes sense since
we already have associativity of ⊞.
Remark 1.12. Viro showed that the general transition of universal relations to P(A) is not as straightforward as it may seem. The analogous argument to Theorem 1.11 unravels for hyperrings, since distributivity does not pass from elements to sets:
(i) [22, Theorem 4.B] (a ⊞ b)(c ⊞ d) ⊆ (ac) ⊞ (ad) ⊞ (bc) ⊞ (bd) in any hyperring; the same argument
shows that (⊞S)(⊞T ) ⊆ ⊞(ST ) for any finite sets S, T ;
(ii) [22, Theorem 5.B] Recall Viro’s “triangle” hyperfield, defined over R+ by the formula
a ⊞ b = {c ∈ R+ : |a − b| ≤ c ≤ a + b}.
In other words, c ∈ a ⊞ b iff there exists an Euclidean triangle with sides of lengths a, b, and c.
The “triangle” hyperfield R does not satisfy “double distributivity,” so P(R) is not distributive.
In other words, the analog of Theorem 1.4 fails for distributivity. To overcome this setback, we need
to modify our underlying algebraic structure both at the hyper level and the power set level, the crux of
the matter being to weaken generalized distributivity on sets.
Actually, many of the important examples of hyperfields are doubly distributive, so we could pass to
(distributive) power semirings without further ado. But even in the absence of doubly distributivity, we
can formulate the weaker theory in terms of universal algebra, in order to have those techniques at our
disposal. On the face of it, this is problematic since the hypersum set could be arbitrarily large. However,
we can get around this by focusing on the monoid of singletons, and using operators instead of elements.
1.5. Pre-semirings. Motivated by the fact that a hyperring is a multiplicative monoid, we bring in the
following definition, keeping the power set in mind:
Definition 1.13. A pre-semiring† is a set (A, +, ·, 1R ) for which (A, +) is an additive Abelian semigroup and (A, ·, 1R ) is a multiplicative monoid but not necessarily satisfying the usual distributive laws.
Given a pre-semiring† A and a (distinguished) multiplicative submonoid T , an (A, T )-module is a
premodule over A that also satisfies the distributivity conditions for ri ∈ T and ai ∈ M :
(i) (r1 + r2 )a = r1 a + r2 a,
(ii) r(a1 + a2 ) = ra1 + ra2 ,
In line with [11], we call T tangible. When these conditions hold for A, we then call (A, T ) a
T -semiring† .
6
L. ROWEN
This can be described in the framework of universal algebra, with elaboration and details given in [19,
§ 5]. We define the left multiplication maps ℓr : M → M by ℓr (a) = ra, a unary operator for each r ∈ T ,
and rewrite these rules as the identical relations
(i) ℓr1 +r2 (x) = ℓr1 (x) + ℓr2 (x),
(ii) ℓr (x1 + x2 ) = ℓr (x1 ) + ℓr (x2 )
Example 1.14. For any hyperring T , taking P (T ), T ) is a T -semiring† .
1.5.1. The categorical approach.
We can improve these results slightly by working in a category with a weaker definition of morphism.
Definition 1.15. Multiplication weakly distributes over addition in a subset S ⊆ P(T ) if for all finite
S, T ⊆ S we have
(⊞S)(⊞T ) ⊆ ⊞(ST ).
(1.7)
In this case we call P(T ) a weak power semiring. (It is not a semiring.)
The restriction of this definition to hyperrings is:
Definition 1.16. Suppose (T , ·) is a monoid and (T , ⊞) a hypermonoid. Multiplication weakly distributes over ⊞ in T if for all ai , b ∈ T we have
(⊞i ai )b ⊆ ⊞i (ai b).
(1.8)
In this case we call T a weak hyperring.
Proposition 1.17. Suppose T is a hyperring. Then multiplication weakly distributes over ⊞ in P(T ).
Proof. We need to verify (1.7). But writing S = {s1 , . . . , sm } we have
[
[
⊞(si T ) ⊆ ⊞(ST )
(⊞S)(⊞T ) = (⊞si )T =
i
i
since each si T ⊆ ⊞ST.
The reverse inclusion fails since we simultaneously encounter varying si T when i varies. Thus, we are
interested in weak power semirings which, strictly speaking, are not quite semirings. So far, the power
set of a hyperring is a weak power semiring. Now we repeat the proof of [22, Theorem 4.B], to show that
at the bottom level we have not lost anything.
Theorem 1.18. Suppose T is a weak hyperring. If (T , ·) also is a group, then T is a hyperring, which
can be identified with the set of singletons of P(T ).
Proof. As in [22, Theorem 4.A], to obtain distributivity, we need to reverse the inequality (1.7) when
S = {a} is a singleton {a}, given multiplicative inverses in T . But
⊞(aT ) = aa−1 (⊞(aT )) ⊆ a(⊞a−1 aT ) = a ⊞ T.
Thus, the theory of hyperfields embeds into the theory of weak power semirings.
Definition 1.19. A weak morphism of weak power semirings is a multiplicative homomorphism f :
P(T1 ) → P(T2 ), satisfying
f (S1 ⊞ S2 ) ⊆ f (S1 ) ⊞ f (S2 ).
By induction, we have f (⊞S) ⊆ ⊞f (S) for any finite set S, and thus for arbitrary S. This is described
in universal algebra in [19, Definition 5.4]. Note that the multiplicative version yields equality, since if
f ({a})f ({b}) ⊆ f ({ab}), then f ({a})f ({b}) = f ({ab}) since they are both singletons, so f (a)f (b) = f (ab).
HYPERGROUPS AND HYPERFIELDS
7
1.6. Weak power modules.
As often is the case, one gets a deeper understanding by turning to modules. This was done in [3,
Definition 2.18], but we take a slightly weaker definition in line with our earlier categorical considerations.
Definition 1.20. Suppose that P(T ) is a weak power semiring. A weak module over P(T ) is a set M
with an element 0M such that (P(M ), ⊞M , 0M ) is a monoid with scalar multiplication T × P(M ) → P(M )
satisfying the following for all Si ⊆ T and T ⊆ M :
(i) (S1 S2 )T = S1 (S2 T );
(ii) (⊞T S)(⊞M T ) ⊆ ⊞(ST );
(iii) r0M = 0M for all r in T ;
(iv) 0R a = 0M , ∀a ∈ M.
This leads to a slight modification for hypermodules:
Definition 1.21. A weak hypermodule over a weak hyperring (R, ⊞R , 0R ) is a hypergroup (M, ⊞M , 0)
together with a binary operation R × M → P(M ) satisfying the following properties for all r, ri ∈ R and
a, aj ∈ M :
(i) (r1 r2 )a = r1 (r2 a);
(ii) (⊞i ri )(⊞j aj ) ⊆ ⊞i,j (ri aj );
(iii) r0M = 0M ;
(iv) 0R a = 0M .
As before, {⊞S : finite S ⊆ M } is a submodule of P(M ).
Definition 1.22. A weak module morphism is a map f : P(M ) → P(N ) of weak modules, satisfying
f (rS1 ) ⊆ rf (S1 );
f (S1 ⊞ S2 ) ⊆ f (S1 ) ⊞ f (S2 ),
∀Si ⊆ M.
The weak module morphism is an example of a morphism as given in [19].
Theorem 1.23. Suppose P(R) is a weak power semiring and (R, ·) also is a group. If M is a weak
module over P(R), then M is an R-module.
Proof. We repeat the proof of Theorem 1.18. First we show that f (rS) = rf (S), i.e., rf (s) = f (rs) ∈
f (rS). for all r ∈ R and s ∈ S. Indeed,
rf (s) = rf (r−1 rs) ⊆ rr−1 f (rs) = f (rs).
To obtain distributivity, we need to reverse the inclusion (1.7), given multiplicative inverses in R.
Taking S = {a} to be a singleton, we are given
⊞(aT ) = aa−1 (⊞(aT )) ⊆ a(⊞(a−1 aT )) = a ⊞ T.
Taking f to be left multiplication by an element a ∈ R, we have a(⊞S) ⊆ ⊞(aS) which is precisely
weak distributivity. In other words, M is weakly distributive over R iff left multiplication is a weak
module morphism, for every element r ∈ R. When (R, ·) is a group, M is distributive over R.
1.6.1. Negation maps.
We also want to treat negation maps from [19, §4] in this perspective. We review the definition.
Definition 1.24. A negation map on an additive semigroup (A, +) is a semigroup homomorphism
(−) : A → A of order ≤ 2, again written a 7→ (−)a.
(Thus (−)(a + b) = (−)a + (−)b.) For all other operators, including multiplication, we have a different
perspective:
Definition 1.25. A negation map (−) : A → A on an operator ωm,j (other than addition) satisfies
(−)ωm,j (a1 , . . . , au−1 , au , au+1 , . . . , am ) = ωm,j (a1 , . . . , au−1 , (−)au , au+1 , . . . , am )
for each au ∈ Au , 1 ≤ u ≤ m.
(1.9)
8
L. ROWEN
Lemma 1.26. Any negation map on A induces a negation map on P(A), via (−)S = {(−)s : s ∈ S}.
Proof. A special case of Theorem 1.4(ii), viewing the properties of the negation map as identical relations.
To see that (−)(a1 + a2 ) = (−)a1 (−)a2 , note that 0 ∈ ai (−)ai for i = 1, 2, so
0 ∈ a1 (−)a1 + a2 (−)a2 = (a1 + a2 )(−)a1 (−)a2 .
Likewise, 0 ∈ a1 (−)a1 implies 0 ∈ a(a1 (−)a1 ) ∈ aa1 (−)aa1 .
When T is a hyperfield, (A, T ) is a T -semiring† with the negation map a 7→ (−)a. In this way we have
embedded the theory of hyperfields into the theory of T -semirings† with a negation map. This is pushed
even further in [19, §7.9].
1.7. Major examples.
Let us see how all of this applies to the major examples of [3]. Since these examples are so important,
we will pay special attention to the set T̃ corresponding to a hyperring T . Although the theory presented
above formally passes to the weak power semiring P(T ), one actually gets distributivity in P(T ) when
the underlying hyperring satisfies generalized distributivity, which happens in many of the examples.
Even better, we can often identify P(T ) with a semiring which we already recognize.
Many of the “good” examples can be put in the framework of [3, Remark 2.7].
Example 1.27. Let R be a commutative semiring. Any multiplicative monoid T , together with a surjection of multiplicative monoids ϕ : R → T , has an induced hyperring structure given by the hyperaddition
law
a1 ⊞ a2 := ϕ(ϕ−1 (a1 ) + ϕ−1 (a2 )).
This extends naturally to P(T ), via
S1 ⊞ S2 := ϕ(ϕ−1 (S1 ) + ϕ−1 (S2 )).
Generalized distributivity on P(T ) and thus on T̃ , is inherited from generalized distributivity on P(R).
Explicitly, for ai ∈ S and bj ∈ T, we have
X
X
X
X
ϕ(ϕ−1 (ai bj )) ⊆ ⊞(ST ),
ϕ(ϕ−1 (ai )ϕ−1 (bj )) =
ϕ(ϕ−1 (bj )) =
ϕ(ϕ−1 (ai ))
(⊞i ai )(⊞j bj ) =
i,j
i,j
j
i
(1.10)
yielding ⊞S ⊞ T ⊆ ⊞(ST ). For the opposite direction, given ⊞i,j ai bj ∈ ⊞(ST ), we reverse (1.10) to get
X
X
X
ϕ(ϕ−1 (bj )) ∈ (⊞S)(⊞T ).
ϕ(ϕ−1 (ai ))
ϕ(ϕ−1 (ai bj )) =
i,j
i
j
Thus P(T ) is a semiring, and its theory can be embedded into semiring theory.
The complications arise when Example 1.27 is not applicable, cf. (vii), (viii) of the next example.
Example 1.28.
J
The tropical hyperfield. Define R∞ = R ∪ {−∞} and define the product a b := a + b and
(
max(a, b) if a 6= b,
a⊞b=
{c : c ≤ a} if a = b.
Thus 0 is the multiplicative identity element, −∞ is the additive identity, and we have a hyperfield
(satisfying Property P), called the tropical hyperfield.
Proposition 1.29. This is easily seen to be isomorphic (as hyperfields) to Izhakian’s extended tropical
arithmetic [10], further expounded as supertropical algebra in [11], where we identify (−∞, a] :=
{c : c ≤ a} with aν , so we have a natural hyperfield isomorphism of this tropical hyperfield with the
d
sub-semiring R
∞ of P(R∞ ), because
b : b > a;
(−∞, a] + b = (−∞, a] : b = a
(−∞, b] ∪ (b, a] = (−∞, a] : b < a.
HYPERGROUPS AND HYPERFIELDS
9
This isomorphism is as semirings. Thus Example 1.27 is applicable.
Example 1.30. The Krasner hyperfield. Let K = {0; 1} with the usual operations of Boolean algebra,
except that now 1 ⊞ 1 = {0; 1}. The Krasner hyperfield satisfies Property P. Again, this generates a subsemiring of P(K) having three elements, and is just the supertropical algebra of the monoid K, where we
identify {0; 1} with 1ν . Example 1.27 is applicable.
Example 1.31. Valuative hyperfields ([3, Example 2.12]) also are isomorphic to the extended semirings
in the sense of [11], in the same way.
Example 1.32. (Hyperfield of signs) Let S := {0, 1, −1} with the usual multiplication law and hyperaddition defined by 1 ⊞ 1 = {1}, −1 ⊞ −1 = {−1}, x ⊞ 0 = 0 ⊞ x = {x}, and 1 ⊞ −1 = −1 ⊞ 1 = {0, 1, −1}.
Then S is a hyperfield (satisfying Property P), called the hyperfield of signs.
In this case, though, we have a natural interpretation for P(S) :
(i) {+1} means “positive,” denoted as >0 .
(ii) {−1} means “negative,” denoted as <0 .
(iii) {0} means “neutral.”
(iv) {0, +1} means “non-negative,” denoted as ≥0 .
(v) {0, −1} means “non-positive,” denoted as ≤0 .
(vi) S = {−1, 0, 1} means “could be anything”
(Note that we have not denoted {−1, +1}.)
Then we have the familiar identifications:
(i) >0 +0 = >0 + ≥0 = >0 + >0 = >0 ;
(ii) ≥0 +0 = ≥0 + ≥0 = ≥0 ;
(iii) 0 + 0 = 0;
(iv) ≤0 +0 = ≤0 + ≤0 = ≤0 ;
(v) <0 +0 = <0 + ≤0 = <0 + <0 = <0 ;
(vi) >0 + <0 = ≥0 + <0 = >0 + ≤0 = ≥0 + ≤0 = S+ <0 = S + 0 = S+ >0 = S+ ≤0 =
S+ ≥0 = S.
These six elements constitute the sub-semiring S̃ of P(S).
Example 1.33. The “triangle” hyperfield A of Remark 1.12 is not doubly distributive but does satisfy
2
1
Property P since |a − b| ≤ a ≤ a + b. Here à = {[a1 , a2 ] : a1 ≤ a2 } since [a1 , a2 ] = a1 +a
+ a2 −a
∈ Ã.
2
2
b
b
◦
Any interval [0, b] is in P(A) , since [0, b] = [0, 2 ] + [0, 2 ].
Example 1.34. The phase hyperfield. Let S 1 denote the complex unit circle, and P := S 1 ∪ {0}. We say
that points a and b are antipodes if a = −b. Multiplication is defined as usual (so corresponds on S 1 to
addition of angles). We call an arc of less than 180 degrees short. The hypersum is given by
all points in the short arc from a to b if a 6= b;
a ⊞ b = {−a, 0, a} if a, b are antipodes;
{a} if b = 0.
Then P is a hyperfield (satisfying Property P), called the phase hyperfield. At the power set level, given
T1 , T2 ⊆ S 1 , one of which having at least two points, we define T1 ⊞ T2 to be the union of all (short) arcs
from a point of T1 to a non-antipodal point in T2 (which together makes a connected arc), together with
{0} if T2 contains an antipode of T1 . Note that any arc of S 1 can be obtained by taking T1 to be a single
point in the middle and T2 to be the two endpoints.
f1 of P(S 1 ) is the set of short arcs, possibly with {0} adjoined. This is not a subIn other words, S
semiringl for this we need the set of all arcs, where ⊞ is concatenation (and filling in the rest of S 1 if the
arcs go more than half way around), and adjoining {0} if the arcs contain an antipode.
Double distributivity fails, when we take a1 and a2 almost to be antipodes, b1 = a2 , and the arc
connecting b1 and b2 just passes the antipode of a1 ; then (a1 ⊞ a2 )(b1 ⊞ b2 ) is the arc from a1 to b2 , a
little more than a semicircle, whereas a1 b1 ⊞ a1 b2 ⊞ a2 b2 is already all of S 1 .
Viro [22] also has a somewhat different version, which is not distributive.
10
L. ROWEN
Example 1.35. Here is another example, suggested by Lopez, also cf. [13]. Consider R, with addition
given by a ⊞ b and b ⊞ a (for a ≤ b) to be the interval [a, b]. This extends to addition on intervals, given by
[a1 , b2 ] + [a2 , b2 ] = {min(a1 , a2 ), max(b1 , b2 )}, which clearly is associative. But the inverse is not unique,
since a + (−a) = [−a, a] contains 0, but so does a2 + a. On the other hand, this does satisfy the restriction
that every set of the form a + (−a) cannot be of the form a + (−b) for b 6= a, so if we modify the condition
of quasi-inverse to stipulate that a + (−a) must be of the form c + (−c) for some c, then it is unique.
This is essentially the general condition set forth in [19].
References
[1] M. Akian, S. Gaubert, and A. Guterman. Linear independence over tropical semirings and beyond. In Tropical and
Idempotent Mathematics, G.L. Litvinov and S.N. Sergeev, (eds.), Contemp. Math., 495:1–38, 2009.
[2] M. Akian, S. Gaubert, and A. Guterman. Tropical Cramer determinants revisited. In Tropical and idempotent mathematics and applications, Contemp. Math. 616, 1–45. Amer. Math. Soc., Providence, RI, 2014.
[3] M. Baker. Matroids over hyperfields, arXiv:1601.01204v2 [math.CO], 2016.
[4] G. Blachar and L. Rowen Symmetrized rank of matrices
[5] A. Connes and C. Consani, From monoids to hyperstructures: in search of an absolute arithmetic. In Casimir Force,
Casimir operators, and the Riemann hypothesis,, Walter de Gruyter, Berlin, 147–198, 2010.
[6] S. Gaubert, Théorie des systèmes linéaires dans les dioı̈des. PhD dissertation, School of Mines. Paris, July 1992.
[7] J. Golan, Semirings and their Applications, Springer-Science + Business, Dordrecht, 1999. (Previously published by
Kluwer Acad. Publ., 1999.)
[8] M. Gondran and M. Minoux. Graphs, dioids and semirings, volume 41 of Operations Research/Computer Science
Interfaces Series. Springer, New York, 2008.
[9] Henry
[10] Z. Izhakian. Tropical arithmetic and matrix algebra. Comm. in Algebra 37(4):1445–1468, 2009.
[11] Z. Izhakian and L. Rowen. Supertropical algebra. Adv. in Math., 225(4):2222–2286, 2010, Preprint at arXiv:0806.1175,
2007.
[12] J. Jaiung, Algebraic geometry over hyperrings. Preprint. Available at arxiv:math.AG/1512.04837, 37 pages, 2015.
[13] M. Gómez, S. López-Permouth , F. Mazariegos, A.J. Vargas De León, and R.Z. Cifuentes, Group Structures on Families
of Subsets of a Group, arXiv:1604.01119 [math.GR], 2015.
[14] A. Niv. On pseudo-inverses of matrices and their characteristic polynomials in supertropical algebra, Linear Algebra
Appl. 471: 264–290, 2015.
[15] A. Niv. Factorization of tropical matrices, J. Algebra Appl., 13(1):1350066:1–26, 2014.
[16] M. Plus. Linear systems in (max; +)-algebra. In Proceedings of the 29th Conference on Decision and Control, Honolulu,
Dec. 1990.
[17] C. Reutenauer and H. Straubing, Inversion of matrices over a commutative semiring, J. Algebra, 88, 350-360, 1984.
[18] L.H. Rowen. Graduate algebra: A noncommutative view Semigroups and Combinatorial Applications. American Mathematical Society, 2008.
[19] L.H. Rowen. Algebras with a negation map, originally entitled Symmetries in tropical algebra, arXiv:1602.00353
[math.RA], 2016.
[20] D.E. Rutherford. Inverses of Boolean matrices, Proceedings of the Glasgow Mathematical Association 6 (1963), 49–53.
[21] H. Straubing. A combinatorial proof of the Cayley-Hamilton theorem. Discrete Math., 43(2-3):273-279, 1983.
[22] O.Y. Viro. Hyperfields for Tropcial Geometry I. Hyperfields and dequantization. ArXiv AG/1006.3034, 2010.
Department of Mathematics, Bar-Ilan University, Ramat-Gan 52900, Israel
E-mail address: [email protected]
| 0math.AC
|
General Video Game AI: a Multi-Track Framework
for Evaluating Agents, Games and Content
Generation Algorithms
arXiv:1802.10363v1 [cs.AI] 28 Feb 2018
Diego Perez-Liebana, Jialin Liu, Ahmed Khalifa, Raluca D. Gaina, Julian Togelius, Simon M. Lucas
Abstract—General Video Game Playing (GVGP) aims at
designing an agent that is capable of playing multiple video
games with no human intervention. In 2014, The General Video
Game AI (GVGAI) competition framework was created and
released with the purpose of providing researchers a common
open-source and easy to use platform for testing their AI
methods with potentially infinity of games created using Video
Game Description Language (VGDL). The framework has been
expanded into several tracks during the last few years to meet the
demand of different research directions. The agents are required
to either play multiples unknown games with or without access
to game simulations, or to design new game levels or rules.
This survey paper presents the VGDL, the GVGAI framework,
existing tracks, and reviews the wide use of GVGAI framework
in research, education and competitions five years after its birth.
A future plan of framework improvements is also described.
I. I NTRODUCTION
Game-based benchmarks and competitions have been used
for testing artificial intelligence capabilities since the inception
of the research field. For the first four or five decades,
such testing was almost exclusively based on board games.
However, since the early 2000s a number of competitions
and benchmarks based on video games have sprung up. The
more well-known include the Ms Pac-Man competition [1],
the Mario AI competition [2], the Simulated Car Racing
competition [3], the Arcade Learning Environment [4] and the
StarCraft [5] competitions.
So far, most competitions and game benchmarks challenge
the agents to play a single game (an exception is the Arcade Learning Environment which contains several dozens
of games, but so far almost all published studies deal with
learning agents for one game at a time). This leads to an overspecialization, or overfitting, of agents to individual games.
This is reflected in the outcome of individual competitions
– for example, over the more than five years the Simulated
Car Racing Competition ran, submitted car controllers got
better at completing races fast, but incorporated more and more
game-specific engineering and arguably less of general AI and
machine learning algorithms. Similarly, the well-performing
bots on the StarCraft competitions are generally highly domain
– and even strategy-specific, and display very little in the way
of learning and adaptation.
It should not come as a surprise that, whenever possible,
researchers will incorporate domain knowledge about the game
they design an agent for. Yet, this trend threatens to negate the
usefulness of game-based AI competitions for spurring and
testing the development of stronger and more general AI.
The General Video Game AI competition [6] was founded
on the belief that the best way to stop AI researchers from
relying on game-specific engineering in their agents is to
make it impossible. The idea is that researchers develop their
agents without knowing what games they will be playing, and
after submitting their agents to the competition all agents are
evaluated using an unseen set of games. Every competition
event requires the design of a new set of games, as reusing
previous games would make it possible to adapt agents to
existing games.
In order to make this possible, the Video Game Description
Language (VGDL) was developed to easily create new games
to be played by agents. VGDL was designed so that it would
be easy to create games both for humans and algorithms, eventually allowing for automated generation of testbed games. A
game engine was also created to allow games specified in this
language to be visualized and played, both by humans and AI
agents, forming the basis of the competition.
While the GVGAI competition was initially focused on
benchmarking AI algorithms for playing the game, the competition and its associated software has multiple uses. In addition
to the competition tracks dedicated to game-playing agents,
there are now competition tracks focused on generating game
levels or rules. There is also the potential to use VGDL and
GVGAI as a game prototyping system, and there is a rapidly
growing body of research using this framework for everything
from building mixed-initiative design tools to demonstrating
new concepts in game design.
This paper gives an overview of the various tracks of the
GVGAI competition, the GVGAI software, and the various
uses the software has been put to. We start with describing
the basics of VGDL and the GVGAI framework. We then
describe the various competition tracks that have launched so
far: planning (single-player and two-player), level generation,
rule generation and learning. The next few sections are devoted
to discussing the various kinds of AI methods that have been
used in the submissions to each track. Especial consideration
is given to the single-player planning track, as it has existed
for longest and received the most submissions up to date.
This is followed by a section cataloging some of the noncompetition research uses of the GVGAI software. The final
few sections look forward to the future use and development
of the framework and competition: how it can be used in
teaching, open research problems (specifically related to the
planning tracks), and the future evolution of the competition
and framework itself.
II. VGDL AND
THE
GVGAI F RAMEWORK
The idea of Video Game Description Language (VGDL)
was initially proposed by Ebner et al. [7] at the Dagstuhl Seminar on Artificial and Computational Intelligence in Games,
during which the first VGDL game, Space Invaders, was
created. Then, Tom Schaul continued this work by completing
and improving the language in a Python framework for modelbased learning and released the first game engine in 2013 [8],
[9]. VGDL is a text description language that allows for
the definition of two-dimensional, arcade, single-player, gridbased physics and (generally) stochastic games and levels.
With an ontology defined in the framework, VGDL permits
the description of objects (sprites), their properties and the
consequences of them colliding with each other. Additionally,
termination conditions and reward signals (in form of game
scores) can be defined for these games. In total, four different
sets are defined: Sprite Set (which defines what kind of sprites
take part in the game), Interaction Set (rules that govern the
effects of two sprites colliding with each other), Termination
Set (which defines how does the game end) and Mapping Set
(that specifies which characters in the level file map to which
sprites from the Sprite set).
M. Ebner et al. [7] and J. Levine et al. [10] described,
in their Dagstuhl 2013 follow-up, the need and interest for
such a framework that could accommodate a competition for
researchers to tackle the challenge of General Video Game
Playing (GVGP). D. Perez et al. [6] implemented a version of
T. Schaul’s initial framework in Java and organized the first
General Video Game AI (GVGAI) competition in 2014 [11].
In the following years, this framework was extended to accommodate two-player games [12], [13], level [14], rule [15]
generation, and real-world physics games [16].
This framework provided an API for creating bots that
would be able to play in any game defined in VGDL, without
giving them access to the rules of the games nor to the
behaviour of other entities defined in the game. However, the
agents do have access to a Forward Model (in order to roll
the current state forward given an action) during the thinking
time, set to 40ms in the competition settings. Controllers also
had access at every game tick to game state variables, such
game status - winner, time step and score -, state of the player
(also referred to in this paper as avatar) - position, orientation,
resources, health points -, history of collisions and positions of
the different sprites in the game identified with a unique type
id. Additionally, sprites are grouped in categories attending to
their general behaviour: Non-Player Characters (NPC), static,
movable, portals (which spawn other sprites in the game, or
behave as entry or exit point in the levels) and resources (that
can be collected by the player).
All this information is also provided to agents for the
learning setting of the framework and competition, with the
exception of the Forward Model. In its last (to date) modification of the framework, GVGAI included compatibility
for creating learning agents, both in Java and in Python,
which would learn to play any game is given in an episodic
manner [17]. At the moment of writing, the framework counts
on 120 single-player and 60 two-player games.
Contest Leg
CIG-14
GECCO-15
CIG-15
CEEC-15
GECCO-16
CIG-16
WCCI-16 (2P)
CIG-16 (2P)
GECCO-17
CEC-17 (2P)
Winner
OLETS
Type
Tree Search Method
Section
IV-B [11]
YOLOBOT
Return42
YBCriber
Hyper-heuristic
Hyper-heuristic
Hybrid
IV-E [18]
IV-E [19]
IV-D [20]
YOLOBOT
MaastCTS2
ToVo2
Number27
Hyper-heuristic
Tree Search Method
Hybrid
Hybrid
IV-E [18]
IV-B [21]
V-A [13]
V-B [13]
YOLOBOT
ToVo2
Hyper-heuristic
IV-E [18]
Hybrid
V-A [13]
TABLE I
W INNERS OF ALL EDITIONS OF THE GVGAI P LANNING COMPETITION .
2P INDICATES 2-P LAYER TRACK . H YBRID DENOTES 2 OR MORE
TECHNIQUES COMBINED IN A SINGLE ALGORITHM . H YPER - HEURISTIC
HAS A HIGH LEVEL DECISION MAKER TO DECIDES WHICH SUB - AGENT
MUST PLAY ( SEE S ECTION IV). TABLE EXTENDED FROM [19].
III. GVGAI C OMPETITION
TRACKS
This section summarizes the different tracks featured in the
GVGAI competition.
A. Single-player planning
The first competition track is Single-player Planning [11],
in which one agent (also referred to as bot or controller) plays
single-player games with the aid of the Forward Model. Each
controller has 1 second for initialization and 40ms at each
game tick as decision time.
All GVGAI tracks follow a similar structure: there are
several public sets of games, included in the framework,
allowing the participants to train their agents on them. For
each edition, there is one validation and one test set. Both
sets are private and stored in the competition server1 . There
are 10 games on each set, with 5 different levels each.
Participants can submit their entries any time before the
submission deadline to all training and validation sets, and
preliminary rankings are displayed in the competition website
(the names of the validation set games are anonymized). All
controllers are run on the test set after the submission deadline
to determine the final rankings of the competition, executing
each agent 5 times on each level.
Rankings are computed by first sorting all entries per game
according to victory rates, scores and time steps, in this order.
These per-game rankings award points to the first 10 entries,
from first to tenth position: 25, 18, 15, 12, 10, 8, 6, 4, 2 and 1.
The winner of the competition is the submission that sums
more points across all games in the test set. For a more
detailed description of the competition and its rules, the reader
is referred to [11].
Table I shows the winners of all editions up to date, along
with the section of this survey in which the method is included
and the paper that describes the approach more in depth.
B. Two-player planning
The Two-player Planning track [12] was added in 2016,
with the aim of testing general AI agents in environments
1 www.gvgai.net;
Intel Core i5 machine, 2.90GHz, and 4GB of memory.
which are more complex and present more direct player
interaction. Games in this setting are played by two-players in
a simultaneous move fashion. The rules of the games included
in the competition are still secret to the players, similar to
the Single-player track, but an additional piece of information
is hidden: whether the game is competitive or cooperative.
Having both types of games ensures that agents are not tuned
to only compete against their opponents, instead having to
correctly judge various situations and identify what the goal
of the other intelligent presence in the game is. However, this
question of an appropriate opponent model remains open, as
all competition entries so far have employed random or very
simple techniques.
Two legs of this track were organized for the first time in
2016, at the IEEE World Congress on Computational Intelligence (WCCI) and the IEEE Conference on Computational
Intelligence and Games (CIG). Although ToVo2 won the first
leg and Number27 the second, the winner (adrienctx) and
runner-up (MaastCTS2) of the overall championship showed
that a general agent may not be the best at a particular
subset of problems, but will perform at a high level on many
different ones. In 2017, ToVo2 did win the overall competition,
organized at the IEEE Congress on Evolutionary Computation
(CEC). The final results of the competition can be seen in
Table I. For more details, the reader is referred to [13].
C. Learning track
The GVGAI Single-Player learning track started in 2017.
Among the GVGAI tracks, this is the only one which accepts
submission of agent written not only in Java, but also in
Python, in order to accommodate for popular machine learning
libraries written in this language. The Single-Player Learning
track is briefly introduced below. More technical details of
the framework and competition procedure are explained in the
track technical manual [17].
The learning track is based on the GVGAI framework, but
a key difference is that no forward model is provided to
the controllers. Hence, learning needs to be achieved by an
episodic repetition of games played. Note that, even while no
forward model is accessible during the learning and validation
phases, controllers receive an observation of the current game
state via a StateObservation object, which is provided as a
Gson object or as a screen-shot of the game screen (in png
format). The agent is free to select either or both forms
of game state observation at any game tick. Similar to the
planning tracks, controllers (either in Java or in Python) inherit
from an abstract class and a constructor and three methods can
be implemented: INIT (called at the beginning of every game,
must finish in no more than 1s of CPU time), ACT (called at
every game tick, determines the next action of the controller
in 40ms of CPU time) and RESULT (called at the end of every
game, with no time limit to allow for processing the outcome
and perform some learning).
As in the planning track, games sets count on 10 games
with 5 levels each. The execution in a given set is split in two
phases: a learning phase, which allows training in the first 3
levels of each game, and a validation phase, which uses the
Agent
kkunan
sampleRandom†
DontUnderestimateUchiha
sampleLearner†
ercumentilhan
YOLOBOT
Training set
Score
Ranking
125
6
154
2
149
3
149
4
179
1
132
5
TABLE II
Test set
Score
Ranking
184
1
178
2
158
3
152
4
134
5
112
6
TABLE SHOWS THE SCORE AND RANKING OF THE SUBMITTED LEARNING
AGENTS HAVE OBTAINED ON THE TRAINING AND TEST SET. † DENOTES A
SAMPLE CONTROLLER .
1-P Planning
1-P Learning
Best score
Best score
Agent
G2
109.00 ± 38.19
31.5 ± 14.65
sampleRandom†
G3
1.00 ± 0.00
0±0
*
G4
1.00 ± 0.00
0.2 ± 0.09
kkunan
G5
216.00 ± 24.00
1±0
*
5.60 ± 0.78
3.45± 0.44
DontUnderestimateUchiha
G6
G7 31696.10 ± 6975.78 29371.95±2296.91
kkunan
G8
1116.90 ± 660.84
35.15±8.48
kkunan
G9
1.00 ± 0.00
0.05± 0.05
sampleRandom†
56.70 ± 25.23
2.75 ± 2.04
sampleLearner†
G10
TABLE III
TABLE COMPARES THE BEST SCORES BY SINGLE - PLAYER PLANNING AND
LEARNING AGENTS ON THE SAME TEST SET. N OTE THAT ONE OF THE
Game
GAMES IN THE TEST SET IS REMOVED FROM THE FINAL RANKING DUE TO
BUGS IN THE GAME ITSELF. † DENOTES A SAMPLE CONTROLLER .
other 2 available levels. The execution of a controller in a set
has therefore 2 phases:
a) Learning phase: In each of the games, each controller
has a limited amount of time, 5 minutes, for learning the first
3 levels. It will play each of the 3 levels once then be free to
choose the next level to play if there is time remaining. The
controller will be able to send an action called ABORT to finish
the game at any moment. The method RESULT is called at the
end of every game regardless if the game has finished normally
or has been aborted by the controller. Thus, the controller can
play as many games as desired, potentially choosing the levels
to play in, as long as it respects the 5 minutes time limit.
b) Validation phase: Immediately after learning phase,
the controller plays 10 times the levels 4 and 5 sequentially.
There is no more total time limit, but the agent respects the
time limits for init, act and result, and can continue learning
during the game playing.
The first GVGAI single-player learning track was held in the
IEEE CIG 2017. Besides two sample random agents written in
Java and Python and one sample agent using Sarsa [22] written
in Java, there were three submissions written in Java and one
in Python [23]. Table II illustrates the score and ranking of
learning agents on the training and test sets. Table III compares
the score achieved by the best planning agent in 2017 on each
of the games in the test set and the score achieved by the best
learning agent.
D. Level Generation
The Level Generation track [14] started in 2016. The aim
of the track is to allow competitors to develop a general level
generation algorithm. To participate in this track, competitors
must implement their own level generator. Competitors have to
provide at least one function that is responsible for generating
the level. The framework provides the generator with all the
information needed about the game such as game sprites,
interaction set, termination conditions and level mapping. Additionally, the framework also provides access to the Forward
Model, in order to allow testing the generated levels via agent
simulation. The levels are generated in the form of 2d matrix
of characters, with each character representing the game sprites
at the specific location determined by the matrix. Competitors
have the choice to either use the provided level mapping or
provide their own one.
The first level generation competition was held at the International Joint Conference on Artificial Intelligence (IJCAI)
in 2016, which received four participants. Each participant
was provided a month to submit a new level generator. Three
different level generators were provided in order to help
the users get started with the system (see Section VII for
a description of these). Three out of the four participants
were simulation-based level generators while the remaining
was based on cellular automata. During the competition day,
people attending the conference were encouraged to try pairs
of generated levels and select which level they liked (one,
both, or neither). Finally, the winner was selected based on
the generator with more votes. The winner of the contest
was the Easablade generator, a cellular automata described
in Section VII-A4.
The competition was run again during IEEE CIG 2017 with
same configuration from previous year (one month for implementation followed by on site judging). Unfortunately, only
one submission was received, hence the the competition was
canceled. This submission used a n-gram model to generate
new constrained levels using a recorded player keystrokes.
E. Rule Generation
The Rule Generation track [15] was introduced and held
during CIG 2017. The aim of the track is to generate the
interaction set and termination conditions for a certain level
with a fixed set of game sprites in a fixed amount of time.
To participate in the track, competitors have to provide their
own rule generator. The framework provides the competitors
with game sprites, a certain level, and a forward model to
simulate generated games, as in the Level Generation case.
The generated games are represented as two arrays of strings.
The first array contains the interaction set, while the second
array contains the termination conditions.
The first rule generation competition was intended to be
run at CIG 2017. Three different sample rule generators were
provided (as discussed in Section VIII) and the contest ran
over a month’s period. Unfortunately, no submissions were
received for this track.
IV. M ETHODS FOR S INGLE P LAYER P LANNING
This section describes the different methods that have been
implemented for Single Player Planning in GVGAI. All the
controllers that face this challenge have in common the possibility of using the forward model to sample future states from
the current game state, plus the fact that they have a limited
action-decision time. While most attempts abide by the 40ms
decision time imposed by the competition, other efforts in the
literature compel their agents to obey a maximum number of
uses of the forward model.
Section IV-A briefly introduces the most basic methods
that can be found within the framework. Then Section IV-B
describes the different tree search methods that have been
implemented for this settings by the community, followed by
Evolutionary Methods in Section IV-C. Often, more than one
method is combined into the algorithm, which gives place to
Hybrid methods (Section IV-D) or Hyper-heurisric algorithms
(Section IV-E). Further discussion on these methods and their
common take-aways has been included in Section XI.
A. Basic Methods
The GVGAI framework contains several agents aimed at
demonstrating how a controller can be created for the singleplayer planning track of the competition [11]. Therefore, these
methods are not particularly strong.
The simplest of all methods is, without much doubt, doNothing. This agent returns the action NIL at every game tick without exception. The next agent in complexity is sampleRandom,
which returns a random action at each game tick. Finally,
onesteplookahead is another sample controller that rolls the
model forward for each one of the available actions in order
to select the one with the highest action value, determined
by a function that tries to maximize score while minimizing
distances to NPCs and portals.
B. Tree Search Methods
One of the strongest and influential sample controllers
is sampleMCTS, which implements the Monte Carlo Tree
Search [24] algorithm for real-time games. Initially implemented in a closed loop version (the states visited are stored
in the tree node, without requiring the use of the Forward
Model during the tree policy phase of MCTS), it achieved the
third position (out of 18 participants) in the first edition of the
competition.
The winner of that edition, Adrien Couëtoux, implemented
Open Loop Expectimax Tree Search (OLETS), which is an
open loop (states visited are never stored in the associated tree
node) version of MCTS which does not include rollouts and
uses Open Loop Expectimax (OLE) for the tree policy. OLE
substitutes the empirical average reward by rM , a weighted
sum of the empirical average of rewards and the maximum of
its children rM values [11].
Torsten Schuster, in his MSc thesis [25], analyzes several
enhancements and variations for MCTS in different sets of
the GVGAI framework. These modifications included different tree selection, expansion and play-out policies. Results
show that combinations of Move-Average Sampling Technique
(MAST) and N-Gram Selection Technique (NST) with Progressive History provided an overall higher rate of victories
than their counterparts without these enhancements, although
this result was not consistent across all games (with some
simpler algorithms achieving similar results).
In a different study, Dennis Soemers [21], [26] explored
multiple enhancements for MCTS: Progressive History (PH)
and NST for the tree selection and play-out steps, tree reuse (by starting at each game tick with the subtree grown in
the previous frame that corresponds to the action taken, rather
than a new root node), bread-first tree initialization (direct
successors of the root note are explored before MCTS starts),
safety pre-pruning (prune those nodes with high number of
game loses found), loss avoidance (MCTS ignores game lose
states when found for the first time by choosing a better alternative), novelty-based pruning (in which states with features
rarely seen are less likely to be pruned), knowledge based
evaluation [27] and deterministic game detection. The authors
experimented with all these enhancements in 60 games of the
framework, showing that most of them improved the performance of MCTS significantly and their all-in-one combination
increased the average win rate of the sample agent in 17
percentage points. The best configuration was winner of one
of the editions of the 2016 competitions (see Table I).
F. Frydenberg studied yet another set of enhancements
for MCTS. The authors showed that using MixMax backups
(weighing average and maximum rewards on each node)
improved the performance in only some games, but its combination with reversal penalty (to penalize visiting the same
location twice in a play-out) offers better results than vanilla
MCTS. Other enhancements, such as macro-actions (by repeating an action several times in a sequence) and partial
expansion (a child node is considered expanded only if its
children have also been expanded) did not improve the results
obtained.
Perez et al. [27] implemented KB-MCTS, a version of
MCTS with two main enhancements. First, distance to different sprites were considered features for a linear combination, where the weights were evolved to bias the MCTS
rollouts. Secondly, a Knowledge Base (KB) is kept about
how interesting for the player the different sprites are, where
interesting is a measure of curiosity (rollouts are biased
towards unknown sprites) and experience (a positive/negative
bias for getting closer/farther to beneficial/harmful entities).
The results of applying this algorithm to the first set of games
of the framework showed that the combination of these two
components gave a boost in performance in most games of
the first training set.
The work in [27] has been extended by other researchers in
the field, which also put a special effort on biasing the Monte
Carlo (MC) simulations. In [28], the authors modified the
random action selection in MCTS rollouts by using potential
fields, which bias the rollouts by making the agent move in
a direction akin to the field. The authors showed that KBMCTS provides a better performance if this potential field
is used instead of the Euclidean distance between sprites
implemented in [27]. Additionally, in a similar study [29],
the authors substituted the Euclidean distance for a measure
calculated by a pathfinding algorithm. This addition achieved
some improvements over the original KB-MCTS, although the
authors noted in their study that using pathfinding does not
provide a competitive advantage in all games.
Another work by H. Park and K. Kim [30] tackles this
challenge by a) determining the goodness of the other sprites
in the game; b) computing an Influence Map (IM) based on
this; and c) using the IM to bias the simulations, in this
occasion by adding a third term to the Upper Confidence
Bound (UCB) [31] equation for the tree policy of MCTS. Although not compared with KB-MCTS, the resultant algorithm
improves the performance of the sample controllers in several
games of the framework, albeit performing worse than these
in some of the games used in the study.
Biasing rollouts is also attempted by E. dos Santos et
al. [32], who introduced Redundant Action Avoidance (RAA)
and Non-Defeat Policy (NDP); RAA analyzes changes in
the state to avoid selecting sequences of actions that do not
produce any alteration on position, orientation, properties or
new sprites in the avatar. NDP makes the recommendation
policy ignore all children of the root node who found at least
one game loss in a simulation from that state. If all children
are marked with a defeat, normal (higher number of visits)
recommendation is followed. Again, both modifications are
able to improve the performance of MCTS in some of the
games, but not in all.
M. de Waard et al. [33] introduced the concept of options
or macro-actions in GVGAI. Each option is associated with
a goal, a policy and a termination condition. The selection
and expansion steps in MCTS are modified so the search
tree branches only if an option is finished, allowing for a
deeper search in the same amount of time. Their results show
that Option MCTS outperforms MCTS in games with small
levels or a few number of sprites, but loses in the comparison
to MCTS when the games are bigger due to these options
becoming too large.
In a similar line, Perez et al. [16] employed macro-actions
for GVGAI games that used continuous (rather than gridbased) physics. These games have a larger state space, which
in turn delays the effects of the player’s actions and modifies
the way agents navigate through the level. Macro-actions are
defined as a sequence or repetition of the same action during
M steps, which is arguably the simplest kind of macro-actions
that can be devised. MCTS performed better without macroactions on average across games, but there are particular games
where MCTS needs macro-actions to avoid losing at every
attempt. The authors also concluded that the length M of
the macro-actions impacts different games distinctly, although
shorter ones seem to provide better results than larger ones,
probably due to a more fine control in the movement of the
agents.
Some studies have brought multi-objective optimization to
this challenge. For instance, Perez et al. [34] implemented
a Multi-Objective version of MCTS, concretely maximizing
score and level exploration simultaneously. In the games
tested, the rate of victories grew from 32.24% (normal MCTS)
to 42.38% in the multi-objective version, showing great
promise for this approach. In a different study, A. Khalifa et
al. [35] apply multi-objective concepts to evolving parameters
for a tree selection confidence bounds equation. A previous
work by I. Bravi [36] (also discussed later in Section IV-D)
provided multiple UCB equations for different games. The
work in [35] evolved, using S-Metric Selection Evolutionary
Multi-objective Optimization Algorithm (SMS-EMOA), the
linear weights of an UCB equation that results of combining
all from [36] in a single one. All these components respond
to different and conflicting objectives, and their results show
that it is possible to find good solutions for the games tested.
A significant exception to MCTS with regards to tree search
methods for GVGAI is that of T. and H. Geffner [20] (winner
of one of the editions of the 2015 competition, YBCriber, as
indicated in Table I), who implemented Iterated Width (IW;
concretely IW(1)). IW(1) is a breadth-first search with a crucial
alteration: a new state found during search is pruned if it does
not make true a new tuple of at most 1 atom, where atoms
are boolean variables that refer to position (and orientations
in the case of avatars) changes of certain sprites at specific
locations. The authors found that IW(1) performed better than
MCTS in many games, with the exception of puzzles, where
IW(2) (pruning according to pairs of atoms) showed better
performance. This agent was declared winner in the CEEC
2015 edition of the Single-player planning track [6].
A. Badabi [37] implemented several versions of Enforced
Hill Climbing (EHC), a breadth-first search method that looks
for a successor of the current state with a better heuristic value.
EHC obtained similar results to KB-MCTS in the first set of
games of the framework, with a few disparities in specific
games of the set.
Finally, Mark J. Nelson [38] ran a study on MCTS in order
to investigate if, giving a higher time budget to the algorithm
(i.e. increasing the number of iterations), MCTS was able to
master most of the games. In other words, if the real-time
nature of the GVGAI framework and competition is the reason
why different approaches fail to achieve a high victory rate.
This study provided up to 30 times more budget to the agent,
but the performance of MCTS only increased marginally even
at that level. In fact, this improvement was achieved by means
of losing less often rather than by winning more games. This
paper concludes that the real-time aspect is not the only factor
in the challenge, but also the diversity in the games. In other
words, increasing the computational budget is not the answer
to the problem GVGAI poses, at least for MCTS.
C. Evolutionary Methods
The second big group of algorithms used for single-player
planning is that of evolutionary algorithms (EA). Concretely,
the use of EAs for this real-time problem is mostly implemented in the form of Rolling Horizon EAs (RHEA). This
family of algorithms evolves sequences of actions with the
use of the forward model. Each sequence is an individual of
an EA which fitness is the value of the state found at the end
of the sequence. Once the time budget is up, the first action of
the sequence with the highest fitness is chosen to be applied
in that time step.
The GVGAI competition includes SampleRHEA as a sample
controller. SampleRHEA has a population size of 10, individual
length of 10 and implements uniform crossover and mutation,
where one action in the sequence is changed for another one
(position and new action chosen uniformly at random) [11].
R. Gaina et al. [39] analyzed the effects of the RHEA
parameters on the performance of the algorithm in 20 games,
chosen among the existent ones in order to have a representative set of all games in the framework. The parameters
analyzed were population size and individual length, and
results showed that higher values for both parameters provided
higher victory rates. This study motivated the inclusion of
Random Search (SampleRS) as a sample in the framework,
which is equivalent to RHEA but with an infinite population
size (i.e. only one generation is evaluated until budget is
consumed) and achieves better results than RHEA in some
games. [39] also compared RHEA with MCTS, showing better
performance for an individual length of 10 and high population
sizes.
A different Evolutionary Computation agent was proposed
by B. Jia et al. [40], [41], which consists of a Genetic
Programming (GP) approach. The authors extract features
from a screen capture of the game, such as avatar location and
the positions and distances to the nearest object of each type.
These features are inputs to a GP system that, using arithmetic
operands as nodes, determines the action to execute as a result
of three trees (horizontal, vertical and action use). The authors
report that all the different variations of the inputs provided to
the GP algorithm give similar results to those of MCTS, on
the three games tested in their study.
D. Hybrids
The previous studies feature techniques in which one technique is predominant in the agent created, albeit they may
include enhancements which can place them in the boundary
of hybrids. This section describes those approaches that, in the
opinion of the authors, would in their own right be considered
as techniques that mix more than one approach in the same,
single algorithm.
An example of one of these approaches is presented by R.
Gaina et al. [42] analyzed the effects of seeding the initial
population of the RHEA using different methods. Part of
the decision time budget is dedicated to initialize a population with sequences that are promising, as determined by
onesteplookahead and MCTS. Results show that both seeding
options provide a boost in victory rate when population size
and individual length are small, but the benefits vanish when
these parameters are large.
Other enhancements for RHEA proposed in [43] are incorporating a bandit-based mutation, a statistical tree, a shift
buffer and rollouts at the end of the sequences. The banditbased mutation breaks the uniformity of the random mutations
in order to choose new values according to suggestions given
by a uni-variate armed bandit. However, the authors reported
that no improvement on performance was noticed. A statistical
tree, previously introduced in [44], keeps a game tree with
visit count and accumulated rewards in the root node, which
are subsequently used for recommending the action to take
in that time step. This enhancement produced better results
with shorter individuals and smaller population sizes. The
shift buffer enhancement provided the best improvement in
performance, which consist of shifting the sequences of the
individuals of the population one action to the left, removing
the action from the previous time step. This variation, similar
to keeping the tree between frames in MCTS, combined with
the addition of rollouts at the end of the sequences provided
an improvement victory rate (20 percentile points over vanilla
RHEA) and scores.
A similar (and previous) study was conducted by H. Horn et
al. [45]. In particular, this study features RHEA with rollouts
(as in [43]), RHEA with MCTS for alternative actions (where
MCTS can determine any action with the exception of the one
recommended by RHEA), RHEA with rollouts and sequence
planning (same approach as the shift buffer in [43]), RHEA
with rollouts and occlusion detection (which removes not
needed actions in a sequence that reaches a reward) and
RHEA with rollouts and NPC attitude check (which rewards
sequences in terms of proximity to sprites that provide a
positive or negative reward). Results show that RHEA with
rollouts improved performance in many games, although all
the other variants and additions performed worse than the
sample agents. It is interesting to see that in this case the
shift buffer did not provide an improvement in the victory
rate, although this may be due to the use of different games.
Torsten Schuster [25] proposes two methods that combine
MCTS with evolution. One of them, (1+1)−EA as proposed
by [27], evolves a vector of weights for a set of game features
in order to bias the rollouts towards more interesting parts
of the search space. Each rollout becomes an evaluation for
an individual (weight vector), using the value of the final
state as fitness. The second algorithm is based on stronglytyped GP (STGP) and uses game features to evolve state
evaluation functions that are embedded within MCTS. These
two approaches join MAST and NST (see Section IV-B) in
a larger comparison, and the study concludes that different
algorithms outperform others in distinct games, without an
overall winner in terms of superior victory rate, although
superior to vanilla MCTS in most cases.
The idea of evolving weight vectors for game features
during the MCTS rollouts introduced in [27] (KB-MCTS2 )
was explored further by Jim van Eeden in his MSc thesis [46].
In particular, the author added A* as a pathfinding algorithm
to replace the euclidean distance used in KB-MCTS for a more
accurate measure and changing the evolutionary approach.
While KB-MCTS used a weight for each pair feature-action,
being the action chosen at each step by the Softmax equation,
this work combines all move actions on a single weight and
picks the action using Gibbs sampling. The author concludes
that the improvements achieved by these modifications are
marginal, and likely due to the inclusion of pathfinding.
Additional improvements on KB-MCTS are proposed by
C. Chu et al. [47]. The authors replace the Euclidean distance
feature to sprites with a grid view of the agent’s surroundings,
and also the (1 + 1) − EA with a Q-Learning approach to bias
the MCTS rollouts, making the algorithm update the weights at
each step in the rollout. The proposed modifications improved
the victory rate in several sets of games of the framework
and also achieved the highest average victory rate among the
algorithms it was compared with.
2 This approach could also be considered an hybrid. Given its influence in
other tree approaches, it has also been partially described in Section IV-B
E. İlhan and A. Etaner-Uyar [48] implemented a combination of MCTS and true online Sarsa (λ) [49]. The authors
use MCTS rollouts as episodes of past experience, executing
true online Sarsa with each iteration with a ǫ-greedy selection
policy. Weights are learnt for features taken as the smallest
euclidean distance to sprites of each type. Results showed that
the proposed approached improved the performance on vanilla
MCTS in the majority of the 10 games used in the study.
Evolution and MCTS have also been combined in different
ways. In one of them, I. Bravi et al. [50] used a GP system
to evolve different tree policies for MCTS. Concretely, the
authors evolve a different policy for each one of the (five)
games employed in the study, aiming to exploit the characteristics of each game in particular. The results showed that the
tree policy plays a very important role on the performance of
the MCTS agent, although in most cases the performance is
poor - none of the evolved heuristics performed better than
the default UCB in MCTS.
Finally, C. Sironi et al. [51] designed three Self-Adaptive
MCTS (SA-MCTS) that tuned the parameters of MCTS (playout depth and exploration factor) on-line, using Naive MonteCarlo, an Evolutionary Algorithm (λ, µ) and the N-Tuple
Bandit Evolutionary Algorithm [52]. Results show that all
tuning algorithms improve the performance of MCTS there
where vanilla MCTS performs poorly, while keeping a similar
rate of victories in those where MCTS performs well.
E. Hyper-heuristics / algorithm selection
Several authors have also proposed agents that use several
algorithms, but rather than combining them into a single one,
there is a higher level decision process that determines which
one of them should be used at each time.
Blaine Ross, in his MSc thesis [53] proposes an agent that
is a combination of two methods. This approach uses A* with
Enforced Hill Climbing to navigate through the game at a
high level and switches to MCTS when in close proximity
to the goal. The work highlights the problems of computing
paths in the short time budget allowed, but indicate that goal
targeting with path-finding combined with local maneuvering
using MCTS does provide good performance in some of the
games tested.
T. Joppen et al. implemented YOLOBOT [18], arguably the
most successful agent for GVGAI up to date, as it has won
several editions of the competition. Their approach consists
of a combination of two methods: a heuristic Best First
Search (BFS) for deterministic environments and MCTS for
stochastic games. Initially, the algorithm employs BFS until
the game is deemed stochastic, an optimal solution is found
or a certain game tick threshold is reached, extending through
several consecutive frames if needed for the search. Unless the
optimal sequence of actions is found, the agent will execute
an enhanced MCTS consistent of informed priors and rollout
policies, backtracking, early cutoffs and pruning. The resultant
agent has shown consistently a good level of play in multiple
game sets of the framework.
Another hyper-heuristic approach, also winner of one of
the 2015 editions of the competition (Return42, see Table I),
determines first if the game is deterministic or stochastic. In
case of the former, A* is used to direct the agent to sprites of
interest. Otherwise, random walks are employed to navigate
through the level [19].
The fact that this type of portfolio agents has shown very
promising results has triggered more research into hyperheuristics and game classification. The work by P. Bontrager
et al. [54] used K-means to cluster games and algorithms
attending to game features derived from the type of sprites
declared in the VGDL description files. The resulting classification seemed to follow a difficulty pattern, with 4 clusters
that grouped games that were won by the agents at different
rates.
Mendes et al. [55] built a hyper-agent which selected
automatically an agent from a portfolio of agents for playing
individual game and tested it on the GVGAI framework. This
approached employed game-based features to train different
classifiers (Support Vector Machines - SVM, Multi-layer Perceptrons, Decision Trees - J48, among others) in order to select
which agent should be used for playing each game. Results
show that the SVM and J48 hyper-heuristics obtained a higher
victory rate than the single agents separately.
H. Horn et al. [45] (described before in Section IV-D) also
includes an analysis on game features and difficulty estimation.
The authors suggest that the multiple enhancements that are
constantly attempted in many algorithms could potentially be
switched on and off depending on the game that is being
played, with the objective of dynamically adapting to the
present circumstances.
D. Ashlock et al. [19] suggest the possibility of creating a
classification of games, based on the performance of multiple
agents (and their variations: different enhancements, heuristics,
objectives) on them. Furthermore, this classification needs
to be stable, in order to accommodate the ever-increasing
collection of games within the GVGAI framework, but also
flexible enough to allow an hyper-heuristic algorithm to choose
the version that better adapts to unseen games.
V. M ETHODS FOR T WO -P LAYER P LANNING
This section approaches agents developed by researchers
within the Two-Player Planning setting. Most of these entries
have been submitted to the Two-Player Planning track of the
competition [12]. Two methods stood out as the base of most
entries received so far, Monte Carlo Tree Search (MCTS) and
Evolutionary Algorithms (EA) [13]. On the one hand, MCTS
performed better in cooperative games, as well as showing the
ability to adapt better to asymmetric games, which involved a
role switch between matches in the same environment. EAs,
on the other hand, excelled in games with long lookaheads,
such as puzzle games, which rely on a specific sequence of
moves being identified.
Counterparts of the basic methods described in Section IV-A
are available in the framework for the Two-Player track as
well, the only difference being in the One Step Lookahead
agent which requires an action to be supplied for the opponent
when simulating game states. The opponent model used by the
sample agent assumes they will perform a random move (with
the exception of those actions that would cause a loss of the
game).
A. Tree Search methods
Most of the competition entries in the 2016 season were
based on MCTS (see Section IV-B).
Some entries employed an Open Loop version, which would
only store statistics in the nodes of the trees and not game
states, therefore needing to simulate through the actions at
each iteration for a potentially more accurate evaluation of the
possible game states. Due to this being unnecessarily costly
in deterministic games, some entries such as MaasCTS2 and
YOLOBOT switched to Breadth-First Search in such games
after an initial analysis of the game type, a method which has
shown ability to finding the optimal solution if the game lasts
long enough.
Enhancements brought to MCTS include generating value
maps, either regarding physical positions in the level, or
higher-level concepts (such as higher values being assigned to
states where the agent is closer to objects it hasn’t interacted
with before; or interesting targets as determined by controllerspecific heuristics). The winner of the 2016 WCCI leg, ToVo2,
also employed dynamic Monte Carlo roll-out length adjustments (increased with the number of iterations to encourage
further lookahead if budget allows) and weighted roll-outs (the
weights per action generated randomly at the beginning of
each roll-out).
All agents use online learning in one way or another
(the simplest form being the base Monte Carlo Tree Search
backups, used to gather statistics about each action through
multiple simulations), but only the overall 2016 Championship
winner, adrienctx, uses offline learning on the training set
supplied to tune the parameters in the Stochastic Gradient
Descent function employed, learning rate and mini batch size.
B. Evolutionary methods
Two of the 2016 competition entries used an EA technique as a base as an alternative to MCTS: Number27 and
CatLinux [13].
Number27 was the winner of the CIG 2016 leg, the
controller placing 4th overall in the 2016 Championship.
Number27 uses a Genetic Algorithm, with one population
containing individuals which represent fixed-length action
sequences. The main improvement it features on top of the
base method is the generation of a value heat-map, used to
encourage the agent’s exploration towards interesting parts of
the level. The heat-map is initialized based on the inverse
frequency of each object type (therefore a lower value the
higher the object number) and including a range of influence
on nearby tiles. The event history is used to evaluate game
objects during simulations and to update the value map.
CatLinux was not a top controller on either of the individual
legs run in 2016, but placed 5th overall in the Championship.
This agent uses a Rolling Horizon Evolutionary Algorithm
(RHEA). A shift buffer enhancement is used to boost performance, specifically keeping the population evolved during
one game tick in the next, instead of discarding it, each action
sequence is shifted one action to the left (therefore removing
the previous game step) and a new random action is added at
the end to complete the individual to its fixed length.
No offline learning used by any of the EA agents, although
there could be scope for improvement through parameter
tuning (offline or online).
C. Opponent model
Most agents submitted to the Two-Player competition use
completely random opponent models. Some entries have
adopted the method integrated within the sample One Step
Lookahead controller, choosing a random but non-losing action. In the 2016 competition, webpigeon assumed the opponent would always cooperate, therefore play a move beneficial
to the agent. MaasCTS2 used the only advanced model at
the time: it remembered Q-values for the opponent actions
during simulations and added them to the statistics stored
in the MCTS tree nodes; an ǫ-greedy policy was used to
select opponent actions based on the Q-values recorded. This
provided a boost in performance on the games in the WCCI
2016 leg, but it did not improve the controller’s position in
the rankings for the following CIG 2016 leg.
Opponent models were found to be an area to explore
further in [13] and Gonzalez and Perez looked at 9 different
models integrated within the sample MCTS agent provided
with the framework [56]. Alphabeta builds a tree incrementally, returning the best possible action in each time tick,
while Minimum returns the worst possible action. Average
uses a similar tree structure, but it computes the average
reward over all the actions and it returns the action closest
to the average. Fallible returns the best possible action with
a probability p = 0.8 and the action with the minimum
reward otherwise. Probabilistic involved offline learning over
20 games in the GVGAI framework in order to determine
the probability of an MCTS agent to select each action,
and then using these to determine the opponent action while
playing online. Same Action returns the same action the agent
plays, while Mirror returns its opposite. Finally, LimitedBuffer
records the last n = 20 actions performed by the player
and builds probabilities of selecting the next action based on
this data, while UnlimitedBuffer records the entire history of
actions during the game. When all 9 opponent models were
tested in a round robin tournament against each other, the
probabilistic models achieve the highest win rates and two
models, Probabilistic and UnlimitedBuffer outperforming a
random opponent model.
VI. M ETHODS FOR S INGLE -P LAYER L EARNING
The GVGAI framework has also been used from an agent
learning perspective. In this setting, the agents do not use the
forward model to plan ahead actions to execute in the real
game. Instead, the algorithms learn the games by repeatedly
playing them multiple times (as episodes in Reinforcement
Learning), ideally improving their performance progressively.
This section first describes those approaches that tackled
the challenge set in the single-player learning track of the
competition (described in Section III-C), to then move to those
the ones not adhered to the competition format (Section VI-B.
A. 2017 competition entries
Some of the entries are available in the framework3.
1) Random agent: A sample random agent, which selects
an action at uniform random at every game tick, is included
in the framework (in both Java and Python) for the purposes
of testing. This agent is also meant to be taken as a baseline:
a learner is expected to perform better than an agent which
acts randomly and does not undertake any learning.
2) Multi-armed bandit algorithms: DontUnderestimateUchiha by K. Kunanusont is based on two popular MultiArmed Bandit (MAB) algorithms, ǫ-Decreasing Greedy Algorithm [57] and UCB [31]. At any game tick T , the current
best action a∗ with probability 1 − ǫT is picked, otherwise an
action is uniformly randomly selected. The best action a∗ at
time T is defined as in Equation 1.
!
r
2
log
T
∗
ˆ a+
,
(1)
a = arg max ∆score
a∈A
ta
where ta denotes the number of times that the action a
ˆ a denotes the empirical mean
has been selected, and ∆score
increment of score by applying the action a so far.
This is a very interesting combination, as the UCB-style
selection (Equation 1) and the ǫ-Decreasing Greedy Algorithm
both aim at balancing the trade-off between exploiting more
the best-so-far action and exploring others. Additionally, ǫ0
is set to 0.5 and it decreases slowly along time, formalized
as ǫT = ǫ0 − 0.0001T . According to the competition setting,
all games will last longer than 2, 000 game ticks, so ∀T ∈
{1, . . . , 2000}, 0.5 ≥ ǫT ≥ 0.3. As a result, random decisions
are made for approximately 40% time.
3) State-Action-Reward-State-Action: sampleLearner and
ercumentilhan are based on the State-Action-Reward-StateAction (SARSA) algorithm. Both agents choose to use a subset
of the whole game state information to build a new state to
reduce the amount of information to be saved and to take
into account similar situations. The main difference is that the
former uses a square region with fixed size centered at the
avatar’s position, while the latter uses a first-person view with
a fixed distance.
4) Q-learning: kkunan, by K. Kunanusont, is a simple Qlearning agent using most of the avatar’s current information
as features, which a few exceptions (as avatar’s health and
screen size, as these elements that vary greatly from game
to game). The reward at game tick t + 1 is defined as the
difference between the game score at game tick t + 1 and
the one at t. The learning rate α and discounted factor γ are
manually set to 0.05 and 0.8. During the learning phase, an
random is performed with probability ǫ = 0.1, otherwise, the
best action is selected. During the validation phase, the best
action is always selected. Despite it’s simplicity, it won the
only edition of this track up to date, being the only entry that
ranked above the sample random agent.
5) Tree search methods: YOLOBOT is an adaption of
the YOLOBOT planning agent (as described previously in
Section IV-E). As the forward model is no more accessible
3 https://github.com/GAIGResearch/GVGAI/tree/master/clients/
in the learning track, the MCTS is substituted by a greedy
algorithm to pick the action that minimizes the distance to
the chosen object at most. According to the authors, the poor
performance of YOLOBOT in the learning track, contrary to
its success in the planning tracks, was due to the collision
model created by themselves that did not work well.
B. Other learning agents
One of the first works that used this framework as a learning
environment was carried out by S. Samothrakis et al. [58], who
employed Neuro-Evolution in 10 games of the benchmark.
Concretely, the authors experimented with Separable Natural
Evolution Strategies (S-NES) using two different policies (ǫgreedy versus softmax) and a linear function approximator
versus a neural network as a state evaluation function. Features
like score, game status, avatar and other sprites information
were used to evolve learners during 1000 episodes. Results
show that ǫ-greedy with a linear function approximator was
the better combination to learn how to maximize scores on
each game.
A. Braylan and R. Miikkulainen [59] performed a study
in which the objective was to learn a forward model on
30 games. The objective was to learn the next state from
the current one plus an action, where the state is defined
as a collection of attribute values of the sprites (spawns,
directions, movements, etc.), by means of logistic regression.
Additionally, the authors transfer the learnt object models from
game to game, under the assumption that many mechanics
and behaviours are transferable between them. Experiments
showed the effective value of object model transfer in the
accuracy of learning forward models, resulting in these agents
being stronger at exploration.
Also in a learning setting, K. Kunanusont et al. [60], [61]
developed agents that were able to play several games via
screen capture. In particular, the authors employed a Deep
Q-Network [62] in 7 games of the framework of increasing
complexity, and included several enhancements to GVGAI to
deal with different screen sizes and a non-visualization game
mode. Results showed that the approach allowed the agent to
learn how to play in both deterministic and stochastic games,
achieving a higher winning rate and game score as the number
of episodes increased.
VII. M ETHODS FOR LEVEL GENERATION
Different researchers used different approaches to generate
levels for the GVGAI framework. The following subsection
describes all known generators either included in the framework or developed during the competition.
A. Constructive methods
Constructive generators are designed to generate the level
based on general knowledge. For example: enemies should be
away from the avatar, walls shouldn’t divide the world into
islands, etc. Based on the game the generator adjusts a couple
of parameters and rules to fit the game as, for example, the
number of non-playable characters (NPCs) in the generated
level. Constructive generators don’t need any simulations after
generating the level. The following are the known constructive
generators.
1) Sample random generator: This is the most naive
method to generate a level. The generator first identifies
solid sprites. Solid sprites block the avatar and all NPCs
from moving and don’t react to anything. The generator adds
a selected solid sprite as a border for the generated level
to prevent sprites from wandering outside the game screen.
Followed by adding one of each character in the level mapping
section to a random location in the level. This step ensures the
game is playable. Finally, it adds a random amount of random
sprites from the level mapping to random locations in the level.
2) Sample constructive generator: This generator uses
some general game knowledge to generate the level. First, the
generator calculates the level dimensions and the number of
sprites in the level, then labels game sprites based on their
interactions and sprite types. After that, it constructs a level
layout using the solid sprites, to later add the avatar to a
random empty location. After knowing the avatar position, the
generator adds harmful sprites (those that can kill the avatar)
in a far location from the avatar and adds other sprites at any
random free locations. Finally, the generator makes sure that
the number of goal sprites is sufficient to prevent winning or
losing automatically when the game starts.
3) Easablade constructive generator: This is the winner
generator for the first level generator competition. The generator is similar to the sample constructive generator but it uses
cellular automata to generate the level instead of layering the
objects randomly. The cellular automata is run on multiple
layers. The first layer is to design the map obstacles, followed
by the exit and the avatar, then the goal sprites, harmful sprites,
and others.
4) N-Gram constructive generator: This generator uses a
n-gram model to generate the level. The generator records
the player actions from a previous play-through. This action
sequence is used to generate the levels using predefined rules
and constraints. For example, if the player uses the USE action
quite often, the generator will include more enemies in the
level. The n-gram is used to specify the rules. Instead of reacting to each separate action, the model reacts to a n-sequence
of actions. During the generation process, the algorithm keeps
track of the number and position of every generated object to
ensure the generated sprites do not overpopulate the level. A
single avatar sprite is placed in the lower half of the level.
B. Search-based methods
Search-based generators use simulations to make sure the
generated level is playable and better than just placing random
objects. The following are the known search-based generators.
1) Sample genetic generator: This is a search-based level
generator based on the Feasible Infeasible 2 Population Genetic Algorithm (FI2Pop) [63]. FI2Pop is a genetic algorithm
which uses 2 populations, one for feasible chromosomes and
the other for infeasible chromosomes. The feasible population
tries to increase the difference between the OLETS agent (see
Section IV-B) and one-step look ahead, while the infeasible
population tries to decrease the number of chromosomes that
violate the problem constraints (i.e. at least one avatar in the
game, the avatar must not die in the first 40 steps, etc.). Each
population evolves on its own, where the children can transfer
between the two populations. This generator initializes the
population using sample constructive generator.
2) Amy12 genetic generator: This generator is built on top
of the sample genetic generator. The main idea is to generate a
level that fits a certain suspense curve. Suspense is calculated
at each point in time, by calculating the number of actions that
leads to death or tie using the OLETS agent. The algorithm
modifies the levels to make sure the suspense curve is not
constant during the life time of the game. Good generators
are aimed at producing 3 suspense peeks with values of 50%
(where half of the actions, on average lead to losing the game).
One of the advantages of using this technique that it makes
sure that the generated level is winnable. Games that are hard
to win will have a higher peak in the suspense curve, which
is not valued highly by the generator.
3) Jnicho genetic generator: This generator [64] uses a
standard genetic algorithm with similar crossover and mutation
operators to the sample genetic algorithm. The fitness function
used is a combination between the score difference and the
constraints specified in the sample genetic generator. The score
difference is calculated between an Monte Carlo Tree Search
agent and One Step Look Ahead agent. The score difference is
normalized between 0 and 1 to make sure it won’t overshadow
the constraint values.
4) Number13 genetic generator: This is a modified version
of the sample genetic generator. These modifications includes
using adaptive crossover mechanism, adaptive mutation rate,
a better agent than OLETS, and allowing crossover between
feasible and infeasible population, which is not allowed in the
sample genetic generator.
C. Constraint-based methods
1) ASP generator: This generator [65] uses Answer Set
Programming (ASP) to generate levels. The main idea is to
generate ASP rules that generate suitable levels for the current
game. The generated rules consists of three different types. The
first type are basic rules, which are based on specific decisions
to keep the levels simple (for instance, levels can only have one
sprite per tile). The second type are game specific rules, which
are extracted from the game description file. An example is
the identification of singleton sprites that should only have
one sprite in the level. The last type are additional rules to
minimize the search space. These rules limit the minimum and
maximum number of each sprite type. All the rules are evolved
using evolutionary strategy with the algorithm performance
difference between sampleMCTS and a random agent as the
fitness function.
VIII. M ETHODS FOR RULE G ENERATION
This section describes the different algorithms that are
included in the framework or have been found in the literature [66] toward generating rules for the GVG-AI framework.
A. Constructive methods
Constructive methods are algorithms that generate the rules
in one pass without the need to play the game. The constructive
methods often incorporate knowledge about game design to
generate more interesting games.
1) Sample random generator: This is the simplest generator
provided with the framework. The main idea is to generate a game that compiles with no errors. For example, the
game shouldn’t contain interactions such as killing the end
of screen (EOS) sprite. The algorithm starts by generating
a random number of interactions by selecting two random
sprites (including EOS) and a random interaction rule one
by one. The algorithm checks that every interaction is valid
before adding it to the generated game. After generating the
random interactions, the algorithm generates two termination
conditions, one for winning and one for losing. The losing
condition is fixed to the avatar being killed, while the winning
is either winning the game after a random amount of frames or
winning the game when the number of a certain sprite reaches
zero.
2) Sample constructive generator: This is a more complex
generator that utilizes knowledge about VGDL language and
level design to generate more interesting games. The algorithm
starts by classifying the game sprites into different categories,
such as wall sprites (sprites that surround the level), collectible/harmful sprites (immovable sprites that cover around
10% of the level), spawner sprites (sprites that spawn another
sprite type), etc. For each type of sprite, the algorithm has rules
to generate interactions based on them. For example, harmful
sprites kill the avatar on collision, wall sprites either prevent
any movable object from passing through or kill the movable
object upon collision, etc. For more details about the rules,
the reader is referred to [15]). After the game interactions are
generated, two termination conditions are generated, one for
winning and one for losing. The losing condition is fixed to
the avatar’s death, while the winning condition depends on the
current sprites. For example: if collectible sprites exist in the
current definition, the winning condition is set to collect all of
them.
B. Search-based methods
Search-based methods use a search based algorithm to find a
game based on certain criteria that ensure the generated game
have better rules than just randomly choosing them.
1) Sample genetic generator: Similar to the level generation track, the search based algorithm uses Feasible Infeasible
2 Population Genetic Algorithm (FI2Pop) [63] to evolve new
games. As discussed before, FI2Pop keeps two populations
one for feasible games and the other for infeasible games.
The infeasible games tries to become feasible by satisfying
multiple constraints such as minimizing the number of bad
frames (frames contains sprites outside the level boundaries)
under certain threshold, the avatar doesn’t die in the first 40frames, etc. On the other hand, the feasible chromosomes try
to maximize its fitness. The fitness consists of two parts, the
first part is to maximize the difference in performance between
the OLETS and MCTS agents, and the difference between
MCTS and random agent. The second part is to maximize the
number of interaction rules that fires during the simulation of
the generated game.
2) Thorbjrn generator: This generator [66] is inspired by
the sample genetic generator. Similar to the sample genetic
generator, it tries to maximize the difference between the
performance of different algorithms. This generator uses evolutionary strategies with mutation and crossover operators to
generate an entire game instead of an interaction set and
termination conditions.
IX. R ESEARCH
THAT BUILDS ON
GVGAI
A. AI-assisted game design
Machado et al. [67] implemented a recommender system
based on the VGDL to recommend game elements, such as
sprites and mechanics. Then, the recommender system was
expanded to Cicero [68], an AI-assisted game design tool
built on top of the GVGAI. Cicero has a statistics tool of
the interactions to help figuring out the unused game rules; a
visualization system to illustrate the information about game
objects and events, a mechanics recommender, a query system
for in-game data and a retrospective analysis application SeekWhence [69]. The gameplay sessions by human players or AI
agents can be recorded and every single frame at every game
tick can be easily extracted for further study and analysis.
Recently, Liu et al. [70] applied a simple Random Mutation
Hill Climber (RMHC) and a Multi-Armed Bandit RMHC
together with resampling methods to tune game parameters
automatically. Games instances with significant skill-depth
have been evolved using GVGAI agents. Futhermore, Kunanusont et al. [52] evolved simultaneously the GVGAI agents as
part of the game (opponent models).
Guerrero et al. [71] explored five GVGAI agents using
four different heuristics separately on playing twenty GVGAI
games, allowing different behaviors according to the diverse
scenarios presented in the games. In particular, this work
explored heuristics that were not focused on winning the game,
but to explore the level or interact with the different sprites
of the games. These agents can be used to evaluate generated
games, thus help evolve them with preferences to particular
behaviors.
Khalifa et al. [72] modified MCTS agents by editing the
UCT formula used in the agent. Human playing data has
been used for modeling to make the modified agents playing
in a human-like way. Primary results showed that one of
the studied agents achieved a similar distribution of repeated
actions to the one by human players. The work was then
extended by Bravi et al. [50], in which game-play data have
been used to evolve effective UCT alternatives for a specific
game. The MCTS agents using new formulas, with none
or limited domain information, are compared to a standard
implementation of MCTS (the sampleMCTS agent of GVGAI)
on the game Missile Command. Applying the UCT alternatives
evolved using game-playing data to a standard MCTS significantly improved its performance.
Besides designing games and the agents used in them,
the automatic generation of video game tutorials (aimed at
helping players understanding how to play a game) is also
an interesting sub-field of study. Green et al. [73] pointed
out that the GVGAI Framework provides an easy testbed for
tutorial generation. The game rules in GVGAI are defined in
VGDL, therefore the tutorial generation can be easily achieved
by reading and translating VGDL files.
A more recent work by Anderson et al. [74] focused on
designing deceptive games to deceive AI agents and lead
the agents away from a globally optimal policy. Designing
such games helps understand the capabilities and weaknesses
of existing AI agents and can serve at a preparation step
for designing a meta-agent for GVGP which combines the
advantages of different agents. The authors categorized the deceptions and imported various types of deception to the existing GVGAI games by editing the corresponding VGDL files.
The agents submitted to the GVGAI single-player planing
competition have been tested on the new games. Interestingly,
the final ranking of the agents on each of the games differed
significantly from the rankings in the GVGAI competition.
The new designed deceptive games successfully explored the
weaknesses of agents which have performed well on the test
set of the official competition.
B. Game generation with RAPP
Nielsen et al. [75] proposed Relative Algorithm Performance Profile (RAPP) as a measure of relative performance
of agents and tested their approach on different general gameplaying AI agents using GVGAI framework. The authors
showed that well-designed games have clear skill-depth, thus
being able to distinct good or bad players. In other words,
a strong agent or human player should perform significantly
better than a weak agent or human player over multiple
playings on well-designed games. For instance, a skillful agent
is expected to perform better than a random agent, or one that
does not move.
Then, Nielsen et al. [66] integrated the differences of
average game scores and win rate between any agent and a
random agent to the evaluation of new games either randomly
generated or generated by editing existing GVGAI games.
Though most of the resulted games are interesting to play,
there are some exceptions, in which the core challenge of the
game has been removed. For instance, the enemy can not heart
the player, which makes it no more an enemy. But it still
provides useful starting points for human designers.
Kunanusont et al. [52] extended the idea of RAPP. Five
GVGAI agents and a deterministic agent designed for the
tested Space Battle Game are used as the candidate opponent,
which is considered as part of the game to be evolved.
Two GVGAI agents, One Step Look Ahead (weak), MCTS
(strong) and the deterministic agent (mediocre), are used to
play multiple times the evolved game for evaluation. The
evaluation function is defined as the minimum of the difference
of game scores between the strong and mediocre agents, and
the difference of game scores between the mediocre and weak
agents, aiming at generating games that can clearly distinguish
stronger agents and weak agents. A recently proposed N-Tuple
Bandit Evolutionary Algorithm [76] has been compare to the
simple RMHC. Two human players have tested the generated
games and preferred the new game evolved using the N-Tuple
Bandit Evolutionary Algorithm.
C. Robustness testing
Perez et al. [77] ran a study on the winners of the 2014
and 2015 editions of the single player planning competition
in order to analyze how robust they were to changes in the
environment with regards to actions and rewards. The aim of
this work was to analyze a different type of generality: controllers for this framework are developed to play in multiple
games under certain conditions, but the authors investigated
which could be the effect of breaking those compromises. An
inaccurate and noisy Forward Model, an agent that does not
execute the move decided by the algorithm, or score penalties
incurred by performing certain actions.
An interesting conclusion on this study is that, once the
conditions have been altered, sample agents climb up to the
top of the rankings and the good controllers behave worse.
Agents that rely on best first search or A* (such as YOLOBOT
or Return42, already described in this paper) handled noise
very badly. MCTS also showed to be quite robust in this
regard, above other Rolling Horizon agents that could not
cope so well with these changes. This work also reinforced
the idea that the GVGAI framework and competition are also
robust. Despite the changes in the performance of the agents,
some controllers do better than others under practically all
conditions. The opposite (rankings depending only on noise
factors, for instance) would mean that the framework is quite
fragile.
X. E DUCATIONAL USE OF GVGAI
The GVGAI framework has been used to provide engaging
assignments for taught modules, and as the basis for many
MSc dissertation projects. The descriptions below give an idea
of the educational uses of GVGAI but are not intended to be
an exhaustive list.
interesting games as the assignment. This was done to good
effect at IT University of Copenhagen, where the students
produced a number of challenging puzzle games that were
later used in the planning track training and validation sets. A
similar approach was taken in a module on AI-Assisted Game
Design at the University of Essex, where planning track games
were also produced.
B. MSc Dissertation Projects
GVGAI offers an extensive range of interesting research
challenges, some of which have been addressed in MSc
dissertation projects. The majority of the ones we are aware of
have focused on the planning track, but this is not surprising
as it was the first track to be developed. The planning track
also has the benefit of providing some good sample agents
as starting points for further work either in the sense of
extending the samples to achieve higher performance, or using
the samples as a useful source of comparison. A good example
is the work on MCTS with options, which used showed the
the version with options significantly outperformed the sample
MCTS agent on most of the games studied: as with many
cases what began as an MSc thesis was later published as a
conference paper [33]. In our experience this usually provides
an excellent educational experience for the student. Other
planning track theses include [25], the real-time enhancements
of [26], knowledge-based variants [46] and goal-oriented
approaches [53].
Beyond the planning track, other examples (already described in this survey) include applying Answer-Set Programming (ASP) [78] or Genetic Algorithms (GA) [64] for the level
generation track and learning from screen capture [60] (this
was essentially a learning track approach before the learning
track was running). Finally, another approach is to extend the
framework in some way, such developing the 2-player learning
track [79].
XI. D ISCUSSION
AND OPEN RESEARCH PROBLEMS ON
SINGLE AND TWO - PLAYER PLANNING
A. Taught Modules
GVGAI has been used in at least two distinct ways within
taught modules. The most typical way is to use design specific
aspects of the course around the framework, teaching the
students about the main concepts of GVGAI with examples
of how to write agents for the selected tracks. This is then
followed up with an assignment, where a significant weight is
given to how well each student or group’s entry performs in the
league. Several institutions have run private leagues for this,
including Otto Von Guericke Universität Magdeburg, University of Essex, University of Muenster, Universidad Carlos III
de Madrid, Universidad de Malaga and New York University.
Running a private league means the course supervisor has full
control over the setup of the league, including when students
can enter and how thoroughly the entries are evaluated, and
the set of games to evaluate them on. For the 2-player track,
this also allows control over the opponents chosen.
Another use-case in taught modules is to teach the VGDL
part of framework, and then set the development of novel and
The single and two-player planning versions of GVGAI
are the ones that have received most attention and research.
Despite their popularity and efforts, the best approaches rarely
surpass an approximately 50% victory rate in competition
game sets, with very low victory percentage in a great number
of games. Similarly, different MCTS and RHEA variants
(including many of the enhancements studied in the literature)
struggle to achieve a higher than 25% victory rate in all (more
than a hundred) single-player games of the framework. Therefore, increasing performance in a great proportion of games is
probably the most challenging problem at the moment.
Literature shows multiple enhancements on algorithms and
methods aiming to improve this performance, but in the vast
majority of cases the improvements only affect a subset of
games or certain configurations of the algorithms. While this
is understandable due to the nature of general video game
playing, it also shows that the current approaches does not
work in order to reach truly general approaches that work
across board.
The work described in this survey has shown, however,
interesting insights that can point us in the right direction. For
instance, several studies show that using more sophisticated
(i.e. with A* or other methods such as potential fields)
distances to sprites as features works better than Euclidean
distances. The downside is that computing these measurements
take an important part of the decision time budget, which can’t
be used in case it is needed for some games or states where
the best action to take is not straight-forward.
In general, one could say that one of the main points to
address is how to use the decision time more wisely. Some
approaches tried to make every use of the forward model
count, like those agents that attempt to learn facts about the
game during the rollouts of MCTS. Again, some attempts in
this direction have provided marginal improvements, but the
problem may be trying to a general feature extractor. In other
words, what we try to learn is influenced by what we know
about existing games (i.e. some sprites are good, other are bad,
some spawn other entities and there are things - resources that the avatar can collect). Some games may require features
that have not been thought of, especially because the challenge
itself presents games that has not been seen before to the
participants.
Another improvement that has been tried in several studies
is the use of macro-actions (in most cases, a repetition of an
action during several consecutive steps) to a) make the action
space coarser; and b) make a better use of the time budget.
Again, these modifications have improved performance in
certain games (including some that had not been won by any
algorithm previously) but they either did not have an impact
in others, or they made the performance worse. It is likely
that different games can benefit from different macro-action
lengths (so work could be done on trying to automatically and
dynamically adapt the number of times the action is repeated)
but also of more complex structures that allow for high-level
planning. In fact, games that require high-level planning are
still and open problem to be solved in this setting.
Games classification and the use of hyper-heuristics are also
an interesting area for research. Some of the best approaches
up to date, as YOLOBOT, do make a differentiation between
stochastic and deterministic games to later use one or another
algorithm. An open challenge is how to make this classification
more accurate and detailed, so an approach could count on a
portfolio of (more than 2) algorithms that adapt to every game.
Attempts have been made to classify with game features, but
results suggest that these classifications and the algorithms
used are not strong enough. Devising more general features
for this clustering, maybe focused on the agent gameplay
experience rather than game features, is a line of future
research.
All these unsolved issues apply to both single and two
player settings, although the latter case adds the difficulty of
having an opponent to compete or collaborate with. There are
two open problems that arise from this: first, no study has
been made that tries to identify the game and behaviour of the
opponent as collaborative of competitive. Analysis of the other
player’s intentions can be seen as a sub-field on its own, only
that in this case we add the general game playing component
to it. Secondly, some advancements have been done in using
opponent models that go beyond random, but investigation in
more complicated opponent models that better capture and
learn the behaviour of the other player could potentially yield
better results.
Beside the development of agents for game playing, AIassisted game design and automatic game testing and debugging using GVGAI agents have attracted researchers’ attention.
Some work around evolving game skill-depth using relative
performance between GVGAI agents have been done recently,
and most of this work has been focused on Relative Algorithm
Performance Profiles (RAPP), where performance is measured
in terms of how well the agents play the given games.
However, it is sensible to explore other aspects of agent game
play to influence game design. Factors like the amount of
level explored by different agents (so a generator favours those
levels or games that allow for a wider exploration, or maybe a
progressive one), their decisiveness [80] on selecting the best
action to take or the entropy of their moves can also be used
to this end.
XII. F UTURE D IRECTIONS
The GVGAI framework and competition are in constant
development. The opportunities that this benchmark provides
for different lines of research and education are varied, and
this section outlines the future directions planned ahead for
the following years.
A. New tracks
As new challenges are proposed, the possibility of organizing them as tracks for the competition arise. Below are listed
some possible new tracks that can attract interesting areas of
research.
1) Automatic game design: The game design involves,
but not limited to, game generation, level generation, rule
generation and playtesting (playing experience, game feeling,
fun, etc.), study of game market, user interface design and
audio design. The automatic game design becomes an active
research topic since the late 2000’s, though, to the best of our
knowledge, Barney Pell firstly automatically generated rules
for chess-like games in 1992 [81]. A review of the state of the
art in automatic game design can be found in [70].
A Game Generation track would aim at providing
AI controllers which automatically generate totally new games
or game instances by varying the game parameters, i.e.,
parameter tuning. How to achieve the former is an open
question. The straight-forward way would be providing a
particular theme, a database of game objects, or searching
spaces of game rules, with which the participants can generate
new games. The ideal case would be that the controllers
automatically create totally new games from nothing. Though
there is a yawning gulf between aspiration and reality, an
interdisciplinary field combining automatic game design and
domain-specific automatic programming is expected. The latter, automatic game tuning, is relatively easier. Some searchbased and population-based methods have been applied to
game parameter optimisation aiming at maximising the depth
of game variants [70] or finding more playable games [82].
2) Multi-Player GVGAI: Multi-agent games has drawn
people’s attention, for instance, real time strategy games (e.g.
StarCraft) and board games (e.g. Mahjong). The study of
multi-agent GVGAI is a fruitful research topic. Atari games
can also be extended to multi-agent games. In particular,
the Pac-Man can be seen as a multi-agent game and related
competitions have been held since 2011 [1]. The most recent
Ms Pac-Man vs Ghost Team Competition [83], which included
partial observability, was held at CIG in 2016. Nevertheless,
a more general multi-agent track is favorable.
The interface of the Two-Player Planning Track was initially
developed for two or more players, so it has the potential to
be expanded to a Multi-player planning track, in which an
agent is allowed to control more than one player or each of
the players is controlled by a separate agent. This future track
can be expanded again as a multi-agent learning framework,
providing a two-or-more player learning track.
3) Turing Test GVGAI: Determining if an agent that is
playing a game is a human or a bot is a challenge that has
been subject of study for many years [84], [85], and the idea
of applying it to a general video game setting is not new [86].
This offers an interesting opportunity to extend the framework
to having a Turing Test track where participants create AI
agents that play like humans for any game that is given. Albeit
the understandable difficulty of this problem, the interest for
research in this area is significant: what are the features that
can make an agent play like a human in any game?
B. General directions
There are several improvements and additions to the framework that can be done and would potentially affect all existent
and future competition tracks. One of these continuous modifications is the constant enlargement of the games library.
Not only new games are added for each new edition of the
competition, but the work done on automatic game design
using the GVGAI framework has the potential to create infinite
number of games that can be integrated perfectly into the
framework.
Adding more games can also be complemented with compatibility with other systems. Other general frameworks like
OpenAI Gym [87], ALE [4] or Microsoft Malmö [88] count on
a great number of single or multi-player, model-free or model
based tasks. Interfacing with these systems would increment
the number of available games which all GVGAI agents could
play via a common API. This would also open the framework
to 3D games, an important section of the environments the
current benchmark does not cover.
With regards to the agents, another possibility could be
to provide them with a wider range of available actions.
For instance, the player could be able to apply more than
one action simultaneously, or these actions could form a
continuous action space (i.e. pressing a throttle in a range
between 0 and 1). In general, this would enhance the number
of legal combinations for the agent to choose from at each
decision step.
Beside the framework itself, the website for GVGAI could
also be improved to provide better and faster feedback to
the competition participants. More data analysis features can
be added, such as visualization of the score changes during
the game playing, the action entropy and the exploration of
the game world (heat-map of visited positions). A related
work is to provide better and more flexible support for game
play metric logging, better support for data mining of results
together with visualization, and better data saving, which will
help enabling to upload replays (i.e., action logs) from AI
agents and human play-throughs.
Another envisaged feature is being able to play the game
in a web browser (without any download or installation) by
an AI agent or human, and visualize the analyzed features
during the game playing in real time. A bonus will be the
easy parametrisation options for games., thus a player or
an AI agent can easily set up the parameters and rules to
define the desired game by inserting values directly or generate
pseudo-randomly a level to play using some pre-implemented
automatic game tuning techniques given some particular goals
or features.
XIII. C ONCLUSIONS
The GVGAI framework offers the most comprehensive
system to date for evaluating the performance of general
video game playing agents, and for testing general purpose
algorithms for creating new games and creating new content
for novel games. The framework has been used in multiple
international competitions, and has been used to evaluate the
performance of hundreds of general video game agents.
The agent tracks cater for planning agents able to exploit
a fast forward model, and learning agents that must learn
to react sensibly without the benefits of a forward model.
The planning track already comes in single and two-player
versions; the learning track is currently single-player only, but
with a two-player version under development. Although longterm learning may also be used within the planning track, the
best-performing agents have, as far as we know, not yet done
this. The success of AlphaGoZero [89] indicates what can be
achieved by combining learning and planning, so applying a
similar system within GVGAI is an interesting prospect.
The main alternative to GVGAI is the Atari Learning
Environment (ALE). At the time of writing, ALE offers
higher-quality games than GVGAI as they were home-console
commercial games of a few decades ago. In GVGAI terms,
ALE offers just two tracks: single-player learning, and singleplayer planning, with the learning track being the more widely
used. For future work on machine learning in video games,
we predict that the 2-player tracks will become the most
important, as they offer open-ended challenges based on an
arms race of intelligence as new players are developed, and
are also outside of the current scope of ALE.
In addition to offering a wider range of tasks, GVGAI
also benefits from being much more easily extensible than
ALE, though ALE has so far had greater uptake within some
sectors of the machine learning community. An immediate
priority is to test the rich set of ALE agents on the equivalent
GVGAI-tracks to gain a sense of the relative difficulty of
each environment and to learn more of the relative challenges
offered by each one.
The content creation tracks offer an extremely hard challenge: creating rules or levels for unseen games. Promising
directions include the further development and exploitation of
a range of general game evaluation measures [80], and greater
use of the best GVGAI agents to perform the play-testing of
the novel rules and levels.
The Video Game Description Language (VGDL) has been
an important part of GVGAI to date, since it makes it possible
to rapidly and concisely specify new games. However, it is
also a source of limitation, as its limited expressivity makes it
hard to make games which are fun for humans to play. VGDL
also limits the ease with which complex game mechanics can
be embedded in games, which in turn limits the depth of
challenge that can be posed for the GVGAI agents. Hence
an important future direction is the authoring of GVGAIcompatible games in any suitable language which conform
to the necessary GVGAI API in order to ensure compatibility
with the desired GVGAI track.
Finally, while the above discussion provides a compelling
case for the future of GVGAI as a tool for academic study, we
also believe that when it reaches a higher level of maturity it
will provide an important tool for game designers. The vision
is to provide an army of intelligent agents with a range of
play-testing abilities, and a diverse set of metrics with which
to analyse a range of important functional aspects of a game.
ACKNOWLEDGEMENTS
The authors would like to thank the participants of all tracks
of the competition for their work and submitted controllers and
generators. This work was partially funded by the EPSRC
CDT in Intelligent Games and Game Intelligence (IGGI)
EP/L015846/1.
R EFERENCES
[1] D. P.-L. Philipp Rohlfshagen, Jialin Liu and S. M. Lucas, “Pac-Man
Conquers Academia: Two Decades of Research Using a Classic Arcade
Game,” IEEE Transactions on Computational Intelligence and AI in
Games, 2017.
[2] S. Karakovskiy and J. Togelius, “The mario ai benchmark and competitions,” IEEE Transactions on Computational Intelligence and AI in
Games, vol. 4, no. 1, pp. 55–67, 2012.
[3] D. Loiacono, P. L. Lanzi, J. Togelius, E. Onieva, D. A. Pelta, M. V.
Butz, T. D. Lonneker, L. Cardamone, D. Perez, Y. Sáez et al., “The
2009 simulated car racing championship,” IEEE Transactions on Computational Intelligence and AI in Games, vol. 2, no. 2, pp. 131–147,
2010.
[4] M. G. Bellemare, Y. Naddaf, J. Veness, and M. Bowling, “The arcade
learning environment: An evaluation platform for general agents.” J.
Artif. Intell. Res.(JAIR), vol. 47, pp. 253–279, 2013.
[5] S. Ontanón, G. Synnaeve, A. Uriarte, F. Richoux, D. Churchill, and
M. Preuss, “A survey of real-time strategy game ai research and competition in starcraft,” IEEE Transactions on Computational Intelligence
and AI in games, vol. 5, no. 4, pp. 293–311, 2013.
[6] D. Perez-Liebana, S. Samothrakis, J. Togelius, S. M. Lucas, and
T. Schaul, “General Video Game AI: Competition, Challenges and
Opportunities,” in Thirtieth AAAI Conference on Artificial Intelligence,
2016, pp. 4335–4337.
[7] M. Ebner, J. Levine, S. M. Lucas, T. Schaul, T. Thompson, and
J. Togelius, “Towards a video game description language,” 2013.
[8] T. Schaul, “A video game description language for model-based or
interactive learning,” in Computational Intelligence in Games (CIG),
2013 IEEE Conference on. IEEE, 2013, pp. 1–8.
[9] ——, “An extensible description language for video games,” IEEE
Transactions on Computational Intelligence and AI in Games, vol. 6,
no. 4, pp. 325–331, 2014.
[10] J. Levine, C. B. Congdon, M. Ebner, G. Kendall, S. M. Lucas, R. Miikkulainen, T. Schaul, and T. Thompson, “General video game playing,”
Dagstuhl Follow-Ups, vol. 6, 2013.
[11] D. Perez, S. Samothrakis, J. Togelius, T. Schaul, S. Lucas, A. Couëtoux,
J. Lee, C.-U. Lim, and T. Thompson, “The 2014 general video game
playing competition,” IEEE Transactions on Computational Intelligence
and AI in Games, vol. 8, pp. 229–243, 2015.
[12] R. D. Gaina, D. Perez-Liebana, and S. M. Lucas, “General Video
Game for 2 Players: Framework and Competition,” in Proceedings of
the IEEE Computer Science and Electronic Engineering Conference
(CEEC), 2016.
[13] R. D. Gaina, A. Couëtoux, D. J. Soemers, M. H. Winands, T. Vodopivec,
F. Kirchgessner, J. Liu, S. M. Lucas, and D. Perez-Liebana, “The
2014 general video game playing competition,” IEEE Transactions on
Computational Intelligence and AI in Games, 2017.
[14] A. Khalifa, D. Perez-Liebana, S. M. Lucas, and J. Togelius, “General
video game level generation,” in Proceedings of the 2015 Annual
Conference on Genetic and Evolutionary Computation. ACM, 2016.
[15] A. Khalifa, M. C. Green, D. Pérez-Liébana, and J. Togelius, “General
Video Game Rule Generation,” in 2017 IEEE Conference on Computational Intelligence and Games (CIG). IEEE, 2017.
[16] D. Pérez-Liébana, M. Stephenson, R. D. Gaina, J. Renz, and S. M.
Lucas, “Introducing Real World Physics and Macro-Actions to General
Video Game AI,” in 2017 IEEE Conference on Computational Intelligence and Games (CIG). IEEE, 2017.
[17] J. Liu, D. Perez-Liebana, and S. M. Lucas, “The Single-Player GVGAI
Learning Framework - Technical Manual,” 2017. [Online]. Available:
http://www.liujialin.tech/publications/GVGAISingleLearning manual.pdf
[18] T. Joppen, M. Moneke, N. Schroder, C. Wirth, and J. Furnkranz,
“Informed Hybrid Game Tree Search for General Video Game Playing,”
IEEE Transactions on Computational Intelligence and AI in Games,
2017.
[19] D. Ashlock, D. Pérez-Liébana, and A. Saunders, “General Video Game
Playing Escapes the No Free Lunch Theorem,” in 2017 IEEE Conference
on Computational Intelligence and Games (CIG). IEEE, 2017.
[20] T. Geffner and H. Geffner, “Width-based planning for general videogame playing,” in Eleventh Artificial Intelligence and Interactive Digital
Entertainment Conference, 2015.
[21] D. Soemers, C. Sironi, T. Schuster, and M. Winands, “Enhancements for
Real-Time Monte-Carlo Tree Search in General Video Game Playing,”
in IEEE Conference on Computational Intelligence and Games (CIG),
2016.
[22] G. A. Rummery and M. Niranjan, On-line Q-learning using connectionist systems. University of Cambridge, Department of Engineering,
1994, vol. 37.
[23] J.
Liu,
“Gvgai
single-player
learning
competition
at
ieee
cig17,”
2017.
[Online].
Available:
https://www.slideshare.net/ljialin126/gvgai-singleplayer-learning-competition-at-ieee-ci
[24] C. B. Browne, E. Powley, D. Whitehouse, S. M. Lucas, P. I. Cowling,
P. Rohlfshagen, S. Tavener, D. Perez, S. Samothrakis, and S. Colton,
“A survey of monte carlo tree search methods,” IEEE Transactions on
Computational Intelligence and AI in games, vol. 4, no. 1, pp. 1–43,
2012.
[25] M. B. A. for General Video Games, “Torsten schuster,” Master’s thesis,
Maastricht University, 2015.
[26] D. J. Soemers, “Enhancements for real-time monte-carlo tree search
in general video game playing,” Master’s thesis, Maastricht University,
2016.
[27] D. Perez, S. Samothrakis, and S. Lucas, “Knowledge-based fast evolutionary mcts for general video game playing,” in 2014 IEEE Conference
on Computational Intelligence and Games. IEEE, 2014, pp. 1–8.
[28] C. Y. Chu, T. Harada, and R. Thawonmas, “Biasing monte-carlo rollouts
with potential field in general video game playing,” , p. 6p, 2015.
[29] C. Y. Chu, H. Hashizume, Z. Guo, T. Harada, and R. Thawonmas,
“Combining pathfmding algorithm with knowledge-based monte-carlo
tree search in general video game playing,” in 2015 IEEE Conference
on Computational Intelligence and Games (CIG). IEEE, 2015, pp.
523–529.
[30] H. Park and K.-J. Kim, “Mcts with influence map for general video game
playing,” in 2015 IEEE Conference on Computational Intelligence and
Games (CIG). IEEE, 2015, pp. 534–535.
[31] P. Auer, “Using confidence bounds for exploitation-exploration tradeoffs,” Journal of Machine Learning Research, vol. 3, no. Nov, pp. 397–
422, 2002.
[32] E. H. dos Santos and H. S. Bernardino, “Redundant action avoidance and
non-defeat policy in the monte carlo tree search algorithm for general
[33]
[34]
[35]
[36]
[37]
[38]
[39]
[40]
[41]
[42]
[43]
[44]
[45]
[46]
[47]
[48]
[49]
[50]
[51]
[52]
[53]
video game playing,” in Proceedings do XVI Simpsio Brasileiro de Jogos
e Entretenimento Digital, 2017.
M. d. Waard, D. M. Roijers, and S. C. Bakkes, “Monte carlo tree search
with options for general video game playing,” in 2016 IEEE Conference
on Computational Intelligence and Games (CIG). IEEE, 2016, pp. 47–
54.
D. Perez-Liebana, S. Mostaghim, and S. M. Lucas, “Multi-objective
tree search approaches for general video game playing,” in Evolutionary
Computation (CEC), 2016 IEEE Congress on. IEEE, 2016, pp. 624–
631.
A. Khalifa, M. Preuss, and J. Togelius, “Multi-objective adaptation of
a parameterized gvgai agent towards several games,” in International
Conference on Evolutionary Multi-Criterion Optimization. Springer,
2017, pp. 359–374.
I. Bravi, A. Khalifa, C. Holmgård, and J. Togelius, “Evolving uct
alternatives for general video game playing,” in The IJCAI-16 Workshop
on General Game Playing, 2016, p. 63.
A. Babadi, B. Omoomi, and G. Kendall, “Enhic: An enforced hill climbing based system for general game playing,” in 2015 IEEE Conference
on Computational Intelligence and Games (CIG). IEEE, 2015, pp.
193–199.
M. J. Nelson, “Investigating vanilla mcts scaling on the gvg-ai game
corpus,” in 2016 IEEE Conference on Computational Intelligence and
Games (CIG). IEEE, 2016, pp. 402–408.
R. D. Gaina, J. Liu, S. M. Lucas, and D. Pérez-Liébana, “Analysis
of vanilla rolling horizon evolution parameters in general video game
playing,” in European Conference on the Applications of Evolutionary
Computation. Springer, 2017, pp. 418–434.
B. Jia, M. Ebner, and C. Schack, “A gp-based video game player,” in
Proceedings of the 2015 Annual Conference on Genetic and Evolutionary Computation. ACM, 2015, pp. 1047–1053.
B. Jia and M. Ebner, “A strongly typed gp-based video game player,”
in 2015 IEEE Conference on Computational Intelligence and Games
(CIG). IEEE, 2015, pp. 299–305.
R. D. Gaina, S. M. Lucas, and D. Pérez-Liébana, “Population Seeding
Techniques for Rolling Horizon Evolution in General Video Game Playing,” in 2017 IEEE Conference on Evolutionary Computation (CEC).
IEEE, 2017.
——, “Rolling Horizon Evolution Enhancements in General Video
Game Playing,” in 2017 IEEE Conference on Computational Intelligence
and Games (CIG). IEEE, 2017.
D. Perez Liebana, J. Dieskau, M. Hunermund, S. Mostaghim, and
S. Lucas, “Open loop search for general video game playing,” in Proceedings of the 2015 Annual Conference on Genetic and Evolutionary
Computation. ACM, 2015, pp. 337–344.
H. Horn, V. Volz, D. Pérez-Liébana, and M. Preuss, “Mcts/ea hybrid
gvgai players and game difficulty estimation,” in Computational Intelligence in Games (CIG), 2013 IEEE Conference on. IEEE, 2013, pp.
1–8.
J. van Eeden, “Analysing and improving the knowledge-based fast
evolutionary mcts algorithm,” 2015.
C.-Y. Chu, S. Ito, T. Harada, and R. Thawonmas, “Position-based
reinforcement learning biased mcts for general video game playing,”
in 2016 IEEE Conference on Computational Intelligence and Games
(CIG). IEEE, 2016, pp. 444–451.
E. Ilhan, Erdem, and Ş. Uyar, “Monte Carlo Tree Search with TemporalDifference Learning for General Video Game Playing,” in 2017 IEEE
Conference on Computational Intelligence and Games (CIG). IEEE,
2017.
H. Van Seijen, A. R. Mahmood, P. M. Pilarski, M. C. Machado, and R. S.
Sutton, “True online temporal-difference learning,” Journal of Machine
Learning Research, vol. 17, no. 145, pp. 1–40, 2016.
I. Bravi, A. Khalifa, C. Holmgård, and J. Togelius, “Evolving gamespecific ucb alternatives for general video game playing,” in European
Conference on the Applications of Evolutionary Computation. Springer,
2017, pp. 393–406.
C. F. Sironi, J. Liu, D. Perez-Liebana, R. D. Gaina, I. Bravi, S. M.
Lucas, and W. M. H. M, “Self-Adaptive MCTS for General Video Game
Playing,” in European Conference on the Applications of Evolutionary
Computation. Springer, 2018, p. to appear.
K. Kunanusont, R. D. Gaina, J. Liu, D. Perez-Liebana, and S. M. Lucas,
“The n-tuple bandit evolutionary algorithm for automatic game improvement,” in Evolutionary Computation (CEC), 2017 IEEE Congress on.
IEEE, 2017, pp. 2201–2208.
B. Ross, “General video game playing with goal orientation,” Master’s
thesis, University of Strathclyde, 2014.
[54] P. Bontrager, A. Khalifa, A. Mendes, and J. Togelius, “Matching games
and algorithms for general video game playing,” in Twelfth Artificial
Intelligence and Interactive Digital Entertainment Conference, 2016.
[55] A. Mendes, A. Nealen, and J. Togelius, “Hyperheuristic general video
game playing,” Proceedings of Computational Intelligence and Games
(CIG). IEEE, 2016.
[56] J. M. Gonzalez-Castro and D. Perez-Liebana, “Opponent Models Comparison for 2 Players in GVGAI Competitions,” in Computer Science
and Electronic Engineering Conference (CEEC), 2017 9th. IEEE, 2017.
[57] R. S. Sutton and A. G. Barto, Reinforcement learning: An introduction.
MIT press Cambridge, 1998, vol. 1, no. 1.
[58] S. Samothrakis, D. Perez-Liebana, S. M. Lucas, and M. Fasli, “Neuroevolution for general video game playing,” in 2015 IEEE Conference
on Computational Intelligence and Games (CIG). IEEE, 2015, pp.
200–207.
[59] A. Braylan and R. Miikkulainen, “Object-model transfer in the general
video game domain,” in Twelfth Artificial Intelligence and Interactive
Digital Entertainment Conference, 2016.
[60] K. Kunanusont, “General video game artificial intelligence: Learning
from screen capture,” Master’s thesis, University of Essex, 2016.
[61] K. Kunanusont, S. M. Lucas, and D. Pérez-Liébana, “General Video
Game AI: Learning from Screen Capture,” in 2017 IEEE Conference on
Evolutionary Computation (CEC). IEEE, 2017.
[62] V. Mnih, K. Kavukcuoglu, D. Silver, A. A. Rusu, J. Veness, M. G.
Bellemare, A. Graves, M. Riedmiller, A. K. Fidjeland, G. Ostrovski
et al., “Human-level control through deep reinforcement learning,”
Nature, vol. 518, no. 7540, p. 529, 2015.
[63] S. O. Kimbrough, G. J. Koehler, M. Lu, and D. H. Wood, “On a feasible–
infeasible two-population (fi-2pop) genetic algorithm for constrained
optimization: Distance tracing and no free lunch,” European Journal
of Operational Research, vol. 190, no. 2, pp. 310–327, 2008.
[64] J. Nichols, “The use of genetic algorithms in automatic level generation,”
Master’s thesis, University of Essex, 2016.
[65] X. Neufeld, S. Mostaghim, and D. Perez-Liebana, “Procedural level generation with answer set programming for general video game playing,”
in Computer Science and Electronic Engineering Conference (CEEC),
2015 7th. IEEE, 2015, pp. 207–212.
[66] T. S. Nielsen, G. A. Barros, J. Togelius, and M. J. Nelson, “Towards
generating arcade game rules with vgdl,” in Computational Intelligence
and Games (CIG), 2015 IEEE Conference on. IEEE, 2015, pp. 185–
192.
[67] T. Machado, I. Bravi, Z. Wang, A. Nealen, and J. Togelius, “Shopping
for game mechanics,” 2016.
[68] T. Machado, A. Nealen, and J. Togelius, “Cicero: A mixed-initiative
AI-assisted game design tool,” in Proceedings of the 12th International
Conference on the Foundations of Digital Games, ser. FDG’17. New
York, NY, USA: ACM, 2017.
[69] ——, “SeekWhence a Retrospective Analysis Tool for General
Game Design,” in Proceedings of the 12th International Conference
on the Foundations of Digital Games, ser. FDG ’17. New
York, NY, USA: ACM, 2017, pp. 4:1–4:6. [Online]. Available:
http://doi.acm.org/10.1145/3102071.3102090
[70] J. Liu, J. Togelius, D. Pérez-Liébana, and S. M. Lucas, “Evolving
game skill-depth using general video game ai agents,” in 2017 IEEE
Conference on Evolutionary Computation (CEC). IEEE, 2017.
[71] C. Guerrero-Romero, A. Louis, and D. Pérez-Liébana, “Beyond Playing
to Win: Diversifying Heuristics for GVGAI,” in 2017 IEEE Conference
on Computational Intelligence and Games (CIG). IEEE, 2017.
[72] A. Khalifa, A. Isaksen, J. Togelius, and A. Nealen, “Modifying mcts
for human-like general video game playing.” in IJCAI, 2016, pp. 2514–
2520.
[73] M. C. Green, A. Khalifa, G. A. Barros, and J. Togelius, ““Press
Space To Fire”: Automatic Video Game Tutorial Generation,” in Fourth
Experimental AI in Games Workshop, 2017.
[74] D. Anderson, M. Stephenson, J. Togelius, C. Salge, J. Levine, and
J. Renz, “Deceptive games,” arXiv preprint arXiv:1802.00048, 2018.
[75] T. S. Nielsen, G. A. Barros, J. Togelius, and M. J. Nelson, “General
video game evaluation using relative algorithm performance profiles,” in
European Conference on the Applications of Evolutionary Computation.
Springer, 2015, pp. 369–380.
[76] S. M. Lucas, J. Liu, and D. Perez-Liebana, “The n-tuple bandit
evolutionary algorithm for game agent optimisation,” arXiv preprint
arXiv:1802.05991, 2018.
[77] D. Pérez-Liébana, S. Samothrakis, J. Togelius, T. Schaul, and S. M.
Lucas, “Analyzing the robustness of general video game playing agents,”
in Proceedings of the IEEE Conference on Computational Intelligence
and Games, 2016.
[78] X. Neufeld, “Procedural level generation with answer set programming
for general video game playing,” Master’s thesis, University of Magdeburg, 2016.
[79] R. D. Gaina, “The 2 player general video game playing competition,”
Master’s thesis, University of Essex, 2016.
[80] V. Volz, D. Ashlock, S. Colton, S. Dahlskog, J. Liu, S. Lucas, D. PerezLiebana, and T. Thompson, “Gameplay evaluation measures,” Dagstuhl
Follow-Ups, 2017.
[81] B. Pell, “Metagame in symmetric chess-like games,” 1992.
[82] A. Isaksen, D. Gopstein, J. Togelius, and A. Nealen, “Discovering unique
game variants,” in Computational Creativity and Games Workshop at the
2015 International Conference on Computational Creativity, 2015.
[83] P. R. Williams, D. Perez-Liebana, and S. M. Lucas, “Ms. Pac-Man Versus Ghost Team CIG 2016 Competition,” in Computational Intelligence
and Games (CIG), 2016 IEEE Conference on, 2016.
[84] P. Hingston, “The 2k botprize,” in Computational Intelligence and
Games, 2009. CIG 2009. IEEE Symposium on. IEEE, 2009, pp. 1–1.
[85] N. Shaker, J. Togelius, G. N. Yannakakis, L. Poovanna, V. S. Ethiraj,
S. J. Johansson, R. G. Reynolds, L. K. Heether, T. Schumann, and
M. Gallagher, “The turing test track of the 2012 mario ai championship:
entries and evaluation,” in Computational Intelligence in Games (CIG),
2013 IEEE Conference on. IEEE, 2013, pp. 1–8.
[86] J. Lehman and R. Miikkulainen, “General video game playing as a
benchmark for human-competitive ai,” in AAAI-15 Workshop on Beyond
the Turing Test, 2015.
[87] G. Brockman, V. Cheung, L. Pettersson, J. Schneider, J. Schulman, J. Tang, and W. Zaremba, “Openai gym,” arXiv preprint
arXiv:1606.01540, 2016.
[88] M. Johnson, K. Hofmann, T. Hutton, and D. Bignell, “The malmo
platform for artificial intelligence experimentation.” in IJCAI, 2016, pp.
4246–4247.
[89] D. Silver, J. Schrittwieser, K. Simonyan, I. Antonoglou, A. Huang,
A. Guez, T. Hubert, L. Baker, M. Lai, A. Bolton et al., “Mastering
the game of go without human knowledge,” Nature, vol. 550, no. 7676,
p. 354, 2017.
| 2cs.AI
|
arXiv:1405.3675v1 [cs.PL] 14 May 2014
Under consideration for publication in Theory and Practice of Logic Programming
1
Abstract Diagnosis for tccp using a Linear
Temporal Logic
MARCO COMINI, LAURA TITOLO
DIMI, Università degli Studi di Udine, Italy
(e-mail: [email protected])
ALICIA VILLANUEVA
DSIC, Universitat Politècnica de València, Spain
(e-mail: [email protected])
submitted 1 January 2003; revised 1 January 2003; accepted 1 January 2003
Abstract
Automatic techniques for program verification usually suffer the well-known state explosion
problem. Most of the classical approaches are based on browsing the structure of some form of
model (which represents the behavior of the program) to check if a given specification is valid.
This implies that a part of the model has to be built, and sometimes the needed fragment is
quite huge.
In this work, we provide an alternative automatic decision method to check whether a given
property, specified in a linear temporal logic, is valid w.r.t. a tccp program. Our proposal (based
on abstract interpretation techniques) does not require to build any model at all. Our results
guarantee correctness but, as usual when using an abstract semantics, completeness is lost.
KEYWORDS: concurrent constraint paradigm, linear temporal logic, abstract diagnosis, decision procedures, program verification
1 Introduction
The Concurrent Constraint Paradigm (ccp, (Saraswat 1989)) is a simple, logic model
which is different from other (concurrent) programming paradigms mainly due to the
notion of store-as-constraint that replaces the classical store-as-valuation model. It is
based on an underlying constraint system that handles constraints on variables and
deals with partial information. Within this family, (de Boer et al. 2000) introduced the
Timed Concurrent Constraint Language (tccp in short) by adding to the original ccp
model the notion of time and the ability to capture the absence of information. With
these features, one can specify behaviors typical of reactive systems such as timeouts or
preemption actions.
It is well-known that modeling and verifying concurrent systems by hand can be an
extremely hard task. Thus, the development of automatic formal methods is essential.
One of the most known techniques for formal verification is model checking, that was
originally introduced in (Clarke and Emerson 1981; Queille and Sifakis 1982) to automatically check if a finite-state system satisfies a given property. It consisted in an exhaustive
analysis of the state-space of the system; thus the state-explosion problem is its main
drawback and, for this reason, many proposals in the literature try to mitigate it.
All the proposals of model checking have in common that a part of the model of the
(target) program has to be built, and sometimes the needed fragment is quite huge. In this
work, we propose a completely different approach to the formal verification of temporal
(LTL) properties of concurrent (reactive) systems specified in tccp. We formalize a method
to validate a specification of the expected behavior of a tccp program P , expressed by a
linear temporal formula φ, which does not require to build any model at all.
The linear temporal logic we use to express specifications, csLTL, is an adaptation of
the propositional LTL logic to the concurrent constraint framework. This logic is also
used as the basis of the abstract domain for a new (abstract) semantics for the language.
In brief, our method is an extension of abstract diagnosis for tccp (Comini et al. 2011)
where the abstract domain F is formed by csLTL formulas. We cannot use the original
abstract diagnosis framework of (Comini et al. 2011) since F is not a complete lattice.
The contributions of this work are the following:
● A new abstract semantics for tccp programs based on csLTL formulas;
● A novel and effective method to validate csLTL properties based on the ideas of
abstract diagnosis. This proposal intuitively consists in viewing P as a formula
transformer by means of an (abstract) immediate consequence operator D̂JP K which
works on csLTL formulas. Then, to decide the validity of φ, we just have to check
if D̂JP Kφ (i.e., the P -transformation of φ) implies φ;
● An automatic decision procedure for csLTL properties that makes our method effective.
With our technique we can check, for instance, that, at a railway crossing system,
each time a train is approaching, the gate is down, or that whenever a train has crossed,
the gate is up. When a property is non valid, the method identifies the buggy process
declaration. Technical results of Sections 3 and 4 can be found in (Comini et al. 2014).
2 The small-step operational behavior of the tccp language
The tccp language (de Boer et al. 2000) is particularly suitable to specify reactive and
time critical systems. As the other languages of the ccp paradigm (Saraswat 1993), it
is parametric w.r.t. a cylindric constraint system which handles the data information of
the program in terms of constraints. The computation progresses as the concurrent and
asynchronous activity of several agents that can accumulate information in a store, or
query information from it. Briefly, a cylindric constraint system1 is an algebraic structure
C = ⟨C, ⪯, ⊗, false, true, Var , ∃ ⟩ composed of a set of constraints C such that (C, ⪯) is a
complete algebraic lattice where ⊗ is the lub operator and false and true are respectively
the greatest and the least element of C; Var is a denumerable set of variables and ∃
existentially quantifies variables over constraints. The entailment ⊢ is the inverse of ⪯.
Given a cylindric constraint system C and a set of process symbols Π, the syntax of
agents is given by the grammar:
A ∶∶= skip ∣ tell(c) ∣ A ∥ A ∣ ∃x A ∣ ∑ni=1 ask(ci ) → A ∣ now c then A else A ∣ p(⃗
x)
1
See (de Boer et al. 2000; Saraswat 1993) for more details on cylindric constraint systems.
2
⟨tell(c), d⟩ → ⟨skip, c ⊗ d⟩
d ≠ false
⟨∑n
i=1 ask(ci ) → Ai , d⟩ → ⟨Aj , d⟩
j ∈ [1, n], d ⊢ cj , d ≠ false
⟨A, d⟩ → ⟨A′ , d′ ⟩
d⊢c
⟨now c then A else B, d⟩ → ⟨A′ , d′ ⟩
⟨A, d⟩ →
/
d ⊢ c, d ≠ false
⟨now c then A else B, d⟩ → ⟨A, d⟩
⟨B, d⟩ → ⟨B ′ , d′ ⟩
d⊬c
⟨now c then A else B, d⟩ → ⟨B ′ , d′ ⟩
⟨B, d⟩ →
/
d⊬c
⟨now c then A else B, d⟩ → ⟨B, d⟩
⟨A, d⟩ → ⟨A′ , d′ ⟩ ⟨B, d⟩ → ⟨B ′ , c′ ⟩
⟨A ∥ B, d⟩ → ⟨A′ ∥ B ′ , d′ ⊗ c′ ⟩
⟨A, d⟩ → ⟨A′ , d′ ⟩ ⟨B, d⟩ →
/
⟨A ∥ B, d⟩ → ⟨A′ ∥ B, d′ ⟩
⟨B ∥ A, d⟩ → ⟨B ∥ A′ , d′ ⟩
⟨A, l ⊗ ∃x d⟩ → ⟨B, l′ ⟩
⟨∃l x A, d⟩ → ⟨∃l′ x B, d ⊗ ∃x l′ ⟩
⟨p(⃗
x), d⟩ → ⟨A, d⟩
p(⃗
x ) ∶− A ∈ D, d ≠ false
Fig. 1. The transition system for tccp.4
where c, c1 , . . . , cn are finite constraints in C; p/m ∈ Π and x⃗ denotes a generic tuple of
m variables. A tccp program is an object of the form D . A, where A is an agent, called
initial agent, and D is a set of process declarations of the form p(⃗
x ) ∶− A (for some agent
A). The notion of time is introduced by defining a discrete and global clock.
The operational semantics of tccp, defined in (de Boer et al. 2000), is formally described by a transition system T = (Conf , →). Configurations in Conf are pairs ⟨A, c⟩
representing the agent A to be executed in the current global store c. The transition relation → ⊆ Conf ×Conf is the least relation satisfying the rules of Figure 1. Each transition
step takes exactly one time-unit.
Example 2.1 (Guiding example) Through the paper, we use as guiding example a
part of the full specification of a railway crossing system introduced in (Alpuente et al.
2006). Let us call Dm the following tccp declaration:
master (C , G) ∶− ∃C ′ , G′ ( now (C = [near ∣ ]) then
′
′
′
′
tell(C = [near ∣ C ]) ∥ tell(G = [down ∣ G ]) ∥ master (C , G )
else now (C = [out ∣ ]) then
tell(C = [out ∣ C ′ ]) ∥ tell(G = [up ∣ G′ ]) ∥ master (C ′ , G ′ )
else master (C , G))
Due to the monotonicity of the store, streams (written in a list-fashion way) are used
to model imperative-style variables (de Boer et al. 2000). The master process uses an
input channel C (implemented as a stream) through which it receives signals from the
environment (trains), and an output channel G through which it sends orders to a gate
process. It checks the input channel for a near signal (the guard in the first now agent),
in which case it sends (tells) the order down through G, links the future values (C ′ )
of the stream C and restarts the check at the following time instant (recursive call
master (C ′ , G ′ )). If the near signal is not detected, then, the else branch looks for the
out signal and (if present) behaves dually to the first branch. Finally, if no signal is
detected at the current time instant (last else branch), then the process keeps checking
from the following time instant.
4
The auxiliary agent ∃l x A makes explicit the local store l of A. This auxiliary agent is linked to the
principal hiding construct by setting the initial local store to true, thus ∃x A ∶= ∃true x A.
3
In this work, we prove the correctness of our technique w.r.t. the denotational concrete semantics of (Comini et al. 2013a), which is fully-abstract (correct and complete)
w.r.t. the small-step operational behavior of tccp. Also csLTL is interpreted over this
denotational model. We thus introduce the most relevant aspects of such semantics.
The denotational semantics of a tccp program consists of a set of conditional (timed)
traces that represent, in a compact way, all the possible behaviors that the program can
manifest when fed with an input (initial store). Conditional traces can be seen as hypothetical computations in which, for each time instant, we have a condition representing
the information that the global store has to satisfy in order to proceed to the next time
instant. Briefly, a conditional trace is a (possibly infinite) sequence t1 ⋯tn ⋯ of conditional
states, which can be of three forms:
conditional store: a pair η ↣ c, where η is a condition and c ∈ C a store;
stuttering: the construct stutt(C), with C ⊆ C ∖ {true};
end of a process: the construct ⊠.
Intuitively, the conditional store η ↣ c means that, provided condition η is satisfied by
the current store, the computation proceeds so that in the following time instant, the
store is c. A condition η is a pair η = (η + , η − ) where η + ∈ C and η − ∈ ℘(C) are called
positive and negative condition, respectively. The positive/negative condition represents
information that a given store must/must not entail, thus they have to be consistent in
the sense that ∀c− ∈ η − η + ⊬ c− . The stuttering construct models the suspension of the
computation when none of the guards in a non-deterministic agent is satisfied. C is the
set of guards in the non-deterministic agent. Conditional traces are monotone (i.e., for
each ti = ηi ↣ ci and tj = ηj ↣ cj such that j ≥ i, cj ⊢ ci ) and consistent (i.e., each store
in a trace does not entail the negative conditions of the following conditional state).
We denote the domain of conditional trace sets as M. (M, ⊑, ⊔, ⊓, M, {ǫ}) is a com¯x r
plete lattice, where M1 ⊑ M2 ⇔ ∀r1 ∈ M1 ∃r2 ∈ M2 . r1 is a prefix of r2 . We define as ∃
the sequence resulting by removing from r ∈ M all the information about the variable x.
We distinguish two special classes of conditional traces. r ∈ M is said to be self-sufficient if
+
−
the first condition is (true, ∅) and, for each ti = (ηi+ , ηi− ) ↣ ci and ti+1 = (ηi+1
, ηi+1
) ↣ ci+1 ,
+
ci ⊢ ηi+1
(each store satisfies the successive condition). Moreover, r is x-self-sufficient if
¯Var ∖{x} r is self-sufficient. Thus, this definition demands that for self-sufficient condi∃
tional traces, no additional information (from other agents) is needed in order to complete
the computation.5
The semantics definition is based on a semantics evaluation function AJAKI (Comini
et al. 2013a) which, given an agent A and an interpretation I , builds the conditional
traces associated to A. The interpretation I is a function which associates to each process symbol a set of conditional traces “modulo variance”. The semantics for a set of
process declarations D is the fixpoint F JDK ∶= lfp (DJDK) of the continuous immediate consequences operator DJDKI (p(⃗
x)) ∶= ⊔p(⃗x )∶−A∈D AJAKI . Proof of full abstraction
w.r.t. the operational behavior of tccp is given in (Comini et al. 2013a).
Example 2.2 Consider the process declaration Dm of Example 2.1. Given an interpretation I , the semantics of master (C , G) is graphically represented in Figure 2, where we
5
The set of all self-sufficient conditional traces can be considered as a generalization (using conditional
states in place of stores) of the traditional strongest postcondition for semantics.
4
(cnear , ∅) ↣ c′near ∧ cdown
¯C ′ ,G′
∃
I (master (C ′ , G ′ ))
(true, {cnear , cout }) ↣ true
(cout , {cnear } ) ↣ c′out ∧ cup
I (master (C , G))
¯C ′ ,G′
∃
′
′
I (master (C , G ))
Fig. 2. Tree representation of DJ{Dm }KI (master (C , G)) of Example 2.2.
have used some shortcuts for characteristic constraints. Namely, cnear ∶= (C = [near ∣ ]),
c′near ∶= ∃C ′ (C = [near ∣ C ′ ]), cdown ∶= ∃G′ (G = [down ∣ G′ ]), cout ∶= (C = [out ∣ ]),
c′out ∶= ∃C ′ (C = [out ∣ C ′ ]), cup ∶= ∃G′ (G = [up ∣ G′ ]).
The branch on the left represents the computation when a near signal arrives. The
first conditional state requires that cnear holds, thus the constraints c′near and cdown are
concurrently added to the store during that computational step. A recursive call is also
concurrently invoked. Process calls do not modify the store when invoked, but they affect
the store from the following time instant, which is graphically represented by the triangle
labeled with the interpretation of the process. The branch in the middle is taken only
if cout is entailed and cnear is not entailed by the initial store (it occurs in the negative
condition of the first conditional state in that branch). Finally, the branch on the right
represents the case when both cnear and cout are not entailed by the initial store.
3 Abstract semantics for tccp over csLTL formulas
In this section, we present a novel abstract semantics over formulas that approximates
the small-step semantics described in Section 2 and, therefore, the small-step operational
behavior of a tccp program. To this end, we first define an abstract domain of logic
formulas which is a variation of the classical Linear Temporal Logic (Manna and Pnueli
1992). Following (Palamidessi and Valencia 2001; de Boer et al. 2001; de Boer et al. 2002;
Valencia 2005), the idea is to replace atomic propositions by constraints of the underlying
constraint system.
Definition 3.1 (csLTL formulas) Given a cylindric constraint system C, c ∈ C and
x ∈ Var , formulas of the Constraint System Linear Temporal Logic over C are:
˙ ∣ c ∣ ¬˙ φ ∣ φ ∧˙ φ ∣ ∃˙ x φ ∣ ◯ φ ∣ φ U φ.
˙ ∣ false
φ ∶∶= true
csLTLC is the set of all temporal formulas over C (we omit C if clear from the context).
˙ ¬,
˙ false,
˙ ∧˙ , ◯ φ, φ1 U φ2 have the classical logical meaning. The atomic formula c ∈
true,
C states that c has to be entailed by the current store. ∃˙ x φ is the existential quantification
over the set of variables Var. As usual, we use φ1 ∨˙ φ2 as a shorthand for ¬˙ φ1 ∧˙ ¬˙ φ2 ;
˙ U φ and ◻ φ for
φ1 →
˙ φ2 for ¬˙ φ1 ∨˙ φ2 ; φ1 ↔
˙ φ2 for φ1 →
˙ φ2 ∧˙ φ2 →
˙ φ1 ; ◇ φ for true
¬˙ ◇ ¬˙ φ. A constraint formula is an atomic formula c or its negation ¬˙ c. Formulas ◯ φ and
¬˙ ◯ φ are called next formulas. Constraint and next formulas are said to be elementary
˙ φ) are called eventualities.
formulas. Finally, formulas of the form φ1 U φ2 , ◇ φ or ¬(◻
We define the abstract domain F ∶= csLTL/↔˙ (i.e., the domain formed by csLTL formulas
˙
˙ , true,
˙, ⋀
˙ false)
modulo logical equivalence) ordered by →.
˙ The algebraic lattice (F, →,
˙ ⋁
˙ always exist just for finite sets of formulas.
˙ and ⋁
is not complete, since both ⋀
5
The semantics of a temporal formula is typically defined in terms of an infinite sequence
of states which validates it. Here we use conditional traces instead. As usually done in
the context of temporal logics, we define the satisfaction relation ⊧ only for infinite
conditional traces. We implicitly transform finite traces (which end in ⊠) by replicating
the last store infinite times.
Definition 3.2 The semantics of φ ∈ F is given by function γ F ∶ F → M defined as γ F (φ) ∶=
⊔{r ∈ M ∣ r ⊧ φ}, where, for each φ, φ1 , φ2 ∈ csLTL, c ∈ C and r ∈ M, satisfaction relation
⊧ is defined as:
˙
r ⊧ true
+
˙
r⊧
/ false
and
−
(3.1a)
′
+
(η , η ) ↣ d ⋅ r ⊧ c
−
iff η ⊢ c
′
−
(3.1b)
−
−
′
stutt (η ) ⋅ r ⊧ c
r ⊧ ¬˙ φ
iff ∀d ∈ η . c ⊬ d and r ⊧ c
(3.1c)
iff r ⊭ φ
(3.1d)
r ⊧ φ1 ∧˙ φ2
iff r ⊧ φ1 and r ⊧ φ2
1
r ⊧ ◯φ
r ⊧ φ1 U φ2
r ⊧ ∃˙ x φ
iff r ⊧ φ
(3.1e)
6
(3.1f)
i
j
iff ∃i ≥ 1. ∀j < i. r ⊧ φ2 and r ⊧ φ1
¯
¯x r, r ′ x-self-sufficient and r ′ ⊧ φ
iff exists r s.t. ∃x r ′ = ∃
′
(3.1g)
(3.1h)
We say that φ ∈ F is a sound approximation of R ∈ M if R ⊑ γ F (φ). φ is said to be
satisfiable if there exists r ∈ M such that r ⊧ φ, while it is valid if, for all r ∈ M, r ⊧ φ.
All the cases are fairly standard except (3.1b) and (3.1c). The conditional trace r = (η + ,
η ) ↣ d ⋅ r′ prescribes that η + is entailed by the current store, thus r models all the
constraint formulas c such that η + ⊢ c. We have to note that, by the monotonicity of
the store of tccp computations, the positive conditions in conditional traces contains all
the information previously added in the constraint store. Furthermore, by the definition
of condition, since η + cannot be in contradiction with η − , it holds that neither c is in
contradiction with η − . Thus, the conditional trace stutt(η − ) ⋅ r′ models all the constraint
formulas c that are not in contradiction with the set η − and such that c holds in the
continuation r′ by monotonicity.
−
Lemma 3.3 The function γ F is monotonic, injective and ⊓-distributive.
3.1 csLTL Abstract Semantics
The technical core of our semantics definition is the csLTL agent semantics evaluation
function ÂJAK which, given an agent A and an interpretation Î (for the process symbols
of A), builds a csLTL formula which is a sound approximation of the (concrete) behavior
Π
of A. In the sequel, we denote by AΠ
C the set of agents and DC the set of sets of process
declarations built on signature Π and constraint system C.
Definition 3.4 Let PC ∶= {p(⃗
x) ∣ p ∈ Π, x
⃗ are distinct variables }. An F-interpretation
is a function PC → F modulo variance7 . Two functions I, J∶ PC → F are variants if for
each π ∈ PC there exists a renaming ρ such that (Iπ)ρ = J(πρ). The semantic domain
IF is the set of all F-interpretations ordered by the point-wise extension of →.
˙
6
7
r k denotes the sub-sequence of r starting from state k.
i.e., a family of elements of F, indexed by PC, modulo variance.
6
Definition 3.5 (csLTL Semantics) Given A ∈ AΠ
C and Î ∈ IF , we define the csLTL
semantics evaluation ÂJAKÎ by structural induction as follows.
ÂJskipKÎ ∶= true
ÂJtell(c)KÎ ∶= ◯ c
ÂJA1 ∥ A2 KÎ ∶= ÂJA1 KÎ ∧˙ ÂJA2 KÎ
ÂJp(⃗
x)KÎ ∶= ◯ Î(p(⃗
x))
ÂJ∃x AKÎ ∶= ∃˙ x ÂJAKÎ
˙ ni=1 ¬˙ ci ) ∨˙ ((⋀
˙ ni=1 (ci ∧˙ ◯ ÂJAi K ))
˙ ni=1 ¬˙ ci ) U ⋁
ÂJ∑ni=1 ask(ci ) → Ai KÎ ∶= ◻(⋀
Î
ÂJnow c then A1 else A2 KÎ ∶= (c ∧˙ ÂJA1 KÎ ) ∨˙ (¬˙ c ∧˙ ÂJA2 KÎ )
Given D ∈ DΠ
C we define the immediate consequence operator D̂JDK∶ IF → IF as
˙ {ÂJAK ∣ p(⃗
x ) ∶− A ∈ D}
D̂JDKÎ (p(⃗
x)) ∶= ⋁
Î
We have that  is a sound approximation of A and D̂ is a sound approximation of D.
Π
Theorem 3.6 (Correctness of  and D̂) Let A ∈ AΠ
C , D ∈ DC and Î ∈ IF . Then,
F
F
AJAKγ F (Î) ⊑ γ (ÂJAKÎ ) and DJDKγ F (Î) ⊑ γ (D̂JDKÎ ).
Example 3.7 Consider the process declaration Dm of Example 2.1 and let us use ◯n
to abbreviate the repetition of ◯ n-times. Given Î ∈ IF , with Definition 3.5 we compute
φM (Î) ∶= D̂J{Dm }KÎ (master(C , G))) = φnear (Î) ∨˙ φout (Î) ∨˙ φcwait (Î)
where
φnear (Î) = ∃˙ C ′ ,G′ (C = [near ∣ ] ∧˙ ◯ C = [near ∣ C ′ ] ∧˙ ◯ G = [down ∣ G′ ] ∧˙ ◯ Î(master (C ′ , G ′ )))
′
˙
= [near ∣ ]) ∧˙ ◯ C = [out ∣ C ] ∧˙
φout (Î) = ∃˙ C ′ ,G′ (¬(C
C = [out ∣ ] ∧˙ ◯ G = [up ∣ G′ ] ∧˙ ◯ Î(master (C ′ , G ′ )))
˙
˙
φcwait (Î) = ¬(C
= [near ∣ ]) ∧˙ ¬(C
= [out ∣ ]) ∧˙ ◯ Î(master (C , G))
The three disjuncts of φM (Î) match the three possible behaviors of master (C , G): when
signal near is emitted by the train, when out is emitted, and when no signal arrives.
4 Abstract diagnosis of tccp with csLTL formulas
Since F is not a complete lattice, it is impossible to find for the function γ F an adjoint
function α which forms a Galois Connection ⟨α, γ⟩, and therefore we cannot use the
abstract diagnosis framework for tccp defined in (Comini et al. 2011). Thus, we propose
in this section a new weaker version of abstract diagnosis that works on F 8 .
Given a set of declarations D and Ŝ ∈ IF , which is the specification of the abstract
intended behavior of D over F, we say that
1. D is (abstractly) partially correct w.r.t. Ŝ if F JDK ⊑ γ F (Ŝ).
2. D is (abstractly) complete w.r.t. Ŝ if γ F (Ŝ) ⊑ F JDK.
The differences between F JDK and γ F (Ŝ) are usually called symptoms. Many of the
8
Actually, the proposal is defined using just γ F only for the sake of simplicity. It could easily be defined
parametrically w.r.t. a suitable family of concretization functions.
7
symptoms are just a consequence of some “originating” ones, those which are the direct consequence of errors. The abstract diagnosis determines exactly the “originating”
symptoms and, in the case of incorrectness, the faulty process declarations in D. This
is captured by the definitions of abstractly incorrect process declaration and abstract
uncovered element :9
Definition 4.1 Let D ∈ DΠ
C , R a process declaration for process p, φt ∈ F and Ŝ ∈ IF .
● R is abstractly incorrect w.r.t. Ŝ (on testimony φt ) if φt →
˙ D̂J{R}KŜ (p(⃗
x)) and φt ∧˙
˙
Ŝ(p(⃗
x)) = false.
● φt is an uncovered element for p(⃗
x) w.r.t. Ŝ if φt →
˙ Ŝ(p(⃗
x)) and φt ∧˙ D̂JDKŜ (p(⃗
x)) =
˙
false.
Informally, R is abstractly incorrect if it derives a wrong abstract element φt from the
intended semantics. Dually, φt is uncovered if the declarations cannot derive it from the
intended semantics.
Theorem 4.2 Let D ∈ DΠ
C and Ŝ ∈ IF . (1) If there are no abstractly incorrect process
declarations in D (i.e., D̂JDKŜ →
˙ Ŝ), then D is partially correct w.r.t. Ŝ. (2) If D is
partially correct w.r.t. Ŝ and D has abstract uncovered elements then D is not complete.
Absence of abstractly incorrect declarations is a sufficient condition for partial correctness, but it is not necessary. Because of the approximation, it can happen that a (concretely) correct declaration is abstractly incorrect. Hence, abstract incorrect declarations
are in general just a warning about a possible source of errors. However, an abstract
correct declaration cannot contain an error; thus, no (manual) inspection is needed for
declarations which are not abstractly incorrect. Moreover, as shown by the following
theorem, all concrete errors—that are “visible”—are indeed detected, as they lead to
an abstract incorrectness or abstract uncovered. Intuitively, a concrete error is visible if
we can express a formula φ whose concretization reveals the error (i.e., if the logic is
expressive enough).
Theorem 4.3 Let R be a process declaration for p(⃗
x), S a concrete specification and
Ŝ a sound approximation for S (i.e., S ⊑ γ F (Ŝ)). (1) If DJ{R}KS ⋢ γ F (Ŝ) and it exists
˙
φt such that γ F (φt ) ⊑ DJ{R}KS (p(⃗
x)) and φt ∧˙ Ŝ(p(⃗
x)) = false,
then R is abstractly
incorrect w.r.t. Ŝ (on testimony φt ). (2) If there exists an abstract uncovered element φ
w.r.t. Ŝ, then there exists r ∈ γ F (φ) such that r ∉ DJ{R}KS (p(⃗
x)).
Point 2 says that the concrete error has an abstract symptom which is not hidden by the
approximation on Ŝ and, moreover, there exists a formula φt which can express it.
In the following examples, we borrow from (Alpuente et al. 2006) the notation for last
entailed value of a stream: X =c
˙ holds if the last instantiated value in the stream X is c.
Example 4.4 We verify (for Example 2.1) that each time a near signal arrives from a
train, the order down is sent to a gate process.10 To model this property, we define the
specification (of the property) Ŝdown as
φordersent ∶= Ŝdown (master (C , G)) ∶= ◻(C =near
˙
→
˙ ◇(G=down))
˙
9
10
It is worth noticing that although the notions defined in this section are similar to those defined for
the standard approach, the formal definitions and proofs are different due to the weaker framework.
A more interesting property, namely that, in addition, the gate is eventually down, is verified in
(Comini et al. 2014). Here we have simplified the property due to space limitations.
8
To check whether the program implies the specification (D̂J{Dm }KŜdown →
˙ Ŝdown ) we have
to check if φM (Ŝdown ) →
˙ φordersent (where φM (⋅) is defined in Example 3.7). Each of the
three disjuncts of φM (Ŝdown ) implies φordersent . Thus, by Theorem 4.2, Dm is partially
correct w.r.t. Ŝdown .
When the check of a process declaration R against a specification S fails, our method
reports that R is not partially correct w.r.t. S. If this occurs, the formula testimony for
the possible incorrectness gives useful information to fix the process declaration or check
whether it corresponds to a false positive.
Example 4.5 Now we show how our technique detects an error in a buggy set of declarations. We remove instruction tell(G = [up ∣ G′ ]) in the process declaration Dm (of
Example 2.1). To avoid misunderstandings, we call the modified process master′ and let
R be the new process declaration.
We aim to verify that the order up is sent whenever the signal out is received:
φ ∶= Ŝup (master ′ (C , G)) ∶= ◻((C =out
˙
)→
˙ ◇(G=up))
˙
We need to compute the (one step) semantics for the (buggy version of the) process:
φ′ ∶= D̂J{R}KŜup (master ′ (C , G)) = φ′near ∨˙ φ′out ∨˙ φ′cwait
where
˙ G = [down ∣ G′ ] ∧˙ ◯ Ŝup (master ′ (C ′ , G ′ )))
φ′near ∶= ∃˙ C ′ ,G′ (C = [near ∣ ] ∧˙ ◯ C = [near ∣ C ′ ] ∧◯
˙
˙
= [out ∣ C ′ ] ∧˙ ◯ Sˆup (master ′ (C ′ , G ′ ))))
= [near ∣ ]) ∧˙ C = [out ∣ ]∧◯(C
φ′out ∶= ∃˙ C ′ ,G′ (¬(C
′
′
˙
˙
= [near ∣ ]) ∧˙ ¬(C
φcwait ∶= ¬(C
= [out ∣ ]) ∧˙ ◯ Ŝup (master (C , G))
We detect an incorrectness of R (in master′ process) w.r.t. Ŝup on testimony φ′out since
˙
φ′out →
˙ φ′ and φ′out ∧˙ φ = false.
The testimony suggests that on channel C we have out
signal but we do not see the corresponding up signal on channel G.
Our technique behaves negatively for sets of declarations D where D̂JDK has more than
one fixpoint. This happens with programs with loops that do not produce contributes at
all (which are in some sense non meaningful programs). In such situations, we can have
that the actual behavior does not model a specification Ŝ which is a non-least fixpoint of
D̂JDK, but, since Ŝ is a fixpoint, we do not detect the abstractly incorrect declaration,
as shown by the following example.
Example 4.6 (Pathological cases) Let Dp ∶= {q(y) ∶− now y = 1 then q(y) else q(y)}
and Ŝp (q(y)) ∶= ◇(y = 1) be the specification. Then, we compute D̂JDp KŜp (q(y)) =
˙ ◇(y = 1), thus Dp is
(y = 1 ∧˙ ◇ y = 1) ∨˙ (¬˙ y = 1 ∧˙ ◇ y = 1). We can see that D̂JDp KŜp →
partially correct w.r.t. Ŝp . However, y = 1 is not explicitly added by the process.
Note that, if Ŝ(p(⃗
x)) is assumed to hold for each process p(⃗
x) defined in D and
D̂JDKŜ →
˙ Ŝ, then F JDK satisfies Ŝ.
4.1 An automatic decision procedure for csLTL
In order to make our abstract diagnosis approach effective, we have defined an automatic
decision procedure to check the validity of the formulas involved in Definition 4.1 (of the
9
Table 1. α- and β-formulas rules.
R1
α
A(α)
¬˙ ¬˙ φ
{φ}
R2 φ1 ∧˙ φ2
R3
¬˙ ◯ φ
B1 (β)
B2 (β)
˙ 1 ∧˙ φ2 )
R4 ¬(φ
{¬˙ φ1 }
{¬˙ φ2 }
{φ1 , φ2 }
˙ 1 U φ2 )
R5 ¬(φ
{¬˙ φ1 , ¬˙ φ2 }
{φ1 , ¬˙ φ2 , ¬˙ ◯(φ1 U φ2 )}
{◯ ¬˙ φ}
R6
{φ2 }
{φ1 , ¬˙ φ2 , ◯((Γ∗ ∧˙ φ1 ) U φ2 )}
β
φ1 U φ2
form ψ →
˙ φ with φ = Ŝ(p(⃗
x)) and ψ = D̂JDKŜ (p(⃗
x))). We adapt to csLTL the tableau
construction for Propositional LTL of (Gaintzarain et al. 2008; Gaintzarain et al. 2009).
(Comini et al. 2013b) contains a preliminary version of the method.
Intuitively, a tableau consists of a tree whose nodes are labeled with sets of formulas.
The root is labeled with the set of formulas which has to be checked for satisfiability.
Branches are built according to rules defined on the syntax of formulas (see Table 1
defining α and β formulas). The basic idea is that a formula from a node is selected and,
depending on its form, a rule of Table 1 is applied. β formulas generate a bifurcation on
the tree and there are specific rules for next and existential quantification formulas.
If all branches of the tree are closed (Definition 4.8), then the formula has no models.
Otherwise, we can obtain a model from the open branches.
Definition 4.7 (csLTL tableau) A csLTL tableau for a finite set of formulas Φ is a
tuple TΦ = (Nodes, nΦ , L, B , R ) such that:
1. Nodes is a finite non-empty set of nodes;
2. nΦ ∈ Nodes is the initial node;
3. L ∶ Nodes → ℘(csLTL) is the labeling function that associates to each node the formulas
which are true in that node; the initial node is labeled with Φ;
4. B is the set of branches such that exactly one of the following points holds for every
branch b = n0 , . . . , ni , ni+1 , . . . , nl ∈ B and every 0 ≤ i < l:
(a) for an α-formula α ∈ L(ni ), L(ni+1 ) = {A(α)} ∪ L(ni ) ∖ {α};
(b) for a β-formula β ∈ L(ni ), L(ni+1 ) = {B1 (β)}∪ L(ni )∖{β} and there exists another branch
in B of the form b′ = n0 , . . . , ni , n′i+1 , . . . , n′k such that L(n′i+1 ) = {B2 (β)} ∪ L(ni ) ∖ {β} ;
(c) for an existential quantified formula ∃˙ x φ′ ∈ L(ni ), L(ni+1 ) = {φ′′ } ∪ L(ni ) ∖ {∃˙ x φ′ } where
φ′′ ∶= φ′ [y/x] with y fresh variable;
(d) in case L(ni ) is a set formed only by elementary formulas, L(ni+1 ) = next(L(ni )), where
next(Φ) ∶= {φ ∣ ◯ φ ∈ Φ} ∪ {¬˙ φ ∣ ¬˙ ◯ φ ∈ Φ} ∪ (Φ ∩ C).
Rules 4a and 4b are standard, replacing α and β-formulas with one or two formulas
according to the matching pattern of rules in Table 1, except for Rule R6 that uses the
so-called context Γ∗ , which is defined in the following. The next operator used in Rule 4d
is different from the corresponding one of PLTL since it also preserves the constraint
formulas. This is needed for guaranteeing correctness in the particular setting of tccp
where the store is monotonic. Finally, Rule 4c is specific for the ∃˙ case: ∃˙ x is removed
after renaming x with a fresh variable11 .
Definition 4.8 A node in the tableau is inconsistent if it contains a couple of formulas
˙
φ, ¬˙ φ, or the formula false,
or a constraint formula ¬˙ c′ such that the merge c of all the
11
The csLTL existential quantification does not correspond to the one of FO logic. It serves to model
local variables, and ∃˙ x φ can be seen just as φ where the information about x is local.
10
(positive) constraint formulas c1 , . . . , cn in the node (i.e., c ∶= c1 ⊗ ⋅ ⋅ ⋅ ⊗ cn ) is such that
c ⊢ c′ . A branch is closed if it contains an inconsistent node.
The last condition for inconsistence of a node is particular to the ccp context.
We now describe the algorithm that automatically builds the csLTL tableau for a given
set of formulas Φ (see (Comini et al. 2014) for the pseudocode). The construction consists in selecting at each step a branch that can be extended by using α or β rules or ∃˙
elimination. When none of these can be applied, the next operator is used to pass to the
next stage. When dealing with eventualities, to determine the context Γ∗ in Rule R6, it is
necessary to distinguish the eventuality that is being unfolded in the path. Given a node
n and φ ∈ L(n), Γ ∶= L(n) ∖ {φ}. Then, when Rule R6 is applied to a distinguished even˙ γ∈Γ ¬˙ γ; otherwise Γ∗ ∶= true. The use of contexts is the mechanism
tuality, we set Γ∗ ∶= ⋁
to detect the loops that allows one to mark branches containing eventuality formulas as
open or to generate inconsistent nodes and mark branches as closed. A node is marked
as closed when it is inconsistent while is marked as open when (1) it is the last node of
the branch and contains just constraint formulas or (2) the branch is cyclic and all the
eventualities in the cycle have been already distinguished.
In order to ensure termination of the algorithm, it is necessary to use a fair strategy
to distinguish eventualities, in the sense that every eventuality in an open branch must
be distinguished at some point. This assumption and the fact that, given a finite set of
initial formulas, there exists only a finite set of possible labels in a systematic tableau,
imply termination of the tableau construction. Moreover, the constructed tableau is sound
and complete. Therefore, to check the validity of a formula of the form ψ →
˙ φ, with φ =
Ŝ(p(⃗
x)) and ψ = D̂JDKŜ (p(⃗
x)), we just have to build the tableau for its negation T¬(ψ
˙ →φ)
˙
and check if it is closed or not. If it is, we have that D is abstractly correct. Otherwise,
we can extract from T¬(ψ
an explicit testimony ϕ of the abstract incorrectness of D.
˙ →φ)
˙
The construction of ψ = D̂JDKŜ (p(⃗
x)) is linear in the size of D. The systematic
˙
tableau construction of ¬(ψ
→
˙ φ) (from what said in (Gaintzarain et al. 2009)) has worst
˙
∣ ¬(ψ
→φ)∣
˙
)
case O(2O(2
). However, we believe that such bound for the worst-case asymptotic
behavior is quite meaningless in this context, since it is not very realistic to think that the
formulas of the specification should grow much (big formulas are difficult to comprehend
and in real situations people would hardly try even to imagine them). Moreover, note that
tableau explosion is due to nesting of eventualities and in practice really few eventualities
are used in specifications. Therefore, in real situations, we do not expect that (extremely)
big tableaux will be built.
5 Related Work
A Constraint Linear Temporal Logic is defined in (Valencia 2005) for the verification of
a different timed concurrent language, called ntcc, which shares with tccp the concurrent
constraint nature and the non-monotonic behavior. The restricted negation fragment of
this logic, where negation is only allowed for state formulas, is shown to be decidable.
However, no efficient decision procedure is given (apart from the proof itself). Moreover,
the verification results are given for the locally-independent fragment of ntcc, which
avoids the non-monotonicity of the original language. In contrast, in this work, we address
the problem of checking temporal properties for the full tccp language.
11
Some model-checking techniques have been defined for tccp in the past (Falaschi and
Villanueva 2006; Alpuente et al. 2005a; Alpuente et al. 2005b; Falaschi et al. 2001).
It is worth noting that the notions of correctness and completeness in these works are
defined in terms of F JDK, i.e., in terms of the concrete semantics, and therefore their
check requires a (potentially infinite) fixpoint computation. In contrast, the notions of
abstractly incorrect declarations and abstract uncovered elements are defined in terms of
just one application of D̂JDK to Ŝ. Moreover, since D̂JDK is defined compositionally, all
the checks are defined on each process declaration in isolation. Hence, our proposal can
be used with partial sets of declarations. When a property is falsified, model checking
provides a counterexample in terms of an erroneous execution trace, leaving to the user
the problem of locating the source of the bug. On the contrary, we identify the faulty
process declaration.
In (Falaschi et al. 2007), a first approach to the declarative debugging of a ccp language is presented. However, it does not cover the particular extra difficulty of the nonmonotonicity behavior, common to all timed concurrent constraint languages. This makes
our approach significantly different. Moreover, although they propose the use of LTL for
the specification of properties, their formulation, based on the depth-k concretization
function, complicates the task of having an efficient implementation.
Finally, this proposal clearly relates to the abstract diagnosis framework for tccp defined for Galois Insertions (Comini et al. 2011). That work can compete with the precision
of model checking, but its main drawback is the fact that the abstract domain did not
allow to specify temporal properties in a compact way. In fact, specifications consisted
of sets of abstract conditional traces. Thus, specifications were big and unnatural to be
written. The use of temporal logic in this proposal certainly overcomes this problem.
6 Conclusion and Future Work
We have defined an abstract semantics for tccp based on the domain of a linear temporal
logic with constraints. The semantics is correct w.r.t. the behavior of the language.
By using this abstract semantics, we have defined a method to validate csLTL formulas
for tccp sets of declarations. Since the abstract semantics cannot be defined by means of
a Galois Connection, we cannot use the abstract diagnosis framework for tccp defined in
(Comini et al. 2011), thus we devised (from scratch) a weak version of the abstract diagnosis framework based only on a concretization function γ. It works by applying D̂JDK to
the abstract specification and then by checking the validity of the resulting implications
(whether that computation implies the abstract specification). The computational cost
depends essentially on the cost of that check of the implication.
We have also presented an automatic decision procedure for the csLTL logic, thus
we can effectively check the validity of that implication. We are currently finishing to
implement a proof of concept tool, which is available online at URL http://safe-tools.
dsic.upv.es/tadi/, that realizes the proposed instance. Then we would be able to
compare with other tools and assess the “real life” goodness of our proposal.
In the future, we also plan to explore other instances of the method based on logics
for which decision procedures or (semi)automatic tools exists. This proposal can also be
immediately adapted to other concurrent (non-monotonic) languages (like tcc and ntcc)
once a suitable fully abstract semantics has been developed.
12
References
Alpuente, M., Falaschi, M., and Villanueva, A. 2005a. A Symbolic Model Checker for
tccp Programs. In First International Workshop on Rapid Integration of Software Engineering
Techniques (RISE 2004), Revised Selected Papers. Lecture Notes in Computer Science, vol.
3475. Springer-Verlag, 45–56.
Alpuente, M., Gallardo, M., Pimentel, E., and Villanueva, A. 2005b. A Semantic
Framework for the Abstract Model Checking of tccp Programs. Theoretical Computer Science 346, 1, 58–95.
Alpuente, M., Gallardo, M., Pimentel, E., and Villanueva, A. 2006. Verifying RealTime Properties of tccp Programs. Journal of Universal Computer Science 12, 11, 1551–1573.
Clarke, E. M. and Emerson, E. A. 1981. Design and synthesis of synchronization skeletons
using branching-time temporal logic. In Logic of Programs, D. Kozen, Ed. Lecture Notes in
Computer Science, vol. 131. Springer, 52–71.
Comini, M., Titolo, L., and Villanueva, A. 2011. Abstract Diagnosis for Timed Concurrent
Constraint programs. Theory and Practice of Logic Programming 11, 4-5, 487–502.
Comini, M., Titolo, L., and Villanueva, A. 2013a. A Condensed Goal-Independent BottomUp Fixpoint Modeling the Behavior of tccp. Tech. rep., DSIC, Universitat Politècnica de
València. Available at http://riunet.upv.es/handle/10251/34328.
Comini, M., Titolo, L., and Villanueva, A. 2013b. Towards an Effective Decision Procedure for LTL formulas with Constraints. In 23rd Workshop on Logic-based methods in
Programming Environments (WLPE 2013). CoRR abs/1308.2055.
Comini, M., Titolo, L., and Villanueva, A. 2014. Abstract Diagnosis for tccp using a
Linear Temporal Logic. Tech. rep., Universitat Politècnica de València. Available at http:
//riunet.upv.es/handle/10251/8351.
de Boer, F. S., Gabbrielli, M., and Meo, M. C. 2000. A Timed Concurrent Constraint
Language. Information and Computation 161, 1, 45–83.
de Boer, F. S., Gabbrielli, M., and Meo, M. C. 2001. A Temporal Logic for Reasoning about
Timed Concurrent Constraint Programs. In TIME ’01: Proceedings of the Eighth International
Symposium on Temporal Representation and Reasoning (TIME’01). IEEE Computer Society,
Washington, DC, USA, 227.
de Boer, F. S., Gabbrielli, M., and Meo, M. C. 2002. Proving correctness of Timed
Concurrent Constraint Programs. CoRR cs.LO/0208042.
Falaschi, M., Olarte, C., Palamidessi, C., and Valencia, F. D. 2007. Declarative Diagnosis
of Temporal Concurrent Constraint Programs. In Logic Programming, 23rd International
Conference, ICLP 2007, Proceedings, V. Dahl and I. Niemelä, Eds. Lecture Notes in Computer
Science, vol. 4670. Springer-Verlag, 271–285.
Falaschi, M., Policriti, A., and Villanueva, A. 2001. Modeling concurrent systems specified in a temporal concurrent constraint language-I. Electronic Notes in Theoretical Computer
Science 48, 197–210.
Falaschi, M. and Villanueva, A. 2006. Automatic verification of timed concurrent constraint
programs. Theory and Practice of Logic Programming 6, 3, 265–300.
Gaintzarain, J., Hermo, M., Lucio, P., and Navarro, M. 2008. Systematic semantic
tableaux for PLTL. Electronic Notes in Theoretical Computer Science 206, 59–73.
Gaintzarain, J., Hermo, M., Lucio, P., Navarro, M., and Orejas, F. 2009. Dual Systems
of Tableaux and Sequents for PLTL. The Journal of Logic and Algebraic Programming 78, 8,
701–722.
Manna, Z. and Pnueli, A. 1992. The temporal logic of reactive and concurrent systems specification. Springer.
Palamidessi, C. and Valencia, F. D. 2001. A Temporal Concurrent Constraint Programming Calculus. In 7th International Conference on Principles and Practice of Constraint
Programming (CP’01). Lecture Notes in Computer Science, vol. 2239. Springer, 302–316.
13
Queille, J. P. and Sifakis, J. 1982. Specification and verification of concurrent systems in
CESAR. In Symposium on Programming, M. Dezani-Ciancaglini and U. Montanari, Eds.
Lecture Notes in Computer Science, vol. 137. Springer, 337–351.
Saraswat, V. A. 1989. Concurrent Constraint Programming Languages. Ph.D. thesis,
Carnegie-Mellon University.
Saraswat, V. A. 1993. Concurrent Constraint Programming. The MIT Press, Cambridge,
Mass.
Valencia, F. D. 2005. Decidability of infinite-state timed CCP processes and first-order LTL.
Theoretical Computer Science 330, 3, 577–607.
14
| 6cs.PL
|
Superrosy fields and valuations
arXiv:1308.3394v1 [math.LO] 15 Aug 2013
Krzysztof Krupiński∗
Abstract
We prove that every non-trivial valuation on an infinite superrosy field of
positive characteristic has divisible value group and algebraically closed residue
field. In fact, we prove the following more general result. Let K be a field such
that for every finite extension L of K and for every natural number n > 0 the
index [L∗ : (L∗ )n ] is finite and, if char(K) = p > 0 and f : L → L is given
by f (x) = xp − x, the index [L+ : f [L]] is also finite. Then either there is a
non-trivial definable valuation on K, or every non-trivial valuation on K has
divisible value group and, if char(K) > 0, it has algebraically closed residue
field. In the zero characteristic case, we get some partial results of this kind.
We also notice that minimal fields have the property that every non-trivial
valuation has divisible value group and algebraically closed residue field.
0
Introduction
A motivation for our work comes from some open structural questions concerning
fields in various model-theoretic contexts.
A fundamental theorem says that each infinite superstable field is algebraically
closed [16, 4]. An important generalization of superstable theories is the class of
supersimple theories and yet more general class of superrosy theories. Superrosy
theories with NIP (the non independence property) also form a generalization of
superstable theories which is “orthogonal” to supersimple theories in the sense that
each supersimple theory with NIP is superstable. It is known from [11] that perfect
PAC (pseudo algebraically closed) fields with small absolute Galois group (i.e. with
absolute Galois group possessing only finitely many closed subgroups of every finite
index) are supersimple. A well-known conjecture predicts the converse:
Conjecture 1 Each infinite supersimple field is perfect PAC with small absolute
Galois group.
Research supported by NCN grant 2012/07/B/ST1/03513
2010 Mathematics Subject Classification: 03C60, 12J10
0
Key words and phrases: superrosy field, valuation
∗
0
1
A complementary conjecture on infinite superrosy fields with NIP was formulated
in [6].
Conjecture 2 Each infinite superrosy field with NIP is either algebraically or real
closed.
Recall that both algebraically closed and real closed fields are superrosy with
NIP. After dropping the NIP assumption, one has to extend the list of possibilities
in the conclusion of the above conjecture. Namely, since perfect PAC fields with
small absolute Galois group as well as orderable PRC (pseudo real closed) fields
with small absolute Galois group are known to be superrosy [18, Appendix A], the
following conjecture is strongest possible. (See Section 4 for the definition of PRC
fields, which is chosen so that PAC fields are PRC).
Conjecture 3 Each infinite superrosy field is perfect PRC with small absolute Galois group.
It is known that a PAC field is simple if and only if its absolute Galois group is
small [11, 2, 3]; it is supersimple if and only if it is perfect and has small absolute
Galois group. Similarly, a PRC field is superrosy if and only if it is perfect and its
absolute Galois group is small (see Fact 4.1). Thus, in Conjectures 1 and 3, once we
know that the field is PAC [resp. PRC], the rest of the conclusion is automatically
satisfied. It is also easy to see that Conjecture 3 implies Conjecture 1, because one
can show that orderable PRC fields have strict order property, and so they are not
simple (see Remark 4.2). By Fact 4.3, Conjecture 3 also implies Conjecture 2.
There are also interesting questions and conjectures concerning NIP fields (without assuming superrosiness). By [12], infinite NIP fields are closed under ArtinSchreier extensions. A. Hasson and S. Shelah formulated some dichotomies between
nice algebraic properties of the field in question and the existence of non-trivial
definable valuations. In particular, one can expect that the following is true.
Conjecture 4 Suppose K is an infinite field with NIP with the property that for
every finite extension L of K and for every natural number n > 0 the index [L∗ :
(L∗ )n ] is finite. Then either there is a non-trivial definable valuation on K, or K is
either algebraically or real closed.
Note that if a pure field K is algebraically or real closed, then there is no non-trivial
definable valuation on K (e.g. because K is superrosy and we have Fact 1.8). Notice
also that by Facts 1.8 and 1.9, Conjecture 4 implies Conjecture 2.
Another interesting problem is to classify strongly dependent fields [20, Section
5].
Independently of the questions of A. Hasson and S. Shelah in the NIP context,
our approach to attack Conjectures 2 and 3 was to assume that the field in question
does not satisfy the conclusion and try to produce a non-trivial definable valuation
(existence of which contradicts rosiness by Fact 1.8). This approach led us to the
2
following conjecture whose assumptions generalize the situations from Conjectures 1,
2, 3 and 4, but whose conclusion is weaker than the conclusions of these conjectures
(see Section 4 for explanations). So, one could say that it is a common approximation
of these conjectures.
Conjecture 5 Let K be a field such that for every finite extension L of K and for
every natural number n > 0 the index [L∗ : (L∗ )n ] is finite and, if char(K) = p > 0
and f : L → L is given by f (x) = xp − x, the index [L+ : f [L]] is also finite. Then
either there is a non-trivial definable valuation on K, or every non-trivial valuation
on K has divisible value group and either algebraically or real closed residue field.
Our main result is the proof of Conjecture 5 in the positive characteristic case.
In fact, we will prove the following theorem.
Theorem 6 Let K be a field such that for every finite extension L of K and for
every natural number n > 0 the index [L∗ : (L∗ )n ] is finite and, if char(K) = p > 0
and f : L → L is given by f (x) = xp − x, the index [L+ : f [L]] is also finite. Then
either there is a non-trivial definable valuation on K, or every non-trivial valuation
on K has divisible value group and, in the case when char(K) > 0, it has algebraically
closed residue field.
By Facts 1.8 and 1.9, one gets the following corollary.
Corollary 7 Every non-trivial valuation on a superrosy field of positive characteristic has divisible value group and algebraically closed residue field.
Since infinite NIP fields are closed under Artin-Schreier extensions [12], we also
get the following corollary.
Corollary 8 Suppose K is a field of positive characteristic, satisfying NIP and with
the property that for every finite extension L of K and for every natural number
n > 0 the index [L∗ : (L∗ )n ] is finite. Then either there is a non-trivial definable
valuation on K, or every non-trivial valuation on K has divisible value group and
algebraically closed residue field.
The proof of Theorem 6 relies on [13], where the appropriate results on the existence of non-trivial definable valuations under the presence of certain multiplicative
or additive subgroups were established. In contrast, directly from the definition of
minimality, we obtain the following variant of Corollary 7 for minimal fields.
Theorem 9 Every non-trivial valuation on a minimal field has divisible value group
and algebraically closed residue field.
3
Recall that the famous Podewski’s conjecture predicts that each minimal field is
algebraically closed [19].
Our results tells us that, in various situations, all non-trivial valuations on the
field in question have divisible value groups and algebraically closed residue fields.
The ultimate goal is to show that the original field (i.e. the residue field with respect
to trivial valuation) is algebraically closed [or PAC]. So a questions arises what
information about the field in question can be deduced from the information that
for all non-trivial valuations the value groups are divisible and the residue fields are
algebraically closed. A discussion and some questions about it are included in the
last section.
The paper is constructed as follows. First we give prelimnaries concerning rosy
theories and valuations, listing all the facts from [13] which are used in the course of
the proof of Theorem 6. Section 2 is devoted to the proof of Theorem 6 and some
partial results concerning Conjecture 5 in the zero characteristic case. A very short
and easy Section 3 is self-contained and yields a proof of Theorem 9. In the last
section, we discuss some facts and questions concerning potential applications of our
results to prove the original conjectures.
Katharina Dupont is currently working around Conjecture 4 in her Ph.D. project
under the supervision of Salma Kuhlmann and in collaboration with Assaf Hasson.
She has a different approach than the one presented in this paper (although also
based on [13]). A few details on this are mentioned in the last section.
The author would like to thank Thomas Scanlon for discussions and suggestions
concerning superrosy fields during the visit at Berkeley in 2007.
1
Preliminaries
1.1
Valuations
In this subsection, we list the definitions and facts from [13] which will be useful in
this paper. But before that, let us recall the definition of valuation and other basic
notions. A good reference for fields with valuations is for example [9].
Definition 1.1 A valuation on a field K is a surjective map v : K → Γ ∪ {∞},
where (Γ, +) is an ordered group and Γ < ∞, satisfying the following axioms. For
all x, y ∈ K:
1. v(x) = ∞ =⇒ x = 0,
2. v(xy) = v(x) + v(y),
3. v(x + y) ≥ min{v(x), v(y)}.
Let v be a valuation on a field K. Define
Ov := {x ∈ K : v(x) ≥ 0}.
4
This is a valuation ring, i.e., a subring of K such that for any x ∈ K, either x ∈ Ov
or x−1 ∈ Ov . We say that the valuation v is trivial if Γ = {0}; equivalently, Ov = K.
The group Ov∗ of units of Ov equals {x ∈ K : v(x) = 0}. Finally,
Mv := {x ∈ K : v(x) > 0}
is a unique maximal ideal of Ov , and
K v := Ov /Mv
is called the residue field of v.
With a valuation v we associated its valuation ring Ov . Conversely, starting from
a valuation ring O, one can define a valuation v so that Ov = O, and one can do it
in such a way that both operations are inverses of each other (after the appropriate
identification of value groups). Namely, having a valuation ring O, we define Γ :=
K ∗ /O∗ with xO∗ + yO∗ := xyO∗ , and we order it by xO∗ ≤ yO∗ ⇐⇒ y/x ∈ O;
then we define a valuation v by v(x) = xO∗ ∈ Γ.
Definition 1.2 We say that a valuation v : K → Γ∪{∞} on a field K (possibly with
an additional structure) is definable if the ordered group Γ is interpretable in K and
after the interpretation of Γ in K, graph(v) is definable. By the above discussion,
this is equivalent to the fact that Ov is definable in K.
If K ⊆ L is a field extension and w is a valuation on L, we say that w extends v if
w↾K = v (after an isomorphic embedding of the value group of v into the value group
of w); equivalently, Ow ∩ K = Ov . In such a situation, we write (K, Ov ) ⊆ (L, Ow ).
Let Γv be the value group of v and Γw the value group of w. Then Γv can be treated
as a subgroup of Γw , and e(Ow /Ov ) := [Γw : Γv ] is called the ramification index
of the extension (K, Ov ) ⊆ (L, Ow ). Similarly, since Mw ∩ Ov = Mv , we get that
K v = Ov /Mv ֒→ Ow /Mw = Lw , and f (Ow /Ov ) := [Lw : K v ] is called the residue
degree of the extension (K, Ov ) ⊆ (L, Ow ).
Fact 1.3 Whenever (K, Ov ) ⊆ (L, Ow ) with n := [L : K] is finite, one has the
following inequality e(Ow : Ov )f (Ow : Ov ) ≤ n.
When v and w are two valuations on the same field K, we say that w is a
coarsening of v if Ov ⊆ Ow . In such a situation, Mw ⊆ Mv is a prime ideal of Ov .
In this way, one gets a 1-1 correspondence between overrings of Ov and prime ideals
of Ov . Going further, one gets a 1-1 correspondence between this set of prime ideals
and the set of convex subgroups of Γv (the value group of v):
∆ 7→ p∆ := {x ∈ K : v(x) > δ for all δ ∈ ∆}
p 7→ ∆p := {γ ∈ Γ : γ, −γ < v(x) for all x ∈ p}.
For details on this, see [9, Chapter 2.3]. Let us add that two valuations v and w on
K are said to be comparable if Ov ⊆ Ow or Ow ⊆ Ov .
5
From the model-theoretic perspective, an important question is when there exists
a non-trivial definable valuation on a given field K. A deep insight into this question
is provided in [13]. In particular, with an additive [resp. multiplicative] subgroup
T the author associates a certain valuation ring, denoted by OT , and he gives a
complete characterization of when OT is first order definable in (K, +, ·, 0, 1, T ).
Here, we recall some definitions and results from [13] which we will use later.
Definition 1.4 Let v be a valuation on K and T an additive [resp. multiplicative]
subgroup.
1. v is compatible with T if Mv ⊆ T [resp. 1 + Mv ⊆ T ].
2. v is weakly
compatible with T if A ⊆ T [resp. 1 + A ⊆ T ] for some Ov -ideal A
√
with A = Mv .
3. v is coarsely compatible with T if it is weakly compatible with T and there is
no proper coarsening w of v such that Ow∗ ⊆ T .
Fact 1.5 [13, Lemma 1.2] Let v be a valuation on K. If either T is a multiplicative
subgroup such that for some n ∈ ω, (K ∗ )n ⊆ T and (n, char(K v )) = 1, or T is
an additive subgroup, char(K) = p and {xp − x : x ∈ K} ⊆ T , then v is (fully)
compatible with T if and only if v is weakly compatible with T .
Fact 1.6 [13, Proposition 1.4] For an additive [resp. multiplicative] subgroup T of a
field K, any two coarsly compatible valuations are comparable, and there is a unique
finest coarsly compatible valuation ring of K which we will denote by OT . Moreover,
OT is non-trivial, whenever T is proper (i.e., T 6= K [resp. T 6= K ∗ ]) and admits
some non-trivial weakly compatible valuation.
The author concludes that for any additive [resp. multiplicative] subgroup T of
a field K exactly one of the following possibilities holds:
• groups case: there is a valuation v on K such that Ov∗ ⊆ T .
In this case, OT is the only coarsely compatible valuation ring with this property, and all weakly compatible valuations are fully compatible.
• weak case: there is a weakly, but not fully compatible valuation on K.
In this case, OT is the only valuation ring with this property; the weakly
compatible valuations are the coarsenings of OT ; there is no valuation v with
Ov∗ ⊆ T .
• residue case: all weakly compatible valuations are fully compatible, and there
is no valuation v with Ov∗ ⊆ T .
In this case, OT is the finest fully compatible valuation ring.
6
Now, we are going to recall [13, Theorem 2.5], which will be the main tool in
this paper. This is a complete characterization of when, for a given additive [resp.
multiplicative] subgroup T of a filed K, the ring OT is definable in the language L :=
{+, ·, 0, 1, T }. In fact, in our applications we will only use the positive part of this
characterization (namely the cases when one has definability). Before we formulate
the full characterization, we should emphasis that here by definability in L we do not
just mean definability in the structure (K, +, ·, 0, 1, T ), but the existence of a formula
ϕ(x) in L which defines OT ′ in every model (K ′ , +, ·, 0, 1, T ′) ≡ (K, +, ·, 0, 1, T ).
Fact 1.7 [13, Theorem 2.5] Let K be a field with an additive [resp. multiplicative]
subgroup T . Denote by MT the maximal ideal of OT , and by T the subgroup induced
by T on the residue field of OT . Then OT is definable in L := {+, ·, 0, 1, T } in the
following cases:
T ≤ (K, +)
T ≤ K∗
group case
iff OT is discrete or (∀x ∈ MT )(x−1 OT * T )
always
weak case
iff OT is discrete
iff OT is discrete
residue case
always
iff T is no ordering
1.2
Rosy theories
In this paper, we will only need two properties of rosy groups and fields. Although
they are a folklore, for the reader’s convenience we recall fundamental definitions
concerning rosiness and we give proofs of these two properties:
Fact 1.8 Let K be a rosy field. Then there is no non-trivial definable valuation on
K.
Fact 1.9 Let G be a commutative superrosy group. Then for every natural number
n > 0, if G[n] := {g ∈ G : g n = e} is finite, the index [G : Gn ] is also finite, where
Gn denotes the subgroup consiting of n-th powers.
In particular, if K is a superrosy field, then for every n > 0 the index [K ∗ : (K ∗ )n ]
is finite and, if char(K) = p 6= 0, then the image of the function f : K → K defined
by f (x) = xp − x is a subgroup of finite index in K + . Since any finite extension of
an elementary extension of K is also superrosy, this holds for all finite extensions of
any elementary extension of K, too.
For details on rosy theories the reader is referred to [1, 7, 18], and on rosy groups
to [6]. More information about rosy groups and fields can be found in [14, 15].
In this subsection, we work in Ceq where C is a monster model of a theory T in a
language L.
A motivation to consider rosy theories is the fact that in a sense it is the largest
class of theories which allows the application of techniques from stability theory,
especially of basic forking calculus. This class contains stable and more generally
simple theories as well as o-minimal theories. The definition of rosiness which justifies
7
what we have just said is the following: T is rosy if there is a ternary relation ⌣
| ∗
on small subsets of Ceq satisfying all the basic properties of forking independence in
simple theories except for the Independence Theorem. Such a relation will be called
an independence relation. There is a concrete, particularly useful independence
relation in rosy theories, called þ-independence, which we are going to define now.
A formula δ(x, a) strongly divides over A if it is non-algebraic and the set of
formulas {δ(x, a′ )}a′ |=tp(a/A) is k-inconsistent for some k ∈ N. We say that δ(x, a)
þ-divides over A if we can find some tuple c such that δ(x, a) strongly divides over
Ac. A formula þ-forks over A if it implies a (finite) disjunction of formulas which
þ-divide over A.
We say that a type p(x) þ-divides over A if there is a formula implied by p(x)
which þ-divides over A; þ-forking is similarly defined. We say that a is þ-independent
from b over A, denoted a ⌣
| þA b, if tp (a/Ab) does not þ-fork over A.
In rosy theories, þ-independence is the weakest independence relation in the sense
that a ⌣
| ∗C b implies a ⌣
| þC b for any independence relation ⌣
| ∗.
By a rosy group [or field] we mean a group [or field], possibly with an additional
structure, whose theory is rosy.
Rosy theories can be also defined by means of local þ-ranks.
Definition 1.10 Given a formula ψ(x), a finite set Φ of formulas with object variables x and parameter variables y, a finite set of formulas Θ in variables y, z, and
natural number k > 0, we define the þΦ,Θ,k -rank of ψ inductively as follows:
1. þΦ,Θ,k (ψ) ≥ 0 if ψ is consistent.
2. For λ a limit ordinal, þΦ,Θ,k (ψ) ≥ λ if þΦ,Θ,k (ψ) ≥ α for all α < λ.
3. þΦ,Θ,k (ψ) ≥ α + 1 if there is ϕ ∈ Φ, some θ(y; z) ∈ Θ and parameter c such
that
(a) þΦ,Θ,k (ψ(x) ∧ ϕ(x; a)) ≥ α for infinitely many a |= θ(y; c), and
(b) {ϕ (x; a)}a|=θ(y;c) is k−inconsistent.
Given a (partial) type π(x) we define þΦ,Θ,k (π(x)) to be the minimum of þΦ,Θ,k (ψ)
for ψ ∈ π(x).
Recall that a theory is rosy if and only if for each ψ, Φ, Θ, k as above, the local
thorn rank þΦ,Θ,k (ψ) is finite. One could prove Fact 1.8 using this characterization.
However, one can give an immediate proof using another characterization of rosiness
in terms of the so-called equivalence ranks considered in [7, Section 5].
Definition 1.11 Let π(x) be a partial type, and let ∆ be a finite set of formulas in
variables x, y, z. Define eq-rk∆ (π(x)) as follows:
1. eq-rk∆ (π(x)) ≥ 0 if π(x) is consistent.
8
2. For λ a limit ordinal, eq-rk∆ (π(x)) ≥ λ if eq-rk∆ (π(x)) ≥ α for all α < λ.
3. eq-rk∆ (π(x)) ≥ α + 1 if there is some equivalence relation E(x, y) defined by
δ(x, y, c) with δ(x, y, z) ∈ ∆ and c ∈ Ceq , and there are representatives bi ,
i < ω, of different equivalence classes, such that eq-rk∆ (π(x) ∧ E(x, bi )) ≥ α.
From [7, Section 5], we know that T is rosy if and only if for every ∆ and π(x)
as above, eq-rk∆ (π(x)) is finite.
Proof of Fact 1.8. Suppose for a contradiction that there is a non-trivial definable
valuation v on the rosy field K. We can assume K is a monster model. For x, y from
the sort of K and z from the sort of Γ consider the formula
δ(x, y, z) = (v(x − y) ≥ z).
Then, for γ ∈ Γ the formula δ(x, y, γ) defines an equivalence relation on K which we
denote by Eγ (x, y). Put ∆ = {δ(x, y, z)}.
It is enough to show that eq-rk∆ (δ(x, b, γ)) ≥ n for all n ∈ ω, γ ∈ Γ and b ∈ K,
because then eq-rk∆ (δ(x, b, γ)) are infinite, so K is not rosy.
We argue by induction on n. The case n = 0 is trivial. Suppose eq-rk∆ (δ(x, b, γ)) ≥
n for all n ∈ ω, γ ∈ Γ and b ∈ K. Consider any γ ∈ Γ and b ∈ K. By saturation,
there is γ ′ ∈ Γ such that γ ′ > γ and Eγ ′ (x, y) refines [b]Eγ into infinitely many classes,
say with representatives bi , i ∈ ω. Then, taking c := γ ′ in the definition of eq-rk∆ ,
we get eq-rk∆ (δ(x, b, γ)) ≥ n + 1.
Using ⌣
| þ , we define Uþ -rank in the same way as U-rank is defined in stable
theories by means of ⌣
| , namely Uþ is a unique function from the collection of all
complete types to the ordinals together with ∞ with the property that for any ordinal
α, Uþ (p) ≥ α + 1 if and only if there is some tuple a and some type q ∈ S(Aa) such
that q ⊃ p, Uþ (q) ≥ α and q þ-forks over A. Uþ -rank in rosy theories has most of the
nice properties that U-rank has in stable theories, e.g. it satisfies Lascar Inequalities:
Uþ (a/b, A) + Uþ (b/A) ≤ Uþ (a, b/A) ≤ Uþ (a/b, A) ⊕ Uþ (b/A).
Assume T is rosy. If D is an A-definable set, then Uþ (D) := sup{Uþ (d/A) : d ∈ D}.
Of course, if this supremum is finite, then it is just the maximum. It turns out that if
D is a definable group, then the supremum is also attained [6, Remark 1.20]. For D a
definable set, Uþ (D) = 0 if and only if D is finite. There are also Lascar inequalities
for groups: For definable groups H ≤ G we have
Uþ (H) + Uþ (G/H) ≤ Uþ (G) ≤ Uþ (H) ⊕ Uþ (G/H).
We say that T is superrosy if Uþ (p) < ∞ for every type p; a group [or field] is
superrosy if its theory is superrosy.
9
Proof of Fact 1.9. Let H = Gn . We claim that Uþ (G) = Uþ (H). To see this, take
g ∈ G with Uþ (g/∅) = Uþ (G). Of course, g n ∈ acl(g), so Uþ (g n /g) = 0. Since G[n]
is finite, we get g ∈ acl(g n ), and so Uþ (g/g n ) = 0. Thus, by Lascar Inequalities, we
get Uþ (g n ) = Uþ (g), so Uþ (H) = Uþ (G). Using Lascar Inequalities for groups, we
immediately conclude that Uþ (G/H) = 0, so G/H is finite.
Definition 1.12 We say that a theory T has the NIP if there is no formula ϕ(x, y)
and sequence hai ii<ω such that for every w ⊆ ω there is bw such that |= ϕ(ai , bw ) iff
i ∈ w.
The main result of [12] says that an NIP field K has no Artin-Schreier extensions,
i.e., if char(K) = p > 0, then the function x 7→ xp − x from K to K is surjective.
The proof in general uses some algebraic geometry, but assuming that the image of
the function x 7→ xp − x is of finite index, this is an immediate consequence of the
existence of (K + )00 (i.e., of the smallest type-definable subgroup of K + of bounded
index (wlog we assume here that K is a monster model)). Indeed, one easily checks
that (K + )00 is a non-trivial ideal of K, so (K + )00 = K. Since the image of x 7→ xp −x
is a definable subgroup of K + of finite index, we get that it contains (K + )00 = K,
so it is equal to K.
2
Superrosy fields
This section is devoted to the proof of Theorem 6 and some observations concerning
Conjecture 5 in the zero characteristic case. In fact, we will not use superrosiness,
but only the assumption of Theorem 6, which is a consequence of superrosiness by
Fact 1.9. For a given field K this assumption is the following:
(A) For every finite extension L of K and for any natural number n > 0 the index
[L∗ : (L∗ )n ] is finite and, if char(K) = p > 0 and f : L → L is given by
f (x) = xp − x, then the index [L+ : f [L]] is also finite.
We start from some preparatory observations.
Remark 2.1 If K ⊆ L is an algebraic field extension and v is a non-trivial valuation
on L, then v↾K is also non-trivial.
Proof. Suppose for a contradiction that v ↾K is trivial. Consider any a ∈ L with
v(a) > 0. Let P (x) = xn + an−1 xn−1 + . . . + a0 ∈ K[x] be the minimal monic
polynomial of a over K. Then v(an ) > v(an−1 an−1 ) > . . . > v(a0 ) = 0, so ∞ =
v(0) = v(P (a)) = v(a0 ) = 0, a contradiction.
Remark 2.2 Let L be a finite extension of a field K. Suppose that any non-trivial
valuation on L has the property that its residue field is algebraically closed. Then the
residue field of any non-trivial valuation on K is either algebraically or real closed.
10
Proof. Consider any non-trivial valuation v on K. By Chevalley’s Extension Theorem, v has an extension to a (non-trivial) valuation w on L. By assumption, Lw is
algebraically closed. Since, by Fact 1.3, Lw is a finite extension of K v , the conclusion
follows.
Lemma 2.3 Let v be a valuation on a field K. Let F be a finite extension of the
residue field K v . Then there is a finite extension L of K and an extension w of v to
L such that the field F is isomorphic over K v to the residue field Lw .
Proof. It is enough to prove it for 1-generated extensions. So, let F = K v (α)
for some α ∈ F . Choose P (x) ∈ Ov [x] a monic polynomial such that P (x) is the
minimal monic polynomial of α over K v , where P (x) ∈ K v [x] denotes the polynomial
obtained from P (x) by reducing the coefficients modulo Mv . Consider a root β of
P (x), put L := K(β) and take any extension of v to a valuation w on L. Since the
coefficients of P (x) are in Ov with the leading coefficient equal to 1, one easily gets
that w(β) ≥ 0, so β ∈ Ow .
Now, for β := β + Mw ∈ Lw , we have that P (β) = 0. Since P is irreducible, we
get
[Lw : K v ] ≥ [K v (β) : K v ] = deg P = deg P ≥ [L : K],
so we have everywhere equalities by Fact 1.3. Therefore, Lw = K v (β) which is isomorphic to F over K v .
The next lemma recalls a standard method of showing that a given field is algebraically closed.
Lemma 2.4 Let K be a field. If for every finite extension L of K and for every
prime number n>0, (L∗ )n = L∗ and, in the case when char(K) = p > 0, the function
f : L → L given by f (x) = xp − x is onto, then K is algebraically closed.
Proof. By assumption, K is perfect. If it is not algebraically closed, then it has a
proper Galois extension F of minimal degree k > 1. There is an intermediate field
L with Gal(F/L) ∼
= Zq for some prime number q.
If q = p = char(K), then Galois theory tells us F is a splitting field over L of a
polynomial of the form xp −x−a for some a ∈ L, but, by assumption, this polynomial
has at least one zero in L, so all its zeros are in L, a contradiction.
Assume q 6= char(K). Since q ≤ k and for a primitive q-th root of unity ζq the
field K(ζq ) is a Galois extension of K of degree less than q, by the choice of k, we
conclude that ζq ∈ K. So ζq ∈ L. By Galois theory, we conclude that F is a splitting
field over L of a polynomial of the form xq − a for some a ∈ L. By assumption, this
polynomial has a zero in L, so all its zeros are in L, a contradiction.
Remark 2.5 Let K be an infinite field of characteristic p > 0. Then either K p = K,
or the index [K ∗ : (K ∗ )p ] of multiplicative groups is infinite and the index [K + :
(K + )p ] of additive groups is also infinite. In particular, if the index [K ∗ : (K ∗ )p ] is
finite, then K is perfect.
11
Proof. Suppose K p 6= K, and let b1 , b2 ∈ K be linearly independent over K p . Then
b1 + k p b2 for k ∈ K are in different cosets modulo (K ∗ )p . So the index [K ∗ : (K ∗ )p ]
is infinite. Similarly, the elements b1 k p for k ∈ K are in different additive cosets
modulo (K + )p , so the index [K + : (K + )p ] is also infinite
Proof of Theorem 6. Suppose there is no non-trivial definable valuation on K.
Part 1. The value group of any non-trivial valuation v on K is divisible.
Proof. Suppose for a contradiction that some γ ∈ Γv is not divisible by n. Put
S := {x(K ∗ )n ∈ K ∗ /(K ∗ )n : Ov∗ ∩ x(K ∗ )n 6= ∅}
and
T :=
D[ E
S ,
that is, T is the subgroup of K ∗ generated by the union of the family S.
Claim 1 T is a proper, definable subgroup of K ∗ such that Ov∗ ≤ T and (K ∗ )n ≤ T .
Proof of Claim 1. Since (K ∗ )n is a finite index subgroup of K ∗ , we get that T is a
union of finitely many cosets of (K ∗ )n , so T is definable and contains (K ∗ )n . The
fact that Ov∗ ≤ T follows directly from the definition of T . To see that T is proper,
suppose for a contradiction that T = K ∗ . Then there is a ∈ T such that v(a) = γ.
By the definition of T , we can write a = aε11 · . . . · aεmm , where ai ∈ Ov∗ · (K ∗ )n and
εi ∈ {−1, 1} for i = 1, . . . , m. Then γ = v(a) = ε1 v(a1 ) + . . . + εm v(am ) ∈ nΓv , a
contradiction.
By Claim 1, we know that T is proper and 1 + Mv ⊆ Ov∗ ≤ T , so v is fully compatible with T . By Fact 1.6, we conclude that OT is non-trivial. We also get that
we are in the group case, so Fact 1.7 tells us that OT is definable in (K, +, ·, 0, 1, T ),
but T is definable in K, so OT is definable in K. This is a contradiction, which
completes the proof of Part 1.
Part 2. Assuming that char(K) = p > 0, the residue field of any non-trivial valuation v on K is algebraically closed.
Proof. By Lemmas 2.3 and 2.4, we will be done if we prove the following two claims.
Claim 2 For every finite extension L of K, for every extension w of v to a valuation
∗
∗
on L, and for every prime number n, one has (Lw )n = Lw .
Claim 3 For every finite extension L of K, one has that the function f : Lw → Lw
given by f (x) = xp − x is onto.
Proof of Claim 2. First, notice that by Remark 2.1, we can assume that L = K.
Indeed, since L is definable in K (living in some Cartesian power of K), by Remark
2.1, we get that there is no non-trivial definable valuation on L; it is also clear that
12
L satisfies (A).
Consider the case n 6= pp.
Subclaim K n (1 + Mv ) = K
Proof of the subclaim. Suppose it is not true. Then T := (K ∗ )n (1 + Mv ) is a proper
subgroup of K ∗ , and v is fully compatible with T . By Fact 1.6, OT is non-trivial.
Since [K ∗ : (K ∗ )n ] is finite, we see that T is definable, so there is no non-trivial
valuation on K definable in (K, +, ·, 0, 1, T ).
Let K T be the residue field corresponding to the valuation ring OT . Since
∗ n
(K ) ≤ T and (n, char(K T )) = 1, Fact 1.5 gives us that we are either in the
group case or in the residue case. By Fact 1.7, in the group case, OT is definable in
(K, +, ·, 0, 1, T ) which is impossible, and, in the residue case, either OT is definable
in (K, +, ·, 0, 1, T ) which is impossible or T is the positive cone of an ordering on
K T . However, the last thing is also impossible as char(K T ) = p > 0.
n
By the subclaim, K v = K v . Indeed, for any a ∈ Ov , a = k n (1 + m) for some
n
n
k ∈ K and m ∈ Mv . Since 1 + m ∈ Ov∗ , we get that k ∈ Ov , so a = k ∈ K v .
p
Now, consider the case n = pp. Suppose for a contradiction that K v is a proper subm
field of K v . As we have already proved that K v = K v whenever (m, char(K v )) = 1,
∗
∗
we see that K v is infinite. Thus, Remark 2.5 implies that [K v : (K v )p ] is infinite.
p
p
Moreover, if aK v 6= bK v for some a, b ∈ Ov , then aK p 6= bK p , because otherwise
p
p
either ak p = b or a = bk p for some k ∈ Ov , and so either ak = b or a = bk , a
contradiction. This implies that [K ∗ : (K ∗ )p ] is infinite, which contradicts (A).
Proof of Claim 3. As in the proof of Claim 2, by Remark 2.1, we can assume that
K = L. Let F : K → K be given by F (x) = xp − x.
Subclaim F [K] + Mv = K.
Proof of the subclaim. Suppose it is not true. Then T := F [K] + Mv is a proper
subgroup of K + , and v is fully compatible with T . By Fact 1.6, OT is non-trivial.
As [K + : F [K]] is finite, T is definable in K, so there is no non-trivial valuation on
K definable in (K, +, ·, 0, 1, T ).
Since F [K] ≤ T , Fact 1.5 ensures that we are either in the group case or in the
residue case. In the residue case, OT is definable by Fact 1.7, a contradiction. In the
group case, since we know that OT is not definable, Fact 1.7 yields some x ∈ MT
with x−1 OT ⊆ T .
Now, we will adopt the argument from the proof of [13, Theorem 3.1]. For the
reader’s convenience, we give all the details.
Let AT be the largest fractional OT -ideal contained in T . We aim at defining
a certain fractional OT -ideal AαT which will properly contain AT and which will be
contained in T . This will contradict the choice of AT .
Denote by vT and by ΓT the valuation and the value group corresponding to OT .
For a real number α and γ1 , γ2 ∈ Γ, by γ1 ≥ αγ2 we mean that γ1 ≥ rγ2 for all
13
rationals r ≤ α if γ2 ≥ 0, or for all rationals r ≥ α if γ2 < 0. This definition makes
sense since ΓT is divisible by Part 1 of the proof. By γ1 < αγ2 we mean the negation
of γ1 ≥ αγ2 . For α > 1 define
AαT := {x ∈ K : vT (x) ≥ αvT (y) for some y ∈ AT }.
This is, of course, a fractional OT -ideal. Since x−1 OT ⊆ T and x ∈ MT , we have
that AT properly contains OT . This easily implies that AT ⊆ AαT . Now, we check
that this inclusion is proper.
Take a natural number n such that α > 1 + n1 . There is y ∈ AT \ OT such
that (1 + n1 )vT (y) ∈
/ vT (AT ), as otherwise one easily gets that for all y ∈ AT \ OT ,
yOT generates a valuation ring properly containing OT and contained in T , which
contradicts the fact that vT is coarsely compatible with T . Choose z ∈ K with
vT (z) = n1 vT (y). Then vT (zy) = (1 + n1 )vT (y) ≥ αvT (y), so zy ∈ AαT \ AT .
The proof of the subclaim will be completed if we show that AαT ⊆ T for α ∈
(1, 2 − 1p ). It is enough to prove that every element t ∈ AαT \ AT belongs to T .
We have vT (t) ≥ αvT (y) for some y ∈ AT \ OT . Since t ∈
/ AT , we immediately get
vT (ty −1 ) = vT (t)−vT (y) < 0. Therefore, ty −1 ∈ K \OT . Thus, by Claim 2, ty −1 = ap
for some a ∈ K \ OT , and so ta−p = y ∈ AT \ OT ⊆ T \ OT , hence ta−p = bp − b + m
for some b ∈ K \ OT and m ∈ Mv ⊆ MT . As α < 2 − p1 , we obtain
vT (ap b) = vT (ap ) + vT (b) = vT (ty −1) + vT (b) = vT (t) − vT (y) + 1p vT (y)
≥ α − 1 + p1 vT (y) > vT (y).
Therefore, vT (ab) > vT (ap b) > vT (y) ≥ αvT (y) and vT (ap m) > vT (ap b) > vT (y) ≥
αvT (y). Hence, ab, ap b, ap m ∈ AT ⊆ T , and so t = ((ab)p −ab)+ab−ap b+ap m ∈ T .
By the subclaim, f [Lv ] = Lv . Indeed, for any a ∈ Ov , a = k p − k + m for some
p
k ∈ K and m ∈ Mv . Then k ∈ Ov , so a = k − k.
So, the proof of Part 2 and of the whole theorem has been completed.
As was pointed out in the introduction, Corollary 7 follows from Theorem 6 and
Facts 1.8 and 1.9, and Corollary 8 follows from Theorem 6 and the fact that NIP
fields are closed under Artin-Schreier extensions.
Proposition 2.6 Let K be a field of characteristic zero satisfying (A) and containing
√
−1. Then either there is a non-trivial definable valuation on K, or for every nontrivial valuation v on K:
n
1. if char(K v ) > 0, then K v = K v for every n > 0,
2. if char(K v ) = 0, then there is a prime number p such that for all primes n
n
different from p, K v = K v .
14
Proof. Suppose there is no non-trivial definable valuation on K.
(1) The same argument (based on Facts 1.5, 1.6 and 1.7) as in √
the proof of Claim 2
in the proof of Theorem 6 works, noting that the assumption −1 ∈ K eliminates
the possibility that T is the positive cone of an ordering on K T .
p
(2) Suppose that for some prime p, K v 6= K v . Then T := (K ∗ )p (1 + Mv ) is a proper,
definable subgroup of K ∗ . If p 6= char(K T ), then we get a contradiction as in the
proof of the subclaim in the proof of Claim 2 in Theorem 6. So p = char(K T ).
Consider any prime n 6= p. Let T ′ := (K ∗ )np (1 + Mv ). Then T ′ ≤ T . On the
other hand, since T is definable and OT is not definable, Fact 1.7 implies that T
does not belong to the group case. Thus, we conclude that OT ⊆ OT ′ , and hence
char(K T ′ ) ∈ {0, p}. However, it is impossible to have char(K T ′ ) = 0, as in this case
once again one gets a contradiction as in the proof of the subclaim in the proof of
Claim 2 in Theorem 6. So char(K T ′ ) = p. Our goal is to show that K n (1+Mv ) = K.
If this is not the case, then T ′′ := (K ∗ )n (1 + Mv ) is a proper, definable subgroup
of K ∗ , and once again Facts 1.5, 1.6 and 1.7 yield that (n, char(KT ′′ )) 6= 1 and T ′′
does not belong to the group case. Since T ′ ≤ T ′′ , we conclude that OT ′′ ⊆ OT ′ , but
char(K T ′ ) = p, and so we get that char(K T ′′ ) = p. This contradicts the assumption
that n is relatively prime to p.
If one was able to strengthen the conclusion of Proposition 2.6(2) by showing
n
that K v = K v for all primes n, then using Lemma 2.4, one would get that, under
the assumption of Proposition 2.6, either there is a non-trivial definable valuation
on K, or for any non-trivial valuation v with char(K v ) = 0 the residue field K v is
algebraically closed. So, arguing as in the proof of Remark 2.2, one would also get
that Assumption (A) implies that either there is a non-trivial definable valuation
on K, or for any non-trivial valuation v with char(K v ) = 0 the residue field K v is
either algebraically or real closed. In fact, strengthening slightly Assumption (A),
this would imply the full conclusion of Conjecture 5.
Proposition 2.7 Suppose that the conclusion of Proposition 2.6(2) can be strengthn
ened to ‘ K v = K v for all primes n’. Let K be a field such that for every finite
extension L of any elementary extension of K and for every natural number n > 0
the index [L∗ : (L∗ )n ] is finite and, if char(K) = p > 0 and f : L → L is given by
f (x) = xp − x, the index [L+ : f [L]] is also finite. Then either there is a non-trivial
definable valuation on K, or every non-trivial valuation on K has divisible value
group and either algebraically or real closed residue field.
Proof. Suppose there is no non-trivial
definable valuation on K. Using Remarks 2.1
√
and 2.2, we can assume that −1 ∈ K. Let v be any non-trivial valuation on K.
Divisibility of Γv was proved in Theorem 6. By this theorem, it remains to consider
the case char(K) = 0. From the discussion above Proposition 2.7, we are done in
the case char(K v ) = 0. So it remains to consider the case char(K v ) = p > 0.
e Γ,
e e
Take a monster model (K,
v) ≻ (K, Γv , v). Let O be a maximal valuation ring
e
/ O. Let M be the maximal ideal of O. Then
in K containing Ove and such that p1 ∈
15
p ∈ M, so k := O/M is of characteristic p.
e such that O ( O1 .
We claim that there is a non-trivial valuation ring O1 in K
e O of Γ
e defined
To see this, recall that O corresponds to the proper convex subgroup Γ
by
eO := {γ ∈ Γ
e : (∀x ∈ M)(e
Γ
v(x) > γ, −γ)}.
eO . Put Γ
e1 := S (−nγ, nγ) a convex subgroup of Γ
e properly
So, there is γ > Γ
n
eO . By saturation, Γ
e1 6= Γ.
e So, Γ
e1 = Γ
eO1 for some non-trivial valuation
containing Γ
ring O1 ) O.
Since O ( O1 , we have that 1p ∈ O1 , hence the residue field k1 := O1 /M1 (where
M1 is the maximal ideal of O1 ) is of characteristic 0. So k1 is algebraically closed
(as the conclusion of Proposition 2.7 holds in the residue zero characteristic case,
e in place of K). As Ove ⊆ O1 , this implies that the residue field K
e ev is
also for K
algebraically closed, and so K v is algebraically closed, too.
3
Minimal fields
We will prove here Theorem 9. In contrast to Theorem 6, where the proof relies on
non-trivial results from [13], the proof of Theorem 9 is a trivial consequence of the
definition of minimality.
Recall that a minimal field is an infinite field whose every definable (in one variable) subset is finite or co-finite.
Proof of Theorem 9. Consider any non-trivial valuation v on K.
By minimality, for every natural number n > 0, (K ∗ )n = K ∗ . (To see this, notice
that (K ∗ )n is an infinite multiplicative subgroup of K, so it is a co-finite subgroup,
so it is everything.) From this, it is clear that Γv is divisible.
Consider any monic polynomial P (x) = xn + an−1 xn−1 + . . . + a0 ∈ K v of positive
degree n; here a0 , . . . , an−1 ∈ Ov . Let P (x) = xn + an−1 xn−1 + . . . + a0 . Since P (x)
takes co-finitely many values in K and Mv is infinite, there exists a ∈ K such that
P (a) ∈ Mv . Then a ∈ Ov , as otherwise v(a) < 0, and so v(ai ai ) = v(ai ) + iv(a) >
nv(a) = v(an ) for i = 0, . . . , n − 1, which implies that v(P (a)) = v(an ) = nv(a) < 0,
a contradiction with the fact that P (a) ∈ Mv . Therefore, P (a) = 0, i.e., P (x) has a
root in K v .
4
Final comments
We will say that a given field K has Property (∗) if every non-trivial valuation on
K has divisible value group and algebraically closed residue field. We will say that
it has Property (∗−) if every non-trivial valuation on K has divisible value group
and either algebraically or real √
closed residue field. By Remark 2.2, a field K has
Property (∗−) if and only if K( −1) has Property (∗).
16
Having the results and conjectures discussed in this paper, a natural questions
arises what can be said about the structure of fields with Property (∗) or (∗−). In
particular, can one deduce from our results Conjectures 1, 2, 3, 4 (at least in positive
characteristic) or Podewski’s conjecture?
Recall that a filed K is bounded if its absolute Galois group is small, i.e., its
absolute Galois group has only finitely many closed subgroups of any finite index;
equivalently, for every natural number n > 0, K has only finitely many extensions
of degree n (up to isomorphism over K). A field is orderable if it can be equipped
with some order making it an ordered field; equivalently, it is formally real (i.e., −1
is not a sum of squares).
Recall that a field K is PAC (pseudo algebraically closed) if every absolutely
irreducible variety over K has a K-rational point. A field K is PRC (pseudo real
closed) if every absolutely irreducible variety over K which has an F -rational point in
every real closed field F containing K has a K-rational point. With such definitions
√
each PAC field is PRC. We know that if K is an orderable PRC field, then K( −1)
is (perfect) PAC [18, Lemma A.1.1.3].
Fact 4.1 A PRC field is superrosy if and only if it is perfect and bounded.
Proof. (←) was proved in [18, Appendix A] for orderable PRC fields and in [11] for
PAC fields. To see the converse, note that the same argument as in [21, Theorem
5.6.5] yields boundedness. Finally, suppose for a contradiction that K is not perfect.
Then, by Remark 2.5, [K ∗ : (K ∗ )p ] is infinite, which contradicts Fact 1.9.
Remark 4.2 An orderable PRC field has the strict order property, so it is not simple.
In particular, Conjecture 3 implies Conjecture 1.
Proof. Let K be a formally real PRC field. We claim that any element which is a sum
finitely many squares is a sum of two squares. For this, consider any non-zero element
a ∈ K which is a sum of finitely many squares. Then the polynomial x2 + y 2 − a has
a zero in any real closed field containing K. Since x2 + y 2 − a is absolutely irreducible
(e.g. by the Eisenstein criterion) and K is PRC, this polynomial has a zero in K,
and so a is a sum of two squares in K.
Now, define the relation ≤ on K by
x ≤ y ⇐⇒ y − x is a sum of squares of finitely many elements of K.
We claim that ≤ is a partial order with an infinite chain. The fact that ≤ is
antisymmetric follows from the fact that K is formally real. Transitivity is clear.
The existence of an infinite chain can be seen as follows. Take any a 6= 0. Then
a2 < 2a2 < 3a2 < . . ..
Since, by the first paragraph of the proof, x ≤ y if and only if y − x is a sum of
two squares, we see that ≤ is definable.
Fact 4.3 A non separably and non real closed PRC field does not have NIP. In
particular, Conjecture 3 implies Conjecture 2.
17
Proof. The main result of [5] says a non separably closed PAC field does not have
NIP. Now, consider a non separably and
√ non real closed PRC field. Suppose for a
contradiction that it has NIP. Then K( −1) still has NIP (as it is interpretable in
K) and it is PAC but not separably closed, a contradiction.
By [8], perfect PAC fields have Property (∗), so, by [18, Lemma A.1.1.3], PRC
fields have Property (∗−). One could ask whether for infinite fields Property (∗−)
[or (∗)] implies that the field in question is PRC (in the non formally real case, PRC
can be replaced by PAC). By virtue of our results and Fact 4.3, the positive answer
would imply Conjectures 1, 2, 3 and 4 in positive characteristic. Unfortunately the
answer is negative, which we briefly explain now. In [8] and [10], the Hasse principle
for Brauer groups is considered. It is shown in [8, Theorem 3.4] that if K is a perfect
PAC field, then any extension F of K of relative transcendence degree 1 satisfies the
Hasse principle for the Brauer groups. On the other hand, it is shown in [8, Theorem
4.1] that whenever K is a perfect field such that the Hasse principle for the Brauer
groups holds for all extensions F of K of relative transcendence degree 1, then the
field K has Property (∗). It was also asked in [8, Question 4.2] whether a non formally
real infinite perfect field K such that the Hasse principle for the Brauer groups holds
for all extensions F of K of relative transcendence degree 1 is necessarily PAC? In
[10], a counter-example was constructed. The authors introduced the class of the
so-called weakly PAC fields, they proved that weakly PAC fields are non formally
real and that (assuming perfectness) they satisfy the Hasse principle for the Brauer
groups for all extensions of relative transcendence degree 1, and they constructed
perfect weakly PAC fields which are not PAC. In particular, perfect weakly PAC
fields have Property (∗).
The above discussion leads to the following questions in our context.
Question 4.4 Does there exist a weakly PAC but not PAC superrosy [resp. supersimple] field?
The positive answer would refute Conjecture 3 [resp. Conjecture 1]. The negative
answer would support (but not prove) these conjectures.
Question 4.5 Let K be an infinite perfect field with NIP satisfying Property (∗) [or
(∗−)]. Is it true that K is either algebraically or real closed?
√
Applying the trick with adding −1, we see that both versions of this question
are equivalent. By Corollaries 7 and 8, the positive answer would imply Conjectures
2 and 4 in positive characteristic.
Question 4.6 Does there exist a non algebraically closed, perfect weakly PAC field
which [is superrosy and] has NIP?
The positive answer would yield the negative answer to Question 4.5. The positive
answer to the extended version would refute Conjecture 2. Notice that since by [5]
18
we know that non separably closed PAC fields do not have NIP, if the answer to the
above question was positive, the witness field would have to be weakly PAC but not
PAC.
In any case, in order to find counter-examples to Conjectures 1, 2, 3 or 4, one could
try to construct suitable non PRC fields with Property (∗−); possible candidates
could be among perfect weakly PAC but not PAC fields. Or, try to prove the
conjectures by showing that there are no such fields.
A generalization of Corollary 7 to the characteristic zero case is an open problem.
Of course, Thereom 6 yields the divisibility of the value groups, but the problem is
with the residue fields.
Conjecture 10 Every non-trivial valuation on a superrosy field has divisible value
group and either algebraically or real closed residue field.
Another idea of attacking Conjecture 2 by means of Corollary 7 (or rather Conjecture 10), suggested by E. Hrushovski, is to try to show that a superrosy field with
NIP is the residue field with respect to some non-trivial valuation on another superrosy field with NIP, and use Corollary 7 (or Conjecture 10 in the zero characteristic
case) to conclude that the original field is either algebraically or real closed.
As was mentioned in the introduction, K. Dupont has
√ a different approach to
Conjecture 4 (and so also to Conjecture 2). After adding −1 to the field, the goal
is to show that either for every finite extension L of K and for every prime number n,
(L∗ )n = L∗ (as then K is algebraically closed by Lemma 2.4), or there is a non-trivial
definable valuation on K. Assume that the first possibility fails for some L and n.
Put T = (L∗ )n . It follows from the assumptions and Remark 2.5 that L is perfect,
so n 6= char(L). The first goal is to show that OT is a definable valuation ring on L.
Further, K. Dupont deduced from [13] that OT is non-trivial if and only if the family
of sets {aT + a : a ∈ L∗ } is a subbasis of a V -topology on L. The second goal of her
project is to show (using the NIP assumption) that the last condition holds.
References
[1] H. Adler. A geometric introduction to forking and thorn-forking, preprint.
[2] Z. Chatzidakis, Simplicity and independence for pseudo-algebraically closed
fields. In Models and computability (Leeds, 1997), London Math. Soc. Lecture
Note Ser., vol. 259, 41-61, Cambridge Univ. Press, Cambridge, 1999.
[3] Z. Chatzidakis and A. Pillay, Generic structures and simple theories, Ann. Pure
Appl. Logic 95, 71-92, 1998.
[4] G. Cherlin, S, Shelah, Superstable fields and groups, Annals of Mathematical
Logic 18, 227-270, 1980.
19
[5] J. Duret, Les corps faiblement algébriquement clos ont la propriété
d’indépendance. In Model Theory of Algebra and Arithmetic (proc. Karpacz),
Lecture Notes in Mathematics, Springer, vol. 834, 136-162, 1979.
[6] C. Ealy, K. Krupiński, A. Pillay. Superrosy dependent groups having finitely
satisfiable generics, Annals of Pure and Applied Logic 151, 1-21, 2008.
[7] C. Ealy and A. Onshuus. Characterizing rosy theories, Journal of Symbolic Logic
72, 919-940, 2007.
[8] I. Efrat, A Hasse Principle for function fields over PAC fields, Israel Journal of
Mathematics 122, 43-60, 2001.
[9] A. J. Engler, A. Prestel, Valued Fields, Springer, The Netherlands, 2005.
[10] W. Geyer, M. Jarden, Non PAC fields whose Henselian closures are separably
closed, Mathematical Research Letters 8, 509-520, 2001.
[11] E. Hrushovski, Pseudo-finite fields and related structures, preprint, 1991.
[12] I. Kaplan, T. Scanlon, F. Wagner, Artin-Schreier extensions in NIP and simple
fields, Israel Journal of Mathematics 185, 141-153, 2011.
[13] J. Koenigsmann, Definable Valuations, In Delon, Dickmann, and Gondard, editors, Seminaire Structures algébriques ordonnées Paris VII, 1994.
[14] K. Krupiński, Fields interpretable in rosy theories, Israel Journal of Mathematics
175, 421-444, 2010.
[15] K. Krupiński, Fields interpretable in superrosy groups with NIP (the non-solvable
case), Journal of Symbolic Logic 75, 372-386, 2010.
[16] A. Macintyre, On ω1 -categorical theories of fields, Fundamenta Mathematicae
71, 1-25, 1971.
[17] A. Onshuus, Thorn-Forking in Rosy Theories, Ph.D. thesis, Berkeley, 2002.
[18] A. Onshuus. Properties and consequences of thorn-independence, Journal of
Symbolic Logic 71, 1-21, 2006.
[19] K. P. Podewski, Minimale Ringe, Math. Phys. Semesterber. 22, 193-197, 1973.
[20] S. Shelah, Strongly dependent theories, preprint, SH863.
[21] F. Wagner, Simple Theories, Kluwer Academic Publishers, Dordrecht, The
Netherlands, 2000.
Address:
Instytut Matematyczny, Uniwersytet Wrocławski,
pl. Grunwaldzki 2/4, 50-384 Wrocław, Poland.
E-mail address: [email protected]
20
| 0math.AC
|
Real Time Impact Control on Charging Energy Storage
For Shipboard Power Systems
arXiv:1801.05797v1 [cs.SY] 17 Jan 2018
Yusheng Luoa , Sanjeev Srivastavab , Manish Mohanpurkara , Svetomir
Stevica , Rob Hovsapiana
a
Idaho National Laboratory
Idaho Falls, Idaho 83415
b
Siemens Corporate Technology
755 College Rd E, Princeton, NJ 08540
Abstract
Medium voltage direct-current based integrated power system is projected as
one of the solutions for powering the all-electric ship. It faces significant challenges for accurately energizing advanced loads, especially the pulsed power
load, which can be rail gun, high power radar, and other state of art equipment. Energy storage based on supercapacitors is proposed as a technique
for buffering the direct impact of pulsed power load on the power systems.
However, the high magnitude of charging current of the energy storage can
pose as a disturbance to both distribution and generation systems. This paper presents a fast switching device based real time control system that can
achieve a desired balance between maintaining the required power quality
and fast charging the energy storage in required time. Test results are shown
to verify the performance of the proposed control algorithm.
Keywords: Medium voltage direct-current based integrated power system,
Pulsed power load, Power quality, Disturbance metric, Real time control
1. Introduction
Research related to navy shipboard power system raise a critical concern
regarding to the system stability due to diverse loads. Similar to microgrids,
navy shipboard power systems do not have a slack bus [1]. It can be viewed as
a microgrid always operating in islanding mode. Compared with typical terrestrial microgrid, the ratio between the overall load and generation is much
higher [2]. Although new avenues such as zonal load architecture [3], high
Preprint submitted to International Journal of Electrical Power and Energy SystemsJanuary 18, 2018
energy/power density energy storage systems [4], and high efficiency generator [5] have been developed for supporting stability control, new challenges
also arise due to the diverse demand from load side.
In order to reduce ship’s size and visibility in future battlefields, an Integrated Power System (IPS) based all-electric ship is developed [6]. All electric
ship is characterized by using electrical motor driven propulsion system to
replace traditional mechanical transmission propulsion system. Therefore,
the propulsion system is integrated with the shipboard power system, and
is termed as IPS. The application of electrical propulsion waives the need
of slow speed gear providing power to mechanical propulsion. The remaining fast gear can drive both the generation and electrical propulsion [7].
Therefore the space occupied by the transmission system is largely reduced.
Nevertheless, the electrical propulsion has high efficiency, which means less
fuel is needed and the ship size can be further reduced. Not only the shape
of the ship is optimized, but the weapons system can also be improved significantly. Phase-array radars are capable of providing detection with higher
resolution and wider area, pulsed power load features in fast, powerful, and
accurate attacks, electromagnetic ejector can inject high launch dynamics
for shipboard aircrafts and so on. These state of the art devices share a
common characteristic, their power demand appears as a large pulse due to
the integration of power electronics interface [8]. The electrical propulsion
and Pulsed Power Load (PPL) place a huge burden on the generation and
distribution capacity of the shipboard power systems.
To address the challenge of reliably powering the all-electric ship, several
efforts have been made. Different distribution architectures such as Medium
Voltage Alternative Current (MVAC), High Frequency Alternative Current
(HFAC), and Medium Voltage Direct Current (MVDC), have been analyzed
[5, 9]. Such research efforts reached a conclusion that MVDC system is projected as the optimal distribution architecture for IPS. This is on account
of MVDC’s no concerns regarding generator synchronization, harmonics distortion, and frequency stability. The electrical devices and related control
strategy for supporting MVDC distribution architecture have also been developed [10].
Besides developing IPS, contributions have also been made to test the performance of IPS in powering advanced loads. Propulsion loads occupy a large
portion of the total shipboard load consumption. In the rest of the loads,
PPLs are usually critical due to their indispensable role in the battle. One of
the most challenging issues in controlling IPS operation is providing expected
2
power and energy to PPL without affecting the operation of other critical
loads. The impact from PPL to IPS with various distribution architectures
are analyzed in [11, 12, 13]. Analysis shows that the impact from energizing
PPL with around 100 MW power can seriously influence the system stability
and power quality. In [14], noteworthy work is presented for an innovative
application of power electronics interfaced Supercapacitor Energy Storage
System (SESS). Energy storage has found wide application in such hybrid
energy systems, for augmenting limited generation and modern loads [15, 16].
From the beginning of the development of IPS, energy storage was used as
auxiliary power supply [17]. Supercapacitor is well known for its high power
density, which can enable fast charging/discharging [18, 19]. By switching
the power source from shipboard generation to SESS, the direct impact from
PPLs to IPS can be eliminated. However, the charging of SESS can induce
high current and still destabilize the IPS. Control methods for driving the
SESS charging circuit within desired limits and rates are proposed. In [11],
limit-based supercapacitor control is proposed. Although this method can
be effective, it requires a precise understanding of the system’s current and
power limits in order to prevent abrupt disturbance to the system. In [20],
a trapezoidal-based control of SESS charging is introduced. This control is
only effective when trapezoidal load profile fits the system requirement, and
vice versa. Profile-Based Control (PBC) in [14] features a minimal impact
to system stability within a fixed time frame, however it still requires the
prerequisite information about the system’s current limit. Furthermore, it is
challenging to set a fixed charging time which is essential for such systems.
Some other papers, such as [21, 22] focus on coordinating SESS charging
control and generation control to mitigate the possible disturbance caused
by SESS charging. However, during mission time, especially battle time, the
quality of communication supporting coordination may not be guaranteed
due to weather or battle damage. Nevertheless, these coordination strategies
are mostly applied to MVAC system, which highlight coordination of generation control and other load control. The MVDC system is proposed to
waiver the concerns of generation control coordination. In another word, it is
more open decentralized control strategy. Henceforth a decentralized control
strategy is considered as more suitable method for SESS charging in MVDC
system.
The proposed work in this paper is based on the contribution in [23],
which proposed a Disturbance Metric Control (DMC). DMC is a decentralized technique developed to control SESS charging according to disturbances
3
monitored in real time from SESS charging in MVDC system. However, it is
needed to prove that existing hardware is capable of realizing the proposed
real time control theory. To achieve the expected performance of DMC, this
study integrates the cutting edge IGBT technology to DMC and proposed
Real Time DMC (RTDMC). The paper comprises of the following sections:
section 2 describes the experimental system; section 3 analyzes the disturbance due to ES charging; section 4 presents DMC strategy guided by two
major concerned disturbance metrics in MVDC system; section 5 describes
the hardware implementation of proposed RTDMC strategy; in section 6 test
results based on the proposed RTDMC are discussed; and the conclusion is
summarized in section 7.
2. System description
2.1. Test Environment
Unpredicted and unaccounted charging/discharging of SESS can damage
hardware leading to failures, making it uneconomical to test in real-world
applications. Hence, a digital real time simulator (DRTS) is used to assess
the performances and generate result. In this environment, the time-step
for simulating electrical circuit is 50 µs. For power electronics simulation,
the time-step can be reduced to 1-2 µs. The real time computation speed
of the power systems can test whether the proposed control strategy is fast
enough to adjust SESS charging so that expected control performane can be
achieved.
2.2. MVDC System
With the maturity of technology in DC circuit breakers [24], DC cables [25], MVDC system is deemed as advantageous for power distribution.
Following the introduction of MVDC in [9], we simulated and studied a shipboard IPS with MVDC distribution architecture as shown in Figure 1. In
this topology, the main power source are two Main Gas Turbine Generators
(MTGs). Each MTG has a generation capacity of 36 MW. In case of need,
each MTG is equipped with a 5 MW Auxiliary gas Turbine Generator (ATG).
Generation output from MTGs and ATGs is fed into a controllable rectifier
and then into a 5 kV MVDC distribution bus. Two 5 kV distribution buses
constitute the MVDC distribution system. Since a ship is equally divided
into two parts i.e., starboard and port, the bus that feeds starboard is called
starboard bus whereas the other one is called port bus. Each bus is directly
4
connected with an MTG, an ATG, and a 36.5 MW propulsion motor. Between two buses, there are zonal loads, stern circuit breaker, and bow circuit
breaker. Zonal loads can use circuit breaker to choose drawing power from
one bus or both buses. Stern and bow circuit breakers are used to connect
or disconnect starboard and port buses. When starboard and port bus are
connected with each other, generation from all four generators are merged
together, the mode for this connection is called Ring Mode (RM). When starboard and port bus are separated, power supplied to each bus can only come
from one MTG and one ATG. We call this operation as Split Plant Mode
(SPM). The PPL module is directly connected to port bus, which means the
power supply for charging SESS in RM is much more than in SPM.
MTG1
GT
GT
AC
Circuit
Breaker
Rectifier
ATG1
AC
Circuit
Breaker
Port
Propulsion
Motor
DC
Disconnect
Rectifier
DC
Disconnect
MVDC
Starboard Bus
Drive Inverter
Zone 1
Load
Center
Stern
Circuit
Breaker
Zone 2
Load
Center
Zone 3
Load
Center
Zone 4
Load
Center
DC
Disconnect
PPL
ES
BUCK
Converter
Zone 5
Load
Center
DC
Disconnect
Starboard
Propulsion
Motor
DC
Disconnect
Bow
Circuit
Breaker
MVDC
Port Bus
Drive Inverter
ATG2
Rectifier
GT
AC
Circuit
Breaker
GT
MTG2
AC
Circuit
Breaker
Rectifier
Figure 1: System topology of MVDC based IPS.
2.3. Pulsed Power Load Module
Inside the PPL module, there is PPL, SESS, and the power electronics
interface serving as charging circuit of SESS, as shown in Figure 2. The
charging circuit is a BUCK converter which is controlled by switch S1 [26].
During each commutation period T , the turn on time for S1 is Ton . During
Ton , power goes through S1 and L, then enters SESS. When S1 is opened,
current goes through SESS can be continued by passing D and L. When the
5
BUCK converter
MVDC bus P S1
L
+
Vin
S2
+
D
Vo
-
SESS
PPL
-
MVDC bus N
Figure 2: Charging circuit for SESS.
charging reaches a steady state, the voltage on SESS, Vo can be maintained
at a fixed value, as shown in Eq (1):
Ton
Vin .
(1)
T
During energizing PPL, power is sent from SESS, then S2, finally reaches
PPL. To prevent PPL directly draw power from MVDC bus, S1 and S2
are not allowed to be closed simultaneously. Therefore S2 is always opened
during charging, and S1 remained opened when PPL is being energized. In
this study, totally 300 MJ energy is expected to be injected into SESS. This
amount of energy can allows PPL at least be fired up for 2 times without
recharging according to [12].
Vo =
3. Analysis of Disturbance due to SESS charging
Figure 3 is a simplified diagram of the MVDC system using the single
line representation, which is used to analyze the possible disturbance caused
by SESS charging. In this figure EA is the emf generated from a single
phase and Xg is generator’s reactance. The resistance in generator can be
ignored during the transient analysis [27]. D1 -D4 are the switches forming
the rectifier to convert power from AC to DC. RLine is the line resistance in
distribution system and RLoad is the resistance of loads other than SESS. We
set the analysis time begins at t0 , when EA is in the positive semi-period,
6
and only D1 and D4 are turned on. It is generator’s terminal current. Before
charging, It can be expressed as Eq.(2):
It,t−0 =
EA
Xg + RLine + RLoad
.
(2)
The terminal voltage which is also the MVDC bus voltage is calculated in
Eq.(3)
(3)
Vbus,t−0 = EA − Xg ILoad .
RLine
It
EA
D1
D2
Xg +
Vbus
-
RLoad
D3
SESS
D4
Figure 3: Simplification diagram of MVDC based IPS.
Right after that, SESS charging is initiated. Since the voltage of supercapacitor cannot suddenly change, and supercapacitor has zero initial voltage,
it is equal to insert a short circuit fault into the system. Rload is henceforth
removed. It calculation is changed to Eq.(4)
It,t+0 =
EA
.
Xg + RLine
(4)
A transient is introduced into the system, the generator stator reactance in
0
Eq.(4) is replaced by transient reactance Xg which has smaller value compared with Xg , we got Eq.(5)
It,t++
=
0
EA
Xg + RLine
0
7
(5)
Compared to (2), both real and imaginary parts of denominator are reduced, the magnitude of denominator is smaller. At this moment, EA ’s value
can be reviewed as constant, the current magnitude is therefore increased.
0
Since usually Xg is 3-4 times of Xg , and RLoad >> RLine , We got the following
equation:
Xg
EA
0
Xg + RLine + RLoad
< Xg
EA
,
Xg + RLine
0
which means Vbus is reduced. We can further conclude that the reactive
power from generator also increases, according to:
Xg (
EA
0
Xg + RLine + RLoad
)2 < Xg (
EA
)2 ,
Xg + RLine
0
Reduced voltage influences the operation of other loads and increased
reactive power can reduce the fuel efficiency of generator. When these two
disturbances can be mitigated within the tolerated range, SESS charging can
be allowed.
4. Disturbance Mitigation control strategy
The critical impact is mainly correlated with the charging current, the
higher the charging current, the more severe its impact. It is expected that
charging control systems must strike a balance between rapid charging and
tolerable impact. According to analysis presented in the previous section,
power quality impacts from SESS charging are reflected in MVDC bus and
generator reactive power output. Two metrics are developed to evaluate the
impact from SESS charging:
leftmargan=*,labelsep=3mm M1 = |Vbus,lim −Vbus |, Vbus,lim is the input limit
of MVDC bus voltage, Vbus is the real time
measured MVDC bus voltage;
leftmbrgbn=*,lbbelsep=3mm M2 = QM T G , QM T G is the real time measured
generator reactive power.
For designing such power systems, standards are employed as guiding
principles. One of the applicable standards is related to the expected power
quality range and its limitations, such as [28]. For any device operating
within this system, their control designs should ensure that as long as the
8
specified power quality is maintained, the devices should properly operate as
well. Thus, if the SESS charging controller can ensure the power quality is not
lowered below the limit required by the standard, all the loads can function
normally and their performance is acceptable forming the basis for the DMC
proposed in this paper. Assuming the DMC can restrain disturbance below
the set limits, the adopters of the proposed work need to modify the limits
defined by applicable standards. The customized upper limits of these two
metrics are input to DMC, and DMC maintains metric values below limit.
Therefore the expected balance can be reached, as shown in Figure 4. DMC
measures the metric value and determines the control signal sent to charging
circuit. The charging current is defined as the current injected from MVDC
bus positive node to SESS module. We assume that during the charging
process, MVDC bus voltage settle at a steady state other than its rated
value. It is expected to maintain charging current with a minor variation,
because the charging power can be close to a constant value and facilitate
IPS reaching a new steady state during charging. Due to the variations in
the characteristics of MVDC bus voltage and MTG reactive power, there
are two different procedures for controlling charging current to according to
changes of M1 and M2 .
4.1. Control procedure for mitigating impact to MVDC bus voltage
The charging current can directly affect MVDC power, therefore, if we
can maintain charging current at a constant value, the bus voltage can also
stay close to a constant value. Hence the key part of mitigating impact to
MVDC bus voltage is to find the maximum charging current. The proposed
control procedure is:
1. Set an alert value, which is less than but close to the preset upper limit
of bus voltage deviation.
2. Start the controlled charging process with above alert and metric value.
3. When the metric value reaches the preset alert value, stop charging and
record the corresponding charging current.
4. Set this recorded maximum charging current as the reference for charging circuit and let the controller closely tracking it.
5. Stop charging when SESS voltage reaches the desired value.
9
M1
S1 firing
DMC pulse
Charging
circuit
Metric value
M2
Expected region
of metric value
metric curve
Charging time
Figure 4: DMC control strategy.
4.2. Control procedure for mitigating impact to MTG reactive power output
Existing test result shows the impact from charging current to MTG
reactive power output is indirect. There is a delay between the charging
current change and consequent reactive power change. Therefore, the control
procedure in mitigating bus voltage cannot be applied to mitigating reactive
power increment. An adaptive control strategy is developed for reactive
power output limitation:
1. Set an alert value, which is less than but close to the preset upper limit
10
Start charging
Stop charging
when metric value
first time hit alert
value
Record the current
value moment before
suspension
Set the current
reference value
based on the
recorded value
Resume charging and
regulate charging
with acquired
reference value
Stop charging when
SESS stored voltage
reaches desired value
Figure 5: Flow chart for mitigating MVDC bus voltage sag
of bus voltage deviation.
2. Start the controlled charging process.
3. When the metric value reaches the preset alert value, stop charging and
record the corresponding charging current. Set this charging current
value as ‘maximum charging value’ for implementing the algorithm.
11
4. If the metric value continues to increase and exceeds the upper limit,
use a suitable attenuation factor α (0 < alpha < 1) to revise the value of
maximum charging value using the formula ‘attenuationf actorXrecordedchargingvalue’.
5. Resume charging and if the charging current increases to the revised
maximum charging value, stop charging and monitor the metric change.
If the metric value can still reach the upper limit, repeat step 4, 5 and
6.
6. Repeat iteratively until the metric does not exceed upper limit again,
set the last value of the product as maximum charging current. Let the
upper limit value times a positive coefficient (less than 1) be the lower
limit for charging current. This coefficient for the proposed work is set
to 0.9.
7. The reference current value is switched between the lower and upper
limit based on the actual, monitored value of the charging current. Although the current is oscillating, this oscillation impact can be tolerated
due the 10% difference between upper limit and lower limit.
8. Stop charging when SESS stored voltage reaches desired value.
The performance of the proposed control strategy depends on how quickly
the critical current value can be captured, how precisely the actual charging
current can track the reference current, and how immediately the charging
current value can be switched from one state to another state. A desired
controller hardware implementation meeting all of the aforementioned prerequisites is expected to be implemented.
5. Fast Switchable IGBT Based Hardware implementation of Real
Time Disturbance Metric Controller
The expected hardware should be capable of being immediately turning
on/off during SESS charging. It should also be capable of withstanding high
voltages up to 5 kV during turn off period. Furthermore, the switch should
also allow monetary overshoots of charing/discharging current. Based on
the these requirements for the fast response of controller, we conducted a
literature review related to the fast switchable power electronics devices. It
was inferred that a chopper IGBT module FD500R65KE3-K from Infineon
has the necessay functionalities for realizing the proposed control strategy.
According to [29], the turn off delay of this device has been reduced to only 7.3
µs. It can sustain up to 6.5 kV exerted voltage during turn off. By connecting
12
Start charging
Stop charging
when metric value
first time hit alert
value
Record the current
value moment before
suspension
When charging
current reaches upper
limit stop charging
Monitor the trend of
MTG reactive power
increase
Resume charging
If it exceeds
preset limit?
Yes
Let the last recorded
current value multiply
an attenuation factor to
generate an upper limit
No
Fix the upper limit
and lower limit for
charging current
Resume charging and
regulate charging
with acquired limit
value
Stop charging when
SESS stored voltage
reaches desired value
Figure 6: Flow chart for limit MTG reactive power output
13
multiple switches in parallel, a BUCK chopper which is capable of fast turn
on/off during high charing/discharing current can be implemented, and the
concerned disturbance can be maintained within desired regions. This IGBT
module is commercially-off-the-shelf and there may be similar products with
the necessary functionalities. The expected energy to be stored in SESS is
300 MW. To reduce the size of the capacitor, the stored voltage should be
as high as possible. On the other hand, the highest voltage stored in SESS
should not be too high to affect the turning on/off performance of S1 and S2
in Figure 2. According to [29], the best performance is verified when VCE of
the switch device is 3600 V. We allow a 10% increase of test voltage, and the
final charged voltage is set at 4 kV. According to Eq. (6), the capacitance of
SESS in this study case is 37.5 F.
1
Ec = CVc2 .
2
(6)
6. Test Result
In order to validate whether the proposed test can properly mitigate the
impact to IPS and ensure fast charging, test case is setup in the DRTS with a
simulation time step of 50 µs. The reaction of test system can well approach
the actual real time power system response. Overall initial generation output
in all the study cases is set to 70 MW and 5 MVAr. The generation margin
left for SESS charging is 30 MVA. 4 test cases are run: M1,limit =0.6 kV,
M1,limit =0.8 kV, M2,limit = 6 MVAr, M2,limit =10 MVAr. SESS charging begins
at the fifth second after a test is initialized.
6.1. Test case for MVDC bus voltage sag mitigation
Figure 7 shows test result when M1,limit is set to 0.8 kV and the alert
value is set to 4.205 kV. When bus voltage goes down and reaches this value,
the charging is suspended and current value is marked as 4.3 kA. Then, charging controller enables the actual current value stay at 4.3 kA. The charging
process lasts 19 second and stops when SESS stored voltage reaches 4 kV.
The minimum bus voltage is 4.201 kV. Average bus voltage value during
charging is 4.208 kV, and the maximum value is 4.3 kV. During the charging
the expected bus voltage value is maintained.
14
Figure 7: Case study: M1,limit = 0.8 kV
6.2. Test case for MTG reactive power output restrain
Figure 8 is used to demonstrate the effect of DMC when the limit of
M2 is set to 10 MVAr with the alert value is set to 9.5 MVAr. Attenuation
factor is set to 0.95. When charging current first reaches 3.3 kA, single MTG
reactive power output reaches 9.5 MVAr and hence charging is suspended.
MTG output continues to increase and peaks at 9.641 MVAr. Therefore
3.3 kA is set as the upper limit for charging current, and 2.97 kA is set as
the lower limit. It can be seen that after the setting, reactive exceeds the
limit twice, this is caused by the adjustment of generator excitation. After
a short transient, reactive power is stays below 10 Mvar. The average value
of M2 is 9.58 MVAr. It takes 26 seconds to inject 300 MJ energy to SESS.
Attenuation factor is set to 0.9.
Table 1 shows the summarization of all 4 test cases. Test results show
the proposed DMC can strike an optimal balance point between fast charging
and required power quality acquisition.
15
Figure 8: Case study: M2,limit = 10 Mvar
Table 1: Test cases result summarization
test setting
M1,limit =0.6 kV
M1,limit =0.8 kV
M2,limit =6 Mvar
M2,limit =10 Mvar
maximum metric value minimum metric value average metric value
4.04 kV
4.401 kV
4.02 kV
4.3 kV
4.201 kV
4.208 kV
6.04 Mvar
5.84 Mvar
5.89 Mvar
11.15 Mvar
9.23 Mvar
9.58 Mvar
charging current value charging time
3.9 kA
21 s
4.3 kA
19 s
(1.8, 2.0) kA
50 s
(2.97, 3.3) kA
26 s
7. Conclusions
In this paper, DMC strategy is proposed based on a real-time analysis of
SESS charing in MVDC based IPS. The proposed control strategy focus on
reaching a reasonable balance between fast charging SESS and maintaining
required power quality. Based on the first stage analysis, the impact from
SESS charging to power quality is assessed and studied. The underlying reason of the impact from SESS charging to power quality is the large charging
current. It is challenging to quantify the relation between charging current
and power quality. One feasible way for limiting impact from charging current is an online real time monitoring charging current and power quality.
16
Based on the real time monitoring result, an optimal charging current value
can be generated. The next challenging part is how to drive the charging current rapidly and precisely track the reference value. A fast switching charging
circuit based on the latest IGBT technology development is proposed. Test
results shows that using fast IGBT to implement SESS charging system can
accurately follow the control command sent from DMC. Realization of fast
SESS charging can maintain the required power quality.
In the future, greater disturbances caused by SESS charging, such as the
AC side harmonic influence on MVDC bus, is expected to be analyzed. The
proposed control strategy highly depends on the rapid monitoring and action.
In this paper, using DRTS, the effect of real time control has been simulated.
Simulation result shows if the device for monitoring and controlling is fast
enough, rapid SESS charging can be realized without sacrificing required
power quality. In the future, hardware-in-the-loop test in real time will
be performed in order to validate the proposed control strategy on actual
hardware. Finally, decentralized coordination between SESS charging control
and generator control will also be studied as future work.
8. Acknowledgements
Proposed work is supported by the Water Power Technology Office, Department of Energy and the INL Laboratory Directed Research & Development (LDRD) Program under DOE Idaho Operations Office Contract DEAC07-05ID14517.
References
[1] J. G. Ciezki, R. W. Ashton, Selection and stability issues associated with
a navy shipboard dc zonal electric distribution system, IEEE Transactions on power delivery 15 (2000) 665–669.
[2] I. Kondratiev, R. Dougal, Invariant based ship dc power system design,
in: Electric Ship Technologies Symposium (ESTS), 2011 IEEE, IEEE,
pp. 15–20.
[3] N. Doerry, Zonal ship design, Naval engineers journal 118 (2006) 39–53.
17
[4] R. Mo, H. Li, Hybrid energy storage system with active filter function
for shipboard mvdc system applications based on isolated modular multilevel dc/dc converter (im2dc), IEEE Journal of Emerging and Selected
Topics in Power Electronics (2016).
[5] N. Doerry, Next generation integrated power systems (ngips) for the
future fleet, in: IEEE Electric Ship Technologies Symposium.
[6] C. McCoy, Powering the 21st-century fleet, PROCEEDINGS-UNITED
STATES NAVAL INSTITUTE 126 (2000) 54–59.
[7] S. Young, J. Newell, G. Little, Beyond electric ship, Naval engineers
journal 113 (2001) 79–92.
[8] A. K. Panda, T. Penthia, Design and modeling of smes based sapf for
pulsed power load demands, International Journal of Electrical Power
& Energy Systems 92 (2017) 114–124.
[9] R. Soman, E. Davidson, S. McArthur, J. Fletcher, T. Ericsen, Modelbased methodology using modified sneak circuit analysis for power electronic converter fault diagnosis, IET Power Electronics 5 (2012) 813–
826.
[10] S. Castellan, R. Menis, A. Tessarolo, F. Luise, T. Mazzuca, A review
of power electronics equipment for all-electric ship mvdc power systems,
International Journal of Electrical Power & Energy Systems 96 (2018)
306–323.
[11] B. Cassimere, C. R. Valdez, S. Sudhoff, S. Pekarek, B. Kuhn, D. Delisle,
E. Zivi, System impact of pulsed power loads on a laboratory scale integrated fight through power (iftp) system, in: Electric Ship Technologies
Symposium, 2005 IEEE, IEEE, pp. 176–183.
[12] M. Steurer, M. Andrus, J. Langston, L. Qi, S. Suryanarayanan,
S. Woodruff, P. Ribeiro, Investigating the impact of pulsed power charging demands on shipboard power quality, in: Electric Ship Technologies
Symposium, 2007. ESTS’07. IEEE, IEEE, pp. 315–321.
[13] Y. Luo, M. Andrus, J. Lian, S. K. Srivastava, D. A. Cartes, Power
impact analysis of pulse power loads in the future integrated shipboard
18
power system, in: ASNE Electric Electric Machines Technology Symposium.
[14] J. M. Crider, S. D. Sudhoff, Reducing impact of pulsed power loads on
microgrid power systems, Smart Grid, IEEE Transactions on 1 (2010)
270–277.
[15] R. Omar, N. Rahim, Voltage unbalanced compensation using dynamic
voltage restorer based on supercapacitor, International Journal of Electrical Power & Energy Systems 43 (2012) 573–581.
[16] J. Chen, H. E. Garcia, J. S. Kim, S. M. Bragg-Sitton, Operations optimization of nuclear hybrid energy systems, Nuclear Technology 195
(2016) 143–156.
[17] J. Amy, Considerations in the design of naval electric power systems,
in: Power Engineering Society Summer Meeting, 2002 IEEE, volume 1,
IEEE, pp. 331–335.
[18] D. De, C. Klumpner, C. Patel, K. Ponggorn, M. Rashed, G. Asher,
Modelling and control of a multi-stage interleaved dc–dc converter with
coupled inductors for super-capacitor energy storage system, IET Power
Electronics 6 (2013) 1360–1375.
[19] Y. Luo, M. Panwar, M. Mohanpurkar, R. Hovsapian, Real time optimal
control of supercapacitor operation for frequency response, in: 2016
IEEE Power and Energy Society General Meeting (PESGM), pp. 1–5.
[20] M. Bash, R. Chan, J. Crider, C. Harianto, J. Lian, J. Neely, S. Pekarek,
S. Sudhoff, N. Vaks, A medium voltage dc testbed for ship power system
research, in: Electric Ship Technologies Symposium, 2009. ESTS 2009.
IEEE, IEEE, pp. 560–567.
[21] W.-S. Im, C. Wang, L. Tan, W. Liu, L. Liu, Cooperative controls for
pulsed power load accommodation in a shipboard power system, IEEE
Transactions on Power Systems 31 (2016) 5181–5189.
[22] L. Tan, Q. Yang, W. Im, W. Liu, Adaptive critic design based cooperative control for pulsed power loads accommodation in shipboard
power system, IET Generation, Transmission & Distribution 10 (2016)
2739–2747.
19
[23] Y. Luo, S. Srivastava, M. Andrus, D. Cartes, Application of distubance
metrics for reducing impacts of energy storage charging in an mvdc based
ips, in: 2013 IEEE Electric Ship Technologies Symposium (ESTS), pp.
287–291.
[24] Y. M. Yeap, N. Geddada, A. Ukil, Capacitive discharge based transient
analysis with fault detection methodology in dc system, International
Journal of Electrical Power & Energy Systems 97 (2018) 127–137.
[25] S. J. Sutton, P. L. Lewin, S. G. Swingler, Review of global hvdc subsea cable projects and the application of sea electrodes, International
Journal of Electrical Power & Energy Systems 87 (2017) 121–135.
[26] A. Khoudiri, K. Guesmi, D. Mahi, Spectral decomposition based approach for dc–dc converters modeling, International Journal of Electrical
Power & Energy Systems 61 (2014) 288–297.
[27] H. Saadat, Power system analysis, WCB/McGraw-Hill, 1999.
[28] Ieee recommended practice for 1 kv to 35 kv medium-voltage dc power
systems on ships, IEEE Std 1709-2010 (2010) 1–54.
[29] IGBT-Module, Infineon Technologies AG, 2014. Rev. 3.0.
20
| 3cs.SY
|
HOMOLOGY OF LITTLEWOOD COMPLEXES
arXiv:1209.3509v2 [math.RT] 8 Feb 2013
STEVEN V SAM, ANDREW SNOWDEN, AND JERZY WEYMAN
Abstract. Let V be a symplectic vector space of dimension 2n. Given a partition λ with at most n
parts, there is an associated irreducible representation S[λ] (V ) of Sp(V ). This representation admits
a resolution by a natural complex Lλ• , which we call the Littlewood complex, whose terms are
restrictions of representations of GL(V ). When λ has more than n parts, the representation S[λ] (V )
is not defined, but the Littlewood complex Lλ• still makes sense. The purpose of this paper is to
compute its homology. We find that either Lλ• is acyclic or it has a unique non-zero homology group,
which forms an irreducible representation of Sp(V ). The non-zero homology group, if it exists, can
be computed by a rule reminiscent of that occurring in the Borel–Weil–Bott theorem. This result
can be interpreted as the computation of the “derived specialization” of irreducible representations
of Sp(∞), and as such categorifies earlier results of Koike–Terada on universal character rings. We
prove analogous results for orthogonal and general linear groups. Along the way, we will see two
topics from commutative algebra: the minimal free resolutions of determinantal ideals and Koszul
homology.
Contents
1. Introduction
2. Preliminaries
3. Symplectic groups
4. Orthogonal groups
5. General linear groups
References
1
4
6
14
24
31
1. Introduction
1.1. Statement of main theorem. Let V be a symplectic vector space over the complex numbers of dimension 2n. Associated to a partition λ with at most n parts there is an irreducible
representation S[λ] (V ) of Sp(V ), and all irreducible representations of Sp(V ) are uniquely of this
form. The space S[λ] (V ) can be defined as the quotient of the usual Schur functor Sλ (V ) by the
sum of the images of all of the “obvious” Sp(V )-linear maps Sµ (V ) → Sλ (V ), where µ can be
obtained by removing from λ a vertical strip of size two. In other words, we have a presentation
M
Sµ (V ) → Sλ (V ) → S[λ] (V ) → 0.
λ/µ=(1,1)
This presentation admits a natural continuation to a resolution Lλ• = Lλ• (V ), which we call the Littlewood complex. It can be characterized as the minimal resolution of S[λ] (V ) by representations
which extend to GL(V ).
Date: February 8, 2013.
2010 Mathematics Subject Classification. 05E10, 13D02, 15A72, 20G05.
S. Sam was supported by an NDSEG fellowship and a Miller research fellowship. A. Snowden was partially
supported by NSF fellowship DMS-0902661. J. Weyman was partially supported by NSF grant DMS-0901185.
1
2
STEVEN V SAM, ANDREW SNOWDEN, AND JERZY WEYMAN
When the number of parts of λ exceeds n, it still makes sense to speak of the complex Lλ• ,
even though there is no longer an associated irreducible S[λ] (V ) (see §1.2 for a simple example).
However, Lλ• is typically no longer exact in higher degrees. A very natural problem is to compute
its homology, and this is exactly what the main theorem of this paper accomplishes:
Theorem 1.1. The homology of Lλ• is either identically zero or else there exists a unique i for
which Hi (Lλ• ) is non-zero, and it is then an irreducible representation of Sp(V ).
In fact, there is a procedure, called the modification rule (see §3.4), which allows one to compute exactly which homology group is non-zero and which irreducible representation it is. This rule
can be phrased in terms of a certain Weyl group action, and, in this way, the theorem is reminiscent
of the classical Borel–Weil–Bott theorem. There is also a more combinatorial description of the
rule in terms of border strips. See Theorem 3.6 for a precise statement.
We prove analogous theorems for the orthogonal and general linear groups, but for clarity of
exposition we concentrate on the symplectic case in the introduction. An analogous result for the
symmetric group can be found in [SS1, Proposition 7.4.3]; this will be elaborated upon in [SS2].
1.2. An example. Let us now give the simplest example of the theorem, namely λ = (1, 1). If
V
n ≥ 2 then the irreducible representation S[1,1] (V ) is the quotient of S(1,1) (V ) = 2 V by the line
spanned by the symplectic form (where we identify V with V ∗ via the form). The complex Lλ• is
thus
V
· · · → 0 → C → 2 V,
the differential being multiplication by the form. This complex clearly makes since even if n < 2.
When n = 0, the differential is surjective, and H1 = C, the trivial representation of Sp(V ).
When n = 1, the differential is an isomorphism and all homology vanishes. And when n ≥ 2, the
differential is injective and H0 = S[1,1] (V ). More involved examples can be found in §3.6.
1.3. Representation theory of Sp(∞). The proper context for Theorem 1.1 lies in the representation theory of Sp(∞). We now explain the connection, noting, however, that the somewhat
exotic objects discussed here are not used in our proof of Theorem 1.1 and do not occur in the
remainder of the paper. Let Rep(Sp(∞)) denote the category of “algebraic” representations1 of
Sp(∞). This category was first identified in [DPS], where it is denoted Tg . It is also studied from
a slightly different point of view in [SS2]. As shown in [SS2], there is a specialization functor
ΓV : Rep(Sp(∞)) → Rep(Sp(V )).
This functor is right exact, but not exact — the category Rep(Sp(∞)) is not semi-simple. The
Littlewood complex Lλ• (C∞ ) makes sense, and defines a complex in Rep(Sp(∞)). It is exact in
positive degrees and its H0 is a simple object S[λ] (C∞ ); all simple objects are uniquely of this
form. The Schur functor Sλ (C∞ ), as an object of Rep(Sp(∞)), has two important properties: it is
projective and it specializes under ΓV to Sλ (V ). We thus see that Lλ• (C∞ ) is a projective resolution
of S[λ] (C∞ ) and specializes under ΓV to Lλ• (V ). We therefore have the following observation, which
explains the significance of the Littlewood complex from this point of view:
Proposition 1.2. We have Lλ• (V ) = LΓV (S[λ] (C∞ )), i.e., Lλ• (V ) computes the derived specialization of the simple object S[λ] (C∞ ) to V .
We can thus rephrase Theorem 1.1 as follows:
Theorem 1.3. Let M be an irreducible algebraic representation of Sp(∞). Then LΓV (M ) is either
acyclic or else there is a unique i for which Li ΓV (M ) is non-zero, and it is then an irreducible
representation of Sp(V ).
1Technically, we should use the “pro” version of the category, which is opposite to the more usual “ind” version
of the category. See [SS2] for details.
HOMOLOGY OF LITTLEWOOD COMPLEXES
3
The symplectic Schur functors S[λ] exhibit stability for large dimensional vector spaces (as explained in [KT], but see also [EW] and [HTW]), but not in general, in contrast to the usual Schur
functors. A general strategy for dealing with problems involving symplectic Schur functors is to
pass to the stable range (e.g., work with C∞ ), take advantage of the simpler behavior there, and
then apply the specialization functor to return to the unstable range. For this strategy to be viable,
one must understand the behavior of the specialization functor. This was one source of motivation
for this project, and is accomplished by Theorem 1.3.
1.4. Relation to results of Koike–Terada. Theorem 1.3 categorifies results of [KT], as we now
explain. In [KT], a so-called universal character ring Λ is defined, and a ring homomorphism π
(“specialization”) from Λ to the representation ring of Sp(V ) is given. A basis s[λ] of Λ is given
and it is shown that the image under π of s[λ] is either 0 or (plus or minus) the character of an
irreducible representation of Sp(V ). In fact, Λ is the Grothendieck group of Rep(Sp(∞)), π is the
map induced by the specialization functor ΓV and s[λ] is the class of the simple object S[λ] (C∞ )
in the Grothendieck group. Thus the K-theoretic shadow of Theorem 1.3 is precisely the result of
[KT] on specialization. However, we note that our proof depends on [KT].
1.5. Koszul homology and classical invariant theory. Theorem 1.1 can be reinterpreted as
the calculation of the homology groups of the Koszul complex on the generators of an ideal which
arises in classical invariant theory. This will be explained in §3.2 (see also §4.2 and §5.2 for the
orthogonal and general linear groups). For now, we remark that Koszul homology seems to be
remarkably difficult to calculate, even for well-behaved classes of ideals, such as determinantal
ideals. Very few cases have been worked out explicitly; we point to [AH] for the case of codimension
2 perfect ideals, and [SW] for the case of codimension 3 Gorenstein ideals. Both of these classes
of ideals are determinantal. They are singled out because their Koszul homology modules are
Cohen–Macaulay (this property fails for all other determinantal ideals).
1.6. Resolutions of determinantal ideals. In §2.5, we will see how the interpretation of Theorem 1.1 in terms of Koszul homology in §1.5 can also be interpreted in terms of the minimal free
resolutions of certain modules Mλ supported on the determinantal varieties defined by the Pfaffians
of a generic skew-symmetric matrix. The coordinate ring of the determinantal variety is the module
M∅ and so the computation of its resolution becomes a special case of Theorem 1.1, and therefore
realizes this classical resolution as the first piece of a much larger structure. The orthogonal group
and general linear group correspond to determinantal varieties in generic symmetric matrices and
generic matrices, respectively. We refer the reader to [Wey, §6] for the calculation of the minimal
free resolutions of the coordinate rings of determinantal varieties.
1.7. Overview of proof. There are three main steps to the proof:
(a) We first establish a combinatorial result, relating Bott’s algorithm for calculating cohomology
of irreducible homogeneous bundles to the modification rule appearing in theVmain theorem.
(b) We then introduce a certain module Mλ over the polynomial ring A = Sym( 2 E) (where E is
an auxiliary vector space), and compute its minimal free resolution. The main tools are step
(a), the Borel–Weil–Bott theorem and the geometric method of the third author.
(c) Lastly, we identify Mλ with the S[λ] (V )-isotypic piece of the ring B = Sym(V ⊗ E). The results
of step (b) and the specialization homomorphism on K-theory (see [Koi], [KT], [Wen]) are used
to get enough control on Mλ to do this. Once the identification is made, the results of step (b)
give the minimal free resolution of B as an A-module.
The theorem then follows, as the Littlewood complex can be identified with a piece of the minimal
free resolution of B over A.
4
STEVEN V SAM, ANDREW SNOWDEN, AND JERZY WEYMAN
1.8. Notation and conventions. We always work over the complex numbers. It is possible to
work over any field of characteristic 0, but there does not seem to be any advantage to doing so. We
write ℓ(λ) for the number of parts of a partition λ. The rank of a partition λ, denoted rank(λ), is
the number of boxes on the main diagonal. We write λ† for the transpose of the partition λ. We will
occasionally use Frobenius coordinates to describe partitions, which we now recall. Let r = rank(λ).
For 1 ≤ i ≤ r, let ai (resp. bi ) denote the number of boxes to the right (resp. below) the ith diagonal
box, including the box itself. Then the Frobenius coordinates of λ are (a1 , . . . , ar |b1 , . . . , br ). We
denote by cλµ,ν the Littlewood–Richardson coefficients, i.e., the coefficient of the Schur function sλ in
the product sµ sν . For the relevant background on partitions, Schur functions, and Schur functors,
we refer to [Mac, Chapter 1] and [Wey, Chapters 1, 2].
2. Preliminaries
2.1. The geometric technique. Let X be a smooth projective variety. Let
0→ξ→ε→η→0
be an exact sequence of vector bundles on X, with ε trivial, and let V be another vector bundle on
X. Put
A = H0 (X, Sym(ε)),
M = H0 (X, Sym(η) ⊗ V).
Then A is a ring — in fact, it is the symmetric algebra on H0 (X, ε) — and M is an A-module. The
following proposition encapsulates what we need of the geometric technique of the third author.
For a proof, and a stronger result, see [Wey, §5.1].
V
Proposition 2.1. Assume Hj (X, i+j (ξ) ⊗ V) = 0 for i < 0 and all j. Then we have a natural
isomorphism
M
V
TorA
Hj (X, i+j (ξ) ⊗ V).
i (M, C) =
j≥0
2.2. The Borel–Weil–Bott theorem. Let U be the set of all integer sequences (a1 , a2 , . . .) which
are eventually 0. We identify partitions with non-increasing sequences in U (such sequences are
necessarily non-negative). For i ≥ 1, let si be the transposition which switches ai and ai+1 , and let
S be the group of automorphisms of U generated by the si . The group S is a Coxeter group (in
fact, the infinite symmetric group), and admits a length function ℓ : S → Z≥0 . By definition, the
length of w ∈ S is the minimum number ℓ(w) so that there exists an expression
(2.1)
w = si1 · · · siℓ (w) .
Alternatively, ℓ(w) is the number of inversions of w, interpreted as a permutation.
We define a second action of S on U , denoted •, as follows. For w ∈ S and λ ∈ U we put
w • λ = w(λ + ρ) − ρ, where ρ = (−1, −2, . . .). In terms of the generators, this action is:
si • (. . . , ai , ai+1 , . . .) = (. . . , ai+1 − 1, ai + 1, . . .).
Let λ be an element of U . Precisely one of the following two possibilities occurs:
• There exists a unique element w of S such that w • λ is a partition. In this case, we call λ
regular.
• There exists an element w 6= 1 of S such that w • λ = λ. In this case, we call λ singular.
Bott’s algorithm [Wey, §4.1] is a procedure for determining if λ is regular. It goes as follows.
Find an index i such that λi+1 > λi . If no such index exists, then λ is a partition and is regular.
If λi+1 − λi = 1 then λ is singular. Otherwise apply si to λ and repeat. Keeping track of the si
produces a minimal word for the element w in the definition (2.1). In particular, it is important
to note that we have a choice of which index i to pick in the first step. Different choices lead
to different minimal words, but the resulting partition and permutation are independent of these
choices.
HOMOLOGY OF LITTLEWOOD COMPLEXES
5
Let E be a vector space and let X be the Grassmannian of rank n quotients of E. (We assume
dim E ≥ n, obviously.) We have a tautological sequence on X
(2.2)
0 → R → E ⊗ OX → Q → 0,
where Q has rank n. For a partition λ with at most n parts and a partition µ, let (λ |n µ) be the
element of U given by (λ1 , . . . , λn , µ1 , µ2 , . . .). The Borel–Weil–Bott theorem [Wey, §4.1] is then:
Theorem 2.2 (Borel–Weil–Bott). Let λ be a partition with at most n parts, let µ be any partition
and let V be the vector bundle Sλ (Q) ⊗ Sµ (R) on X.
• Suppose (λ |n µ) is regular, and write w • (λ |n µ) = α for a partition α. Then
(
Sα (E) if i = ℓ(w)
Hi (X, V) =
0
otherwise.
• Suppose (λ |n µ) is singular. Then Hi (X, V) = 0 for all i.
Remark 2.3. If ℓ(µ) > rank(R) then V = 0. Similarly, if ℓ(α) > dim(E) then Sα (E) = 0.
These problems disappear if dim(E) is sufficiently large compared to λ and µ; in fact, the situation
becomes completely uniform when dim(E) = ∞.
2.3. Resolution of the second Veronese ring. In our treatment of odd orthogonal groups, we
need to know the resolution for the second Veronese ring in a relative setting. We state the relevant
result here, so as not to interrupt the discussion later.
Let X be a variety and let R be a vector bundle on X. Let π : P(R) → X be the associated
projective space bundle of one dimensional quotients of R. Let π ∗ (R) → L be the universal rank
one quotient. Define ξ to be the kernel of the map Sym2 (π ∗ R) → Sym2 (L). The result we need is
the following:
Proposition 2.4. Let a be 0 or 1. We have
M
M
V
Sµ (R),
Rj π∗ ( i+j (ξ) ⊗ La ) =
j∈Z
µ
where the sum is over those partitions µ such that µ = µ† , rank(µ) = a (mod 2) and i = 12 (|µ| −
rank(µ)).
Proof. This is a relative version of the calculation
of the minimal free resolution (over Sym(U ))
L
2d
of the second Veronese ring M (U )0 =
d≥0 Sym (U ) (a = 0) and its odd Veronese module
L
M (U )1 = d≥0 Sym2d+1 (U ) (a = 1), where U is some vector space.
The case a = 0 is contained in [Wey, Theorem 6.3.1(c)]. Now we calculate the case a = 1. Note
that it is functorial in R, so due to the stability properties of Schur functors, if we calculate the
resolution for rank U = N , the same result also holds for rank U < N . So it is enough to handle
the case that rank U is odd, but arbitrarily large.
Consider the total space of O(−2) on P(U ) with structure map π ′ : O(−2) → P(U ), and define
′
L = π ′ ∗ OP(U ) (1). Also consider the map p : O(−2) → Spec(Sym(U )). Then M (U )a = p∗ (L′ ⊗a ).
Using this setup and [Wey, Theorem 5.1.4], we see that the Ext dual [Wey, Proposition 1.2.5] of
M (U )0 is M (U )1 when rank U is odd. All of the partitions µ in the free resolution of M (U )0 fit in
a square of size rank U , and on the level of the partitions that index the Schur functors appearing
in the free resolution, this duality amounts to taking complements within this square, and then
reversing the direction of the arrows, hence the result follows.
Remark 2.5. Let ε = Sym2 (π ∗ R) and η = Sym2 (L), so that we have an exact sequence
0 → ξ → ε → η → 0.
6
STEVEN V SAM, ANDREW SNOWDEN, AND JERZY WEYMAN
The ring π∗ (Sym(η)) is identified with Sym(Sym2 (R)), i.e., the projective coordinate ring of the
second Veronese of P(R). By aVrelative version of the geometric method, its minimal locally free
resolution is computed by Rπ∗ ( • (ξ)), i.e., the sheaves appearing in the proposition.
2.4. A criterion for degeneration of certain spectral sequences. Let π : X ′ → X be a map
of proper varieties and let V be a coherent sheaf on X ′ . We say that (π, V) is degenerate if the
Leray spectral sequence
i
j
i+j
Ei,j
(X, V)
2 = H (X, R π∗ (V)) ⇒ H
degenerates at the second page. The following is a simple criterion for degeneracy that applies in
our one case of interest:
Lemma 2.6. Suppose that a group G acts
X ′ and that π and V are G-equivariant.
L oni X and
j
Suppose furthermore that the G-module
i,j H (X, R π∗ (V)) is semi-simple and multiplicity-free.
Then (π, V) is degenerate.
Proof. The differentials of the spectral sequence are G-equivariant, and thus forced to vanish.
2.5. A lemma from commutative algebra. We now give a very simple lemma that allows us
to interpret Koszul homology groups as Tor’s. This is useful since we are ultimately interested in
certain Koszul homology groups, but the geometric technique computes Tor’s.
Let B be a graded C-algebraVand let U be a homogeneous subspace of B. We can then form
the Koszul complex K• = B ⊗ • U . If f1 , . . . , fn is a basis for U then K• is the familiar Koszul
complex on the fi . Let A = Sym(U ), so that there is a natural homomorphism A → B. We then
have the following result:
Lemma 2.7. There is a natural identification TorA
i (B, C) = Hi (K• ).
Proof. We can resolve C as an A-module using the Koszul resolution A ⊗
with B gives K• , and is also how one computes TorA
• (B, C).
V•
U . Tensoring over A
3. Symplectic groups
3.1. Representations of Sp(V ). Let (V, ω) be a symplectic space of dimension 2n (here ω ∈
V2 ∗
V is the symplectic form, and gives an isomorphism V ∼
= V ∗ ). As stated in the introduction,
the irreducible representations of Sp(V ) are indexed by partitions λ with ℓ(λ) ≤ n (see [FH,
§17.3]). We call such partitions admissible. For an admissible partition λ, we write S[λ] (V ) for
the corresponding irreducible representation of Sp(V ).
V
3.2. The Littlewood complex. Let E be a vector space. Put U = 2 E, A = Sym(U ) and
B = Sym(E ⊗ V ). Consider the inclusion U ⊂ B given by
V
V
V2
E ⊂ 2 E ⊗ 2 V ⊂ Sym2 (E ⊗ V ),
where the first inclusion is multiplication by ω. This inclusion defines an algebra homomorphism
A → B. Put C = B ⊗A C; this is the quotient of B by the ideal generated by U . We have maps
Spec(C) → Spec(B) → Spec(A).
We have a natural identification of Spec(B) with the space Hom(E, V ) of linear maps ϕ : E → V and
V
of Spec(A) with the space 2 (E)∗ of anti-symmetric forms on E. The map Spec(B) → Spec(A)
takes a linear map ϕ to the form ϕ∗ (ω). The space Spec(C), which we call the Littlewood
variety, is the scheme-theoretic fiber of this map above 0, i.e., it consists of those maps ϕ for
which ϕ∗ (ω) = 0. In other words, Spec(C) consists of maps ϕ : E → V such that the image of ϕ is
an isotropic subspace of V .
HOMOLOGY OF LITTLEWOOD COMPLEXES
7
V
Let K• (E) = B ⊗ • U be the Koszul complex of the Littlewood variety. We can decompose this
complex under the action of GL(E):
M
K• (E) =
Sλ (E) ⊗ Lλ• .
ℓ(λ)≤dim E
Lλ•
is the Littlewood complex, and is independent of E (so long as dim E ≥ ℓ(λ)).
The complex
By [How, Theorem 3.8.6.2], its zeroth homology is
(
S[λ] (V ) if λ is admissible
λ
(3.1)
H0 (L• ) =
0
otherwise.
By Lemma 2.7, we have Hi (K• ) = TorA
i (B, C), and so we have a decomposition
M
(3.2)
TorA
Sλ (E) ⊗ Hi (Lλ• ).
i (B, C) =
ℓ(λ)≤dim E
Applied to i = 0, we obtain
(3.3)
C=
M
Sλ (E) ⊗ S[λ] (V ).
admissible λ
Remark 3.1. It is possible to compute the terms of Lλ• explicitly. Let Q−1 be the set of partitions
λ whose Frobenius coordinates (a1 , . . . , ar |b1 , . . . , br ) satisfy ai = bi − 1 for each i (see §3.5 for
further discussion of this set). Then
M
Lλi =
Sλ/µ (V ).
µ∈Q−1 ,
|µ|=2i
If λ is admissible then the higher homology of Lλ• vanishes (see Proposition 3.2 below), and so,
taking Euler characteristics, we get an equality in the representation ring of Sp(V ):
X
(−1)|µ|/2 [Sλ/µ (V )].
[S[λ] (V )] =
µ∈Q−1
The significance of this identity is that it expresses the class of the irreducible S[λ] (V ) in terms of
representations which are restricted from GL(V ). It is due to Littlewood [Lit, p.295] (see also [KT,
Prop. 1.5.3(2)]), and is why we name the complexes Lλ• after him.
3.3. A special case of the main theorem. Our main theorem computes the homology of the
complex Lλ• . We now formulate and prove the theorem in a particularly simple case. We mention
this here only because it is worthwhile to know; the argument is not needed to prove the main
theorem.
Proposition 3.2. Suppose λ is admissible. Then
(
S[λ] (V )
Hi (Lλ• ) =
0
if i = 0
otherwise.
Proof. Choose E to be of dimension n. By Lemma 3.3 below, K• (E) has no higher homology. It
follows that Lλ• does not either. The computation of H0 (Lλ• ) is given in (3.1).
Lemma 3.3. Suppose dim E ≤ n. Then U ⊂ B is spanned by a regular sequence.
Proof. It suffices to show that dim Spec(C) = dim Spec(B) − dim U . Put d = dim E. Observe that
the locus in Spec(C) where ϕ is injective is open. Let IGr(d, V ) be the variety of d-dimensional
isotropic subspaces of V , which comes with a rank d tautological bundle R ⊂ V ⊗ OIGr(d,V ) . There
is a natural birational map from the total space of Hom(E, R) to Spec(C), and thus Spec(C) has
dimension 2nd − 21 d(d − 1). As dim Spec(B) = 2nd and dim U = 12 d(d − 1), the result follows.
8
STEVEN V SAM, ANDREW SNOWDEN, AND JERZY WEYMAN
3.4. The modification rule. We now associate to a partition λ two quantities, i2n (λ) and τ2n (λ),
which will be used to describe the homology of Lλ• . (Recall that 2n = dim V .) We give two
equivalent definitions of these quantities, one via a Weyl group action and one via border strips.
We begin with the Weyl group definition, following [Wen, §1.5]. Recall that in §2.2 we defined
automorphisms si of the set U of integer sequences, for i ≥ 1. We now define an additional
automorphism: s0 negates a1 . We let W be the group generated by the si , for i ≥ 0. Then W is a
Coxeter group of type BC∞ , and, as such, is equipped with a length function ℓ : W → Z≥0 , which
is defined just as in (2.1). Let ρ = (−(n + 1), −(n + 2), . . .). Define a new action of W on U by
w • λ = w(λ + ρ) − ρ. On S this action agrees with the one defined in §2.2, despite the difference
in ρ. The action of s0 is given by
s0 • (a1 , a2 , . . .) = (2n + 2 − a1 , a2 , . . .).
Given a partition λ ∈ U , exactly one of the following two possibilities hold:
• There exists a unique element w ∈ W such that w•λ† = µ† is a partition and µ is admissible.
We then put i2n (λ) = ℓ(w) and τ2n (λ) = µ.
• There exists a non-identity element w ∈ W such that w • λ† = λ† . We then put i2n (λ) = ∞
and leave τ2n (λ) undefined.
Note that if λ is an admissible partition then we are in the first case with w = 1, and so i2n (λ) = 0
and τ2n (λ) = λ.
We now give the border strip definition, following [Sun, §5] (which is based on [Kin]). If ℓ(λ) ≤ n
we put i2n (λ) = 0 and τ2n (λ) = λ. Suppose ℓ(λ) > n. Recall that a border strip is a connected
skew Young diagram containing no 2 × 2 square. Let Rλ be the connected border strip of length
2(ℓ(λ) − n − 1) which starts at the first box in the final row of λ, if it exists. If Rλ exists, is nonempty and λ \ Rλ is a partition, then we put i2n (λ) = c(Rλ ) + i2n (λ \ Rλ ) and τ2n (λ) = τ2n (λ \ Rλ ),
where c(Rλ ) denotes the number of columns that Rλ occupies; otherwise we put i2n (λ) = ∞ and
leave τ2n (λ) undefined.
Remark 3.4. There is an alternative way to think about removing Rλ in terms of hooks. Given
a box b in the Young diagram of λ, recall that the book of b is the set of boxes which are either
directly below b or directly to the right of b (including b itself). The border strips R of λ that
begin at the last box in the first column, and have the property that λ \ R is a Young diagram,
are naturally in bijection with the boxes in the first column: just take the box bR in the same row
where R ends. The important point is that the size of this border strip is the same as size of the
hook of bR , and removing R is the same as removing the hook of bR and shifting all boxes below
this hook one box in the northwest direction. This is illustrated in the following diagram:
The shaded boxes indicate the border strip (left diagram) and hook (right diagram).
The agreement of the above two definitions may be known to some experts, but we are unaware
of a reference, so we provide a proof.
Proposition 3.5. The above two definitions agree.
Proof. Suppose that we are removing a border strip Rλ of size 2(ℓ(λ) − n − 1) from λ which begins
at the first box in the final row of λ. Let c = c(Rλ ) be the number of columns of Rλ . The sequence
HOMOLOGY OF LITTLEWOOD COMPLEXES
9
(sc−1 sc−2 · · · s1 s0 ) • λ† is
(λ†2 − 1, λ†3 − 1, . . . , λ†c − 1, 2n + 2 − λ†1 + c − 1, λ†c+1 , λ†c+2 , . . . ),
and these are the same as the column lengths of λ \ Rλ .
Conversely, if we use the Weyl group modification rule with w ∈ W , then the expression (2.1)
for w must begin with s0 : if we apply any si with i > 0, then we increase the number of inversions
of the sequence, so if we write wsi = v, then ℓ(v) = ℓ(w) + 1 [Hum, §5.4, Theorem], so the resulting
expression for w will not be minimal. If we choose i maximal so that w = w′ si−1 · · · s1 s0 with
ℓ(w) = ℓ(w′ ) + i, then we have replaced the first column of λ with 2n + 2 − λ†1 and then moved it
over to the right as much as possible (adding 1 to it each time we pass a column and subtracting
1 from the column we just passed) so that the resulting shape is again a Young diagram. This is
the same as removing a border strip of length 2(ℓ(λ) − n − 1) with i columns.
Finally, there is a third modification rule, defined in [KT, §2.4]. We will not need to know the
statement of the rule, but we will cite some results from [KT], so we need to know that their rule
is equivalent to the previous two. The equivalence of the rule from [KT, §2.4] with the border strip
rule comes from the fact that both rules were derived from the same determinantal formulas (see
[KT, Theorem 1.3.3] and [Kin, Footnote 18]).
3.5. The main theorem. Our main theorem is the following:
Theorem 3.6. For a partition λ and an integer i we have
(
S[τ2n (λ)] (V ) if i = i2n (λ)
λ
Hi (L• ) =
0
otherwise.
In particular, if i2n (λ) = ∞ then Lλ• is exact.
Remark 3.7. Consider the coordinate ring R of rank ≤ 2n skew-symmetric matrices; identifying
E with its dual, this is the quotient of A by the ideal generated by 2(n + 1) × 2(n + 1) Pfaffians. A
description of the resolution of R over A can be found in [Wey, §6.4] and [JPW, §3]. On the other
hand, R is the Sp(V )-invariant part of B, and so the above theorem, combined with (3.2), shows
that Sλ (E) appears in its resolution if and only if τ2n (λ) = ∅. Thus the modification rule gives
an alternative description of the resolution of R. It is a pleasant combinatorial exercise to show
directly that these two descriptions agree. While the description in terms of the modification rule
is more complicated, it has the advantage that it readily generalizes to our situation.
The proof of the theorem will take the remainder of this section. We follow the three-step plan
outlined in §1.7. Throughout, the space V is fixed and n = 21 dim(V ).
Step a. Let Q−1 be the set of partitions λ whose Frobenius coordinates (a1 , . . . , ar |b1 , . . . , br ) satisfy
ai = bi − 1 for all i. This set admits an inductive definition that will be useful for us and which
we now describe. The empty partition belongs to Q−1 . A non-empty partition µ belongs to Q−1 if
and only if the number of rows in µ is one more than the number of columns, i.e., ℓ(µ) = µ1 + 1,
and the partition obtained by deleting the first row and column of µ, i.e., (µ2 − 1, . . . , µℓ(µ) − 1),
belongs to Q−1 . The significance of this set is the plethysm
M
V• V2
Sµ (E)
( (E)) =
µ∈Q−1
(see [Mac, I.A.7, Ex. 4]).
Let λ be a partition with ℓ(λ) ≤ n. We write (λ|µ) in place of (λ |n µ) in this section. Define
S1 (λ) = {µ ∈ Q−1 such that (λ|µ) is regular}
S2 (λ) = {partitions α such that τ2n (α) = λ}.
10
STEVEN V SAM, ANDREW SNOWDEN, AND JERZY WEYMAN
Lemma 3.8. Let µ be a non-zero partition in S1 (λ) and let ν be the partition obtained by removing
the first row and column of µ. Then ν also belongs to S1 (λ). Furthermore, let w (resp. w′ ) be the
unique element of W such that α = w • (λ|µ) (resp. β = w′ • (λ|ν)) is a partition. Then the border
strip Rα is defined (see §3.4) and we have the following identities:
|Rα | = 2µ1 ,
α \ Rα = β,
c(Rα ) = µ1 + ℓ(w′ ) − ℓ(w).
Proof. Suppose that in applying Bott’s algorithm to (λ|µ) the number µ1 moves r places to the
left. Thus, after the first r steps of the algorithm, we reach the sequence
(λ1 , . . . , λn−r , µ1 − r, λn−r+1 + 1, . . . , λn + 1, µ2 , . . . , µℓ(µ) ).
Notice that the subsequence starting at λn−r+1 + 1 is the same as the subsequence of (λ|ν) starting
at λn−r+1 , except 1 has been added to each entry of the former. It follows that Bott’s algorithm
runs in exactly the same manner on each. In particular, if (λ|ν) were not regular then (λ|µ)
would not be either; this shows that ν belongs to S1 (λ). Suppose that Bott’s algorithm on (λ|ν)
terminates after N = ℓ(w′ ) steps. By the above discussion, Bott’s algorithm on (λ|µ) terminates
after N + r = ℓ(w) steps, and we have the following formula for α:
1≤i≤n−r
λi
αi = µ1 − r
i=n−r+1
βi−1 + 1 n − r + 2 ≤ i ≤ n + µ1 + 1
Since ℓ(α) = n + µ1 + 1, the border strip Rα has 2µ1 boxes. Using Remark 3.4, we see that Rα
exists since the box in the (n − r + 1)th row and the first column has a hook of size 2(ℓ(α) − n − 1).
Furthermore, α \ Rα = β and c(Rα ) = µ1 − r. Since r = ℓ(w) − ℓ(w′ ), the result follows.
Lemma 3.9. Let ν belong to S1 (λ) and suppose that w′ ∈ W is such that w′ • (λ|ν) = β is a
partition. Let α be a partition such that Rα is defined and α \ Rα = β. Then there exists a partition
µ ∈ S1 (λ) and an element w ∈ W such that w • (λ|µ) = α, and the partition obtained from µ by
removing the first row and column is ν.
Proof. Reverse the steps of Lemma 3.8.
Proposition 3.10. There is a unique bijection S1 (λ) → S2 (λ) under which µ maps to α if there
exists w ∈ S such that w • (λ|µ) = α; in this case, ℓ(w) + i2n (α) = 12 |µ|.
Proof. Let µ be an element of S1 (λ) and let w ∈ W be such that w • (λ|µ) = α is a partition. We
show by induction on |µ| that α belongs to S2 (λ) and that ℓ(w) + i2n (α) = 12 |µ|. For |µ| = 0 this
is clear: w = 1 and α = λ. Suppose now that µ is non-empty. In what follows, we tacitly employ
Lemma 3.8. Let ν be the partition obtained by removing the first row and column of µ. Then ν
belongs to S1 (λ), and so we can choose w′ ∈ W such that w′ •(λ|ν) = β is a partition. By induction
we have τ2n (β) = λ and ℓ(w′ ) + i2n (β) = 12 |ν|. Since α \ Rα = β, we have τ2n (α) = τ2n (β) = λ.
Furthermore, i2n (α) = c(Rα ) + i2n (β), and so
i2n (α) = µ1 + ℓ(w′ ) − ℓ(w) + i2n (β) = 21 |µ| − ℓ(w).
This completes the induction.
We have thus shown that µ 7→ α defines a map of sets S1 (λ) → S2 (λ). We now show that
this map is injective. Suppose µ and µ′ are two elements of S1 (λ) that both map to α. Then
the sequences (λ|µ) + ρ and (λ|µ′ ) + ρ are identical as multisets of numbers. In particular, we
can rearrange the sequence of numbers to the right of the bar of (λ|µ) + ρ to get the sequence of
numbers to the right of the bar of (λ|µ′ ) + ρ. But both of these sequences (to the right of the bar)
are strictly decreasing, so we see that µ = µ′ .
Finally, we show that µ 7→ α is surjective. The partition α = λ has the empty partition as its
preimage. Suppose now that α 6= λ belongs to S2 (λ), and let β = α \ Rα . By induction on size, we
HOMOLOGY OF LITTLEWOOD COMPLEXES
11
can find ν ∈ S1 (λ) mapping to β. Applying Lemma 3.9, we find a partition µ ∈ S1 (λ) mapping to
α. This completes the proof.
Step b. Let E be a vector space of dimension at least n. Let X be the Grassmannian
V of rank n
quotients of E. Let R and Q be the tautological bundles on X as in (2.2). Put ε = 2 (E) ⊗ OX ,
V
ξ = 2 R and define η by the exact sequence
0 → ξ → ε → η → 0.
Finally, for a partition λ with ℓ(λ) ≤ n, put Mλ = Sym(η) ⊗ Sλ (Q) and Mλ = H0 (X, Mλ ). Note
that A = H0 (X, Sym(ε)), and so Mλ is an A-module.
Lemma 3.11. Let λ be a partition with ℓ(λ) ≤ n and µ ∈ S1 (λ) correspond to α ∈ S2 (λ). Then
(
Sα (E) if i = 12 |µ| − i2n (α)
i
H (X, Sλ (Q) ⊗ Sµ (R)) =
0
otherwise.
Proof. This follows immediately from Proposition 3.10 and the Borel–Weil–Bott theorem.
Lemma 3.12. Let λ be a partition with ℓ(λ) ≤ n and let i be an integer. We have
M
M
V
Sα (E),
Hj (X, i+j (ξ) ⊗ Sλ (Q)) =
α
j∈Z
where the sum is over partitions α with τ2n (α) = λ and i2n (α) = i. In particular, when i < 0 the
left side above vanishes.
Proof. We have
Vi+j
(ξ) =
Vi+j V2
( (R)) =
M
Sµ (R),
µ∈Q−1 ,
|µ|=2(i+j)
and so
M
j∈Z
Hj (X,
Vi+j
(ξ) ⊗ Sλ (Q)) =
M
H|µ|/2−i (X, Sµ (R) ⊗ Sλ (Q)).
µ∈Q−1
The result now follows from the previous lemma.
Proposition 3.13. We have
TorA
i (Mλ , C) =
M
Sα (E),
α
where the sum is over partitions α with τ2n (α) = λ and i2n (α) = i.
Proof. This follows immediately from the previous lemma and Proposition 2.1.
Step c. For a partition λ with at most n parts put Bλ = HomSp(V ) (S[λ] (V ), B). Note that Bλ is
an A-module and has a compatible action of GL(E). Our goal is to show that Bλ is isomorphic to
Mλ .
Lemma 3.14. The spaces Bλ and Mλ are isomorphic as representations of GL(E) and have finite
multiplicities.
L
Proof. Let M =
ℓ(λ)≤n Mλ ⊗ S[λ] (V ). It is enough to show that M and B are isomorphic as
representations of GL(E) × Sp(V ) and have finite multiplicities. In fact, it is enough to show that
the Sθ (E) multiplicity spaces of M and B are isomorphic as representations of Sp(V ) and have
finite multiplicities. This is what we do.
12
STEVEN V SAM, ANDREW SNOWDEN, AND JERZY WEYMAN
The Sθ (E) multiplicity space of B is Sθ (V ). The decomposition of this in the representation ring
of Sp(V ) can be computed by applying the specialization homomorphism to [KT, Thm. 2.3.1(1)].
The result is
X
(−1)i2n (ν) cθ(2µ)† ,ν [S[τ2n (ν)] (V )].
µ,ν
Note that for a fixed θ there are only finitely many values for µ and ν which make the Littlewood–
Richardson coefficient non-zero, which establishes finiteness of the multiplicities. Now, we have an
equality
X
[Mλ ] = [A]
(−1)i [TorA
i (Mλ , C)]
i≥0
in the representation ring of GL(E). Applying Proposition 3.13, we find
X
X
(−1)i [TorA
(−1)i2n (ν) [Sν (E)].
i (Mλ , C)] =
i≥0
As [A] =
P
µ [S(2µ)† (E)]
τ2n (ν)=λ
(see [Mac, I.A.7, Ex. 2]), we obtain
X
(−1)i2n (ν) cθ(2µ)† ,ν [Sθ (E)].
[Mλ ] =
ν,µ,θ
τ2n (ν)=λ
We therefore find that the Sθ (E)-component of M is given by
X
(−1)i2n (ν) cθ(2µ)† ,ν [S[λ] (V )].
λ,µ
τ2n (ν)=λ
The result now follows.
Proposition 3.15. We have an isomorphism Mλ → Bλ which is A-linear and GL(E)-equivariant.
Proof. According to Proposition 3.13, we have
TorA
0 (Mλ , C) = Sλ (E),
TorA
1 (Mλ , C) = S(λ,12n+2−2ℓ(λ) ) (E),
since the only ν for which τ2n (ν) = λ and i2n (ν) ≤ 1 must agree with λ everywhere except possibly
the first column. We therefore have a presentation
S(λ,12n+2−2ℓ(λ) ) (E) ⊗ A → Sλ (E) ⊗ A → Mλ → 0.
Note that S(λ,12n+2−2ℓ(λ) ) (E) occurs with multiplicity one in Sλ (E) ⊗ A, and thus does not occur in
Mλ ; it therefore does not occur in Bλ either, since Mλ and Bλ are isomorphic as representations
of GL(E) by Lemma 3.14.
Now, TorA
0 (B, C) is the coordinate ring of the Littlewood variety, and its S[λ] (V ) multiplicity
space is Sλ (E) by (3.3). We therefore have a surjection f : Sλ (E)⊗A → Bλ . Since S(λ,12n+2−2ℓ(λ) ) (E)
does not occur in Bλ , the copy of S(λ,12n+2−2ℓ(λ) ) (E) in Sλ (E)⊗A lies in the kernel of f , and therefore
f induces a surjection Mλ → Bλ . Finally, since the two are isomorphic as GL(E) representations
and have finite multiplicity spaces, this surjection is an isomorphism.
Combining this proposition with Proposition 3.13, we obtain the following corollary.
Corollary 3.16. We have
TorA
i (B, C) =
M
Sλ (E) ⊗ S[τ2n (λ)] (V ).
i2n (λ)=i
Combining this with (3.2) yields the main theorem. (We can choose E to be arbitrarily large.)
HOMOLOGY OF LITTLEWOOD COMPLEXES
13
Remark 3.17. The arguments of step c made no use of the construction of the module Mλ , simply
that it satisfied Proposition 3.13. More precisely, say that a GL(E)-equivariant A-module M is of
“type λ” (for an admissible partition λ) if
M
(M,
C)
=
Sα (E),
TorA
i
α
where the sum is over all partitions α with τ2n (α) = λ and i2n (α) = i. Then the arguments of
step c establish the following statement: if a type λ module exists then it is isomorphic to Bλ , and
thus Bλ has type λ. (Actually the argument is a bit weaker, since it works with all λ at once.)
Step b can be thought of as simply providing a construction of a module of type λ.
3.6. Examples. We now give a few examples to illustrate the theorem.
V
V
Example 3.18. Suppose λ = (1i ). Then Lλ• is the complex i−2 V → i V , where the differential
V
is the multiplication by the symplectic form on V ∗ treated as an element of 2 V .
• If i ≤ n then the differential is injective, and H0 (Lλ• ) = S[1i ] (V ) is an irreducible representation of V .
• If i = n + 1 then the differential is an isomorphism, and all homology of Lλ• vanishes.
• If n + 2 ≤ i ≤ 2n + 2 then the differential is surjective and H1 (Lλ• ) = S[12n−i+2 ] (V ).
• If i > 2n + 2 then the complex Lλ• is identically 0.
Example 3.19. Suppose λ = (2, 1, 1). Then Lλ• is the complex C → S(2,1,1)/(1,1) (V ) → S(2,1,1) (V ),
where
the differential is the multiplication by the symplectic form on V ∗ treated as an element of
V2
V.
• If n ≥ 3 then the differential is injective, and H0 (Lλ• ) = S[2,1,1] (V ) is an irreducible representation of V .
• If n = 2 then the complex is exact, and all homology of Lλ• vanishes.
• If n = 1 then H1 (Lλ• ) = S[2] (V ).
• Finally when n = 0 then H2 (Lλ• ) = C.
The reader will check easily that in both instances the description of the homology agrees with
the rule given by the Weyl group action.
Example 3.20. Suppose λ = (6, 5, 4, 4, 3, 3, 2) and n = 2 (so dim(V ) = 4). The modification rule,
using border strips, proceeds as follows:
We start on the left with λ0 = λ. As ℓ(λ0 ) = 7, we are supposed to remove the border strip R0 of
size 2(ℓ(λ0 ) − n − 1) = 8; this border strip is shaded. The result is the second displayed partition,
λ1 = (6, 5, 3, 2, 2, 1). As ℓ(λ1 ) = 6, the border strip R1 we remove from it has length 6. The result
of removing this strip is the third partition λ2 = (6, 5, 1, 1). As ℓ(λ2 ) = 4, the border strip R2 has
length 2. The result of removing it is the final partition λ3 = (6, 5). This satisfies ℓ(λ3 ) ≤ n, so the
algorithm stops. We thus see that τ4 (λ) = (6, 5) and
i4 (λ) = c(R0 ) + c(R1 ) + c(R2 ) = 4 + 3 + 1 = 8.
It follows that Hi (Lλ• ) = 0 for i 6= 8 and H8 (Lλ• ) = S[6,5] (C4 ).
14
STEVEN V SAM, ANDREW SNOWDEN, AND JERZY WEYMAN
s
i
Now we illustrate the modification rule using the Weyl group action. We write α −
→
β if
β = si (α). The idea for getting the Weyl group element is to apply s0 if the first column length
is too long, then sort the result, and repeat as necessary. We start with λ† + ρ = (7, 7, 6, 4, 2, 1) +
(−3, −4, −5, . . . ):
s
s
s
s
0
1
(4, 3, 1, −2, −5, −7) −→
(−4, 3, 1, −2, −5, −7) −→
(3, −4, 1, −2, −5, −7)
2
3
−→
(3, 1, −4, −2, −5, −7) −→
(3, 1, −2, −4, −5, −7)
s
s
s
s
0
1
−→
(−3, 1, −2, −4, −5, −7) −→
(1, −3, −2, −4, −5, −7)
2
0
−→
(1, −2, −3, −4, −5, −7) −→
(−1, −2, −3, −4, −5, −7).
Subtracting ρ from the result, we get (6, 5)† .
4. Orthogonal groups
4.1. Representations of O(V ). Let (V, ω) be an orthogonal space of dimension m (here ω ∈
Sym2 V ∗ is the orthogonal form, and gives an isomorphism V ∼
= V ∗ ). We write m = 2n if it is
even, or m = 2n + 1 if it is odd. We now recall the representation theory of O(V ); see [FH, §19.5]
for details. The irreducible representations of O(V ) are indexed by partitions λ such that the first
two columns have at most m boxes in total, i.e., λ†1 + λ†2 ≤ m. We call such partitions admissible.
For an admissible partition λ, we write S[λ] (V ) for the corresponding irreducible representation of
O(V ).
Given an admissible partition λ, we let λσ be the partition obtained by changing the number
of boxes in the first column of λ to m minus its present value; that is, (λσ )†1 = m − λ†1 . We call
λσ the conjugate of λ. Conjugation defines an involution on the set of admissible partitions. On
irreducible representations, conjugating the partition corresponds to twisting by the sign character:
S[λσ ] (V ) = S[λ] (V ) ⊗ sgn. It follows that S[λ] (V ) and S[λσ ] (V ) are isomorphic when restricted
to SO(V ). In fact, these restrictions remain irreducible, unless λ = λσ (which is equivalent to
ℓ(λ) = n and m = 2n), in which case S[λ] (V ) decomposes as a sum of two non-isomorphic irreducible
representations.
For an admissible partition λ, exactly one element of the set {λ, λσ } has at most n boxes in its
first column. We denote this element by λ. Thus λ = λ if λ†1 ≤ n, and λ = λσ otherwise.
4.2. The Littlewood complex. Let E be a vector space. Put U = Sym2 (E), A = Sym(U ) and
B = Sym(E ⊗ V ). Consider the inclusion U ⊂ B given by
Sym2 (E) ⊂ Sym2 (E) ⊗ Sym2 (V ) ⊂ Sym2 (E ⊗ V ),
where the first inclusion is multiplication with ω. This inclusion defines an algebra homomorphism
A → B. Put C = B ⊗A C; this is the quotient of B by the ideal generated by U . We have maps
Spec(C) → Spec(B) → Spec(A).
We have a natural identification of Spec(B) with the space Hom(E, V ) of linear map ϕ : E → V
and of Spec(A) with the space Sym2 (E)∗ of symmetric forms on E. The map Spec(B) → Spec(A)
takes a linear map ϕ to the form ϕ∗ (ω). The space Spec(C), which we call that Littlewood
variety, is the scheme-theoretic fiber of this map above 0, i.e., is consists of those maps ϕ such
that ϕ∗ (ω) = 0. In other words, Spec(C) consists of maps ϕ : E → V such that the image of ϕ is
an isotropic subspace
V of V .
Let K• = B ⊗ • U be the Koszul complex of the Littlewood variety. We can decompose this
complex under the action of GL(E):
M
K• (E) =
Sλ (E) ⊗ Lλ• .
ℓ(λ)≤dim E
HOMOLOGY OF LITTLEWOOD COMPLEXES
15
The complex Lλ• is the Littlewood complex, and is independent of E (so long as dim E ≥ ℓ(λ)).
By [How, Proposition 3.6.3], its zeroth homology is
(
S[λ] (V ) if λ is admissible
(4.1)
H0 (Lλ• ) =
0
otherwise.
By Lemma 2.7, we have Hi (K• ) = TorA
i (B, C), and so we have a decomposition
M
(4.2)
TorA
Sλ (E) ⊗ Hi (Lλ• ).
i (B, C) =
ℓ(λ)≤dim E
Applied to i = 0, we obtain
(4.3)
C=
M
Sλ (E) ⊗ S[λ] (V ).
admissible λ
4.3. A special case of the main theorem. Our main theorem computes the homology of the
complex Lλ• . We now formulate and prove the theorem in a particularly simple case. We mention
this here only because it is worthwhile to know; the argument is not needed to prove the main
theorem.
Proposition 4.1. Suppose λ is admissible. Then
(
S[λ] (V )
λ
Hi (L• ) =
0
if i = 0
otherwise.
Proof. Choose E to be of dimension n. By Lemma 4.2 below, K• (E) has no higher homology. It
follows that Lλ• does not either. The computation of H0 (Lλ• ) is given in (4.1).
Lemma 4.2. Suppose dim E ≤ n. Then U ⊂ B is spanned by a regular sequence.
Proof. The proof is the same as Lemma 3.3. The only difference worth pointing out (but which
does not affect the proof) is that when dim V = 2n and dim E = n, the Grassmannian of isotropic
n-dimensional subspaces of V has two connected components, and the variety cut out by U has
two irreducible components.
4.4. The modification rule. As in the symplectic case, we now associate to a partition λ two
quantities im (λ) and τm (λ). We again give two equivalent definitions.
We begin with the Weyl group definition, following [Wen, §1.4]. Let s0 be the automorphism of
the set U which negates and swaps the first and second entries, and let W be the group generated
by the si with i ≥ 0. This is a Coxeter group of type D∞ . Let ℓ : W → Z≥0 be the length function,
which is defined just as in (2.1). Note that this group W , as a subgroup of Aut(U ), is equal to the
one from §3.4, but that the length function is different since we are using a different set of simple
reflections. Let ρ = (−m/2, −m/2 − 1, . . .). Define a new action of W on U by w • λ = w(λ + ρ) − ρ.
On S this agrees with the one defined in §2.2. The action of s0 is given by
s0 • (a1 , a2 , a3 , . . .) = (m + 1 − a2 , m + 1 − a1 , a3 , . . .).
The definitions of im (λ) and τm (λ) are now exactly as in the first half of §3.4.
We now give the border strip definition. This is motivated by [Sun, §5] (which is based on [Kin]),
but [Sun] only focuses on the special orthogonal group, so we have to modify the definition to get
the correct answer for the full orthogonal group. This is the same as the one given in §3.4, except
for three differences:
(D1) the border strip Rλ has length 2ℓ(λ) − m,
(D2) in the definition of im (λ), we use c(Rλ ) − 1 instead of c(Rλ ), and
(D3) if the total number of border strips removed is odd, then replace the end result µ with µσ .
16
STEVEN V SAM, ANDREW SNOWDEN, AND JERZY WEYMAN
One can stop applying the modification rule either when λ becomes admissible or when ℓ(λ) ≤ n;
the resulting values of τ and i are the same. For instance, if λ is admissible but ℓ(λ) > n then one
can stop immediately with i = 0 and τ = λ. Instead, one could remove a border strip. This border
strip occupies only the first column and when removed yields λσ . Thus i = 0 and by (D3), since
we removed an odd number of strips, τ = (λσ )σ = λ.
Proposition 4.3. The above two definitions agree.
Proof. Suppose that we are removing a border strip R1 of size 2ℓ(λ) − m from λ which begins at
the first box in the final row of λ. Let c1 = c(R1 ) be the number of columns of R1 . The first two
column lengths of λ \ R1 are (λ†2 − 1, λ†3 − 1). We have two cases depending on which of the two
quantities λ†2 + λ†3 − 2 and m is bigger.
First suppose that λ†2 +λ†3 −2 > m. Then we remove another border strip R2 of size 2(λ†2 −1)−m
from λ \ R1 which begins at the first box in the final row. Let c2 = c(R2 ) be the number of columns
of R2 . The sequence
(sc2 −1 sc2 −2 · · · s2 s1 sc1 −1 sc1 −2 · · · s3 s2 s0 ) • λ†
gives the column lengths of (λ \ R1 ) \ R2 .
Now suppose that λ†2 + λ†3 − 2 ≤ m. Then we have only removed 1 border strip, which is an odd
number, so we have to replace λ \ R1 with (λ \ R1 )σ according to (D3) above. In this case, the
sequence
(sc1 −1 sc1−2 · · · s3 s2 s0 ) • λ†
gives the column lengths of (λ \ R1 )σ .
Conversely, if we use the Weyl group modification rule with w ∈ W , then the expression (2.1)
for w must begin with s0 : if we apply any si with i > 0, then we increase the number of inversions
of the sequence, so if we write wsi = v, then ℓ(v) = ℓ(w) + 1 [Hum, §5.4, Theorem], so the resulting
expression for w will not be minimal. If we choose i maximal so that w = w′ si−1 · · · s3 s2 s0 with
ℓ(w) = ℓ(w′ ) + i − 1, then we have replaced the first two columns of λ with (m + 1 − λ†2 , m + 1 − λ†1 )
and then moved the column of length m + 1 − λ†1 over to the right as much as possible (adding 1
to it each time we pass a column and subtracting 1 from the column we just passed) so that the
resulting shape (minus the first column) is again a Young diagram.
Now there are two possibilities: if the whole shape is a Young diagram, then it is the result of
first removing a border strip of length 2ℓ(λ) − m with i columns, and then replacing the resulting
µ with µc . Otherwise, the first column length of the resulting shape is less than the second column
length. If we choose j maximal so that w′ = w′′ sj−1 · · · s2 s1 with ℓ(w′ ) = ℓ(w′′ ) + j − 1, then we
have moved the first column over to the right as much as possible (adding 1 to it each time we
pass a column and subtracting 1 from the column we just passed) so that the resulting shape is
again a Young diagram. In this case, then we have removed two border strips of size 2ℓ(λ) − m and
2(λ† − 1) − m with i and j columns, respectively.
Finally, there is a third modification rule, defined in [KT, §2.4]. As in the symplectic case,
we will not need to know the statement of the rule, but we will cite some results from [KT]. The
equivalence of the rule from [KT, §2.4] with the border strip rule comes from the fact that both rules
were derived from the same determinantal formulas (see [KT, Theorem 1.3.2] and [Kin, Footnote
17]). We remark that both rules are only stated for the special orthogonal group, but this will be
enough for our purposes.
As a matter of notation, we write τ m (λ) in place of τm (λ).
4.5. The main theorem. Our main theorem is exactly the same as in the symplectic case:
HOMOLOGY OF LITTLEWOOD COMPLEXES
17
Theorem 4.4. For a partition λ and an integer i we have
(
S[τm (λ)] (V ) if i = im (λ)
Hi (Lλ• ) =
0
otherwise.
In particular, if im (λ) = ∞ then Lλ• is exact.
Remark 4.5. Consider the coordinate ring R of rank ≤ m symmetric matrices; identifying E with
its dual, this is the quotient of A by the ideal generated by (m + 1)× (m + 1) minors. The resolution
of R over A is known, see [Wey, §6.3] or [JPW, §3]. On the other hand, R is the O(V )-invariant part
of B, and so the above theorem, combined with (4.2), shows that Sλ (E) appears in its resolution if
and only if τm (λ) = ∅. Thus the modification rule gives an alternative description of the resolution
of R. It is a pleasant combinatorial exercise to show directly that these two descriptions agree. As
in the symplectic case, the description in terms of the modification rule is more complicated, but
has the advantage of generalizing to our situation.
We separate the proof of the theorem into two cases, according to whether m is even or odd. In
each case, we follow the three-step plan from §1.7.
4.6. The even case. Throughout this section, m = 2n is the dimension of the space V .
Step a. Let Q1 be the set of partitions λ whose Frobenius coordinates (a1 , . . . , ar |b1 , . . . , br ) satisfy
ai = bi + 1 for each i. This set admits an inductive definition, as follows. The empty partition
belongs to Q1 . A non-empty partition µ belongs to Q1 if and only if the number of columns in µ is
one more than the number of rows, i.e., ℓ(µ) = µ1 − 1, and the partition obtained by deleting the
first row and column of µ, i.e., (µ2 − 1, . . . , µℓ(µ) − 1), belongs to Q1 . The significance of this set is
the plethysm
M
V•
Sµ (E)
(Sym2 (E)) =
µ∈Q1
(see [Mac, I.A.7, Ex. 5]).
Let λ be a partition with ℓ(λ) ≤ n. We write (λ|µ) in place of (λ |n µ). Define
S 1 (λ) = {µ ∈ Q1 such that (λ|µ) is regular}
S 2 (λ) = {partitions α such that τ m (α) = λ}.
Lemma 4.6. Let µ be a non-zero partition in S 1 (λ) and let ν be the partition obtained by removing
the first row and column of µ. Then ν also belongs to S 1 (λ). Furthermore, let w (resp. w′ ) be the
element of W such that α = w • (λ|µ) (resp. β = w′ • (λ|ν)) is a partition. Then Rα is defined and
we have the following identities:
|Rα | = 2µ1 − 2,
α \ Rα = β,
c(Rα ) = µ1 + ℓ(w′ ) − ℓ(w).
Proof. Suppose that in applying Bott’s algorithm to (λ|µ) the number µ1 moves r places to the
left. Thus, after the first r steps of the algorithm, we reach the sequence
(λ1 , . . . , λn−r , µ1 − r, λn−r+1 , . . . , λn + 1, µ2 , . . . , µℓ(µ) ).
As before, Bott’s algorithm on this sequence runs just like the algorithm on (λ|ν), and so (λ|ν)
is regular and ν belongs to S 1 (λ). Suppose the algorithm on (λ|ν) terminates after N = ℓ(w′ )
steps. Then the algorithm on (λ|µ) terminates after N + r = ℓ(w) steps, and we have the following
formula for α:
1≤i≤n−r
λi
αi = µ1 − r
i=n−r+1
βi−1 + 1 n − r + 2 ≤ i ≤ n + µ1 − 1
18
STEVEN V SAM, ANDREW SNOWDEN, AND JERZY WEYMAN
Since ℓ(α) = n + µ1 − 1, the border strip Rα has 2µ1 − 2 boxes. Using Remark 3.4, we see that
Rα exists since the box in the (n − r + 1)th row and the first column has a hook of size 2ℓ(α) − m.
Furthermore, α \ Rα = β and c(Rα ) = µ1 − r. Since r = ℓ(w) − ℓ(w′ ), we are done.
Proposition 4.7. There is a unique bijection S 1 (λ) → S 2 (λ) under which µ maps to α if there
exists w ∈ S such that w • (λ|µ) = α; in this case, ℓ(w) + im (α) = 12 |µ| and
(
λ
if rank(µ) is even
τm (α) =
σ
λ
if rank(µ) is odd.
Proof. Except for the computation of τm (α), the proof is exactly like that of Proposition 3.10. In
the proof of Lemma 4.6, we see that the number of border strips removed from α is rank(µ), so the
determination of τm (α) follows from (D3) in §4.4.
Step b. Let E be a vector space of dimension at least n. Let X be the Grassmannian of rank n
quotients of E. Let R and Q be the tautological bundles on X as in (2.2). Put ε = Sym2 (E) ⊗ OX ,
ξ = Sym2 (R) and define η by the exact sequence
0 → ξ → ε → η → 0.
Finally, for a partition λ with ℓ(λ) ≤ n, put Mλ = Sym(η) ⊗ Sλ (Q) and M λ = H0 (X, Mλ ). Note
that A = H0 (X, Sym(ε)), and so M λ is an A-module.
Proposition 4.8. We have
TorA
i (M λ , C) =
M
Sα (E),
α
where the sum is over partitions α with τ m (α) = λ and im (α) = i.
Proof. The proof is exactly like that of Proposition 3.13.
Step c. For an admissible partition λ put Bλ = HomO(V ) (S[λ] (V ), B). Note that Bλ is an A-module
and has a compatible action of GL(E). Let λ be a partition with at most n rows. If ℓ(λ) = n, put
B λ = Bλ ; otherwise, put B λ = Bλ ⊕ Bλσ . Note that the decomposition
M
B=
B λ ⊗ S[λ] (V )
ℓ(λ)≤n
holds SO(V )-equivariantly.
Lemma 4.9. Let λ be a partition with ℓ(λ) ≤ n. The spaces B λ and M λ are isomorphic as
representations of GL(E) and have finite multiplicities.
L
Proof. Put M =
ℓ(λ)≤n S[λ] (V ) ⊗ M λ . It is enough to show that M and B are isomorphic as
representations of GL(E) × SO(V ) and have finite multiplicities. In fact, it is enough to show that
the Sθ (E) multiplicity spaces of M and B are isomorphic as representations of SO(V ) and have
finite multiplicities. This is what we do.
The Sθ (E) multiplicity space of B is Sθ (V ). The decomposition of this in the representation ring
of SO(V ) can be computed by applying the specialization homomorphism to [KT, Thm. 2.3.1(2)].
The result is
X
(−1)im (ν) cθ2µ,ν [S[τ m (ν)] (V )].
µ,ν
For a fixed θ there are only finitely many values for µ and ν which make the Littlewood–Richardson
coefficient non-zero, which establishes finiteness of the multiplicities. Now, we have an equality
X
(−1)i [TorA
[M λ ] = [A]
i (M λ , C)]
i≥0
HOMOLOGY OF LITTLEWOOD COMPLEXES
19
in the representation ring of GL(E). Applying Proposition 4.8, we find
X
X
(−1)i [TorA
(−1)im (ν) [Sν (E)].
i (M λ , C)] =
i≥0
As [A] =
P
µ [S2µ (E)]
τ m (ν)=λ
(see [Mac, I.A.7, Ex. 1]), we obtain
X
[M λ ] =
(−1)im (ν) cθ2µ,ν [Sθ (E)].
µ,ν
τ m (ν)=λ
We therefore find that the Sθ (E) multiplicity space of M is given by
X
(−1)im (ν) cθ2µ,ν [S[λ] (V )].
λ,µ,ν
τm (ν)=λ
The result now follows.
Proposition 4.10. Let λ be a partition with ℓ(λ) ≤ n. We have an isomorphism M λ → B λ which
is A-linear and GL(E)-equivariant.
Proof. Suppose first that ℓ(λ) = n. Let µ be the partition given by µ†1 = 2n + 1 − λ†2 , µ†2 = λ†1 + 1 =
n + 1 and µ†i = λ†i for i > 2. Proposition 4.8 provides the following presentation for M λ :
Sµ (E) ⊗ A → Sλ (E) ⊗ A → M λ → 0.
Note that Sµ (E) occurs with multiplicity one in Sλ (E) ⊗ A by the Littlewood–Richardson rule,
and thus does not occur in M λ ; it therefore does not occur in B λ either, since M λ and B λ are
isomorphic as representations of GL(E).
A
Since TorA
0 (B, C) = C, we see from (4.3) that Tor0 (B λ , C) = Sλ (E). It follows that we have
a surjection f : A ⊗ Sλ (E) → B λ . Since Sµ (E) does not occur in B λ , we see that f induces a
surjection M λ → B λ . Since the two are isomorphic as GL(E) representations and have finite
multiplicities, this surjection is an isomorphism.
Now consider the case where ℓ(λ) < n. Define µ as above. Define ν using the same recipe as for
µ but applied to λσ ; thus νi† = µ†i for i 6= 2 and ν2† = 2n − λ†1 + 1. Proposition 4.8 provides the
following presentation for M λ :
(Sµ (E) ⊗ A) ⊕ (Sν (E) ⊗ A) → (Sλ (E) ⊗ A) ⊕ (Sλσ (E) ⊗ A) → M λ → 0.
Each of Sµ (E) and Sν (E) occur with multiplicity one in the middle module by the Littlewood–
Richardson rule, and thus neither occurs in M λ ; therefore neither occurs in B λ either.
As B λ = Bλ ⊕ Bλσ , we see from (4.3) that TorA
0 (B λ , C) = Sλ (E) ⊕ Sλσ (E). We therefore have
a surjection
f : (A ⊗ Sλ (E)) ⊕ (A ⊗ Sλσ (E)) → B λ .
Since neither Sµ (E) nor Sν (E) occurs in B λ , we see that f induces a surjection M λ → B λ . Since
these spaces are isomorphic as GL(E) representations and have finite multiplicities, this surjection
is an isomorphism.
Remark 4.11. In the second case in the above proof, Sµ (E) does not occur in Sλ (E) ⊗ A and
Sν (E) does not occur in Sλσ (E) ⊗ A. It follows that the presentation of M λ is a direct sum, and
so we have a decomposition M λ = Mλ ⊕ Mλσ . The argument in the proof shows that Mλ = Bλ
and Mλσ = Bλσ . It would be interesting if the modules Mλ and Mλσ could be constructed more
directly.
Combining the above proposition with Proposition 4.8, we obtain the following corollary.
20
STEVEN V SAM, ANDREW SNOWDEN, AND JERZY WEYMAN
Corollary 4.12. We have
TorA
i (B, C) =
M
Sλ (E) ⊗ S[τ m (λ)] (V )
im (λ)=i
as GL(E) × SO(V ) representations.
Combining this with (4.2) shows that
Hi (Lλ• ) =
(
S[τ m (λ)] (V )
0
if i = im (λ)
otherwise
as representations of SO(V ). We thus see that the O(V )-module Him (λ) (Lλ• ) is isomorphic to
S[τm (λ)] (V ) when restricted to SO(V ), and is therefore either isomorphic to S[τm (λ)] (V ) or S[τm (λ)σ ] (V ).
In fact, it is isomorphic to S[τm (λ)] (V ) by [Wen, Theorem 1.9]. This finishes the proof of the main
result.
4.7. The odd case. Throughout this section, m = 2n + 1 is the dimension of the space V .
Step a. Let Q0 be the set of partitions λ whose Frobenius coordinates (a1 , . . . , ar |b1 , . . . , br ) satisfy
ai = bi . Equivalently, Q0 is the set of partitions λ such that λ = λ† . This set admits an inductive
definition, as follows. The empty partition belongs to Q0 . A non-empty partition λ belongs to Q0
if and only if the number of rows and columns of λ are equal, i.e., ℓ(λ) = λ1 , and the partition
obtained by deleting the first row and column from λ belongs to Q0 .
Let λ be a partition with ℓ(λ) ≤ n. We write (λ|µ) in place of (λ |n µ). Define
S 1 (λ) = {µ ∈ Q0 such that (λ|µ) is regular}
S 2 (λ) = {partitions α such that τ m (α) = λ}.
Lemma 4.13. Let µ be a non-zero partition in S1 (λ) and let ν be the partition obtained by removing
the first row and column of µ. Then ν also belongs to S1 (λ). Furthermore, let w (resp. w′ ) be the
element of W such that α = w • (λ|µ) (resp. β = w′ • (λ|ν)) is a partition. Then Rα is defined and
we have the following identities:
|Rα | = 2µ1 + 1,
α \ Rα = β,
c(Rα ) = µ1 + ℓ(w′ ) − ℓ(w).
Proof. Suppose that in applying Bott’s algorithm to (λ|µ) the number µ1 moves r places to the
left. Thus, after the first r steps of the algorithm, we reach the sequence
(λ1 , . . . , λn−r , µ1 − r, λn−r+1 , . . . , λn + 1, µ2 , . . . , µℓ(µ) ).
As before, Bott’s algorithm on this sequence runs just like the algorithm on (λ|ν), and so (λ|ν)
is regular and ν belongs to S 1 (λ). Suppose the algorithm on (λ|ν) terminates after N = ℓ(w′ )
steps. Then the algorithm on (λ|µ) terminates after N + r = ℓ(w) steps, and we have the following
formula for α:
1≤i≤n−r
λi
αi = µ1 − r
i = n−r+1
βi−1 + 1 n − r + 2 ≤ i ≤ n + µ1
Since ℓ(α) = n + µ1 , the border strip Rα has 2µ1 + 1 boxes. Using Remark 3.4, we see that Rα
exists since the box in the (n − r + 1)th row and the first column has a hook of size 2ℓ(α) − m.
Furthermore, α \ Rα = β and c(Rα ) = µ1 − r. Since r = ℓ(w) − ℓ(w′ ), we are done.
Proposition 4.14. There is a unique bijection S 1 (λ) → S 2 (λ) under which µ maps to α if there
exists w ∈ S such that w • (λ|µ) = α; in this case, ℓ(w) + im (α) = 21 (|µ| − rank(µ)) and
(
λ
if rank(µ) is even
τm (α) =
σ
λ
if rank(µ) is odd.
HOMOLOGY OF LITTLEWOOD COMPLEXES
21
Proof. The proof is exactly the same as for Proposition 4.7.
Step b. Let E be a vector space of dimension at least n + 1. Let X ′ be the partial flag variety of
quotients of E of ranks n + 1 and n. Thus on X ′ we have vector bundles Qn+1 and Qn of ranks
n + 1 and n, and surjections E ⊗ OX ′ → Qn+1 → Qn . Let Rn+1 be the kernel of E ⊗ OX ′ → Qn+1
and let Rn be the kernel of E ⊗ OX ′ → Qn . Put L = Rn /Rn+1 . Let X be the Grassmannian of
rank n quotients of E, let π : X ′ → X be the natural map and let R and Q be the usual bundles
on X. Then Rn = π ∗ (R) and Qn = π ∗ (Q). The space X ′ is naturally identified with P(R), with
L being the universal rank one quotient of R.
Let ξ be the kernel of Sym2 (Rn ) → Sym2 (L) = L⊗2 , let ε = Sym2 (E) ⊗ OX ′ , which contains ξ
as a subbundle, and define η by the exact sequence
0 → ξ → ε → η → 0.
Let λ be an admissible partition and let a = a(λ) be 0 if ℓ(λ) ≤ n and 1 otherwise. Put
Mλ = Sym(η) ⊗ Sλ (Qn ) ⊗ L⊗a
and Mλ = H0 (X ′ , Mλ ). Note that H0 (X ′ , ε) = A, and so Mλ is an A-module. Finally, put
M
V
Wiλ =
Hj (X ′ , Sλ (Qn ) ⊗ L⊗a ⊗ i+j (ξ)),
j∈Z
Wiλ
=
M
Rj π∗ (Sλ (Qn ) ⊗ L⊗a ⊗
j∈Z
We wish to compute
Wiλ .
Lemma 4.15. We have
Wiλ =
M
Vi+j
(ξ)).
Sλ (Q) ⊗ Sµ (R),
µ
where the sum is over those partitions µ with µ = µ† , rank(µ) = a (mod 2) and i = 12 (|µ|−rank(µ)).
Proof. Since Qn = π ∗ (Q), and Schur functors commute with pullback, the projection formula gives
M
V
Wiλ =
Sλ (Q) ⊗ Rj π∗ (L⊗a ⊗ i+j (ξ)).
j∈Z
The result now follows from Proposition 2.4.
Lemma 4.16. Let µ ∈ S1 (λ) correspond to α ∈ S2 (λ). Then
(
Sα (E) if i = 21 (|µ| − rank(µ)) − im (α)
Hi (X, Sλ (Q) ⊗ Sµ (R)) =
0
otherwise.
Proof. This follows from Proposition 4.14 and the Borel–Weil–Bott theorem.
Lemma 4.17. We have
M
j∈Z
λ
Hj (X, Wi+j
)=
M
Sα (E),
α
where the sum is over those partitions α for which τm (α) = λ and im (α) = i.
Proof. This follows immediately from Lemmas 4.15 and 4.16, and Proposition 4.14.
V
Lemma 4.18. The pair (π, Sλ (Qn ) ⊗ L⊗a ⊗ • (ξ)) is degenerate (in the sense of §2.4).
Proof. We have
M
M
V
λ
Hi (X, Rj π∗ (Sλ (Qn ) ⊗ L⊗a ⊗ • (ξ))) =
Hj (X, Wi+j
)=
i,j
i,j
M
Sα (E).
τm (α)=λ
This is multiplicity-free as a representation of GL(E), so the criterion of Lemma 2.6 applies.
22
STEVEN V SAM, ANDREW SNOWDEN, AND JERZY WEYMAN
Lemma 4.19. We have
Wiλ =
M
Sα (E),
α
where the sum is over those partitions α for which τm (α) = λ and im (α) = i. In particular, Wiλ = 0
for i < 0.
Proof. By Lemma 4.18, we have an isomorphism of GL(E) representations
M
λ
Wiλ =
Hj (X, Wi+j
).
j∈Z
and so the result follows from Lemma 4.17
Proposition 4.20. We have
TorA
i (Mλ , C) =
M
Sα (E),
α
where the sum is over those partitions α for which τm (α) = λ and im (α) = i.
Proof. This follows immediately from the previous lemma and Proposition 2.1.
Step c. For an admissible partition λ, put Bλ = HomO(V ) (S[λ] (V ), B). Note that Bλ is an A-module
with a compatible action of GL(E).
Lemma 4.21. The spaces Bλ ⊕ Bλσ and Mλ ⊕ Mλσ are isomorphic as representations of GL(E)
and have finite multiplicities.
L
Proof. Let M =
λ Mλ ⊗ S[λ] (V ), where the sum is over admissible partitions λ. It is enough
to show that M and B are isomorphic as representations of GL(E) × SO(V ) and have finite
multiplicities. In fact, it is enough to show that the Sθ (E) multiplicity spaces of M and B are
isomorphic as representations of SO(V ) and have finite multiplicities. The proof goes exactly as
that of Lemma 4.9.
Proposition 4.22. Let λ be an admissible partition. We have an isomorphism Mλ → Bλ which
is A-linear and GL(E)-equivariant.
Proof. Arguing exactly as in the proof of Proposition 3.15 or 4.10, we obtain a surjection f : Mλ →
Bλ . Of course, we also have a surjection f ′ : Mλσ → Bλσ . By the previous lemma, f ⊕ f ′ is an
isomorphism, and so f and f ′ are isomorphisms.
Combining the above proposition with Proposition 4.20, we obtain the following corollary.
Corollary 4.23. We have
TorA
i (B, C) =
M
Sλ (E) ⊗ S[τm (λ)] (V )
im (λ)=i
as GL(E) × O(V ) representations.
4.8. Examples. We now give a few examples to illustrate the theorem.
Example 4.24. Suppose λ = (i). Then Lλ• is the complex Symi−2 V → Symi V , where the
differential is multiplication by the symmetric form on V ∗ treated as an element of Sym2 V .
• If m ≥ 2, or m = 1 and i ≤ 1, or m = 0 and i = 0 then the differential is injective, and
H0 (Lλ• ) = S[i] (V ) is a non-zero irreducible representation of V .
• If m = 1 and i ≥ 2, or m = 0 and i ≥ 3 or i = 1, the differential is an isomorphism and all
homology of Lλ• vanishes.
• If m = 0 and i = 2 then the differential is surjective and H1 (Lλ• ) = C is the trivial
representation of the trivial group O(0).
HOMOLOGY OF LITTLEWOOD COMPLEXES
23
Example 4.25. Suppose λ = (3, 1). Then Lλ• is the complex C → S(3,1)/(2) V → S(3,1) V , where
the differentials are multiplication by the symmetric form on V ∗ treated as an element of Sym2 V .
• If m ≥ 3 then the differential is injective, and H0 (Lλ• ) = S[3,1] (V ) is an irreducible representation of V .
• If m = 1, 2 then the complex is exact, and all homology of Lλ• vanishes.
• Finally when m = 0 then H2 (Lλ• ) = C.
The reader will check easily that in both instances the description of the homology agrees with
the rule given by the Weyl group action.
Example 4.26. Let us consider the same situation as in Example 3.20, i.e., λ = (6, 5, 4, 4, 3, 3, 2)
and m = 4. The modification rule, using border strips, proceeds as follows:
Starting with λ = λ0 we remove the border strip R0 of size 2ℓ(λ) − m = 10. Doing so we obtain
the partition λ1 = (6, 3, 3, 2, 2, 1). We now are supposed to remove the border strip R1 of size 8.
This border strip is shaded. However, upon removing this strip we do not have a Young diagram.
It follows that all homology of Lλ• vanishes.
In the Weyl group version, this amounts to
λ† + ρ = (7, 7, 6, 4, 2, 1) + (−2, −3, −4, . . . ) = (5, 4, 2, −1, −4, −6)
having a nontrivial stabilizer: if σ is the transposition that swaps the first and fifth entries, then
the stabilizer contains s0 σs0 , and this is a non-identity element.
Example 4.27. Suppose λ = (4, 4, 4, 4, 3, 3, 2) and m = 4. The border strip algorithm runs as
follows:
We have removed three border strips R0 , R1 , R2 . Thus according to rule (D3) of §4.4, τ4 (λ) is not
the final partition (2), but its conjugate, i.e., τ4 (λ) = (3, 1). We have
i4 (λ) = (c(R0 ) − 1) + (c(R1 ) − 1) + (c(R2 ) − 1) = 3 + 2 + 1 = 6.
We thus see that Hi (Lλ• ) = 0 if i 6= 6 and H6 (Lλ• ) = S[3,1] (V ).
si
Now we illustrate the modification rule using the Weyl group action. We write α −
→ β if
β = si (α). The idea for getting the Weyl group element is to apply s0 if sum of the first two
column lengths is too big, then sort the result, and repeat as necessary. We start with λ† + ρ =
(7, 7, 6, 4) + (−2, −3, −4, −5, . . . ):
s
s
s
s
s
s
0
2
(5, 4, 2, −1) −→
(−4, −5, 2, −1) −→
(−4, 2, −5, −1)
3
1
−→
(−4, 2, −1, −5) −→
(2, −4, −1, −5)
2
0
−→
(2, −1, −4, −5) −→
(1, −2, −4, −5).
Subtracting ρ from the result, we get (3, 1) = (2, 1, 1)† .
24
STEVEN V SAM, ANDREW SNOWDEN, AND JERZY WEYMAN
5. General linear groups
5.1. Representations of GL(V ). Let V be a vector space of dimension n. The irreducible representations of GL(V ) are indexed by pairs of partitions (λ, λ′ ) such that ℓ(λ) + ℓ(λ′ ) ≤ n (see
[Koi, §1] for more details). We call such pairs admissible. Given an admissible pair (λ, λ′ ), we
denote by S[λ,λ′ ] (V ) the corresponding irreducible representation of GL(V ). Identifying weights
of GL(V ) with elements of Zn , the representation S[λ,λ′ ] (V ) is the irreducible with highest weight
(λ1 , . . . , λr , 0, . . . , 0, −λ′s , . . . , −λ′1 ), where r = ℓ(λ) and s = ℓ(λ′ ). The representation S[λ,0] (V ) is
the usual Schur functor Sλ (V ), while the representation S[0,λ] is its dual Sλ (V )∗ .
5.2. The Littlewood complex. Let E and E ′ be vector spaces. Put U = E ⊗ E ′ , A = Sym(U )
and B = Sym((E ⊗ V ) ⊕ (E ′ ⊗ V ∗ )). Let U ⊂ B be the inclusion given by
E ⊗ E ′ ⊂ (E ⊗ V ) ⊗ (E ′ ⊗ V ∗ ) ⊂ Sym2 (E ⊗ V ⊕ E ′ ⊗ V ∗ ),
where the first inclusion is multiplication with the identity element of V ⊗ V ∗ . This inclusion
defines an algebra homomorphism A → B. Let C = B ⊗A C; this is the quotient of B by the ideal
generated by U . We have maps
Spec(C) → Spec(B) → Spec(A).
We have a natural identification of Spec(B) with the space Hom(E, V ∗ ) × Hom(E ′ , V ) of pairs of
maps (ϕ : E → V ∗ , ψ : E ′ → V ). The space Spec(A) is naturally identified with the space (E ⊗ E ′ )∗
of bilinear forms on E × E ′ . The map Spec(B) → Spec(A) takes a pair of maps (ϕ, ψ) to the
form (ϕ ⊗ ψ)∗ ω, where ω : V ⊗ V ∗ → C is the trace map. The space Spec(C), which we call the
Littlewood variety, is the scheme-theoretic fiber of this map above 0, i.e., it consists of those
pairs of maps (ϕ, ψ) such that (ϕ ⊗ ψ)∗ ω = 0.
Remark 5.1. One can modify the definitions of the rings A, B and C by replacing E with its dual
everywhere. The space Spec(A) is then identified with Hom(E ′ , E), while Spec(B) is identified
with the set of pairs of maps (ϕ : V → E, ψ : E ′ → V ). The map Spec(B) → Spec(A) takes (ϕ, ψ)
to ϕψ. The space Spec(C) consists of those pairs (ϕ, ψ) such that ϕψ = 0; thus Spec(C) is the
space of complexes of the form E ′ → V → E.
V
•
Let K• (E, E ′ ) = B ⊗ U be the Koszul complex of the Littlewood variety. We can decompose
this complex under the action of GL(E) × GL(E ′ ):
M
′
Sλ (E) ⊗ Sλ′ (E ′ ) ⊗ Lλ,λ
K• (E, E ′ ) =
• .
ℓ(λ)≤dim E
ℓ(λ′ )≤dim E ′
′
is the Littlewood complex, and is independent of E and E ′ (so long as
The complex Lλ,λ
•
dim E ≥ ℓ(λ) and dim E ′ ≥ ℓ(λ′ )). By [Bry, Theorem 3.3], its zeroth homology is
(
S[λ,λ′ ] (V ) if (λ, λ′ ) is admissible
λ,λ′
(5.1)
H0 (L• ) =
0
otherwise.
By Lemma 2.7, we have Hi (K• ) = TorA
i (B, C), and so we have a decomposition
M
′
Sλ (E) ⊗ Sλ′ (E ′ ) ⊗ Hi (Lλ,λ
(5.2)
TorA
• ).
i (B, C) =
ℓ(λ)≤dim E
ℓ(λ′ )≤dim E ′
Applied to i = 0, we obtain
(5.3)
C=
M
admissible (λ, λ′ )
Sλ (E) ⊗ Sλ′ (E ′ ) ⊗ S[λ,λ′ ] (V ).
HOMOLOGY OF LITTLEWOOD COMPLEXES
25
5.3. A special case of the main theorem. Our main theorem computes the homology of the
complex Lλ• . We now formulate and prove the theorem in a particularly simple case. We mention
this here only because it is worthwhile to know; the argument is not needed to prove the main
theorem.
Proposition 5.2. Suppose (λ, λ′ ) is admissible. Then
(
S[λ,λ′ ] (V ) if i = 0
λ,λ′
Hi (L• ) =
0
otherwise.
Proof. Choose E, E ′ so that dim E = ℓ(λ) and dim E ′ = ℓ(λ′ ). By Lemma 5.3 below, K• (E, E ′ ) has
′
′
does not either. The computation of H0 (Lλ,λ
no higher homology. It follows that Lλ,λ
• ) is given
•
in (5.1).
Lemma 5.3. Suppose dim E + dim E ′ ≤ n. Then U ⊂ B is spanned by a regular sequence.
Proof. It suffices to show that dim Spec(C) = dim Spec(B) − dim U . Put d = dim E and d′ =
dim E ′ . Observe that the locus of (ϕ, ψ) in Spec(C) where ϕ is injective and ψ is surjective is
open. Let Fl(d, d + d′ , V ) be the variety of partial flags Wd ⊂ Wd+d′ ⊂ V , where the subscript
indicates the dimension of the subspace. This variety comes with a tautological partial flag Rd ⊂
Rd+d′ ⊂ V ⊗ OFl(d,d+d′ ,V ) . There is a natural birational map from the total space of Hom(E, Rd ) ⊕
Hom(Rd+d′ /Rd , E ′ ) to Spec(C), and thus Spec(C) has dimension n(d+d′ )−dd′ . As dim Spec(B) =
n(d + d′ ) and dim U = dd′ , the result follows.
5.4. The modification rule. We now associate to a pair of partitions (λ, λ′ ) two quantities,
in (λ, λ′ ) and τn (λ, λ′ ). As usual, we give two equivalent definitions.
We begin with Koike’s definition of the Weyl group definition from the discussion preceding [Koi,
Prop. 2.2]. First, consider −(λ′† )op , which is the sequence obtained by taking λ′† and reversing
and negating its entries. We add n to each entry, and call the result σ(λ′ ). For example, if
λ′ = (3, 2, 2, 1, 1) and n = 4, then −(λ′† )op = (−1, −3, −5) and σ(λ′ ) = (3, 1, −1). Note that if
µ is any weakly decreasing sequence of nonnegative integers with n ≥ µ1 , then it makes sense to
reverse this procedure, and so we can define σ −1 (µ). Now set α = (σ(λ′ ) | λ† ) and ρ = (. . . , 3, 2, 1 |
0, −1, −2, . . . ). We let W ′ be the group of permutations on the index set Z of the coordinates of ρ
and α which differ from the identity permutation in only finitely many places. Given a permutation
w ∈ W ′ , we define w • α = w(α + ρ) − ρ as usual. One of two possibilities occurs:
• There exists a unique element w ∈ W ′ so that w • α = (β ′ | β) is weakly decreasing. In this
case, we set in (λ, λ′ ) = ℓ(w) and τn (λ, λ′ ) = (β, σ −1 (β ′ )).
• There exists a non-identity element w ∈ W ′ such that w • α = α. In this case, we put
in (λ, λ′ ) = ∞ and leave τn (λ, λ′ ) undefined.
We now give a modified (but equivalent) description of this Weyl group action. The group S × S
acts on the set U × U via the • action. We define a new involution t of U × U by
t • ((a1 , a2 , . . .), (b1 , b2 , . . .)) = ((n + 1 − b1 , a2 , . . .), (n + 1 − a1 , b2 , . . .))
Let W be the subgroup of Aut(U × U ) generated by S × S and t. Then W is isomorphic to an
infinite symmetric group, and comes equipped with a length function ℓ : W → Z≥0 with respect
to its set of generators, as in (2.1). Given a pair of partitions (λ, λ′ ) ∈ U × U , exactly one of the
following two possibilities hold:
• There exists a unique element w ∈ W such that w • (λ, λ′ )† = (µ, µ′ )† is a pair of partitions
and (µ, µ′ ) is admissible. We then put in (λ, λ′ ) = ℓ(w) and τn (λ, λ′ ) = (µ, µ′ ). (The
notation (λ, λ′ )† simply means (λ† , (λ′ )† ).)
• There exists a non-identity element w ∈ W such that w • (λ, λ′ )† = (λ, λ′ )† . We then put
in (λ, λ′ ) = ∞ and leave τn (λ, λ′ ) undefined.
26
STEVEN V SAM, ANDREW SNOWDEN, AND JERZY WEYMAN
As always, if (λ, λ′ ) is already admissible then we are in the first case, and in (λ, λ′ ) = 0 and
τn (λ, λ′ ) = (λ, λ′ ).
We now give the border strip definition, which we could not find in the literature. If (λ, λ′ ) is
admissible then we define in (λ, λ′ ) = 0 and τn (λ, λ′ ) = (λ, λ′ ). Assume now that (λ, λ′ ) is not
admissible. Let Rλ (resp. Rλ′ ) be the border strip of length ℓ(λ) + ℓ(λ′ ) − n − 1 starting in the first
box of the final row of λ (resp. λ′ ), if it exists. If both Rλ and Rλ′ exist and are non-empty and
both λ \ Rλ and λ′ \ Rλ′ are partitions, we put
in (λ, λ′ ) = c(Rλ ) + c(Rλ′ ) − 1 + in (λ \ Rλ , λ′ \ Rλ′ )
and τn (λ, λ′ ) = τn (λ \ Rλ , λ′ \ Rλ′ ). Otherwise, we put in (λ, λ′ ) = ∞ and leave τn (λ, λ′ ) undefined.
Proposition 5.4. The above two definitions agree.
Proof. Consider the diagram
µc
λ
µ
where µc is the complement of µ in a n × µ1 rectangle (we are allowing the possibility that ℓ(µ) > n,
in which case, we need to consider negative column lengths, but for the purposes of explanation,
we assume ℓ(µ) ≤ n). According to Koike’s rule, we are supposed to apply Bott’s algorithm to the
columns of the union of µc and λ.
We describe two procedures:
(1) Start with a shape that is a single column of length d (d could be negative) union a partition
λ with ℓ(λ) > d and apply Bott’s algorithm. Then we end up with a single column of length
ℓ(λ) − 1 union the shape obtained by removing a border strip of size ℓ(λ) − d − 1 if that is
possible, and 0 otherwise.
(2) Dually, start with a shape ν union a column of length e with e > νν†1 and apply Bott’s
algorithm. Then we get the shape η union a column of length νν†1 + 1 where η is obtained
from ν by adding a border strip of length e − νν†1 − 1 starting from the bottom box in the
last column of ν if it exists, and 0 otherwise.
Now go back to our shape, which is (µc , λ). Apply (1) with the column being the last column
of µc so that d = n − ℓ(µ). Then the last column of µc becomes ℓ(λ) − 1 and we have removed
a border strip of size ℓ(λ) + ℓ(µ) − n − 1 from λ. In the process, we used c(λ) simple reflections.
Now apply (2) with ν being µc minus its last column and e = ℓ(λ) − 1. The result is the shape
obtained by adding a border strip (starting from the right, not left) of length ℓ(λ) − νν†1 to ν union
a column of length νν†1 + 1. We can also describe this shape as follows: we added to µc a border
strip (starting from the right) of size ℓ(λ) + ℓ(µ) − n − 1. In this second step, we have used c(µ) − 1
simple reflections.
Finally, note that adding a border strip, starting from the right, to µc is equivalent to removing
a border strip from µ in the usual sense.
5.5. The main theorem. Our main theorem is the following:
Theorem 5.5. For a pair of partitions (λ, λ′ ) and an integer i, we have
(
S[τn (λ,λ′ )] (V ) if i = in (λ, λ′ )
′
Hi (Lλ,λ
)
=
•
0
otherwise.
HOMOLOGY OF LITTLEWOOD COMPLEXES
27
′
is exact.
In particular, if in (λ, λ′ ) = ∞ then Lλ,λ
•
Remark 5.6. Consider the coordinate ring R of rank ≤ n matrices; identifying E with its dual,
this is the quotient of A by the ideal generated by (n + 1) × (n + 1) minors. The resolution of R over
A is known, see [Wey, §6.1] or [Las]. On the other hand, R is the GL(V )-invariant part of B, and
so the above theorem, combined with (5.2), shows that Sλ (E) ⊗ Sλ′ (E ′ ) appears in its resolution
if and only if τn (λ, λ′ ) = ∅. Thus the modification rule gives an alternative description of the
resolution of R. It is a pleasant combinatorial exercise to show directly that these two descriptions
are equivalent. As in previous situations, the modification rule is more complicated, but has the
advantage that it readily generalizes to our situation.
The proof will take the rest of this section, and will follow the three-step plan given in §1.7.
Step a. Fix integers a and b with n = a + b. Let (λ, λ′ ) be a pair of partitions with ℓ(λ) ≤ a and
ℓ(λ′ ) ≤ b. We write (λ|µ) in place of (λ |a µ) and (λ′ |µ′ ) in place of (λ′ |b µ′ ). Define
S1 (λ, λ′ ) = {partitions µ such that (λ|µ) and (λ′ |µ† ) are regular}
S2 (λ, λ′ ) = {pairs of partitions (α, α′ ) such that τn (α, α′ ) = (λ, λ′ )}.
Lemma 5.7. Let µ be a non-zero partition in S1 (λ, λ′ ) and let ν be the partition obtained by
removing the first row and column of µ. Then ν also belongs to S1 (λ, λ′ ). Furthermore, let (w1 , w1′ )
(resp. (w2 , w2′ )) be elements of W such that α = w1 •(λ|µ) and α′ = w1′ •(λ|µ† ) (resp. β = w2 •(λ|ν)
and β ′ = w2′ • (λ|ν † )) are partitions. Then Rα and Rα′ are defined and we have the identities
|Rα | = |Rα′ | = ℓ(µ) + µ1 − 1,
and
α \ Rα = β,
α′ \ Rα′ = β ′
c(Rα ) + c(Rα′ ) = ℓ(µ) + µ1 + ℓ(w2 ) + ℓ(w2′ ) − ℓ(w1 ) − ℓ(w1′ ).
Proof. Reasoning as in the proof of Lemma 3.8, we see that ν belongs to S1 (λ, λ′ ). Suppose that
in applying Bott’s algorithm to (λ|µ) (resp. (λ′ |µ† )) the number µ1 (resp. µ†1 ) moves r (resp. s)
places to the left. Let N = ℓ(w2 ) (resp. N ′ = ℓ(w2′ )) be the number of steps in Bott’s algorithm
applied to (λ|ν) (resp. (λ|ν † )). Then, as in the proof of Lemma 3.8, we can trace Bott’s algorithm
on (λ|µ) and (λ′ |µ† ). We find
1≤i≤a−r
λi
αi = µ1 − r
i=a−r+1
βi−1 + 1 a − r + 1 ≤ i ≤ a + ℓ(µ)
′
1≤i≤b−s
λi
†
′
αi = µ1 − s
i = b−s+1
′
βi−1 + 1 b − s + 2 ≤ i ≤ b + µ1 .
We now examine the border strips used in the modification rule for (α, α′ ). Both Rα and Rα′ have
ℓ(α) + ℓ(α′ ) − n − 1 = ℓ(µ) + µ1 − 1 boxes. Using Remark 3.4, we see that Rα (resp. Rα′ ) exists
since the box in (a − r + 1)th (resp. (b − s + 1)th) row and the first column has a hook of size
ℓ(µ) + µ1 − 1. Furthermore, α \ Rα = β and α′ \ Rα′ = β ′ and
c(Rα ) = ℓ(µ) + µ1 − ℓ(α) + a − r,
c(Rα′ ) = ℓ(µ) + µ1 − ℓ(α′ ) + b − s,
which completes the proof.
(λ, λ′ )
(λ, λ′ )
Proposition 5.8. There is a unique bijection S1
→ S2
under which µ maps to (α, α′ )
if there exists w and w′ in S such that w • (λ|µ) = α and w′ • (λ′ |µ† ) = α′ ; in this case, ℓ(w) +
ℓ(w′ ) + in (α, α′ ) = |µ|.
28
STEVEN V SAM, ANDREW SNOWDEN, AND JERZY WEYMAN
Proof. The proof is essentially the same as that of Proposition 3.10.
Step b. Let E and E ′ be vector spaces of dimensions at least a and b, respectively. Let X (resp.
X ′ ) be the Grassmannian of rank a (resp. b) quotients of E (resp. E ′ ). Let R and Q (resp. R′
and Q′ ) be the tautological bundles on X (resp. X ′ ) as in (2.2); we regard all four as bundles on
X × X ′ . Put ε = E ⊗ E ′ ⊗ OX×X ′ , ξ = R ⊗ R′ and define η by the exact sequence
0 → ξ → ε → η → 0.
Finally, for a pair of partitions (λ, λ′ ) with ℓ(λ) ≤ a and ℓ(λ′ ) ≤ b, put
Mλ,λ′ = Sym(η) ⊗ Sλ (Q) ⊗ Sλ′ (Q′ )
and Mλ,λ′ = H0 (X × X ′ , Mλ,λ′ ). Note that A = H0 (X, Sym(ε)), and so Mλ,λ′ is an A-module.
Lemma 5.9. Let λ and λ′ be partitions with ℓ(λ) ≤ a and ℓ(λ′ ) ≤ b. Let ν ∈ S1 (λ, λ′ ) correspond
to (α, α′ ) in S2 (λ, λ′ ). Let V be the vector bundle on X × X ′ given by
V = Sλ (Q) ⊗ Sλ′ (Q′ ) ⊗ Sν (R) ⊗ Sν † (R′ ).
Then
(
Sα (E) ⊗ Sα′ (E ′ )
H (X × X , V) =
0
i
′
if i = |ν| − in (α, α′ )
otherwise.
Proof. The Künneth formula shows that
H• (X × X ′ , V) = H• (X, Sλ (Q) ⊗ Sν (R)) ⊗ H• (X ′ , Sλ′ (Q′ ) ⊗ Sν † (R′ )).
By definition, (λ|ν) and (λ′ |ν † ) are both regular, and so we can find w and w′ in S such that
α = w • (λ|ν) and α′ = w • (λ′ |ν † ) are partitions. By the Borel–Weil–Bott theorem, we get
(
Sα (E) if i = ℓ(w)
Hi (X, Sλ (Q) ⊗ Sν (R)) =
0
otherwise
(
Sα′ (E ′ ) if i = ℓ(w′ )
Hi (X ′ , Sλ (Q′ ) ⊗ Sν † (R′ )) =
0
otherwise.
By Proposition 5.8, we have ℓ(w) + ℓ(w′ ) = |ν| − in (α, α′ ), which completes the proof.
Lemma 5.10. Let λ and λ′ be partitions with ℓ(λ) ≤ a and ℓ(λ′ ) ≤ b and let i be an integer. We
have
M
M
V
Sα (E) ⊗ Sα′ (E ′ ),
Hj (X × X ′ , i+j (ξ) ⊗ Sλ (Q) ⊗ Sλ′ (Q′ )) =
(α,α′ )
j∈Z
where the sum is over pairs (α, α′ ) with in (α, α′ ) = i and τn (α, α′ ) = (λ, λ′ ). In particular, when
i < 0 the left side above vanishes.
Proof. We have
Vi+j
and so
M
j∈Z
=
M
(ξ) =
Vi+j
Hj (X × X ′ ,
(R ⊗ R′ ) =
Vi+j
M
Sν (R) ⊗ Sν † (R′ ),
|ν|=i+j
(ξ) ⊗ Sλ (Q) ⊗ Sλ′ (Q′ ))
H|ν|−i (X × X ′ , Sλ (Q) ⊗ Sλ′ (Q′ ) ⊗ Sν (R) ⊗ Sν † (R′ )).
ν
Of course, we only need to sum over ν ∈ S1 (λ, λ′ ). The result now follows from Lemma 5.9.
HOMOLOGY OF LITTLEWOOD COMPLEXES
29
Proposition 5.11. Let λ and λ′ be partitions with ℓ(λ) ≤ a and ℓ(λ′ ) ≤ b. We have
M
Sα (E) ⊗ Sα′ (E ′ ),
TorA
i (Mλ,λ′ , C) =
(α,α′ )
where the sum is over pairs (α, α′ ) with in (α, α′ ) = i and τn (α, α′ ) = (λ, λ′ ).
Proof. This follows immediately from the Lemma 5.10 and Proposition 2.1.
Lemma 5.12. The module Mλ,λ′ is independent of the choice of a and b, provided that ℓ(λ) ≤ a
and ℓ(λ′ ) ≤ b.
Proof. Applying Proposition 5.11 with i = 0 and i = 1 shows that we have a presentation
A ⊗ S(λ,1d ) (E) ⊗ S(λ′ ,1d ) (E ′ ) → A ⊗ Sλ (E) ⊗ Sλ′ (E ′ ) → Mλ,λ′ → 0,
where d = n + 1 − ℓ(λ) − ℓ(µ). The left map is unique up to scalar multiple, since S(λ,1d ) (E) ⊗
S(λ′ ,1d ) (E ′ ) occurs with multiplicity one in the middle group, and so the lemma follows.
We thus have a well-defined A-module Mλ,λ′ for any admissible pair (λ, λ′ ).
Step c. For an admissible pair (λ, λ′ ), put
Bλ,λ′ = HomGL(V ) (S[λ,λ′ ] (V ), B).
Note that Bλ,λ′ is an A-module and has a compatible action of GL(E) × GL(E ′ ). Our goal is to
show that Bλ,λ′ is isomorphic to Mλ,λ′ .
Lemma 5.13. Let (λ, λ′ ) be an admissible pair. Then the spaces Bλ,λ′ and Mλ,λ′ are isomorphic
as representations of GL(E) × GL(E ′ ), and have finite multiplicities.
Proof. As we have done before, put
M=
M
admissible
Mλ,λ′ ⊗ S[λ,λ′ ] (V ).
(λ, λ′ )
It is enough to show that M and B are isomorphic as representations of GL(E) × GL(E ′ ) × GL(V )
and have finite multiplicities. In fact, it is enough to show that the Sθ (E) ⊗ Sθ′ (E ′ ) multiplicity
spaces of M and B are isomorphic as representations of GL(V ) and have finite multiplicities. This
is what we do.
The Sθ (E) ⊗ Sθ′ (E) multiplicity space of B is Sθ (V ) ⊗ Sθ′ (V ∗ ). The decomposition of this in
the representation ring of GL(V ) can be computed using [Koi, Thm. 2.4]. The result is
X
′
′
(−1)in (ν,ν ) cθν,µ cθν ′ ,µ [S[τn (ν,ν ′ )] (V )].
ν,ν ′ ,µ
[ν,ν ′ ]
(In the notation of [Koi], we are computing M[θ,0],[0,θ′ ] . These zeros lead to the massive simplification
of the general formula given there.) Note that for fixed (θ, θ ′ ) there are only finitely many values for
(ν, ν ′ , µ) which make the product of Littlewood–Richardson coefficients non-zero, which establishes
finiteness of the multiplicities. Now, we have an equality
X
(−1)i [TorA
[Mλ,λ′ ] = [A]
i (Mλ,λ′ , C)]
i≥0
in the representation ring of GL(E). Applying Proposition 5.11, we find
X
X
′
′
,
C)]
=
(−1)i [TorA
(M
(−1)in (ν,ν ) [Sν (E)][Sν ′ (E ′ )].
λ,λ
i
i≥0
ν,ν ′
τn (ν,ν ′ )=(λ,λ′ )
30
STEVEN V SAM, ANDREW SNOWDEN, AND JERZY WEYMAN
As [A] =
P
µ [Sµ (E)][Sµ (E
′ )],
[Mλ,λ′ ] =
we obtain
X
′
′
(−1)in (ν,ν ) cθν,µ cθν ′ ,µ [Sθ (E)][Sθ′ (E ′ )].
µ,θ,θ ′ ,ν,ν ′
τn (ν,ν ′ )=(λ,λ′ )
We therefore find that the Sθ (E) ⊗ Sθ′ (E ′ ) component of M is given by
X
′
′
(−1)in (ν,ν ) cθν,µ cθν ′ ,µ [S[λ,λ′ ] (V )].
λ,λ′ ,µ,ν,ν ′
τn (ν,ν ′ )=(λ,λ′ )
The result now follows.
Proposition 5.14. Let (λ, λ′ ) be an admissible pair. Then we have an isomorphism Bλ,λ′ → Mλ,λ′
which is A-linear and GL(E) × GL(E ′ ) equivariant.
Proof. As in the proof of Lemma 5.12, we have a presentation
A ⊗ S(λ,1d ) (E) ⊗ S(λ′ ,1d ) (E ′ ) → A ⊗ Sλ (E) ⊗ Sλ′ (E ′ ) → Mλ,λ′ → 0,
where d = n + 1 − ℓ(λ) − ℓ(λ′ ). Note that S(λ,1d ) (E) ⊗ S(λ′ ,1d ) occurs with multiplicity one in the
middle module, and thus does not occur in Mλ,λ′ ; it therefore does not occur in Bλ,λ′ either, since
Mλ,λ′ and Bλ,λ′ are isomorphic as representations of GL(E) × GL(E ′ ).
A
′
Since TorA
0 (B, C) = C, we see from (5.3) that Tor0 (Bλ,λ′ , C) = Sλ (E) ⊗ Sλ′ (E ). It follows that
′
we have a surjection f : A ⊗ Sλ (E) ⊗ Sλ′ (E ) → Bλ,λ′ . Since S(λ,1d ) (E) ⊗ S(λ′ ,1d ) (E ′ ) does not occur
in Bλ,λ′ , we see that f induces a surjection Mλ,λ′ → Bλ,λ′ . Since the two are isomorphic as GL(E)×
GL(E ′ ) representations and have finite multiplicities, this surjection is an isomorphism.
Combining the above proposition with Proposition 5.11, we obtain the following corollary. Combined with (5.2), this proves the main theorem.
Corollary 5.15. We have
TorA
i (B, C) =
M
Sλ (E) ⊗ Sλ′ (E ′ ) ⊗ S[τn (λ,λ′ )] (V ).
in (λ,λ′ )=i
5.6. Examples. We now give a few examples to illustrate the theorem.
′
is the complex
Example 5.16. Suppose λ = (1i ), λ′ = (1j ) with i and j positive. Then Lλ,λ
•
Vj ∗
Vi
Vj−1 ∗
Vi−1
V → V ⊗ V ,
V ⊗
where the differential is the multiplication by the identity, treated as an element of V ⊗ V ∗ . Let
n = dim V .
′
• If i + j ≤ n then the differential is injective, and H0 (Lλ,λ
• ) = S[1i ,1j ] (V ) is the irreducible
i
n−i−j
j
representation with highest weight (1 , 0
, (−1) ).
′
vanishes.
• If i + j = n + 1 then the differential is an isomorphism, and all homology of Lλ,λ
•
′
)
=
S
(V
).
• If i + j > n + 1 but i ≤ n + 1 and j ≤ n + 1 we have H1 (Lλ,λ
n+1−j
n+1−i
•
[1
,1
]
′
vanishes identically.
• Finally, if i > n + 1 or j > n + 1 then the complex Lλ,λ
•
The reader will easily check that the description of homology given above agrees with the rule
given by the Weyl group action.
Example 5.17. Suppose λ = λ0 = (4, 3, 2, 2) and λ′ = λ′0 = (5, 2, 2, 1, 1) and n = 3. We are
supposed to remove border strips of size ℓ(λ0 ) + ℓ(λ′0 ) − n − 1 = 4 + 5 − 4 = 5 from each partition.
HOMOLOGY OF LITTLEWOOD COMPLEXES
31
Let R0 and R0′ be these border strips. The picture is as follows:
The partition λ0 is on the left, with R0 shaded, and λ′0 is on the right with R0′ shaded. Let λ1 = λ0 \
R0 and λ′1 = λ′0 \R0′ ; these are the unshaded boxes in the above diagrams. As ℓ(λ1 )+ℓ(λ′1 ) = 5 > n,
the pair (λ, λ′ ) is not admissible and the algorithm continues. The border strips R1 and R1′ have
size 1. The picture is thus:
Removing these border strips, we obtain the partitions λ2 = (4, 1) and λ′2 = (4). As ℓ(λ2 ) + ℓ(λ′2 ) ≤
n, the pair (λ, λ′ ) is admissible and the algorithm terminates. So τ3 (λ, λ′ ) = ((4, 1), (4)) and
i3 (λ, λ′ ) = (c(R0 ) + c(R0′ ) − 1) + (c(R1 ) + c(R1′ ) − 1) = 4 + 1 = 5.
′
′
λ,λ
It follows that Hi (Lλ,λ
• ) = 0 for i 6= 5, while H5 (L• ) = S[(4,1),(4)] (V ) is the irreducible of GL(3)
with highest weight (4, 1, −4).
Now we illustrate the modification rule using Koike’s original Weyl group action. Using the
notation of §5.4, we have σ(λ′ ) = (2, 2, 2, 0, −2), α = (2, 2, 2, 0, −2 | 4, 4, 2, 1), and ρ = (. . . , 2, 1 |
0, −1, . . . ). If we sort α + ρ = (7, 6, 5, 2, −1 | 4, 3, 0, −2), we get (7, 6, 5, 4, 3 | 2, 0, −1, −2). The
permutation that does this sorting has length 5 (we made 5 consecutive swaps), and subtracting
ρ, we get (2, 2, 2, 2, 2 | 2, 1, 1, 1). Then (2, 1, 1, 1)† = (4, 1) is our first partition, and our second
partition is σ −1 (25 ) = (15 )† = (5).
References
[AH]
Luchezar Avramov, Jürgen Herzog, The Koszul algebra of a codimension 2 embedding, Math. Z. 175 (1980),
no. 3, 249–260.
[Bry] Ranee Kathryn Brylinski, Matrix concomitants with the mixed tensor model, Adv. Math. 100 (1993), no. 1,
28–52.
[DPS] Elizabeth Dan-Cohen, Ivan Penkov, Vera Serganova, A Koszul category of representations of finitary Lie
algebras, arXiv:1105.3407v2.
[EW] Thomas J. Enright, Jeb F. Willenbring, Hilbert series, Howe duality and branching for classical groups, Ann.
of Math. (2) 159 (2004), no. 1, 337–375.
[FH]
William Fulton, Joe Harris, Representation Theory: A First Course, Graduate Texts in Mathematics 129,
Springer-Verlag, New York, 1991.
[How] Roger Howe, Perspectives on invariant theory: Schur duality, multiplicity-free actions and beyond, Israel
Mathematical Conference Proceedings 8, 1995.
[HTW] Roger Howe, Eng-Chye Tan, Jeb F. Willenbring, Stable branching rules for classical symmetric pairs, Trans.
Amer. Math. Soc. 357 (2005), no. 4, 1601–1626, arXiv:math/0311159v2.
[Hum] James E. Humphreys, Reflection Groups and Coxeter Groups, Cambridge Studies in Advanced Mathematics
29, Cambridge University Press, Cambridge, 1990.
[JPW] T. Józefiak, P. Pragacz, J. Weyman, Resolutions of determinantal varieties and tensor complexes associated
with symmetric and antisymmetric matrices, Young tableaux and Schur functors in algebra and geometry
(Toruń, 1980), pp. 109–189, Astérisque, 87-88, Soc. Math. France, Paris, 1981.
[Kin] R. C. King, Modification rules and products of irreducible representations of the unitary, orthogonal, and
symplectic groups, J. Mathematical Phys. 12 (1971), 1588–1598.
[Koi] Kazuhiko Koike, On the decomposition of tensor products of the representations of the classical groups: by
means of the universal characters, Adv. Math. 74 (1989), no. 1, 57–86.
[KT]
Kazuhiko Koike, Itaru Terada, Young-diagrammatic methods for the representation theory of the classical
groups of type Bn , Cn , Dn , J. Algebra 107 (1987), no. 2, 466–511.
[Las]
Alain Lascoux, Syzygies des variétés déterminantales, Adv. in Math. 30 (1978), no. 3, 202–237.
32
STEVEN V SAM, ANDREW SNOWDEN, AND JERZY WEYMAN
[Lit]
[Mac]
[SS1]
[SS2]
[SW]
[Sun]
[Wen]
[Wey]
Dudley E. Littlewood, The Theory of Group Characters and Matrix Representations of Groups, reprint of
the second (1950) edition, AMS Chelsea Publishing, Providence, RI, 2006.
I. G. Macdonald, Symmetric Functions and Hall Polynomials, second edition, Oxford Mathematical Monographs, Oxford, 1995.
Steven V Sam, Andrew Snowden, GL-equivariant modules over polynomial rings in infinitely many variables,
arXiv:1206.2233v1.
Steven V Sam, Andrew Snowden, Stability patterns in representation theory, in preparation.
Steven V Sam, Jerzy Weyman, Koszul homology of codimension 3 Gorenstein ideals, Proc. Amer. Math. Soc.,
to appear, arXiv:1203.3168v1.
Sheila Sundaram, Tableaux in the representation theory of the classical Lie groups, Invariant theory and
tableaux (Minneapolis, MN, 1988), 191–225, IMA Vol. Math. Appl., 19, Springer, New York, 1990.
Hans Wenzl, Quotients of representation rings, Represent. Theory 15 (2011), 385–406, arXiv:1101.5887v1.
Jerzy Weyman, Cohomology of Vector Bundles and Syzygies, Cambridge University Press, Cambridge, 2003.
Department of Mathematics, University of California, Berkeley, CA
E-mail address: [email protected]
Department of Mathematics, MIT, Cambridge, MA
E-mail address: [email protected]
Department of Mathematics, Northeastern University, Boston, MA
E-mail address: [email protected]
| 0math.AC
|
Insulin Regimen ML-based control for T2DM patients
arXiv:1710.07855v1 [q-bio.QM] 21 Oct 2017
Mark Shifrin, Hava Siegelmann
October 24, 2017
Abstract
We model individual T2DM patient blood glucose level (BGL) by stochastic process with discrete
number of states mainly but not solely governed by medication regimen (e.g. insulin injections). BGL
states change otherwise according to various physiological triggers which render a stochastic, statistically
unknown, yet assumed to be quasi-stationary, nature of the process. In order to express incentive for being
in desired healthy BGL we heuristically define a reward function which returns positive values for desirable
BG levels and negative values for undesirable BG levels. The state space consists of sufficient number of
states in order to allow for memoryless assumption. This, in turn, allows to formulate Markov Decision
Process (MDP), with an objective to maximize the total reward, summarized over a long run. The
probability law is found by model-based reinforcement learning (RL) and the optimal insulin treatment
policy is retrieved from MDP solution.
1
Introduction
Diabetes mellitus type 2 (T2DM) makes up for about 90% of all Diabetes cases [15], with increasing rates
since as early as 1960 [9]. The goal of therapy for T2DM is to bring the average blood glucose (BG) as close
to the normal range as possible, which can be done by various oral medications or/and insulin injections [12].
The exact dosage and frequency of the medication intakes are coordinated by blood glucose measurements
normally being performed by designated glucose meters, see , e.g., [14]. The recent recommendations are
such that T2DM patients are advised to measure their blood glucose level (BGL) at least 3 times a day [8].
In this work, we address individually adopted insulin regimen control. For this purpose, we harness tools
from the area of Machine Learning (ML) and stochastic control theory. In particular, we build Markov
Decision Process of blood glucose level. The solution to the defined MDP is expressed by a policy which
assigns an action according to the measured BGL.
Note that we pose no contradiction to other known methods employed by medical care. The ML theory
is agnostic to physiological processes, e.g., biochemical insulin impacts e.g. renal, hepatic effects of insulin
resistance, effect of other possibly malfunctioning processes inside beta-cells, impairments of insulin pathways
and so on. We merely taking a different approach, modeling the insulin treatment as a controllable stochastic
process. In contrary to other known, including ML-based works, we exploit no mathematical or biological
models which express chemical dependencies.
One of the well-known difficulties related to insulin regimen control is associated with short and ultradian
insulin secretion peaks, which, at the recent time, are not well understood, [5]. The BGL is known to increase
after meal intake and to decrease during fasting periods and after physical activity. However, the peaks
render the BGL to vary according to the pattern which is hard to mathematically characterize.
In particular, patients with insulin resistance and especially those who had progressed to T2DM, have
abnormal insulin oscillations, no first phase and diminished/scattered 2nd phase of insulin secretion out of
1
beta-cells. Even more important, these oscillations may not exactly follow glucose oscillations, like it is known
to be in healthy people [5].
In order to characterize the BGL state of a patient, we assume continuous and bounded BG axes and
perform discretization by dividing the entire region into finite BG states. For example, between 70 and 300
mg/dL (see, e.g., [6]). Each state refers to continuous region, such that any BG in this region is assigned to
the corresponding state. The state can be augmented by additional individually measured parameters. In
particular, we may also account for glucose absorption rate (GAT) and HbA1c (glycosylated haemoglobin)
values.
The patient’s BGL, hence, is viewed as a stochastic process which varies according to physiological
dynamics and according to the patient activities. This process is the controllable subject such that the control
is applied by means of medications, e.g. insulin injections (IJ). The finesse of the discretization is set in a
way to assure that the next state will solely depend on the previous state and the action, i.e., IJ, if any, taken
in that state. While the BGL process is clearly continuously fluctuating through the time, the measurements
are performed in discrete time slots. Hence, we view the state of the patient at certain time marks. We are
interested in transition law, i.e. transition probabilities from one state, at some time mark t, to the next state
which the patient sees (i.e, feels) right after the next measurement which occurs at time mark t + 1. This law
is expressed by the transition probability from a state, associated with the corresponding BGL, given the set
of actions taken by the patient in that state, to the next state.
The aforementioned state transition probabilities are learned by a model-based reinforcement learning
(RL). The input data for the RL is taken from existing samples of measurements individually performed
under effect of medication applied by physician administration. We assume that such samples exist for each
individual patient. Alternatively, the individual can be assigned to a category of patients for which the
initial insulin policy is already defined. This policy will be improved henceforth and individually adjusted by
applying reinforcement learning on the individual’s measurements. In this document, we show results from
database open for purposes of academic usage [7].
The methods of ML were already addressed in [2] for the treatment of Type I diabetes Mellitis (T1DM).
The paper [2] proposes linearizion of MDP solution. The model is based on high dependence of insulin and
glucose levels. However, for T2DM this assumption is not necessarily true. T1DM treatment is facilitated
by very frequent (e.g., every several minutes) BG measurements and actions, while T2DM measurements
are normally performed several times a day. This makes T2DM control an appropriate candidate for the
stochastic modeling by discrete MDP with discrete actions control. The prediction of BGL was addressed
in [3] and references therein, giving a stronger emphasis to T1DM. Neural Network (NN) based model is
presented in [16], aimed to predict BGL in T2DM patients. The training was performed on other patients for
the purpose of the NN convergence. In this work, the crucial learning part is performed using individual data,
with objective to individually adjust to each patient. Authors of this model suggest that convergence of NN
by sample data from other patients may cause a bias and dependency on that specific data. (Note that we do
use data samples from various patients for research phases of the algorithm presented in this paper, while
adopting to the specific patient is fulfilled in the phase of on-line control.) Block-oriented Wiener modeling
for BG prediction is done in [13].
Previous works which apply ML in the context of Diabetes employ a great deal of system parameters,
which is especially helpful for the prediction. Our approach is control-oriented and is designed to completely
hide the complex impact of multiple parameters behind the stochastic process of the BGL, which we consider
as the main indication for the insulin control purpose. That is, we allow the assumption that BGLs’ stochastic
process embodies in itself the impact of all relevant physical parameters (including, in most extremely
simplified cases, carbohydrates intakes and physical activities), hence we rely on learning the statistical
properties of the process.
Using the states, actions and transition laws, we formulate Markov Decision Process (MDP), which aims
to maximize the average reward over time. For this purpose we heuristically define a mapping which transfers
2
the BGL state into reward, such that a positive reward is obtained for the healthiest states and negative
rewards are obtained at undesired BG levels. We also introduce non-linear components which aim to fine for
being in dangerous states, e.g., overly high BGL or hypoglycemia.
We formally define the MDP and describe the RL algorithm in the next section.
2
The probabilistic model
We start by defining the state space, next we define the reward function, the reward functional and the action
space.
2.1
2.1.1
Definition of spaces
The measurement space
We consider blood glucose interval of glmin to glmax , where the lower bound means an extremely low BGL
(possibly corresponding to hypoglycemia) while the upper bound means extremely high level of glucose in
blood. The BGL finesse coefficient (BFC) defines the size of the interval which is mapped to the single BG
state. For example for BF C = 1 and glmin = 70mg/dL, the lowest interval of [70, 71) corresponds to the
lowest state. The set of all possible BGL forms the space:
L = {1, · · · , L}, L =
glmax − glmin
.
BF C
Denote the BGL-state by gl, where gl ∈ L,
We allow to augment state space by additional measured parameters, e.g., glucose absorption rate (GAT),
Similarly to BGL, we perform discretization for GAT by defining the bounds of the GAT axes gamin to
gamax . The GAT finesse coefficient (GFC) is used in order to define the GAT-state:
G = {1, · · · , G}, G =
gamax − gamin
.
GF C
The measurement space M defines the set of all possible measurements of the patient. In the case both
BGL and GAT can be measured at all times we have M = L × G. In the simplest scenario where the only
component of the measurement space is BGL, we merely have M = L.
2.1.2
The daytime space
We differentiate the BGL process instantiations during the day at different measurement points. These points
are expressed by daytime space, denoted by
T = {τ1 , · · · , τT },
where T is the number of measurements per day. For simplicity we assume this number is constant.
Example 2.1. Consider approximately constant measurement timings, corresponding to morning, noon,
afternoon, evening, night.
Then, T = 5 and {τ1 , τ2 , τ3 , τ4 , τ5 } = {morning, noon, afternoon, evening, night}.
The detailed time is denoted by a couple {t, τ }. In this sense t ∈ {0, 1, · · · } represents the date, while τ ∈ T
represents the time during the day, corresponding to one of the measurements. The absolute measurement
points can be counted by denoting ι = t · T + τ . At any ι we consider ms(ι) ∈ M.
3
2.1.3
Carbohydrates intake history space
The carbohydrates finesse coefficient (CFC) is used in order to define the quantity of carbohydrates consumed
at some time mark by a patient:
C = {1, · · · , C}, C =
chmax − chmin
.
CF C
In order to wisely incorporate the Carbohydrates intake history (CIH) we will assume that at each ι last H
intakes of carbohydrates are relevant for the insulin treatment. That is, at any measurement point ι we care of
all measurement in the range [ι, ι − H], while measurements of older points are presumed of no longer having
impact on the individual. The CIH at time ι is denoted by CH(ι) = {chι }, where chi ∈ C, ∀i ∈ [ι, ι − H].
See that in example 2.1, for H = 3, at time {t, noon} we have CH({t, noon}) = {ch{t, noon}, ch{t, morning}, ch{t−
1, night}, ch{t−1, evening}}, while CH({t, afternoon}) = {ch{t, af ternoon}, ch{t, noon}, ch{t, morning}, ch{t−
1, night}}.
Such a structure allows for impact of the past carbohydrate intakes on one hand and preserves memoryless
property on th other hand. That is, CH(ι) depends only on CH(ι − 1) and ch(ι).
2.1.4
Physical activity history space
We assume physical activities have a prolonged effect, similarly to CIH. The measurements of energy spent
for a given activity can be done in metabolic equivalents, or METs, see, e.g., the table of translation of
activity to METs in [4]. The physical activity finesse coefficient (PFC) is used in order to define the number
of METs corresponding to the energy spent by an individual at some measurement (or in a period between
two consecutive measurements):
PH = {1, · · · , P H}, C =
phmax − phmin
.
PFC
To define the Physical activity history (PAH), denote at time ι the set P H(ι) = {phι }, where phi ∈ PH, ∀i ∈
[ι, ι − Y ], and Y is the size of the relevant history.
2.1.5
Patient activity space
We define the patient activity (PA) space by
PA = CH × PH
At any ι we consider pa(ι) ∈ PA. We reason we differ between PA space and the measurement space is
summarized as follows
• ph is directly known by a patient and can be foreseen by conducted activities. In contrary, ms is
assumed to be purely stochastic and can be only measured.
• The insulin medication is applied in order to control the ms, i.e., the BGL, while pa may only effect
the dosage.
• We aim to apply optimal control for given ms and ,optionally, for a given pa. The pa can be seen as
the secondary means of control, however the maintaining certain level of activities is not a part of the
insulin policy (see definition below).
4
2.1.6
Overall state space definition
We now ready to define the state space and the state. Denote
S = PA × M × T
Denote by sι ∈ S, a state at absolute measurement point ι. That is, the state consists of BGL and other
measurements according to definition of M, patient activities history according to definition of PA and
daytime of ι, according to definition of T.
Note that in the simplest scenario where the only component of the measurement space is BGL and no
patient activities are accounted for, we merely have S = L × T, and s = {gl, τ }, where gl ∈ L and τ ∈ T.
Also note, that even though ι already carries the information about τ , we will still sometimes explicitly
specify τ , for the clarity.
2.2
Reward function
Denote the optimal (the most healthy) BGL by glh . Define reward function r(s(t,τ ) ) which maps the BGL at
s at some given time {t, τ } to the value which quantifies the quality of the healthiness of a patient with BGL
at s(t,τ ) . The linear reward component, normalized to 1, is given by
rl (s) = rl (g) = 1 −
|gl − glh |
gl1 − glL
Define a subset of dangerous states Lhyp ∈ L of being close to hypoglycemia by Lhyp = {1, · · · , Lc }, Lc < L.
Normally, Lc is small. It defines the number of states with critically low BGL. Define the non-linear reward
component which fines for being close to hypoglycemia
rn (s) = rn (g) = Igl<glLc · [1 −
gl − glLc
],
glLc
where I is the standard indicator function. The joint instantaneous reward is given by
r(g) = rl (g) − rn (g)
.
2.3
Action Space
The action space consists of set of possible insulin related medical actions. We allow for finite number
of treatments, in particular, the most known of them - short-acting (Actrapid-like) insulin preparations,
intermediate-acting (NPH-like) and long-acting (Ultratard-like) insulin preparations. Denote the sets of
possible dosages
As = {as1 , · · · , as|As | }, Ai = {ai1 , · · · , ai|Ai | }, Al = {al1 , · · · , al|Al | },
corresponding the short-acting, intermediate-acting and long-acting insulin medications. Note that the first
values of the spaces as1 , ai1 , al1 are equal to 0, meaning no mediation of that type is given. As long as the
medications can be administered concurrently we assume that the action space is given by
A = A s × A i × Al
At each state s ∈ S, an action a ∈ A is chosen. At time mark ι, a(ι) is a triplet {as(ι) , ai(ι) , al(ι) }.
5
2.4
Transition Probabilities
The transitions between states occurs at every absolute measurement point according to transition probabilities
given by p(s0(ι+1) |s(ι) , a(ι) ). That is, the probability of getting to the next state s0 , given that the previous
state was s, and action a was taken. Clearly,
X
p(s0(ι+1) |s(ι) , a(ι) ) = 1,
s0
where the summation is over all possible states at ι + 1.
Clearly, these probabilities are unknown a priori. Hence, the objective is to statistically learn them using
a patient’s measurements, under previously applied insulin regime. This regime could be applied according to
doctor’s purely medical (i.e., non-mathematical) considerations. Once these probabilities are learned, the
new, optimal insulin policy is derived.
2.5
Total reward functional
Define the total discounted reward, with positive discount factor γ < 1.
J =E
∞
X
γ ι r(s(ι) ),
(1)
ι=0
The discount factor γ has two important interpretations. The first one follows from the known analytical
property which allows the sum in (1) to converge. The second one advocates the reasoning of ”feeling better
now” is more important than ”feeling better in a distant future”. Therefore, BGL rewards for the distant
time marks are discounted with geometrically decreasing discount multiplicative. The alternative approach,
which values equally the feeling throughout the day employs the following definition
J =E
τT
∞ X
X
t
γ r(s(t,τ ) ) = E
t=0 τ =τ1
3
∞
X
γ
t=0
τT
X
t
r(s(t,τ ) ).
(2)
τ =τ1
MDP and RL solution
3.1
The Bellman equation
Write the equation (1), for the given initial state s(0) as follows
J π (s(0) ) = E
∞
X
γ t r(s(t) ) = r(s(0) ) + E
t=0
∞
X
γ t r(s(t) )
(3)
t=1
Denote by π a policy which selects action at each state. State s(1) is set at time mark t = 1, according to
the transition probability from state s(0) , given action aπ (s(0) ) = {aπ,s (s(0) ), aπ,i (s(0) ), aπ,l (s(0) )}, which is
chosen according to π.
Define the value function given an initial state:
V (s(0) ) = max J π (s(0) )
π
6
(4)
V (s(0) ) indicates the optimal total reward obtained for the BGL throughout time, when starting BGL was as
in s(0) . We use next the dynamic programming (DP) principle. Expand (3) as follows
J π (s(0) ) = r(s(0) ) + E
∞
X
γ t r(s(t) ) = r(s(0) ) +
X
p(s1 |a(s0 ), s0 )J π (s(1) )
(5)
s1
t=1
Applying DP on the value function gives the Bellman equation (see e.g., [1])
X
V (s(0) ) = r(s(0) ) + max
p(s1 |a(s0 ), s0 )V (s(1) )
a(s0 )
(6)
s1
Denote V the vector of all V (s). The Bellman equation (6) is solved by the value function, a result of a
mapping from the state space to the space of the total discounted reward, which is complete metric space,
denote it by V:
V : S 7→ V,
equipped with a suitable metric kVa − Vb k. Observe that (6) can be written with operator T , that is, V = T V .
By a well-known result (see, e.g., [1]), operator T is strict contraction (i.e., αkVa − Vb k < kVa − Vb k, for
some 0 < α < 1). Therefore, T has a fixed point (see, e.g. [11, Theorem V.18]). This fixed point is V and it
constitutes the unique solution to the Bellman equation (6). The reasoning above means V is found by a
well-known method of value iteration (see, again, [1]). That is, T is iteratively applied, starting with initial
values of some U ∈ V, till the arbitrarily close convergence to the fixed point V .
3.2
Algorithm formulation
Observe that (6) can only be solved by using transition probabilities, as were defined in previous section.
These probabilities constitute the dynamics of the stochastic process of BGL and the response to the actions,
i.e. the insulin preparations. The straightforward approach to calculate them is by mere statistical learning
over the existing vector of samples of measurements of BGL paired with the corresponding actions.
We formulate the complete Insulin Optimization Policy Algorithm next. See Figure 3.2 below for the
complete formulation. The phase A is the initialization of the vector of statistics of ν(s0 , a, s) which counts all
occurrences of being in state s, acting by applying a and passing to the next state s0 . The phase B comprises
the learning phase where acting is done by initial policy π 0 , which normally taken from the same samples,
and stems from doctor’s heuristic consideration treating the individual patient. The output of this phase is
the transition probabilities. These probabilities are exploited in the phase C to perform the value iteration
and to retrieve the optimal policy π ∗ . The number of iterations depends on the size of the vector V and is
heuristically set, in order to achieve the desired level of convergence. Finally, phase D is the exploitation
part, which exploits the latest found optimal policy π. The phase keeps tracking the transition occurrences
and updates the policy every M transitions.
The final phase reflects possible changes in the BGL process properties through the time.
4
Optimal policy result
In this section we use data from [7], in order to demonstrate the activation of the algorithm described in
the previous section. Both examples presented in this section are solely based on short vectors of several
hundreds of measurements of BGL. Anything else, e.g., carbohydrates intakes and physical activities are
presumed to be embodied in the BGL process. Clearly, the obtained policy is only a presumable outcome, as
the learning was done over short-sized vectors. While the BGL (simply denoted as Glucose in the graphs
below) scale is presented starting from minimal measurement and ending by maximal BGL measurement, not
7
Insulin Optimization Policy Algorithm
A. Initialization
1. Initialize statistical data ν(s0 , a, s) = 0, for all triplets {s0 , a, s}
2. Initialize the value function V (s) for all initial states s.
B. Learning from the training sequence of length N
1. set n = 0
2. Visit s = sn , see the action a = an , a ∈ π 0 , taken in that state and the next state
s0 = sn+1 . Update ν(s0 , a, s) = ν(s0 , a, s) + 1
3. Increment n, if n < N go to 1.
4. Calculate p(s0 |a, s) =
0
Pν(s ,a,s)
0
s0 ν(s ,a,s)
C. Value function and policy calculation
1. Use p(s0 |a, s) to perform value iteration till desired level of convergence of V .
2. Apply the max operator at each state s to find the best action. Set argmaxa ∈ π ∗ , the
optimal policy.
D. Policy on-line tracking and calibration
1. Initialize M , set n = 0.
2. Visit s = sn , see the action a = an , a ∈ π ∗ , taken in that state and the next state
s0 = sn+1 . Update ν(s0 , a, s) = ν(s0 , a, s) + 1
3. Increment n, if n < M go to 1.
4. Calculate new p(s0 |a, s)
5. Use new p(s0 |a, s) to perform again value iteration
6. Update the new π ∗ .
at all points in the BGL scale the policy any policy is indicated. The reason for this stems from the limited
data size. In particular, the action outcome cannot be learned in the case no occurrence of any action for
that specific BGL was ever recorded. For this reason the preliminary policy π 0 should be carefully selected.
The best anticipated practice would be having an assessment of a particular individual and assigning them
to the specific category of policies. Once the successful assignment is done, the correction is expected to be
effective even if only limited period of individualized measurements is offered. We further elaborate in the
Future Work section.
5
Future Work
The future work is aimed at two directions.
• Producing a set of categories of individuals. The individuals would be assigned into the finite number
of policy sets. Each set will contain patients with resembling (but clearly not similar) insulin regime.
8
Figure 1: Policy example 1, 943 measurement points.
Figure 2: Policy example 2, 340 measurement points.
9
The patients may also have other resembling parameters, e.g., body weight, age, etc. The categorizing
process will be done by both heuristic assignment and correlation study.
• The second direction relates tot the pure ML approach and is directed towards coping with the size of
of the state-space. For this purpose, approach from approximate MDP (AMDP) will be introduced.
See, e.g., methods described in [10, Chapter 7].
References
[1] D. Bertsekas. Dynamic programming and optimal control, volume 2. Athena Scientific Belmont, MA,
1995.
[2] E. Daskalaki, P. Diem, and S. G. Mougiakakou. Model-free machine learning in biomedicine: Feasibility
study in type 1 diabetes. PloS one, 11(7):e0158722, 2016.
[3] E. I. Georga, V. C. Protopappas, and D. I. Fotiadis. Glucose prediction in type 1 and type 2 diabetic
patients using data driven techniques. In Knowledge-oriented applications in data mining. InTech, 2011.
[4] Harvard School of Public Health. Measuring Physical Activity. https://www.hsph.harvard.edu/
nutritionsource/mets-activity-table/, 2017.
[5] R. I. Holt, C. Cockram, A. Flyvbjerg, and B. J. Goldstein. Textbook of diabetes. John Wiley & Sons,
2016.
Blood glucose
[6] Joslin Diabetes Center.
Goals-for-Blood-Glucose-Control.html, 2017.
chart.
http://www.joslin.org/info/
[7] M. Lichman, University of California, Irvine, School of Information and Computer Sciences. UCI machine
learning repository. https://archive.ics.uci.edu/ml/datasets/Diabetes+130-US+hospitals+
for+years+1999-2008, 2013.
[8] Mayo Clinic.
Blood sugar testing: Why, when and how.
http://www.mayoclinic.org/
diseases-conditions/diabetes/in-depth/blood-sugar/art-20046628, 2017.
[9] S. Moscou. Getting the word out: advocacy, social marketing, and policy development and enforcement.
Public Health Nursing, page 285, 2010.
[10] W. B. Powell. Approximate Dynamic Programming: Solving the curses of dimensionality, volume 703.
John Wiley & Sons, 2007.
[11] M. Reed and B. Simon. Methods of modern mathematical physics: Functional analysis, volume 1. Gulf
Professional Publishing, 1980.
[12] C. M. Ripsin, H. Kang, R. J. Urban, et al. Management of blood glucose in type 2 diabetes mellitus.
Am Fam Physician, 79(1):29–36, 2009.
[13] D. K. Rollins, N. Bhandari, J. Kleinedler, K. Kotz, A. Strohbehn, L. Boland, M. Murphy, D. Andre,
N. Vyas, G. Welk, et al. Free-living inferential modeling of blood glucose level using only noninvasive
inputs. Journal of process control, 20(1):95–107, 2010.
[14] U.S. National Library of Medicine. Blood Glucose Self-Monitoring. https://meshb.nlm.nih.gov/
record/ui?ui=D015190, 2017.
[15] World Health Organization. Diabetes facts list. https://web.archive.org/web/20130826174444/
http://www.who.int/mediacentre/factsheets/fs312/en/, 2013.
10
[16] R. A. Zitar and A. Al-Jabali. Towards neural network model for insulin/glucose in diabetics-ii. Informatica,
29(2), 2005.
11
| 3cs.SY
|
DEFORMATIONS OF FUNDAMENTAL GROUP REPRESENTATIONS AND
EARTHQUAKES ON SO(n, 1) SURFACE GROUPS
arXiv:1609.02644v1 [math.GT] 9 Sep 2016
SON LAM HO
A BSTRACT. In this article we construct a type of deformations of representations π1 (M ) → G
where G is an arbitrary lie group and M is a large class of manifolds including CAT(0) manifolds.
The deformations are defined based on codimension 1 hypersurfaces with certain conditions, and
also on disjoint union of such hypersurfaces, i.e. multi-hypersurfaces. We show commutativity
of deforming along disjoint hypersurfaces. As application, we consider Anosov surface groups in
SO(n, 1) and show that the construction can be extended continuously to measured laminations,
thus obtaining earthquake deformations on these surface groups.
1. I NTRODUCTION
The Fenchel-Nielsen twist is one of the most fundamental tools in studying the deformation
space of a hyperbolic surface M , equivalently the Teichmueller space T (M ). Geometrically it can
be described as cutting the surface along a simple closed geodesic, do a twist of some length t and
then glue back. However these deformations can also be described algebraically and perhaps more
naturally so. That is the point of view of this paper: we will define a type of algebraic deformations
of representations ρ : π1 (M ) → G where G is an arbitrary lie group and M is a manifold with
contractible universal cover. It turns out that many geometric deformations are special cases of this
construction if we consider only the holonomy representation, for example, bending a quasifuchsian
surface group defined by Thurston [17], Johnson-Millson bending [13], the twist-bulge deformation
of a convex real projective surface by Goldman [6], etc. Interestingly, the starting representation ρ
from which we deform need not be discrete.
1.1. Construction and results. Let M be an oriented surface of genus g > 1 and N ,→ M a
directed simple closed curve, this is the case of most interest to us even though our construction
applies more generally to aspherical manifolds of higher dimensions with two-sided aspherical
hypersurfaces. Let ρ : π1 (M ) → G be faithful, and suppose that the centralizer CG (ρ(π1 (N )))
f, we assign
is non-trivial. For each lift Ni (where i = 0, 1, 2, ..) of N in the universal cover M
a transformation γi in the G-centralizer of the ρ-image of the cyclic subgroup of π1 (M ) which
preserve Ni under deck transformation. This assignment of γi also has to satisfy a condition of
equivariance, which means that the choice of γ0 for N0 determines the rest of the γi ’s. Fix a base
f, our construction can be informally described as follows. Given A ∈ π1 (M ), we
point x̃0 ∈ M
consider a directed path from x̃0 to x̃0 A where x̃0 A is the image of x̃0 under the right-action of
deck transformations. This path will cross an ordered collection of lifts N1 , ..., Nk , we can define a
new function EN,γ (ρ) : π1 (M ) → G such that:
EN,γ (ρ)(A) = ρ(A)γksk ...γ1s1
2010 Mathematics Subject Classification. 57M50, 20F65.
Key words and phrases. CAT(0) manifold, Deformation, Surface group, Quasifuchsian, Hyperbolic, Convex cocompact, Fenchel-Nielsen, Earthquakes, Bending.
1
2
SON LAM HO
where si is the intersection sign between Ni and the path from x̃0 . In fact the result does not depend
on the particular path from x̃0 to x̃0 A, as long as this path intersects Ni transversely. We can then
show that the resulting map EN,γ (ρ) : π1 (M ) → G is indeed a homomorphism. Thus for example,
when G = P SL(2, R) and ρ a holonomy representation of a hyperbolic surface, we can choose
γi inside a 1-parameter group of hyperbolic transformations and obtain a 1-parameter family of
representations corresponding to the Fenchel-Nielsen twists.
In section 2.2 we generalize the above construction to multi-curves on surfaces (and multihypersurfaces on manifolds). We will also prove a commutativity result, Theorem 2.9, which we
will state below. Let L be a multi-hypersurface, or a union of disjoint hypersurfaces on M , satisfying generic conditions. Suppose that for each component Li of lifts of L we have γi and αi chosen
equivariantly such that the deformations EL,γ (ρ) and EL,α (ρ) can be defined. Then
Theorem. (Theorem 2.9) If we have γi αi = αi γi then
EL,γ (EL,α (ρ)) = EL,α (EL,γ (ρ)).
This implies the following important corollaries: 1-parameter deformations along a hypersurface
is a flow, that is, EL,γ t (EL,γ s (ρ)) = EL,γ t+s (ρ); and deformations along disjoint hypersurfaces
commute. For example, given a P SL(2, C) surface group, bending it along a curve and twisting it
along another (disjoint) curve are commutative.
The above construction and theorem are presented in section 2. A large part of this paper is in
section 3 where we switch attention to the case of surface groups in G = SO(n, 1) = Isom(Hn )
and study the limit of deformations along weighted simple closed curves that approach a measured
lamination. Let S be a surface of genus g > 1. Thurston introduced the space of measured laminations ML(S) which can be thought of as a completion of the space of weighted simple closed C(S)
which is in fact dense in ML(S) [17]. The classical earthquake deformation is then defined to be
the limit of Fenchel-Nielsen twists as a sequence in C(S) approach a measured lamination [11]. We
aim to generalize this method for Anosov surface groups in SO(n, 1) in order to define earthquakes.
Guichard-Wienhard [7] showed that for SO(n, 1) surface groups, being Anosov is equivalent to
being convex cocompact (See also the work of Bowditch [4], Kapovich-Leeb-Porti [10]). We state
below a reduced version of their theorem.
Theorem 1.1. (Guichard-Wienhard) Let π be a finitely generated word hyperbolic group and G
a real semisimple Lie group of real rank 1. For a representation ρ : π → G, the following are
equivalent:
(i) ρ is Anosov.
(ii) There exists a continuous ρ-equivariant and injective map L : ∂∞ π → G/P
(iii) ker ρ is finite and ρ(π) is convex cocompact.
Indeed since S is a closed surface, this coincide with the notion of 1-quasifuchsian group in [9].
When π = π1 (S) and G = SO(n, 1), condition (ii) above implies that there exists a continuous
e → ∂∞ Hn whose image is the limit set Λ, a quasi-circle
ρ-equivariant injective map L : ∂∞ (S)
n
in ∂∞ H . Moreover, ρ(π) is purely loxodromic, each element of ρ(π) has an attracting and a
repelling fixed point on Λ, and the centralizer of this element must contain a 1-parameter group of
hyperbolic transformations having the same pair of fixed points. We use these 1-parameter groups
to define twist deformations of ρ along weighted simple closed curves, where the weight determines
the translation length.
To pass from simple closed curves to measured laminations, we use the geodesic currents view
e =
of measured laminations [2]. Thus a measured lamination is a π1 (S)-invariant measure on G(S)
e × ∂∞ (S)
e − ∆)/Z2 with 0 self-intersection. Convergence of weighted simple closed curves
(∂∞ (S)
DEFORMATIONS OF REPRESENTATIONS AND EARTHQUAKES ON SO(n, 1) SURFACE GROUPS
3
e Together with the homeoto a measured lamination becomes convergence of measures on G(S).
e → Λ we can show the following convergence result.
morphism L : ∂∞ (S)
Theorem. (Corollary 3.7) For any > 0, λ ∈ ML(S) an Anosov surface group representation ρ
into SO(n, 1), there is a neighborhood U 3 λ such that for any two weighted simple closed curves
l1 , l2 ∈ U , the corresponding representations El1 (ρ) and El2 (ρ) are close.
This means that we can simply define Eλ (ρ) to be limi→∞ Eli (ρ) whenever (li ) → λ, and we have a
continuous map from ML(S) to the space of representations near ρ. This is a direct generalization
of the classical earthquake on hyperbolic surfaces. But unlike in dimension 2, simple dimension
count implies that for n > 2, these earthquakes cannot take ρ to all nearby points in the moduli
space.
1.2. Implications and further directions. In [14], McMullen defined a notion of complex earthquake: starting from a fuchsian group Γ ⊂ P SL(2, R) ⊂ P SL(2, C) it combines classical earthquake and Thurston’s bending/grafting deformation and deform the original group to become quasifuchsian. Theorem 2.9 and Corollary 3.7 implies that starting from a quasifuchsian Γ, along any
measured lamination, both earthquake and Thurston’s bending can be defined and they are commutative. Thus we can combine them into quake-bend deformations with complex parameters like
in [3]. One can then ask whether it is possible to connect 2 quasifuchsian surface groups by 2
quake-bend operations. Thurston showed that we can do that starting from a fuchsian group.
Theorem. (Thurston, [8]) The projective grafting map Gr : ML(S) × T (S) → P(S) is a homeomorphism.
where T (S) is the Teichmuller space of hyperbolic structures whose holonomy are fuchsian, and
P(S) is the space of CP 1 structures whose holonomy include all quasifuchsian groups.
We have Anosov surface groups in G = SO(n, 1) form an open subset QF G (S) of the moduli
space Hom(π1 (S), G)/G of representations up to conjugations, but they are in general not a whole
component of this space. An interesting question is whether earthquake paths Etλ (ρ) can go outside
of the closure of QF G (S).
Another interesting direction is the question of Fenchel-Nielsen coordinates for SO(n, 1) surface groups, in particular SO(4, 1). Tan [15] and Kourouniotis [12] constructed complex FenchelNielsen coordinates for the case of P SL(2, C) = SO+ (3, 1). The situation in SO(4, 1) is more
complicated since in some cases the bending parameter can be any SO(3) rotation, and SO(3) is
not abelian. Moreover, by dimension count there may be up to 4 dimensions of internal parameter
for each pair-of-pants in the decomposition, these parameters specify the arrangement of rotation
axes. Some work has been done in [16] to study pair-of-pants groups in dimension 4, but the whole
picture remains mysterious.
Acknowledgement. I would like to thank Jean-Marc Schlenker and Virginie Charette for being my
mentors during the writing of this article, and I thank my advisor Bill Goldman whose works are
such an inspiration.
2. D EFORMATIONS ALONG HYPERSURFACES
f is con2.1. The Construction. A manifold M is said to be aspherical if its universal cover M
tractible, or equivalently πk (M ) = 0 for all k > 1. For the rest of this section let M be a connected
aspherical manifold (possibly with boundary) of dimension at least 2 with finitely generated fundamental group, and let ι : N ,→ M be a connected properly embedded two-sided aspherical
hypersurface such that ι∗ : π1 (N ) → π1 (M ) is injective.
4
SON LAM HO
Let ρ : π1 (M ) → G be a representation into an arbitrary Lie group G. We will define deformations of ρ along the hypersurface N . An important example is when M is a hyperbolic surface and
N is a simple closed geodesic, or more generally when N is a totally geodesic hypersurface in a
hyperbolic manifold M , in these cases our construction corresponds to Johnson-Millson’s bending
in [13].
b ⊂M
f denote the preimage of N in the universal cover. Then N
b is a disjoint union of a
We let N
e which is
countable collection N = {N0 , N1 , ...} of connected components, each Ni is a copy of N
f
f
contractible and of codimension 1 in M . So M − Ni has two components with a common boundary
Ni .
Definition 2.1. We define the dual tree T to N ⊂ M to be the tree whose vertices are connected
f−N
b and edges are Ni . The adjacency of vertices and edges corresponds to the
components of M
f−N
b with Ni ’s.
adjacency of components M
Each vertex of the tree T possibly has an infinite number of edges attached. Deck transformation
f induce an action on T which is transitive (but not free) on the set of vertices
action of π1 (M ) on M
and the set of edges.
f, with a corresponding base point x0 ∈ M so that x0 6∈ N . For all
Choose a base point x̃0 ∈ M
f as a right action of deck transformations, so in our notation we
A, B ∈ π1 (M, x0 ), they act on M
f. Now we will define an algebraic twist deformation of the
have x̃.(AB) = (x̃.A).B for x̃ ∈ M
representation ρ : π1 (M, x0 ) → G.
Let y0 ∈ N a base point and choose a path from x0 to y0 . By path concatenation this induces
an injective homomorphism π1 (N, y0 ) → π1 (M, x0 ) whose image is a subgroup T0 ⊂ π1 (M, x0 ).
f. We have ỹ0 is on a component of N
b and
The path x0 to y0 lifts uniquely to a path x̃0 to ỹ0 in M
we label this component N0 . The action of T0 preserves N0 , that is, N0 .A = N0 for all A ∈ T0 .
b we have Ni = N0 .Ai , so the subgroup Ti = A−1 T0 Ai preserves Ni .
For each component Ni of N
i
The following is the crucial point of our construction. We define a map γ : N → G with the
following properties:
(1) γ(Ni )ρ(A) = ρ(A)γ(Ni ) for all A ∈ Ti , that is, γ(Ni ) is in the centralizer of ρ(Ti ) in G,
(2) If A ∈ π1 (M, x0 ) such that Nj = Ni .A, then γ(Nj ) = ρ(A−1 )γ(Ni )ρ(A).
For the rest of this article, γ will be reserved to denote this map. Note that property (2) implies
property (1), since Ni = Ni .Ti . We want to check that if the centralizer of ρ(Ti ) is non-trivial then
a non-trivial γ satisfying (2) exists.
We claim that choosing γ(N0 ) 6= 1 in the centralizer of ρ(T0 ) uniquely determines γ(Ni ) for all
Ni ∈ N . For Ni ∈ N , suppose that Ni = N0 .Ai , then we can use property (2) to define γ(Ni ) =
0
ρ(A−1
i )γ(N0 )ρ(Ai ). Note that Ai is not unique for each Ni . If we have Ni = N0 .Ai = N0 .Ai then
0
−1
0
−1
N0 .Ai (Ai ) = N0 , so Ai (Ai ) ∈ T0 . Since we chose γ(N0 ) in the centralizer of ρ(T0 ) we have
0 −1
γ(N0 ) = ρ(A0i A−1
i )γ(N0 )ρ(Ai (Ai ) )
ρ(Ai0−1 )γ(N0 )ρ(A0i ) = ρ(A−1
i )γ(N0 )ρ(Ai )
So γ(Ni ) is well-defined, not depending on different choices of Ai .
Now suppose that Nj = Ni .B, and we have Ni = N0 .Ai . So we can let Aj = Ai B, thus
B = A−1
i Aj . We have
γ(Nj ) = ρ(A−1
j )γ(N0 )ρ(Aj )
−1
−1
= ρ(A−1
j Ai )ρ(Ai )γ(N0 )ρ(Ai )ρ(Ai Aj )
−1
= ρ(B )γ(Ni )ρ(B).
DEFORMATIONS OF REPRESENTATIONS AND EARTHQUAKES ON SO(n, 1) SURFACE GROUPS
5
This is property (2). Therefore γ as above is well-defined and satisfies (1) and (2).
We are now ready to define the twist deformation of representation ρ with respect to γ. First we
f.
choose an orientation of N ⊂ M , which induces orientation of each Ni ⊂ M
Definition 2.2. Let γ : N → G be defined as above. Given A ∈ π1 (M, x0 ), a (minimal) path
f with direction from x̃0 to x̃0 .A will cross a sequence of hypersurfaces in N which we name
in M
Na1 , ...Nak in order. We define
EN,γ (ρ)(A) = E(ρ)(A) := ρ(A) [γ(Nak )sk ...γ(Na1 )s1 ]
where each si = ±1 and equals the intersection sign between Ni and the directed path x̃0 to x̃0 .A.
We say that the deformation is non-trivial if γ(Ni ) 6= 1.
Remark 2.1. Note that it is not necessary for the path from x̃0 to x̃0 A to be minimal since any overlapping would be cancelled out in the expression for E(ρ)(A), this is because of the tree structure
f−N
b . The intersection sign between Ni and the path depends on the chosen orientation of
of M
M and N . There is a dual definition of E(ρ) obtained by changing the orientation of M , or changing the orientation of N ⊂ M . Both are equivalent to switching between a “left” and a “right”
Fenchel-Nielsen twist for the case of surface groups in G = Isom+ (H2 ).
From the above definition we have a well-defined map E(ρ) : π1 (M, x0 ) → G, because a
minimal path from x̃0 to x̃0 .A induces a minimal path in the dual tree T and which is unique, so
the sequence of lines Na1 , ..., Nak is uniquely determined. The following proposition shows that
E(ρ) : π1 (M, x0 ) → G is a homomorphism.
Proposition 2.3. Let A, B ∈ π1 (M, x0 ) and E(ρ) as defined above. Then
(i) E(ρ)(AB) = E(ρ)(A)E(ρ)(B),
(ii) E(ρ)(A−1 ) = E(ρ)(A)−1 .
Therefore E(ρ) is a homomorphism.
Proof.
(i) We use notations as before in this section. Let Na1 , ..., Nap be the sequence of hypersurfaces
between x̃0 and x̃0 A, and Nb1 , ..., Nbq be the sequence of hypersurfaces between x̃0 and x̃0 B, and
let Nc1 , ..., Ncr be the sequence of hypersurfaces between x̃0 and x̃0 AB. Suppose that the definition
of E(ρ) gives us
E(ρ)(A) = ρ(A)αp ...α1
E(ρ)(B) = ρ(B)βq ...β1
E(ρ)(AB) = ρ(A)ρ(B)δr ...δ1
Where αi = γ(Nai )±1 , βi = γ(Nbi )±1 , and δi = γ(Nci )±1 , with the signs determined by the
corresponding intersection signs.
We have
E(ρ)(A)E(ρ)(B)
= ρ(A)αp ...α1 ρ(B)βq ...β1
= ρ(A)ρ(B)[(ρ(B −1 )αp ρ(B)) ... (ρ(B −1 )α1 ρ(B)) βq ...β1 ]
{z
}
|
{z
} |
Indeed β1 , ..., βq are elements in G associated with hypersurfaces Nb1 , ..., Nbq which are between
x̃0 and x̃0 .B, and ρ(B −1 )α1 ρ(B), ..., ρ(B −1 )αp ρ(B) are elements associated with hypersurfaces
Na1 .B, ..., Nap .B which are between x̃0 .B and x̃0 .AB. On the other hand, δ1 , ..., δr are elements
associated with the sequence of hypersurfaces between x̃0 and x̃0 .AB. Consider the tree structure
of the dual tree T (see Figure 1). Indeed the difference between a minimal path from x̃0 to x̃0 AB
6
SON LAM HO
βq
x̃0 B
ρ(B −1 )α1 ρ(B)
ρ(B −1 )α2 ρ(B)
β2
β1
ρ(B −1 )αp ρ(B)
x̃0
δ2
δ1
δr
x̃0 AB
F IGURE 1.
and a concatenation of 2 minimal paths x̃0 to x̃0 B to x̃0 AB is a segment of back tracking, and
along that segment which crosses let’s say m hypersurfaces, we must have cancellation between
ρ(B −1 )αm ρ(B)...ρ(B −1 )α1 ρ(B) and βq ...βq−m+1 .
So
[ρ(B −1 )αp ρ(B)...ρ(B −1 )α1 ρ(B)]βq ...β1 = δr ...δ1 .
Therefore E(ρ)(A)E(ρ)(B) = E(ρ)(AB).
(ii) A similar argument as above shows that
E(ρ)(A−1 )
= ρ(A−1 ) ρ(A)α1−1 ρ(A)−1 ... ρ(A)αk−1 ρ(A)−1
{z
} |
|
{z
}
= α1−1 ...αk−1 ρ(A−1 )
= E(ρ)(A)−1 .
For the case of simple closed curves in surfaces we have the following.
Proposition 2.4. Let M be a closed surface of genus g > 1 and let N, L be disjoint simple closed
curves in distinct homotopy classes, and let ρ : π1 (M ) → G be a representation such that ρ(π1 (N ))
and ρ(π1 (L)) have non-trivial centralizers. Then non-trivial deformations (as constructed in definition 2.2) of ρ along N and L are distinct.
Proof. We cut the surface M along N and obtain a surface M 0 with boundary which still contain
L, so we can choose a base point x0 ∈ M in the same component containing L (if N is a separating
curve). Note that the representation Eρ only changes by a conjugation as the base point changes.
Now we can choose A ∈ π1 (M, x0 ) represented by a curve intersecting L and not touching the
boundary. Thus EN (ρ)(A) = ρ(A) while EL (ρ)(A) 6= ρ(A). Thus EN (ρ) 6= EL (ρ).
2.2. Deformation along multi-hypersurfaces. In this section we will define deformations along
a set of disjoint hypersurfaces - the generalization of deforming a surface group along multicurves.
A theorem on commutativity of deformations will be shown.
DEFORMATIONS OF REPRESENTATIONS AND EARTHQUAKES ON SO(n, 1) SURFACE GROUPS
7
Let L = N(1) ∪ ... ∪ N(n) be a collection of mutually disjoint connected hypersurfaces of M ,
each satisfies properties of N in previous section. Moreover we require that no pair N(i) , N(j) are
homotopic. Let
n
[
L = {L1 , L2 , ...} =
{N(i),1 , N(i),2 , ...}
i=1
f. As before, Ti ⊂ π1 (M, x0 )
be the collection of components of pre-images of N(1) , ..., N(n) in M
is the subgroup preserving Li .
Let γ : L → G be such that
(1) γ(Li )ρ(A) = ρ(A)γ(Li ) for all A ∈ Ti
(2) If A ∈ π1 (M, x0 ) such that Lj = Li .A, then γ(Lj ) = ρ(A−1 )γ(Li )ρ(A).
Suppose L1 , ..., Ln are N(1),0 , ..., N(n),0 in order. Then indeed, γ is determined by γ(L1 ), ..., γ(Ln ).
The following definition and proposition are straight forward generalization of previous section.
Definition 2.5. Let L = N(1) ∪...∪N(n) and γ : L → G be defined as above. Given A ∈ π1 (M, x0 ),
f with direction from x̃0 to x̃0 .A will cross a sequence of hypersurfaces in L
a (minimal) path in M
which we name La1 , ..., Lak in order. We define
EL,γ (ρ)(A) = E(ρ)(A) := ρ(A) [γ(Lak )sk ...γ(La1 )s1 ]
where each si = ±1 and equals the intersection sign between Li and the directed path x̃0 to x̃0 .A.
Proposition 2.6. EL,γ (ρ)(AB) = EL,γ (ρ)(A)EL,γ (ρ)(B). Deformation along multiple disjoint
hypersurfaces gives EL,γ (ρ) a homomorphism.
We will need the following discussion to prove commutativity of deformations. Let x0 , x00 be
different base points on M , x0 , x00 6∈ L. Choose a path from x0 to x00 which determine a lift x̃0 to
f. We will deform ρ : π1 (M, x0 ) → G in two different ways: the first using definition 2.5
x̃00 on M
and base point x0 resulting in EL,γ (ρ); the second way of deforming is by using x̃00 as base point.
0 (ρ) be the resulting representation obtained by using x̃0 as the base point as follows: for
Let EL,γ
0
A ∈ π1 (M, x0 ), we use the (possibly non-minimal) path from x̃00 to x̃00 A obtained by going from
x̃00 to x̃0 , then to x̃0 A, then to x̃00 A. By remark 2.1 we see that taking a non-minimal path does not
0 (ρ) and E
change the deformation result. We will now show that the difference between EL,γ
L,γ (ρ)
is conjugating by some transformation in G.
Lemma 2.7. Let the minimal path from x̃0 to x̃00 cross L1 , ..., Lk with intersection sign s1 , ..., sk .
Then
!
!−1
1
1
Y
Y
E 0 L,γ (ρ)(A) =
γ(Li )si EL,γ (ρ)(A)
γ(Li )si
i=k
i=k
Proof. L1 , ..., Lk are the hypersurfaces between x̃0 and x̃00 in order, so L1 A, ..., Lk A are between
x̃0 A and x̃00 A. Let La1 , ..., Lar be the hypersurfaces between x̃0 and x̃0 A. We have the path from
x̃00 to x̃0 to x̃0 A to x̃00 A crosses the following hypersurfaces
Lk , ..., L1 , La1 , ..., Lar , L1 A, ..., Lk A.
So using property (2) of γ we get
0
EL,γ
(ρ)(A) = ρ(A)
1
Y
i=k
!
ρ(A−1 )γ(Li )si ρ(A)
1
Y
i=r
!
γ(Lai )sai
k
Y
i=1
!
γ(Li )−si
8
SON LAM HO
=
1
Y
!
γ(Li )
si
ρ(A)
1
Y
i=r
i=k
!
sai
γ(Lai )
1
Y
!−1
si
γ(Li )
i=k
and we get the result.
Lemma 2.8. Let La ∈ L be hypersurface and Ta ⊂ π1 (M, x0 ) be the subgroup preserving it.
Suppose the minimal path from x̃0 to La crosses L1 , ..., Lk in order (not including La ), then
!−1
!
1
1
Y
Y
EL,γ (ρ)(Ta ) =
γ(Li )si
ρ(Ta )
γ(Li )si
i=k
i=k
Proof. We can choose x̃00 near La so that the minimal path from x̃0 to x̃00 cross L1 , ..., Lk . That is,
f−S∞ Li adjacent to La and closest to x̃0 . So x̃0 A is still in C for all
x̃00 is in the component C of M
0
i=1
0 (ρ)(A) = ρ(A) for all A ∈ T (because of path independence for definiA ∈ Ta , which means EL,γ
a
Q
−1
Q
1
1
si
si
γ(L
)
γ(L
)
tion 2.5). Applying lemma 2.7 we get EL,γ (ρ)(A) =
ρ(A)
i
i
i=k
i=k
The following theorem implies commutativity of deforming along disjoint (multi-) hypersurfaces, and it also shows that choosing γ values in a 1-parameter subgroup of the centralizer induces
a flow path of representations.
Theorem 2.9. Let γ and α be maps L → G satisfying (1) and (2), and so that γ(Li )α(Li ) =
α(Li )γ(Li ). Then
EL,γ (EL,α (ρ)) = EL,α (EL,γ (ρ)).
Proof. Note that γ on the left hand side is different from the one on the right because the starting
representations are different. When distinction is required we write γρ for the γ on the right hand
side and γE for the left hand side, similarly for α. The equation can be written more unambiguously
as
EL,γE (EL,αρ (ρ)) = EL,αE (EL,γρ (ρ)).
Note that we have yet to define γE and αE . Since, by lemma 2.8, EL,α (ρ)(Ti ) = ξi−1 ρ(Ti )ξi for
some transformation ξi , we can naturally define γE so that γE (Li ) = ξi−1 γρ (Li )ξi . Similarly can
relate αρ and αE this way.
Let A ∈ π1 (M, x0 ), let L1 , ..., Lk be the hypersurfaces between x̃0 and x̃0 A and Ti be the
subgroup preserving Li , for i = 1, .., k. Since L1 , ..., Li−1 is between x̃0 and Li , by lemma 2.8 we
have
EL,α (ρ)(Ti ) = αρ (L1 )−s1 ...αρ (Li−1 )−si−1 ρ(Ti )αρ (Li−1 )si−1 ...αρ (L1 )s1
So for i = 1, ...k
(1)
γE (Li ) = αρ (L1 )−s1 ...αρ (Li−1 )−si−1 γρ (Li )αρ (Li−1 )si−1 ...αρ (L1 )s1 .
Moreoever, by definition
(2)
EL,γ (EL,α (ρ))(A) = ρ(A) [αρ (Lk )sk ...αρ (L1 )s1 γE (Lk )sk ...γE (L1 )s1 ] .
From (1) and (2) we get
EL,γ (EL,α (ρ))(A) = ρ(A) [αρ (Lk )sk γρ (Lk )sk ...αρ (L1 )s1 γρ (L1 )s1 ] .
An analogous formula can be shown for EL,α (EL,γ (ρ))(A), and together with the commutativity
assumption αρ (Li )γρ (Li ) = γρ (Li )αρ (Li ), the theorem follows.
Corollary 2.10. Deformations along disjoint hypersurfaces commute.
DEFORMATIONS OF REPRESENTATIONS AND EARTHQUAKES ON SO(n, 1) SURFACE GROUPS
9
2.3. Infinitesimal deformation. Let g be the lie algebra of G, let π = π1 (M ).
Definition 2.11. Suppose s : (−, ) → G is an analytic path such that s(0) = 1, that is, s(t) =
d
exp(a1 t + a2 t2 + ...) for ai ∈ g; and suppose that a1 6= 0. Then we let dt
|t=0 s(t) = a1 and it will
be called the derivative of s(t). We may also write ṡ(0) = a1 .
With the notation as in previous section, suppose we have a 1-parameter subgroup etX (where
X ∈ g) contained in the centralizer of ρ(N0 ). Thus we have a family of choices for γ(N0 ), we
write γ t (N0 ) = etX . Applying the above construction we get a differentiable 1-parameter family
of deformations of the representation ρ ∈ Hom(π, G).
ρ
Ad
Composition with the adjoint representation gives us π → G → Aut(g). An infinitesimal
deformation of Adρ is given by a 1-cocycle u : π → g. See [5].
Let A ∈ π. For simplicity of notation suppose we have a family of deformations as in previous
section given by E t (ρ)(A) = ρ(A)γkt ...γ1t where each γit = etXi is a 1-parameter group. Of
course the collection X1 , ..., Xk ∈ g depends on A. Let E t (ρ)(A) = ρ(A) exp(tu(A) + O(t2 )) for
u(A) ∈ g, in other words we have
Definition 2.12. For a smooth deformation E t (ρ) of ρ such that E 0 (ρ) = ρ, the map u : π1 (M ) → g
defined by
d
u(A) := |t=0 ρ(A)−1 E t (ρ)(A)
dt
is called the infinitesimal deformation corresponding to E t (ρ).
Indeed u is a 1-cocycle corresponding to the 1-parameter family of representations E t (ρ). We
have
exp(tu(A) + O(t2 )) = γkt ...γ1t
= etXk ...etX1
2
= et(Xk +...+X1 )+O(t )
where by the Baker–Campbell–Hausdorff formula, the coefficients for higher order t-terms on the
right hand side include various combinations nested of lie bracket [, ] between X1 , ..., Xk . Therefore
considering the linear coefficients of t we get
Remark 2.2. u(A) = X1 + ... + Xk .
3. A LGEBRAIC EARTHQUAKE ALONG MEASURED LAMINATIONS
We now switch our attention to the case where S is a closed hyperbolic surface, and Lie group
G = SO(n, 1) = Isom(Hn ), with Lie algebra g = so(n, 1). For the rest of this section, let ρ :
π1 (S) → SO(n, 1) be an Anosov representation. In this section we will define twist deformations
ρ along weighted simple closed curves, and then earthquakes along measured laminations on S. We
have
Theorem 3.1. (Guichard-Wienhard [7]) Let π be a finitely generated word hyperbolic group and
G a real semisimple Lie group of real rank 1. For a representation ρ : π → G, the following are
equivalent:
(i) ρ is Anosov.
(ii) There exists a continuous ρ-equivariant and injective map L : ∂∞ π → G/P
(iii) ker ρ is finite and ρ(π) is convex cocompact.
Considering that G = SO(n, 1) is of rank 1, and π1 (S) is word hyperbolic, we will use condition
(ii) above as the definition of Anosov representations in our case. Thus there is a ρ-equivariant
homeomorphism L : S1∞ → Λ where S1∞ is the boundary at infinity of Se and Λ is the limit set of
ρ(π1 (S)). For the rest of this article L is reserved to denote this limit set map.
10
SON LAM HO
e Following [1] and [2], we can
3.1. Measured laminations. Let A ∈ π1 (S), λ ∈ ML(S), x˜0 ∈ S.
1
1
view λ as an π1 (S)-invariant measure on G(S̃) = (S∞ × S∞ − ∆)/Z2 where Z2 acts by swapping
the two coordinates. Indeed G(S̃) is the space of unoriented geodesic on S̃.
Also considering the geodesic arc A from x˜0 to x˜0 A, and we can define λ|A to be the measure on
the space of geodesics intersecting A transversely. We have the support of λ|A is contained inside
the interior of a compact set I1 × I2 where Ii ’s are disjoint closed intervals of S1∞ , indeed I1 , I2
are separated by (neighborhoods of) the end points of the extended geodesic containing A. If a
sequence (λi ) converges to λ in ML(S) then λi |A converges to λ|A in the weak topology on the
space of measures on I1 × I2 .
Given ρ : π1 (S) → SO(n, 1) an Anosov representation, recall that the homeomorphism L :
∼
∼
1
S∞ → Λρ gives us G(S̃) → (Λρ × Λρ − ∆)/Z2 and in particular I1 × I2 → L(I1 ) × L(I2 ). Thus
any measure on I1 × I2 pushes forward to a measure on L(I1 ) × L(I2 ). In particular λ|A on I1 × I2
induces L∗ λ|A a measure on L(I1 ) × L(I2 ).
3.2. Deforming Anosov representations along weighted simple closed curves. Suppose we have
a weighted simple closed geodesic l ⊂ S, and ρ : π1 (S) → SO(n, 1) an Anosov representation.
We will now define a (right) twist of ρ along l. Following the construction in section 2.1, let x̃0 be a
base point in Se not lying on a lift of l. We choose an orientation for S and an orientation for l. The
way we construct γ, it will turn out that the orientation of l does not matter.
Notation 3.1. For x, y ∈ ∂∞ Hn , x 6= y, t ∈ R, let H(x, y, t) be the hyperbolic transformation
(loxodromic without rotation) fixing x, y with translation length t in the direction from x to y. So
x is the repelling and y is the attracting fixed point.
We have H(x, y, t) preserve the geodesic in Hn whose end-points are x, y, and it translate along
this geodesic a distance t from x to y.
e and suppose T0 ⊂ π1 (S) is the subgroup preserving
Let l0 be a lift of l in the universal cover S,
l0 under deck transform action. Then ρ(T0 ) is an infinite cyclic group, ρ(T0 ) = hτ0 i. Moreover by
properties of Anosov representations in SO(n, 1), we have τ0 is loxodromic, thus τ0 can be written
uniquely as a composition τ0 = σ0 θ0 such that σ0 = H(p0 , q0 , t0 ) is a hyperbolic transformation
in the direction of l0 , and θ0 is an elliptic transformation with p0 , q0 among its fixed points. It’s
important to note that we choose the generator τ0 such that L−1 (p0 ), L−1 (q0 ) ∈ S1∞ is respectively
the starting and ending point of the directed infinite geodesic l0 .
We have σ0 and θ0 commutes with H(p0 , q0 , t) for any t ∈ R. (This is because up to conjugation
p0 = 0 and q0 = ∞ in Rn−1 ∪ {∞} = ∂∞ Hn , so H(p0 , q0 , t) acts as scalar multiplication, θ0 acts
as an SO(n − 1) rotation on Rn−1 .) So we can choose
γ t (l0 ) = H(p0 , q0 , tw)
where w is the weight of l and t ∈ R+ , and indeed γ t (l0 ) commutes with τ0 as required by condition
(1) in section 2.1, and thus this choice of γ t (l0 ) gives us γ t (li ) = H(pi , qi , tw) for any lift li . Note
that we also have L−1 (pi ), L−1 (qi ) are respectively the starting and ending points of li . Following
Definition 2.2 we have constructed a 1-parameter family of algebraic twist deformations of ρ.
Definition 3.2. Let Elt (ρ) be the 1-parameter family of representations obtained by deforming of ρ
along the weighted simple closed curve l by choosing γ t (l0 ) = H(p0 , q0 , tw) as above. We simply
write El (ρ) for when t = 1.
Considering the boundary at infinity S1∞ = ∂∞ Se which is mapped homeomorphically to the
limit set Λρ , the picture of these deformations is very similar to the classical Fenchel-Nielsen twist
DEFORMATIONS OF REPRESENTATIONS AND EARTHQUAKES ON SO(n, 1) SURFACE GROUPS
11
I1
x̄2 ...
x̄1
x̄k
x̃0
A
x̃0 A
ȳk
ȳ1 ȳ2 ...
I2
F IGURE 2. I1 × I2 ⊂ S1∞
(2)
. Here x̄i , ȳi are L−1 (xi ), L−1 (yi ) respectively.
deformation of a hyperbolic surface. Let A ∈ π1 (S) and let A be the geodesic arc from x̃0 to x̃0 A.
By definition we have
ρ(A)−1 Elt (ρ)(A) = γ t (lk )sk ...γ t (l1 )s1 =
L−1 (x
γ t (li )si
1
Y
H(xi , yi , tw)
i=k
−1
i ), L (yi )
if we let
= H(xi , yi , tw). Then indeed
are the end points of the lift li . If
we changed the orientation of l, the signs s1 , ..., sk would be flipped, but we would also invert the
choice of γ t (l0 ) and hence of each γ t (li ), because H(y0 , x0 , t) = H(x0 , y0 , t)−1 , thus we would
keep the deformation the same.
We will now show that all the repelling fixed points xi ’s are on one side of A and the attracting
fixed points yi ’s are on the otherside. Let x∗ , y∗ ∈ S1∞ be the endpoints of the bi-infinite geodesic
that is A extended. Then L−1 (xi ), L−1 (yi ) are in S1∞ − {L−1 (x∗ ), L−1 (y∗ )} which is the union
of two disjoint open intervals I1o , I2o . We name these intervals such that a geodesic from I1o to I2o
intersects A at a positive intersection point. (Again refer to definition 2.2.) This is because if li is
from I1o to I2o then si = 1, so γ t (li ) = H(xi , yi , tw) and so L−1 (xi ) is the starting point of li which
is in I1o . If lj is from I2o to I1o then sj = −1, so γ t (lj )−1 = H(xj , yj , tw) so γ t (lj ) = H(yj , xj , tw)
which makes L−1 (yj ) the starting point of lj which is in I2o .
Remark 3.2. xi ∈ L(I1o ) and yi ∈ L(I2o ) for i = 1, ..., k.
3.3. Proof of convergence. For the rest of this section let λ ∈ ML(S) and A ∈ π1 (S). Our goal
is to prove that if a sequence of weighted simple closed curves (li ) converges to λ, then (Eli (ρ)(A))
converges. The limit can then be defined to be Eλ (ρ)(A). Our approach for the proof of convergence
loosely follow Kerckhoff’s [11].
Notation 3.3. For x a point in a metric space X, we denote Br (x) the open ball of radius r centered
at x.
Notation 3.4. For a subset X of a topological space, we denote cl(X) the closure of X.
Notation 3.5. Let X be any set, we denote
X (2) := ((X × X) − {(x, x)|x ∈ X})/Z2
12
SON LAM HO
the set of unordered distinct pairs of points in X.
Notation 3.6. Let v(x, y) ∈ g be such that et.v(x,y) = H(x, y, t) for all t ∈ R.
Note that for w ∈ R+ we have
d
|t=0 H(x, y, tw) = w.v(x, y).
dt
Remark 3.7. Let Rn−1 be the stereographic projection coordinate of ∂∞ Hn − {∞}, and let x, y ∈
Rn−1 . Then H(x, y, t) varies analytically in SO(n, 1) as x, y, t vary. That is, the map H :
(∂∞ Hn )(2) × R → G is analytic, where ∂∞ Hn is represented by either Rn−1 ∪ {∞} or S n−1 .
We also have v : (∂∞ Hn )(2) → g is analytic.
From this point we work with a chosen positive definite inner product on g, the corresponding
left invariant metric dG on G, and the standard metric dS on S n−1 = ∂∞ Hn which induces the
(2)
product metric on S n−1 .
(2)
Lemma 3.3. Let T > 0 be a fixed constant. Let (x0 , y0 ) ∈ S n−1 , there is ε̄ > 0 and K > 0
depending continuously on (x0 , y0 ) such that for any 0 < ε ≤ ε̄ and x, x0 ∈ Bε (x0 ) and y, y 0 ∈
Bε (y0 ) we have
ktv(x, y) − tv(x0 , y 0 )k < Ktε,
and also
dG (H(x, y, t), H(x0 , y 0 , t)) < Ktε
for 0 < t ≤ T , where K depends continuously on (x0 , y0 ).
Proof. We let ε̄ = (1/3)dS (x0 , y0 ) which ensures that cl(Bε̄ (x0 )) ∩ cl(Bε̄ (y0 )) = ∅. Since
v : (∂∞ Hn )(2) → g is analytic, it is locally lipschitz, and thus lipschitz on compact subsets of
(∂∞ Hn )(2) , in particular on cl(Bε̄ (x0 )) × cl(Bε̄ (y0 )). Let K 0 be the infimum of all lipschitz constant for v on this compact set. Indeed
kv(x, y) − v(x0 , y 0 )k
|(x, y), (x0 , y 0 ) ∈ cl(Bε̄ (x0 )) × cl(Bε̄ (y0 ))}
K 0 = sup{
d((x, y), (x0 , y 0 ))
which depends continuously on (x0 , y0 ).
Let 0 < ε < ε̄, for all (x, y), (x0 , y 0 ) ∈ Bε (x0 ) × Bε (y0 ),
√
kv(x, y) − v(x0 , y 0 )k ≤ K 0 d((x, y), (x0 , y 0 )) < K 0 ε 2
√
We let K1 = K 0 2 and scale both sides by t > 0 to get the first inequality.
The inverse exponential map from a neighborhood of I ∈ G to g is analytic and bijective, so it is
lipschitz on compact subsets, in particular on the set Cx0 ,y0 ,T = {H(x, y, t)|(x, y) ∈ cl(Bε̄ (x0 )) ×
cl(Bε̄ (y0 )), t ∈ [−T, T ]}. Again we let K 00 be the infimum of all lipschitz constants for the inverse
exponential map on Cx0 ,y0 ,T which depends continuously on (x0 , y0 ) once T is fixed. Then for all
(x, y), (x0 , y 0 ) ∈ Bε (x0 ) × Bε (y0 ) and 0 < t ≤ T ,
dG (H(x, y, t), H(x0 , y 0 , t)) < K 00 ktv(x, y) − tv(x0 , y 0 )k < K 00 K1 tε.
We let K = K 00 K1 .
Notation 3.8. Let A be the geodesic path from x̃0 to x̃0 A, and let λ(A) be the λ-measure of this
path. For our purpose, the constant T in the above lemma is λ(A) + 1.
Following the discussion in section 3.1, A and λ and a choice of base point x̃0 ∈ Se determines a
(2)
compact set I1 ×I2 ⊂ S1∞ , and Anosov representation ρ determines the limit set homeomorphism
L.
DEFORMATIONS OF REPRESENTATIONS AND EARTHQUAKES ON SO(n, 1) SURFACE GROUPS
13
Lemma 3.4. For a compact set D ⊂ G, there exists a constant C > 0 such that for any β ∈ D,
(x, y), (x0 , y 0 ) ∈ L(I1 ) × L(I2 ), and weights |t|, |t0 | < λ(A) + 1 we have
dG (H(x, y, t)β, H(x0 , y 0 , t0 )β) < CdG (H(x, y, t), H(x0 , y 0 , t0 )).
Proof. We have that the map (x, y, t, β) 7→ H(x, y, t)β is analytic with respect to any reasonable
coordinates. So the result follows from locally lipschitz argument and compactness of the domain
of (x, y, t, β). Note that C depends on D, L(I1 ) × L(I2 ) and λ(A).
Lemma 3.5. There is a constant C > 0 depending on L(I1 ) × L(I2 ) and λ(A) such that if we have
0
0
0
(x1 , P
y1 ), ...,
P(x0k , yr ) ∈ L(I1 )×L(I2 ) and real positive weights t1 , t1 , ..., tr , tr such that |ti −ti | < δ,
and ti , ti < λ(A) + 1. Then
r
r
Y
Y
dG ( H(xi , yi , ti ),
H(xi , yi , t0i )) < Crδ.
i=1
i=1
Proof. First note that by left-invariance, dG (H(xi , yi , ti ), H(xi , yi , t0i )) = dG (I, H(xi , yi , t0i − ti ).
By compactness and locally lipschitz argument, there exists a constant C 0 > 0 depending on
L(I1 ) × L(I2 ) and λ(A) such that dG (I, H(x, y, t)) < C 0 |t| for all (x, y) ∈ L(I1 ) × L(I2 ) and
|t| < λ(A). So
dG (H(xi , yi , ti ), H(xi , yi , t0i )) < C 0 δ
for all i.
|(x, y) ∈ L(I1 ) × L(I2 ) and |t| ≤ λ(A) + 1}. Let D
Let constant R = max{ dG (I,H(x,y,t))
|t|
Q
be the compact subset of distance at most R(λ(A) + 1) from I. We have ki=1 H(xi , yi , ti ) and
Qk
0
i=1 H(xi , yi , ti ) are in D. Using lemma 3.4 we get for any β ∈ D,
dG (H(xi , yi , ti )β, H(xi , yi , t0i )β) < C 00 dG (H(xi , yi , ti ), H(xi , yi , t0i )) < C 00 C 0 δ.
Therefore by triangle inequality and replacing terms of the product one by one from the right, we
have
r
r
Y
Y
dG ( H(xi , yi , ti ),
H(xi , yi , t0i )) < rC 00 C 0 δ = Crδ
i=1
i=1
where C depends only on L(I1 ) × L(I2 ) and λ(A).
Theorem 3.6. Given ρ : π → SO(n, 1) Anosov, A ∈ π1 (S), λ ∈ ML(S), for all > 0 there exists
a neighborhood U ⊂ ML(S) of λ such that for any two weighted simple closed curves l1 , l2 ∈ U
we have
dG (El1 (ρ)(A), El2 (ρ)(A)) < .
Proof. Recall our notation, here Eli (ρ) is the representation obtained from ρ by algebraic deformation along the weighted simple closed curve li .
Notation 3.9. For any µ ∈ ML(S), we abbreviate µ∗ = L∗ µ|A the induced measure on L(I1 ) ×
L(I2 ).
From lemma 3.3 we have a continuous function ε̄ : L(I1 ) × L(I2 ) → R+ , and on this compact
set ε̄ reaches a non-zero minimum value ε̄min = min{ε̄(x, y)|(x, y) ∈ L(I1 ) × L(I2 )} > 0. The
K in lemma 3.3 depends continuously on the centers of the balls considered. So we let Kmax be a
constant that works for all (x, y) in the compact set L(I1 ) × L(I2 ). Following the proof of lemma
Q
3.5 we let D be the compact subset that contains
all transformation of the form ki=1 H(xi , yi , ti )
P
as long as (xi , yi ) ∈ L(I1 ) × L(I2 ) and
ti ≤ λ(A). Let C0 be the constant of lemma 3.4 that
works for D.
14
SON LAM HO
We choose ε0 < ε̄min , and also such that
4C0 Kmax (λ(A) + 1)
The reason for choosing as such will be apparent by the end.
We have for i = 1, 2, the compact set L(Ii ) can be partitioned into m disjoint intervals L(Ii )1 , ..., L(Ii )m
such that each interval is of diameter less than ε0 , in particular each interval lie inside an open ball
of radius ε0 center at some point in L(Ii ). We may use half-open-half-closed (topologically (0, 1])
intervals to ensure they are disjoint. So L(I1 ) × L(I2 ) is partitioned into m2 squares which we
name Si,j = L(I1 )i × L(I2 )j for 1 ≤ i, j ≤ m.
ε0 <
Notation 3.10. For a square S ∈ {Si,j } we let (xS , yS ) be its center. That is, (xS , yS ) ∈ S such
that S ⊂ Bε0 (xS ) × Bε0 (yS )
Since only a finite number of geodesic leaf of λ in I1 × I2 can have positive measure, we can
assume each Si,j to be a continuity set with respect to λ∗ , that is, their boundaries have λ∗ -measure
0.
Choose U ⊂ ML(S) to be a neighborhood of λ containing all measured laminations µ such
that |λ(A) − µ(A)| < 1 and λ∗ measure and µ∗ measure are δ/2 close on every square Si,j . We
choose δ such that
δ<
6(2m − 1)C
where the constant C is from lemma 3.5. Note that m depends on ε0 which in turn depends on
, ρ, λ, A.
Let l1 , l2 be weighted simple closed curves in U with weights t1 , t2 ∈ R+ respectively. We have
∗
l1 and l2∗ are δ close on Si,j . Since l1 is a simple curve, its lifts are disjoint from each other. So,
suppose Si,j contains the endpoints of a lift of l1 , then Si0 ,j 0 cannot contain endpoints of any lift
of l1 if either i < i0 , j > j 0 or i > i0 , j < j 0 . In other words if both Si,j , Si0 ,j 0 contain endpoints
of lifts of l1 then either i ≤ i0 , j ≤ j 0 or i ≥ i0 , j ≥ j 0 . So there are at most 2m − 1 squares
containing endpoints of lifts of l1 and there is a complete ordering on this set of 2m − 1 squares.
Same statements can be made about l2 or any other geodesic lamination. Let S1 be the ordered set
of squares in {Si,j } which contain the endpoints of lifts of l1 . Similarly we define the set S2 for l2 .
The set S1 ∩ S2 also has a complete ordering compatible with both S1 and S2 .
Note that for S ∈ S1 − (S1 ∩ S2 ), l2∗ (S) = 0, so l1∗ (S) < δ. Similarly for S ∈ S2 − (S1 ∩ S2 ).
Then by lemma 3.5 we have the following:
Y
Y
H(xS , yS , l1∗ (S)) < (2m − 1)Cδ,
dG
H(xS , yS , l1∗ (S)),
S∈S1 ∩S2
S∈S1
Y
dG
Y
H(xS , yS , l1∗ (S)),
S∈S1 ∩S2
S∈S1 ∩S2
dG
H(xS , yS , l2∗ (S)) < (2m − 1)Cδ,
Y
S∈S1 ∩S2
Y
H(xS , yS , l2∗ (S)),
S∈S2
H(xS , yS , l2∗ (S)) < (2m − 1)Cδ.
Here the order in the products are determined by the ordering of S1 , S2 and S1 ∩ S2 . Therefore
we have
Y
Y
(3)
dG
H(xS , yS , l1∗ (S)),
H(xS , yS , l2∗ (S)) < 3(2m − 1)Cδ <
2
S∈S1
S∈S2
DEFORMATIONS OF REPRESENTATIONS AND EARTHQUAKES ON SO(n, 1) SURFACE GROUPS
15
By the choice of δ.
Q
Now we will show that ρ(A)−1 El1 (ρ)(A) is close S∈S1 H(xS , yS , l1∗ (S)) and similarly for
l2 .Suppose (x11 , y11 ), (x12 , y21 ), ..., (x1k1 , yk11 ) (in order) are the pairs of end points in L(I1 ) × L(I2 )
corresponding to the lifts of l1 which intersect the geodesic path from x̃0 to x̃0 A, and let t1 ∈ R+
be the weight of l1 . We have
k1
Y
Y
Y
ρ(A)−1 El1 (ρ)(A) =
H(x1i , yi1 , t1 ) =
H(x1i , yi1 , t1 )
(4)
i=1
S∈S1
(x1i ,yi1 )∈S
On the other hand,
(5)
Y
H(xS , yS , l1∗ (S)) =
Y
S∈S1
S∈S1
Y
H(xS , yS , t1 )
(x1i ,yi1 )∈S
Q
Recall theP
we have the compact set D ⊂ G and that S∈S1 ∪S2 H(xS , yS , tS ) ∈ D for any tS
such that S∈S1 tS < λ(A) + 1. Applying lemma 3.4 and 3.3 we have for any β ∈ D and
(x1i , yi1 ) ∈ S ∈ S1 ,
dG (H(xS , yS , t1 )β, H(x1i , yi1 , t1 )β) < C0 dG (H(xS , yS , t1 ), H(x1i , yi1 , t1 ))
< C0 Kt1 ε0 .
We can now estimate distance between (4) and (5) by replacing the all the terms one by one from
the right, each time adding C0 Kt1 ε0 to the distance. So
k1
Y
X
dG ρ(A)−1 El1 (ρ)(A),
H(xS , yS , l1∗ (S)) <
C0 Kt1 ε0 = C0 Kl1∗ (A)ε0
(6)
i=1
S∈S
1
< C0 Kmax (λ(A) + 1)ε0 < /4
The analogous inequality holds for l2 as well. Therefore from (3) and (6) we have
dG (El1 (ρ)(A), El2 (ρ)(A)) = dG ρ(A)−1 El1 (ρ)(A), ρ(A)−1 El2 (ρ)(A) <
Following [11] we choose a set of generators A1 , ..., A2g for π1 (S) and say that two representations ρ1 , ρ2 are close if dG (ρ1 (Ai ), ρ2 (Ai )) < for each Ai . The induced topology on the space
of representations is independent of the choice of generating set. We have the following corollary
by applying Theorem 3.6 to each generator and taking the finite intersection of open neighborhoods.
Corollary 3.7. For any > 0, λ ∈ ML(S) an Anosov surface group representation ρ into
SO(n, 1), there is a neighborhood U 3 λ such that for any two weighted simple closed curves
l1 , l2 ∈ U , the corresponding representations El1 (ρ) and El2 (ρ) are close.
This implies that for any sequence of weighted simple closed curves (li ) converging to λ ∈
ML(S), the corresponding sequence of representations Eli (ρ) is Cauchy and thus converges. Moreover, the limit of this sequence does not depend on the choice of sequence converging to λ, so we
can define Eλ (ρ) := limi→∞ Eli (ρ) which is also a representation of π1 (S). Thus we have a continuous map from ML(S) to the connected components of representations near ρ.
16
SON LAM HO
R EFERENCES
[1] J. Aramayona and C. J. Leininger, Hyperbolic structures on surfaces and geodesic currents, Lecture notes (2014).
[2] F. Bonahon, The geometry of Teichmüller space via geodesic currents, Invent. Math. 92 (1988), 139–162.
[3] F. Bonahon, Shearing hyperbolic surfaces, bending pleated surfaces and Thurstons symplectic form, Ann. Fac. Sci.
Toulouse Math. 6 (1996), 5(2): 233–297
[4] B. H. Bowditch, Geometrical finiteness with variable negative curvature, Duke Math. J. 77 (1995), no. 1, 229–274.
[5] W. M. Goldman, The symplectic nature of fundamental groups of surfaces, Adv. in Math. 54 (1984), no. 2, 200–225.
[6] W. M. Goldman, Convex projective structures on compact surfaces, J. Differential Geom. 31 (1990), 791–845.
[7] O. Guichard and A. Wienhard, Anosov representations: domains of discontinuity and applications, Invent. math.
(2012) 190: 357. doi:10.1007/s00222-012-0382-7
[8] Y. Kamishima and S. P. Tan, Deformation spaces on geometric structures, Aspects of low-dimensional manifolds,
Kinokuniya, Tokyo (1992), 263–299.
[9] M. Kapovich, Kleinian groups in higher dimensions, Progress in Math. Vol. 265, 485–562.
[10] M. Kapovich and B. Leeb and J. Porti, Lectures on Anosov representations I: Dynamical and geometric characterizations, preprint (2016).
[11] S. P. Kerckhoff, The Nielsen realization problem, The Annals of Math. 117 (1983), no. 2, 235–265.
[12] C. Kourouniotis, Complex length coordinates for quasifuchsian groups, Mathematika, 41 (1994), 173–188.
[13] D. Johnson and J. J. Millson, Deformation spaces associated to compact hyperbolic manifolds, Discrete Groups in
Geometry and Analysis, Vol. 67 Progress in Math., 48–106.
[14] C. McMullen, Complex earthquakes and Teichmuller theory, J. Amer. Math. Soc., 11, (1998), no. 2, 283–320.
[15] S. P. Tan, Complex Fenchel-Nielsen coordinates for quasi-Fuchsian structures, Internat. J. Math 5 (1994), no. 2,
239–251.
[16] S. P. Tan and Y. L. Wong and Y. Zhang, Generalized Delambre-Gauss formulas for oriented, augmented, rightangled hexagons in hyperbolic 4-space, Adv. Math. 230, (2012), 927–956.
[17] W. P. Thurston, The geometry and topology of 3-manifolds, Princeton University Lecture Notes, 1982, online at
http://www.msri.org/publications/books/gt3m
D EPARTMENT OF M ATHEMATICS , U NIVERSIT É DE S HERBROOKE , S HERBROOKE , QC C ANADA .
| 4math.GR
|
Nonparametric intensity estimation from indirect point process
observations under unknown error distribution
Martin Kroll∗
arXiv:1703.05619v1 [math.ST] 16 Mar 2017
January 19, 2018
Synopsis
We consider the nonparametric estimation of the intensity function of a Poisson point process in a circular model from indirect observations N1 , . . . , Nn . These observations emerge
from hidden point process realizations with the target intensity through contamination with
additive error. Under the assumption that the error distribution is unknown and only available by means of an additional sample Y1 , . . . , Ym we derive minimax rates of convergence
with respect to the sample sizes n and m under abstract smoothness conditions and propose
an orthonormal series estimator which attains the optimal rate of convergence. The performance of the estimator depends on the correct specification of a dimension parameter whose
optimal choice relies on smoothness characteristics of both the intensity and the error density. Since a priori knowledge of such characteristics is a too strong assumption, we propose
a data-driven choice of the dimension parameter based on model selection and show that
the adaptive estimator either attains the minimax optimal rate or is suboptimal only by a
logarithmic factor.
Key words and phrases: Poisson point process, intensity function, nonparametric estimation,
minimax theory, adaptive estimation, statistical inverse problem
AMS 2010 Subject Classification: 60G55, 62G05
1. Introduction
Point process models are used in a wide variety of applications, including, amongst others,
stochastic geometry [Sto+13], extreme value theory [Res87], and queueing theory [Bré81]. Each
realization of a point process is a random set of points {xj } which can alternatively be represented
P
as an N0 -valued random measure j δxj where δ• denotes the Dirac measure concentrated at •.
Poisson point processes (PPPs) are of particular importance since they serve as the elementary
building blocks for more complex point process models. Let X be a locally compact second
countable Hausdorff space, X the corresponding Borel σ-field and Λ a locally finite measure on
the measurable space (X, X ), i.e., Λ(C) < ∞ for all relatively compact sets C in X . A random
P
set of points N = {xj } from X (resp. the random measure N = j δxj ) is called Poisson point
process with intensity measure Λ if
• the number NC = |N ∩ C| of points located in C follows a Poisson distribution with
parameter Λ(C) for all relatively compact C ∈ X , and
∗
Universität Mannheim. Institut für Mathematik. A5, 6. D-68131 Mannheim, Germany.
B: [email protected]. Financial support by the Deutsche Forschungsgemeinschaft
(DFG) through the Research Training Group RTG 1953 is gratefully acknowledged. I am indebted to Jan
Johannes and Martin Schlather for helpful comments on the paper.
• for all n ∈ N and disjoint sets A1 , . . . , An ∈ X , the random variables NA1 , . . . , NAn are
independent.
It is well-known that the distribution of a PPP is completely determined by its intensity measure.
Hence, from a statistical point of view, the (non-parametric) estimation of the intensity measure or its Radon-Nikodym derivative (the intensity function) with respect to some dominating
measure from observations of the point process is of fundamental importance.
Inference and testing problems for Poisson and more general point processes have been tackled in a wide range of scenarios. The monographs [Kar91] and [Kut98] offer a comprehensive
overview and discuss both parametric and nonparametric methods. From a methodological
point of view, our approach in this paper is related to the article [RB03] where the estimation
of the intensity function from direct observations was considered. As in the present paper, the
analysis of the adaptive estimator in [RB03] is based on appropriate concentration inequalities for Poisson point processes. Whereas the concentration inequalities developed and applied
in [RB03] represent analogues of results for random variables due to [Mas00], one main tool of
our approach are analogues of concentration inequalities due to [KR05] that we have derived in
a separate manuscript [Kro16]. Other approaches to nonparametric intensity estimation from
direct observations, without making a claim to be exhaustive, can be found in [BB09] (where
the performance of a histogram estimator under Hellinger loss is analysed), [Bir07] (using a testing approach to model selection), [GN00] (using a minimum complexity estimator in the Aalen
model), and [PW04] (suggesting a wavelet estimator in the multiplicative intensity model).
Theoretical work on intensity estimation has recently been motivated by applications to genomic data. The model considered in the article [Big+13] is motivated by data arising throughout the processing of DNA ChIP-seq data. The article [San14] takes its motivation from the
analysis of genomic data as well. In that paper, the author studies nonparametric inference
of a so-called reproduction function from one realization of an aggregated point process and
additional observations related to the model. The observations in that paper are not of the
same type as the ones that we will consider here, although the estimation problem can also
be ascribed to the area of intensity estimation from indirect observations. In addition, let us
mention two further articles where the development of nonparametric statistical methods for the
analysis of point processes was inspired by applications from biology: first, motivated through
DNA sequencing techniques, the article [SZ12] introduces a change-point model for nonhomogeneous Poisson processes occurring in molecular biology. Second, the article [ZK10] considered
the nonparametric inference of Cox process data by means of a kernel type estimator.
e1 , . . . , N
en
Usually one aims to estimate the intensity function λ from direct observations N
where
X
ei =
N
δxij
(1.1)
j
are realizations of a PPP with the target intensity λ.
In this paper, however, we assume that we are interested in the nonparametric estimation
of the intensity function λ without having access to the observations in (1.1). Instead, we are
in the setup of a Poisson inverse problem [AB06] where we can only observe N1 , . . . , Nn given
through
X
δyij .
(1.2)
Ni =
j
ei by the identity yij = xij + εij − ⌊xij +
The indirect observations Ni are related to the hidden N
εij ⌋. The definition of the yij as the fractional part of the additively contaminated xij yields a
circular model by means of the usual topological identification of the interval [0, 1) and the circle
of perimeter 1. Assumptions concerning the intensity and the errors will be discussed below in
more detail.
2
Note that the circular definition of the general model (1.2) is convenient to model the case of
periodic intensity functions which is of particular importance in applications: periodic intensity
functions are suitable to model the occurence of events that are subject to a natural temporal
period (day, week, . . . ), for instance financial transactions or gun crimes. We refer the interested
reader to [HMZ03] for references concerning the wide range of applications.
The model (1.2) is closely related to (circular) deconvolution models. In the circular deconvolution problem (CDP) one aims at estimating the density g of a random variable X with values
in [0, 1) from repeated observations Yi = Xi + εi − ⌊Xi + εi ⌋. Here, εi i.i.d. ∼ f for some error
density f . The CDP and its analogue extension on the real line (where X is real-valued and the
observations are Yi = Xi + εi ) have been treated in a wide range of research articles (cf. [Mei09]
for a comprehensive introduction to the subject).
Note that, in contrast to our approach, the few existing papers on Poisson inverse problems
([CJ02], [AB06], [Big+13]) assume the error distribution to be known. This conservative assumption is also present in most of the research literature dealing with (circular) deconvolution
problems. If the error density is unknown, even identifiability of the statistical model is not
guaranteed. Thus, several remedies have been introduced to overcome this problem: for instance, it is possible to impose additional assumptions on the statistical model (cf., for instance,
the article [SVB10] which deals with blind convolution under additive centred Gaussian noise
with unknown variance). Alternatively, one can consider a framework with panel data [Neu07].
Finally, one can assume the availability of an additional sample from the error density (cf., for
instance, [DH93], [Joh09], [CL10], [CL11]) to guarantee identifiability and enable inference. In
this paper, we will stick to this last option.
Let us assume that the errors εij contributing to the general model (1.2) are stationary in the
sense that εij ∼ f for some unknown error density f . Under this assumption, two models are of
particular interest:
1. the errors εij are i.i.d. ∼ f (this case will be referred to as the Poisson model throughout
the paper),
2. the errors coincide, that is εij ≡ εi ∼ f for all j (this case will be referred to as the Cox
model)
(the names Poisson resp. Cox model will be justified below). Note that in the classical (circular)
deconvolution problem a differentiation between distinct models as in the point process case is
not possible. We will study nonparametric estimation of the intensity function λ under the
Poisson and the Cox model from observations
N1 , . . . , Nn
i.i.d.
and
Y 1 , . . . , Ym
i.i.d. ∼ f
(1.3)
where the Ni are given as in (1.2). A natural aim here is to detect optimal rates of convergence
in terms of the sample sizes n and m and to construct adaptive estimators attaining these rates.
It has to be remarked that the Cox model has already been studied intensively in [Big+13]:
the authors of that paper exclusively consider the case that the error density f is known (making
the Y sample in (1.3) needless) and ordinary smooth, that is, the Fourier coefficients of f are
polynomially decreasing. Note that our investigation in this paper allows to consider error
densities with faster rates of decay. Under the stated assumptions, the article [Big+13] provides
a remarkable proof of a minimax lower bound over Besov spaces and suggests a hard-thresholding
wavelet series estimator which automatically adapts to unknown smoothness and attains a rate
which is optimal up to a logarithmic factor. Moreover the authors show (cf. Theorems 6.1
and 6.2 in [Big+13]) that it is impossible to construct a consistent estimator of the intensity on
the basis of a preceding ’alignment step’.
From a methodological point of view, our approach is inspired by the one conducted in [JS13].
In contrast to that paper, the proof of the minimax lower bound in our setup has to deal with
3
the different nature of the observations. The key argument here is that the Hellinger distance
between Poisson point processes can be bounded from above by the Hellinger distance between
the corresponding intensity measures. We consider orthonormal series estimators of the form
X
b =
λ
k
0≤|j|≤k
c e
[λ]
j j
(1.4)
c is an appropriate estimator of the Fourier coefficient [λ]
where ej (·) = exp(2πij·) and [λ]
j
j
corresponding to the basis function ej (·) (see Section 2 for details). Of course, this estimator
P
is motivated by the L2 -convergent representation λ = j∈Z [λ]j ej for square-integrable λ. It
b crucially depends on the choice of the
will turn out that the performance of the estimator λ
k
dimension parameter k and that its optimal value depends on smoothness characteristics of the
intensity that are usually not available in practice. In order to choose k in a completely datadriven manner, we follow an approach based on model selection (cf., for instance, [BBM99] or
[Com15]) and select the dimension parameter as the minimizer of a penalized contrast criterion,
b ) + pen },
kb := argmin {Υ(λ
k
k
0≤k≤Knm
where Knm ∈ N0 (depending both on the samples sizes n, m ∈ N and the random observations)
determines the set of admissible models, Υ is a contrast function and penk a penalty term
which will be specified in more detail below. Our choice of the penalty term is non-deterministic
but random already in the partially adaptive case where one assumes only the smoothness of
the intensity to be unknown but the smoothness of the error density to be known. For the
theoretical analysis of the adaptive estimator we need Talagrand type concentration inequalities
tailored to the framework with PPP observations which cannot be directly transferred from
results applied in the usual density estimation or deconvolution frameworks (cf. Remark 2.3
in [Kro16]). As already mentioned above, these inequalities have already been derived in a
separate manuscript [Kro16], and only the specific results needed for our application in this
paper are stated in the appendix.
The article is organized as follows: in Section 2 we introduce our methodological approach and
in Section 3 we consider our nonparametric estimation problems from a minimax point of view.
Then, Section 4 considers adaptive estimation of the intensity for the Poisson model, whereas
Section 5 deals with adaptive estimation for the Cox model.
2. Methodology
2.1. Notation
Throughout this work we will assume that the intensity λ and the density f belong to the space
L2 := L2 ([0, 1), dx) of square-integrable functions on the interval [0, 1). Let {ej }j∈Z be the
complex trigonometric basis of L2 given by ej (t) = exp(2πijt). The Fourier coefficients of a
function g ∈ L2 are denoted as follows:
[g]j =
Z
0
1
g(t)ej (−t)dt.
For a strictly positive symmetric sequence ω = (ωj )j∈Z we introduce the weighted norm k · kω
P
defined via kgk2ω := j∈Z ωj |[g]j |2 . The corresponding scalar product is denoted with hg, hiω =
P
j∈Z ωj [g]j [h]j . Throughout the paper, we use the notation a(n, m) . b(n, m) if a(n, m) ≤
C · b(n, m) for some numerical constant C independent of n and m.
4
2.2. The minimax point of view
e of λ by means of the mean inWe will evaluate the performance of an arbitrary estimator λ
e − λk2 ]. We will take up the minimax point of view and
tegrated weighted squared loss E[kλ
ω
consider the maximum risk defined by
e − λk2 ]
sup sup E[kλ
ω
(2.1)
λ∈Λ f ∈F
where Λ and F are classes of potential intensity functions λ and densities f , respectively. Note
that in the case of a known error density f one would only consider the supremum over all λ ∈ Λ
for this fixed error density. If one is able to show an upper risk bound which holds uniformly
for all f belonging to some class F, one obtains an upper bound for the maximum risk in (2.1).
In order to show minimax lower bounds in terms of the sample size m in (1.3), it will turn out
sufficient to reduce the class F even to a subset of two elements (see the proof of Theorem 3.3).
The minimax risk is defined via
e − λk2 ]
inf sup sup E[kλ
ω
e
λ
λ∈Λ f ∈F
e of λ. An estimator λ∗ is called rate optimal if
where the infimum is taken over all estimators λ
e − λk2 ].
sup sup E[kλ∗ − λk2ω ] . inf sup sup E[kλ
ω
e
λ
λ∈Λ f ∈F
λ∈Λ f ∈F
The classes Λ of intensity functions and F of densities to be considered in this article will be
specified in Section 3 below where we derive lower bounds on the minimax risk for these specific
choices and prove that this lower bound is attained up to a numerical constant by a suitably
defined orthonormal series estimator.
2.3. Sequence space representation for the Poisson model
The orthogonal series estimator that we will introduce in the next subsection is crucially motivated by a sequence space representation of the considered models. This representation forms
c of the Fourier coefficients in (1.4). We
the point of origin for the definition of the estimators [λ]
j
address ourselves now to the derivation of this sequence space model.
Under the Poisson model, the observed point processes N1 , . . . , Nn in (1.3) are generated
e1 , . . . , N
en with intensity function λ by independent
from independent Poisson point processes N
random contaminations of the individual points. We emphasize again that the (unobserved)
contaminations are assumed to follow a probability law given by an unknown density f and are
to be understood additively modulo 1. Thus, the observations Ni under the Poisson model are
given by
X
Ni =
δxij +εij −⌊xij +εij ⌋
j
P
ei =
where N
j δxij is the realization of a Poisson point process with intensity function λ and
the errors εij are i.i.d. ∼ f . Note that under the Poisson model each Ni is again a realization
of a Poisson point process whose intensity function is given by the circular convolution λ ⋆ f
modulo 1 of λ with the error density f . More precisely, ℓ = λ ⋆ f is given by the formula
ℓ(t) =
Z
0
1
λ((t − ε) − ⌊t − ε⌋)f (ε)dε,
t ∈ [0, 1).
(2.2)
By the convolution theorem, we have [ℓ]j = [λ]j · [f ]j for all j ∈ Z. From Campbell’s theorem
5
(cf. [Ser09], Chapter 3, Theorem 24) it can be deduced that for measurable functions g we have
Z
E
1
0
g(t)dNi (t) =
Z
1
ℓ(t)g(t)dt
0
provided that the integral on the right-hand side exists. Exploiting this equation for g(t) =
ej (−t) and setting
n Z 1
1X
c
ej (−t)dNi (t)
(2.3)
[ℓ]j :=
n i=1 0
c = [λ] · [f ] for all j ∈ Z. More precisely, we have
we thus obtain that E[ℓ]
j
j
j
c = [λ] · [f ] + ξ
[ℓ]
j
j
j
j
for all j ∈ Z
(2.4)
with centred random variables
n Z
1X
c
c
ξj = [ℓ]j − E[ℓ]j =
n
i=1
1
ej (−t)dNi (t) −
0
Z
0
1
ℓ(t)ej (−t)dt .
2.4. Sequence space representation for the Cox model
The Cox model permits a sequence space representation similar to the one of the Poisson
model which was obtained in [Big+13]. Under the Cox model, the independent point processes N1 , . . . , Nn are generated by a random contamination of the individual points of the
e1 , . . . , N
en as for the Poisson model. In contrast to the
unobservable Poisson point processes N
ei ,
Poisson model, however, the additive error is the same for all the points of one realization N
that is the observations Ni are given by
Ni =
X
j
δxij +εi −⌊xij +εi ⌋
(2.5)
P
ei =
where N
j δxij is the realization of a Poisson point process with intensity function λ and
εi ∼ f . Alternatively, the generation of the point processes Ni can be described by the following
two-step procedure: in the first step, random shifts εi ∼ f are generated. In the second step,
conditionally on εi , Ni is drawn as the realization of a Poisson point process on [0, 1) whose
intensity function is λ(t − εi − ⌊t − εi ⌋). Thus, in this second model the observations follow the
distribution of a Cox process which is directed by the random measure with random intensity
λ(t − ε − ⌊t − ε⌋) for ε ∼ f . Note that for i = 1, . . . , n and integrable functions g we have
E
Z
0
1
g(t)dNi (t) | εi =
Z
1
0
g(t)λ(t − εi − ⌊t − εi ⌋)dt,
which implies
E
Z
0
1
g(t)dNi (t) =
Z
0
1
g(t)
Z
1
0
λ(t − ε − ⌊t − ε⌋)f (ε)dεdt =
Z
1
g(t)ℓ(t)dt,
0
where ℓ = λ ⋆ f denotes the circular convolution of the function λ and the density f as in (2.2).
Thus, the mean measure of a generic realization N obeying the Cox model has the RadonNikodym derivative ℓ with respect to the Lebesgue measure. Note that the mean measures of
the observed point processes under the Poisson and the Cox model coincide, but the observations
c defined as in (2.3)
for the Cox model stem from a Cox instead of a Poisson process. With [ℓ]
j
6
the relation
f] =
holds with [f
j
model:
P
1
n
Pn
c | ε , . . . , ε ] = [λ] · [f
f]
E[[ℓ]
1
n
j
j
j
i=1 ej (−εi ).
Thus, we get the following representation as a sequence space
c = [λ] · [f
f] + ξ ,
[ℓ]
j
j
j
j
R
(2.6)
R
where ξj = n1 ni=1 [ 01 ej (−t)dNi (t)− 01 ej (−t)λ(t−εi −⌊t−εi ⌋)dt] are centred random variables
for all j ∈ Z. The connection between the sequence space model at hand and the stated sequence
space model formulation for statistical linear inverse problems is discussed in detail in Section
2.1 of [Big+13].
2.5. Orthonormal series estimator
Recall that each λ ∈ L2 can be represented by its Fourier series representation as λ =
Hence, in view of (2.4) and (2.6), a natural estimator of λ is given by
b =
λ
k
c
[ℓ]
j
X
c
0≤|j|≤k [f ]j
1 Ωj e j
P
j∈Z [λ]j ej .
(2.7)
P
m
c as defined in (2.3), [f
c] := 1
c 2
with [ℓ]
i=1 ej (−Yi ) and Ωj := {|[f ]j | ≥ 1/m}. Note that
j
j
m
[f ]j in (2.4) is not directly available and thus has to be estimated from the sample Y1 , . . . , Ym
f] in (2.6) cannot be observed. In contrast to the setup
in (1.3). For the Cox model, note that [f
j
f] cannot be substituted
in [Big+13] we do not assume the density f to be known and hence [f
j
by its expectation [f ]j which was suggested in [Big+13]. Since we have the additional sample
f] by [f
c] which
Y1 , . . . , Ym from f in (1.3) at hand, the most natural idea seems to substitute [f
j
j
leads to the same estimator as for the Poisson model. The additional threshold occurring in
b through the indicator function over the set Ω compensates for ’too small’
the definition of λ
j
k
c
absolute values of [f ]j and is imposed in order to avoid unstable behaviour of the estimator. The
definition of the event Ωj is in accordance with [NH97] and the chosen threshold corresponds
to the parametric rate at which [f ]j can be estimated from the sample Y1 , . . . , Ym . The optimal
choice kn∗ of the dimension parameter in the minimax framework will be determined in Section 3
b ∗ is not completely data-driven
and depends on the classes Λ and F. Thus, the estimator λ
kn
which is unacceptable in applications where the degree of smoothness of the functions λ and f
will usually not be available in advance. The data-driven choice of the dimension parameter is
discussed in Sections 4 and 5.
3. Minimax theory
3.1. Model assumptions
Let γ = (γj )j∈Z and α = (αj )j∈Z be strictly positive symmetric sequences and fix r > 0, d ≥ 1.
In this section, we derive minimax rates of convergence concerning the maximum risk defined
in (2.1) with respect to the classes
Λrγ := {λ ∈ L2 : λ ≥ 0 and
and
X
j∈Z
γj |[λ]j |2 =: kλk2γ ≤ r}
Fαd := {f ∈ L2 : f ≥ 0, [f ]0 = 1 and d−1 ≤ |[f ]j |2 /αj ≤ d}
7
of intensity functions and error densities, respectively. The regularity conditions imposed on the
sequences γ and α are summarized in the following assumption.
Assumption A γ = (γj )j∈Z , α = (αj )j∈Z and ω = (ωj )j∈Z are strictly positive symmetric
sequences such that γ0 = ω0 = α0 = 1, γj ≥ 1 for all j ∈ Z and the sequences (ωn /γn )n∈N0 and
P
(αn )n∈N0 are both non-increasing. Finally, ρ := j∈Z αj < ∞.
Note that in the special case ω ≡ 1, that is, the weighted norm k · kω coincides with the usual
the additional assumption γj ≥ 1 is contained in the requirement that the sequence
(ωn /γn )n∈N0 is non-increasing.
L2 -norm,
3.2. Minimax lower bounds for the Poisson model
We first derive lower bounds in terms of the sample sizes n and m in (1.3) for Poisson model.
Our first theorem provides such a lower bound Ψn in terms of the sample size n under mild
assumptions.
Theorem 3.1 For n ∈ N, set
ω
ωj
,
kn∗ := argmin max
γk
nαj
k∈N0
0≤|j|≤k
X
k
and
Ψn := max
Let Assumption A hold, and further assume that
(C1) Γ :=
P
−1
j∈Z γj
< ∞, and
(C2) 0 < η −1 = inf n∈N Ψ−1
n · min
Then, for any n ∈ N,
ωk ∗ P
ωj
n
∗
0≤|j|≤kn
γk∗ ,
nαj
n
e − λk2 ] ≥
inf sup sup E[kλ
ω
e
λ
λ∈Λrγ
f ∈Fαd
ω
∗
kn
γkn∗
,
X
∗
0≤|j|≤kn
ωj
.
nαj
(3.1)
for some 1 ≤ η < ∞.
ζr
· Ψn
16η
1
1
e of
√ } with δ = 1 − √
where ζ = min{ 2Γdη
, d2δ
, and the infimum is taken over all estimators λ
2
r
2 2
λ based on the observations from (1.3) under the Poisson model.
Before proceeding with the lower bound in m, let us state some remarks concerning Theorem 3.1. Firstly, it follows immediately from the proof that the lower bound Ψn holds already
P
in case of a known error density. Secondly, assuming the convergence of the series j∈Z γj−1
through condition (C1) is necessary only in order to establish the non-negativity of the hypotheses λθ considered in the proof. Finally, it is noteworthy that the proof of Theorem 3.1 can be
adapted to the case of direct observations of Poisson point processes with intensity function λ
(cf. [Kro16], Theorem 3.3 where the corresponding lower bound is obtained by replacing all the
αj with 1).
Remark 3.2 The key ingredient for the proof of Theorem 3.1 is the fact that the Hellinger
distance between two PPPs is bounded by the Hellinger distance of the corresponding intensity
measures. Since this relation holds only for the special case of Poisson processes, it cannot be
directly extended to the Cox model and the given proof fails. Thus, the derivation of minimax
lower bounds based on Fourier expansions for the Cox model in our framework remains open
and needs to be addressed in future research. Note that [Big+13] give a lower bound proof in
case of a known and ordinary smooth error density f .
The following theorem provides a lower bound Φm in terms of the sample size m.
8
Theorem 3.3 For m ∈ N, set
Φm := max
j∈N
(
1
1,
mαj
ωj γj−1 min
!)
.
(3.2)
Let Assumption A hold, and in addition assume that
√
(C3) there exists a density f in Fα d with f ≥ 1/2.
Then, for any m ∈ N,
√
e − λk2 ] ≥ 1 − 3/2 · ζ 2 rd−1/2 · Φm
inf sup sup E[kλ
ω
8
e
λ λ∈Λrγ f ∈Fαd
1
e of λ based on the
, 1 − d−1/4 } and the infimum is taken over all estimators λ
where ζ = min{ 4√
d
observations from (1.3) under the Poisson model.
Let us state some remarks concerning Theorem 3.3. First, for the proof of the theorem it
was sufficient to construct two hypotheses which are statistically indistinguishable but already
establish the lower bound Φm . This is in contrast to the proof of Theorem 3.1 where we had
∗
to construct 2kn +1 hypotheses. Second, the condition (C3) has to be imposed in order to
guarantee that the hypotheses fθ considered in the proof
belong to Fαd . It is easy to check that
√
P
condition (C3) is satisfied if ρ0 := j6=0 αj satisfies d ≥ max 4ρ20 , 1 .
Remark 3.4 The proof of Theorem 3.3 cannot be transferred to the Cox model neither. In the
proof, the equality λ−1 ⋆ f−1 = λ1 ⋆ f1 would imply only the equality of the mean measures of the
two Cox process hypotheses but not equality of their distributions (which holds true for PPPs).
The next corollary is an immediate consequence of Theorems 3.1 and 3.3.
Corollary 3.5 Under the assumptions of Theorems 3.1 and 3.3, for any n, m ∈ N,
e − λk2 ] & Ψ + Φ .
inf sup sup E[kλ
n
m
ω
3.3. Upper bound
e
λ
λ∈Λrγ f ∈Fαd
Let us now establish an upper bound for the maximum risk in terms of n and m for the estimator
b in (2.7) under a suitable choice of the dimension parameter k. More precisely, the following
λ
k
b ∗ with k ∗ defined in (3.1)
theorem establishes an upper bound for the rate of convergence of λ
kn
n
for the Poisson and the Cox model (an analysis of the proofs indicates only a slightly different
numerical constant). Thus, due to the lower bound proofs in the preceding subsection it is shown
b ∗ attains the minimax rates of convergence in terms of the samples sizes n and m under
that λ
kn
the Poisson model. Note that the optimal choice of the dimension parameter does not depend
on the sample size m.
Theorem 3.6 Let Assumption A hold and further assume that the samples N1 , . . . , Nn and
Y1 , . . . , Ym in (1.3) are drawn in accordance with the Poisson or the Cox model. Then, for any
n, m ∈ N,
b ∗ − λk2 ] . Ψ + Φ .
sup sup E[kλ
n
m
kn
ω
λ∈Λrγ f ∈Fαd
9
γ
α
Θ(Ψn )
m−
Restrictions
(p−s)∧a
a
(pol)
(pol)
(exp)
(pol)
(log n)2s+2a+1 · n−1
m−1
(pol)
(exp)
(log n)−2(p−s)
(log m)−2(p−s)
(exp)
(exp)
n
Θ(Φm )
2(p−s)
− 2p+2a+1
p
(log n)2s · n− p+a
(log m)
2s
m
·m
−p/a
−1
p ≥ s, a >
a>
if a ≥ p
1
2
1
2
p≥s
if a < p
Table 1: Exemplary rates of convergence for nonparametric intensity estimation. The rates are given in the
framework of Theorems 3.1, 3.3 and 3.6 which impose the given restrictions. In all examples ω0 = 1,
ωj = |j|2s for j 6= 0, whereas the choices (pol) and (exp) for the sequences γ and α are explained in
Section 3.4.
3.4. Examples of convergence rates
In order to flesh out the abstract results of this section, we consider special choices for the
sequences ω, γ and α and state the resulting rates of convergence with respect to both sample
sizes n and m. For the sequence ω, we will assume throughout that ω0 = 1 and ωj = |j|2s for
j 6= 0. The resulting weighted norm corresponds to the usual L2 -norm of the sth weak derivative.
Choices for the sequence γ:
narios:
Concerning the sequence γ we distinguish the following two sce-
(pol): γ0 = 0 and γj = |j|2p for all j =
6 0 and some p ≥ 0. This corresponds to the case
when the unknown intensity function belongs to some Sobolev space.
(exp): γj = exp(2p|j|) for all j ∈ Z and some p ≥ 0. In this case, λ belongs to some space
of analytic functions.
Choices for the sequence α:
Concerning the sequence α we consider the following scenarios:
(pol): α0 = 0 and αj = |j|−2a for all j 6= 0 and some a ≥ 0. This corresponds to the case
when the error density is ordinary smooth.
(exp): αj = exp(−2a|j|) for all j ∈ Z and some a ≥ 0.
Table 1 summarises the rates Ψn and Φm corresponding to the different choices of γ and α.
The rates with respect to n coincide with the classical rates for nonparametric inverse problems
(see for instance Table 1 in [Cav08] where the error variance ε2 corresponds to n−1 in our setup
and only the case s = 0 is considered).
4. Adaptive estimation for the Poisson model
The estimator considered in Theorem 3.6 is obtained by specializing the orthonormal series
estimator in (2.7) with kn∗ defined in (3.1). Thus, this procedure suffers from the apparent
drawback that it depends on the smoothness characteristics of both λ and f , namely on the
sequences γ and α. Since such characteristics are typically unavailable in advance, there is a
need for an adaptive selection of the dimension parameter which does not require any a priori
knowledge on λ and f . In order to reach such an adaptive definition for the Poisson model (the
Cox model will be dealt with in Section 5) we proceed in two steps. In the first step (treated in
Subsection 4.1), we assume that the class Λrγ is unknown but assume the class Fαd of potential
10
error densities f to be known. This assumption allows us to define a partially adaptive choice
ke of k. In the second step (treated in Subsection 4.2), we dispense with any knowledge on the
smoothness both of λ and f and propose a fully adaptive choice kb of the dimension parameter.
4.1. Partially adaptive estimation (Λrγ unknown, Fαd known)
In this subsection, we aim at choosing k equal to some ke that, in contrast to kn∗ in Section 3,
does no longer depend on the sequence γ but only on the sequence α. For the definition of ke
some terminology has to be introduced: for any k ∈ N0 , let
∆αk := max ωj α−1
j
0≤j≤k
and δkα := (2k + 1)∆αk
log(∆αk ∨ (k + 3))
.
log(k + 3)
For all n, m ∈ N, setting ωj+ := max0≤i≤j ωi , we define
Nnα
(
log(n + 3)ωj+
αj
:= inf 1 ≤ j ≤ n :
<
2j + 1
n
)
− 1 ∧ n,
α
:= inf{1 ≤ j ≤ m : αj < 640dm−1 log(m + 1)} − 1 ∧ m,
Mm
α := N α ∧ M α . Now, consider the contrast
and set Knm
n
m
b
Υ(t) := ktk2ω − 2ℜhλ
n∧m , tiω ,
t ∈ L2 .
g k )k∈N0 via
and define the random sequence of penalties (pen
g k :=
pen
δα
165 −1 c
dη · ([ℓ]0 ∨ 1) · k ,
4
n
where η ∈ (0, 1) is some additional tuning parameter. The parameter η finds its way into
the upper risk bound only as a numerical constant and does not have any effect on the rate
of convergence. The dependence of the adaptive estimator on the specific choice of η will be
suppressed for the sake of convenience in the sequel. Building on our definition of contrast and
penalty, we define the partially adaptive selection of the dimension parameter k as
b ) + pen
g k }.
ke := argmin {Υ(λ
k
α
0≤k≤Knm
b .
The following theorem provides an upper bound for the partially adaptive estimator λ
e
k
Theorem 4.1 Let Assumption A hold. Then, for any n, m ∈ N,
b − λk2 ] .
sup sup E[kλ
ω
e
k
r
d
λ∈Λγ f ∈Fα
ωk δkα
,
min α max
0≤k≤Knm
γk n
+ Φm +
1
1
+
m n
where the observations in (1.3) stem from the Poisson model.
4.2. Fully adaptive estimation (Λrγ and Fαd unknown)
We now also dispense with the knowledge of the smoothness of the error density f and propose
b
a fully data-driven selection kb of the dimension parameter such that the resulting estimator λ
b
k
adapts to the unknown smoothness of both λ and f and attains the optimal rate of convergence
in a wide range of scenarios. As in the case of partially adaptive estimation, we have to introduce
11
some notation first. For k ∈ N0 , let
b := max
∆
k
0≤j≤k
For n, m ∈ N, set
ωj
c] |2
|[f
j
1 Ωj
b
δbk := (2k + 1)∆
k
and
b ∨ (k + 4))
log(∆
k
.
log(k + 4)
c] |2 /(2j + 1) < log(n + 4)ω + /n} − 1 ∧ n,
bn := inf{1 ≤ j ≤ n : |[f
N
j
j
c] |2 < m−1 log(m)} − 1 ∧ m,
cm := inf{1 ≤ j ≤ m : |[f
M
j
cnm := N
bn ∧ M
cm . We consider the same contrast function as in the partially adaptive case
and K
d k )k∈N0 of penalities now by
but define the random sequence (pen
c ∨ 1) ·
d k := 1375η −1 · ([ℓ]
pen
0
δbk
.
n
Note that this definition does not depend on the knowledge of the sequence α. Using this
definition of a completely data-driven penalty, we define the fully adaptive selection kb of the
dimension parameter k by means of
b ) + pen
d k }.
kb := argmin {Υ(λ
k
bnm
0≤k≤K
b , we have to introduce some
In order to state and prove the upper risk bound of the estimator λ
b
k
α
further notation. We keep the definition of ∆k from Subsection 4.1 but slightly redefine δkα as
δkα := (2k + 1)∆αk
log(∆αk ∨ (k + 4))
.
log(k + 4)
For k ∈ N0 , we also define
ωj
0≤j≤k |[f ]j |2
∆k := max
and
δk := (2k + 1)∆k
log(∆k ∨ (k + 4))
,
log(k + 4)
which can be regarded as analogues of ∆αk and δkα in Subsection 4.1 in the case of a known error
density f . Finally, for n, m ∈ N, define
Nnα− := inf{1 ≤ j ≤ n : αj /(2j + 1) < 4d log(n + 4)ωj+ /n} − 1 ∧ n,
Nnα+ := inf{1 ≤ j ≤ n : αj /(2j + 1) < log(n + 4)ωj+ /(4dn)} − 1 ∧ n,
α−
:= inf{1 ≤ j ≤ m : αj < 4dm−1 log m} − 1 ∧ m,
Mm
α+
:= inf{1 ≤ j ≤ m : 4dαj < m−1 log m} − 1 ∧ m,
Mm
α− := N α− ∧ M α− , K α+ := N α+ ∧ M α+ . In contrast to the proof of Theorem 4.1 we
and set Knm
m
n
n
m
nm
b :
have to impose an additional assumption for the proof of an upper risk bound of λ
b
k
−5 for all m ∈ N.
Assumption B exp(−mαMm
α+
+1 /(128d)) ≤ C(α, d)m
Under this additional assumption, the following theorem establishes a uniform upper risk
bound for the completely data-driven estimator.
Theorem 4.2 Let Assumptions A and B hold. Then, for any n, m ∈ N,
b − λk2 ] .
sup sup E[kλ
ω
b
k
r
d
λ∈Λγ f ∈Fα
ωk δkα
,
min max
α−
γk n
0≤k≤Knm
12
+ Φm +
1
1
+
m n
where the observations in (1.3) stem from the Poisson model.
Note that the only additional prerequisite of Theorem 4.2 in contrast to 4.1 is the validity of
Assumption B.
4.3. Examples of convergence rates (continued from Subsection 3.4)
We consider the same configurations for the sequences ω, γ and α as in Subsection 3.4. In
particular, we assume that ω0 = 1 and ωj = |j|2s for all j 6= 0. The different configurations for γ
and α will be investigated in the following (compare also with the minimax rates of convergence
Assumption B is satisfied in all the considered cases.
given in Table 1). Note that the additional
n
αo
ωk δk
⋄
:=
argmink∈N0 max γk , n , that is, kn⋄ realizes the best compromise between
Let us define kn
squared bias and penalty.
Scenario (pol)-(pol):
1
1
In this scenario, it holds kn⋄ ≍ n 2p+2a+1 and Nnα− ≍ (n/ log n) 2s+2a+1 .
2(p−s)
α− . In case that s < p, the rate w.r.t. n is n− 2p+2a+1 which is the
First assume that Nnα− ≤ Mm
2(p−s)
minimax optimal rate. In case that s = p, it holds Nnα− kn⋄ and the rate is (n/ log n)− 2p+2a+1
α− ≤ N α− . If
which is minimax optimal up to a logarithmic factor. Assume now that Mm
n
α− , then the estimator obtains the optimal rate with respect to n. Otherwise, M α− ≍
kn⋄ . Mm
m
p−s
(m/ log m)1/(2a) yields the contribution (m/ log m)− a to the rate.
Scenario (exp)-(pol): Nnα− ≍ (n/ log n)1/(2a+2s+1) as in scenario (pol)-(pol). Since kn⋄ ≍ log n,
α− .
it holds kn⋄ . Nnα− and the optimal rate with respect to n holds in case that kn⋄ . Mm
2s
α−
α−
Otherwise, the bias-penalty tradeoff generates the contribution (Mm ) · exp(−2p · Mm ) to the
rate.
Scenario (pol)-(exp): It holds that kn⋄ ≍ Nnα− and again the sample size n is no obstacle for
α− , the optimal rate holds as well. If
attaining the optimal rate of convergence. If kn⋄ . Mm
α− k ⋄ , we get the rate (log m)−2(p−s) which coincides with the optimal rate with respect to
Mm
n
the sample size m.
Scenario (exp)-(exp): We have Nnα− ≍ log n and k1 ≤ kn⋄ ≤ k2 where k1 is the solution of
k12 exp((2a + 2p)k1 ) ≍ n and k2 the solution of exp((2a + 2p)k2 ) ≍ n. Thus, we have kn⋄ Nnα−
ω
and computation of γkk1 resp. δkα2 /n shows that only a loss by a logarithmic factor can occur as
1
α− . If M α− ≤ k ⋄ , the contribution to the rate arising from the trade-off
far as kn⋄ ≤ Nnα− ∧ Mm
m
n
α− )2s · exp(−2pM α− ) which deteriorates
between squared bias and penalty is determined by (Mm
m
the optimal rate with respect to m at most by a logarithmic factor.
Remark 4.3 In this paper, we have not considered the case that the Fourier coefficients of the
error density obey a power-exponential decay, that is αj = exp(−2κ|j|a ) for some κ > 0 and
α+ , Assumption B is not satisfied in
arbitrary a > 0. Indeed, for our definition of the quantity Mm
this case. This shortage can be removed by considering a more elaborate choice of the quantities
cm and M α+ as was considered in [JS13] but we do not include this here.
M
m
5. Adaptive estimation for the Cox model
Unfortunately, the approach from Section 4 cannot be transferred in order to obtain an upper
risk bound for an adaptive estimator in the case of Cox observations. Thus, in this section we
follow another approach. The price we have to pay is that we can only obtain rates that contain
13
an additional logarithmic factor. Again we split our investigation into the partially adaptive
and the fully adaptive case.
5.1. Partially adaptive case
P
ω
We define Dαk := 0≤|j|≤k αjj which might be interpreted as the dimension of the model corresponding to the linear subspace of L2 spanned by the ej with j ∈ {−k, . . . , k} in the considered
α , and K α as well as the contrast
inverse problem. In addition, we define the quantities Nnα , Mm
nm
function Υ exactly as in Section 4.1. However, we replace the definition of the penalty given for
the case of Poisson observations with
c ∨ 1) ·
g k := 2000η −1 · ([ℓ]
pen
0
α
dDαk log(n + 2)
c 2 ∨ 1) · dDk log(n + 2)
+ 2000η −2 · ([ℓ]
0
n
n
where η ∈ (0, 1). Based on this updated definition of penalty we define the adaptive choice of
the dimension parameter in the case of Cox observations by means of
b ) + pen
g k }.
ke := argmin {Υ(λ
k
α
0≤j≤Knm
Theorem 5.1 Let Assumption A hold. Then, for any n, m ∈ N,
b − λk2 ] .
sup sup E[kλ
ω
e
k
r
d
λ∈Λγ f ∈Fα
ωk Dαk log(n + 2)
,
min α max
0≤k≤Knm
γk
n
+ Φm +
1
1
+
m n
where the observations in (1.3) stem from the Cox model.
Note that the proof of Theorem 5.1 is more intricate than the one of Theorem 4.1 due to the
fact that we need to introduce some additional terms in the proof. In order to deal with these
terms we have to apply Talagrand type concentration inequalities both for Poisson processes
and random variables.
5.2. Fully adaptive case
In the fully adaptive case, we replace the dimension parameter Dαk from the previous subsection
by the estimate
X
ωj
b :=
D
1 Ωj .
k
c
2
0≤|j|≤k |[f ]j |
Moreover, as in the Poisson case, we have to adapt the definition of the penalty
c ∨ 1) ·
d k := 8000η −1 · ([ℓ]
pen
0
b
b log(n + 2)
D
k
c 2 ∨ 1) · Dk log(n + 2) .
+ 8000η −2 · ([ℓ]
0
n
n
We define the contrast function Υ exactly as in Subsection 4.1. For n, m ∈ N, set
c] |2 /(2j + 1) < log(n + 3)ω + /n} − 1 ∧ n
bn := inf{1 ≤ j ≤ n : |[f
N
j
j
c] |2 < m−1 log(m)} − 1 ∧ m,
cm := inf{1 ≤ j ≤ m : |[f
M
j
b of k in analogy to the approach
cnm := N
bn ∧ M
cm . We define the fully data-driven choice k
and K
for the Poisson model via
b ) + pen
d k }.
kb := argmin {Υ(λ
k
bnm
0≤k≤K
14
For the statement and the proof of the following theorem, define for n, m ∈ N the quantities
Nnα− := inf{1 ≤ j ≤ n : αj /(2j + 1) < 4d log(n + 3)ωj+ /n} − 1 ∧ n,
Nnα+ := inf{1 ≤ j ≤ n : αj /(2j + 1) < log(n + 3)ωj+ /(4dn)} − 1 ∧ n,
α−
:= inf{1 ≤ j ≤ m : αj < 4dm−1 log m} − 1 ∧ m,
Mm
α+
:= inf{1 ≤ j ≤ m : 4dαj < m−1 log m} − 1 ∧ m,
Mm
α− := N α− ∧ M α− , and K α+ := N α+ ∧ M α+ . Note that the proof of the following theorem
Knm
m
n
nm
m
n
requires the validity of Assumption B again.
Theorem 5.2 Let Assumptions A and B hold. Then, for any n, m ∈ N,
b − λk2 ] .
sup sup E[kλ
ω
b
k
r
d
λ∈Λγ f ∈Fα
min α− max
0≤k≤Knm
ωk Dαk log(n + 2)
,
γk
n
+ Φm +
1
1
+
m n
where the observations in (1.3) stem from the Cox model.
Remark 5.3 Using the approach presented in this subsection we are not able to dispense with
the additional logarithmic factor in the rates in case of the Cox model. Note that in case that
the error density f is known (which is informally equivalent to m = ∞) we regain the adaptive
rate established in [Big+13] for the case that the unknown intensity is ordinary smooth and the
Fourier coefficients of f obey a polynomial decay. However, our results are more general since
we do not exclusively consider the case of polynomially decreasing Fourier coefficients.
Remark 5.4 Needless to say, the numerical constants in the definition of the penalty are ridiculously large which makes our rate optimal estimator hardly applicable for small sample sizes.
Hence there is still research necessary to establish an estimator which performs well both from a
theoretical point of view and also yields good results for simulations with relatively small sample
sizes. Another approach would be to calibrate numerical constants in the penalty by means of
a simulation study as way done for example in [CRT06].
Remark 5.5 Of course, the approach presented in this subsection can also be applied to the
case of Poisson observations but since the logarithmic factor in the rates is unavoidable we would
obtain worse rates than using the approach from Section 4.
5.3. Examples of convergence rates (continued from Subsections 3.4 and 4.3)
Note that in all the scenarios considered in Table 1 we have kn⋄ . Nnα− where kn⋄ denotes the
optimal trade-off between the squared bias ωk /γk and the term Dαk log(n + 2)/n. Computations
similar to the ones leading to the rates in Table 1 show that the rates with respect to the sample
size n are those from the minimax framework in Table 1 with n replaced with n/ log(n + 2)
α− . If M α− ≤ k ⋄ , M α− determines the rate exactly with the same
as long as kn⋄ ≤ Nnα− ∧ Mm
m
n
m
contribution as in 4.3. It seems noteworthy to mention that in the scenario (pol)-(exp) we attain
the upper bound from the minimax theory in Theorem 3.6. In the case (pol)-(pol), however, one
can observe a loss by a logarithmic factor which was also observed in case of adaptive estimation
in [Big+13].
15
A. Proofs of Section 3
A.1. Proof of Theorem 3.1
∗
Let us define ζ as in the statement of the theorem and for each θ = (θj )0≤j≤kn∗ ∈ {±1}kn +1 the
function λθ through
λθ :=
=
1/2
r
4
1/2
r
4
+ θ0
+
rζ
4n
rζ
4n
1/2
1/2
+
rζ
4n
X
1/2
X
∗
1≤|j|≤kn
−1/2
∗
0≤|j|≤kn
−1/2
θ|j| αj
θ|j| αj
ej
ej .
Then each λθ is a real-valued function by definition which is non-negative since we have
rζ
4n
1/2
X
∗
0≤|j|≤kn
−1/2
θ|j|αj
ej
∞
≤
rζ
4n
1/2
≤
rζ
4
1/2
≤
≤
rζΓ
4
−1/2
αj
∗
0≤|j|≤kn
X
∗
0≤|j|≤kn
1/2
rζηΓ
4
X
γ ∗
kn
ωkn∗
1/2
≤
1/2
γj−1
X
∗
0≤|j|≤kn
1/2
r
4
1/2
X
γj
nαj
X
γj
nαj
∗
0≤|j|≤kn
1/2
ωj
nαj
.
∗
Moreover kλθ k2γ ≤ r holds for each θ ∈ {±1}kn +1 due to the estimate
kλθ k2γ =
X
∗
0≤|j|≤kn
|[λθ ]j |2 γj =
"
1/2
r
4
+ θ0
rζ
4n
≤
rζ
rζ γkn∗
r
+
+
2 2n
4 ωkn∗
≤
r rζ γkn∗
+
2
2 ωkn∗
X
1/2 #2
X
∗
1≤|j|≤kn
∗
0≤|j|≤kn
+
rζ
4
ωj
nαj
∗
1≤|j|≤kn
ωj
≤ r.
nαj
∗
This estimate and the non-negativity of λθ together imply λθ ∈ Λrγ for all θ ∈ {±1}kn +1 .
From now on let f ∈ Fαd be fixed and let Pθ denote the joint distribution of the i.i.d. samples
i
N1 , . . . , Nn and Y1 , . . . , Ym when the true parameters are λθ and f , respectively. Let PN
θ denote
the corresponding one-dimensional marginal distributions and Eθ the expectation with respect
e be an arbitrary estimator of λ. The key argument of the proof is the following
to Pθ . Let λ
reduction scheme:
e − λk2 ≥
sup sup Ekλ
ω
λ∈Λrγ
f ∈Fαd
=
sup
∗ +1
kn
θ∈{±1}
1
2kn∗ +1
X
∗
e − λ k2 ] ≥
Eθ [kλ
θ ω
X
∗
θ∈{±1}kn +1 0≤|j|≤kn
1
2kn∗ +1
X
e − λ ]j |2 ]
ωj Eθ [|[λ
θ
16
∗
θ∈{±1}kn +1
e − λ k2 ]
Eθ [kλ
θ ω
=
X
1
∗ +1
kn
2
∗
0≤|j|≤kn
X
ωj
2
∗
θ∈{±1}kn +1
e − λ (|j|) ]j |2 ]}, (A.1)
e − λ ]j |2 ] + E (j) [|[λ
{Eθ [|[λ
θ
θ
θ
(|j|)
∗
∗
where for θ ∈ {±1}kn +1 and j ∈ {−kn∗ , . . . , kn∗ } the element θ (|j|) ∈ {±1}kn +1 is defined by θk =
Rp
(|j|)
dPθ dPθ(|j|) .
θk for k 6= |j| and θ|j| = −θ|j|. Consider the Hellinger affinity ρ(Pθ , Pθ(|j|) ) :=
e of λ we have
For an arbitrary estimator λ
ρ(Pθ , Pθ(|j|) ) ≤
Z
≤
e − λ ]j | q
|[λ
θ
dPθ dPθ(|j|) +
|[λθ − λθ(|j|) ]j |
Z
e − λ ]j |2
|[λ
θ
dPθ
|[λθ − λθ(|j|) ]j |2
!1/2
+
Z
e − λ (|j|) ]j | q
|[λ
θ
dPθ dPθ(|j|)
|[λθ − λθ(|j|) ]j |
Z
e − λ (|j|) ]j |2
|[λ
θ
dP (|j|)
|[λθ − λθ(|j|) ]j |2 θ
!1/2
from which we conclude by means of the elementary inequality (a + b)2 ≤ 2a2 + 2b2 that
1
e − λ (|j|) ]j |2 ].
e − λ ]j |2 ] + E (|j|) [|[λ
|[λθ − λθ(|j|) ]j |2 ρ2 (Pθ , Pθ(|j|) ) ≤ Eθ [|[λ
θ
θ
θ
2
Introduce
the Hellinger distance between two probability measures P and Q as H(P, Q) :=
√
R √
two finite
ν and
( [ dP − dQ]2 )1/2 and, analogously, the Hellinger distance between
R √
√ measures
2
1/2
:=
µ (that not necessarily have total mass equal to one) by H(ν, µ)
( [ dν − dµ] ) (as usual,
the integral is formed with respect to any measure dominating both ν and µ). Let νθ denote
the intensity measure of a Poisson point process N on [0, 1) whose Radon-Nikodym derivative
with respect to the Lebesgue measure is given by ℓθ := λθ ⋆ f . Note that we have the estimate
√
∗
1
due to
ℓθ ≥ δ r for all θ ∈ {±1}kn +1 with δ = 12 − 2√
2
rζ
4n
1/2
X
+
∗
1≤|j|≤kn
|[λθ ]j · [f ]j | ≤
rdζ
4n
1/2
X
∗
0≤|j|≤kn
−1/2
αj
√
r
≤ √
2 2
which can be realized in analogy to the non-negativity of λθ shown above. We obtain
2
H (νθ , νθ(|j|) ) =
Z p
( ℓθ −
q
2
ℓθ(|j|) ) =
Z
√
kℓθ − ℓθ(|j|) k22
|ℓθ − ℓθ(|j|) |2
ζd r
1
√
p
√
≤
=
≤ .
2
4δ r
4δn
n
( ℓθ + ℓθ(|j|) )
Since the distribution of the sample Y1 , . . . , Ym does not depend on the choice of θ we obtain
H 2 (Pθ , Pθ(|j|) ) ≤
n
X
i=1
Ni
i
H 2 (PN
θ , Pθ (|j|) ) ≤
n
X
i=1
H 2 (νθ , νθ(|j|) ) ≤ 1,
where the first estimate follows from Lemma 3.3.10 (i) in [Rei89] and the second one is due to
Theorem 3.2.1 in [Rei93] which can be applied since each Ni is a Poisson point process for the
Poisson model. Thus, the relation ρ(Pθ , Pθ(|j|) ) = 1 − 12 H 2 (Pθ , Pθ(|j|) ) implies ρ(Pθ , Pθ(|j|) ) ≥ 21 .
Finally, putting the obtained estimates into the reduction scheme (A.1) leads to
e − λk2 ≥
sup sup Ekλ
ω
λ∈Λrγ f ∈Fαd
≥
1
2kn∗ +1
X
X
∗
X
∗
θ∈{±1}kn +1 0≤|j|≤kn
∗
0≤|j|≤kn
ωj
e − λ (|j|) ]j |2 ]}
e − λ ]j |2 ] + E (|j|) [|[λ
{Eθ [|[λ
θ
θ
θ
2
ωj
ζr X
ωj
ζr
|[λθ − λθ(|j|) ]j |2 =
· Ψn
≥
16
16 0≤|j|≤k∗ nαj
16η
n
e was arbitrary.
which finishes the proof of the theorem since λ
17
A.2. Proof of Theorem 3.3
The following reduction scheme follows along a general strategy that is well-known for the
establishment of lower bounds in nonparametric estimation (for a detailed account cf. [Tsy08],
e of λ and an
Chapter 2). Note that by Markov’s inequality we have for an arbitrary estimator λ
arbitrary A > 0 (which will be specified below)
2
2
e
e
E[Φ−1
m kλ − λkω ] ≥ A · P(kλ − λkω ≥ AΦm ),
which by reduction to two hypotheses implies
2
2
e
e
sup sup E[Φ−1
m kλ − λkω ] ≥ A sup sup P(kλ − λkω ≥ AΦm )
λ∈Λrγ f ∈Fαd
λ∈Λrγ f ∈Fαd
e − λ k2 ≥ AΦm )
≥ A sup Pθ (kλ
θ ω
θ∈{±1}
where Pθ denotes the distribution when the true parameters are λθ and fθ . The specific hypotheses λ1 , λ−1 and f1 , f−1 will be specified below. If λ−1 and λ1 can be chosen such that
kλ1 − λ−1 k2ω ≥ 4AΦm , application of the triangle inequality yields
e − λ k2 ≥ AΦ ) ≥ P (τ ∗ 6= θ)
Pθ (kλ
m
θ
θ ω
e − λ k2 .
where τ ∗ denotes the minimum distance test defined through τ ∗ = arg minθ∈{±1} kλ
θ ω
Hence, we obtain
e − λ k2 ≥ AΦm )
e − λk2 ≥ AΦm ) ≥ inf sup P (kλ
inf sup sup P(kλ
θ
θ ω
ω
e
λ
e
λ
λ∈Λrγ f ∈Fαd
θ∈{±1}
≥ inf sup Pθ (τ 6= θ)
τ θ∈{±1}
=: p∗ ,
(A.2)
where the infimum is taken over all {±1}-valued functions τ based on the observations. Thus,
it remains to find hypotheses λ1 , λ−1 ∈ Λrγ and f1 , f−1 ∈ Fαd such that
kλ1 − λ−1 k2ω ≥ 4AΦm ,
(A.3)
and which allow us to bound p∗ by a universal constant (independent of m) from below.
ωj
1
−1/2 α−1/2 ),
∗ := arg max
For this purpose, set km
∗
j≥1 { γj min(1, mαj )} and am := ζ min(1, m
km
where ζ is defined as in the statement of the theorem. Take note of the inequalities
1/d1/2 = (1 − (1 − 1/d1/4 ))2 ≤ (1 − am )2 ≤ 1
and
1 ≤ (1 + am )2 ≤ (1 + (1 − 1/d1/4 ))2 = (2 − 1/d1/4 )2 ≤ d1/2
which in combination imply 1/d1/2 ≤ (1 + θam )2 ≤ d1/2 for θ ∈ {±1}. These inequalities will be
used below without further reference. For θ ∈ {±1}, we define
λθ =
1/2
r
2
+ (1 − θam )
1/2
r
8
−1/2
∗ + e−k ∗ ).
(ekm
d−1/4 γkm
∗
m
Note that λθ is real-valued by definition. Furthermore, we have
kλθ k2γ =
r
r
3r
r
2
∗ |[λθ ]k ∗ | ≤
+ 2γkm
+ (1 + am )2 d−1/2 ≤
m
2
2
4
4
18
and
|λθ (t)| ≥
1/2
r
2
−2
1/2
r
8
≥0
∀t ∈ [0, 1),
which together imply that λθ ∈ Λrγ for θ ∈ {±1}. The identity
2 −1/2
−1
∗ γ ∗ = ζ rd
kλ1 − λ−1 k2ω = ra2m d−1/2 ωkm
· Φm
km
√
shows that the
condition in (A.3) is satisfied with A = ζ 2 r/(4 d).
√
Let f ∈ Fα d be such that f ≥ 1/2 (the existence is guaranteed through condition (C4)) and
define for θ ∈ {±1}
∗ ek ∗ + [f ]−k ∗ e−k ∗ ).
fθ = f + θam ([f ]km
m
m
m
∗ ≥ 1 we have
Since km
R1
0
fθ (x)dx = 1 and fθ ≥ 0 holds because of the estimate
1/2
1/2
≥0
|fθ (t)| ≥ 1/2 − 2am αkm
∗ d
for all t.
∗ since
∗ , we have [f ] = [f ] and thus trivially 1/d ≤ |[f ] |2 /α ≤ d for |j| =
6 km
For
|j| =
6 km
j
j
θ j
θ j
√
Fα d ⊂ Fαd . Moreover
1/d ≤ d−1/2
2
2
2
∗ |
∗ |
∗ |
|[f ]±km
(1 + θam )2 |[f ]±km
|[f ]±km
≤
≤ d1/2
≤d
∗
∗
∗
α±km
α±km
α±km
and hence fθ ∈ Fαd for θ ∈ {±1}.
To obtain a lower bound for p∗ defined in (A.2) consider the joint distribution Pθ of the
samples N 1 , . . . , N n and Y1 , . . . , Ym under λθ and fθ . Note that due to our construction we have
i
Ni
λ−1 ⋆ f−1 = λ1 ⋆ f1 . Thus PN
−1 = P1 for all i = 1, . . . , n (due to the fact that the distribution of
a Poisson point process is determined by its intensity) and the Hellinger distance between P−1
and P1 does only depend on the distribution of the sample Y1 , . . . , Ym . More precisely,
1 ,...,Ym
1
, PY1 1 ),
, PY1 1 ,...,Ym ) ≤ mH 2 (PY−1
H 2 (P−1 , P1 ) = H 2 (PY−1
1
and we proceed by bounding H 2 (PY−1
, PY1 1 ) from above. Recall that f ≥ 1/2 which is used to
obtain the estimate
1
H 2 (PY−1
, PY1 1 ) =
Z
1
0
|f1 (x) − f−1 (x)|2
dx ≤
2f (x)
Z
∗ ≤
|f1 (x) − f−1 (x)|2 dx ≤ 8da2m αkm
1
.
m
Hence we have H 2 (P−1 , P1√
) ≤ 1 and application of statement (ii) of Theorem 2.2 in [Tsy08] with
α = 1 implies p∗ ≥ 12 (1 − 3/2) which finishes the proof of the theorem.
A.3. Proof of Theorem 3.6
We give the proof for the Poisson model only. The proof for the Cox model follows in complete
analogy by exploiting statement ii) instead of i) in part a) of Lemma A.1.
e ∗ := P
Set λ
∗ [λ]j 1Ωj ej . The proof consists in finding appropriate upper bounds for
kn
0≤|j|≤kn
the quantities and △ in the estimate
b ∗ − λk2 ] ≤ 2 E[kλ
b ∗ −λ
e ∗ k2 ] + 2 E[kλ − λ
e ∗ k2 ] =: 2 + 2△.
E[kλ
kn ω
kn
kn
kn ω
ω
19
(A.4)
c = [f ] [λ] we obtain
Upper bound for : Using the identity E[ℓ]
j
j
j
=
X
∗
0≤|j|≤kn
≤2
X
c /[f
c] − [λ] |2 1 ]
ωj E[|[ℓ]
j
Ωj
j
j
∗
0≤|j|≤kn
c /[f
c] − E[ℓ]
c /[f
c] |2 1 ] + 2
ωj E[|[ℓ]
Ωj
j
j
j
j
X
∗
0≤|j|≤kn
=: 21 + 22 .
c] − 1|2 1 ]
ωj |[λ]j |2 E[|[f ]j /[f
Ωj
j
c] , the definition of Ω and the independence
Using the estimate |a|2 ≤ 2 |a − 1|2 +2 for a = [f ]j /[f
j
j
c and [f
c] we get
of [ℓ]
j
j
X
c /[f
c] − E[ℓ]
c /[f
c] |2 · [f ]j
1 =
ωj E |[ℓ]
j
j
j
j
[f ]j
0≤|j|≤k ∗
≤2
n
X
mωj
∗
0≤|j|≤kn
c ) Var([f
c] )
Var([ℓ]
j
j
|[f ]j |2
+2
2
X
1 Ωj
ωj
∗
0≤|j|≤kn
c )
Var([ℓ]
j
|[f ]j |2
.
Applying statements a) and b) from Lemma A.1 together with f ∈ Fαd yields
1 ≤ 4d
X
∗
0≤|j|≤kn
ωj
[λ]0
nαj
which using that γj ≥ 1 (which holds due to Assumption A) implies
√
1 ≤ 4d r
X
∗
0≤|j|≤kn
√
ωj
≤ 4d r · Ψn .
nαj
c] and the definition of
Now consider 2 . Using the estimate |a|2 ≤ 2 |a − 1|2 + 2 for a = [f ]j /[f
j
Ωj yields
c] − [f ] |4 ]
c] )
E[|[f
Var([f
j
j
j
c] − 1|2 1 ] ≤ 2m
E[|[f ]j /[f
+
2
(A.5)
Ω
j
j
2
2 .
|[f ]j |
|[f ]j |
c] −
Notice that Theorem 2.10 in [Pet95] implies the existence of a constant C > 0 with E[|[f
j
[f ]j |4 ] ≤ C/m2 . Using this inequality in combination with assertion b) from Lemma A.1 and
f ∈ Fαd implies
c] − 1|2 1 ] ≤ 2d(C + 1)/(mα ).
E[|[f ]j /[f
(A.6)
Ωj
j
j
c] − 1|2 1 ] ≤ m Var([f
c] ) ≤ 1 which in combination with (A.6) implies
In addition, E[|[f ]j /[f
Ωj
j
j
X
1
2 ≤ 2d(C + 1)
ωj |[λ]j | min 1,
mαj
0≤|j|≤k ∗
2
n
!
.
Exploiting the fact that λ ∈ Λrγ and the definition of Φm in (3.2) we obtain
2 ≤ 2dr(C + 1)(1 + γ1 /ω1 ) · Φm .
Putting together the estimates for 1 and 2 yields
√
≤ 8d r · Ψn + 4d(C + 1)(1 + γ1 /ω1 )r · Φm .
20
Upper bound for △: △ can be decomposed as
△=
X
j∈Z
ωj |[λ]j |2 E(1 − 1{0≤|j|≤kn∗ } · 1Ωj ) =
X
∗
|j|>kn
ωj |[λ]j |2 +
X
∗
0≤|j|≤kn
ωj |[λ]j |2 · P(Ωcj )
= △1 + △2 .
λ ∈ Λrγ implies △1 ≤ rωkn∗ /γkn∗ ≤ r · Ψn and Lemma A.1 yields the estimate △2 ≤ 4dr · Φm which
together imply △ ≤ r · Ψn + 4dr · Φm . Putting the obtained estimates for and △ into (A.4)
finishes the proof of the theorem.
A.4. Auxiliary results for the proof of Theorem 3.6
Lemma A.1 With the notations introduced in the main part of the article, the following assertions hold:
a)
c ) ≤ [λ] /n under the Poisson model and
i) Var([ℓ]
0
j
c ) ≤ 2(|[λ]|2 + [λ] )/n under the Cox model.
ii) Var([ℓ]
0
j
j
c] ) ≤ 1/m,
b) Var([f
j
c] |2 < 1/m) ≤ min {1, 4d/(mα )}
c) P(Ωcj ) = P(|[f
j
j
∀f ∈ Fαd .
Proof. The proof of statement i) in a) is given by the identity
c ) = 1 Var
Var([ℓ]
j
n
Z
1
ej (t)dN1 (t) =
0
1
n
Z
0
1
|ej (t)|2 (λ ⋆ f )(t)dt =
1
· [λ]0 .
n
c = [λ] [f ] implies
To prove ii), the identity E[ℓ]
j
j
j
c ) := E[|[ℓ]
c − E[ℓ]
c |2 ] ≤ 2E[|[f
f] [λ] − [f ] [λ] |2 ] + 2 E[|ξ |2 ] =: 2V + 2V
Var([ℓ]
j
j
j
j
1
2
j
j
j
j
f] ) ≤ |[λ] |2 /n. Here, the estimate Var([f
f] ) ≤ 1/n is easily derived in
where V1 ≤ |[λ]j |2 · Var([f
j
j
j
analogy to the proof of part b). In order to bound V2 from above, notice
" "Z
1
E[|ξj | ] = E E
n
2
Z
0
2
1
ej (−t) {dN1 (t) − λ(t − ε1 − ⌊t − ε1 ⌋)dt}
1
1
= E
|ej (−t)|2 λ(t − ε1 − ⌊t − ε1 ⌋)dt
n
0
Z 1
1
= E
λ(t − ε1 − ⌊t − ε1 ⌋)dt
n
0
= [λ]0 /n.
| ε1
##
The assertion follows now by combining the obtained bounds for V1 and V2 .
c] ) = 1 Var (e (−Y )) and the assertion follows
For the proof of b), note that we have Var([f
j
1
j
m
from the estimate
Var(ej (−Y1 )) = E[|ej (−Y1 )|2 ] − |E [ej (−Y1 )]|2 ≤ E[|ej (−Y1 )|2 ] = 1.
For the proof of c), we consider two cases: if |[f ]j |2 < 4/m we have 1 <
and the statement is evident. Otherwise, |[f ]j |2 ≥ 4/m which implies
4d
mαj
because f ∈ Fαd
c] |2 < 1/m) ≤ P(|[f
c] |/ |[f ] | < 1/2) ≤ P(|[f
c] /[f ] − 1| > 1/2).
P(|[f
j
j
j
j
j
21
Applying Chebyshev’s inequality and exploiting the definition of Fαd yields
c] |2 < 1/m) ≤ 4/ |[f ] |2 · Var([f
c] ) ≤ 4d/(mα )
P(|[f
j
j
j
j
and statement c) follows.
B. Proofs of Section 4
B.1. Proof of Theorem 4.1
c ∨ 1 ≤ η −1 ([ℓ] ∨ 1)} and
Define the events Ξ1 := {η([ℓ]0 ∨ 1) ≤ [ℓ]
0
0
Ξ2 :=
(
∀ 0 ≤ |j| ≤
α
Mm
c]−1 − [f ]−1 | ≤
: |[f
j
j
1
2|[f ]j |
and
The identity 1 = 1Ξ1 ∩Ξ2 + 1Ξc2 + 1Ξc1 ∩Ξ2 provides the decomposition
c] | ≥ 1
|[f
j
m
)
.
b − λk2 1Ξc ] + E[kλ
b − λk2 1Ξc ∩Ξ ],
b − λk2 ] = E[kλ
b − λk2 1Ξ ∩Ξ ] + E[kλ
E[kλ
ω
ω
2
ω
ω
1
2
e
e
e
e
2
1
k
k
k
k
{z
|
}
=:1
|
{z
}
=:2
|
{z
=:3
}
and we will establish uniform upper bounds over the ellipsoids Λrγ and Fαd for the three terms
on the right-hand side separately.
Uniform upper bound for 1 : Denote by Sk the linear subspace of L2 spanned by the functions
b k2 holds for all t ∈ S ,
b k2 − kλ
ej (·) for j ∈ {−k, . . . , k}. Since the identity Υ(t) = kt − λ
k
k ω
k ω
b . Using this identity and
k ∈ {0, . . . , n ∧ m}, we obtain for all such k that argmint∈Sk Υ(t) = λ
k
α } that
the definition of ke yields for all k ∈ {0, . . . , Knm
P
b ) + pen
b ) + pen
g e ≤ Υ(λ
g k ≤ Υ(λk ) + pen
gk
Υ(λ
k
e
k
k
where λk := 0≤|j|≤k [λ]j ej denotes the projection of λ on the subspace Sk . Elementary computations imply
b k2 ≤ kλ k2 + 2ℜhλ
b
b − λ i + pen
g k − pen
ge
kλ
(B.1)
n∧m , λe
k ω
k ω
e
k ω
k
k
α }. In addition to λ defined above, introduce the further abbreviations
for all k ∈ {0, . . . , Knm
k
e :=
λ
k
as well as
c
[ℓ]
j
ej
[f ]j
0≤|j|≤k
X
b − λ̌ − λ
e +λ ,
Θk := λ
k
k
k
k
and
λ̌k :=
X
0≤|j|≤k
e −λ ,
e := λ
Θ
k
k
k
and
[ℓ]j
c]
[f
j
1 Ωj e j ,
Θ̌k := λ̌k − λk .
b n∧m − λn∧m = Θn∧m + Θ
e n∧m + Θ̌n∧m , we deduce
Using these abbrevations and the identity λ
from (B.1) that
b − λk2 ≤ kλ − λ k2 + pen
e n∧m , λ
b − λ iω
g k − pen
g e + 2ℜhΘ
kλ
k ω
k
ω
e
e
k
k
k
b − λ i + 2ℜhΘ̌
b −λ i
+ 2ℜhΘn∧m , λ
n∧m , λe
k ω
k ω
e
k
k
(B.2)
α }. Define B := {λ ∈ S : kλk ≤ 1}. For every τ > 0 and t ∈ S , the
for all k ∈ {0, . . . , Knm
k
k
k
ω
estimate 2uv ≤ τ u2 + τ −1 v 2 implies
2 |hh, tiω | ≤ 2 ktkω sup |hh, tiω | ≤ τ ktk2ω +
t∈Bk
22
1
sup |hh, tiω |2 .
τ t∈Bk
b −λ ∈S
Because λ
, combining the last estimate with (B.2) we get
k
e
e
k
k∨k
b − λk2 ≤ kλ − λ k2 + 3τ kλ
b − λ k2 + pen
g k − pen
g e+
kλ
k ω
k ω
ω
e
e
k
k
k
e n∧m , tiω |2 + τ −1 sup |hΘn∧m , tiω |2 + τ −1 sup |hΘ̌n∧m , tiω |2 .
+ τ −1 sup |hΘ
t∈B
t∈B
e
t∈B
e
e
k∨k
k∨k
k∨k
b − λk2 + 2 kλ − λk2 and kλ − λ k2 ≤ rω γ −1 for all λ ∈ Λr since
b − λ k2 ≤ 2kλ
Note that kλ
k
k ω
k ω
k k
ω
γ
ω
e
e
k
k
ωγ −1 is non-increasing due to Assumption A. Specializing with τ = 1/8, we obtain
b − λk2 ≤ 7rω γ −1 + 4pen
e n∧m , tiω |2
g k − 4pen
g e + 32 sup |hΘ
kλ
k k
ω
e
k
k
t∈B
e
k∨k
+ 32 sup |hΘn∧m , tiω | + 32 sup |hΘ̌n∧m , tiω |2 .
2
t∈B
e
k∨k
t∈B
(B.3)
e
k∨k
α and K α ≤ M α by definition, we
Combining the facts that 1Ωj 1Ξ2 = 1Ξ2 for 0 ≤ |j| ≤ Mm
m
nm
α
α
obtain for all j ∈ {−Knm , . . . , Knm } the estimate
c] 1 − 1|2 1 = |[f ] |2 |1/[f
c] − 1/[f ] |2 1 ≤ 1/4.
|[f ]j /[f
Ξ2
j
j
Ξ2
j Ωj
j
Hence, supt∈Bk |hΘn∧m , tiω |2 1Ξ2 ≤
we obtain
b − λk2 1Ξ ∩Ξ
kλ
ω
1
2
e
k
1
4
e n∧m , tiω |2 for all 0 ≤ k ≤ K α . Thus, from (B.3)
supt∈Bk |hΘ
nm
33d([ℓ]0 ∨ 1)δα e
−1
2
k∨k
e
≤ 7rωk γk + 40 sup |hΘn∧m , tiω | −
8n
t∈B
+ (165d([ℓ]0 ∨
e
k∨k
1)δα e/n
k∨k
+
g k − 4pen
g e)1Ξ1 ∩Ξ2 + 32 sup |hΘ̌n∧m , tiω |2 .
+ 4pen
k
α
t∈BKnm
g and the event Ξ1 , we obtain
Exploiting the definition of both the penalty pen
ωk δkα
2
b
E[kλek − λkω 1Ξ1 ∩Ξ2 ] ≤ C(d, r) minα max
,
0≤k≤Knm
γk n
"
! #
α
Knm
X
33([ℓ]0 ∨ 1)dδkα
2
e
E
sup |hΘn∧m , tiω | −
+ 40
8n
t∈Bk
k=0
+ 32E sup |hΘ̌n∧m , tiω |2 .
α
t∈BKnm
+
(B.4)
Applying Lemma B.2 with δk∗ = dδkα and ∆∗k = d∆αk yields
E
"
α
e n∧m , tiω |2 − 33d([ℓ]0 ∨ 1)δk
sup |hΘ
8n
t∈Bk
! #
+
≤ K1
+
"
δkα
dkf kkλk∆αk
exp −K2
n
kf k2 kλk2 ∆αk
!
√
dδkα
exp(−K3 n) .
2
n
α ≤ n by definition, we obtain that
Using statement a) of Lemma B.1 and the fact that Knm
α
Knm
X
k=0
E
"
α
e n∧m , tiω | − 33d([ℓ]0 ∨ 1)δk
sup |hΘ
t∈Bk
! #
8n
+
√
∞
3/2
X
√
d
rρ
2K2 k log(∆αk ∨ (k + 3))
α
∆k exp − √
+ exp(−K3 n),
·
.
n
log(k + 3)
drρ
k=0
23
where the last estimate is due to the fact that kf k2 ≤ dρ for all f ∈ Fαd and kλk2 ≤ r for all
λ ∈ Λrγ . Note that we have
∞
X
∆αk exp
k=0
2K2 k log(∆αk ∨ (k + 3))
·
−√
log(k + 3)
drρ
≤C<∞
with a numerical constant C which implies
α
Knm
X
E
k=0
"
α
e n∧m , tiω | − 33d([ℓ]0 ∨ 1)δk
sup |hΘ
8n
t∈Bk
! #
1
.
n
.
+
The last term in (B.4) is bounded by means of Lemma B.3 which immediately yields
E sup |hΘ̌n∧m , tiω |2 . Φm .
α
t∈BKnm
Combining the preceeding estimates, which hold uniformly for all λ ∈ Λrγ and f ∈ Fαd , we
conclude from equation (B.4) that
h
i
b − λk2 1Ξ ∩Ξ .
sup sup E kλ
ω
1
2
e
k
r
d
λ∈Λγ f ∈Fα
ωk δkα
,
min α max
0≤k≤Knm
γk n
+ Φm +
1
.
n
P
b − λ̆ k2 ≤ kλ
b ′−
Uniform upper bound for 2 : Define λ̆k := 0≤|j|≤k [λ]j 1Ωj ej . Note that kλ
k
k ω
k
α
2
′
2
2
λ̆k′ kω for k ≤ k and kλ̆k − λkω ≤ kλkω for all k ∈ N0 . Consequently, since k ∈ {0, . . . , Knm },
we obtain the estimate
b − λk2 1Ξc ] ≤ 2E[kλ
b − λ̆ k2 1Ξc ] + 2 E[kλ̆ − λk2 1Ξc ]
E[kλ
ω
ω
e
e
e
e
2
2
2
k
k
k ω
k
b α − λ̆ α k2 1 c ] + 2 kλk2 P(Ξc ),
≤ 2E[kλ
Knm
Knm ω Ξ2
2
ω
and due to Assumption A and Lemma B.5 it is easily seen that kλk2ω · P(Ξc2 ) . m−4 . Using the
definition of Ωj , we further obtain
b α − λ̆ α k2 1 c ] ≤ 2m
E[kλ
Knm
Knm ω Ξ2
≤ 2m
X
α
0≤|j|≤Knm
X
α
0≤|j|≤Knm
+ 2m
X
c − [ℓ] |2 1 c ] + E[|[f ] [λ] − [f
c] [λ] |2 1 c ]}
ωj {E[|[ℓ]
j
Ξ2
j
j
j
Ξ2
j
j
c − [ℓ] |4 ])1/2 P(Ξc )1/2
ωj (E[|[ℓ]
j
2
j
α
0≤|j|≤Knm
. mP(Ξc2 )1/2
c] − [f ] |4 ])1/2 P (Ξc )1/2
ωj |[λ]j |2 (E[|[f
j
2
j
X
α
0≤|j|≤Knm
X
ωj
+ P(Ξc2 )1/2
ωj |[λ]j |2 ,
n
α
0≤|j|≤K
(B.5)
nm
where the last estimate follows by applying Theorem 2.10 from [Pet95] with p = 4 two times. If
α = 0, Lemma B.5 implies
Knm
b K α − λ̆K α k2 1Ξc ] .
E[kλ
nm
nm ω
2
1
1
+
.
nm m2
α > 0, we exploit ω ≤ ω + α−1 , K α ≤ N α and the definition of N α to bound
Otherwise, if Knm
j
n
n
nm
j j
the first term in (B.5). The second term in (B.5) can be bounded from above by noting that
24
ωj ≤ γj thanks to Assumption A, and we obtain
b K α − λ̆K α k2 1Ξc ] . mP(Ξc )1/2
E[kλ
2
nm
nm ω
2
X
0≤|j|≤Nnα
1
1
+ P(Ξc2 )1/2 .
2|j| + 1 log(n + 3)
Thanks to the logarithmic increase of the harmonic series, Nnα ≤ n and Lemma B.5, the last
estimate implies
b K α − λ̆K α k2 1Ξc ] . 1 + 1 ,
E[kλ
nm
nm ω
2
m m2
α
if Knm > 0, and thus
b K α − λ̆K α k2 1Ξc ] . 1 + 1 ,
E[kλ
nm
nm ω
2
m m2
α
independent of the actual value of Knm . Using the obtained estimates, we conclude
b − λk2 1Ξc ] .
E[kλ
ω
e
2
k
1
.
m
Uniform upper bound for 3 : In order to find a uniform upper bound for 3 , first recall the
P
definition λ̆k := 0≤|j|≤k [λ]j 1Ωj ej and consider the estimate
b − λk2 1Ξc ∩Ξ ] ≤ 2E[kλ
b − λ̆ k2 1Ξc ∩Ξ ] + 2E[kλ̆ − λk2 1Ξc ∩Ξ ].
E[kλ
ω
2
2
ω
2
e
e
e
e
1
1
1
k
k
k ω
k
(B.6)
Using the estimate kλ̆ek − λk2ω ≤ kλk2ω , we obtain for λ ∈ Λrγ by means of Lemma B.4 that
E[kλ̆ek − λk2ω 1Ξc1 ∩Ξ2 ] ≤ rP(Ξc1 ) .
1
n
which controls the second term on the right-hand side of (B.6). We now bound the first term on
α = 0, we have k
e = 0, and by means of the Cauchy-Schwarz
the right-hand side of (B.6). If Knm
inequality and Theorem 2.10 from [Pet95] it is easily seen that
b − λ̆ k2 1Ξc ∩Ξ ] .
E[kλ
2
e
e
1
k
k ω
1
.
n
α > 0, and we need the following further estimate, which is easily verified:
Otherwise, Knm
b − λ̆ k2 1Ξc ∩Ξ ] ≤ 3
E[kλ
2
e
e
1
k
k ω
X
α
0≤|j|≤Knm
+3
X
c] − [ℓ] /[f ] |2 1 c
ωj E[|[ℓ]j /[f
j
j
Ξ1 ∩Ξ2 ]
j
α
0≤|j|≤Knm
+3
X
α
0≤|j|≤Knm
c − [ℓ] |2 /|[f ] |2 1 c
ωj E[|[ℓ]
j
j
Ξ1 ∩Ξ2 ]
j
c − [ℓ] |2 · |1/[f
c] − 1/[f ] |2 1 c
ωj E[|[ℓ]
j
j
Ξ1 ∩Ξ2 ].
j
j
(B.7)
We start by bounding the first term on the right-hand side of (B.7). Using the definition of Ξ2
and ωj ≤ γj , we obtain for all λ ∈ Λrγ that
X
α
0≤|j|≤Knm
c] − [ℓ] /[f ] |2 1 c
ωj E[|[ℓ]j /[f
j
j
Ξ1 ∩Ξ2 ] ≤
j
1
r
· P(Ξc1 ) . .
4
n
Since |[f ]j |−2 ≤ dαj for f ∈ Fαd , the Cauchy-Schwarz inequality in combination with Theo-
25
rem 2.10 from [Pet95] implies for the second term on the right-hand side of (B.7) that
X
α
0≤|j|≤Knm
c 1/2
c − [ℓ] |2 /|[f ] |2 1 c
ωj E[|[ℓ]
j
j
Ξ1 ∩Ξ2 ] . P(Ξ1 )
j
X
α
0≤|j|≤Knm
ωj+
.
nαj
α ≤ N α to obtain
We exploit the definition of Nnα together with Knm
n
X
α
0≤|j|≤Knm
c − [ℓ] |2 /|[f ] |2 1 c
ωj E[|[ℓ]
j
j
Ξ1 ∩Ξ2 ] .
j
X
P(Ξc1 )1/2
1
,
log(n + 3) 0≤|j|≤N α 2|j| + 1
n
from which by the logarithmic increase of the harmonic series and Lemma B.4 we conclude that
X
α
0≤|j|≤Knm
c − [ℓ] |2 /|[f ] |2 1 c
ωj E[|[ℓ]
j
j
Ξ1 ∩Ξ2 ] .
j
1
,
n
α . Finally, the third and last term on the right-hand side
independent of the actual value of Knm
of (B.7) can be bounded from above the same way after exploiting the definition of Ξ2 , and we
obtain
X
1
c − [ℓ] |2 · |1/[f
c] − 1/[f ] |2 1 c
ωj E[|[ℓ]
.
j
j
Ξ1 ∩Ξ2 ] .
j
j
n
α
0≤|j|≤K
nm
Putting together the derived estimates, we obtain
b − λk2 1Ξc ∩Ξ ] .
E[kλ
2
e
1
k
1
.
n
Finally, the statement of the theorem follows by combining the obtained uniform upper bounds
for 1 , 2 , and 3 .
B.2. Proof of Theorem 4.2
Consider the event
α−
cnm ≤ N α+ ∧ M α+ }
Ξ3 := {Nnα− ∧ Mm
≤K
n
m
in addition to the event Ξ1 introduced in the proof of Theorem 4.1 and the slightly redefined
event Ξ2 defined as
α+
c] − 1/[f ] | ≤ 1/(2|[f ] |) and |[f
c] | ≥ 1/m}.
Ξ2 := {∀0 ≤ |j| ≤ Mm
: |1/[f
j
j
j
j
Defining Ξ := Ξ1 ∩ Ξ2 ∩ Ξ3 , the identity 1 = 1Ξ + 1Ξc2 + 1Ξc1 ∩Ξ2 + 1Ξ1 ∩Ξ2 ∩Ξc3 motivates the
decomposition
b − λk2 ] = E[kλ
b − λk2 1Ξ ] + E[kλ
b − λk2 1Ξc ]
E[kλ
ω
ω
ω
b
b
b
2
k
k
k
b − λk2 1
b − λk2 1 c
+ E[kλ
ω Ξ1 ∩Ξ2 ∩Ξc3 ]
ω Ξ1 ∩Ξ2 ] + E[kλb
b
k
k
=: 1 + 2 + 3 + 4 ,
and we establish uniform upper risk bounds for the four terms on the right-hand side separately.
b ≤ 9 ∆ , and thus
Uniform upper bound for 1 : On Ξ we have the estimate 41 ∆k ≤ ∆
k
4 k
1
b ∨ (k + 4) ≤ 9 [∆ ∨ (k + 4)]
[∆k ∨ (k + 4)] ≤ ∆
k
k
4
4
26
α+ }. This last estimate implies
for all k ∈ {0, . . . , Mm
log(k + 4)
log(∆k ∨ (k + 4))
log 4
2k + 1
∆k
1−
≤ δbk
4
log(k + 4)
log(k + 4) log(∆k ∨ (k + 4))
log(k + 4)
log(∆k ∨ (k + 4))
log(9/4)
9(2k + 1)
∆k
1+
,
≤
4
log(k + 4)
log(k + 4) log(∆k ∨ (k + 4))
from which we conclude
that on Ξ2 the estimate
3
100
· δk ≤ δbk ≤
17
5
· δk . Putting penk :=
dk ≤
penk ≤ pen
340
penk
3
165 −1 c
4 η ([ℓ]0
∨ 1) · δnk , we observe
α+ }. Note that on Ξ we have k
b ≤ M α+ which implies
holds for all k ∈ {0, . . . , Mm
m
d k − pen
d b)1Ξ ≤ (penk + penb + pen
d k − pen
d b)1Ξ ≤
(penk∨bk + pen
k
k
k
343
penk 1Ξ .
3
(B.8)
Now, we can proceed by mimicking the derivation of (B.4) in the proof of Theorem 4.1. More
g k used in that proof by pen
d k , using the definition of
precisely, replacing the penalty term pen
penk above and (B.8), we obtain
b − λk2 1 ] ≤ 7rω γ −1 + 40
E[kλ
k k
ω Ξ
b
k
α+
N
n
X
E
k=0
"
e n∧m , tiω | − 33([ℓ]0 ∨ 1)δk
sup |hΘ
8n
t∈Bk
2
! #
+
d k − pen
d b)1Ξ ]
+ 32E[ sup |hΘ̌n∧m , tiω |2 ] + 4E[(penk∨bk + pen
k
t∈BK α+
nm
≤ 7rωk γk−1 + 40
α+
N
n
X
k=0
E
"
e n∧m , tiω |2 − 33([ℓ]0 ∨ 1)δk
sup |hΘ
8n
t∈Bk
! #
+
1372
penk .
+ 32E[ sup |hΘ̌n∧m , tiω |2 ] +
3
t∈B α+
Knm
As in the proof of Theorem 4.1, the second and the third term are bounded applying Lemmata B.2 (with δk∗ := δk and ∆∗k := ∆) and B.3, respectively. Hence, by means of an obvious
adaption of statement a) in Lemma B.1 (with Nnα replaced by Nnα+ ) and the estimates
∆k ≤ d∆αk ,
δk ≤ dζd δkα ,
log(∆αk ∨ (k + 4))
δk
≥ 2kζd−1
∆k
log(k + 4)
with ζd = log(4d)/ log(4), we obtain in analogy to the way of proceeding in the proof of Theorem 4.1 that
b − λk2 1 ] .
sup sup E[kλ
ω Ξ
b
k
r
d
λ∈Λγ f ∈Fα
ωk δkα
min α− max
,
γk n
0≤k≤Knm
+ Φm +
1
.
n
(B.9)
Upper bound for 2 : The uniform upper bound for 2 can be derived in analogy to the bound
for 2 in the proof of Theorem 4.1 using Assumption B instead of statement b) from Lemma B.1
in the proof of Lemma B.5. Hence, we obtain
b − λk2 1Ξc ] . 1 .
sup sup E[kλ
ω
e
2
k
m
λ∈Λrγ f ∈Fαd
(B.10)
Upper bound for 3 : The term 3 can also be bounded analogously to the bound established for
3 in the proof of Theorem 4.1 (here, we do not have to exploit the additional Assumption B),
27
and we get
1
b − λk2 1 c
.
sup sup E[kλ
Ξ1 ∩Ξ2 ] .
e
k
n
λ∈Λrγ f ∈Fαd
(B.11)
Upper bound for 4 : To find a uniform upper bound for the term 4 , one can use exactly
the same decompositions as in the proof of the uniform upper bound for 3 in Theorem 4.1 by
replacing the probability of Ξc1 with the one of Ξc3 . Doing this, we obtain by means of Lemma B.6
that
1
b − λk2 1
.
(B.12)
sup sup E[kλ
Ξ1 ∩Ξ2 ∩Ξc3 ] .
e
k
r
m
d
λ∈Λγ f ∈Fα
The result of the theorem now follows by combining (B.9), (B.10), (B.11) and (B.12).
B.3. Auxiliary results
Lemma B.1 Let Assumption A hold. Then the following assertions hold true.
a) δjα /n ≤ 1 for all n ∈ N and j ∈ {0, . . . , Nnα },
−5 for all m ∈ N, and
α /(128d) ≤ C(d)m
b) exp −mαMm
2
−1 for all m ∈ N.
α |[f ]j | ≥ 2m
c) min1≤j≤Mm
α = 1 and there is nothing to show. Otherwise 0 < N α ≤
Proof. a) In case Nnα = 0, we have δN
α
n
n
α
n, and by definition of Nn we have (2j + 1)∆αj ≤ n/ log(n + 3) for 0 ≤ j ≤ Nnα which by the
definition of δjα implies that
δjα ≤
n
log(n/((2j + 1) log(n + 3)) ∨ (j + 3))
·
.
log(n + 3)
log(j + 3)
We consider two cases: In the first case, n/((2j + 1) log(n + 3)) ∨ (j + 3) = j + 3. Then n ≥ 1
directly implies the estimate δjα ≤ n. In the second case, we have n/((2j +1) log(n+3))∨(j +3) =
n/((2j + 1) log(n + 3)) and therefrom
δjα ≤ n log(n)/(log(n + 3) log(j + 3)) ≤ n,
and thus δjα ≤ n in both cases. Division by n yields the assertion of the lemma.
α > 0 for all sufficiently large m and that it is
b) Note that, due to Assumption A, we have Mm
α , we have
sufficient to show the desired inequality for such values of m. By the definition of Mm
−1
α ≥ 640dm
αMm
· log(m + 1) which implies
−5
α /(128d)) ≤ exp(−5 log m) = m
exp(−mαMm
,
and the assertion follows.
c) Take note of the observation that
min α |[f ]j |2 ≥
1≤j≤Mm
min α
1≤j≤Mm
α
αMm
αj
=
≥ 640m−1 · log(m + 1)
d
d
and 640m−1 · log(m + 1) ≥ 2m−1 for all m ≥ 1.
Lemma B.2 Let (δk∗ )k∈N0 and (∆∗k )k∈N0 be sequences such that for all k ≥ 1,
δk∗ ≥
X
ωj
|[f ]j |2
0≤|j|≤k
and
28
ωj
.
0≤|j|≤k |[f ]j |2
∆∗k ≥ max
Then, for any k ∈ {1, . . . , n ∧ m}, we have
E
"
∗
e n∧m , ti|2 − 33δk ([ℓ]0 ∨ 1)
sup |hΘ
8n
t∈Bk
≤ K1
(
! #
+
∗
kf k kλk ∆k
n
δk∗
exp −K2 ·
kf k kλk ∆∗k
!
)
√
δ∗
+ k2 exp −K3 n ,
n
with positive numerical constants K1 , K2 , and K3 .
Proof. The proof is a combination of the proofs of Lemma A.1 in [Kro16] (which deals with
the case ω ≡ 1) and Lemma A.4 in [JS13] (where general sequences ω are considered in the
framework with random variables instead of point processes). More precisely, one can apply
Proposition C.1 in [Kro16] with c(ε) from that statement replaced with c(ε) = 4(1 + 2ε) (this
δ∗
makes the proposition applicable also for complex-valued functions), M12 = δk∗ , H 2 = nk ([ℓ]0 ∨1),
1
υ := kλkkf k∆∗k ([ℓ]0 ∨ 1) and setting ε = 64
.
Lemma B.3 Let m ∈ N and k ∈ N0 . Then
"
2
sup E sup |hΘ̌n∧m , tiω |
λ∈Λrγ
t∈Bk
#
≤ C(d, r) · Φm .
Proof. Note that λ ∈ Λrγ implies
c] · 1 − 1|2 ]
E[ sup |hΘ̌n∧m , tiω |2 ] ≤ r sup ωj γj−1 E[|[f ]j /[f
Ωj
j
t∈Bk
−k≤j≤k
Thus, recalling the definition of Φm in (3.2), it suffices to show that
c] · 1 − 1|2 ] ≤ C(d, r) min{1, 1/(mα )},
E[|[f ]j /[f
Ωj
j
j
which can be realised by means of the identity
c] · 1 − 1|2 ] = E[|[f ] /[f
c] 1 − 1|2 · 1 ] + P(Ωc ) =: + △.
E[|[f ]j /[f
Ωj
j
Ωj
j
j
j Ωj
The bound ≤ C(d, r) min{1, 1/(mαj )} was already derived in the proof of Theorem 3.6. For
△, the corresponding upper bound can be obtained from statement c) of Lemma A.1.
Lemma B.4 Let Assumption A hold and consider the event Ξ1 defined in Theorem 4.1. Then,
for any n ∈ N, P(Ξc1 ) ≤ 2 exp(−Cn) with a numerical constant C = C(η) > 0.
Proof. Note that
c ∨ 1 < η([ℓ] ∨ 1)) + P([ℓ]
c ∨ 1 > η −1 ([ℓ] ∨ 1)),
P(Ξc1 ) = P([ℓ]
0
0
0
0
and the two terms on the right-hand side can be bounded by Chernoff bounds for Poisson
distributed random variables (see [MU05], Theorem 5.4). More precisely, we have
c ∨ 1 < η([ℓ] ∨ 1)) ≤ exp(−ω (η)n),
P([ℓ]
0
1
0
and
c ∨ 1 > η −1 ([ℓ] ∨ 1)) ≤ exp(−ω (η)n)
P([ℓ]
0
2
0
with ω1 (η) = 1 − η + η log η > 0 and ω2 (η) = 1 − η −1 − η −1 log η > 0 for all η ∈ (0, 1).
Lemma B.5 Let Assumption A hold and consider the event Ξ2 defined in the proof of Theorem 4.1. Then, for any m ∈ N, P(Ξc2 ) ≤ C(d)m−4 .
29
Proof. The complement Ξc2 of Ξ2 is
α
c] − 1| > 1/2 or |[f
c] |2 < 1/m}.
Ξc2 =:= {∃1 ≤ |j| ≤ Mm
: |[f ]j /[f
j
j
α }. In
Owing to statement c) from Lemma B.1 we have |[f ]j |2 ≥ 2/m for all j ∈ {1, . . . , Mm
c] |2 < 1/m a direct calculation using the reverse triangle inequality shows that
case that |[f
j
c] /[f ] −1| > 1/3,
c] /[f ] −1| ≥ 1/√2−1 > 1/4. In case that |[f ] /[f
c] −1| > 1 , one obtains |[f
|[f
j
j
j
j
j
j
2
and thus together we have
α
c] /[f ] − 1| > 1/4}.
Ξc2 ⊆ {∃1 ≤ |j| ≤ Mm
: |[f
j
j
Now, Hoeffding’s inequality implies
2
c] /[f ] − 1| > 1/4) ≤ 4 exp − m|[f ]j |
P(|[f
j
j
128
!
α
mαMm
≤ 4 exp −
128d
,
and the statement of the lemma follows from statement b) of Lemma B.1 and the estimate
α ≤ m which holds by definition of M α .
Mm
m
Lemma B.6 Let Assumptions A and B hold. The event Ξ3 defined in (B.2) satisfies P(Ξc3 ) ≤
C(α, d)m−4 for all m ∈ N.
Proof. Let us consider the random sets
α−
cnm } and
Ξ31 := {Nnα− ∧ Mm
>K
cnm > N α+ ∧ M α+ }.
Ξ32 := {K
n
m
Then, Ξc3 = Ξ31 ∪ Ξ32 and we establish bounds for P (Ξ31 ) and P (Ξ32 ), separately.
cm < K α− }. Owing to
bn < K α− } ∪ {M
Upper bound for P (Ξ31 ): We use the equality Ξ31 = {N
nm
nm
2
+
α−
the definition of Nn , we have |[f ]j | /((2j + 1)ωj ) ≥ 4 log(n + 4)/n for all j ∈ {0, . . . , Nnα− },
which yields
c] |2 /((2j + 1)ω + ) < log(n + 4)/n}
bn < K α− } ⊆ {∃1 ≤ j ≤ K α− : |[f
{N
nm
nm
j
j
⊆
⊆
[
α−
1≤j≤Knm
[
α−
1≤j≤Knm
c] |/|[f ] | ≤ 1/2}
{|[f
j
j
c] /[f ] − 1| ≥ 1/2}.
{|[f
j
j
cm < K α− } ⊆
In a similar way, we obtain {M
nm
α− ≤ M α+ by definition, we have
Mm
m
Ξ31 ⊆
[
α+
1≤j≤Mm
S
c
α− {|[f ] /[f ]j
j
0≤j≤Knm
− 1| ≥ 1/2}. Thus, since
c] /[f ] − 1| ≥ 1/2}.
{|[f
j
j
Applying Hoeffding’s inequality as in the proof of Lemma B.5 and exploiting Assumption B
yields
!
X
m|[f ]j |2
P(Ξ31 ) ≤ 4
exp −
≤ C(α, d) · m−4 .
(B.13)
128
α+
1≤|j|≤Mm
cm > K α+ }. In particular,
bn > K α+ } ∩ {M
Upper bound for P (Ξ32 ): First, note that Ξ32 = {N
nm
nm
α+
α+
α+
Knm < n ∧ m. If Knm = Nn < n, we obtain
c] |2 /((2j + 1)ω + ) ≥ log(n + 4)/n}
bn > N α+ } ⊆ {∀1 ≤ j ≤ N α+ + 1 : |[f
Ξ32 ⊆ {N
n
n
j
j
30
c] α+ |/|[f ] α+ | ≥ 2} ⊆ {|[f
c] α+ /[f ] α+ − 1| ≥ 1}.
⊆ {|[f
Nn +1
Nn +1
Nn +1
Nn +1
2
α+ = M α+ < m, using m−1 log m ≥ 4|[f ]
Analogously, if Knm
α+
m
Mm
+1 | yields
c] α+ /[f ] α+ − 1| ≥ 1}
cm > M α+ } ⊆ {|[f
Ξ32 ⊆ {M
m
Mm +1
Mm +1
c] α+ /[f ] α+
and thus Ξ32 ⊆ {|[f
Knm +1
Knm +1 − 1| ≥ 1}. Application of Hoeffding’s inequality and
exploiting Assumption B yields
P(Ξ32 ) ≤ 4 exp −
2
m|[f ]Knm
α+
+1 |
128
!
≤ 4 exp −
mαMm
α+
+1
128d
≤ C(α, d)m−5 .
(B.14)
The statement of the lemma follows by combining Equations (B.13) and (B.14).
C. Proofs of Section 5
C.1. Proof of Theorem 5.1
We define all the quantities appearing in this proof exactly as in the proof of Thereom 4.1 unless
otherwise stated. We use the decomposition
b − λk2 ] = + +
E[kλ
1
2
3
ω
e
k
established in the proof of Theorem 4.1 and use exactly the same arguments as in that proof to
bound the terms 2 and 3 . Thus, it remains to find an appropriate bound for 1 . In order to
get such a bound, we first proceed as in the proof of Theorem 4.1 in order to obtain on Ξ1 ∩ Ξ2
the estimate
b − λk2 ≤ 7rω γ −1 + 4pen
e n∧m , tiω |2 + 32 sup |hΘ̌n∧m , tiω |2 . (C.1)
g k − 4pen
g e + 40 sup |hΘ
kλ
k k
ω
e
k
k
t∈B
t∈B
e
Let us now introduce the quantity
X
λ̈k :=
0≤|j|≤k
e
k∨k
k∨k
c |ε]
E[[ℓ]
j
[f ]j
ej
where ε = (ε1 , . . . , εn ) is the vector containing the unobserved shifts εi in (2.5). Using the
e n∧m = λ
e n∧m − λn∧m = λ
e n∧m − λ̈n∧m + λ̈n∧m − λn∧m and setting
decomposition Θ
(1)
e n∧m − λ̈n∧m
Θn∧m := λ
and
(2)
Θn∧m := λ̈n∧m − λn∧m
we obtain from (C.1) that on Ξ1 ∩ Ξ2
(1)
b − λk2 ≤ 7rω γ −1 + 4pen
g k − 4pen
g e + 80 sup |hΘn∧m , tiω |2
kλ
k k
ω
e
k
k
t∈B
e
k∨k
+ 80 sup
t∈B
e
k∨k
(2)
|hΘn∧m , tiω |2
+ 32 sup |hΘ̌n∧m , tiω |2 .
t∈B
e
k∨k
Taking expectations, we obtain in analogy to the proof of Theorem 4.1 that
b − λk2 1Ξ ∩Ξ ] ≤ C
E[kλ
ω
1
2
e
k
min α max
0≤k≤Knm
ωk Dαk log(n + 2)
,
γk
n
31
X
E
"
X
E
"
α
Knm
+ 80
k=0
α
Knm
+ 80
k=0
E
100 log(n + 2)dDαk ([ℓ]0 ∨ 1)
−
n
! #
100 log(n + 2)dDαk ([ℓ]20 ∨ 1)
n
! #
(2)
sup |hΘn∧m , tiω |2 −
t∈Bk
+
+
+ 32E sup |hΘ̌n∧m , tiω |2 .
We have
"
sup
t∈Bk
(1)
|hΘn∧m , tiω |2
sup
t∈Bk
(1)
|hΘn∧m , tiω |2
t∈B
e
k∨k
(C.2)
! #
100 log(n + 2)dDαk ([ℓ]0 ∨ 1)
−
n
" "
=E E
+
100 log(n + 2)dDαk ([ℓ]0 ∨ 1)
(1)
sup |hΘn∧m , tiω |2 −
n
t∈Bk
!
+
##
|ε
We apply Lemma C.2 with δk∗ = dDαk in order to obtain
E
"
sup
t∈Bk
(1)
|hΘn∧m , tiω |2
100 log(n + 2)dDαk ([ℓ]0 ∨ 1)
−
n
!
+
.
|ε
#
q
Dαk
Dαk
+
exp(−K
n log(n + 2)).
2
n3
n2
Hence
E
"
sup
t∈Bk
(1)
|hΘn∧m , tiω |2
100 log(n + 2)dDαk ([ℓ]0 ∨ 1)
−
n
! #
.
+
q
Dαk
Dαk
+
exp(−K
n log(n + 2)).
2
n3
n2
α ≤ N α and hence by the definition of N α that
By definition of Ξ we have Knm
n
n
Dαk ≤ DαNnα =
X
0≤|j|≤Nnα
X
ωj
n
1
≤
.n
αj
log(n + 3) 0≤|j|≤N α 2|j| + 1
n
where we obtain the last estimate thanks to the logarithmic increase of the harmonic series. Due
α ≤ n we obtain
to Knm
α
Knm
X
E
k=0
"
sup
t∈Bk
(1)
|hΘn∧m , tiω |2
100 log(n + 2)dDαk ([ℓ]0 ∨ 1)
−
n
! #
.
+
1
.
n
Applying Lemma C.4 with δk∗ = dDαk we obtain that
E
"
sup
t∈Bk
(2)
|hΘn∧m , tiω |2
100 log(n + 2)dDαk ([ℓ]20 ∨ 1)
−
n
! #
+
.
Dαk
exp (−2 log(n + 2))
n
+
q
Dαk
exp(−K
n log(n))
2
n2
Using the relation Dαk . n from above we obtain
α
Knm
X
k=0
E
"
sup
t∈Bk
(2)
|hΘn∧m , tiω |2
100 log(n + 2)dDαk ([ℓ]20 ∨ 1)
−
n
32
! #
+
.
1
.
n
Finally, bounding the last term in (C.2) by means of Lemma B.3 we obtain from (C.2) using
the obtained estimates that
b − λk2 1Ξ ∩Ξ ] ≤
E[kλ
ω
1
2
e
k
min α max
0≤k≤Knm
ωk Dαk log(n + 2)
,
γk
n
+ Φm +
1
.
n
This shows the desired lower bound for 1 and combining it with the bounds for 2 and 3
yields the result.
C.2. Proof of Theorem 5.2
We use the same decomposition as in the Proof of Theorem 4.2. In particular, all the quantities
arising in the sequel are defined as in the proof of Theorem 4.2 unless otherwise stated. The
terms 2 , 3 , and 4 are bounded exactly as in the proof of Theorem 4.2 and it remains find
P
ω
an appropriate bound for 1 . Set Dk := 0≤|j|≤k |[f ]jj |2 and
c ∨ 1) ·
penk = 2000η −1 · ([ℓ]
0
Dk log(n + 2)
c 2 ∨ 1) · Dk log(n + 2) .
+ 2000η −2 · ([ℓ]
0
n
n
d k one immediately obtains that on Ξ
From the definition of penk and pen
d k ≤ 9penk
penk ≤ pen
from which one follows that
d k − pen
d b)1Ξ ≤ (penk + penb + pen
d k − pen
d b)1Ξ ≤ 10penk .
(penk∨bk + pen
k
k
k
Now, combining the arguments from the proofs of Theorem 4.2 and 5.1 one can show
b − λk2 1 ] .
sup sup E[kλ
ω Ξ
b
k
r
d
λ∈Λγ f ∈Fα
min α− max
0≤k≤Knm
ωk Dαk log(n + 2)
,
γk
n
+ Φm +
1
.
n
The statement of the Theorem follows now by combining the bounds established for 1 , 2 ,
3 , and 4 .
C.3. Auxiliary results
The following result is a conditional version of Proposition C.1 in [Kro16]. Since the proof is
exactly the same as the one in the unconditional case (with Poisson processes instead of Cox
processes) we omit its proof.
Proposition C.1 Let N1 , . . . , Nn be Cox processes onRa Polish space XR driven by finite random
P
measures η1 , . . . , ηn , respectively. Set νn (r) = n1 nk=1 { X r(x)dNk (x)− X r(x)dηk (x)} for r contained in a countable class R of complex-valued measurable functions. Denote η = (η1 , . . . , ηn ).
Then, for any ε > 0, there exist constants c1 , c2 = 16 , and c3 such that
E
"
2
sup |νn (r)| − c(ε)H
r∈R
2
!
≤ c1
+
(
|η
#
nH 2
2υ
exp −c2 ε
n
υ
33
!
√ nH
M2
+ 2 1 2 exp −c3 C(ε) ε
C (ε)n
M1
)
√
where C(ε) = ( 1 + ε − 1) ∧ 1, c(ε) = 4(1 + 2ε) and M1 , H and υ are such that
"
sup krk∞ ≤ M1 ,
#
E sup |νn (r)||η ≤ H, and sup Var
r∈R
r∈R
r∈R
Lemma C.2 Let (δk∗ )k∈N0 be a sequence such that δk∗ ≥
E
"
sup
t∈Bk
(1)
|hΘn∧m , ti|2
100 log(n + 2)δk∗ ([ℓ]0 ∨ 1)
−
n
≤ K1
!
+
P
#
Z
X
r(x)dNk (x)|η ≤ υ ∀k.
ωj
0≤|j|≤k |[f ]j |2
for all k ∈ N0 . Then,
|ε
q
δk∗ ([ℓ]0 ∨ 1)
δ∗
exp (−2 log(n + 2)) + k2 exp −K2 n log(n + 2)
n
n
,
with strictly positive numerical constants K1 and K2 .
Proof. Putting rt =
P
0≤|j|≤k
ωj [f ]−1
−j [t]−j ej , it is easy to check that given ε
(1)
hΘn∧m , tiω
n
1X
=
n i=1
Z
1
0
rt (x)(dNi (x) − λεi (x)dx)
where λε (x) = λ(x − ε − ⌊x − ε⌋). Thus, we are in the framework of Proposition C.1 and it
remains to find suitable constants M1 , H, and υ satisfying its preconditions.
Condition concerning M1 : We have
sup krt k2∞ = sup sup |rt (y)|2 ≤
t∈Bk y∈[0,1)
t∈Bk
X
ωj
≤ δk∗
2
|[f
]
|
j
0≤|j|≤k
1
and one can choose M1 = (δk∗ ) 2 .
Condition concerning H: We have
(1)
E[ sup |hΘn∧m , tiω |2 | ε] =
t∈Bk
([ℓ] ∨1)δ∗ log(n+2) 1/2
0
k
and one can choose H =
n
Condition concerning υ: It holds that
Var
Z
0
1
X
([ℓ]0 ∨ 1)δk∗ log(n + 2)
ωj
[ℓ]0
·
≤
n 0≤|j|≤k |[f ]j |2
n
rt (x)Nk (x) | ε =
Z
0
1
.
X
ωj
· [ℓ]0 ≤ δk∗ · ([ℓ]0 ∨ 1)
|rt (x)|2 λεk (x)dx ≤
2
|[f
]
|
j
0≤|j|≤k
and one can choose υ = δk∗ · ([ℓ]0 ∨ 1). The statement of the Lemma follows now by applying
Proposition C.1 with ε = 12.
Proposition C.3 Let X1 , . . . , Xn random variables with values in some Polish space. Let
P
νn (r) = n1 ni=1 r(Xi ) − E[r(Xi )] for r belonging to some countable set R of measurable complexvalued functions. Then, for all ε > 0, there are constants c1 , c2 = 61 and c3 such that
E
"
2
sup |νn (r)| − c(ε)H
r∈R
2
! #
+
≤ c1
(
nH 2
2υ
exp −c2 ε
n
υ
34
!
√ nH
M2
+ 2 1 2 exp −c3 C(ε) ε
C (ε)n
M1
)
,
√
with C(ε) = ( 1 + ε − 1) ∧ 1, c(ε) = 4(1 + 2ε) and
"
sup krk∞ ≤ M1 ,
#
E sup |νn (r)| ≤ H,
r∈R
r∈R
sup Var(r(Xk )) ≤ υ∀k.
r∈R
Proof. The proof consists of an easy adaption of the proof given in [Cha13] to the complex-valued
case and the statement follows by bookkeeping of the occuring numerical constants.
Lemma C.4 Let (δk∗ )k∈N0 be a sequence such that δk∗ ≥
E
"
sup
t∈Bk
(2)
|hΘn∧m , ti|2
100 log(n + 2)δk∗ ([ℓ]20 ∨ 1)
−
n
≤ K1
(
δk∗ ([ℓ]20 ∨
n3
1)
+
P
! #
ωj
0≤|j|≤k |[f ]j |2
+
([ℓ]20 ∨ 1)δk∗
n2
for all k ∈ N0 . Then
q
)
· exp(−K2 n log(n + 2)) .
P
Proof. We define rt′ (x) = 0≤|j|≤k ωj [f ]−1
−j [t]−j ej which coincides with the definition of rt in the
proof of Lemma C.2. Then, we have
(2)
hΘn∧m , tiω
R
n
1X
=
n i=1
Z
1
0
rt′ (x)λεi (x)dx
−
Z
1
0
rt′ (x)ℓ(x)dx.
Setting rt (εi ) = 01 rt′ (x)λεi (x)dx, we are in the framework of Proposition C.3 and its remains
to find suitable constants M1 , H and υ satisfying the preconditions of this proposition.
Condition concerning M1 : Note that the definition of rt′ is the same as the definition of rt in the
proof of Lemma C.2. Thus we obtain
sup krt k∞ = sup sup |
ε∈[0,1) t∈Bk
t∈Bk
Z
1
0
rt′ (x)λε (x)dx| ≤ (δk∗ )1/2 · sup
Z
1
ε∈[0,1) 0
λε (x)dx = (δk∗ )1/2 · ([ℓ]0 ∨ 1)
and we can take M1 = (δk∗ )1/2 · ([ℓ]0 ∨ 1).
Condition concerning H: We have
E[ sup
t∈Bk
(2)
|hΘn∧m , tiω |2 ]
ωj 1
E[|
≤
|[f ]j |2 n
0≤|j|≤k
≤
δ∗ log(n+2) 1/2
Var(rt (εi )) ≤ E
0
1
Z
1
0
ej (x)(λε1 (x)dx − ℓ(x)dx)|2 ]
δk∗ [ℓ]20
δ∗ [ℓ]2 log(n + 2)
≤ k 0
n
n
and we can set H = k n
Condition concerning υ: It holds
"Z
X
· ([ℓ]0 ∨ 1).
2
rt′ (x)λεi (x)dx
#
≤ [ℓ]20 · E
Z
1
0
|rt′ (x)|2
λεi (x)
dx ≤ ([ℓ]20 ∨ 1) · δk∗
[λ]0
and we define υ = ([ℓ]20 ∨ 1) · δk∗ . Now that statement of the lemma follows from Proposition C.3
with ε = 12.
References
[AB06]
Antoniadis, A. and Bigot, J. Poisson inverse problems. Ann. Statist. 34 (2006),
2132–2158.
35
[BB09]
Baraud, Y. and Birgé, L. Estimating the intensity of a random measure by histogram type estimators. Probab. Theory Related Fields 143 (2009), 239–284.
[BBM99]
Barron, A., Birgé, L., and Massart, P. Risk bounds for model selection via
penalization. Probab. Theory Related Fields 113 (1999), 301–413.
[Big+13]
Bigot, J. et al. Intensity estimation of non-homogeneous Poisson processes from
shifted trajectories. Electron. J. Stat. 7 (2013), 881–931.
[Bir07]
Birgé, L. Model selection for Poisson processes. In: IMS Lecture Notes - Monograph
Series. Institute of Mathematical Statistics, 2007, 32–64.
[Bré81]
Brémaud, P. Point processes and queues, martingale dynamics. Springer, 1981.
[Cav08]
Cavalier, L. Nonparametric statistical inverse problems. Inverse Problems 24 (2008).
[Cha13]
Chagny, G. Estimation adaptative avec des données transformées ou incomplètes.
Application à des modèles de survie. PhD thesis. 2013. url: https://tel.archives-ouvertes.fr/t
[CJ02]
Cavalier, L. and Ja-Yong Koo. Poisson intensity estimation for tomographic data
using a wavelet shrinkage approach. IEEE T. Inform. Theory 48 (2002), 2794–2802.
[CL10]
Comte, F. and Lacour, C. Pointwise deconvolution with unknown error distribution. C. R. Math. Acad. Sci. Paris 348 (2010), 323–326.
[CL11]
Comte, F. and Lacour, C. Data-driven density estimation in the presence of additive noise with unknown distribution: Data-driven Density Estimation. J. Roy.
Statist. Soc. Ser. B 73 (2011), 601–627.
[Com15]
Comte, F. Estimation non-paramétrique. Spartacus, Paris, 2015.
[CRT06]
Comte, F., Rozenholc, Y., and Taupin, M.-L. Penalized contrast estimator for
adaptive density deconvolution. Can. J. Stat. 34 (2006), 431–452.
[DH93]
Diggle, P. J. and Hall, P. A Fourier approach to nonparametric deconvolution of
a density estimate. J. Roy. Stat. Soc. Ser. B 55 (1993), 523–531.
[GN00]
Grègoire, G. and Nembè, J. Convergence rates for the minimum complexity estimator of counting process intensities. Nonparametric Statistics 12 (2000), 611–643.
[HMZ03]
Helmers, R., Mangku, I. W., and Zitikis, R. Consistent estimation of the intensity function of a cyclic Poisson process. J. Multivariate Anal. 84 (2003), 19–39.
[Joh09]
Johannes, J. Deconvolution with unknown error distribution. Ann. Statist. 37 (2009),
2301–2323.
[JS13]
Johannes, J. and Schwarz, M. Adaptive circular deconvolution by model selection
under unknown error distribution. Bernoulli 19 (2013), 1576–1611.
[Kar91]
Karr, A. F. Point processes and their statistical inference. Marcel Dekker, New
York, 1991.
[KR05]
Klein, T. and Rio, E. Concentration around the mean for maxima of empirical
processes. Ann. Probab. 33 (2005), 1060–1077.
[Kro16]
Kroll, M. Concentration inequalities for Poisson point processes with an application to intensity estimation from indirect observations. ArXiv e-prints (2016). arXiv:
1612.07901.
[Kut98]
Kutoyants, Y. A. Statistical inference for spatial Poisson processes. Springer, New
York, 1998.
[Mas00]
Massart, P. About the constants in Talagrand’s concentration inequalities for empirical processes. Ann. Probab. 28 (2000), 863–884.
36
[Mei09]
Meister, A. Deconvolution problems in nonparametric statistics. Springer, Berlin,
2009.
[MU05]
Mitzenmacher, M. and Upfal, E. Probability and computing: randomized algorithms and probabilistic analysis. Cambridge University Press, Cambridge, 2005.
[Neu07]
Neumann, M. H. Deconvolution from panel data with unknown error distribution.
J. Multivariate Anal. 98 (2007), 1955–1968.
[NH97]
Neumann, M. and Hössjer, O. On the effect of estimating the error density in
nonparametric deconvolution. J. Nonparametr. Stat. 7 (1997), 307–330.
[Pet95]
Petrov, V. V. Limit theorems of probability theory. Clarendon Press, Oxford, UK,
1995.
[PW04]
Patil, P. N. and Wood, A. T. A. Counting process intensity estimation by orthogonal wavelet methods. Bernoulli 10 (2004), 1–24.
[RB03]
Reynaud-Bouret, P. Adaptive estimation of the intensity of inhomogeneous Poisson processes via concentration inequalities. Probab. Theory Related Fields 126 (2003),
103–153.
[Rei89]
Reiss, R.-D. Approximate distributions of order statistics. Springer, New York, 1989.
[Rei93]
Reiss, R.-D. A course on point processes. Springer, New York, 1993.
[Res87]
Resnick, S. I. Extreme values, regular variation and point processes. Springer, 1987.
[San14]
Sansonnet, L. Wavelet thresholding estimation in a Poissonian interactions model
with application to genomic data. Scand. J. Statist. 41 (2014), 200–226.
[Ser09]
Serfozo, R. Basics of applied stochastic processes. Springer, Berlin, 2009.
[Sto+13]
Stoyan, D. et al. Stochastic Geometry and its Applications. Third edition. John
Wiley & Sons, 2013.
[SVB10]
Schwarz, M. and Van Bellegem, S. Consistent density deconvolution under partially known error distribution. Statist. Probab. Lett. 80 (3-4) (2010), 236–241.
[SZ12]
Shen, J. J. and Zhang, N. R. Change-point model on nonhomogeneous Poisson processes with application in copy number profiling by next-generation DNA sequencing.
Ann. Appl. Statist. 6 (2012), 476–496.
[Tsy08]
Tsybakov, A. B. Introduction to nonparametric estimation. Springer, New York,
2008.
[ZK10]
Zhang, T. and Kou, S. Nonparametric inference of doubly stochastic Poisson process data via the kernel method. Ann. Appl. Statist. 4 (2010), 1913–1941.
37
| 10math.ST
|
Community detection with spiking neural networks for
neuromorphic hardware∗
arXiv:1711.07361v1 [cs.NE] 20 Nov 2017
Kathleen E. Hamilton
Oak Ridge National Laboratory
Computing & Computational
Sciences Dir
One Bethel Valley Road
Oak Ridge, Tennessee 37831-6015
[email protected]
Neena Imam
Oak Ridge National Laboratory
Computing & Computational
Sciences Dir
One Bethel Valley Road
Oak Ridge, Tennessee 37831-6015
[email protected]
Travis S. Humble
Oak Ridge National Laboratory
Computing & Computational
Sciences Dir
One Bethel Valley Road
Oak Ridge, Tennessee 37831-6015
[email protected]
ABSTRACT
1
We present results related to the performance of an algorithm for
community detection which incorporates event-driven computation. We define a mapping which takes a graph G to a system
of spiking neurons. Using a fully connected spiking neuron system, with both inhibitory and excitatory synaptic connections,
the firing patterns of neurons within the same community can be
distinguished from firing patterns of neurons in different communities. On a random graph with 128 vertices and known community
structure we show that by using binary decoding and a Hammingdistance based metric, individual communities can be identified
from spike train similarities. Using bipolar decoding and finite rate
thresholding, we verify that inhibitory connections prevent the
spread of spiking patterns.
Graph partitioning and community detection are ubiquitous tasks
encountered in a diverse set of sciences and many methods have
been developed to sort the vertices of a graph, or nodes of a network into classes based on similarity measures. These methods
utilize graphical characteristics and structures, such as transitivity,
modularity or betweenness, and spectral-based analysis and often
require large scale matrix analysis [2, 4, 5, 24]. These methods can
be parallelized and many algorithms exist for the analysis of very
large networks. However, the emergence of unconventional processors, such as neuromorphic processors, requires approaches which
utilize event-based computation. We present work in this paper
related to the development of an algorithm which incorporates
event-based computation in the identification of related vertices in
networks or graphs.
Inspired by the recent work of Shaub et al [33] in which the
different approaches to community detection are organized according to the problem they are designed to solve (e. g. partitioning
problems, clustering problems, dynamical problems), we describe
our approach as a hybrid dynamical clustering method which incorporates the discrete time signals of a spiking neuron system. In this
paper we present results that serve as a proof-of-concept related
to the mapping of recurrent neural networks based on interacting
spin dynamics (Hopfield networks) to spiking neural systems.
Our goal is to construct a system of spiking neurons which can be
used to generate a set of spike responses which can identify vertex
communities in a graph. We choose to characterize a community
in a graph G(V, E) as a subset of vertices v ∈ V such that the
density of edges internal to this subset is higher than the density of
edges connecting to the remainder of the graph. This is similar to
definitions used in other cluster based models [27]. There is a well
known aphorism in Hebbian learning, attributed to Siegrid Löwel
[21]: “neurons that fire together are wired together.” We use the
statement inversion:“neurons that are wired together, fire together”
to construct our approach to community detection.
The vertices and edges of a given graph G are mapped to a network of symmetrically connected spiking neurons, which is then
selectively driven by time-dependent external currents. There must
be a degree of similarity between the resulting spike trains which
can be used to distinguish individual communities of G. We binary decoding and bipolar decoding of spike trains with similarity
between trains measured using a Hamming distance metric. The
ability to identify individual communities is dependent on the linear
CCS CONCEPTS
•Mathematics of computing → Graph algorithms; Random
graphs; •Hardware → Emerging tools and methodologies; Neural systems;
KEYWORDS
neuromorphic, community detection, spiking neural networks
ACM Reference format:
Kathleen E. Hamilton, Neena Imam, and Travis S. Humble. 2017. Community detection with spiking neural networks for neuromorphic hardware.
In Proceedings of ORNL Neuromorphic workshop, Oak Ridge, Tennessee USA,
July 2017 (Neuromorphic Computing), 7 pages.
DOI:
∗ This
work was supported by the United States Department of Defense and used
resources of the Computational Research and Development Programs at Oak Ridge
National Laboratory. This manuscript has been authored by UT-Battelle, LLC, under
Contract No. DE-AC0500OR22725 with the U.S. Department of Energy. The United
States Government retains and the publisher, by accepting the article for publication,
acknowledges that the United States Government retains a non-exclusive, paid-up,
irrevocable, world-wide license to publish or reproduce the published form of this
manuscript, or allow others to do so, for the United States Government purposes. The
Department of Energy will provide public access to these results of federally sponsored
research in accordance with the DOE Public Access Plan.
Permission to make digital or hard copies of part or all of this work for personal or
classroom use is granted without fee provided that copies are not made or distributed
for profit or commercial advantage and that copies bear this notice and the full citation
on the first page. Copyrights for third-party components of this work must be honored.
For all other uses, contact the owner/author(s).
Neuromorphic Computing, Oak Ridge, Tennessee USA
© 2017 Copyright held by the owner/author(s). .
DOI:
INTRODUCTION
Neuromorphic Computing, July 2017, Oak Ridge, Tennessee USA
separability of the Hamming distance metric values. This is controlled by the size of the bin width ∆t used in the binary decoding
of individual spike trains.
This approach incorporates Hopfield networks [13–15] and spin
glass models [1, 31, 32]. It has been shown that recurrent networks
with steady states can be used to find solutions to the problem
of graph partitioning [12, 36]. The use of positive and negatively
weighted edges is needed to drive the system to a solution which
meets the optimization conditions. However, these approaches
are often limited to finding partitions of equal sizes in a graph, or
require prior knowledge of the number of communities to find.
Hopfield networks have been implemented using spiking neurons for the task of content addressable memories and pattern
retrieval [23, 34], in this work we focus on the application of the
recurrent neural network model to a task related to graph partitioning. We show in this paper that in conjunction with carefully
chosen parameters for our neuron model, these spin-glass based
models can be used to find un-equal sized groups. We establish that
spiking data can be used to identify communities.
Recent works on community detection have focused on how
real-world networks may have ambiguous community structure
and there are limitations how much information can be inferred
from metadata based approaches [29]. In this work we focus on a
discussion of how spiking data can be used to identify communities
and we work with graphs generated with known communities and
fixed labels. The graphs analyzed in this paper are all instances
of a specific class of random graphs: Girvan-Newman benchmark
graphs [9]. These graphs have 128 vertices organized into 4 equalorder communities of 32 vertices each. The average degree is hdi =
16. Using these graphs we demonstrate our spike-based approach
to community detection, focusing on binary decoding and bipolar
decoding of spike trains.
A graph is mapped to a system of spiking neurons and driven
in such a manner that the generated spike trains can be used to
reconstruct the known community labels of the original graph.
Two neurons (ni , n j ) in Q i=j must have firing patterns that exhibit
a degree of similarity which distinguishes them from the spike
trains generated by neurons (ni , n j ) in Q j,i . In Section 2, we derive
the theoretical structure which underlies the main components
of our approach. We begin with how a graph G is mapped to a
spiking neural network (SNN). Then, we discuss how the physical
parameters associated with the spiking neuron system must be set
and how the selective driving of this SNN must be done in order to
generate a set of spike trains are generated that can distinguish a
single community Q i from the remaining Q j,i . In Sections 3 and 4
we describe how spike responses are decoded and we discuss the
metrics we use to analyze decoded spike trains. This paper presents
on results which establish proof-of-concept that a mapping and
driving pattern exists which can be used to generate spike trains
characteristic of a graph’s known community structure. In Section 5
we introduce how our spiking neural networks can be incorporated
into existing community detection algorithms.
K. Hamilton et al.
2
SPIKING NEURON MODEL
CONSTRUCTION
We focus on community detection in undirected, unweighted graphs.
A graph is defined by a vertex set and an edge set G = G(V, E).
Multiple edges and self-loops are not allowed. In this initial work,
we study graphs with clearly delineated (non-overlapping) communities.
SNNs are dynamical systems which compute without the use of
steady states [22, 23]. Information is transmitted through electrical pulses. The neurons which compose the spiking network are
nonlinear units and can exhibit a rich set of dynamics and firing
patterns based on the physical parameters. We use few parameters
to build our SNNs; they are leaky-integrate and fire neurons defined by a threshold voltage (vth ), a refractory period (tr ), and a time
constant (τ ). The full spiking network itself S(N, W) is defined by a
set of homogeneous neurons N = {ni }, and a set of symmetrically
connected synapsesW = {s(w)i j }, weighted edges which connect
neurons ni ↔ n j .
For a graph G(V, E), we assume there is a set of known vertex
Ð
communities {Q i } such that i Q i = V(G) and no vertex exists
in more than one community. The remainder of this paper will
develop how spiking systems can identify individual communities
Qi .
The mapping of an undirected, unweighted graph to SNN is analogous to the construction of a Hopfield recurrent neural network;
using symmetric connections which are either positive or negatively
weighted. The two-step mapping first defines a SNN by defining
a spiking neuron for each vertex on the graph: vi ∈ V → ni ∈ N,
and each edge on the graph defines a symmetric pair of excitatory
synapses ei j ∈ E → s(w + )i j , s(w + )ji ∈ W, w + > 0. This SNN is
then transformed to a fully connected SNN by the addition of symmetric pairs of inhibitory synapses s(w − )jk , s(w − )k j ∈ W, w − < 0
for any edges which do not exist on the original graph. The magnitudes of the excitatory and inhibitory synapses are equal |s(w + )| =
|s(w − )|.
Leaky integrate and fire neurons are simplified models of neuronal behavior [8]. We work with these models because of their
close proximity to the behavior of the IBM TrueNorth processor
[3, 25]. The resulting networks of spiking neurons do not require
extensive detail about the biological behavior, nor are they an attempt to describe cortical network dynamics. We focus on a model
of the membrane potential in which the potential vi (t) of a single
neuron is a time-dependent function which changes depending on
discrete or continuous inputs. Discrete signals are measured upon
the arrival of positively and negatively weighted spikes ∆v = s(w ± ),
any external driving force term in Vex t (t) is assumed to be continuous. Additionally, we use the existence of a non-zero leak τ to
continuously relax v(t) to an equilibrium value vr .
d(vi (t)) Vex t (t) − vi (t)
=
dt
τ
Õ
Vex t (t) = Iex t (t)R +
s(w)i j δ (t − t 0 )
(1)
(2)
t 0, j→i
If a neuron’s potential exceeds the firing threshold vth , it fires a
spike along its synaptic connections and enters a refractory period
t R during which its potential is not changed. The choice of the
Community detection with spiking neurons
Neuromorphic Computing, July 2017, Oak Ridge, Tennessee USA
neuron system parameters is determined by whether they are used
to generate similar spike trains, or to enhance dissimilarity between
spike trains. The refractory period t R is needed to impose a sense
of directionality to the spread of spiking patterns. Aided by the
positively weighted synapses, spike responses spread to neurons
that have not yet fired. These correspond to connected vertices
on the graph and lead to similarities in spike trains of neurons
ni , n j in the same community Q i . The parameters that lead to
dissimilar spike trains are those which help to impede the spread
of spiking patterns across several neurons in a system, primarily
the negatively weighted synaptic connections. Many parameters
play a dual role, necessary for both creating similar spike trains and
to inhibit spike pattern spread. The time constant (τ ) is balanced
to ensure arriving spikes can accumulate and lead to secondary
firing while also ensuring that the effects of spike impulses are not
long-lived, the firing threshold vt h ensures that a neuron fires in
response to incoming spike impulses only when multiple impulses
arrive in a short time window.
For a set of homogeneous spiking neurons, there are many system parameters which are tunable in order to generate spiking
patterns which are informative about the original graph’s community structure. Previously, we have studied how the simplest
Hopfield network could be mapped to a system of spiking networks
[11] we considered graphs with fully connected communities connected by a single bridge bond (barbell graphs). These networks
could be driven using a pair of sinusoidal currents out of phase
by 180 degrees. The negative driving current, as well as a careful tuning of neuron system parameters (refractory period, time
constant, firing threshold and synaptic weight) was sufficient to
generate spike trains characteristic of the two communities. In this
work we generalize our approach to a system connected as a spin
glass, the role of the negative driving current is replaced by the
negatively weighted synapses, which inhibit the growth and limit
the occurrences of spike cascades.
The neuron dynamics under square pulse driving are relatively
simple. Only one neuron is actively driven by a square pulse at
any time (t), and the effects from the refractory period t R allow
for the equations of motion in Eq. (2) to be separated as: a set of
equation to describe the firing dynamics under active driving, and
one set of equations to describe the firing dynamics as a reaction to
a neighbor being actively driven. The equation for the square pulse
is,
be found by integrating the equation of motion:
Vex t (t) = Amax [tanh (β[t − t 1 ]) − tanh (β[t 2 − t])]
(3)
The shape of the square pulse is determined by the pulse height
(Amax ), pulse width (t A = t 2 − t 1 ), and the parameter β which
determines the sharpness of the pulse’s rise. The gap between
subsequent pulses applied to different neurons is sufficiently large
such that any ∆V induced by Vex t on neuron ni has decayed away
before neuron n j is actively driven. The parameter β controls the
sharpness of the square pulse. A sharp step is needed to ensure
that spikes are fired only when the external current is a constant
driving force, Imax = 2Amax /R, however the function Vex t (t) must
remain continuous. Under active square pulse driving, effect of the
constant driving force leads to a constant firing rate (δ ), which can
d(v(t)) Vex t (t) − v(t)
=
,
(4)
dt
τ
and is found in terms of the time constant τ , the square pulse amplitude Amax , the reset voltage v 0 and the spike threshold voltage
vth :
δ
2Amax − v 0
.
(5)
= log
τ
2Amax − vth
The firing rate δ can be used to determine the synaptic weight sw
with this assumption that a neuron (n j ) will fire a spike in response
to one of its nearest neighbors (ni , s(w + )(i→j) > 0) being actively
driven when 2 spikes arrive at n j in a short time span. Setting
sw = αvth (0 < α < 1), the first spike arrives at t 0 and increases
the potential v j (t 0 ) = v 0 + s(w + ). The arrival of the second spike
happens at t = δ and the resulting potential must exceed the spike
threshold:
v j (t 0 + δ ) > vth .
(6)
Again integrating the dynamical equation for v j during time t :
[t 0 , t 0 + δ ] is straightforward since there is no active driving. The
inequality in Eq. (6) becomes:
sw e −δ /τ + 1 > vth ,
(7)
α
δ
< log
.
(8)
τ
1+α
3
BINARY DECODING AND SPIKE TRAIN
SIMILARITY
For each neuron ni ∈ N(S), there is an associated set of spike
times called a spike train. The spike train data is analyzed using a
comparison matrix, where similarity between trains is quantified
by a similar metric as one used in Ref. [16]. Binary decoding
(m)
converts the spike trains to binary vectors x i (∆t) using a discrete
time step ∆t; x i has a value of 1 if at least one spike occurs in
the time window ∆t. The label (m) assigned to a given neuron is
(m)
included for completeness. The set of all {x i (∆t)} binary vectors
of length |x i | are compared pairwise to construct entries in an
|N| × |N| comparison matrix H with entries Hi j defined by the
normalized Hamming distance between two binary decoded spike
(n) (m)
trains: Hi j = 1 − h(x i , x j )/|x i | [16]. We introduce a modified
version of this comparison matrix:
h
i
(n)
(m)
h x i (∆t), x j (∆t) ª
©
® (1T xi )(1T xj ).
Hi j = 1 −
(9)
®
|x i |
«
¬
The inclusion of the terms: 1T xi , 1T xj down weights the entries
of the matrix Hi j in which the spike train of a firing neuron is
compared to a non-firing neuron (e.g. x i = {0} ⊗ |x i | ). Later (see
Sec. 5), as we begin to develop scalable methods for real-world
analysis, these terms will be omitted and the Hamming metric will
be defined as 0 when either x i or x j corresponds to a non-firing
neuron.
In the initial tests of our algorithm, we use the benchmark graphs
of Girvan-Newman and Fortunato (see [9, 19, 20, 26]) which have
equal-sized, strongly-connected communities with minimal overlap
Neuromorphic Computing, July 2017, Oak Ridge, Tennessee USA
K. Hamilton et al.
Q1
Q2
Q3
Figure 1: The Girvan-Newman benchmark graph instance
generated by software available from [6] and used to generate the spiking data analyzed in this paper.The different
communities are distinguished by vertex color: Community
1 (red), Community 2 (blue), Community 3 (black), Community 4 (grey).
(see Fig. 1). Using the software available at [6], we generate instances of the random Girvan-Newman graphs (and the known community memberships), map them to a SNN of homogeneous neurons, and simulate the spiking dynamics using the Brian2 Python
library [10]. A SNN is constructed with the parameters: τ = 25 ms,
vth = 0.8V, |s(w + )| = |s(w − )| = 0.75 V, v 0 = 0 V and t R = 20 ms.
The spiking data analyzed in this paper are all generated from
the same graph instance (shown in Fig. 1). Three of the four communities are driven: Q 1 , Q 2 , Q 4 . Neurons are driven by individual
square pulses of maximum height Amax V, and width τA = 200 ms
with a gap between subsequent pulses of ∆tpul se = 800 ms. The
complete set of neurons which are driven is ordered by community:
Q 1 is driven first, then Q 2 , then Q 3 . During driving, the primary
firing neuron fires 10 spikes, separated by nearly uniform time
intervals: δ 1 ≈ 21 ms. The secondary firing neuron fires 5 spikes at
nearly uniform time intervals: δ 2 ≈ 42 ms. Slight variations in the
time interval between spikes are possible due to spikes fired when
Vex t / Imax R, and approximations introduced during numerical
integration of the equations of motion.
From the generated spike trains, the comparison matrix is constructed using binary decoding with ∆t = 8.00 sec. In Fig. 2 the
three driven communities can be identified by the magnitude of
Hi j and the block matrix structure. Along the diagonal, the value of
(n)
Hi j is significantly higher between a driven neuron x i
(n)
neuron x j
and target
in the same community, compared to an off-diagonal
(n)
(m)
block in which a driven neuron x i and target neuron x j are
different communities. The un-driven neurons of Q 3 do not exhibit
a strong similarity to any of the driven neurons in Q 1 , Q 2 , Q 4 .
4
BIPOLAR DECODING AND FIRING RATE
THRESHOLDING
The concept of a community in Section 1 is now used to decode
spiking output. During the community-ordered driving sequence
used in Section 3, we conjecture that neurons undergoing active
Q4
Figure 2: The comparison matrix for a Girvan-Newman
benchmark graph instance. Community three (Q 3 )was not
driven and is difficult to discern from the noisy background
as a coherent spiking pattern is not able to spread from one
community to the next. The Hamming metric used to generate this matrix included terms to down-weight low spiking
neurons.
driving and contained in the same community will exhibit higher
firing rates that those outside of the community. Only two neuronal
states, corresponding to the Hopfield network states +1 and −1, are
allowed. These states are mapped to spike train patterns through
the use of time window binning. This is a variation on the binary
code discussed in Section 3 but now we consider the total number
of spikes fired by a neuron during ∆t. The Hopfield network state
+1 is mapped to a neuron which has a firing rate over a fixed time
window ∆t which exceeds a pre-determined threshold value fi ≥ f 0 ,
this is considered to be an active state. The Hopfield network state
−1 is mapped to a neuron which has a firing threshold over a
fixed time window ∆t which does not exceed a pre-determined
threshold value fi < f 0 , this is an inactive state. Similar approaches
to mapping Hopfield networks to spiking neural networks have
been investigated for the task of pattern retrieval using FitzHughNagumo neuron models [17, 18].
An upper bound on f 0 can be found by looking at the spiking
response of a fully-connected K 32 community. For a fully connected
clique, where each neuron is actively driven once, the maximum
number of spikes that a neuron can fire is,
fmax = r 1 + (n − 1)r 2 ,
(10)
where r 1 is the active firing rate and r 2 is the response firing rate.
The minimum firing threshold value is still under investigation,
but we use a simple heuristic to approximate f 0 . For the GirvanNewman benchmark graphs studied in this paper, the communities
are known to each be of order 32 and hdi = 16. For a lower bound on
the minimum number of spikes a neuron can fire and still be counted
as being a member of a community, we make the assumption that
at least half of a vertex’s neighbors must be in the same community.
Neuromorphic Computing, July 2017, Oak Ridge, Tennessee USA
Spike count
1.0
200
Spike count for neurons with label 1
Spike count for neurons with label 2
200
175
175
150
150
125
0.8
100
125
Spike count
Community detection with spiking neurons
75
100
75
50
50
0.6
25
25
0
0
[0:00-0:33] [0:33-1:06] [1:06-1:39]
Time interval [mm:ss]
Figure 3: The full spike raster for a Girvan-Newman benchmark graph instance. Only communities Q 1 , Q 2 and Q 4 were
driven. Each neuron in the communities was actively driven
once. The vertical lines mark: t = 1 (sec), the start of Q 1 driving; t = 33 (sec), the end of Q 1 driving and the start of Q 2
driving; t = 65 (sec), the end of Q 2 driving and the start of Q 4
driving; and t = 97 (sec) the end of Q 4 driving.
We replace the factor of (n − 1) in Equation (10) with 16/2 + 1,
fmin = r 1 + 9r 2 .
(11)
Further study into a robust lower bound would require a stricter
definition of what is a community. For our task of reconstructing
a known set of labels, and with the non-overlapping community
structure of the Girvan-Newman benchmarks, this heuristic will
suffice. Inhibitory synapses impede the spread of spike cascades
throughout the entire neuron system, as shown in Fig. 2 and 4. For
an instance of the Girvan-Newman benchmark graph it is shown
that when 3 of the 4 communities are driven, the similarity between
spike trains in the same driven community is significantly higher
than the similarity between spike trains in different communities.
Additionally, the un-driven community never exhibits a significant degree of spike response, showing that the spread of spiking
synchronicity throughout the entire network is impeded.
5
IDENTIFYING UNKNOWN COMMUNITIES
Future development of spike-based community detection is dependent on how well spiking neuron systems can be incorporated
into existing algorithms. There exists a near linear algorithm for
community identification called “label propagation,” introduced in
2007 by Raghavan, Albert and Kumara [30]. The similarity between
this method and a Potts spin model have been noted [35], and we
consider label propagation to be an algorithm that can incorporate
spike-based data. However there are several obstacles to overcome
before spiking neuron systems can be incorporated into real-world
community identification tasks.
First, we need to know how a neuron’s spike response is affected
when the neurons are driven in a random order. We use the GirvanNewman benchmark graph shown in Fig. 1, but instead of the
200
0.4
175
200
150
150
125
125
100
0.2
75
175
100
75
50
50
25
25
0.00
0.0 [0:00-0:33] [0:33-1:06]
0.2
[1:06-1:39]
0.4
Time interval [mm:ss]
[0:00-0:33] [0:33-1:06] [1:06-1:39]
Time interval [mm:ss]
Spike count for neurons with label 4
Spike count
Spike count
Spike count for neurons with label 3
0
[0:00-0:33]
0.6
[0:33-1:06]
0.8
[1:06-1:39] 1.0
Time interval [mm:ss]
Figure 4: The spike counts for neurons in each labelled community during three time windows of width ∆t ≥ 32 seconds
showing how neurons in each community react when active
driving is applied to neurons in Q 1 , Q 2 , Q 4 (respectively). The
(red, dashed) horizontal is an upper bound on the total number of spikes that would be generated under square pulse
driving of a K 32 community. The (black,dashed) line is a
lower bound generated by using the mean degree hdi = 16
and the assumption that at least half of a neuron’s neighbors have the same community label.
community-ordered driving used in Sections 3 and 4, we randomly
permute and drive the entire neuron set. Randomly permuting
the neuron set reduces the usefulness of bipolar decoding and the
spiking data analysis in this section only uses binary decoding.
Additionally, when randomly driving neurons, ∆t must be carefully
chosen such that the response between subsequent driven neurons
are not covered by a single time window. If that happens, then any
distinction between signals belonging to individual neurons which
may have different community labels may be lost. There is now an
upper bound on the time window: ∆t < t A .
Second, the scalability of the Hamming metric of Eq. 9 is quite
poor. Analysis of larger graphs will require driving a large number of neurons and the length of the binary decoded spike trains
will rapidly increase. As the length a binary decoded spike train
increases, the similarity between two neurons that fire frequently
will become quite large. To remedy this, we return to the original
implementation of Hamming metric as defined in [16].
Returning to the Hamming metric as originally defined in [16]
does not affect the comparison matrix structure (see Fig. 5) and
using the Hamming metric, we show how spiking data can be
used in label propagation algorithms. Label propagation is done by
choosing a vertex to act as a “source,” fixing its label, then updating
Neuromorphic Computing, July 2017, Oak Ridge, Tennessee USA
Q1
Q2
Q3
Q4
Figure 5: The comparison metric of a Girvan-Newman graph
instance on 128 vertices: the complete neuron set is first randomly permuted, then every neuron is individually driven
by a square pulse. The time window ∆t = 30 ms was chosen
such that δ 1 < ∆t < δ 2 t A = 200 ms. The Hamming metric
does not include any down-weighting terms, and the scale
is now limited to the range [0.0, 1.0]. Along the matrix diagonal, it is seen that H (x i , x i ) = 1.00.
the labels of all its neighboring vertices. For spiking data to be
incorporated into this method, we need to ensure that the choice
of source does not significantly impact the linear separability of
the Hamming metric values. The role of ∆t in tuning this quality is
shown in Fig. 6: at ∆t = 3.2 (sec) the mean Hamming metric value
is defined as:
1 Õ
hH (x im , x jn )ii =
H (x im , x jn ).
(12)
|Qm |
i ∈Q n
i,j
If m = n, the self-similarity value H (x im , x im ) = 1.00 is excluded.
When the source and target are in different labelled communities
(n , m) the mean metric is nearly overlapping, but when the source
and target are in the same community, the mean is nearly 1.00. The
linear separability of the mean metric value is dependent on the
size of ∆t, as seen in Fig. 6.
6
CONCLUSIONS
Our approach to a community detection using spike-based computing depended on vertices within the same community exhibiting similar spiking patterns. We have studied how SNNs can be
constructed, driven by external current, and decoded in order to
generate such spiking patterns. An undirected graph is mapped to
a fully connected set of homogeneous spiking neurons, with both
excitatory and inhibitory synapses and selectively driven by square
pulse currents. The neurons contained within densely connected
regions exhibit synchronicity in their firing patterns.
Spike trains are decoded by time window binning, using varying
window widths (∆t). For ∆t < t A the resulting binned vectors
K. Hamilton et al.
are converted into binary codes, with entries {0, 1} only. From
these binary vectors and a Hamming-distance based metric, we
construct a comparison matrix which had dimension n × n for a
graph of order n. The degree of similarity between spike trains can
be quantified and used to infer the membership of communities
on the original graph. For ∆t ≈ |Q i |(t A + δpulse ) the resulting
vectors tabulate the number of spikes fired during ∆t and use a
pre-determined threshold to determine the appropriate label to
assign to the neurons.
Using binary decoding, the driven communities could be distinguished by the Hamming metric similarity. Simulations of spiking neuron dynamics demonstrated the ability to identify nonoverlapping, communities of equal order (see Figs. 2, 5) for graphs
with known community structure. In Fig. 2, it was seen that the
use of inhibitory synaptic connections prevented any large scale
spike cascade. The Hamming metric can distinguish a single community when a neuron system is driven sequentially according to
community, or when the neuron set is randomly permuted. Bipolar decoding, assigning community labels according to the spike
counts, can distinguish the driven communities by the dominant
spiking behavior. However, the bipolar decoding is only used when
a neuron system is driven sequentially according to community.
The results presented here are generated from one instance of the
Girvan-Newman benchmark graph (shown in Fig. 1), multiple
instances have been tested with similar results. Additional simulations have also been tested on graphs with unequal-sized communities. We explore how our method performs on un-ordered data,
by randomly permuting and driving the entire neuron set. In Fig. 6,
the linear separability of the Hamming metric value shows that for
small enough ∆t value, the Hamming metric can distinguish one
community from the remaining 3 communities.
Our results show that spiking data can be used to identify communities in undirected graphs. The separability of the Hamming
similarity measure (see Fig. 6) could be incorporated into a label
propagation workflow [7, 28, 30]. Community detection is a unique
use of neuromorphic hardware, utilizing neural systems that do
not have hidden units, or require training over large data sets.
Future work related to spike-based community detection will
investigate two areas: the applicability, and the scalability of this
method. Applicability focuses on quantifying the limits of a spikebased approach: effectiveness on other graph classes, identification
of overlapping communities, and establishing a resolution limit
(smallest community that can be identified in a large network).
Scalability addresses how this method can be applied to large network analysis. In particular, as presented in this work, the binary
analysis has the potential for poor scaling: a system of n vertices is
mapped to a densely connected system of n neurons and produces
a matrix of dimension n × n.
REFERENCES
[1] Marcelo Blatt, Shai Wiseman, and Eytan Domany. 1996. Superparamagnetic
Clustering of Data. Phys. Rev. Lett. 76 (Apr 1996), 3251–3254. Issue 18. https:
//doi.org/10.1103/PhysRevLett.76.3251
[2] Stefano Boccaletti, Vito Latora, Yamir Moreno, Martin Chavez, and D-U Hwang.
2006. Complex networks: Structure and dynamics. Physics reports 424, 4 (2006),
175–308.
[3] Andrew S Cassidy, Paul Merolla, John V Arthur, Steve K Esser, Bryan Jackson,
Rodrigo Alvarez-Icaza, Pallab Datta, Jun Sawada, Theodore M Wong, Vitaly
Feldman, et al. 2013. Cognitive computing building block: A versatile and efficient
Community detection with spiking neurons
Neuromorphic Computing, July 2017, Oak Ridge, Tennessee USA
Figure 6: For an instance of the Girvan-Newman graph on 128 vertices, the complete neuron set is randomly permuted, then
every neuron is individually driven by a square pulse. The mean Hamming metric hH (x im , x m
j )ii for source and target in the
same community (black) can be distinguished from the mean Hamming metric mean Hamming metric hH (x im , x jn )ii for source
and target in different communities (red, blue, green) if the bin width ∆t is small. The standard deviation in the mean is shown
by the shaded regions.
[4]
[5]
[6]
[7]
[8]
[9]
[10]
[11]
[12]
[13]
[14]
[15]
[16]
[17]
[18]
[19]
[20]
[21]
digital neuron model for neurosynaptic cores. In Neural Networks (IJCNN), The
2013 International Joint Conference on. IEEE, 1–10.
Aaron Clauset, M. E. J. Newman, and Cristopher Moore. 2004. Finding community
structure in very large networks. Phys. Rev. E 70 (Dec 2004), 066111. Issue 6.
https://doi.org/10.1103/PhysRevE.70.066111
Santo Fortunato. 2010. Community detection in graphs. Physics reports 486, 3
(2010), 75–174.
Santo Fortunato. 2017. Santo Fortunato’s Website: Software. https://sites.google.
com/site/santofortunato/inthepress2. (2017). Accessed: 2017-05-10.
Brian Gallagher, Hanghang Tong, Tina Eliassi-Rad, and Christos Faloutsos. 2008.
Using Ghost Edges for Classification in Sparsely Labeled Networks. In Proceedings
of the 14th ACM SIGKDD International Conference on Knowledge Discovery and
Data Mining (KDD ’08). ACM, New York, NY, USA, 256–264. https://doi.org/10.
1145/1401890.1401925
Wulfram Gerstner and Werner M Kistler. 2002. Spiking neuron models: Single
neurons, populations, plasticity. Cambridge university press.
Michelle Girvan and Mark EJ Newman. 2002. Community structure in social
and biological networks. Proceedings of the national academy of sciences 99, 12
(2002), 7821–7826.
Dan Goodman and Romain Brette. 2008. Brian: A Simulator for Spiking Neural
Networks in Python. Frontiers in Neuroinformatics 2 (2008).
Kathleen E. Hamilton, Neena Imam, and Travis S. Humble. 2017. Community
Identification with spiking neural networks. Poster presented at SIAM’s Network
Science Workshop 2017 July 13–14, Pittsburgh, PA.. (2017).
John Hertz, Anders Krogh, and Richard G Palmer. 1991. Introduction to the theory
of neural computation. Santa Fe Institute studies in the sciences of complexity,
Vol. 1. Addison-Wesley.
John J Hopfield. 1982. Neural networks and physical systems with emergent
collective computational abilities. Proceedings of the national academy of sciences
79, 8 (1982), 2554–2558.
John J Hopfield. 1984. Neurons with graded response have collective computational properties like those of two-state neurons. Proceedings of the national
academy of sciences 81, 10 (1984), 3088–3092.
John J Hopfield and David W Tank. 1985. Neural computation of decisions in
optimization problems. Biological cybernetics 52, 3 (1985), 141–152.
Mark D Humphries. 2011. Spike-train communities: finding groups of similar
spike trains. Journal of Neuroscience 31, 6 (2011), 2321–2336.
Takashi Kanamaru and Yoichi Okabe. 2000. Associative memory retrieval induced
by fluctuations in a pulsed neural network. Physical Review E 62, 2 (2000), 2629.
Takashi Kanamaru and Yoichi Okabe. 2001. Fluctuation-induced memory retrieval in a pulsed neural network storing sparse patterns with hierarchical
correlations. Physical Review E 64, 3 (2001), 031904.
Andrea Lancichinetti and Santo Fortunato. 2009. Benchmarks for testing community detection algorithms on directed and weighted graphs with overlapping
communities. Physical Review E 80, 1 (2009), 016118.
Andrea Lancichinetti, Santo Fortunato, and Filippo Radicchi. 2008. Benchmark
graphs for testing community detection algorithms. Physical review E 78, 4 (2008),
046110.
S Lowel and W Singer. 1992.
Selection of intrinsic horizontal connections in the visual cortex by correlated neuronal activity.
Science 255, 5041 (1992), 209–212. https://doi.org/10.1126/science.1372754
arXiv:http://science.sciencemag.org/content/255/5041/209.full.pdf
[22] Wolfgang Maass and Christopher M Bishop. 2001. Pulsed neural networks. MIT
press.
[23] Wolfgang Maass and Thomas Natschläger. 1997. Networks of spiking neurons
can emulate arbitrary Hopfield nets in temporal coding. Network: Computation
in Neural Systems 8, 4 (1997), 355–371.
[24] Fragkiskos D Malliaros and Michalis Vazirgiannis. 2013. Clustering and community detection in directed networks: A survey. Physics Reports 533, 4 (2013),
95–142.
[25] Paul A Merolla, John V Arthur, Rodrigo Alvarez-Icaza, Andrew S Cassidy, Jun
Sawada, Filipp Akopyan, Bryan L Jackson, Nabil Imam, Chen Guo, Yutaka
Nakamura, et al. 2014. A million spiking-neuron integrated circuit with a scalable
communication network and interface. Science 345, 6197 (2014), 668–673.
[26] Mark EJ Newman and Michelle Girvan. 2004. Finding and evaluating community
structure in networks. Physical review E 69, 2 (2004), 026113.
[27] M. E. J. Newman. 2010. Networks: An Introduction. Oxford University Press.
[28] L. Peel. 2016. Graph-based semi-supervised learning for relational networks.
ArXiv e-prints (Dec. 2016). arXiv:1612.05001
[29] Leto Peel, Daniel B Larremore, and Aaron Clauset. 2017. The ground truth about
metadata and community detection in networks. Science Advances 3, 5 (2017),
e1602548.
[30] Usha Nandini Raghavan, Réka Albert, and Soundar Kumara. 2007. Near linear
time algorithm to detect community structures in large-scale networks. Physical
Review E 76, 3 (2007), 036106.
[31] Jörg Reichardt and Stefan Bornholdt. 2004. Detecting fuzzy community structures
in complex networks with a Potts model. Physical Review Letters 93, 21 (2004),
218701.
[32] Jörg Reichardt and Stefan Bornholdt. 2006. Statistical mechanics of community
detection. Phys. Rev. E 74 (Jul 2006), 016110. Issue 1. https://doi.org/10.1103/
PhysRevE.74.016110
[33] Michael T. Schaub, Jean-Charles Delvenne, Martin Rosvall, and Renaud Lambiotte.
2017. The many facets of community detection in complex networks. Applied
Network Science 2, 1 (2017), 4. https://doi.org/10.1007/s41109-017-0023-6
[34] Hideki Tanaka, Takashi Morie, and Kazuyuki Aihara. 2005. Associative memory
operation in a Hopfield-type spiking neural network with modulation of resting
membrane potential. In Int. Symp. on Nonlinear Theory and its Applications, Bruges,
Belgium.
[35] Gergely Tibély and János Kertész. 2008. On the equivalence of the label propagation method of community detection and a Potts model approach. Physica A:
Statistical Mechanics and its Applications 387, 19 (2008), 4982–4984.
[36] David E Van Den Bout and Thomas K Miller. 1990. Graph partitioning using
annealed neural networks. IEEE Transactions on neural networks 1, 2 (1990),
192–203.
| 9cs.NE
|
Generalized SURE for optimal shrinkage of singular values in
low-rank matrix denoising
arXiv:1605.07412v3 [math.ST] 2 Oct 2017
Jérémie Bigot, Charles Deledalle & Delphine Féral
Institut de Mathématiques de Bordeaux et CNRS (UMR 5251)
Université de Bordeaux
October 3, 2017
Abstract
We consider the problem of estimating a low-rank signal matrix from noisy measurements
under the assumption that the distribution of the data matrix belongs to an exponential family. In this setting, we derive generalized Stein’s unbiased risk estimation (SURE) formulas
that hold for any spectral estimators which shrink or threshold the singular values of the
data matrix. This leads to new data-driven spectral estimators, whose optimality is discussed using tools from random matrix theory and through numerical experiments. Under
the spiked population model and in the asymptotic setting where the dimensions of the data
matrix are let going to infinity, some theoretical properties of our approach are compared
to recent results on asymptotically optimal shrinking rules for Gaussian noise. It also leads
to new procedures for singular values shrinkage in finite-dimensional matrix denoising for
Gamma-distributed and Poisson-distributed measurements.
Keywords: matrix denoising, singular value decomposition, low-rank model, Gaussian spiked
population model, spectral estimator, Stein’s unbiased risk estimate, random matrix theory,
exponential family, optimal shrinkage rule, degrees of freedom.
AMS classifications: 62H12, 62H25.
1
1.1
Introduction
Low rank matrix denoising in an exponential family
In various applications, it is of interest to estimate a signal matrix from noisy data. Typical
examples include the case of data that are produced in a matrix form, while others are concerned
with observations from multiple samples that can be organized in a matrix form. In such setting,
a typical inference problem involves the estimation of an unknown (non-random) signal matrix
X ∈ Rn×m from a noisy data matrix Y satisfying the model:
Y = X + W,
(1.1)
where W is an n × m noise matrix with real entries W ij assumed to be independent random
variables with E[W ij ] = 0 and Var(W ij ) = τij2 for 1 ≤ i ≤ n and 1 ≤ j ≤ m. In this paper, we
focus on the situation where the signal matrix X is assumed to have a low rank structure, and
we consider the general setting where the distribution of Y belongs to a continuous exponential
family parametrized by the entries of the matrix X = E[Y ]. For discrete observations (count
data), we also consider the specific case of Poisson noise.
The low rank assumption on X is often met in practice when there exists a significant
correlation between the columns of X. This can be the case when the columns of X represent
2D images at different wavelength of hyperspectral data, since images at nearby wavelengths are
strongly correlated [CSLT13]. Further applications, where low-rank modeling of X is relevant,
can be found in genomics [WDB01, ABB00], NMR spectroscopy [NPDL11], collaborative filtering
[CR09] or medical imaging [BD06, LBH+ 12], among many others.
Low-rank matrix estimation is classically done in the setting where the additive noise is
Gaussian with homoscedastic variance. The more general case of observations sampled from
an exponential family is less developed, but there exists an increasing research interest in the
study of low rank matrix recovery beyond the Gaussian case. Examples of low-rank matrix
recovering from Poisson distributed observations can be found in applications with count data
such as network traffic analysis [BMG13] or call center data [SH05]. A theory for low-rank matrix
recovery and completion in the case of Poisson observations has also been recently proposed in
[CX16]. Matrix completion under a low rank assumption with additive errors having a subexponential distribution and belonging to an exponential family has also been considered in
[Laf15]. The recent work [UHZB16] proposes a novel framework to approximate, by a low rank
matrix, a tabular data set made of numerical, Boolean, categorical or ordinal observations.
1.2
The class of spectral estimators
A standard approach to estimate a low rank matrix relies on the singular value decomposition
(SVD) of the data matrix
min(n,m)
Y =
X
σ̃k ũk ṽ tk ,
(1.2)
k=1
where σ̃1 ≥ σ̃2 ≥ . . . ≥ σ̃min(n,m) ≥ 0 denote its singular values, and ũk , ṽ k denote the associated
f
singular vectors. In this paper, we propose to consider the class of spectral estimators X̂ =
f (Y ), where f : Rn×m → Rn×m is a (possibly data-dependent) mapping that acts on the singular
values of the data matrix Y while leaving its singular vectors unchanged. More precisely, these
estimators take the form
f
X̂ = f (Y ) =
min(n,m)
X
fk (Y )ũk ṽ tk ,
(1.3)
k=1
where, for each 1 ≤ k ≤ min(n, m), fk (Y ) are real positive values that may depend only on σ̃k
(hence we write fk (σ̃k )) or on the whole matrix Y .
2
1.3
Investigated spectral estimators
Typical examples of spectral estimators include the classical principal component analysis (PCA)
applied to matrix denoising defined, for some 1 ≤ r ≤ min(n, m), as
r
X̂ =
r
X
σ̂k ũk ṽ tk
with σ̂k = fk (σ̃k ) = σ̃k
(1.4)
k=1
for all 1 ≤ k ≤ r and where it is implicitely understood that fk (σ̃k ) = 0 for k ≥ r + 1.
Another typical spectral estimator in matrix denoising with Gaussian measurements is the softthresholding [CSLT13] which corresponds to the choice
min(m,n)
X
X̂ soft =
σ̂k ũk ṽ tk
k=1
λ(Y )
σ̃k ,
with σ̂k = fk (Y ) = 1 −
σ̃k
+
(1.5)
for all 1 ≤ k ≤ min(n, m) and where λ(Y ) > 0 is a possibly data-dependent threshold parameter,
and (x)+ = max(x, 0) for any x ∈ R. Finaly, we will consider a more general class of shrinkage
estimators, encompassing the PCA and the soft-thresholding, that perform
min(m,n)
X̂ w =
X
σ̂k ũk ṽ tk
with σ̂k = fk (Y ) = wk (Y )σ̃k ,
(1.6)
k=1
where wk (Y ) ∈ [0, 1] is a possibly data-dependent shrinking weight.
1.4
Main contributions
Under the assumption that the distribution of Y belongs to an exponential family, the goal of this
paper is to derive data-driven choices for the weights wk (Y ) in (1.3). We construct estimators
via a two-step procedure. First, an active set of non-zero singular values is defined. Then, in a
second step, weights wk (Y ) associated with non-zero singular values are optimized, and shown
to reach desired asymptotical properties in the Gaussian spiked population model. The main
contributions of the paper are then the following ones.
1.4.1
An AIC inspired criterion for rank and singular values locations estimation
When no a priori is available on the rank of the signal matrix X, optimizing for the weights wk ,
for all 1 ≤ k ≤ min(m, n), can lead to estimators with large variance (i.e., overfitting the noise).
We propose an automatic rule to prelocalize the subset of non-zero singular values. An active
set s? ⊆ I = {1, 2, . . . , min(n, m)} of singular values is defined as the minimizer of a penalized
log-likelihood criterion that is inspired by the Akaike information criterion (AIC)
s
s∗ ∈ arg min − 2 log q(Y ; X̃ ) + 2|s|pn,m
with pn,m =
s⊆I
√ 2
1 √
m+ n ,
2
(1.7)
P
s
s
where X̃ = k∈s σ̃k ũk ṽ tk , |s| is the cardinal of s, and q(Y ; X̃ ) is the likelihood of the data in a
s
given exponential family with estimated parameter X̃ . For the case of Gaussian measurements
3
s
s
with homoscedastic variance τ 2 , one has that q(Y ; X̃ ) = kY − X̃ k2F /2τ 2 , where k · kF denotes
the Frobenius norm of a matrix, and we show that the active set of singular values boils down to
s? = {k ; σ̃k > cn,m
+ },
(1.8)
√
√
where cn,m
= τ ( m + n). For Gamma and Poisson measurements, we resort to a greedy
+
optimization procedure described in Section 4.
Once the active set has been determined, the subsequent shrinkage estimator is obtained by
optimizing only for the weights within this subset while setting the other ones to zero.
1.4.2
Novel data-driven shrinkage rules minimizing SURE-like formulas
We use the principle of Stein’s unbiased risk estimation (SURE) [Ste81] to derive unbiased
estimation formulas for the mean squared error (MSE) risk and mean Kullback-Leibler (MKL)
risks of spectral estimators. Minimizing such SURE-like formulas over an appropriate class of
spectral estimators is shown to lead to novel data-driven shrinkage rules of the singular values
of the matrix Y . In particular, our approach leads to novel spectral estimators in situations
where the variances τij2 of the entries W ij of the noise matrix are not necessarily equal, and may
depend on the signal matrix X.
As an illustrative example, let us consider spectral estimators of the form
1
X̂ w = f (Y ) = w1 (Y )σ̃1 ũ1 ṽ t1 ,
(1.9)
which only act on the first singular value σ̃1 of the data while setting all the other ones to zero.
In this paper, examples of data-driven choices for the weight w1 (Y ) are the following ones:
• for Gaussian measurements with n ≤ m and known homoscedastic variance τ 2
!!
n
X
τ2
σ̃12
w1 (Y ) = 1 − 2 1 + |m − n| + 2
11{σ̃1 >cn,m } ,
+
σ̃1
σ̃12 − σ̃`2
`=2
(1.10)
+
• for Gamma measurements with τij2 = X 2ij /L and L > 2 (see Section 2.1 for a precise definition),
−1
1
min(n,m)
n X
m
2
X
X
X̂ ij
L
−
1
1
σ̃
1
1 + |m − n| + 2
11{1∈s∗ } ,
w1 (Y ) = min 1,
+
Lmn
Y ij
Lmn
σ̃12 − σ̃`2
i=1 j=1
`=2
(1.11)
• for Poisson measurements with τij2 = X ij (see Section 2.1 for a precise definition)
Pn Pm
w1 (Y ) = min 1, P
j=1 Y ij
i=1
11{1∈s∗ } .
1
n Pm
X̂
ij
i=1
j=1
4
(1.12)
Beyond the case of rank one, closed-form solutions for the weights cannot be obtained, except
for the case of Gaussian measurements with homoscedastic variance τ 2 . In this latter case, the
rule for w1 (Y ) in (1.10) generalizes to other eigenvalues wk (Y ) as
min(n,m)
2
2
X
σ̃k
τ
11{σ̃ >cn,m } .
(1.13)
wk (Y ) = 1 − 2 1 + |m − n| + 2
2
k
+
σ̃k
σ̃k − σ̃`2
`=1;`6=k
+
For Gamma or Poisson distributed measurements, we propose fast algorithms to get numerical
approximations of the weights wk (Y ) (see Section 5.2 for more details).
1.4.3
Asymptotic properties in the Gaussian spiked population model
Another contribution of the paper is to discuss the optimality of the shrinking weights (1.13)
for Gaussian noise in the asymptotic setting where the dimensions of the matrix Y are let going
to infinity. These theoretical results are obtained for the so-called spiked population model that
has been introduced in the literature on random matrix theory and high-dimensional covariance
matrix estimation (see e.g. [BS06, BN12, DS07, SN13]). All the theoretical and asymptotic
results of the paper (other than derivation of proposed estimators) assume this model.
Definition 1.1. The Gaussian spiked population model corresponds to the following setting:
• the W ij in (1.1) are iid Gaussian random variables with zero mean and variance τ 2 = 1/m,
• the X ij ’s in (1.1) are the entries of an unknown n×m matrix X that has a low rank structure,
Pr∗
t
meaning that it admits the SVD X =
k=1 σk uk v k , where uk and v k are the left and
right singular vectors associated to the singular value σk > 0, for each 1 ≤ k ≤ r∗ , with
σ1 > σ2 > . . . > σr∗ ,
• the rank r∗ of the matrix X is assumed to be fixed,
• the dimensions of the data matrix Y = X + W are let going to infinity in the asymptotic
n
framework where the sequence m = mn ≥ n is such that limn→+∞ m
= c with 0 < c ≤ 1.
In the Gaussian spiked population model, the asymptotic locations of the empirical singular
values σ̃1 ≥ . . . ≥ σ̃min(n,m) are well understood in the random matrix theory (further details are
given in Section 3.1). Note that the setting where the rank r∗ is not held fixed but allowed to
grow with min(n, m) is very different, see e.g. [LW12] and references therein.
Under the Gaussian spiked population model, our contributions are then as follows:
• we prove the convergence of the SURE formula when the dimensions of Y tend to infinity,
• it is shown that minimizing the asymptotic value of SURE leads to the same estimator as the
limiting value of the estimator obtained by minimizing the SURE,
• this model allows to show that the novel data-driven spectral estimators derived in this paper
are asymptotically connected to existing optimal shrinkage rules [SN13, GD14a, Nad14] for
low-rank matrix denoising,
• in this setting, we are also able to connect the choice of the penalty function 2|s|pn,m in (1.7)
with Stein’s notion of degrees of freedom (see e.g. [Efr04]) for spectral estimators.
5
1.4.4
Numerical experiments and publicly available source code
As the theoretical properties of our estimators are studied in an asymptotic setting, we report the
results of various numerical experiments to analyze the performances of the proposed estimators
for finite-dimensional matrices. These experiments allow the comparison with existing shrinkage
rules for Gaussian-distributed measurements and they are also used to shed some lights on the
finite sample properties of the method for Gamma-distributed or Poisson-distributed measurements. We also exhibit the settings where the signal matrix X is either easy or more difficult to
recover. From these experiments, the main findings are the following ones:
• the use of an appropriate active set s of singular values is an essential step for the quality
of shrinkage estimators whose weights are data-driven by SURE-like estimators; taking
s = {1, . . . , min(n, m)} leads to poor results while the choice of s = s∗ minimizing the AIC
criterion (1.7) appears to yield the best performances,
• for Gaussian noise, the performances of our approach are similar to those obtained by the
asymptotically optimal spectral estimator proposed in [GD14a] when the true rank r∗ of
the signal matrix X is sufficiently small, but for large to moderate values of the signal-tonoise ratio our approach may perform better than existing methods in the literature,
• for Gamma or Poisson distributed measurements, the spectral estimators proposed in this
paper give better results than estimators based on PCA (restricted to the active set s∗ ) or
soft-thresholding of singular values.
Beyond the case of Gaussian noise, the implementation of the estimators is not straightforward, and we thus provide publicly available source code at
https://www.math.u-bordeaux.fr/~cdeledal/gsure_low_rank
to reproduce the figures and the numerical experiments of this paper.
1.5
Related results in the literature
Early work on singular value thresholding began with the work in [EY36] on the best approximation of fixed rank to the data matrix Y . Spectral estimators with different amounts of shrinkage
for each singular value of the data matrix have then been proposed in [EM72, EM76]. In the case
of Gaussian measurements with homoscedastic variance, the problem of estimating X under a
low-rank assumption has recently received a lot of attention in the literature on high-dimensional
statistics, see e.g. [CSLT13, DG14, JS15, SN13]. Recent works [GD14a, Nad14] also consider the
more general setting where the distribution of the additive noise matrix W is orthogonally invariant, and such that its entries are iid random variables with zero mean and finite fourth moment.
In all these papers, the authors have focused on spectral estimators which shrink or threshold the
singular values of Y , while its singular vectors are left unchanged. In this setting, the main issue
is to derive optimal shrinkage rules that depends on the class of spectral estimators that is considered, on the loss function used to measure the risk of an estimator of X, and on appropriate
assumptions for the distribution of the additive noise matrix W .
6
1.6
Organization of the paper
Section 2 is devoted to the analysis of a data matrix whose entries are distributed according to
a continuous exponential family. SURE-like formula are first given for the mean squared error
risk, and then for the Kullback-Leibler risk. As an example of discrete exponential family, we
also derive such risk estimators for Poisson distributed measurements. The computation of datadriven shrinkage rules is then discussed for Gaussian, Gamma and Poisson noises. In Section 3,
we restrict our attention to the Gaussian spiked population model in order to derive asymptotic
properties of our approach. We study the asymptotic behavior of the SURE formula proposed
in [CSLT13, DG14] for spectral estimators using tools from RMT. This result allows to make a
connection between data-driven spectral estimators minimizing the SURE for Gaussian noise, and
the asymptotically optimal shrinkage rules proposed in [SN13, Nad14] and [GD14a]. In Section
4, we study the penalized log-likelihood criterion (1.7) used to select an active set of singular
values. Its connection to the degrees of freedom of spectral estimators and rank estimation in
matrix denoising is discussed. Various numerical experiments are finally proposed in Section 5
to illustrate the usefulness of the approach developed in this paper for low-rank denoising and
to compare its performances with existing methods. The proofs of the main results of the paper
are gathered in a technical Appendix A, and numerical implementation details are described in
Appendix B.
2
SURE-like formulas in exponential families
For an introduction to exponential families, we refer to [Bro86]. The idea of unbiased risk
estimation in exponential families dates back to [Hud78]. More recently, generalized SURE
formulas have been proposed for the estimation of the MSE risk, for denoising under various
continuous and discrete distributions in [RS07], and for inverse problems whithin the continuous
exponential families in [Eld09]. In [Del17], SURE-like formula are derived for the estimation of
the Kullback-Leibler risk that applies to both continuous and discrete exponential families. In
what follows, we borrow some ideas and results from these works. We first treat the case of
continuous exponential families, and then we focus on Poisson data in the discrete case.
2.1
Data sampled from a continuous exponential family
We recall that Y is an n×m matrix with independent and real entries Y ij . For each 1 ≤ i ≤ n and
1 ≤ j ≤ m, we assume that the random variable Y ij is sampled from a continuous exponential
family, in the sense that each Y ij admits a probability density function (pdf) q(y; X ij ) with
respect to the Lebesgue measure dy on the real line Y = R. The pdf q(y; X ij ) of Y ij can thus
be written in the general form:
q(y; X ij ) = h(y) exp (η(X ij )y − A(η(X ij ))) , y ∈ Y,
(2.1)
where η (the link function) is a one-to-one and smooth function, A (the log-partition function)
is a twice differentiable mapping, h is a known function, and X ij is an unknown parameter of
interest belonging to some open subset X of R. Throughout the paper, we will suppose that the
following assumption holds:
7
Assumption 2.1. The link function η and the log-partition function A are such that
A0 (η(x)) = x for all x ∈ X ,
where A0 denotes the first derivative of A.
Since E[Y ij ] = A0 (η(X ij )) for exponential families in the general form (2.1), Assumption 2.1
implies that E[Y ij ] = X ij , and thus the data matrix satisfies the relation Y = X + W where
W is a centered noise matrix, which is in agreement with model (1.1). Now, if we let Θ = η(X ),
it will be also convenient to consider the expression of the pdf of Y ij in the canonical form:
p(y; θ ij ) = h(y) exp (θ ij y − A(θ ij )) , y ∈ Y,
(2.2)
where θ ij = η(X ij ) ∈ Θ is usually called the canonical parameter of the exponential family.
Finally, we recall the relation Var(Y ij ) = A00 (θ ij ) = A00 (η(X ij )) where A00 denotes the second
derivative of A. Then, we denote by θ the n × m matrix whose entries are the θ ij ’s.
Examples of data satisfying model (2.1) are the following ones:
Gaussian noise with known variance τ 2 :
(y − X ij )2
1
q(y; X ij ) = √ exp −
, E[Y ij ] = X ij , Var(Y ij ) = τ 2 ,
2τ 2
2π
1
y2
x
θ2
√
Y = R, X = R, Θ = R, h(y) =
exp − 2 , η(x) = 2 , A(θ) = τ 2 .
2τ
τ
2
2πτ
Gamma-distributed measurements with known shape parameter L > 0:
X 2ij
LL y L−1
y
q(y; X ij ) =
exp −L
,
11]0,+∞[ (y), E[Y ij ] = X ij , Var(Y ij ) =
X ij
L
Γ(L)X L
ij
LL y L−1
L
θ
Y = R, X =]0, +∞[, Θ =]−∞, 0[, h(y) =
11
(y), η(x) = − , A(θ) = −L log −
.
Γ(L) ]0,+∞[
x
L
f
f
The matrix θ = η(X) can then be estimated via the n × m matrix θ̂ = θ̂ (Y ) whose entries
are given by
f
f
θ̂ ij (Y ) = η X̂ ij , for all 1 ≤ i ≤ n, 1 ≤ j ≤ m,
(2.3)
f
where X̂ ij is a spectral estimator as defined in eq. (1.3).
In the rest of this section, we follow the arguments in [Eld09] and [Del17] to derive SURE-like
f
f
formulas under the exponential family for the estimators θ̂ and X̂ , using either the meansquared error (MSE) risk or the Kullback-Leibler (KL) risk.
8
2.1.1
Unbiased estimation of the MSE risk
We consider the following MSE risk which provides a measure of discrepancy in the space Θ of
natural parameters, and then indirectly in the space of interest X .
f
f
f
Definition 2.1. The squared error (SE) risk of θ̂ is SE(θ̂ h, θ) = kθ̂ i− θk2Fh, and the meani
f
f
f
f
squared error (MSE) risk of θ̂ is defined as MSE(θ̂ , θ) = E SE(θ̂ , θ) = E kθ̂ − θk2F .
f
Using the above MSE risk to compare θ̂ and θ implies that the discrepancy between the
f
f
estimator X̂ and the matrix of interest X is measured by the quantity MSEη (X̂ , X) =
f
f
f
MSE(η(
), η(X))
h X̂
i which is different from MSE(X̂ , X). For Gaussian noise, MSEη (X̂ , X) =
f
1
2
E kX̂ − XkF , while for Gamma distributed measurements with known shape parameter
τ2
L > 0, it follows that
2
f
n X
m
X
X
−
X̂
f
ij
ij
.
MSEη (X̂ , X) = L2
f
X ij X̂ ij
i=1 j=1
The following proposition gives a SURE formula for the MSE risk introduced in Definition 2.1.
Proposition 2.1. Suppose that the data are sampled from a continuous exponential family.
Assume that the function h, in the definition (2.2) of the exponential family, is twice continuously
differentiable on Y = R. If the following condition holds
h f
i
E θ̂ ij (Y ) < +∞, for all 1 ≤ i ≤ n, 1 ≤ j ≤ m,
(2.4)
then, the quantity
n X
m
X
h00 (Y ij )
h0 (Y ij ) f
f
θ̂ ij (Y ) +
+ 2 div θ̂ (Y ),
GSURE(θ̂ ) = kθ̂ (Y )k +
2
h(Y ij )
h(Y ij )
f
f
2
(2.5)
i=1 j=1
f
where div θ̂ (Y ) =
f
n X
m
X
∂ θ̂ ij (Y )
∂Y ij
i=1 j=1
f
, is an unbiased estimator of MSE(θ̂ , θ)
f
f
f
Note that GSURE(θ̂ ) is an unbiased estimator of MSE(θ̂ , θ) and not of MSE(X̂ , X). It
is shown in Section 3.3 that the results of Proposition 2.4 coincide with the approach in [CSLT13]
on the derivation of a SURE formula in the case of Gaussian noise for smooth spectral estimators.
In the case of Gamma noise, assuming L > 2 implies that the conditions on the function h in
Proposition 2.1 is satisfied, hence assuming that conditions (2.4) holds as well, and using that
f
θ̂ ij (Y
L
)=−
fij (Y )
f
and
it follows that
f
GSURE(θ̂ ) =
n X
m
X
i=1 j=1
∂ θ̂ ij (Y )
∂fij (Y )
L
=
,
2
∂Y ij
|fij (Y )| ∂Y ij
∂fij (Y ) (L − 1)(L − 2)
L2
2L(L − 1)
2L
−
+
−
.
|fij (Y )|2 Y ij fij (Y ) |fij (Y )|2 ∂Y ij
|Y ij |2
(2.6)
9
2.1.2
Unbiased estimation of KL risks
Following the terminology in [Del17], let us now introduce two different notions of KullbackLeibler risk, which arise from the non-symmetry of this discrepancy measure.
Definition 2.2. Let f : Rn×m → Rn×m be a smooth spectral function. Consider the estimator
f
θ̂ defined by (2.3), where Y is a matrix whose entries Y ij are independent random variables
sampled from the exponential family (2.2) in canonical form:
f
• the Kullback-Leibler synthesis (KLS) risk of θ̂ is defined as
f
n X
m Z
X
p(y; θ̂ ij )
f
p(y; θ̂ fij ) dy,
KLS(θ̂ , θ) =
log
p(y; θ ij )
R
i=1 j=1
h
i
f
f
f
and the mean KLS risk of θ̂ is defined as MKLS(θ̂ , θ) = E KLS(θ̂ , θ) ,
f
• the Kullback-Leibler analysis (KLA) risk of θ̂ is defined as
n X
m Z
X
p(y;
θ
)
f
ij
p(y; θ ij ) dy,
KLA(θ̂ , θ) =
log
f
R
p(y;
θ̂
)
i=1 j=1
ij
h
i
f
f
f
and the mean KLA risk of θ̂ is defined as MKLA(θ̂ , θ) = E KLA(θ̂ , θ) .
A key advantage of the Kullback-Leibler risk is that it measures the discrepancy between the
f
unknown distribution p(y; θ ij ) and its estimate p(y; θ̂ ij ). It is thus invariant with respect to the
f
f
f
reparametrization θ̂ = η(X̂ ) (unlike the MSE risk), and we may also write MKLS(θ̂ , θ) =
f
f
f
MKLS(X̂ , X) and MKLA(θ̂ , θ) = MKLA(X̂ , X). As suggested in [Del17], the MKLA risk
f
represents how well the distribution p(y; θ̂ ij ) explain a random variable Y ij sampled from the
pdf p(y; θ ij ). The MKLA risk is a natural loss function in many statistical problems since it
takes as a reference measure the true distribution of the data, see e.g. [Hal87]. The MKLS risk
represents how well one may generate an independent copy of Y ij by sampling a random variable
f
from the pdf p(y; θ̂ ij ). The MKLS risk has also been considered in various inference problems in
statistics [HL06, Yan94].
By simple calculation, it follows that
f
MKLS(θ̂ , θ) =
f
and MKLA(θ̂ , θ) =
n X
m
X
i=1 j=1
n X
m
X
i=1 j=1
E
i
h
i
h f
f
f
θ̂ ij − θ ij A0 (θ̂ ij ) + A(θ ij ) − E A(θ̂ ij ) ,
(2.7)
E
h
i
h
i
f
f
θ ij − θ̂ ij A0 (θ ij ) + E A(θ̂ ij ) − A(θ ij ).
(2.8)
10
Hence, in the case of Gaussian measurements with known
τi2 , we easily retrieve that
h variance
f
f
f
f
2
MKLS(θ̂ , θ) = MKLA(θ̂ , θ) = τ2 MSE(θ̂ , θ) = 2τ12 E kX̂ − Xk2F . In the case of Gamma
distributed measurements with known shape parameter L > 0, it follows that
f
f
n X
m
X
X̂
X̂
f
ij
ij
− log
− 1 ,
MKLS(θ̂ , θ) = L
E
X ij
X ij
i=1 j=1
n X
m
X
X
X
f
ij
ij
MKLA(θ̂ , θ) = L
E f − log f − 1 .
X̂ ij
X̂ ij
i=1 j=1
Below, we use some of the results in [Del17] whose main contributions are the derivation of
new unbiased estimators of the MKLS and MKLA risks. For continuous exponential family, the
risk estimate derived in [Del17] is unbiased for the MKLS risk, while it is only asymptotically
unbiased for the MKLA risk with respect to the signal-to-noise ratio. For data sampled from a
continuous exponential family, this makes simpler the use of the MKLS risk to derive data-driven
shinkage in low rank matrix denoising, and we have therefore chosen to concentrate our study
on this risk in this setting. The following proposition establishes a SURE formula to estimate
the MKLS risk in the continuous case.
Proposition 2.2. Suppose that the data are sampled from a continuous exponential family.
Assume that the function h, in the definition (2.2) of the exponential family, is continuously
differentiable on Y = R. Suppose that the function A, in the definition (2.2) of the exponential
family, is twice continuously differentiable on Θ. If the following condition holds
h
i
f
E A0 (θ̂ ij (Y )) < +∞, for all 1 ≤ i ≤ n, 1 ≤ j ≤ m,
(2.9)
then, the quantity
f
SUKLS(θ̂ ) =
n X
m
X
f
θ̂ ij (Y
i=1 j=1
h0 (Y ij )
)+
h(Y ij )
A
0
f
(θ̂ ij (Y
)) −
f
A(θ̂ ij (Y
)) + div f (Y ),
(2.10)
where div f (Y ) =
m X
n
X
∂fij (Y )
∂Y ij
i=1 j=1
f
, is an unbiased estimator of MKLS(θ̂ , θ) −
n X
m
X
A(θ ij ).
i=1 j=1
A key difference in the formula of unbiased estimates for the MSE and the KL risks is the
Pmin(n,m)
f
computation of the divergence term in (2.5) and (2.10), when X̂ = k=1
fk (σ̃k )ũk ṽ tk is
a smooth spectral estimator in the sense where each function fk : R+ → R+ is assumed to be
(almost everywhere) differentiable for 1 ≤ k ≤ min(n, m). In this setting, the divergence term in
f
f
f
the expression of GSURE(θ̂ ) depends upon the matrix θ̂ (Y ) = η(X̂ ). Therefore, when η is a
f
nonlinear mapping, it is generally not possible to obtain a simpler expression for div θ̂ (Y ). To
f
the contrary, for SUKLS(θ̂ ), the divergence term is div f (Y ) which has the following closed-form
11
expression for any smooth spectral estimators
min(n,m)
div f (Y ) = |m − n|
X
k=1
fk (σ̃k )
+
σ̃k
min(n,m)
X
min(n,m)
fk0 (σ̃k )
k=1
+2
X
k=1
min(n,m)
fk (σ̃k )
X
σ̃ 2
`=1;`6=k k
σ̃k
,
− σ̃`2
(2.11)
thanks to the results from Theorem IV.3 in [CSLT13].
r
r
2
Note that SUKLS(X̂ w ) = τ2 SURE(X̂ w ) for Gaussian measurements, hence, the GSURE
and SUKLS strategies match in this case. In the case of Gamma measurements, assuming that
L > 2 implies that the conditions on the function h in Proposition 2.2 is satisfied, and by
assuming that condition (2.9) holds as well, it follows that
n X
m
X
fij (Y )
f
SUKLS(θ̂ ) =
(L − 1)
− L log (fij (Y )) − Lmn + div f (Y ),
Y ij
i=1 j=1
where the expression of div f (Y ) is given by (2.11).
Note that it is implicitly understood in the definition of div f (Y ) that each mapping fij :
Rn×m → R is differentiable. The differentiability of the spectral function f (and thus of its
components fij ) is a consequence of the assumption that the functions f1 , . . . , fmin(n,m) (acting on
the singular values) are supposed to be differentiable. For further details, on the differentiability
of f and the fij ’s, we refer to Section IV in [CSLT13]. From the arguments in [CSLT13], it
follows that formula (2.11) for the divergence of f is also valid under the assumption that each
function fk is differentiable on R+ except on a set of Lebesgue measure zero.
2.2
The case of Poisson data
For Poisson data, the key result to obtain unbiased estimate of a given risk is the following lemma
which dates back to the work in [Hud78].
Lemma 2.1. Let f : Zn×m → Rn×m be a measurable mapping. Let 1 ≤ i ≤ n and 1 ≤ j ≤ m,
and denote by fij : Zn×m → R a measurable function. Let Y ∈ Zn×m be a matrix whose entries
are independently sampled from a Poisson distribution on Z. Then,
n X
m
n X
m
X
X
E
X ij fij (Y ) = E
Y ij fij (Y − ei etj ) ,
i=1 j=1
i=1 j=1
where, for each 1 ≤ i ≤ n and 1 ≤ j ≤ m, fij (Y ) denotes the (i, j)-th entry of the matrix f (Y ),
and ei (resp. ej ) denotes the vector of Zn (resp. Zm ) with the i-th entry (resp. j-th entry) equals
to one and all others equal to zero.
Hudson’s lemma provides a way to estimate (in an unbiased way) the expectation of the
Frobenius inner product between the matrix X and the matrix f (Y ). To see the usefulness of
this result, one may consider the following mean-squared error
n X
m
X
2
2
f
f
f
f
MSE(X̂ , X) = E X̂ − X
= E X̂
−2
X ij X̂ ij (Y ) + kXk2F .
F
F
12
i=1 j=1
Therefore, by Lemma 2.1, one immediately obtains that
f
PURE(θ̂ ) = X̂
f 2
F
−2
n X
m
X
i=1 j=1
Y ij fij (Y − ei etj ),
(2.12)
f
is an unbiased estimate for the quantity MSE(X̂ , X) − kXk2F .
For Poisson data, one may also define the following KL risks
n X
m
X
X ij
f
f
f
MKLS(θ̂ , θ) =
E X ij − X̂ ij − X̂ ij log f ,
X̂ ij
i=1 j=1
f
n X
m
X
X̂
f
f
ij
MKLA(θ̂ , θ) =
,
E X̂ ij − X ij − X ij log
X ij
(2.13)
(2.14)
i=1 j=1
which are in agreement with Definition 2.2 of KL risks for data sampled from a Poisson distribution. From the arguments in [Del17], there does not currently exist an approach to derive a
SURE formula for the MKLS risk in the Poisson case since they are no unbiased formula for
f
X̂ ij log X ij . Nevertheless, as shown in [Del17], Hudson’s Lemma 2.1 provides an unbiased estif
mator for X ij log X̂ ij , and then it is possible to unbiasedly estimate the MKLA risk as follows.
Proposition 2.3. For data sampled from a Poisson distribution, the quantity
f
PUKLA(θ̂ ) =
n X
m
X
i=1 j=1
f
X̂ ij − Y ij log fij (Y − ei etj ) ,
f
is an unbiased estimator of MKLA(θ̂ , θ) +
n X
m
X
i=1 j=1
2.3
(2.15)
X ij − X ij log (X ij ).
Data-driven shrinkage in low-rank matrix denoising
For a matrix X with entries X ij ∈ X = R, we consider shrinkage estimators of the form
X
s
X̂ w = f (Y ) =
wk σ̃k ũk ṽ tk ,
(2.16)
k∈s
with s ⊆ I = {1, 2, . . . , min(n, m)} and wk ∈ [0, 1], for all k ∈ s.
When the underlying matrix X is constrained to have positive entries, e.g. X =]0, +∞[ in
the Gamma and Poisson cases, we consider instead estimators of the form
"
#
X
s
X̂ w = f (Y ) = max
wk σ̃k ũk ṽ tk , ε ,
(2.17)
k∈s
where ε > 0 is an a priori lower bound on the smallest value of X ij , where for any matrix X,
max[X, ε]ij = max[X ij , ε], for all 1 ≤ i ≤ m and 1 ≤ j ≤ n.
13
The construction of the subset s is postponed to Section 4, and we focus here in selecting
the weights in a data-driven way for a fixed given s. In the following,
we denote by sc the
s
s
complementary set of s in I, i.e., sc = I \ s, and we let θ̂ w = η X̂ w . When X =]0, +∞[, we
have found that considering estimators of the form (2.17) is more appropriate
than trying to find
P
shrinking weights (wk )k∈s such that all the entries of the matrix k∈s wk σ̃k ũk ṽ tk are positive,
for a given subset s.
Gaussian noise with known homoscedastic variance τ 2
By applying the GSURE formula (2.5) for Gaussian distributed measurements and thanks to the
s
expression (2.11) for the divergence of smooth spectral estimators, we obtain for X̂ w , as defined
in (2.16), the SURE expression given by
min(n,m)
s
2
X
X
X
X
σ̃
s
k
1 + |m − n| + 2
wk
SURE(X̂ w ) = −mnτ 2 +
(wk − 1)2 σ̃k2 +
σ̃k2 + 2τ 2
2 − σ̃ 2
σ̃
`
k∈s
k∈sc
k=1
`=1;`6=k k
s
which unbiasedly estimate MSE(X̂ w , X). Hence, for each k ∈ s, by differentiating the above
expression with respect to wk , it follows that a data-driven weight for the k-th empirical singular
value is given by
min(n,m)
2
2
X
σ̃k
τ
wk (Y ) = 1 − 2 1 + |m − n| + 2
,
(2.18)
2
σ̃k
σ̃k − σ̃`2
`=1;`6=k
+
s
2
s
which fullfils the requirement that wk (Y ) ∈ [0, 1]. Note that as SUKLS(X̂ w ) = τ2 SURE(X̂ w )
for Gaussian measurements, the exact same data-driven weight would be obtained by minimizing
s
an estimate of the MKLS(X̂ w , X).
The case of estimators with rank one. Consider the case of estimators with rank 1, i.e., let
1
{1}
1
s = {1}. It follows that X̂ w = X̂ w = w1 X̂ where w1 ∈ [0, 1] is given by
min(n,m)
2
2
X
τ
σ̃1
w1 (Y ) = 1 − 2 1 + |m − n| + 2
.
σ̃1
σ̃12 − σ̃`2
`=1;`6=1
+
Gamma and Poisson distributed measurements
In Gamma and Poisson cases, it is not possible to follow the same strategy as in the Gaussian case
to derive optimal weights for (2.17) in a closed-form using the established SURE-like formulas.
We shall investigate how data-driven shinkage can be approximated in Section 5 on numerical
experiments using fast algorithms. Nevertheless, when the estimator is restricted to rank 1,
optimizing KL risk estimators lead to closed-form expressions under the assumption that all the
entries of the data matrix Y are strictly positive.
14
The case of estimators with rank one under Gamma noise. Consider again the case of esti1
mators with rank 1, i.e., let s = {1}, and let X̂ = σ̃1 ũ1 ṽ t1 denote the PCA approximation of
rank 1 of X. If all the entries of the matrix Y are strictly positive, by the Perron-Frobenius
theorem, all the entries of the first singular vectors ũ1 and ṽ 1 are strictly positive. Therefore, all
1
1
{1}
the entries of X̂ belong to the set X =]0, +∞[, and we can consider X̂ w = X̂ w = w1 σ̃1 ũ1 ṽ t1
as defined in (2.16) instead of (2.17). Assuming L > 2 for the SUKLS formula to hold, it follows
by simple calculations that
1
1 !
n X
m
X
X̂ ij
X̂ ij
1
SUKLS(θ̂ w ) =
(L − 1)w1
− mnL log (w1 ) − L log
− Lmn
Y ij
Y ij
i=1 j=1
min(n,m)
+ (1 + |m − n|)w1 + 2w1
X
`=2
σ̃12
.
σ̃12 − σ̃`2
Hence, by differentiating the above expression with respect to w1 and as it is monotonic on both
1
sides of its unique minimum, the optimal value of w1 ∈ [0, 1] minimizing SUKLS(θ̂ w ) is given by
−1
1
min(n,m)
n X
m
2
X
X
X̂
σ̃1
L−1
1
ij
,
1 + |m − n| + 2
w1 (Y ) = min 1,
+
2 − σ̃ 2
Lmn
Y ij
Lmn
σ̃
1
`
i=1 j=1
`=2
which yields the shrinking rule (1.11) stated in the introduction of this paper. Note that it is
not possible to obtain, in a closed-form, the optimal value of the weight w1 that minimizes the
1
criterion GSURE(θ̂ w ).
The case of estimators with rank one under Poisson noise. Using again that all the assumption that the entries of Y are positive, we can consider (by the Perron-Frobenius theorem)
1
{1}
X̂ w = X̂ w = w1 σ̃1 ũ1 ṽ t1 as defined in (2.16) instead of (2.17).
Then, the PURE formula (2.12)
1
1
and Proposition 2.3 apply to the estimator θ̂ w = log X̂ w which yield to
1
PURE(θ̂ w ) = w12 σ̃12 − 2
1
and PUKLA(θ̂ w ) =
n X
m
X
i=1 j=1
1
(ij)
where X̂ ij = σ̃1 ũ1,i ṽ 1,j , σ̃1
(ij)
n X
m
X
(ij) (ij) (ij)
Y ij w1 σ̃1 ũ1,i ṽ 1,j ,
i=1 j=1
1
(ij) (ij) (ij)
w1 X̂ ij − Y ij log (w1 ) + log σ̃1 ũ1,i ṽ 1,j
,
(ij)
is the largest singular value of the matrix Y − ei etj , and ũ1
(resp. ṽ 1 ) denotes its left (resp. right) singular vectors. Therefore, by differentiating the above
expression with respect to w1 and as it is monotonic on both sides of its unique minimum, an
1
optimal value for w1 ∈ [0, 1] which minimizes PURE(θ̂ w ) is given by
n X
m
X
1
(ij) (ij) (ij)
w1 (Y ) = min 1, 2
Y ij σ̃1 ũ1,i ṽ 1,j .
σ̃1 i=1 j=1
15
However, this optimal shrinking rule cannot be used in practice since evaluating the values of
(ij)
(ij) (ij)
σ̃1 , ũ1 , ṽ 1 for all 1 ≤ i ≤ n and 1 ≤ j ≤ m is not feasible from a computational point of view
for large values of n and m. Nevertheless, a fast algorithm to find a numerical approximation of
the optimal value w1 (Y ) is proposed in Section 5.
1
To the contrary, using again that all the X̂ ij are positive by the Perron-Frobenius theorem,
1
the value of w1 ∈ [0, 1] minimizing PUKLA(θ̂ w ) is
P P
m
n
Y
ij
j=1
i=1
,
w1 (Y ) = min 1, P P
1
n
m
X̂
ij
i=1
j=1
which is straightforward to compute. This corresponds to the shrinkage rule (1.12) given in the
introduction.
3
Gaussian spiked population model
In this section, we restrict our analysis to the Gaussian spiked population model and the asymptotic setting introduced in Definition 1.1.
3.1
Asymptotic location of empirical singular values
We summarize below the asymptotic behavior of the singular values of the data matrix Y =
Pmin(n,m)
σ̃k ũk ṽ tk in the Gaussian spiked population model.
k=1
In the case where X = 0, it is well known [AGZ10, BS10] that the empirical distribution
of the singular values of Y = W (with τ = √1m ) converges, as n → +∞, to the quarter circle
distribution if c = 1 and to its generalized version if c < 1. This distribution is supported on the
compact interval [c− , c+ ] with
√
c± = 1 ± c
where c+ is the so-called bulk (right) edge.
When X 6= 0 has a low rank structure, the asymptotic behavior of the singular values of
Y = X + W is also well understood [BN12, DS07, SN13], and generalizations to noise matrix
W whose distribution is orthogonally invariant have also been recently considered in [BN12].
Below, we recall some of these results that will be needed in this paper. To this end, let us
introduce the real-valued function ρ defined by
r
(1 + σ 2 )(c + σ 2 )
ρ (σ) =
for any σ > 0.
σ2
Then, the following result holds (see e.g. Theorem 2.8 in [BN12] and Proposition 9 in [SN13]).
Proposition 3.1. Assume that Y = X + W isPa random matrix sampled from the Gaussian
∗
spiked population model with τ = √1m and X = rk=1 σk uk v tk . Then, for any fixed k ≥ 1, one
16
has that, almost surely,
ρ (σk )
lim σ̃k =
n→+∞
c+
if k ≤ r∗ and σk > c1/4 ,
otherwise.
Moreover,
lim σ̃min(n,m) = c− .
n→+∞
In what follows, we shall also use the relation
p
ρ2 (σ) − (c + 1) − (ρ2 (σ) − (c + 1))2 − 4c
1
=
that holds for any σ > c1/4 ,
σ2
2c
which is a consequence of e.g. the results in Section 3.1 in [BN12].
3.2
(3.1)
Existing asymptotically optimal shrinkage rules
Below, we briefly summarize some results in [GD14a] and [Nad14] on the construction of asymptotically optimal spectral estimators. Let
min(n,m)
f
X
X̂ = f (Y ) =
fk (σ̃k )ũk ṽ tk
(3.2)
k=1
f
be a given smooth spectral estimator, and consider the standard squared error SE(X̂ , X) =
f
f
kX̂ − Xk2F as a measure of risk. The set of spectral functions minimizing SE(X̂ , X) is given
by fk (σ̃k ) = ũtk X ṽ k , for 1 ≤ k ≤ min(n, m). However, it cannot be used in practice since X is
obviously unknown. A first alternative suggested in [GD14a] and [Nad14] is to rather study the
asymptotic risk
f
f
SE∞ (X̂ ) = lim SE(X̂ , X) (in the almost sure sense)
(3.3)
n→∞
in the Gaussian spiked population model. Then, it is proposed in [GD14a] and [Nad14] to find
f
an asymptotically optimal choice of f by minimizing SE∞ (X̂ ) among a given class of smooth
spectral functions. The results in [GD14a] show that, among spectral estimators of the form
Pmin(n,m)
η
X̂ = k=1
η(σ̃k )ũk ṽ tk , where η : R+ → R+ is a continuous shrinker such that η(σ) = 0
whenever σ ≤ c+ , an asymptotically optimal shrinkage rule is given by the choice
( q
1
(σ 2 − (c + 1))2 − 4c if σ > c+ ,
σ
η ∗ (σ) =
(3.4)
0
otherwise.
P
δ
In [Nad14], it is proposed to consider spectral estimators of the form X̂ = rk=1 δk ũk ṽ tk where
δ1 , . . . , δr are positive weights. By Theorem 2.1 in [Nad14], it follows that, if σk > c1/4 for all
δ
1 ≤ k ≤ r with r ≤ r∗ , then the weights which minimize SE∞ (X̂ ) over Rr+ are given by
δk∗ = δk (σk ) =
σk
q
σk4 − c
(1 + σk2 )(c + σk2 )
, for all 1 ≤ k ≤ r.
17
(3.5)
In what follows, the shrinkage rules (3.4) and (3.5) are shown to be equivalent, and they
will serve as a reference of asymptotic optimality. It should be stressed that the estimators in
[GD14a] and [Nad14] are not equivalent. Indeed, the method in [Nad14] requires an estimate
of the rank, while the approach in [GD14a] applies the same shrinker to all empirical singular
values. Nevertheless, the shrinkage function that is applied to significant singular values (either
above the bulk edge in [GD14a] or up to a given rank in [Nad14]) is the same.
3.3
Asymptotic behavior of data-driven estimators based on SURE
Following the principle of SURE, a second alternative to choose a smooth spectral estimator
of the form (3.2) is to study the problem of selecting a set of functions (fk )1≤k≤min(n,m) that
h
i
f
f
minimize an unbiased estimate of MSE(X̂ , X) = E kX̂ − Xk2F . For any 1 ≤ i ≤ m and
f
1 ≤ j ≤ n, we recall that fij (Y ) denotes the (i, j)-th entry of the matrix X̂ = f (Y ). Under
the condition that
∂fij (Y )
E |Y ij fij (Y )| +
< +∞, for all 1 ≤ i ≤ n, 1 ≤ j ≤ m.
(3.6)
∂Y ij
it follows from the results in [CSLT13] (or equivalently from Proposition 2.1 for Gaussian noise
with τ 2 = 1/m) that
f
2
SURE X̂ = −n + kf (Y ) − Y k2F + div f (Y ),
m
(3.7)
f
is an unbiased estimate of MSE(X̂ , X), where the divergence div f (Y ) admits the closed-form
expression (2.11). The SURE formula (3.7) has been used in [CSLT13] to find a data-driven value
for λ = λ(Y ) in the the case of singular values shrinkage by soft-thresholding which corresponds
to the choice
fk (σ̃k ) = (σ̃k − λ)+ , for all 1 ≤ k ≤ min(n, m).
We study now the asymptotic behavior of the SURE formula (3.7). To this end, we shall
use Proposition 3.1, but we will also need the following result (whose proof can be found in the
Appendix) to study some of the terms in expression (2.11) of the divergence of f (Y ).
Proposition 3.2. Assume that Y = X + W isPa random matrix sampled from the Gaussian
∗
spiked population model with τ = √1m and X = rk=1 σk uk v tk . Then, for any fixed 1 ≤ k ≤ r∗
such that σk > c1/4 , one has that, almost surely,
n
1 X
σ̃k
1
1
lim
=
1+ 2 .
n→+∞ n
ρ (σk )
σ̃k2 − σ̃`2
σk
`=1;`6=k
In what follows, we restrict our analysis to the following class of spectral estimators (the
terminology in the definition below is borrowed from [GD14a]).
18
Pmin(n,m)
f
Definition 3.1. Let X̂ = f (Y ) = k=1
fk (σ̃k )ũk ṽ tk be a smooth spectral estimator. For
a given 1 ≤ r ≤ min(n, m), the estimator f is said to be a spectral shrinker of order r that
collapses the bulk to 0 if
fk (σ) = 0 whenever σ ≤ c+ and 1 ≤ k ≤ r,
fk (σ) = 0 for all σ ≥ 0 and k > r.
The reason for restricting the study to spectral estimators such that fk (σ̃k ) = 0 whenever
σ̃k < c+ is linked to the choice of the active set s∗ (1.8) of singular values in the Gaussian case,
f
as detailed in Section 4. Now, for a spectral shrinker X̂ of order r that collapses the bulk to 0,
we study the asymptotic behavior of the terms in expression (3.7) that only depend on f , namely
r
r
r
f
X
2 X 0
n X fk (σ̃k )
SURE X̂
=
+
fk (σ̃k )
(fk (σ̃k ) − σ̃k )2 + 2 1 −
m
σ̃k
m
k=1
k=1
k=1
r
n
X
1
σ̃
n X
k
fk (σ̃k )
+4
m
n
σ̃k2 − σ̃`2
(3.8)
`=1;`6=k
k=1
f
The reason for studying SURE X̂ is that finding an optimal shrinkage rule that minimizes
f
SURE X̂
is equivalent to minimizing expression (3.8) over spectral shrinkers of order r that
f
f
P
f
collapses the bulk to 0, since SURE X̂ − SURE X̂ = −n + nk=r+1 σ̃k2 for such X̂ .
Then, using Proposition 3.1, Proposition 3.2, and the assumption that the fk ’s are continuously differentiable functions on R+ , we immediately obtain the following result.
Lemma 3.1. Assume that Y = X + W is a random matrix sampled from the Gaussian spiked
P∗
f
population model with τ = √1m and X = rk=1 σk uk v tk . Let X̂ be a spectral shrinker of order
r ≤ r∗ that collapses the bulk to 0, such that each function fk , for 1 ≤ k ≤ r, is continuously
differentiable on ]c+ , +∞[. Moreover, assume that σk > c1/4 for all 1 ≤ k ≤ r. Then, one has
that, almost surely,
lim SURE X̂
n→+∞
f
2
r
X
σk (1 + c) + 2c
2
(fk (ρ(σk )) − ρ(σk )) + 2fk (ρ(σk ))
=
σk2 ρ(σk )
(3.9)
k=1
Asymptotically optimal shrinkage of singular values. Thanks to Lemma 3.1, one
may
f
determine an asymptotic optimal spectral shrinker as the one minimizing limn→+∞ SURE X̂ .
For this purpose, let us define the class of estimators
r
X̂ w =
r
X
wk σ̃k 11{σ̃k >c+ } ũk ṽ tk ,
(3.10)
k=1
where 1 ≤ r ≤ r∗ is a given integer, and the wk ’s are positive weights. In practice,
p n the estimator
r
X̂ w is computed by replacing the bulk edge c+ by its approximation cn,m
=
1+
+
m in eq. (3.10).
19
For moderate to large values of n and m, the quantities c+ and cn,m
are very close, and this
+
r
replacement does not change the numerical performances of X̂ w .
Then, provided that σk > c1/4 for all 1 ≤ k ≤ r, it follows from Lemma 3.1 that
2
r
r X
σk (1 + c) + 2c
2
2
.
lim SURE X̂ w =
ρ (σk )(wk − 1) + 2wk
n→+∞
σk2
k=1
Differentiating the above expression with respect to each weight wk leads to the following choice
of asymptotically optimal weights
wk∗ = 1 −
σk2 (1 + c) + 2c
for all 1 ≤ k ≤ r.
σk2 ρ2 (σk )
(3.11)
Therefore, if the singular values of the matrix X to be estimated are sufficiently large (namely
σk > c1/4 for all 1 ≤ k ≤ r), by using Proposition 3.1 and eq. (3.11), one has that an asymptotically optimal spectral shrinker of order r ≤ r∗ is given by the choice of functions
(
σk2 (1+c)+2c
ρ(σk ) if ρ(σk ) > c+ ,
1
−
2
σk ρ2 (σk )
for all 1 ≤ k ≤ r.
(3.12)
fk∗ (ρ(σk )) =
0
otherwise,
Using, the relation (3.1) one may also express the asymptotically optimal shrinking rule (3.12)
either as a function of ρ(σk ) only,
(
p
1
(ρ2 (σk ) − (c + 1))2 − 4c if ρ(σk ) > c+ ,
∗
ρ(σk )
fk (ρ(σk )) =
(3.13)
0
otherwise.
or as function of σk only (using that ρ(σk ) > c+ is equivalent to σk > c1/4 ),
(
4
√ σk −c
if σk > c1/4 ,
∗
2
2
σ
(1+σ
k
fk (ρ(σk )) =
k )(c+σk )
0
otherwise.
(3.14)
Therefore, for spectral shrinker of order r, we remark that the shrinkage rule (3.13) coincides
with the rule (3.4) which has been obtained in [GD14a]. Similarly, when the quantity fk∗ (ρ(σk ))
is expressed as a function of σk only in (3.14), then we retrieve the shrinking rule (3.5) derived
in [Nad14]. Therefore, itappears
that minimizing either the asymptotic behavior of the SURE,
f
that is limn→+∞ SURE X̂ , or the limit of SE risk (3.3) leads to the same choice of an
asymptotically optimal spectral estimator.
Data-driven shrinkage of empirical singular values. From the results in Section 2.3, the
principle of SURE minimisation leads to the following data-driven choice of spectral shrinker of
order r that collapses the bulk to 0
r
X̂ w =
r
X
fk (σ̃k )ũk ṽ tk ,
(3.15)
k=1
20
where fk (σ̃k ) = wk (Y )σ̃k 11{σ̃k >c+ } , for all 1 ≤ k ≤ r, with wk (Y ) given by (2.18). From Proposition 3.1 and Proposition 3.2 it follows that, if σk > c1/4 , then, almost surely,
σk2 (1 + c) + 2c
lim fk (σ̃k ) = 1 −
ρ(σk ), for all 1 ≤ k ≤ r ≤ r∗ .
n→+∞
σk2 ρ2 (σk )
r
Therefore, the data-driven spectral estimator X̂ w (3.15) asymptotically leads to the optimal
shrinking rule of singular values given by (3.12) which has been obtained by minimizing the
asymptotic behavior of the SURE.
√
√
√
Note that when τ 6= 1/ m, it suffices to replace the condition σ̃k > c+ by σ̃k > τ ( m + n)
r
in the definition of X̂ w , which yields the shrinking rule (1.13) stated in the introduction of this
paper.
4
Estimating active sets of singular values in exponential families
In this section, we propose to formulate a new Akaike information criterion (AIC) to select
an appropriate set of singular values over which aPshrinkage procedure might be applied. To
s
t
this end, we shall consider the estimator X̃ =
k∈s σ̃k ũk ṽ k defined for a subset s ⊆ I =
{1, 2, . . . , min(n, m)}, and we address the problem of selecting an optimal subset s? from the
data Y .
In the case of Gaussian measurements, the shrinkage estimators that we use in our numerical
P
f
experiments are of the form X̂ = k∈s? fk (σ̃k )ũk ṽ tk where
r
n
n,m
n,m
?
s = {k ; σ̃k > c+ } with c+ = 1 +
,
m
for some (possibly data-dependent) shrinkage functions fk . The set s? is based on the knowledge
of an approximation cn,m
of the bulk edge c+ . Thanks to Proposition 3.1, the bulk edge c+
+
is interpreted as the threshold which allows to distinguish the locations of significant singular
values in the data from those due to the presence of additive noise. Interestingly, the following
result shows that the active set s? may be interpreted through the prism of model selection using
the minimisation of a penalized log-likelihood criterion.
Proposition 4.1. Assume that Y = X + W where the entries of W are iid Gaussian variables
√
with zero mean and standard deviation τ = 1/ m. Then, we have
√ 2
1 √
s 2
∗
,
(4.1)
s = arg min mkY − X̃ kF + 2|s|pn,m with pn,m =
m+ n
2
s⊆I
s
t
k∈s σ̃k ũk ṽ k
P
for s ∈ I = {1, 2, . . . , min(n, m)}, and |s| is the cardinal of s.
P
s
t
Proof. We remark that Y − X̃ = k∈s
/ σ̃k ũk ṽ k . It results that
where X̃ =
mkY −
s
X̃ k2F
+ 2|s|pn,m = m
X
σ̃k2
+ 2|s|pn,m
k∈s
/
n
X
mσ̃k2 if k ∈
/s
=
.
2pn,m otherwise
k=1
21
(4.2)
p
n,m
?
Using that 2pn,m /m = cn,m
+ , it follows that the set s = {k ; σ̃k > c+ } is by definition such
s
that k ∈ s? if and only if 2pn,m < mσ̃k2 . Therefore, by (4.2), the criterion s 7→ mkY − X̃ k2F +
2|s|pn,m is minimum at s = s? which concludes the proof.
In the model Y = X + W , where the entries of W are iid Gaussian variables with zero mean
and variance τ 2 , it is well known that the degrees of freedom (DOF) of a given estimator X̂ is
defined as
n m
n m
1 XX
1 XX
DOF(X̂) = 2
Cov(X̂ ij , Y ij ) = 2
E[X̂ ij W ij ].
τ
τ
i=1 j=1
i=1 j=1
The DOF is widely used in statistics to define various criteria for model selection among a
collection of estimators, see e.g. [Efr04]. In low rank matrix denoising, the following proposition
shows that it is possible to derive the asymptotic behavior of the DOF of spectral estimators.
Proposition 4.2. Assume that Y = X + W is a random matrix sampled from the Gaussian
P∗
f
spiked population model with τ = √1m and X = rk=1 σk uk v tk . Let X̂ be a spectral shrinker of
order r ≤ r∗ that collapses the bulk to 0, such that each function fk , for 1 ≤ k ≤ r, is continuously
differentiable on ]c+ , +∞[. Moreover, assume that σk > c1/4 for all 1 ≤ k ≤ r. Then, one has
that, almost surely,
r
X
1
fk (ρ(σk ))
2c
f
lim
DOF(X̂ ) =
1+c+ 2 .
n→+∞ m
ρ(σk )
σk
k=1
Proof. Thanks to the derivation of the SURE in [Ste81] and formula (2.11) on the divergence of
spectral estimators, one has that
r
r
r
n
h
i
X
X
X
X
f
(σ̃
)
σ̃
f
f
k k
k
.
DOF(X̂ ) = E div X̂ = E |m − n|
+
fk0 (σ̃k ) + 2
fk (σ̃k )
σ̃k
σ̃k2 − σ̃`2
k=1
k=1
k=1
`=1;`6=k
By Proposition 3.1, Proposition 3.2, and our assumptions on the fk ’s, one has that, almost surely,
r
X
1
fk (ρ(σk ))
2c
f
lim
div X̂ =
1+c+ 2 .
n→+∞ m
ρ(σk )
σk
k=1
which completes the proof.
√
Hence, in the Gaussian spiked population model, by Proposition 4.2 and using that σk2 > c
for all 1 ≤ k ≤ r, it follows that if s ⊆ {1, . . . , r} then
√ 2
1
2c
s
lim
DOF(X̃ ) = |s| 1 + c + 2 ≤ |s| 1 + c = |s|c2+ .
(4.3)
n→+∞ m
σk
√
√ 2
s
1
Hence, the quantity 2|s| 2 ( m + n) is asymptotically an upper bound of DOF(X̃ ) (when
normalized by 1/m) for any given set s ⊆ {1, . . . , r}.
22
Let us now consider the more general case where the entries of Y are sampled from an
exponential family. To the best of our knowledge, extending the notion of the bulk edge to nonGaussian data sampled from an exponential family has not been considered so far in the literature
on random matrices and low rank perturbation model. Therefore, except in the Gaussian case,
it is far from being trivial to find an appropriate threshold value c̄ to define an active set in the
form s̄ = {k ; σ̃k > c̄}.
Nevertheless, to select an appropriate active set of singular values, we introduce the following
s
criterion that is inspired by the previous results on the DOF of the estimator X̃ in the Gaussian
case and the statistical literature on the well known AIC for model selection [Aka74].
P
s
Definition 4.1. The AIC associated to X̃ = k∈s σ̃k ũk ṽ tk is
s
s
AIC(X̃ ) = −2 log q(Y ; X̃ ) + 2|s|pn,m
with
pn,m =
√ 2
1 √
m+ n .
2
(4.4)
Q Q
s
s
where |s| is the cardinal of s, and q(Y ; X̃ ) = ni=1 m
j=1 q(Y ij ; X̃ ij ) is the likelihood of the data
s
in the general form (2.1) at the estimated parameters X ij = X̃ ij .
s
In the above definition of AIC(X̃ ), the quantity 2|s|pn,m is an approximation of the degree
s
of freedom of X̃ , i.e., of the numbers of its free parameters as it is justified by Proposition 4.2
in the case of Gaussian measurements. The AIC allows us to define an optimal subset of active
variables as
s
s∗ = arg min AIC(X̃ ).
s⊆I
For Gaussian measurements, Proposition 4.1 gives the value of the optimal set s∗ in a closed-form.
Following the arguments in Section 2.3, for Gamma or Poisson measurements and for a given
subset s, we consider the estimator
#
"
X
s
(4.5)
X̃ = max
σ̃k ũk ṽ tk , ε ,
k∈s
when > 0 is an a priori value to satisfy the positivity constraint on the entries of an estimator
in this setting. However, contrary to the case of Gaussian noise, the search of an optimal subset
s
s? ⊂ arg min AIC(X̃ ) becomes a combinatorial problem in this context. In our numerical
s⊆I
experiments, we thus choose to construct an approximation s̃ of s? with a greedy search strategy
that reads as follows
n
o
I\{k}
I
s̃ = I \ k ∈ I ; AIC(X̃
) ≤ AIC(X̃ ) .
(4.6)
For Gaussian measurements, s̃ = s? since the optimisation problem (4.6) becomes separable.
In our numerical experiments, we have found that s̃ selects a relevant set of active singular values
which separates well the structural content of X while removing most of the noise component.
Further details are given in Section 5 below.
23
For Gaussian noise, the computation of the active set s∗ of singular values may also be
interpreted as a way to estimate the unknown rank r∗ of the signal matrix X. In this setting,
one has that s? = {k ; σ̃k > cn,m
+ } which suggests the choice
r̂ = max{k ; σ̃k > cn,m
+ },
(4.7)
as an estimator of r∗ .
There exists an abundant literature of the problem of estimating the rank of an empirical
covariance matrix for the purpose of selecting the appropriate number of significant components
to be kept in PCA or factor analysis. It is much beyond the scope of this paper to give an
overview of this topic. We point to the review in [Jol02] for a summary of existing methods to
determine the number of components in PCA that are grouped into three categories: subjective
methods, distribution-based test tools, and computational procedures. For recent contributions
in the matrix denoising model (1.1) with Gaussian noise, we refer to the works [CTT14, GD14b]
and references therein. For example for Gaussian data with know variance τ 2 = 1/m, Eq. (11)
in [GD14b] on optimal hard thresholding of singular values suggest to take
s
8c
√
r̂ = max{k ; σ̃k > λ(c)}, with λ(c) = 2(c + 1) +
,
(4.8)
(c + 1) + c2 + 14c + 1
as a simple method to estimate the rank. It should be remarked that the problem of estimating
the true rank r∗ of X in model (1.1) is somewhat ill-posed as, in the Gaussian spiked population
model, Proposition 3.1 implies that one may only expect to estimate the so-called effective rank
reff = max{k ; σk > c1/4 } (see e.g. Section II.D in [Nad14]).
In our numerical experiments, we shall compare different choices for the active set of singular
values of the form ŝ = {1, . . . , r̂} where r̂ is either given by (4.7), (4.8), or by the “oracle choices”
r̂ = r∗ and r̂ = reff .
Other methods based on hypothesis testing [CTT14] could be used for rank estimation in the
Gaussian model (1.1), but it is beyond the purpose of this paper to give a detailed comparison.
For Poisson or Gamma noise, it is more difficult to interpret the computation of s∗ as a way
to estimate the rank of X since, in our numerical experiments, we have found that the cardinality
of s∗ is generally not equal to max{k ; k ∈ s∗ }. Moreover, to the best of our knowledge, there
is not so much work on the estimation of the true rank of a noisy matrix beyond the Gaussian
case. Therefore, we have not included a numerical comparison with other methods for the choice
of the active set of singular values in these two cases.
5
Numerical experiments
In this section, we assess of the performance of data-driven srhinkage rules under various numerical experiments involving Gaussian, Gamma and Poisson measurements.
24
5.1
The case of a signal matrix of rank one
We consider the simple setting where the rank r∗ of the matrix X is known and equal to one
meaning that
X = σ1 u1 v t1 ,
where u1 ∈ Rn and v 1 ∈ Rm are vectors with unit norm that are fixed in this numerical
experiment, and σ1 is a positive real that we will let varying. We also choose to fix n = m = 100,
n
= 1 and c+ = 2. For the purpose of sampling data from Gamma and Poisson
and so to take c = m
distribution, we took singular vectors u1 and v 1 with positive entries. The i-th entry (resp. j-th
entry) of u1 (resp. v 1 ) is chosen to be proportional to 1 − (i/n − 1/2)2 (resp. 1 − (j/m − 1/2)2 ).
Pmin(n,m)
Let Y = k=1
σ̃k ũk ṽ tk be an n × m matrix whose entries are sampled from model (2.1) and
then satisfying E[Y ] = X.
Gaussian measurements
We first consider the case of Gaussian measurements, where Y = X + W with E[W ij ] = 0,
Var(W ij ) = τ 2 with τ = √1m . In this context, we compare the following spectral shrinkage
estimators:
• Rank-1 PCA shrinkage
1
X̂ = σ̃1 ũ1 ṽ t1 ,
• Rank-1 SURE-driven soft-thresholding
1
X̂ soft = σ̂1 ũ1 ṽ t1
with σ̂1 = (σ̃1 − λ(Y ))+ ,
• Rank-1 asymptotically optimal shrinkage proposed in [Nad14] and [GD14a]
q
1
X̂ ∗ = σ̂1 ũ1 ṽ t1 with σ̂1 = σ̃12 − 4 11{σ̃1 >2} ,
• Rank-1 SURE-driven weighted estimator that we have derived in Section 2.3
!!
n
1
1
2 X σ̃12
1
t
X̂ w = σ̂1 ũ1 ṽ 1 with σ̂1 = 1 − 2
+
σ̃1 11{σ̃1 >2} ,
σ̃1 m m
σ̃12 − σ̃`2
`=2
+
where the above formula follows from the results in Section 3.3 using that c = 1 and c+ = 2
in these numerical experiments, and where, for the soft-thresholding, the value λ(Y ) > 0 is
obtained by a numerical solver in order to minimize the SURE. As a benchmark, we will also
consider the oracle estimator X 1∗ that performs shrinkage by using the knowledge of the true
singular-value σ1 defined as
p
X 1∗ = σ̂1 ũ1 ṽ t1 with σ̂1 = ρ(σ1 )2 − 4 11{ρ(σ1 )>2}
25
which corresponds to the asymptotically optimal shrinking rule (3.13) as a function of ρ(σk ) in
the setting c = 1 and c+ = 2. Note that form the formula above ŵ1 = σˆ1 /σ˜1 is necessary in the
range [0, 1] for all considered estimators.
In Figure 1, we compare the estimated singular-values σ̂1 and the estimated weights ŵ1 =
σ̂1 /σ̃1 as functions of σ1 for the four aforementionned estimators. Because all estimators are
subject to noise variance, we display, for all estimators, the median values and the 80% confidence
intervals obtained from M = 100 noise realizations. It can be seen that the median curves for
1
1
the eigenvelues and the weights of X̂ w and X̂ ∗ coincide (up to variations that are slightly larger
for the former) which is in agreement with the asymptotic analysis of shrinkage rules that has
been carried out in Section 3.3. Spectral estimator obtained by SURE-driven soft-thresholding
also leads to an optimal shrinkage rule.
In Figure 2, for each of the four spectral estimators above, we display for M = 100 noise
realizations, as functions of σ1 , the following normalized MSE
NMSE(X̂) =
kX̂ − Xk2F
.
kXk2F
1
1
1
The normalized MSE of the estimators X̂ soft , X̂ ∗ and X̂ w are the same for values of σ1 larger than
c1/4 = 1, and they only differ for values of σ1 close or below the threshold c1/4 = 1 (corresponding
to values of ρ(σ1 ) below the bulk edge c+ = 2). More remarkably, above c1/4 = 1, they offer
similar NMSE values to the oracle shrinkage estimator X 1∗ , not only in terms of median but also in
1
terms of variability, as assessed by the confidence intervals. The performances of the estimator X̂
(standard PCA) are clearly poorer. These numerical experiments also illustrate that, for finitedimensional low rank matrix denoising with r∗ = 1, data-driven spectral estimators obtained by
minimizing a SURE criterion achieve performances that are similar to asymptotically optimal
shrinkage rules.
Gamma and Poisson distributed measurements
Let us now consider the case where the entries of Y ij ≥ 0 of the data matrix Y are independently
sampled from a Gamma or Poisson distribution with mean X ij > 0. To satisfy the constraint
that the estimators must be matrices with positive entries, we consider estimators of the form
(2.17). In this context, we compare the following spectral shrinkage estimators, set for ε = 10−6 ,
as:
• Rank-1 PCA shrinkage
1
X̂ = max σ̃1 ũ1 ṽ t1 , ε ,
• Rank-1 GSURE/SUKLS/PURE/PUKLA-driven soft-thresholding
1
X̂ soft = max σ̂1 ũ1 ṽ t1 , ε
with σ̂1 = (σ̃1 − λ(Y ))+ ,
26
5
X̂1
X̂1w
X1∗
X̂1soft
X1∗
4
Estimated singular value
Estimated singular value
4
5
X̂1
3
2
0
1
2
3
4
3
2
3
2
1
0
5
X̂1∗
X1∗
4
1
1
1
2
3
4
0
5
1
(a)
(b)
0.8
0.8
X̂1
0.2
2
3
4
Singular value σ1 of X
(d)
0.6
0.4
X̂1
0.2
X̂1w
X1∗
1
Estimated weight
0.8
Estimated weight
1
0.4
5
0
2
5
3
4
0.6
0.4
X̂1
0.2
X̂1soft
X1∗
1
4
(c)
1
0.6
3
Singular value σ1 of X
1
0
2
Singular value σ1 of X
Singular value σ1 of X
Estimated weight
X̂1
Estimated singular value
5
5
0
Singular value σ1 of X
(e)
X̂1∗
X1∗
1
2
3
4
5
Singular value σ1 of X
(f)
Figure 1: The case of Gaussian measurements with m = n = 100. Estimated first
singular value σ̂1 as a function of the true underlying one σ1 , for (a) our proposed
1
1
1
estimator X̂ w , (b) the soft-thresholding X̂ soft and (c) the asymptotical one X̂ ∗ . All
of them are compared to the first singular value σ̃1 of Y 1 and the one of the oracle
asymptotical estimator X 1∗ . (c,d,e) Same but for the corresponding weight ŵ1 = σ̂1 /σ̃1 .
Curves have been computed on M = 100 noise realizations, only the median and an
80% confidence interval are represented respectively by a stroke and a shadded area of
the same color.
• Rank-1 GSURE/SUKLS/PURE/PUKLA-driven weighted estimator
1
X̂ w = max σ̂1 ũ1 ṽ t1 , ε
with σ̂1 = w1 (Y )σ̃1 11{1∈s̃} ,
where s̃ is the approximated active subset as defined in Section 4. For the soft-thresholding, the
value λ(Y ) > 0 is obtained by a numerical solver in order to minimize either the GSURE or the
SUKLS criterion (in the Gamma case) and either the PURE or the PUKLA criterion (in the
Poisson case). The weight w1 (Y ) ∈ [0, 1] is obtained by a numerical solver in order to minimize
the GSURE and the PURE, as described in Section B. According to Section 2.3, the weight
27
1.6
1.4
1.4
1.2
1.2
1.2
1
1
1
0.8
0.6
0.8
0.6
0.4
X̂1
0.2
X̂1w
X1∗
0
NMSE
1.6
1.4
NMSE
NMSE
1.6
0.5
1
1.5
2
2.5
3
0.8
0.6
0.4
X̂1
0.4
X̂1
0.2
X̂1soft
X1∗
0.2
X̂1∗
X1∗
0
0.5
Singular value σ1 of X
1
1.5
2
2.5
3
Singular value σ1 of X
(a)
0
0.5
1
1.5
2
2.5
3
Singular value σ1 of X
(b)
(c)
Figure 2: Same as Fig. 1 but for the normalized MSE of the corresponding estimators.
w1 (Y ) ∈ [0, 1], minimizing the SUKLS criterion, has the following closed-form formula
!−1
1
n X
m
n
2
X
X
X̂ ij
L
−
1
1
σ̃
1
,
w1 (Y ) = min 1,
+
1+2
Lmn
Y ij
Lmn
σ̃12 − σ̃`2
i=1 j=1
`=2
and for the PUKLA criterion, we have
P P
n
m
Y
ij
i=1
j=1
.
w1 (Y ) = min 1, P P
1
n
m
X̂
ij
i=1
j=1
To evaluate the performances of these estimators, we perform again a study involving M = 100
noise realizations.
In the Gamma case with shape parameter L = 3, results are reported in Figure 3 where σ1
ranges from 0.1 to 5. In the Poisson case, results are reported in Figure 4. To generate data
from a Poisson distribution with mean value X = σ1 u1 v t1 , we took σ1 ranging from 25 to 400.
In this context, the entries X i,j are in average ranging from 0.25 to 4. When σ1 = 25, about
78% of the entries of Y are 0 and 20% are equals to 1 which correspond to an extreme level of
noise, while when σ1 = 400, the entries of Y concentrate around 4 with a standard deviation of
2 which correspond to a simpler noisy setting.
In these experiments, it can be seen that all the data-dependent spectral estimators achieve
comparable results with really small errors in terms of MSE and MKL risks. Their performances
1
are similar to X̂ = σ̃1 ũ1 ṽ t1 meaning that optimizing either SURE-like criteria leads to a spectral
estimator closed to correspond to matrix denoising by ordinary PCA. However, unlike the Gamma
case, it might be observed in the Poisson case that when reaching a stronger noise level, i.e, for
small value of σ1 , the NMSE of all estimator increases as the denoising problem becomes more
1
challenging. Nevertheless, only the weight of X̂ w driven by PUKLA does not present a drop wich
allows reaching a slightly smaller MKLA. In the Gamma case, the noise level being proportional
to the signal level, the NMSE/MKLS remain constant for all σ1 .
28
2.0
1.0
1
2
3
0.8
0.1
1
4
2
3
0
4
Singular value σ1 of X
(a)
1
2
3
1.0
1
2
3
4
Singular value σ1 of X
(e)
1
2
4
(d)
X̂1
MKLS
0.2
0.8
3
Singular value σ1 of X
(c)
NMSE
Estimated weight
2.0
0.1
0
4
1
3.0
GSURE X̂1soft
Singular value σ1 of X
(b)
4.0
GSURE X̂1w
0.6
Singular value σ1 of X
Estimated singular value
MKLS
3.0
X̂1
0.2
NMSE
Estimated weight
Estimated singular value
1
4.0
0.1
SUKLS X̂1w
SUKLS X̂1soft
0.1
0.6
1
2
3
0
4
Singular value σ1 of X
(f)
1
2
3
4
Singular value σ1 of X
(g)
0
1
2
3
4
Singular value σ1 of X
(h)
Figure 3: The case of Gamma measurements with m = n = 100. (a) Estimated first
1
eigenvalue σ̂1 as a function of the true underlying one σ1 for our proposed estimator X̂ w
1
and the soft-thresholding X̂ soft when both are guided by the GSURE. Both of them
are compared to the first singular value σ̃1 of Y 1 . Same but for (b) the corresponding
weights ŵ1 = σ̂1 /σ̃1 , (c) the NMSE risk and (d) the MKLS risk. (e-h) Exact same
esperiments but when our proposed estimator and the soft-thresholding are both guided
by SUKLS. Curves have been computed on M = 100 noise realizations, only the median
and an 80% confidence interval are represented respectively by a stroke and a shadded
area of the same color.
1
Finally, as mentionned by [GD14a], to use the estimator X̂ ∗ in a Gaussian model with
√
√
1
1
homoscedastic variance τ 2 6= m
, one may take the estimator X̂ ∗ = mτ f1∗ (σ̃1 /( mτ ))ũ1 ṽ t1 .
Hence, provided the variance of the entries of the data matrix Y is known, it is always possible
1
to use a scaled version of the shrinkage rule from [GD14a] when τ 2 6= m
. However, in the setting
of Gamma or Poisson noise, the variance of the additive noise varies from one entry to another
and depends on the entries of the unknown signal matrix X to recover. For this reason, it is not
possible to use a scaled version of the shrinkage rule from [GD14a] as this would require to use
scaling factors depending on the unknown values of the entries of X. Therefore, a comparison
between our approach and the asymptotically optimal shrinkage proposed in [Nad14] and [GD14a]
(for Gaussian noise) is not possible in the case of Gamma or Poisson measurements. Note that
for Gamma measurements one has that Var(Y ij ) = X 2ij /L, and, in our numerical experiments,
it is assumed that the constant L is known.
A typical example where this assumption is reasonable, is the one of the statistical models of
29
150
0.8
150
250
50
350
150
250
0
350
Singular value σ1 of X
Singular value σ1 of X
(a)
50
150
250
0.8
150
250
350
Singular value σ1 of X
(e)
250
350
(d)
X̂1
0.2
0.1
PUKLA X̂1w
PUKLA X̂1soft
0.1
0.6
50
50
150
Singular value σ1 of X
MKLA
150
50
(c)
NMSE
250
0.1
0
350
1
350
PURE X̂1soft
Singular value σ1 of X
(b)
Estimated weight
Estimated singular value
0.1
PURE X̂1w
0.6
50
50
MKLA
250
X̂1
0.2
NMSE
Estimated weight
Estimated singular value
1
350
50
150
250
350
Singular value σ1 of X
(f)
0
50
100
150
200
Singular value σ1 of X
(g)
0
50
150
250
350
Singular value σ1 of X
(h)
Figure 4: The case of Poisson measurements with m = n = 100. (a) Estimated first
eigenvalue σ̂1 as a function of the true underlying one σ1 for our proposed estimator
1
1
X̂ w and the soft-thresholding X̂ soft when both are guided by the PURE. Both of them
are compared to the first singular value σ̃1 of Y 1 . Same but for (b) the corresponding
weights ŵ1 = σ̂1 /σ̃1 , (c) the NMSE risk and (d) the MKLA risk. (e-h) Exact same
esperiments but when our proposed estimator and the soft-thresholding are both guided
by PUKLA. Curves have been computed on M = 100 noise realizations, only the
median and an 80% confidence interval are represented respectively by a stroke and a
shadded area of the same color.
speckle used in coherent imagery, such as, Synthetic Aperture Radar (SAR) and SOund Navigation And Ranging (SONAR) imagery. In such imaging systems, the observed irradiance Y ij of a
pixel with indices (i, j) is obtained as the square modulus of a complex signal modeled as being
zero-mean circular complex Gaussian distributed (consequence of the Central Limit Theorem)
[Goo76]. It follows that Y ij has an exponential distribution1 with mean X ij corresponding to the
underlying irradiance to be estimated. In order to improve the contrast of such images (namely,
the signal to noise ratio), an average of L independent and identically distributed images is often
performed, and the resulting pixel value becomes Gamma distributed with parameter L [UD89].
Because the number L of images to be averaged is chosen by the practitioner, the parameter L is
absolutely known without uncertainties, and for this reason it does not require to be estimated.
Nevertheless the variance Var(Y ij ) = X 2ij /L remains unknown.
1
The exponential distribution is a particular instance of the Gamma distribution with parameter L = 1
30
While all estimators behave similarly in the rank 1 setting, we will see in the next section
that they can significantly differ when the rank is let to be larger than 2.
5.2
The case of a signal matrix of rank larger than two
We now consider the more complex an realistic setting where the rank r∗ of the matrix X is
unknown and potentially larger than two, i.e.,
∗
X=
r
X
σk uk v tk ,
k=1
where uk ∈ Rn and v k ∈ Rm are vectors with unit norm that are fixed in this numerical
experiment, and σk are positive real values also fixed in this experiment. We also choose to fix
n = 100 and m = 200, while the true rank is r∗ = 9 as shown by the red curve in Figure 5(i).
Pmin(n,m)
Again, let Y = k=1
σ̃k ũk ṽ tk be an n × m matrix whose entries are sampled from model
(2.1) and then satisfying E[Y ] = X.
Gaussian distributed measurements
We first consider the case of Gaussian measurements, where Y = X + W with E[W ij ] = 0,
Var(W ij ) = τ 2 with τ = 80. In the following numerical experiments, we study the behavior of
the spectral estimator:
• PCA shrinkage
r
X̂ =
r
X
σ̃k ũk ṽ tk 11{k ≤ r̂} ,
k=1
• SURE-driven soft-thresholding
min(m,n)
X
X̂ soft =
σ̂k ũk ṽ tk
k=1
with σ̂k = (σ̃k − λ(Y ))+ ,
• Asymptotically optimal shrinkage proposed in [Nad14] and [GD14a]
r
X̂ ∗ =
r
X
σ̂k ũk ṽ tk
k=1
with σ̂k =
1
σ̃k
q
σ̃k2 − (c + 1)
2
− 4c 11{k ≤ r̂} ,
• SURE-driven weighted estimator that we have derived in Section 2.3
!!
r
n
X
1
k
2 X σ̃k2
r
t
X̂ w =
σ̂k ũk ṽ k with σ̂k = 1 − 2
+
σ̃k 11{k ≤ r̂} ,
σ̃k m m
σ̃k2 − σ̃`2
k=1
`=2
31
+
•
•
(e)
(f)
X
Y
105
X̂∗
Xsoft
X̂soft
Xrw
3
10
•
◦
(c)
•
(d)
•
◦
(g)
(h)
10−2
X̂r
104
•
10−2
NMSE
Singular values
(b)
NMSE
(a)
X̂rw
102
10−3
10−3
101
100
101
100
102
101
102
100
Rank r (restricted to the bulk edge)
Indices
(i)
(j)
NMSE
10−2
NMSE
10−3
10−3
100
101
Rank r (restricted to the oracle/true rank)
(l)
102
102
(k)
10−2
NMSE
10−2
101
Rank r (no restriction)
100
10−3
101
102
Rank r (restricted to the effective rank [Nad14])
(m)
100
101
Rank r (restricted as in [GD14b])
(n)
Figure 5: (a) Zoom on a 100 × 200 noise-free matrix and (e) a single realization of
corrupted version by Gaussian noise (τ = 80). (b,c) Oracle soft-thresholding X soft and
rmax
data-driven soft-thresholding X̂ soft . (d) PCA full rank X̂
, i.e., rmax = min(n, m).
rmax
(f,g,h) Oracle full rank approximation X w , and data-driven full rank estimation
rmax
rmax
X̂ w
and X̂ ∗ . (i) Their corresponding singular values. (j) NMSE of the various
approximations as a function of the rank r. (k) Same but without knowledge the bulk
edge, namely c+ = 0. (l,m,n) Same when the active set of singular values is of the form
ŝ = {1, . . . , r̂} where r̂ is either given by r̂ = r∗ (oracle/true rank), r̂ = reff (effective
rank) or by (4.8). In all the figures, the solid curves correspond to oracle estimators
and the dashed curves correspond to data-driven estimators, obtained over M = 1, 000
noise realizatrions. The grey areas represent a 80% confidence interval.
32
102
where r ∈ [1, min(n, m)], and for the soft-thresholding, the value λ(Y ) > 0 is obtained
by a numerical solver in order to minimize the SURE. Otherwise specified, we consider
r̂ = max{k ; σ̃k > cn,m
+ }, i.e., an estimator of the rank using knowledge of the bulk edge
n,m
c+ ≈ c+ , hence, 11{k ≤ r̂} = 11{σ̃k >cn,m } . As discussed in Section 4, we compare, in these exper+
iments, the influence of rank estimation by analyzing the performances of the same estimators
when either r̂ = rmax = min(n, m) (i.e. without knowledge the bulk edge, namely c+ = 0),
r̂ = r∗ (oracle/true rank), r̂ = reff (effective rank [Nad14]) or by (4.8) (from hard-thresholding
of singular values in [GD14b]).
In order to assess the quality of SURE as an estimator of the MSE, we also compare the
aforementioned approach with their oracle counterparts given by
min(m,n)
X soft =
X
σ̂k ũk ṽ tk
k=1
X rw =
r
X
σ̂k ũk ṽ tk
with σ̂k = (σ̃k − λoracle (Y ))+ ,
and
with σ̂k = ṽ tk X ũk ,
k=1
where λoracle (Y ) minimizes the squared error SE (non-expected risk) over the sets and softthresholding approximations respectively. Note that X rw and X soft are ideal approximations of
X that cannot be used in practice but serve as benchmarks to evaluate the performances of the
r
r
r
data-driven estimators X̂ , X̂ soft , X̂ ∗ and X̂ w . In order to shed some light on the variance of
these estimators, and indirectly on the variance of the SURE, we perform this experiments over
M = 1000 independent realizations of Y .
The results are reported on Figure 5. For an estimator of the rank given either by r̂ =
∗
max{k ; σ̃k > cn,m
+ } (knowledge of the bulk edege), r̂ = r (oracle/true rank), r̂ = reff (effective
r
r
rank) or by (4.8), it can be observed that X̂ w , X̂ ∗ and X rw achieve comparable performances for
all r ∈ [1, min(m, n)] even though the two first do not rely on the unknown matrix X. Similarly
X̂ soft and X soft achieve also comparable performances showing again that the SURE accurately
r
r
estimates the MSE. In terms of error bands for the NMSE, X̂ w , X̂ ∗ and X rw outperform X̂ soft
r
and X soft provided that r is large enough. Moreover, the performance of X̂ w plateaus to its
optimum when the rank r becomes large. This allows us to choose r = min(n, m) when we do
not have a priori on the true or effective rank.
Interestingly, Fig. 5.(k) shows that when the above estimators are used without thepknowledge
n,m
n
of the bulk edge (i.e. by taking cn,m
+ = 0 in their computation instead of c+ = 1 +
m , which
r
corresponds to the choice r̂ = rmax = min(n, m)), the performance of X̂ w actually decreases
when the rank r becomes too large. Indeed, it is clear from Fig. 5.(k), that the the error band of
r
the NMSE of X̂ w becomes much larger as the rank r increases. This illustrates that the SURE
suffers from estimation variance in the case of over parametrization when r becomes too large,
and thus it cannot be used to estimate jointly a too large number of weights. Therefore, the
knowledge of an appropriate estimator r̂ of the rank (e.g. using the bulk edge) seems to provide
a relevant upper bound on the number of weights that can be jointly and robustly estimated
with the SURE.
33
•
X
Y
104
X̂r
SEη Xsoft
103
GSURE X̂soft
KLS Xsoft
102
SUKLS X̂soft
SEη Xrw
GSURE X̂rw
KLS Xrw
101
SUKLS X̂rw
101
(d)
(i)
10−2
◦
(e)
◦
(j)
10−2
10−3
10−3
100
102
•
•
◦
(h)
NMSE
105
101
102
100
Rank r (restricted to the active set)
Indices
(k)
101
102
Rank r (no restriction)
(l)
(m)
10−0.5
10−0.5
MKLS
100
•
(g)
◦
(c)
MKLS
Singular values
(f)
•
(b)
NMSE
•
(a)
10−1
10−1
100
101
Rank r (restricted to the active set)
(n)
102
100
101
Rank r (no restriction)
(o)
Figure 6: (a) A single realization of corrupted version by Gamma noise (L = 80)
with zoom on a 100 × 200 matrix. (b,c,d,e) Oracle soft-thresholding X soft and datadriven soft-thresholding X̂ soft respectively for SEη , GSURE, KLS and SUKLS. (f)
rmax
PCA X̂
with full rank approximation i.e. rmax = min(n, m). (g,h,i,j) Oracle full
rmax
rank approximation X rwmax , and data-driven full rank estimation X̂ w respectively for
SEη , GSURE, KLS and SUKLS. (k) Their corresponding singular values averaged over
M = 100 noise realizations. (l,m) NMSE averaged over M = 100 noise realizations as
a function of the rank r with and without using the active set. (n,o) Same but with
respect to MKLS.
34
102
•
•
105
X
Y
104
X̂r
SEη Xsoft
103
PURE X̂soft
KLA Xsoft
PUKLA X̂soft
SEη Xrw
PURE X̂rw
KLA Xrw
102
101
NMSE
(g)
◦
(c)
(d)
◦
(h)
(i)
10−4
•
◦
(e)
•
◦
(j)
10−4
PUKLA X̂rw
101
100
102
101
102
100
Rank r (restricted to the active set)
Indices
(k)
101
102
Rank r (no restriction)
(l)
(m)
0.8
0.8
MKLA
100
100
MKLA
Singular values
(f)
•
(b)
NMSE
•
(a)
0.2
100
0.2
101
Rank r (restricted to the active set)
(n)
102
100
101
Rank r (no restriction)
(o)
Figure 7: (a) A single realization of corrupted version by Poisson noise with zoom
on a 100 × 200 noise-free matrix (b,c,d,e) Oracle soft-thresholding X soft and datadriven soft-thresholding X̂ soft respectively for SE, PURE, KLA and PUKLA. (f) PCA
rmax
X̂
with full rank approximation i.e. rmax = min(n, m). (g,h,i,j) Oracle full rank
rmax
approximation X rwmax , and data-driven full rank estimation X̂ w respectively for SE,
PURE, KLA and PUKLA. (k) Their corresponding singular values averaged over 200
noise realizations. (l,m) NMSE averaged over 200 noise realizations as a function of the
rank r with and without using the active set. (n,o) Same but with respect to MKLA.
(Matrix entries are displayed in log-scale for better visual assessment.)
35
102
Gamma and Poisson measurements
Let us now consider the case where the entries of Y ij > 0 of the data matrix Y are independently
sampled from a Gamma or Poisson distribution with mean X ij > 0. We again consider estimators
of the form (2.17). In this context, we compare the following spectral shrinkage estimators, set
for ε = 10−6 , as:
• PCA shrinkage
r
X̂ =
r
X
max σ̃k ũk ṽ tk , ε
with σ̂k = σ̃k 11{k∈s̃} ,
k=1
• GSURE/SUKLS/PURE/SUKLA driven soft-thresholding
min(m,n)
X
X̂ soft =
max σ̂k ũk ṽ tk , ε
k=1
with σ̂k = (σ̃k − λ(Y ))+ ,
• GSURE/SUKLS/PURE/SUKLA driven weighted estimator
r
X̂ w =
r
X
max σ̂k ũk ṽ tk , ε
with σ̂k = wk (Y )σ̃k 11{k∈s̃} ,
k=1
where r ∈ [1, min(n, m)], and s̃ is the approximated active subset as defined in Section 4. For the
soft-thresholding, the value λ(Y ) > 0 is obtained by a numerical solver in order to minimize either
the GSURE or the SUKLS criterion (in the Gamma case) and either the PURE or the PUKLA
criterion (in the Poisson case). As shown in Section 2.3, in the case of Gamma (resp. Poisson)
measurements, the value of wk (Y ) for k ∈ s̃ which minimizes the GSURE (resp. PURE) or the
SUKLS (resp. PUKLA), cannot be obtained in closed form. As an alternative, we adopt a greedy
one-dimensional optimization strategy starting from the matrix σ̃1 ũ1 ṽ t1 and next updating the
weights w` sequentially by starting ` = 1 to ` = min(n, m), with the constraint that, for all ` ∈
/ s̃,
the weight w` is set to zero. To this end, we resort to one-dimensional optimization techniques in
the interval [0, 1] using Matlab’s command fminbnd. This strategy is used for GSURE, SUKLS,
PURE and PUKLA by evaluating them as described in Section B. As in the Gaussian setting,
we compare this spectral estimators with their oracle counterparts given by
min(m,n)
X soft =
X
max σ̂k ũk ṽ tk , ε
k=1
X rw =
r
X
max σ̂k ũk ṽ tk , ε
with σ̂k = (σ̃k − λoracle (Y ))+ ,
and
with σ̂k = wkoracle (Y )σ̃k 11{k∈s̃} .
k=1
where wkoracle (X) and λoracle (Y ) minimizes one of the objective SEη , KLS, SE or KLA (nonexpected risks) over the set of matrices sharing with Y the same r first left and right singular
vectors, and soft-thresholding approximations respectively. Note again that X rw and X soft are
36
ideal approximations of X that cannot be used in practice but serve as benchmarks to evaluate
r
the performances of the data-driven estimators X̂ w and X̂ soft .
The results for the Gamma noise are reported on Figure 6. As in the Gaussian setting, it
r
can be observed that X̂ w and X rw achieve comparable performances, as well as X̂ soft and X soft
showing that the GSURE (resp. SUKLS) accurately estimates the MSEη (resp. KLS). Visual
inspection of the restored matrices tends to show that the estimators driven by MSEη or GSURE
produce less relevant results compared to KLS or SUKLS, as confirmed by the curves of NMSE
and MKLS. Performance in terms of NMSE also illustrates that minimizers of SEη do not
r
coincides with those of SE. As in the Gaussian setting, X̂ w and X r outperform X̂ soft , X soft and
r
r
standard PCA X̂ provided that r is large enough. Moreover, the performance of X̂ w obtained
with KL objectives plateaus to its optimum when the rank r becomes large. Again, this allows
us to choose r = min(n, m) when we do not have a priori on the true rank r? .
The results for the Poisson noise are reported on Figure 7. The conclusions are similar to the
Gaussian and Gamma cases. Obviously, the NMSE is smaller for approximations that minimizes
SE (or PURE) than for those minimizing KLA (or PUKLA). However, visual inspection of the
obtained matrices tends to demonstrate that minimizing such objectives might be less relevant
r
than minimizing KL objectives. In this setting, the performance of X̂ w is on a par with the one
r
of X̂ soft based on PUKLA. In fact, for other choices of matrices X, X̂ w based on PUKLA might
r
improve, in terms of MKLS, much more on X̂ soft , and might improve not as much on X̂ w based
r
on PURE. Nevertheless, whatever X, we observed that X̂ w driven by PUKLA always reaches
r
at least as good performance in terms of MKLS as the best of X̂ w driven by SE and X̂ soft .
Fig. 6.(m), Fig. 6.(o), Fig. 7.(m) and Fig. 7.(o) show that when the above estimators are
r
used without the active set (i.e., by choosing s̃ = [1, min(n, m]), the performance of X̂ w actually
decreases when the rank r becomes too large. As in the Gaussian setting, this can be explained
by the fact that the GSURE, SUKLS, PURE and PUKLA suffer from estimation variance in the
case of over parametrization, hence, they cannot be used to estimate jointly a too large number
of weights. The active set s̃ (in the same manner as the bulk edge) seems to provide a relevant
selection of the weights that can be jointly and robustly estimated in a data driven way.
5.3
Signal matrix with equal singular values and increasing rank
We propose now to highlight
limitations of our approach in the situation where the
Prpotential
∗
∗
t
rank r of the matrix X = k=1 σk uk v k is let growing and all positive singular values σk of X
are equal, namely
∗
Y =
r
X
k=1
σk uk v tk + W
1/4
with σk = γcn,m
for all 1 ≤ k ≤ r∗ ,
(5.1)
n
where uk ∈ Rn and v k ∈ Rm are vectors with unit norm that are fixed, cn,m = m
and W
is centered random matrix whose entries are iid Gaussian variables with variance τ 2 = 1/m.
We again choose to fix n = 100 and m = 200, while the true rank is r∗ let growing from 1 to
min(n, m) in the following numerical experiments. The constant γ is chosen to be larger than
1. Hence, eq. (5.1) corresponds to the Gaussian spiked population model in the setting where
37
1/4
1/4
all positive singular values are equal and larger than the threshold cn,m . The choice σk = γcn,m
with γ > 1 is motivated by the results from Proposition 3.1.
For a given value of the true rank r∗ , we performed experiments involving M = 1000 realizations from model (5.1) to compare the NMSE of the estimators by oracle soft-thresholding
rmax
X soft , data-driven soft-thresholding X̂ soft , PCA full rank X̂
i.e. rmax = min(n, m), oracle
rmax
rmax
rmax
full rank approximation X w , and data-driven full rank estimation X̂ w and X̂ ∗ . All these
estimators have been introduced in Section 5.2.
In Figure 8, we report the results of numerical experiments by displaying errors bars of the
NMSE of these estimators as functions of the true rank r∗ . For low values of the true rank
rmax
rmax
(r∗ ≤ 20), the data-driven estimators X̂ w
(our approach) and X̂ ∗
(shrinkage rule from
[GD14a]) achieve the best performances that are similar in term of median value of the NMSE.
However, our approach has some limitations with respect to the performances of the estimator
from [GD14a] or data-driven soft-thresholding [CSLT13] in the setting where the signal matrix
has equal positive singular values and when its rank is increasing. Moreover, the error bands
of the NMSE for our approach becomes significantly larger than those of the other data-driven
estimators when the true rank r∗ increases. This illustrates that SURE minimization may lead
to estimators with a high variance in the case of over parametrization, that is, when there exists
a large number of significant and close singular values in the signal matrix.
5.4
Influence of the dimension and the signal-to-noise ratio
In Section 5.3, we used simulated data consisting of a signal matrix with equal positive singular
values and an increasing rank. In such a setting , it is likely that the empirical weights wk (Y ),
Pmin(n,m) σ̃k2
used in our approach, will have a high variance due to the term `=1;`6=k σ̃2 −σ̃
2 in their expresk
`
sion (1.13). However, the numerical experiments carried out in Section 5.3 correspond to a very
specific configuration of the signal matrix (with many equal singular values and a high rank)
which is not likely to be encountered with real data.
To conclude these numerical experiments, we finally analyze the influence of the dimension
of the data and the signal-to-noise ratio on the performances of our approach and the estimator
from [GD14a] in a more realistic setting (with Gaussian noise). These two estimators are the
ones giving the best results, and it is thus of interest to compare them with further experiments.
We use real and square signal matrices X ∈ Rn×n having a relatively fast decay of their
singular values, see Figure 9 and Figure 10. We choose to re-size them to let n varying from 20
to 250, and we define the root of the signal-to-noise ratio (RSNR) as
q P
n
1
2
n
i,j=1 (X ij − X̄)
1 X
n2
RSNR =
with X̄ = 2
X ij .
τ
n
i,j=1
For each value of n and RSNR (ranging from 5 to 10), we performed experiments involving
M = 400 realizations from model (5.1) to compare the NMSE of the estimators by data-driven full
rmax
rmax
rank estimation X̂ w (our approach) and X̂ ∗
(shrinkage rule from [GD14a]) with rmax = n.
In Figure 9 and Figure 10, we report the results of these numerical experiments by displaying
38
X̂
X̂
0.1
X̂rw
0.1
8 · 10−2
8 · 10−2
6 · 10−2
6 · 10−2
6 · 10−2
4 · 10−2
4 · 10−2
4 · 10−2
20
40
60
80
100
20
40
True rank
(a) γ = 6
80
20
0.15
X̂soft
Xrw
X̂rw
X̂
X̂rw
0.2
80
0.15
100
0.15
0.1
20
40
60
80
100
20
True rank
(d) γ = 4
r
Xrw
0.7
0.5
0.4
0.3
0.3
0.3
60
True rank
(g) γ = 2
80
100
20
40
60
True rank
(h) γ = 2
80
r
100
X̂soft
0.5
0.4
40
X̂
0.6
0.4
20
100
X̂∗
Xsoft
NMSE
NMSE
X̂rw
0.5
80
0.7
X̂rw
0.6
X̂soft
Xrw
60
(f) γ = 4
X̂∗
Xsoft
0.6
40
True rank
(e) γ = 4
X̂
r
X̂soft
True rank
0.7
100
X̂∗
Xsoft
0.2
0.1
60
80
Xrw
0.1
40
60
(c) γ = 6
NMSE
X̂∗
Xsoft
20
40
True rank
r
0.2
NMSE
100
(b) γ = 6
NMSE
NMSE
60
True rank
X̂
r
X̂∗
Xsoft
X̂soft
NMSE
X̂soft
Xrw
8 · 10−2
0.12
Xrw
X̂rw
X̂∗
Xsoft
0.1
NMSE
0.12
r
NMSE
0.12
20
40
60
80
True rank
(i) γ = 2
Figure 8: Comparison of NMSE as a function of the true rank r∗ in model (5.1) for
different values of γ for the estimator by roracle soft-thresholding X soft , data-driven
max
soft-thresholding X̂ soft , PCA full rank X̂
i.e. rmax = min(n, m), oracle full rank
rmax
rmax
rmax
approximation X w , and data-driven full rank estimation X̂ w
and X̂ ∗ . The
active set set of singular values is of the form ŝ = {1, . . . , r̂} where r̂ = max{k ; σ̃k >
n,m
cn,m
(a),
+ } is an estimator of the rank using knowledge of the bulk edge c+ ≈ c+
(d), (g) Median value of the NMSE of the various estimators over M = 1000 Gaussian
noise realizations in model (5.1) as a function of the true rank r∗ . (b), (c), (e), (f),
(h), (i) The grey areas represent error bands of the NMSE of data-driven and oracle
estimators.
39
100
10 4
10 2
10 0
10 0
(a) Signal matrix X
1.05
·10−3
2
10 1
10 3
(b) Singular values of X
·10−3
3.4
·10−3
X̂∗
X̂∗
X̂∗
X̂rw
1
10 2
X̂rw
1.9
X̂rw
3.2
0.9
1.8
NMSE
NMSE
NMSE
0.95
1.7
3
2.8
0.85
1.6
0.8
0.75
50
100
150
200
250
1.5
2.6
50
100
(c) RSNR = 10
150
Size n
Size n
(d) RSNR = 7
200
250
2.4
50
100
150
200
Size n
(e) RSNR = 5
Figure 9: Comparison of NMSE as a function of the dimension n in model (5.1) with
rmax
rmax
a square matrix X for different values of RSNR for the estimator X̂ w
and X̂ ∗
with rmax = n. The active set set of singular values is of the form ŝ = {1, . . . , r̂} where
r̂ = max{k ; σ̃k > cn,m
+ } is an estimator of the rank using knowledge of the bulk edge
c+ ≈ cn,m
.
(a)
Signal
matrix of size 250 × 250, (b) Decay of the singular values of
+
rmax
rmax
X in log-log scale, (c), (d), (e) Median value of the NMSE of X̂ w
and X̂ ∗
over
M = 400 Gaussian noise realizations in model (5.1) as a function of the dimension n.
The orange and grey areas represent error bands of the NMSE of these two estimators.
errors bars of the NMSE of these estimators as functions of the dimension n. It can be seen that
our approach dominates numerically the estimator from [GD14a] (for all values of n and RSNR)
in settings that are more likely to be encountered in practice than the simulated data used in
Section 5.3.
A
A.1
Proof of the main results
Proof of Proposition 3.2
Let us first introduce some notation and definitions to be used in the proof. For all 1 ≤ ` ≤ n,
let λ̃` be the eigenvalues of Y Y t namely λ̃` = σ̃`2 . For a fixed 1 ≤ k ≤ r∗ such that σk > c1/4 ,
40
250
10 4
10 2
10 0
10 0
(a) Signal matrix X
·10−3
5.5
1
2.2
2
1.8
100
150
200
Size n
250
0.8
4
0.7
3.5
0.6
3
0.5
2.5
X̂rw
0.9
4.5
2.4
50
·10−2
X̂∗
X̂rw
5
NMSE
NMSE
·10−3
X̂∗
X̂rw
2.6
1.6
10 3
(b) Singular values of X
X̂∗
2.8
10 2
NMSE
3
10 1
50
100
150
200
250
0.4
50
Size n
(c) RSNR = 10
(d) RSNR = 7
100
150
200
Size n
(e) RSNR = 5
Figure 10: Same as Fig. 9 with another signal matrix X.
let us introduce the complex-valued function gk defined by
gk (z) =
n
1
1 X
n
z − λ̃`
`=1;`6=k
for z ∈ C \ supp(µk ),
n
o
where supp(µk ) = λ̃` ; 1 ≤ ` ≤ n, ` 6= k is the support of the random measure µk =
1 Pn
`=1;`6=k δλ̃` on R+ , where δλ denotes the Dirac measure at λ. It is clear that
n
Z
gk (z) =
1
dµk (λ).
z−λ
The main difficulty in the proof is to show that, almost surely,
1
1
2
1+ 2 ,
lim gk (σ̃k ) = 2
n→+∞
ρ (σk )
σk
which is the purpose of what follows.
For a matrix A ∈ Rn×m (with n ≤ m), we denote its singular values by σ1 (A) ≥ σ2 (A) ≥
. . . ≥ σn (A) ≥ 0. Hence, one has that σ̃` = σ` (Y ) for all 1 ≤ ` ≤ n. Now, we recall that
Y = X + W where X is a fixed matrix of rank r∗ and W is a random matrix with iid entries
41
250
1
sampled from a Gaussian distribution with zero mean and variance m
. The first step in the proof
is to show that the random measure µk behaves asymptotically as the almost sure limit of the
empirical spectral measure µW W t of the Wishart matrix W W t . By definition, the eigenvalues
of W W t are λ` (W ) = σ`2 (W ) for all 1 ≤ ` ≤ n and µW W t is thus defined as
n
µW W t
1X
=
δλ` (W ) .
n
`=1
n
It is well know (see e.g. Theorem 3.6 in [BS10]) that, once m = mn ≥ n and limn→+∞ m
=c
with 0 < c ≤ 1, then, almost surely, the empirical spectral measure µW W t converges weakly to
the so-called Marchenko-Pastur
distribution µM P which is deterministic and has the following
q
1
(c2+ − λ)(λ − c2− ) 1I[c2− ,c2+ ] (λ). We recall that such a convergence can
density dµMdλP (λ) = 2πcλ
also be characterized through the so-called Cauchy or Stieltjes transform which is defined for
any probability measure µ on R as
Z
1
dµ(λ).
∀z ∈ C outside the support of µ, gµ (z) =
z−λ
By eq. (3.3.2) in [BS10], one obtains that, almost surely,
Z
1
t (λ) = gM P (z) for any z ∈ C \ R,
dµ
lim
n→∞
z − λ WW
where gM P is the Cauchy transform of µM P and
p
Z
z − (1 − c) − (z − (c + 1))2 − 4c
1
dµM P (λ) =
gM P (z) =
z−λ
2cz
(A.1)
for all z ∈ C \ [c2− , c2+ ].
Moreover, by Proposition 6 in [PL03], the convergence (A.1) is uniform over any compact subset
of C \ R.
Then, it follows from the so-called Weyl’s interlacing inequalities (see e.g. Theorem 3.1.2 in
[HJ91]) that for all 1 ≤ ` ≤ n
σ`+r∗ (W ) ≤ σ` (Y ) ≤ σ`−r∗ (W ),
(A.2)
with the convention that σk (W ) = −∞ if k > n and σk (W ) = +∞ if k ≤ 0. Thanks to
the results that have been recalled above on the asymptotic properties of µW W t , one may use
inequalities (A.2) to prove that, almost surely, the random measure µk converges weakly to
the Marchenko-Pastur distribution µM P . Under the assumptions of Proposition 3.2 and using
Proposition 3.1, it can be shown that there exists ηk > 0 such that, almost surely and for all
sufficiently large n
λ̃` ∈
/ Kk := [ρ2 (σk ) − ηk , ρ2 (σk ) + ηk ]
fornany 1 ≤ ` ≤ n with `o6= k. Now, recall that the support supp(µk ) of the random measure µk
is λ̃` ; 1 ≤ ` ≤ n, ` 6= k , and that supp(µMP ) = [c2− , c2+ ]. Hence, for all sufficiently large n, one
has that
supp(µk ) ∩ Kk = ∅
and
supp(µMP ) ∩ Kk = ∅.
42
Therefore, thanks to the weak convergence of µk to µM P and using Ascoli’s Theorem, one may
prove that
lim sup |gk (z) − gM P (z)| = 0 almost surely.
n→∞ z∈K
(A.3)
k
Thanks to our assumptions, one has that, almost surely, limn→+∞ σ̃k2 = ρ2 (σk ) by Proposition
3.1. Hence, almost surely and for all sufficiently large n, one has that σ̃k2 ∈ Kk and so
|gk (σ̃k2 ) − gM P (ρ2 (σk ))| ≤ sup |gk (z) − gM P (z)| + |gM P (σ̃k2 ) − gM P (ρ2 (σk ))|.
z∈Kk
Therefore, using the uniform convergence (A.3) of gk to gM P and the continuity of gM P at
z = ρ2 (σk ), one obtains that, almost surely,
p
2 (σ ) − 1 + c −
ρ
(ρ2 (σk ) − (c + 1))2 − 4c
1
k
lim gk (σ̃k2 ) = gM P (ρ2 (σk )) = 2
×
.
n→+∞
ρ (σk )
2c
P
1
Since gk (σ̃k2 ) = n1 n`=1;`6=k σ̃2 −σ̃
2 , using the above equation and relation (3.1), it follows immek
`
1
1
1
+
so that, almost surely,
diately that gM P (ρ2 (σk )) = ρ2 (σ
2
)
σ
k
k
n
1 X
σ̃k
1
1
2
2
lim
1+ 2 ,
= lim σ̃k gk (σ̃k ) = ρ (σk ) gM P (ρ (σk )) =
n→+∞ n
ρ (σk )
σ̃k2 − σ̃`2 n→+∞
σk
`=1;`6=k
which completes the proof.
A.2
A technical result to prove SURE-like formulas
We recall the key lemma needed to prove the SURE-like formulas in an exponential family in
the continuous case. Similar results have already been formulated in different papers in the
literature, see e.g. the review proposed in [Del17].
Lemma A.1. Let Y ∈ Rn×m be a random matrix whose entries Y ij are independently sampled
from the continuous exponential family (2.2) in canonical form (that is the distribution of Y ij is
absolutely continuous with respect to the Lebesgue measure dy on R). Suppose that the function
h is continuously differentiable on Y = R. Let 1 ≤ i ≤ n and 1 ≤ j ≤ m, and denote by
Fij : Rn×m → R a continuously differentiable function such that
E [|Fij (Y )|] < +∞.
(A.4)
Then, the following relation holds
0
∂Fij (Y )
h (Y ij )
Fij (Y ) +
.
E [θ ij Fij (Y )] = −E
h(Y ij )
∂Y ij
43
Proof. Using the expression (2.2) of the pdf of the random varibles Y ij , one has that
Z
n
Y
p(yk` ; θ k` ) dyk` .
Fij (Y )h(yij )θ ij exp (θ ij yij − A(θ ij )) dyij
E [θ ij Fij (Y )] =
Rn×m
1≤k≤n
1≤`≤m
(k,`)6=(i,j)
where Y = (yk` )1≤k≤n,1≤`≤m . Thanks to condition (A.4), it follows that
Z
n
Y
p(yk` ; θ k` ) dyk` < +∞.
Fij (Y )h(yij ) exp (θ ij yij − A(θ ij )) dyij
Rn×m
(A.5)
1≤k≤n
1≤`≤m
(k,`)6=(i,j)
∂ exp(θ y −A(θ ))
ij ij
ij
Therefore, given that θ ij exp (θ ij yij − A(θ ij )) =
, an integration by part and
∂yij
eq. (A.5) imply that
Z
n
Y
∂Fij (Y )h(yij )
exp (θ ij yij − A(θ ij )) dyij
p(yk` ; θ k` ) dyk` .
E [θ ij Fij (Y )] = −
∂yij
Rn×m
1≤k≤n
1≤`≤m
(k,`)6=(i,j)
∂F (Y )
∂Fij (Y )h(yij )
∂yij
ij
= h0 (yij )Fij (Y ) + ∂y
h(yij ), we finally obtain that
ij
0
h (Y ij )
∂Fij (Y )
E [θ ij Fij (Y )] = −E
Fij (Y ) +
,
h(Y ij )
∂Y ij
Now, since
which completes the proof.
A.3
Proof of Proposition 2.1
We remark that
n X
m h
i
X
f
f
MSE(θ̂ , θ) =
E |θ̂ ij (Y )|2 − 2θ ij θ̂ ij (Y ) + θ 2ij .
f
(A.6)
i=1 j=1
f
Using Lemma A.1 with Fij (Y ) = θ̂ ij (Y ) and condition (2.4), it follows that
f
0
h
i
∂ θ̂ ij (Y )
h (Y ij ) f
f
.
θ̂ (Y ) + E
E θ ij θ̂ ij (Y ) = E
h(Y ij ) ij
∂Y ij
(A.7)
Then, by definition (2.2) of the exponential family, we remark that
00
Z
h (Y ij )
E
=
h00 (yij ) exp (θ ij yij − A(θ ij )) dyij .
h(Y ij )
R
Hence, using an integration by parts twice, we arrive at
00
Z
h (Y ij )
2
E
= θ ij
h(yij ) exp (θ ij yij − A(θ ij )) dyij = θ 2ij .
h(Y ij )
R
To complete the proof, it suffices to insert equalities (A.7) and (A.8) into (A.6).
44
(A.8)
A.4
Proof of Proposition 2.2
Thanks to eq. (2.8), one has that
f
MKLS(θ̂ , θ) =
n X
m
X
i=1 j=1
h f
i
f
f
f
E θ̂ ij (Y )A0 (θ̂ ij (Y )) − θ ij A0 (θ̂ ij (Y )) − A(θ̂ ij (Y )) + A(θ ij ).
(A.9)
f
Using Lemma A.1 with Fij (Y ) = A0 (θ̂ ij (Y )) and condition (2.9), it follows that
f
0
h
i
∂
θ̂
(Y
)
h
(Y
)
f
f
f
ij
ij
A00 (θ̂ ij (Y )) .
A0 (θ̂ ij (Y )) − E
E θ ij A0 (θ̂ ij (Y )) = −E
h(Y ij )
∂Y ij
(A.10)
Thus, inserting equality (A.10) into (A.9) implies that
f
X
n X
m
n X
m
X
∂ θ̂ ij (Y )
h0 (Y ij )
f
f
0 f
00 f
A (θ̂ ij (Y )) − A(θ̂ ij (Y )) +
SUKLS(θ̂ ) =
A (θ̂ ij (Y ))
θ̂ ij (Y ) +
h(Y ij )
∂Y ij
f
i=1 j=1
i=1 j=1
P P
f
A(θ ij ). Now recall that fij (Y ) =
is an unbiased estimator of MKLS(θ̂ , θ) − ni=1 m
j=1
f
f
f
∂f (Y )
η −1 θ̂ ij (Y ) and that A0 (θ̂ ij (Y )) = η −1 θ̂ ij (Y ) by Assumption 2.1. Therefore, ∂ijY
=
ij
f
f
∂ θ̂ (Y )
A00 (θ̂ ij (Y )) ∂ijY , and thus
ij
X
n X
m
n X
m
X
h0 (Y ij )
∂fij (Y )
f
f
f
0 f
SUKLS(θ̂ ) =
A (θ̂ ij (Y )) − A(θ̂ ij (Y )) +
θ̂ ij (Y ) +
,
h(Y ij )
∂Y ij
i=1 j=1
i=1 j=1
which completes the proof.
A.5
Proof of Proposition 2.3
Thanks to the expression (2.14) of the MKLA risk for data sampled from a Poisson distribution,
it follows that
n X
m
n X
m
h f
f i
X
X
f
X ij − X ij log (X ij ) =
E X̂ ij − X ij log X̂ ij
MKLA(θ̂ , θ) +
i=1 j=1
i=1 j=1
h(Y −1)
In the case of Poisson data, one has that exp (θ ij ) = X ij and h(Yij ) = Y ij . Therefore, by
ij
f
applying Hudson’s Lemma 2.1 with Fij (Y ) = log X̂ ij , it follows that
m
n X
m
n X
f
X
X
X ij log X̂ ij = E
Y ij log fij (Y − ei etj ) ,
E
i=1 j=1
i=1 j=1
which completes the proof.
45
B
Implementation details
We discuss below an algorithmic approach to find data-driven spectral estimators.
First, we discuss on how to compute data-driven spectral estimators from the expression of
risk estimators. For SUKLS in continuous exponential families, and for SURE in the Gaussian
case only, eq. (3.7) and (2.10) provide respectively a closed-form solution that can be evaluated
in linear time O(nm). On the contrary, the computations of GSURE (beyond the Gaussian
case), PURE and PUKLA, given respectively in eq. (2.6), (2.12) and (2.15), cannot be evaluated
f
in reasonable time. They rely respectively
of the divergence div θ̂ (Y ),
on the computation
PP
PP
t
t
Y ij fij (Y − ei ej ) and
Y ij log fij (Y − ei ej ) . Without further assumptions, such
quantities requires O(n2 m2 ) operations in general. A standard approach for the computation
of the divergence, suggested in [Gir89, RBU08], is to unbiasedly estimate it with Monte-Carlo
simulations by sampling the following relation
!#
"
f
∂
θ̂
(Y
)
f
δ
div θ̂ (Y ) = Eδ tr δ t
∂Y
at random directions δ ∈ Rn×m satisfying E[δ] = 0, E[δ i δ i ] = 1 and E[δ i δ j ] = 0. Following
[Del17], a similar first order approximation can be used for the other two quantities as
"
#
XX
XX
∂f
(Y
)
Y ij fij (Y − ei etj ) ≈
Y ij fij (Y ) − δ i,j
δ
, and
∂Y
i,j
"
#
XX
XX
∂f
(Y
)
δ
Y ij log fij (Y − ei etj ) ≈
Y ij log fij (Y ) − δ i,j
∂Y
i,j
where the entries of δ should be chosen Bernoulli distributed with parameter p = 0.5. The
advantage of these three approximations is that they can be computed in linear time O(nm) by
making use of the results of [LS01, SS03, Ede05, CSLT13, DVP+ 12] that provide an expression
for the directional derivative given by
∂f (Y )
t
δ = Ũ (D + S + A)Ṽ
∂Y
(B.1)
where Ũ and Ṽ are the matrices whose columns are ũk and ṽ k , and D, S and A are n × m
matrices defined, for all 1 ≤ i ≤ n and 1 ≤ j ≤ m, as
0
fi (σ̃i ) if i = j
Di,j = δ̄ i,j ×
0
otherwise,
(
0
if i = j
δ̄ i,j + δ̄ j,i
S i,j =
×
fi (σ̃i )−fj (σ̃j )
otherwise,
2
σ̃i −σ̃j
(
0
if i = j
δ̄ i,j − δ̄ j,i
Ai,j =
×
fi (σ̃i )+fj (σ̃j )
otherwise,
2
σ̃i +σ̃j
t
where σ̃k and fk (σ̃k ) are extended to 0 for k > min(n, m) and δ̄ = Ũ δ Ṽ ∈ Rn×m .
46
References
[ABB00]
O. Alter, P. O. Brown, and D. Botstein. Singular value decomposition for genomewide expression data processing and modeling. Proceedings of the National Academy
of Sciences (PNAS), 97(18), august 2000.
[AGZ10]
G. W. Anderson, A. Guionnet, and O. Zeitouni. An introduction to random matrices,
volume 118 of Cambridge Studies in Advanced Mathematics. Cambridge University
Press, Cambridge, 2010.
[Aka74]
H. Akaike. A new look at the statistical model identification. Automatic Control,
IEEE Transactions on, 19(6):716–723, 1974.
[BD06]
M. Bydder and J. Du. Noise reduction in multiple-echo data sets using singular value
decomposition. Magn Reson Imaging, 24(7):849–56, 2006.
[BMG13] J. Bazerque, G. Mateos, and G. Giannakis. Inference of Poisson count processes using
low-rank tensor data, pages 5989–5993. ICASSP, IEEE International Conference on
Acoustics, Speech and Signal Processing - Proceedings, 10 2013.
[BN12]
F. Benaych-Georges and R. R. Nadakuditi. The singular values and vectors of low
rank perturbations of large rectangular random matrices. J. Multivariate Analysis,
111:120–135, 2012.
[Bro86]
L. D. Brown. Fundamentals of statistical exponential families: with applications in
statistical decision theory. Institute of Mathematical Statistics, 1986.
[BS06]
J. Baik and J. W. Silverstein. Eigenvalues of large sample covariance matrices of
spiked population models. Journal of Multivariate Analysis, 97(6):1382 – 1408, 2006.
[BS10]
Z. Bai and J. W. Silverstein. Spectral analysis of large dimensional random matrices.
Springer Series in Statistics. Springer, New York, second edition, 2010.
[CR09]
D. J. Candès and B. Recht. Exact matrix completion via convex optimization. Found.
Comput. Math., 9(6):717–772, 2009.
[CSLT13] E. J. Candès, C. A. Sing-Long, and J. D. Trzasko. Unbiased risk estimates for singular
value thresholding and spectral estimators. IEEE Trans. Signal Process., 61(19):4643–
4657, 2013.
[CTT14]
Y. Choi, J. Taylor, and R. Tibshirani. Selecting the number of principal components:
Estimation of the true rank of a noisy matrix. Preprint arXiv:1405.7511, 2014.
[CX16]
Y. Cao and Y. Xie. Poisson matrix recovery and completion. IEEE Trans. Signal
Processing, 64(6):1609–1620, 2016.
[Del17]
C.-A. Deledalle. Estimation of kullback-leibler losses for noisy recovery problems
within the exponential family. Electronic Journal of Statistics, 11(2):3141–3164, 2017.
47
[DG14]
D. Donoho and M. Gavish. Minimax risk of matrix denoising by singular value
thresholding. Ann. Statist., 42(6):2413–2440, 12 2014.
[DS07]
R. B. Dozier and J. W. Silverstein. On the empirical distribution of eigenvalues of large
dimensional information-plus-noise-type matrices. J. Multivariate Anal., 98(4):678–
694, 2007.
[DVP+ 12] C.-A. Deledalle, S. Vaiter, G. Peyré, J. Fadili, and C. Dossal. Risk estimation for
matrix recovery with spectral regularization. In arXiv:1205.1482, 2012. Presented at
ICML’2012 workshop on Sparsity, Dictionaries and Projections in Machine Learning
and Signal Processing, Edinburgh, United Kingdom, 2012.
[Ede05]
A. Edelman. Matrix jacobians with wedge products. MIT Handout for 18.325, 2005.
[Efr04]
B. Efron. The estimation of prediction error: Covariance penalties and crossvalidation. Journal of the American Statistical Association, pages 99–467, 2004.
[Eld09]
Y. C. Eldar. Generalized sure for exponential families: Applications to regularization.
IEEE Transactions on Signal Processing, 57(2):471–481, 2009.
[EM72]
B. Efron and C. Morris. Empirical Bayes on Vector Observations: An Extension of
Stein’s Method. Biometrika, 59(2):335–347, 1972.
[EM76]
B. Efron and C. Morris. Multivariate empirical bayes and estimation of covariance
matrices. Ann. Statist., 4(1):22–32, 01 1976.
[EY36]
C. Eckart and G. Young. The approximation of one matrix by another of lower rank.
Psychometrika, 1(3):211–218, 1936.
[GD14a]
M. Gavish and D. Donoho.
arXiv:1405.7511, 2014.
[GD14b]
M. Gavish and D. L. Donoho. The optimal hard threshold for singular values is
\(4/\sqrt {3}\). IEEE Trans. Information Theory, 60(8):5040–5053, 2014.
[Gir89]
A. Girard. A fast monte-carlo cross-validation procedure for large least squares problems with noisy data. Numerische Mathematik, 56(1):1–23, 1989.
[Goo76]
J. W. Goodman. Some fundamental properties of speckle. JOSA, 66(11):1145–1150,
1976.
[Hal87]
P. Hall. On Kullback-Leibler loss and density estimation. Ann. Statist., 15(4):1491–
1519, 1987.
[HJ91]
R. A. Horn and C. R. Johnson. Topics in matrix analysis. Cambridge University
Press, Cambridge, New York, Melbourne, 1991. Suite de : Matrix analysis. 1985.
[HL06]
J. Hannig and T. C. M. Lee. On Poisson signal estimation under Kullback-Leibler
discrepancy and squared risk. J. Statist. Plann. Inference, 136(3):882–908, 2006.
Optimal shrinkage of singular values.
48
Preprint
[Hud78]
H. M. Hudson. A natural identity for exponential families with applications in multiparameter estimation. Ann. Statist., 6(3):473–484, 05 1978.
[Jol02]
I. T. Jolliffe. Principal component analysis. Springer Series in Statistics. SpringerVerlag, New York, second edition, 2002.
[JS15]
J. Josse and S. Sardy. Adaptive shrinkage of singular values. Statistics and Computing,
pages 1–10, 2015.
[Laf15]
J. Lafond. Low rank matrix completion with exponential family noise. In Proceedings
of The 28th Conference on Learning Theory, COLT 2015, Paris, France, July 3-6,
2015, pages 1224–1243, 2015.
[LBH+ 12] F. Lam, S. D. Babacan, J. P. Haldar, N. Schuff, and Z.-P. Liang. Denoising diffusionweighted MR magnitude image sequences using low rank and edge constraints. In
ISBI, pages 1401–1404. IEEE, 2012.
[LS01]
A. Lewis and H. Sendov. Twice differentiable spectral functions. SIAM Journal on
Matrix Analysis on Matrix Analysis and Applications, 23:368–386, 2001.
[LW12]
O. Ledoit and M. Wolf. Nonlinear shrinkage estimation of large-dimensional covariance matrices. Ann. Statist., 40(2):1024–1060, 04 2012.
[Nad14]
R. R. Nadakuditi. OptShrink: an algorithm for improved low-rank signal matrix
denoising by optimal, data-driven singular value shrinkage. IEEE Trans. Inform.
Theory, 60(5):3002–3018, 2014.
[NPDL11] H. M. Nguyen, X. Peng, M. N. Do, and Z.-P. Liang. Spatiotemporal denoising of
mr spectroscopic imaging data by low-rank approximations. In ISBI, pages 857–860.
IEEE, 2011.
[PL03]
L. Pastur and A. Lejay. Matrices aléatoires: statistique asymptotique des valeurs
propres. In Séminaire de Probabilités, XXXVI, volume 1801 of Lecture Notes in
Math., pages 135–164. Springer, Berlin, 2003.
[RBU08]
S. Ramani, T. Blu, and M. Unser. Monte-Carlo SURE: a black-box optimization of
regularization parameters for general denoising algorithms. IEEE Trans. on Image
Processing, 17(9):1540–1554, 2008.
[RS07]
M. Raphan and E. P. Simoncelli. Learning to be Bayesian without supervision. In
Advances in Neural Inf. Process. Syst. (NIPS), volume 19, pages 1145–1152. MIT
Press, 2007.
[SH05]
H. Shen and J. Z. Huang. Analysis of call centre arrival data using singular value
decomposition: Research articles. Appl. Stoch. Model. Bus. Ind., 21(3):251–263, May
2005.
[SN13]
A. A. Shabalin and A. B. Nobel. Reconstruction of a low-rank matrix in the presence
of Gaussian noise. J. Multivariate Anal., 118:67–76, 2013.
49
[SS03]
D. Sun and J. Sun. Nonsmooth matrix valued functions defined by singular values.
Technical report, Department of Decision Sciences, National University of Singapore,
2003.
[Ste81]
C. M. Stein. Estimation of the mean of a multivariate normal distribution. Ann.
Statist., 9(6):1135–1151, 1981.
[UD89]
F. T. Ulaby and M. C. Dobson. Handbook of Radar Scattering Statistics for Terrain.
Norwood, MA: Artech House, 1989.
[UHZB16] M. Udell, C. Horn, R. Zadeh, and S. Boyd. Generalized low rank models. Foundations
and Trends in Machine Learning, 9(1):1–118, 2016.
[WDB01] M. Wall, P. Dyck, and T. Brettin. Svdman-singular value decomposition analysis of
microarray data. Bioinformatics, 17(6):566–568, 2001.
[Yan94]
T. Yanagimoto. The Kullback-Leibler risk of the Stein estimator and the conditional
MLE. Ann. Inst. Statist. Math., 46(1):29–41, 1994.
50
| 10math.ST
|
On the complexity of computing the
k-restricted edge-connectivity of a graph?
Luis Pedro Montejano1 and Ignasi Sau2
arXiv:1502.07659v2 [cs.DS] 17 Sep 2016
1
Département de Mathématiques, Université de Montpellier, Montpellier, France.
[email protected]
2
AlGCo project team, CNRS, LIRMM, Montpellier, France.
[email protected]
Abstract. The k-restricted edge-connectivity of a graph G, denoted by
λk (G), is defined as the minimum size of an edge set whose removal leaves
exactly two connected components each containing at least k vertices.
This graph invariant, which can be seen as a generalization of a minimum edge-cut, has been extensively studied from a combinatorial point
of view. However, very little is known about the complexity of computing λk (G). Very recently, in the parameterized complexity community
the notion of good edge separation of a graph has been defined, which
happens to be essentially the same as the k-restricted edge-connectivity.
Motivated by the relevance of this invariant from both combinatorial and
algorithmic points of view, in this article we initiate a systematic study of
its computational complexity, with special emphasis on its parameterized
complexity for several choices of the parameters. We provide a number
of NP-hardness and W[1]-hardness results, as well as FPT-algorithms.
Keywords: Graph cut; k-restricted edge-connectivity; Good edge separation; Parameterized complexity; FPT-algorithm; Polynomial kernel.
1
Introduction
Motivation. The k-restricted edge-connectivity is a graph invariant that has
been widely studied in the literature from a combinatorial point of view [1, 4,
17, 22, 32, 33]. Since the classical edge-connectivity may not suffice to measure
accurately how connected a graph is after deleting some edges, Esfahanian and
Hakimi [16] proposed in 1988 the notion of restricted edge-connectivity. Given
a graph G, a non-empty set S ⊆ E(G) is an edge-cut if G − S has at least
two connected components. An edge-cut S is called a restricted edge-cut if there
are no isolated vertices in G − S. The restricted edge-connectivity λ0 (G) is the
minimum cardinality over all restricted edge-cuts S.
Inspired by the above definition, Fàbrega and Fiol [17] proposed in 1994 the
notion of k-restricted edge-connectivity, where k is a positive integer, generalizing
this notion. An edge-cut S is called a k-restricted edge-cut if every component
?
An extended abstract of this article appeared in the Proceedings of the 41st International Workshop on Graph-Theoretic Concepts in Computer Science (WG), pages
219-233, volume 9224 of LNCS, Garching, Germany, June 2015.
2
L. P. Montejano and I. Sau
of G − S has at least k vertices. Assuming that G has k-restricted edge-cuts, the
k-restricted edge-connectivity of G, denoted by λk (G), is defined as the minimum
cardinality over all k-restricted edge-cuts of G, i.e.,
λk (G) = min{|S| : S ⊆ E(G) is a k-restricted edge-cut}.
Note that for any graph G, λ1 (G) is the size of a minimum edge-cut, and
λ2 (G) = λ0 (G). A connected graph G is called λk -connected if λk (G) exists. Let
[X, Y ] denote the set of edges between two disjoint vertex sets X, Y ⊆ V (G),
and let X denote the complement X = V (G) \ X of vertex set X. It is clear that
for any k-restricted cut [X, X] of size λk (G), the graph G − [X, X] has exactly
two connected components.
Very recently, Chitnis et al. [6] defined the notion of good edge separation for
algorithmic purposes. For two positive integers k and `, a partition (X, X) of
the vertex set of a connected graph G is called a (k, `)-good edge separation if
|X|, |X| > k, |[X, X]| ≤ `, and both G[X] and G[X] are connected. That is, it
holds that λk (G) ≤ ` if and only if G admits a (k − 1, `)-good edge separation.
Thus both notions, which have been defined independently and for which there
existed no connection so far, are essentially the same.
Good edge separations turned out to be very useful for designing parameterized algorithms for cut problems [6], by using a technique known as recursive understanding, which basically consists in breaking up the input graph into highly
connected pieces in which the considered problem can be efficiently solved. It
should be mentioned that Kawarabayashi and Thorup [23] had defined before a
very similar notion for vertex-cuts and introduced the idea of recursive understanding. This technique has also been subsequently used in [10, 24, 29].
Very little is known about the complexity of computing the k-restricted edgeconnectivity of a graph, in spite of its extensive study in combinatorics. In this
article we initiate a systematic analysis on this topic, with special emphasis on
the parameterized complexity of the problem. In a nutshell, the main idea is
to identify relevant parameters of the input of some problem, and study how
the running time of an algorithm solving the problem depends on the chosen
parameters. See [8, 13, 18, 28] for introductory textbooks to this area.
Our results. We consider the following two problems concerning the krestricted edge-connectivity of a graph.
Existential Restricted Edge-connectivity (EREC)
Instance: A graph G = (V, E) and a positive integer k.
Question: Is G λk -connected ?
Restricted Edge-connectivity (REC)
Instance: A connected graph G = (V, E) and a positive integer k.
Output: λk (G), or a correct report that G is not λk -connected.
Computing the k-restricted edge-connectivity of a graph
3
The latter problem can be seen as a generalization of computing a Minimum Cut in a graph, which is polynomial-time solvable [30]. In Section 2 we
prove that it is NP-hard, even restricted to λk -connected graphs. In Section 3
we study the parameterized complexity of the REC problem. More precisely,
given a connected graph G and two integers k and `, we consider the problem of
determining whether λk (G) ≤ `. Existing results concerning good edge separations imply that the problem is FPT when parameterized by k and `. We prove
that it is W[1]-hard when parameterized by k, and that it is FPT when parameterized by ` but unlikely to admit polynomial kernels. Moreover, we prove that
the EREC problem is FPT when parameterized by k. Finally, in Section 4 we
also consider the maximum degree ∆ of the input graph as a parameter, and
we prove that the EREC problem remains NP-complete in graphs with ∆ ≤ 5,
and that the REC problem is FPT when parameterized by k and ∆. Note that
this implies, in particular, that the REC problem parameterized by k is FPT in
graphs of bounded degree. Table 1 below summarizes the results of this article.
Problem
Is G
λk -connected ?
λk (G) ≤ ` ?
Classical
complexity
Parameterized complexity with parameter
k+`
k
`
k+∆
NPc, even
FPT
FPT
if ∆ ≤ 5
?
(Thm 4)
?
(Thm 8)
(Thm 7)
NPh, even if G
FPT W[1]-hard FPT (Thm 5)
FPT
is λk -connected (Thm 2, (Thm 3) No poly kernels (Thm 8)
(Thm 1)
by [6])
(Thm 6)
Table 1. Summary of our results, where ∆ denotes the maximum degree of the
input graph G, and NPc (resp. NPh) stands for NP-complete (resp. NP-hard).
The symbol ‘?’ denotes that the problem is not defined for that parameter.
Further research. Some open questions are determining the existence of
polynomial kernels for the REC problem with parameters k + ` or k + ∆,
speeding-up the FPT algorithm of Theorem 5 (which is quite inefficient), improving the bound on the maximum degree in Theorem 7, and studying the
(parameterized) complexity of the REC problem in planar graphs and other
sparse graph classes.
Notation. We use standard graph-theoretic notation; see for instance [11].
For a graph G, let ∆(G) denote its maximum degree, and for a vertex v, its
degree in G is denoted by dG (v). If S ⊆ V (G), we define G − S = G[V (G) \ S],
and if S ⊆ E(G), we define G − S = (V (G), E(G) \ S). Unless stated otherwise,
throughout the article n denotes the number of vertices of the input graph of
the problem under consideration. We will always assume that the input graphs
are connected.
4
2
L. P. Montejano and I. Sau
Preliminary results
Clearly, any connected graph G is λ1 -connected, and λ1 (G) can be computed in
polynomial time by a Minimum Cut algorithm (cf. [30]). However, for k ≥ 2,
there exist infinitely many connected graphs which are not λk -connected, such
as the graphs containing a cut vertex u such that every component of G − u
has at most k − 1 vertices (these graphs are called flowers in the literature [4],
and correspond exactly to stars when k = 2). Moreover, the EREC problem is
hard. Indeed, given a graph G, if n is even and k = n/2, by [14, Theorem 2.2] it
is NP-complete to determine whether G contains two vertex-disjoint connected
subgraphs of order n/2 each. We can summarize this discussion as follows.
Remark 1. The EREC problem is NP-complete.
In Section 4 we will strengthen the above hardness result to the case where
the maximum degree of the input graph is at most 5.
Note that Remark 1 implies that the REC problem problem is NP-hard. Furthermore, even if the input graph G is guaranteed to be λk -connected, computing
λk (G) remains hard, as shown by the following theorem.
Theorem 1. The REC problem is NP-hard restricted to λk -connected graphs.
Proof: We prove it for n even and k = n/2. The reduction is from the Minimum
Bisection problem1 restricted to connected 3-regular graphs, which is known
to be NP-hard [2]. Given a 3-regular connected graph G with even number of
vertices as instance of Minimum Bisection, we construct from it an instance
G0 of REC by adding two non-adjacent universal vertices v1 and v2 . Note that
G0 is λn/2 -connected, since any bipartition of V (G0 ) containing v1 and v2 in
different parts induces two connected subgraphs.
We claim that v1 and v2 should necessarily belong to different connected
subgraphs in any optimal solution in G0 . Indeed, let (V1 , V2 ) be a bipartition of
V (G) such that |[V1 , V2 ]| = λn/2 (G0 ), and assume for contradiction that v1 , v2 ∈
V1 . Since G is connected, there is a vertex u ∈ V2 with at least one neighbor
in V1 \ {v1 , v2 }. Let V10 := V1 ∪ {u} \ {v2 } and V20 := V2 ∪ {v2 } \ {u}, and note
that both G[V10 ] and G[V20 ] are connected. Since u has at least one neighbor in
V1 \ {v1 , v2 }, G is 3-regular, and v1 and v2 are non-adjacent and adjacent to all
other vertices of G0 , it can be checked that |[V10 , V20 ]| ≤ |[V1 , V2 ]|−1 = λn/2 (G0 )−1,
contradicting the definition of λn/2 (G0 ).
Therefore, solving the REC problem in G0 corresponds exactly to solving the
Minimum Bisection problem in G, concluding the proof.
1
Given a graph G with even number of vertices, the Minimum Bisection problem
consists in partitioning V (G) into two equally-sized parts minimizing the number of
edges with one endpoint in each part.
Computing the k-restricted edge-connectivity of a graph
3
5
A parameterized analysis
The NP-hardness results of the previous section naturally lead to considering
parameterized versions of the problem. In this section we consider the following
three distinct parameterizations.
Parameterized Restricted Edge-connectivity (p-REC)
Instance: A connected graph G = (V, E) and two integers k and `.
Parameter 1: The integers k and `.
Parameter 2: The integer k.
Parameter 3: The integer `.
Question: λk (G) ≤ ` ?
As mentioned in the introduction, determining whether λk (G) ≤ ` corresponds exactly to determining whether G admits a (k − 1, `)-good edge separation. This latter problem has been recently shown to be solvable in time
2O(min{k,`} log(k+`)) · n3 log n by Chitnis et al. [6, Lemma II.2].
Theorem 2 (Chitnis et al. [6]). The p-REC problem is FPT when parameterized by both k and `.
We would like to note that any improvement on the running time of the
algorithm behind Theorem 2 would answer an open question raised in [7,9], and
would have direct consequences and improve the algorithms described in [6, 10,
24].
As pointed out in [16,21], the p-REC problem can be solved in time O∗ (n2k ).
Roughly, the idea is to guess two sets of k vertices inducing a connected subgraph, contract them into two vertices s and t, and then call a polynomial-time
Minimum Cut algorithm between s and t (cf. [30]). In other words, it is in XP
when parameterized by k. The following theorem shows that this is essentially
the best algorithm we can hope for when the parameter is only k. Indeed, since
the blow-up of the parameter in the reduction is linear, the fact that k-Clique
cannot be solved in time f (k) · no(k) unless an unlikely collapse occurs in parameterized complexity theory [5] implies that the p-REC problem cannot be
solved in time f (k) · no(k) either.
Theorem 3. The p-REC problem is W[1]-hard when parameterized by k.
Proof: We reduce from k-Clique, which is known to be W[1]-hard [13]. The
parameterized reduction is the same as the one given by Downey et al. in [12,
Theorem 2] to show the W[1]-hardness of the Cutting k Vertices from a
Graph problem, only the analysis changes.
Let G = (V, E) be an n-vertex graph for which we wish to determine whether
it has a k-clique. We construct a graph G0 as follows:
(1) We start with a clique C of size n3 and n representative vertices corresponding bijectively with the vertices of G.
6
L. P. Montejano and I. Sau
(2) Every representative vertex v is connected to n2 − dG (v) arbitrary vertices
of C.
(3) If uv ∈ E(G) then uv ∈ E(G0 ).
C = Kn3
n2 − dG(v)
v
G
n representative vertices
Fig. 1. Illustration of the graph G0 in the proof of Theorem 3.
See Fig. 1 for an illustration of G0 . Consider ` = kn2 −2 k2 and take k ≤ n/2.
We claim that G has a k-clique if and only if G0 is a Yes-instance of p-REC.
Suppose first that K ⊆ V (G) is a k-clique in G. Obviously, K is connected in
G0 and has k vertices. On the other hand, G0 − K is also connected with at least
n3 −|K| >k vertices. Finally, it is straightforward to check that |[K, V (G0 )\K]| =
kn2 − 2 k2 = `.
In the other direction, suppose G0 has a k-restricted edge-cut with at most
` edges, i.e., there exists K ⊆ V (G0 ) such that G[K] and G0 − K are connected,
|K| ≥ k, |V (G0 ) \ K| ≥ k, and |[K, V (G0 ) \ K]| ≤ `. Two cases need to be
distinguished.
Case 1. C ∩ K = ∅. Then every vertex of K must be a representative vertex.
Hence |[K, V (G0 ) \ K]| = |K|n2 − 2|E(G0 [K])| since every representative
vertex
has degree n2 . As by hypothesis |[K, V (G0 ) \ K]| ≤ ` = kn2 − 2 k2 , it follows
that |E(G0 [K])| = k2 , hence K must be a k-clique.
Case 2. C ∩ K 6= ∅. Note that for every bipartition (C1 , C2 ) of C we have
that |[C1 , C2 ]| ≥ n3 − 1 > `. Now suppose C ∩ (V (G0 ) \ K) 6= ∅ and consider the
bipartition C1 = C ∩ K and C2 = C ∩ (V (G0 ) \ K) of C. Then |[K, V (G0 ) \ K]| ≥
|[C1 , C2 ]| ≥ n3 −1 > `, a contradiction. Therefore, we have that C ∩(V (G0 )\K) =
∅. The proof concludes by applying Case 1 to V (G0 ) \ K instead of K.
In contrast to Theorem 3 above, we now prove that the EREC problem
(which is NP-complete by Remark 1) is FPT when parameterized by k. The
proof uses the technique of splitters introduced by Naor et al. [27], which has
also been recently used for designing parameterized algorithms in [6,10,24]. Our
main tool is the following lemma.
Lemma 1 (Chitnis et al. [6]). There exists an algorithm that given a set U of
size n and two integers a, b ∈ [0, n], outputs in time 2O(min{a,b}·log(a+b)) · n log n
a set F ⊆ 2U with |F| = 2O(min{a,b}·log(a+b)) · log n such that for every two sets
A, B ⊆ U , where A ∩ B = ∅, |A| ≤ a, and |B| ≤ b, there exists a set S ∈ F with
A ⊆ S and B ∩ S = ∅.
Computing the k-restricted edge-connectivity of a graph
7
Theorem 4. The EREC problem is FPT when parameterized by k. More precisely, it can be solved by an algorithm running in time 2O(k log k) · n2 log n.
Proof: We use the easy property that G is λk -connected if and only if G contains
two vertex-disjoint trees T1 and T2 such that |V (T1 )| ≥ k and |V (T2 )| ≥ k. In
order to apply Lemma 1, we take U = V (G) and a = b = k, obtaining in time
k O(k) · n log n the desired family F of subsets of vertices of G. Now, if such trees
T1 and T2 exist, then necessarily there exists a set S ∈ F such that V (T1 ) ⊆ S
and V (T2 ) ∩ S = ∅. Therefore, in order to determine whether G is λk -connected
or not, it suffices to check, for each set S ∈ F, whether both G[S] and G − S
contain a connected component with at least k vertices. (Note that for each such
set S ∈ F, this can be done in linear time.) Indeed, if such a set S exists, then
clearly G is λk -connected. Otherwise, by the property of the family F, G does
not contain two disjoint trees T1 and T2 of size k each, and therefore G is not
λk -connected.
Concerning the parameterized complexity of the p-REC problem, in view of
Theorems 2 and 3, it just remains to settle the case when the parameter is `
only. The following theorem provides an answer to this question. We will need
the following result, which is a reformulation of [31, Corollary 1].
Lemma 2 (van Bevern et al. [31]). Given a graph G on n vertices and an
integer `, determining whether λbn/2c (G) ≤ ` can be solved in time f (`) · n11 for
some explicit function f depending only on `.
Theorem 5. The p-REC problem is FPT when parameterized by `.
Proof: If k ≤ `, we solve the problem using the FPT algorithm given by Theorem 2 with parameter k + ` ≤ 2`. Otherwise, in the case where k > `, we proceed
to Turing-reduce the problem to the particular case where k = bn/2c, which is
FPTwith parameter ` by Lemma 2 above, as follows. For each vertex v of the
n-vertex input graph G, and for each integer p with 0 ≤ p ≤ n − 2k, let Gpv be
the graph obtained from G by adding a clique Kp on p vertices and all the edges
between vertex v and the vertices in Kp .
Claim. It holds that λk (G) ≤ ` if and only if there exist a vertex v ∈ V (G) and
an integer p with 0 ≤ p ≤ n − 2k such that λb n+p c (Gpv ) ≤ `.
2
Proof: Assume first that λk (G) ≤ `. Let (X, X) be a partition of V (G)
achieving λk (G), where we assume without loss of generality that X ≥ X, let
p = |X| − |X|, and let v be any vertex in X. Then p ≤ (n − k) − k = n − 2k and
by construction it holds that λb n+p c (Gpv ) ≤ λk (G) ≤ `.
2
Conversely, suppose that there exist a vertex v ∈ V (G) and an integer p
with 0 ≤ p ≤ n − 2k such that λb n+p c (Gpv ) ≤ `. Let (X, X) be a partition of
2
V (Gpv ) achieving λb n+p c (Gpv ), and assume without loss of generality that v ∈ X.
2
We claim that the clique Kp is entirely contained in X. Indeed, suppose for
contradiction that Kp ∩ X 6= ∅, and let K = Kp ∩ X. Since Gpv [X] is connected
8
L. P. Montejano and I. Sau
and v ∈ X, necessarily X = K. We distinguish
two cases. If p < k, then
|X| = |K| ≤ p − 1 ≤ k − 2 ≤ n/2 − 2 < n+p
, contradicting the definition
2
of λb n+p c (Gpv ). Otherwise, if p ≥ k, then we use that the number of edges in
2
[X, X] is at least the minimum cut of the clique of size p+1 induced by Kp ∪{v},
which is equal to p. That is, |[X, X]| ≥ p ≥ k > `, contradicting the hypothesis
that (X, X) is a partition of V (Gpv ) achieving λb n+p c (Gpv ) ≤ `.
2
We now claim that the partition of V (G) given by (X, X \ Kp ) defines a
k-restricted edge-cut of G with at most ` edges, concluding the proof. First note
that both G[X] and G[X \ Kp ] are connected, since both Gpv [X] and Gpv [X] are
p
connected by hypothesis, and the removal of Kp from G
v [X]
n+p
clearly
n preserves
connectivity. On the other hand, we have that |X| ≥
≥
≥ k and
2
2
n+p
n−(n−2k)−1
n−p−1
1
≥
= k − 2 , and since
|X \ Kp | = |X| − p ≥
−p ≥
2
2
2
both |X \ Kp | and k are integers, the latter inequality implies that |X \ Kp | ≥ k.
Finally, since Kp ⊆ X, it holds that |[X, X \ Kp ]| = |[X, X]| ≤ `.
3
Note that the above claim yields the theorem, as it implies that the problem
of deciding whether λk (G) ≤ ` can be solved by invoking O(n2 ) times the FPT
algorithm given by Lemma 2.
To complement Theorem 5, in the next theorem we prove that the p-REC
problem does not admit polynomial kernels when parameterized by `, unless
coNP ⊆ NP/poly.
Theorem 6. Unless coNP ⊆ NP/poly, the p-REC problem does not admit
polynomial kernels when parameterized by `.
Proof: The proof is strongly inspired by the one given by van Bevern et al. [31,
Theorem 3] to prove that the Minimum Bisection problem does not admit
polynomial kernels, which in turn resembles the proof given by Garey et al. [20]
to prove the NP-hardness of Minimum Bisection. The main difference with
respect to the proof given in [31] is that we need to make the appropriate modifications to guarantee that both parts left out by the edge-cut are connected,
which is not an issue in the Minimum Bisection problem.
We will first rule out the existence of polynomial kernels for the generalization
of p-REC where the edges have non-negative integer weights, and the objective
is to decide whether the input graph can be partitioned into two connected
subgraphs with at least k vertices each by removing a set of edges whose total
weight does not exceed `. We call this problem Edge-Weighted p-REC. Then
it will just remain to get rid of the edge weights. This is done at the end of the
proof of the theorem.
As shown by Bodlaender et al. [3], in order to prove that Edge-Weighted
p-REC does not admit polynomial kernels when parameterized by ` (assuming
that coNP ⊆ NP/poly), it is sufficient to define a cross composition from an NPhard problem to Edge-Weighted p-REC. In our case, the NP-hard problem is
Maximum Cut (see [19]), which is defined as follows. Given a graph G = (V, E)
and an integer p, one has to decide whether V can be partitioned into two sets
Computing the k-restricted edge-connectivity of a graph
9
A and B such that there are at least p edges with an endpoint in A and an
endpoint in B.
A cross composition from Maximum Cut to Edge-Weighted p-REC parameterized by ` consists of a polynomial-time algorithm that, given t instances
(G1 , p1 ), . . . , (Gt , pt ) of Maximum Cut, constructs an instance (G∗ , k, `) of
Edge-Weighted p-REC such that (G∗ , k, `) is a Yes-instance if and only
if one of the t instances of Maximum Cut is a Yes-instance, and such that ` is
polynomially bounded as a function of max1≤i≤t |V (Gi )|. Similarly to [31], we
may safely assume that for each 1 ≤ i ≤ t we have |V (Gi )| =: n and pi =: p
(by the arguments given in [3]), that 1 ≤ p ≤ n2 (as if p = 0 all instances
are Yes-instances, and if p > n2 all instances are No-instances), and that t is
odd (otherwise, we can add a No-instance consisting of an edgeless graph on n
vertices).
Given (G1 , p), . . . , (Gt , p), we create G∗ as follows; see Fig. 2 for an illustration. Let w1 := 5n2 and w2 := 5. For each graph Gi = (Vi , Ei ) add to G∗ the
vertices in Vi and a clique Vi0 with |Vi | = n vertices whose edges have weight
w1 . Add an edge of weight w1 between each vertex in Vi and each vertex in
Vi0 . For each pair of vertices u, v ∈ Vi , add the edge {u, v} to G∗ with weight
w1 − w2 if {u, v} ∈ Ei , and with weight w1 otherwise. Let s1i , s2i be two arbitrary distinct vertices in Vi0 , which we call link vertices. For 1 ≤ i ≤ t − 1,
add two edges with weight 1 between s1i and s1i+1 and between s2i and s2i+1 ,
and two edges with weight 1 between s1t and s11 and between s2t and s21 . This
completes the construction of G∗ . These 2t edges among distinct Vi0 ’s are called
chain edges (cf. the thicker edges in Fig. 2). Finally, we set k := |V (G∗ )|/2 and
` := w1 n2 − w2 p + 4. Note that k is not polynomially bounded in terms of n, but
this is not a problem since the parameter we consider is `, which is bounded by
5n4 . This construction can be clearly performed in polynomial time in t · n. We
claim that (G∗ , k, `) is a Yes-instance of Edge-Weighted p-REC if and only
if there exists i ∈ {1, . . . , t} such that (Gi , p) is a Yes-instance of Maximum
Cut.
V10
V1
V20
V2
1
Vi0
Vi
1
s11
s12
w1
w1−w2
s22
w1
1
s1t
w1
w1
w1
w1
s21
Vt0
Vt
s1i
s2i
1
Fig. 2. Illustration of the graph G∗ in the proof of Theorem 6.
s2t
10
L. P. Montejano and I. Sau
Assume first that there exists i ∈ {1, . . . , t} such that (Gi , p) is a Yesinstance of Maximum Cut. Assume without loss of generality that i = 1, and
let V1 = A ] B such that there are at least p edges in E1 between A and B.
We proceed to partition V (G∗ ) into two equally-sized sets A0 and B 0 such that
both G∗ [A0 ] and G∗ [B 0 ] are connected, and such that the total weight of the
edges in G∗ with one endpoint in A0 and one endpoint in B 0 is at most `. The
set A0 contains V1 ∩ A, any set of |B| vertices in V10 containing exactly one of
Sdt/2e
s11 and s21 (this is possible since 1 ≤ |B| ≤ n − 1), and i=2 Vi ∪ Vi0 . Then
B 0 = V (G∗ ) \ A0 . Since t is odd, |A0 | = |B 0 |. Let us now see that G∗ [A0 ] is
connected. As 1 ≤ |A| ≤ n − 1, the set V1 ∩ A0 is connected to V10 ∩ A0 , which is
connected to V20 since A0 contains exactly one of the link vertices s11 and s21 . The
Sdt/2e
graph G∗ [ i=2 Vi ∪ Vi0 ] is clearly connected because of the chain edges, which
implies that G∗ [A0 ] is indeed connected. The proof for the connectivity of G∗ [B 0 ]
is similar, using that V1 ∩ B 0 is connected to V10 ∩ B 0 since 1 ≤ |B| ≤ n − 1, which
is in turn connected to Vt0 since B 0 contains exactly one of the link vertices s11
and s21 . Finally, let us show that the total weight of the edges between A0 and
B 0 is at most `. Note first that two chain edges incident to V10 and two chain
0
edges incident to Vdt/2e
belong to the cut defined by A0 and B 0 , and no other
chain edge belongs to the cut. Beside the chain edges, only edges in the graph
G∗ [V1 ∪ V10 ] are cut. Note that G∗ [V1 ∪ V10 ] is a clique on 2n vertices and each of
(V1 ∪ V10 ) ∩ A0 and (V1 ∪ V10 ) ∩ B 0 contains n vertices. Since |[A, B]| ≥ p, at least
p of the edges of weight w1 − w2 belong to the cut. Therefore, the total weight
of the cut is at most w1 n2 − w2 p + 4 = `.
Conversely, assume that for all i ∈ {1, . . . , t}, (Gi , p) is a No-instance of
Maximum Cut, and we want to prove that (G∗ , k, `) is a No-instance of EdgeWeighted p-REC. Let A ] B be a partition of V (G∗ ) such that |A| = |B| and
both G∗ [A] and G∗ [B] are connected, and such that the weight of the edges
between A and B is minimized among all such partitions. For 1 ≤ i ≤ t, we let
ai := |(Vi ∪ Vi0 ) ∩ A|. Since for 1 ≤ i ≤ t, (Gi , p) is a No-instance of Maximum
Cut, any bipartition of Vi cuts at most p−1 edges. Therefore, the total weight of
the edges between (Vi ∪Vi0 )∩A and (Vi ∪Vi0 )∩B is at least w1 ai (2n−ai )−(p−1)w2 .
Since t is odd, necessarily at least one of the graphs Gi is cut by A]B. Assume
first that exactly one graph Gi is cut by A ] B. Since |A| = |B|, we have that
ai = n, so the value of the cut is at least w1 n2 − (p − 1)w2 = w1 n2 − pw2 + w2 >
w1 n2 − pw2 + 4 = `, and thus (G∗ , k, `) is a No-instance of Edge-Weighted
p-REC.
We claim that there is always exactly one graph Gi cut by A ] B. Assume for
contradiction that it is not the case, that is, that there are two strictly positive
values ai , aj for some i 6= j. By symmetry between A and B, we may assume
that ai +aj ≤ 2n. The total weight of the edges cut in G∗ [Vi ∪Vi0 ] and G∗ [Vj ∪Vj0 ]
is at least
w1 ai (2n − ai ) − (p − 1)w2 + w1 aj (2n − aj ) − (p − 1)w2 =
2nw1 (ai + aj ) − w1 (a2i + a2j ) − 2w2 (p − 1).
Computing the k-restricted edge-connectivity of a graph
11
Now we construct another solution A0 ] B 0 of Edge-Weighted p-REC in
G where the ai + aj vertices are cut in only one of Vi ∪ Vi0 and Vj ∪ Vj0 , say
Vi ∪ Vi0 (note that this is possible since ai + aj ≤ 2n). In order to do so, as far
as there exists such a pair ai , aj , we proceed as follows. If ai + aj = 2n, then
Vi ∪ Vi0 is entirely contained in A0 or B 0 . Otherwise, if ai + aj < 2n, then we
choose (Vi ∪ Vi0 ) ∩ A0 such that it contains exactly one of s1i and s2i . At the end
of this procedure, exactly one graph Gi is cut. Finally, we arrange the other
G` ’s, with ` 6= i, consecutively into A0 and B 0 . That is, we put the vertices of
Si+bt/2c
Si−1
0
0
0
0
`=i+dt/2e V` ∪V` into B , where the
`=i+1 V` ∪V` into A , and the vertices of
indices are counted cyclically from 1 to t. The connectivity of both G∗ [A0 ] and
G∗ [B 0 ] is guaranteed by the chain edges and the choice of the selector vertices
s1i and s2i .
Let i, j be two indices for which the above procedure has been applied. Taking
into account that each Vi0 has four incident chain edges, the total weight of the
edges cut in G∗ [Vi ∪ Vi0 ] and G∗ [Vj ∪ Vj0 ] by the new solution is at most
∗
w1 (ai + aj )(2n − ai − aj ) + 8 =
2nw1 (ai + aj ) − w1 (a2i + a2j ) − 2w1 ai aj + 8.
That is, the weight of the cut defined by A ] B minus the weight of the cut
defined by A0 ] B 0 is at least
−2w2 (p − 1) + 2w1 ai aj − 8 =
2(w1 ai aj − w2 (p − 1) − 4) ≥
2(w1 − w2 (n2 − 1) − 4) > 0,
where we have used that ai , aj ≥ 1, p ≤ n2 , w1 = 5n2 , and w2 = 5. In
other words, A0 ] B 0 defines a cut of strictly smaller weight, contradicting the
definition of A ] B.
To conclude the proof of the theorem, it just remains to deal with the
edge weights. As in [31], we show how to convert the instance (G∗ , k, `) of
Edge-Weighted p-REC that we just constructed into an equivalent instance
of p-REC such that the resulting parameter remains polynomial in n. Given
(G∗ , k, `), we define (Ĝ, k, `) as the instance of p-REC, where Ĝ is an unweighted
graph obtained from G∗ as follows. We replace each vertex v of G∗ with a clique
Cv of size w1 + ` + 1, and for each edge {u, v} of G∗ with weight w, we add w
pairwise disjoint edges between the cliques Cu and Cv . Since no cut of size at
most ` in Ĝ can separate a clique Cv introduced for a vertex v, it follows that
(G∗ , k, `) is a Yes-instance of Edge-Weighted p-REC if and only if (Ĝ, k, `)
is a Yes-instance of p-REC. Finally, it is clear that the desired cut size ` is still
polynomial in n.
4
Considering the maximum degree as a parameter
Towards understanding the parameterized complexity of the REC problem, one
may wonder whether considering the maximum degree of the input graph as
12
L. P. Montejano and I. Sau
an extra parameter turns the problem easier (this is a classical approach in
parameterized complexity, see for instance [25, 26]). We first prove that, from a
classical complexity point of view, bounding the degree of the input graph does
not turn the problem easier. Before stating the hardness result, we need the
define the 3-Dimensional Matching problem, 3DM for short.
An instance of 3DM consists of a set W = R ∪ B ∪ Y , where R, B, Y are
disjoint sets with |R| = |B| = |Y | = m, and a set of triples T ⊆ R × B × Y . The
question is whether there exists a matching M ⊆ T covering W , i.e., |M | = m
and each element of W = R ∪ B ∪ Y occurs in exactly one triple of M .
An instance of 3DMScan be represented by a bipartite graph GI = (W ∪
T, EI ), where EI =
{{r, t}, {b, t}, {y, t}}; see Fig. 3.
t=(r,b,y)∈T
r
b
y
Elements W = R ∪ B ∪ Y
Triples T
(r, b, y)
Fig. 3. Representation of an instance of 3DM.
It is known that 3DM is NP-complete even if each element of W appears in
2 or 3 triples only [14, 15]. In [14, Theorem 2.2] it is proved that partitioning a
graph G into two connected subgraphs of equal size is NP-hard, using a reduction
from 3DM. It is worth noting that the graph constructed in the NP-hardness
reduction contains only two vertices of degree greater than five. In Theorem 7 we
appropriately modify the reduction of [14, Theorem 2.2] so that the constructed
graph has maximum degree at most 5.
Theorem 7. The EREC problem is NP-complete even if the maximum degree
of the input graph is 5.
Proof: Given an instance (W, T ) of 3DM with W = R ∪ B ∪ Y , |R| = |B| =
|Y | = m, and T ⊆ R × B × Y such that each element of W appears in 2 or 3
triples only, we define an n-vertex graph G = (V, E) with maximum degree 5 as
follows (see Fig. 4 for an illustration).
The set of vertices of G is
V = W ∪ T ∪ Ta ∪ Tb ∪ P ∪ {a},
Computing the k-restricted edge-connectivity of a graph
13
nb
Elements
W
Triples
T
Pa
a
Ta
na
Tb
b
nb
Fig. 4. Construction of the graph G in the proof of Theorem 7, with ∆(G) = 5.
where Ta = {ta1 , . . . , ta|T | }, Tb = {b = tb1 , tb2 , . . . , tb|T | }, T = {t1 , . . . , t|T | } is the
S
set of triples, and P =
Pσ , where Pσ = {(σ, t) : t = 1, . . . , nσ } with
σ∈W ∪Tb ∪{a}
na = (3m + |T |)nb + 5m − |T | − 1, nb = 2m3 , and nσ = nb for every σ ∈ W ∪ Tb .
The set of edges of G is
[
E = EI ∪ ETa ∪ ETb ∪ ET +
Eσ ,
σ∈W ∪Tb ∪{a}
where ETa = {{tai , tai+1 } : 1 ≤ i ≤ |T | − 1}, ETb = {{tbi , tbi+1 } : 1 ≤ i ≤ |T | − 1},
ET + = {{ti , tai }, {ti , tbi } : 1 ≤ i ≤ |T |}, and Eσ = {{σ, (σ, 1)}} ∪ {{(σ, t), (σ, t +
1)} : 1 ≤ t ≤ nσ − 1} ∪ {a, ta1 } for every σ ∈ W ∪ Tb ∪ {a}.
Note that the maximum degree of G is indeed 5 (only vertices of T could get
degree 5, all other vertices have degree at most 4). Since n = 1 + 3m + 3|T | +
na + (3m + |T |)nb , we can observe that
n = 2(na + 1 + 2|T | − m).
Next, we show that for k = n/2, G is Yes-instance of the REC problem if
and only if T contains a matching covering W .
One direction is easy. Suppose first that T contains a matching M covering
W . Let S = {a} ∪ Pa ∪ Ta ∪ (T \ M ). It is straightforward to check that |S| = n/2
and that G[S], G[V \ S] are both connected.
Conversely, suppose that G can be partitioned into 2 connected subgraphs
G[S], G[V \S] with |S| = n/2. We can assume that a ∈ S, and then it follows that
14
L. P. Montejano and I. Sau
Pa ⊆ S. Now |S \ (Pa ∪ {a})| = 2|T | − m < 2m3 = nb since |T | ≤ m3 . As Pσ ⊆ S
if and only if σ ∈ S ∩ (W ∪ Tb ), then S ∩ (W ∪ Tb ) = ∅ since |S \ (Pa ∪ {a})| < nb
and |Pσ | = nb for every σ ∈ W ∪ Tb . Hence S \ (Pa ∪ {a}) ⊆ T ∪ Ta . Let
M = (V \ S) ∩ T . Then |M | ≤ m since |S \ (Pa ∪ {a})| = 2|T | − m. Finally,
as G[V \ S] is connected and W ∪ Tb ⊆ V \ S, it follows that |M | ≥ m. Hence
|M | = m and M must be a matching covering W .
In order to understand to which extent the vertices of high degree make the
complexity of computing the restricted edge-connectivity of a graph hard, we
also consider the maximum degree of the input graph as a parameter for the
p-REC problem.
Theorem 8. The p-REC problem is FPT when parameterized by k and the
maximum degree ∆ of the input graph.
Proof: The algorithm is based on a simple exhaustive search. We use the property that, for any graph G and any two integers k, `, λk (G) ≤ ` if and only if G
contains two vertex-disjoint trees T1 and T2 with |V (T1 )| ≥ k and |V (T2 )| ≥ k,
such that there exists an edge set S in G with |S| ≤ ` such that in G − S the
trees T1 and T2 belong to different connected components. Hence, we just have
to determine whether these trees exist in G or not. For doing so, for every pair
of distinct vertices v1 and v2 of G, we exhaustively consider all trees T1 and T2
with k vertices containing v1 and v2 , respectively. Note that the number of such
trees is at most ∆2k . For every pair of vertex-disjoint trees T1 and T2 , we proceed
as follows. We contract tree T1 (resp. T2 ) to a single vertex t1 (resp. t2 ), keeping
edge multiplicities, and then we run in the resulting graph a polynomial-time
Minimum Cut algorithm between t1 and t2 (cf. [30]). If the size of the returned
edge-cut is at most `, then T1 and T2 are the desired trees. Otherwise, we continue searching. It is clear that the overall running time of this algorithm is
O(∆2k · nO(1) ).
Acknowledgement. We would like to thank the anonymous referees for helpful
remarks that improved the presentation of the manuscript. We are particularly
grateful for the ideas to prove Theorem 5, a result that we had left as an open
question in the conference version of this article.
References
1. C. Balbuena, A. Carmona, J. Fàbrega, and M. A. Fiol. Extraconnectivity of graphs
with large minimum degree and girth. Discrete Mathematics, 167:85–100, 1997.
2. P. Berman and M. Karpinski. Approximation hardness of bounded degree MINCSP and MIN-BISECTION. Electronic Colloquium on Computational Complexity,
8(26), 2001.
3. H. L. Bodlaender, B. M. P. Jansen, and S. Kratsch. Kernelization lower bounds by
cross-composition. SIAM Journal on Discrete Mathematics, 28(1):277–305, 2014.
4. P. Bonsma, N. Ueffing, and L. Volkmann. Edge-cuts leaving components of order
at least three. Discrete Mathematics, 256(1):431–439, 2002.
Computing the k-restricted edge-connectivity of a graph
15
5. J. Chen, X. Huang, I. A. Kanj, and G. Xia. Strong computational lower bounds via
parameterized complexity. Journal of Computer and System Sciences, 72(8):1346–
1367, 2006.
6. R. Chitnis, M. Cygan, M. Hajiaghayi, M. Pilipczuk, and M. Pilipczuk. Designing
FPT algorithms for cut problems using randomized contractions. SIAM Journal
on Computing, 45(4):1171–1229, 2016.
7. M. Cygan, F. Fomin, B. M. Jansen, L. Kowalik, D. Lokshtanov, D. Marx,
M. Pilipczuk, and M. Pilipczuk. Open problems from School on Parameterized
Algorithms and Complexity, http://fptschool.mimuw.edu.pl/opl.pdf, 2014.
8. M. Cygan, F. V. Fomin, L. Kowalik, D. Lokshtanov, D. Marx, M. Pilipczuk,
M. Pilipczuk, and S. Saurabh. Parameterized Algorithms. Springer, 2015.
9. M.
Cygan,
L.
Kowalik,
and
M.
Pilipczuk.
Open
problems
from
Update
Meeting
on
Graph
Separation
Problems,
http://worker2013.mimuw.edu.pl/slides/update-opl.pdf, 2013.
10. M. Cygan, D. Lokshtanov, M. Pilipczuk, M. Pilipczuk, and S. Saurabh. Minimum
bisection is fixed parameter tractable. In Proc. of the 46th ACM Symposium on
Theory of Computing (STOC), pages 323–332, 2014.
11. R. Diestel. Graph Theory. Springer-Verlag, Berlin, 3rd edition, 2005.
12. R. G. Downey, V. Estivill-Castro, M. R. Fellows, E. Prieto, and F. A. Rosamond.
Cutting up is hard to do: the parameterized complexity of k-cut and related problems. Electronic Notes in Theoretical Computer Science, 78:209–222, 2003.
13. R. G. Downey and M. R. Fellows. Parameterized Complexity. Springer-Verlag,
1999.
14. M. E. Dyer and A. M. Frieze. On the complexity of partitioning graphs into
connected subgraphs. Discrete Aplied Mathematics, 10:139–153, 1985.
15. M. E. Dyer and A. M. Frieze. Planar 3DM is NP-complete. Journal of Algorithms,
7(2):174–184, 1986.
16. A.-H. Esfahanian and S. L. Hakimi. On computing a conditional edge-connectivity
of a graph. Information Processing Letters, 27(4):195–199, 1988.
17. J. Fàbrega and M. A. Fiol. Extraconnectivity of graphs with large girth. Discrete
Mathematics, 127:163–170, 1994.
18. J. Flum and M. Grohe. Parameterized Complexity Theory. Springer-Verlag, 2006.
19. M. R. Garey and D. S. Johnson. Computers and intractability. A guide to the
theory of NP-completeness. W. H. Freeman and Co., 1979.
20. M. R. Garey, D. S. Johnson, and L. J. Stockmeyer. Some simplified NP-complete
graph problems. Theoretical Computer Science, 1(3):237–267, 1976.
21. A. Holtkamp. Connectivity in Graphs and Digraphs. Maximizing vertex-, edgeand arc-connectivity with an emphasis on local connectivity properties. PhD thesis,
RWTH Aachen University, 2013.
22. A. Holtkamp, D. Meierling, and L. P. Montejano. k-restricted edge-connectivity in
triangle-free graphs. Discrete Applied Mathematics, 160(9):1345–1355, 2012.
23. K. Kawarabayashi and M. Thorup. The Minimum k-way Cut of Bounded Size is
Fixed-Parameter Tractable. In Proc. of the 52nd Annual Symposium on Foundations of Computer Science (FOCS), pages 160–169, 2011.
24. E. J. Kim, C. Paul, I. Sau, and D. M. Thilikos. Parameterized algorithms for minmax multiway cut and list digraph homomorphism. In Proc. of the 10th International Symposium on Parameterized and Exact Computation (IPEC), volume 43
of LIPIcs, pages 78–89, 2015.
25. D. Marx. Parameterized graph separation problems. Theoretical Computer Science,
351(3):394–406, 2006.
16
L. P. Montejano and I. Sau
26. D. Marx and M. Pilipczuk. Everything you always wanted to know about the
parameterized complexity of subgraph isomorphism (but were afraid to ask). In
Proc. of the 31st International Symposium on Theoretical Aspects of Computer
Science (STACS), volume 25 of LIPIcs, pages 542–553, 2014.
27. M. Naor, L. J. Schulman, and A. Srinivasan. Splitters and near-optimal derandomization. In Proc. of the 36th Annual Symposium on Foundations of Computer
Science (FOCS), pages 182–191, 1995.
28. R. Niedermeier. Invitation to Fixed-Parameter Algorithms. Oxford University
Press, 2006.
29. A. Rai, M. S. Ramanujan, and S. Saurabh. A parameterized algorithm for mixedcut. In Proc. of the 12th Latin American Symposium on Theoretical Informatics
(LATIN), volume 9644 of LNCS, pages 672–685, 2016.
30. M. Stoer and F. Wagner. A simple min-cut algorithm. Journal of the ACM,
44(4):585–591, 1997.
31. R. van Bevern, A. E. Feldmann, M. Sorge, and O. Suchý. On the parameterized complexity of computing graph bisections. In Proc. of the 39th International
Workshop on Graph-Theoretic Concepts in Computer Science (WG), volume 8165
of LNCS, pages 76–87, 2013.
32. J. Yuan and A. Liu. Sufficient conditions for λk -optimality in triangle-free graphs.
Discrete Mathematics, 310:981–987, 2010.
33. Z. Zhang and J. Yuan. A proof of an inequality concerning k-restricted edgeconnectivity. Discrete Mathematics, 304(1-3):128–134, 2005.
| 8cs.DS
|
AN EFFICIENT DEEP LEARNING HASHING NEURAL NETWORK FOR MOBILE VISUAL
SEARCH
Heng Qi, Wu Liu, Liang Liu
arXiv:1710.07750v1 [cs.CV] 21 Oct 2017
Beijing Key Laboratory of Intelligent Telecommunication Software and Multimedia,
Beijing University of Posts and Telecommunications, Beijing 100876, China
{qiheng, liuwu, liangliu}@bupt.edu.cn
ABSTRACT
Mobile visual search applications are emerging that enable
users to sense their surroundings with smart phones. However, because of the particular challenges of mobile visual
search, achieving a high recognition bitrate has becomes a
consistent target of previous related works. In this paper,
we propose a few-parameter, low-latency, and high-accuracy
deep hashing approach for constructing binary hash codes for
mobile visual search. First, we exploit the architecture of
the MobileNet model, which significantly decreases the latency of deep feature extraction by reducing the number of
model parameters while maintaining accuracy. Second, we
add a hash-like layer into MobileNet to train the model on labeled mobile visual data. Evaluations show that the proposed
system can exceed state-of-the-art accuracy performance in
terms of the MAP. More importantly, the memory consumption is much less than that of other deep learning models. The
proposed method requires only 13 MB of memory for the neural network and achieves a MAP of 97.80% on the mobile
location recognition dataset used for testing.
Index Terms— Mobile visual search, Supervised hashing, Binary code, Deep learning
1. INTRODUCTION
With the proliferation of mobile devices, it is becoming possible to use mobile perception functionalities (e.g., cameras,
GPS, and Wi-Fi) to perceive the surrounding environment [1].
Among such techniques, mobile visual search plays a key role
in mobile localization, mobile media search, and mobile social networking. However, rather than simply porting traditional visual search methods to mobile platforms, for mobile
visual search, one must face the challenges of a large auralvisual variance of queries, stringent memory and computation
constraints, network bandwidth limitations, and the desire for
an instantaneous search experience.
This work was partially supported by the CCF-Tencent Open Research
Fund (No. AGR20160113), the National Natural Science Foundation of
China (No. 61632008), and the Fundamental Research Funds for the Central
Universities (No. 2016RCGD32).
Input image
Convolutional layers
latent layer
loss
256
256
K
DK
DK
M kernel
Depthwise Convolutional Filters
1
1
N kernel
Pointwise Convolutional Filters
Fig. 1. The architecture of the proposed deep learning hashing
neural network for mobile visual search
Most research on mobile visual search has predominantly
focused on achieving high recognition bitrates [2, 3, 4]. Recently, an increasing number of researchers are attempting to
exploit feature signatures produced through hashing in mobile
visual search because of the good balance that can be achieved
among computation and memory requirements, training efficiency, quantization complexity, and search performance.
However, most of the existing hashing-based mobile visual
search methods attempt to compress existing classical handcrafted features into binary code. These methods all focus on
how to decrease the loss suffered during compression. Only
a few of them attempt to automatically learn effective binary
code features from a large-scale image dataset using a deep
neural network. There are two main reasons for this: 1) the
lack of effective deep learning hashing methods for mobile
visual search and 2) the high computational complexity of existing deep neural networks.
For traditional visual search, convolutional neural networks have become ubiquitous [5][6]. Studies have shown
that the deep features learned using such networks capture
rich image representations and enable better performance
than handcrafted features in visual classification [7][8][9]
[10], object detection [11][12], semantic segmentation [13],
and image retrieval [14][15]. The general trend in research on
deep learning methods is to construct deeper and more complicated networks to achieve higher accuracy [16][17][18].
In the field of mobile visual search, however, a deeper neural network model consumes more memory and time, which
cannot be easily supplied by mobile devices. To adapt deeper
neural networks for mobile devices, a new network architecture called MobileNet has recently been proposed by Google
[19]. In the MobileNet model, a standard convolutional layer
is decomposed into a depthwise convolutional layer and a
pointwise convolutional layer, thereby greatly reducing the
amount of calculation necessary and the model size. This
model works very well on image classification problems.
Inspired by the works presented in [1] and [19], to develop
a more effective and efficient mobile visual search system,
this paper proposes to combine a few-parameter, low-latency,
and high-accuracy network architecture with a hash function.
We incorporate the hash function as a latent layer between the
image representations and the classification outputs in MobileNet, which allows us not only to maintain the accuracy
of the visual search but also to adapt the model to the mobile environment. An overview of the system is illustrated in
Figure 1. The final binary hash codes can be learned by minimizing an objective function defined over the classification
error. Experimental results on a mobile location recognition
dataset show that our method achieves superior performance
compared with other hashing approaches.
2. ARCHITECTURE
In this section, we first describe the MobileNet network structure and the convolution layer structure, which is based on
depthwise separable filters and pointwise separable filters.
Then, we describe the hash layer in our convolutional neural
network and train a convolutional neural network model that
exploits semantic labels to automatically create binary codes.
2.1. MobileNet
The MobileNet model is built on depthwise separable convolutions. These convolutions factorize a standard convolution
into a depthwise convolution and a 1 × 1 convolution called
a pointwise convolution. The MobileNet architecture is defined in Table 1. In MobileNet, the depthwise convolution
involves the application of a single filter to each input channel. The pointwise convolution then involves the application
of a 1×1 convolution to combine the outputs of the depthwise
convolution. A standard convolutional layer takes as input a
DF × DF × M feature map F and produces a DF × DF × N
feature map G, where DF is the spatial width and height of
the square input feature map, M is the number of input chan-
Table 1. The MobileNet-based hashing network for mobile
visual search
Type / Stride
Conv / s2
Conv dw / s1
Conv / s1
Conv dw / s2
Conv / s1
Conv dw / s1
Conv / s1
Conv dw / s2
Conv / s1
Conv dw / s1
Conv / s1
Conv dw/ s2
Conv / s1
Conv dw / s1
Conv / s1
Conv dw / s2
Conv / s1
Conv dw/ s2
Conv / s1
Avg Pool / s1
FC / s1
Sigmoid
FC / s1
Softmax
Filter Shape
3 × 3 × 3 × 32
3 × 3 × 32 dw
1 × 1 × 32 × 64
3 × 3 × 64 dw
1 × 1 × 64 × 128
3 × 3 × 128 dw
1 × 1 × 128 × 128
3 × 3 × 128 dw
1 × 1 × 128 × 256
3 × 3 × 256 dw
1 × 1 × 256 × 256
3 × 3 × 256 dw
1 × 1 × 256 × 512
3 × 3 × 512 dw
1 × 1 × 512 × 512
3 × 3 × 512 dw
1 × 1 × 512 × 1024
3 × 3 × 1024 dw
1 × 1 × 1024 × 1024
Pool 7 × 7
1024 × 64
In place
1024 × 162
Classifier
Input Size
224 × 224 × 3
112 × 112 × 32
112 × 112 × 32
112 × 112 × 64
56 × 56 × 64
56 × 56 × 128
56 × 56 × 128
56 × 56 × 128
28 × 28 × 128
28 × 28 × 256
28 × 28 × 256
28 × 28 × 256
14 × 14 × 256
14 × 14 × 512
14 × 14 × 512
14 × 14 × 512
7 × 7 × 512
7 × 7 × 1024
7 × 7 × 1024
7 × 7 × 1024
1 × 1 × 64
1 × 1 × 64
1 × 1 × 162
1 × 1 × 162
nels (input depth), DG is the spatial width and height of the
square output feature map, and N is the number of output
channel.
A standard convolution has the following computational
cost:
DK · DK · M · N · DF · DF
(1)
The corresponding cost of depthwise separable convolutions
is as follows:
DK · DK · M · DF · DF + M · N · DF · DF
(2)
which is the sum of the costs of the depthwise and 1×1 pointwise convolutions. By expressing a standard convolution as
a two-step process of filtering and combining, we achieve the
following reduction in computational cost:
DK · DK · M · DF · DF + M · N · DF · DF
DK · DK · M · N · DF · DF
(3)
MobileNet uses 3 × 3 depthwise separable convolutions,
which require 8∼9 times less computation than standard convolutions with only a small reduction in accuracy.
Table 2 compares MobileNet with other popular models
in terms of accuracy as assessed on the ImageNet database,
number of multi-adds (million), and number of parameters
(million). MobileNet is nearly as accurate as VGG-16 [16]
while being 32 times smaller and 27 times less compute intensive. It is more accurate than GoogleNet [20] while being
smaller and requiring more than 2.5 times less computation.
Table 2. Comparisons between MobileNet and other popular
deep learning models [19].
Model
Accuracy
Multi-Adds (106 )
Parameters (106 )
AlexNet
Squeezenet
GoogleNet
VGG-16
MobileNet
57.2%
57.5%
69.8%
71.5%
70.6%
720
1700
1550
15300
569
60
1.25
6.8
138
4.2
MobileNet [19] is also more accurate than Squeezenet [21]
and AlexNet [14] with a nearly identical model size and less
computation.
2.2. Hash Function
Hash mapping, which is required for learning from images,
is based on the following principles. The hash codes should
respect the semantic similarity between image labels. Images
that share the same class labels should be mapped to similar binary codes. Let I = {In }N
n=1 denote N images, and
let Y = yn M ×N be their associated label vectors, where
M is the total number of class labels. An entry in yn has
a value of 1 if the image In belongs to the corresponding
class and is 0 otherwise. Our goal is to learn a mapping
F : I → {0, 1}K×N that maps the images to their k-bit
binary codes B = bn ∈ {0, 1}K×N while preserving the
semantic similarity relationships among the image data [22].
Our network is built on MobileNet. Each layer is followed
by a batchnorm layer and a nonlinear ReLU layer, with the
exception of the fully connected layer for classification. A final average pooling layer reduces the spatial resolution to 1
before the fully connected layer. To incorporate the learned
features into the binary codes, we add a latent layer H with K
units to the top of layer pool6, as illustrated in Figure 1. This
latent layer is fully connected to pool6 and uses sigmoid units
so that the activations are bounded between 0 and 1 [23]. Let
WH ∈ Rd×K denote the weights (i.e., the projection matrix)
in the latent layer. For a given image In with the feature vector a6 ∈ Rd in layer pool6, the activations of the units in H
6
H
can be computed as aH
+ eH ). Here, eH is the
n = σ(an · W
bias term, and σ(·) is the logistic sigmoid function, which is
defined as σ(z) = 1/(1 + exp(−z)), where z is a real value.
The binary encoding function is given by
bn = sign(σ(a6n · W H + eH ) − 0.5) = sign(aH
n − 0.5) (4)
where sign(v) = 1 if v > 0 and sign(v) = 0 otherwise.
sign(·) performs element-wise operations on a matrix or a
vector.
3. EXPERIMENTS
To evaluate the proposed method, we conducted experiments
on the mobile location recognition dataset presented in [24].
Table 3. The MAPs of different hashing methods for mobile
location recognition.
Method
VHB [4]
SSFS [24]
DLBH [25]
SSDH [22]
Our method
16-bit
59.80
78.26
96.66
MAP
32-bit
78.68
91.82
97.61
64-bit
19.36
20.22
87.15
92.43
97.80
The dataset contains 8,062 images, which were captured from
162 locations. We used the same experimental parameters
as in [24] and [1]. We implemented our approach using the
open-source CAFFE [26] package. We initialized the network
parameters by adopting the parameters of a MobileNet trained
on the 14 million images of the ImageNet [14] dataset. The
parameters in the hash layer were randomly initialized. The
learning rate was initialized as 0.01 and was decreased to 1/10
of its previous value after every 10,000 iterations. The entire training procedure terminated after 30,000 iterations. For
the learning of the network parameters, in conjunction with
backpropagation, we exploited the mini-batch stochastic gradient descent algorithm with a mini-batch size of 32 images to
minimize the classification error. Our model is a lightweight
modification of MobileNet and thus is easy to implement.
In the evaluation, we used the Mean Average Precision
(MAP) as the evaluation criterion. We ranked all of the images according to their Hamming distances from the query
image, selected the top k images from the ranked list as the
retrieval results, and computed the MAP on these retrieved
images. We set k to 100 in these experiments. We used the
class labels as the ground truth and adopted the common
settings for computing the MAP by examining whether the
retrieved images and the query shared common class labels.
We compared our method with Visual Hash Bits (VHB) [4]
and Space-Saliency Fingerprint Selection-based hash codes
(SSFS) [24], which are both traditional hashing methods for
mobile location recognition. The other two methods considered for comparison, Deep Learning of Binary Hash Codes
(DLBH) [25] and Supervised Semantics-Preserving Deep
Hashing (SSDH) [22], are both deep-learning-based hashing
methods.
Table 3 compares the performances on the dataset of the
different hashing methods for different hash code lengths.
From the results, we can see that the accuracies of the deep
learning hashing methods greatly exceed those of the traditional hashing methods, which demonstrates the power of
deep learning technology for binary code learning. Furthermore, because SSDH learns the feature representations and
binary codes simultaneously and imposes more constraints
for binary code learning, it achieves higher performance than
DLBH. Our method achieves the best results on this dataset
for all of the different hash code lengths. Especially when the
hash code length is relatively short, our method can achieve
extremely high accuracy. This can be attributed to the fact
that MobileNet networks enable the joint learning of representations and hash functions from images. Moreover, the
learned representations are more effective and stable than
those of AlexNet-based models such as DLBH and SSDH.
Finally, the sizes of the DLBH and SSDH models are greater
than 230 MB, whereas for MobileNet, the model size is only
13 MB. Thus, the memory demand for a mobile device is
reduced by a factor of approximately 18.
[10] Chuang Gan, Ming Lin, Yi Yang, Gerard de Melo, and Alexander G Hauptmann, “Concepts not alone: Exploring pairwise relationships for zero-shot video activity recognition,” in AAAI,
2016.
4. CONCLUSION
[12] R. Girshick, J. Donahue, T. Darrell, and J. Malik, “Rich feature
hierarchies for accurate object detection and semantic segmentation,” in CVPR, 2014, pp. 580–587.
We present an effective deep learning framework based on
MobileNet for creating hash-like binary codes for mobile visual search. In this framework, we incorporate the hash function as a latent layer between the feature layer and the output layer in the MobileNet network. By optimizing an objective function defined over the classification error, our method
jointly learns the binary codes, features, and classification results. To evaluate the performance of the proposed network,
we applied it to a mobile location recognition dataset. The
experimental results demonstrate that it can achieve a MAP
improvement of 5.5% compared with state-of-the-art methods. The model requires only 13 MB of memory.
5. REFERENCES
[1] W. Liu, H. Ma, H. Qi, D. Zhao, and Z. Chen, “Deep learning hashing for mobile visual search,” EURASIP J. Image and
Video Processing, vol. 2017, pp. 17, 2017.
[2] W. Liu, T. Mei, and Y.D. Zhang, “Instant mobile video search
with layered audio-video indexing and progressive transmission,” IEEE Trans. on Multimedia, vol. 16, no. 8, pp. 2242–
2255, 2014.
[3] V. Chandrasekhar, G. Takacs, and D. Chen, “Compressed histogram of gradients: A low-bitrate descriptor,” IJCV, vol. 96,
no. 3, pp. 384–399, 2012.
[4] J. He, J. Feng, and X. Liu, “Mobile product search with bag of
hash bits and boundary reranking,” in CVPR, 2012, pp. 3005–
3012.
[5] O. Russakovsky, J. Deng, H. Su, and J. Krause, “Imagenet
large scale visual recognition challenge,” IJCV, vol. 115, no.
3, pp. 211–252, 2015.
[6] W. Liu, T. Mei, Y.D. Zhang, C. Che, and J.B. Luo, “Multitask deep visual-semantic embedding for video thumbnail selection,” in CVPR, 2015, pp. 3707–3715.
[7] R. Xia, Y. Pan, and H. Lai, “Supervised hashing for image
retrieval via image representation learning,” in AAAI, 2014,
pp. 2156–2162.
[8] C. Gan, T. Yao, K.Y. Yang, Y. Yang, and T. Mei, “You lead, we
exceed: Labor-free video concept learning by jointly exploiting web videos and images,” in IEEE Computer Vision and
Pattern Recognition, 2016, pp. 923–932.
[9] C. Gan, Y. Yang, L.C. Zhu, D.L. Zhao, and Y.T. Zhuang, “Recognizing an action using its name: A knowledge-based approach,” International Journal of Computer Vision, vol. 120,
no. 1, pp. 61–77, 2016.
[11] D. Ciresan, U. Meier, and J. Schmidhuber, “Multi-column deep
neural networks for image classification,” in CVPR, 2012, pp.
3642–3649.
[13] P. Sermanet, D. Eigen, and X. Zhang, “Overfeat: Integrated
recognition, localization and detection using convolutional networks,” CoRR, vol. abs/1312.6229, 2013.
[14] A. Krizhevsky, I. Sutskever, and G. Hinton, “Imagenet classification with deep convolutional neural networks,” in NIPS,
2012, pp. 1106–1114.
[15] L.Y. Chu, S.Q. Jiang, S.H. Wang, Y.Y. Zhang, and Q.M.
Huang, “Robust spatial consistency graph model for partial
duplicate image retrieval,” IEEE Trans. on Multimedia, vol.
15, no. 8, pp. 1982–1996, 2013.
[16] K. Simonyan and A. Zisserman, “Very deep convolutional
networks for large-scale image recognition,” CoRR, vol.
abs/1409.1556, 2014.
[17] C. Szegedy, V. Vanhoucke, and S. Ioffe, “Rethinking the
inception architecture for computer vision,” CoRR, vol.
abs/1512.00567, 2015.
[18] C. Gan, N.Y. Wang, Y. Yang, D.Y. Yeung, and A.G. Hauptmann, “Devnet: A deep event network for multimedia event
detection and evidence recounting,” in CVPR, 2015, pp. 2568–
2577.
[19] A. Howard, M. Zhu, B. Chen, and D. Kalenichenko, “Mobilenets: Efficient convolutional neural networks for mobile
vision applications,” CoRR, vol. abs/1704.04861, 2017.
[20] C. Szegedy, W. Liu, and Y. Jia, “Going deeper with convolutions,” in CVPR, 2015, pp. 1–9.
[21] F. N. Iandola, M. Moskewicz, and K. Ashraf, “Squeezenet:
Alexnet-level accuracy with 50x fewer parameters and <1mb
model size,” CoRR, vol. abs/1602.07360, 2016.
[22] H. Yang, K. Lin, and C. Chen, “Supervised learning of
semantics-preserving hashing via deep neural networks for
large-scale image search,” CoRR, vol. abs/1507.00101, 2015.
[23] X. Glorot, A. Bordes, and Y. Bengio, “Deep sparse rectifier
neural networks,” in AIS, 2011, pp. 315–323.
[24] H. Wang, D. Zhao, H.D. Ma, and H. Xu, “SSFS: A spacesaliency fingerprint selection framework for crowdsourcing
based mobile location recognition,” in AMIP, 2016, pp. 650–
659.
[25] K. Lin, H. Yang, J. Hsiao, and C.Chen, “Deep learning of binary hash codes for fast image retrieval,” in CVPR Workshops,
2015, pp. 27–35.
[26] Y. Jia, E. Shelhamer, and J. Donahue, “Caffe: Convolutional
architecture for fast feature embedding,” in ACM MM, 2014,
pp. 675–678.
| 1cs.CV
|
Boltzmann machines for time-series
Takayuki Osogami
arXiv:1708.06004v2 [cs.NE] 22 Sep 2017
IBM Research - Tokyo
[email protected]
version 1.0.1
Abstract
We review Boltzmann machines extended for time-series. These models often have recurrent
structure, and back propagration through time (BPTT) is used to learn their parameters. The perstep computational complexity of BPTT in online learning, however, grows linearly with respect
to the length of preceding time-series (i.e., learning rule is not local in time), which limits the
applicability of BPTT in online learning. We then review dynamic Boltzmann machines (DyBMs),
whose learning rule is local in time. DyBM’s learning rule relates to spike-timing dependent
plasticity (STDP), which has been postulated and experimentally confirmed for biological neural
networks.
1
Introduction
The Boltzmann machine is a stochastic model for representing probability distributions over binary
patterns [28]. In this paper, we review Boltzmann machines that have been studied as stochastic
(generative) models of time-series. Such Boltzmann machines define probability distributions over
time-series of binary patterns. They can also be modified to deal with time-series of real-valued
patterns, similar to Boltzmann machines modified for real-valued patterns (e.g., Gaussian Boltzmann
machines; see Section 6.3 from [28]). We will follow the probabilistic representations of [28] for intuitive
interpretations in terms of probabilities.
In Section 3, we start with a Conditional Restricted Boltzmann Machine (CRBM) [40], which is
a conditional Boltzmann machine (Section 4 from [28]) that gives conditional probability of the next
pattern given a fixed number of preceding patterns. A limitation of a CRBM is that it can take into
account only the dependency within a fixed horizon, and as we increase the length of this horizon, the
complexity of learning grows accordingly.
To overcome this limitation of CRBMs, researchers have proposed Boltzmann machines having
recurrent structures, which we review in Section 4. These include spiking Boltzmann machines [12],
temporal restricted Boltzmann machines (TRBMs) [37], recurrent temporal restricted Boltzmann machines (RTRBMs) [38], and extensions of those models. A standard approach to learning those models
having recurrent structures is back propagation through time (BPTT).
However, BPTT is undesirable when we learn time-series in an online manner, where we update the
parameters of a model every time a new pattern arrives. Such online learning is needed when we want
to quickly adapt to a changing environment or when we do not have sufficient memory to store the
time-series. Unfortunately, the per-step computational complexity of BPTT in online learning grows
linearly with respect to the length of preceding time-series. This computational complexity limits the
applicability of BPTT to online learning.
In Section 5, we review the dynamic Boltzmann machine (DyBM) [32, 31] and its extensions. The
DyBM’s per-step computational complexity in online learning is independent of the length of preceding
1
time-series. We discuss how the learning rule of the DyBM relates to spike-timing dependent plasticity
(STDP), which has been postulated and experimentally confirmed for biological neural networks.
This survey paper is based on a personal note prepared for the third of the four parts of a tutorial
given at the 26th International Joint Conference on Artificial Intelligence (IJCAI-17) held in Melbourne, Australia on August 21, 2017. See a tutorial webpage1 for information about the tutorial. A
survey corresponding to the first part of the tutorial (Boltzmann machines and energy-based models)
can be found in [28]. We follow the definitions and notations used in [28].
2
Learning energy-based models for time-series
Consider a possibly multi-dimensional time-series:
x ≡ (x[t] )Tt=0 ,
(1)
where x[t] denotes the binary pattern (vector) at time t. We will use x[s,t] to denote the time-series of
the patterns from time s to t.
A goal of learning time-series is to maximize the log-likelihood of a given time-series x (or a
collection of multiple time-series) with respect to the distribution Pθ (·) defined by a model under
consideration, where we use θ to denote the set of the parameters of the model:
f (θ) ≡ log Pθ (x) =
T
X
log Pθ (x[t] | x[0,t−1] ),
(2)
t=0
where Pθ (x[t] | x[0,t−1] ) denotes the conditional probability that the pattern at time t is x[t] given that
the patterns up to time t − 1 is x[0,t−1] . Here, Pθ (x[0] | x[0,−1] ) denotes the probability that the pattern
at time 0 is x[0] , where x[0,−1] should be interpreted as an empty history.
We study models where the probability is represented with energy Eθ (·) as follows:
X
Pθ (x) =
Pθ (x, h̃),
(3)
h̃
where
exp − Eθ (x, h)
,
Pθ (x, h) = X X
exp − Eθ (x̃, h̃)
x̃
(4)
h̃
the summation with respect to x̃ is over all of the possible binary time-series of length T , and the
summation with respect to h̃ is over all of the possible hidden values.
The gradient of f (θ) can then be represented as follows (see (73)):
∇f (θ) = −Etarget [Eθ [∇Eθ (X, H) | X]] + Eθ [∇Eθ (X, H)] ,
(5)
where X represents the random time-series, H represents the random hidden values, Eθ denotes the
expectation with respect to the model distribution Pθ , and Etarget denotes the expectation with respect
to the target distribution, which in our case is the empirical distribution of time-series. When a single
time-series x is given as the target, (5) is reduced to
∇f (θ) = −Eθ [∇Eθ (x, H)] + Eθ [∇Eθ (X, H)] ,
1 https://researcher.watson.ibm.com/researcher/view
group.php?id=7834
2
(6)
In other words, f (θ) can be maximized by maximizing the sum of
ft (θ) ≡ log Pθ (x[t] | x[0,t−1] ),
(7)
where
Pθ (x[t] | x[0,t−1] ) =
X
Pθ (x[t] , h̃ | x[0,t−1] )
(8)
h̃
exp − Eθ (x[t] , h | x[0,t−1] )
,
Pθ (x[t] , h | x[0,t−1] ) = X X
exp − Eθ (x̃[t] , h̃ | x[0,t−1] )
x̃[t]
(9)
h̃
and Eθ (x[t] , h | x[0,t−1] ) is the conditional energy of (x[t] , h) given x[0,t−1] .
The gradient of ft (θ) is given analogously to ∇f (θ):
h h
ii
h
i
∇ft (θ) = −Etarget Eθ ∇Eθ (X [t] , H | X [t] , x[0,t−1] ) + Eθ ∇Eθ (X [t] , H | x[0,t−1] ) .
(10)
When the target is a single time-series x, we have
h
i
h
i
∇ft (θ) = −Eθ ∇Eθ (x[t] , H | x[0,t−1] ) + Eθ ∇Eθ (X [t] , H | x[0,t−1] ) .
(11)
3
Non-recurrent Boltzmann machines for time-series
By (2), any model that can represent the conditional probability Pθ (x[t] | x[0,t−1] ) can be used for
time-series. In this section, we start with a Boltzmann machine that can be used to model a D-th
order Markov model for an arbitrarily determined D. In D-th order Markov models, the conditional
probability can be represented as
Pθ (x[t] | x[0,t−1] ) = Pθ (x[t] | x[t−D,t−1] ).
3.1
(12)
Conditional restricted Boltzmann machines
Figure 1a shows a particularly structured Boltzmann machine called Conditional Restricted Boltzmann
Machine (CRBM) [40]. A CRBM represents the conditional probability on the right-hand side of (12).
In the figure, we set D = 2.
The CRBM consists of D + 1 layers of visible units and a layer of hidden units. The units within
each layer have no connections, but units between different layers may be connected to each other.
Each visible layer corresponds to a pattern at a time s ∈ [t − D, t].
The CRBM is a conditional Boltzmann machine shown in Figure 2c from [28] but with a particular
structure to represent time-series. The visible layers corresponding to x[t−D,t−1] are the input, and
the visible layer corresponding to x[t] is the output. The parameters θ of the CRBM are independent
of t.
More formally, the energy of a CRBM is given by
D
X
Eθ x[t] , h | x[t−D,t−1] = −(bV )> x[t] − (bH )> h − h> WHV x[t] −
(x[t−d] )> W[d] x[t] ,
(13)
d=1
where x[t] is output, x[t−D,t−1] is input, and h is hidden.
We can then represent the conditional probability as follows (see (14) from [28]):
X
Pθ (x[t] | x[t−D,t−1] ) =
Pθ (x[t] , h̃ | x[t−D,t−1] ),
h̃
3
(14)
...
Hidden
...
...
...
𝑡−2
𝑡−1
𝑡−2
𝑡
(a) Single hidden layer
Visible
Output
Input
Output
Input
...
...
...
Visible
...
...
Hidden
𝑡−1
𝑡
(b) Multiple hidden layers
Figure 1: Conditional restricted Boltzmann machines.
where
exp − Eθ x[t] , h̃ | x[t−D,t−1]
Pθ (x[t] , h̃ | x[t−D,t−1] ) = X
,
exp − Eθ x̃[t] , h̃ | x[t−D,t−1]
(15)
x̃[t]
and the summation with respect to h̃ is over all of the possible binary hidden patterns, and the
summation with respect to x̃[t] is defined analogously.
One can then learn the parameters θ = (bV , hh , WHV , W[1] , . . . , WD ) of the model by following a
gradient-based method in Section 4 from [28].
3.2
Extensions of conditional restricted Boltzmann machines
The CRBM has been extended in various ways. Taylor et al. study a CRBM with multiple layers of
hidden units [40] (see Figure 1b). Memisevic and Hinton study a CRBM extended with three-way
interactions (i.e., a higher order Boltzmann machine), which they refer to as a gated CRBM [24].
Specifically, the energy of the gated CRBM involves
X
−
wi,j,k xi yj hk ,
(16)
i,j,k
where x denotes input values, y denotes output values, and h denotes hidden values. A drawback of
the gated CRBM is its increased number of parameters due to the three-way interactions. Taylor and
Hinton study a factored CRBM, where the three-way interaction is represented with a reduced number
of parameters as follows [39]:
XX
y
v
h
−
wi,f
wj,f
wk,f
xi yj hk ,
(17)
f
i,j,k
where the summation with respect to f is over a set of factors under consideration.
4
that these forward connections are not required f
e, but only for the purposes of extrapolating into
0.16
0.14
...
...
...
...
Hidden
0.12
r(t)
0.1
0.08
0.06
0.04
...
...
...
...
Input
Visible
Output
0.02
0
𝑡−2
𝑡−1
𝑡
0
5
10
15
20
25
30
Time
(a) Structure of a spiking Boltzmann machine
(b) r(·) used in [12]
Figure 3: The form of the temporal k
1
Figure 2: A spiking Boltzmann machine studied in [12]. In (b), we use the figure in the version available
at http://www.cs.toronto.edu/∼fritz/absps/nips00-ab.pdf.
4
4.1
Boltzmann machines for time-series with recurrent structures
Spiking Boltzmann machines
A spiking Boltzmann machine studied in [12] can be shown to be essentially equivalent to the Boltzmann machine illustrated in Figure 2. This Boltzmann machine consists of input units, output units,
and hidden units. The input units represent historical values of visible units and hidden units. Although hidden units are random and cannot be simply given as input, Hinton and Brown make the
approximation of using sampled values H (−∞,t−1] (ω) as the input hidden units [12].
Specifically, given the visible values x[<t] ≡ x(−∞,t−1] and sampled hidden values H [<t] (ω) up to
time t − 1, the energy with the visible values x[t] and hidden values h[t] at time t can be represented
as follows:
Eθ (x[t] , h[t] | x[<t] , H [<t] (ω)) = −(bV )> x[t] − (bH )> h[t] − (h[t] )> r(τ ) WHV x[t]
∞
∞
X
X
−
(H [t−τ ] (ω))> r(τ ) WHH h[t] −
(x[t−τ ] )> r(τ ) WVH h[t]
−
τ =1
∞
X
∞
X
τ =1
τ =1
τ =1
(H [t−τ ] (ω))> r(τ ) WHV x[t] −
(x[t−τ ] )> r(τ ) WVV x[t] , (18)
where r(·) is an arbitrarily chosen function and is not the target of learning. Namely, the Boltzmann
machine has an infinite number of units but can be characterized by a finite number of parameters
θ ≡ (bV , bH , WVV , WVH , WHV , WHH ). Figure 2(b) shows the specific r(·) used in [12]2 .
Notice that the Boltzmann machine in Figure 2 can be seen as a restricted Boltzmann machine
(RBM) whose bias and weight can depend on x[<t] and H [<t] (ω), because (18) can be represented as
Eθ (x[t] , h[t] | x[<t] , H[<t] (ω)) = −bH (t, ω) h[t] − bV (t, ω) x[t] − h[t] W x[t] ,
(19)
2 Although it is not clear from the descriptions in [12], the labels in the horizontal axis should probably be shifted by
one, so that r(0) = 0, r(1) ≈ 0.1, and so on.
5
where bH (t, ω) is the time-varying bias for hidden units, bV (t, ω) is the time-varying bias for visible
units, and W is the weight between visible units and hidden units:
H
H
b (t, ω) ≡ b +
bV (t, ω) ≡ bV +
∞
X
(H
τ =1
∞
X
>
(ω)) r(τ ) W
HH
+
(H [t−τ ] (ω))> r(τ ) WHV +
τ =1
HV
W ≡ r(0) W
[t−τ ]
∞
X
(x[t−τ ] )> r(τ ) WVH
τ =1
∞
X
(x[t−τ ] )> r(τ ) WVV
(20)
(21)
τ =1
(22)
.
We can then represent the conditional probability as follows:
X
Pθ (x[t] | x[<t] , H (−∞,t−1 (ω)) =
Pθ x[t] , h̃[t] | x[<t] , H (−∞,t−1] (ω)
(23)
h̃[t]
where
exp − Eθ x[t] , h̃[t] | x[<t] , H (−∞,t−1] (ω)
Pθ (x[t] , h̃[t] | x[<t] , H [<t] (ω)) = X
.
exp − Eθ x̃[t] , h̃[t] | x[<t] , H [<t] (ω)
(24)
x̃[t]
We now discuss the choice of r(0) = 0, which appears to be the case in Figure 2(b). In this case,
the energy is reduced to
Eθ (x[t] , h[t] | x[<t] , H [<t] (ω)) = −(bH (t, ω))> h[t] − (bV (t, ω))> x[t] .
(25)
Because there are no connections between visible units and hidden units at time t, the hidden values
at t do not affect the distribution of the visible values at t. The only role of the hidden units is that
the sampled hidden values are used to update the time-varying bias, bV (s, ω) and bH (s, ω) for s > t.
A problem is that there is no mechanism that allows us to learn appropriate values of WVH and
WHH until we observe succeeding visible values. Namely, the hidden values h[t] are sampled with the
dependency on WVH and WHH , but whether the sampled hidden values are good or not can only be
known when those hidden values are used as input. This helps us to learn appropriate values of WHV ,
but not WVH or WHH . See [30] for further discussion.
4.2
Temporal restricted Boltzmann machines
Sutskever and Hinton study a model related to a CRBM, which they refer to as a temporal restricted
Boltzmann machine (TRBM) [37]. While a CRBM defines the conditional distribution of the (visible
and hidden) values at time t given only the visible values from time t − D to t − 1, a TRBM defines the
corresponding conditional probability given both the visible values and the hidden values from time
t − D to t − 1. See Figure 3a. Similar to the CRBM, the parameters θ of the TRBM do not depend
on time t.
Unlike the CRBM, the TRBM is not a conditional RBM. This is because the TRBM with a single
parameter is used for every t, and the distribution of hidden values is shared among those TRBM at
varying t. In particular, hidden values of the TRBM can depend on the future visible values.
Because this dependency makes learning and inference hard, it is ignored in [37]. Namely, the
values at each time t is conditionally independent of the values after time t given the values at and
before time t. In particular, the distribution of the hidden values at time t is completely determined
by the visible values up to time t. The distribution of the hidden values before time t can thus be
considered as input when we use the TRBM to define the conditional distribution of the values at time
t (see Figure 3b). For each sampled values of hidden units, TRBM in 3b is a CRBM.
6
...
...
...
...
𝑡−1
...
...
...
𝑡−2
...
...
...
Visible
Output
Input
...
...
Hidden
𝑡−2
(a) TRBM
Visible
Output
Input
𝑡
Hidden
𝑡−1
𝑡
(b) TRBM with approximations in [37]
Figure 3: Temporal restricted Boltzmann machines. In (b), the gray circles indicate that expected
values are used for the input hidden units.
Furthermore, in [37], the expected values (see Section 5.4 from [28]) are used for the hidden values.
Then the input hidden units in Figure 3b takes real values in [0, 1] that are completely determined by
the visible values before time t.
More formally, with the approximations in [37], the TRBM with parameter θ defines the probability
distribution over the time-series of visible and hidden values as follows:
Pθ (x) =
T X
Y
Pθ (x[t] , h̃[t] | x[t−D,t−1] , r[t−D,t−1] ),
(26)
t=0 h̃[t]
where
exp − Eθ (x[t] , h̃[t] | x[t−D,t−1] , r[t−D,t−1] )
Pθ (x[t] , h̃[t] | x[t−D,t−1] , r[t−D,t−1] ) = X
exp − Eθ (x̃[t] , h̃[t] | x[t−D,t−1] , r[t−D,t−1] )
(27)
x̃[t]
is the conditional distribution defined by the Boltzmann machine shown in Figure 3b, where r[t−D,t−1]
are expected hidden values. Specifically, (27) is used to compute
r[t] = Eθ [H [t] | x[0,t] ],
(28)
which is subsequently used with (27) for t ← t + 1, where the expectation in (28) is with respect to
Pθ (x[t] , h[t] | x[t−D,t−1] , r[t−D,t−1] )
.
Pθ (h[t] | x[0,t−1] ) = X
Pθ (x̃[t] , h[t] | x[t−D,t−1] , r[t−D,t−1] )
(29)
x̃[t]
Notice that r[t] can be computed from x[0,t] in a deterministic manner with dependency on θ.
However, this dependency on θ is ignored in learning TRBMs.
4.3
Recurrent temporal restricted Boltzmann machines
To overcome the intractability of the TRBM without approximations, Sutskever et al. study a refined
model of TRBM, which they refer to as a recurrent temporal restricted Boltzmann machine (RTRBM)
7
...
...
Hidden
...
...
Input
Visible
Output
𝑡−1
𝑡
Figure 4: A recurrent temporal restricted Boltzmann machine.
[38]. The RTRBM simplifies the TRBM by removing connections between visible layers and connections between hidden layers that are separated by more than one lag. This means that the (visible and
hidden) values at time t are conditionally independent of the the visible values before time t and the
hidden values before time t − 1 given the hidden values at time t − 1. Similar to the approximation
made for the TRBM in Figure 3b, the RTRBM uses the expected values for the hidden values at time
t − 1 but defines the conditional distribution of the (visible and hidden) values at time t over their
binary values. See Figure 4.
More formally, let r[t−1] denote the expected values of the hidden units at time t − 1:
r[t−1] ≡ Eθ H[t−1] | x[0,t−1] ,
(30)
where H[t−1] is the random vector representing the hidden values at time t, and Eθ [· | x[0,t−1] ] represents
the conditional expectation given the visible values up to time t − 1. The probability distribution of
the values at time t is then given by
exp − Eθ (x[t] , h[t] | r[t−1] )
,
(31)
Pθ (x[t] , h[t] | r[t−1] ) = X
exp − Eθ (x̃[t] , h̃[t] | r[t−1] )
x̃,h̃
where the conditional energy is given by
Eθ (x[t] , h[t] | r[t−1] ) ≡ −(bV )> x[t] − (bH )> h[t] − (r[t−1] )> U h[t] − (x[t] )> W h[t]
(32)
for parameters
θ ≡ (bV , bH , W, U),
(33)
where bV is the bias for visible units, bH is the bias for hidden units, W is the weight matrix between
visible units and hidden units, and U is the weight matrix between previous expected value of hidden
units (at t − 1) and hidden units (at t).
8
4.3.1
Inference
The marginal conditional distribution of visible values at time t − 1 can then be represented as follows:
exp − Fθ (x[t] | r[t−1] )
,
Pθ (x[t] | r[t−1] ) = X
(34)
exp − Fθ (x̃[t] | r[t−1] )
x̃[t]
where the conditional free-energy is given by
Fθ (x[t] | r[t−1] ) ≡ − log
X
exp − Eθ (x[t] , h̃[t] | r[t−1] ) .
(35)
h̃[t]
Once the visible values at time t is given, the conditional probability distribution of the hidden
values at time t can be represented as follows:
exp − Eθ (h[t] | r[t−1] , x[t] )
,
Pθ (h[t] | r[t−1] , x[t] ) = X
(36)
exp − Eθ (h̃[t] | r[t−1] , x[t] )
h̃[t]
where the (bV )> x[t] term is canceled out between the numerator and the denominator, and the conditional energy is given by
>
Eθ (h[t] | r[t−1] , x[t] ) ≡ − bH + U> r[t−1] + W> x[t] h[t] .
(37)
By Corollary 1 from [28], the hidden values at time t are conditionally independent of each other given
r[t−1] and x[t] :
Y
[t]
Pθ (h[t] | r[t−1] , x[t] ) =
Pθ (hi | r[t−1] , x[t] ),
(38)
i
where
[t]
[t]
Pθ (hi
[t−1]
[t]
[t]
exp − bi hi
,
(39)
b[t] ≡ bH + U> r[t−1] + W> x[t]
(40)
b[0] ≡ binit + W> x[0] .
(41)
|r
,x ) =
[t]
1 + exp − bi
[t]
where bi is the i-th element of
for t ≥ 1, and
where we now follow [38] and allow the hidden units at time 0 to have own bias binit that can differ
from bH . The expected values are thus given by
r[t] =
1
,
1 + exp b[t]
where the operations are defined elementwise.
9
(42)
...
𝐡𝑡
...
...
...
...
...
...
...
...
...
...
𝑡=1
...
...
...
...
𝑡=0
𝐫𝑡
𝐱𝑡
𝑡=𝑇
Figure 5: A recurrent temporal restricted Boltzmann machine unfolded through time, where T = 4.
4.3.2
Learning
The parameters of an RTRBM can be trained through back propagation through time, analogous to
recurrent neural networks, but with contrastive divergence. To understand how we can train RTRBMs,
Figure 5 shows an RTRBM unfolded through time. Recall that the expected values of hidden units
are deterministically updated from r[t−1] to r[t] according to (40)-(42). Hence, r[t] can be understood
as hidden values of a recurrent neural network (RNN) [34]. An RTRBM can then be seen as an RNN
but gives an RBM as an output instead of real values, which would be given as an output from the
standard RNN. We will derive the learning rule of the RTRBM, closely following [25] but using our
notations.
When an RTRBM is unfolded through time, its energy can be represented as follows:
Eθ (x, h) = −
T
T
T
T
X
X
X
X
(bV )> x[t] − (binit )> h[0] −
(bH )> h[t] −
(x[t] )> W h[t] −
(r[t−1] )> U h[t] .
t=0
t=1
t=0
t=1
(43)
By (5), we can maximize the log-likelihood of a given time-series x with a gradient-based approach.
What we need in (5) is the gradient of the energy with respect to the parameter. A caveat is that the
energy in (43) depends on r[·] , which in turn depends on θ in a recursive manner. Also, expectation
with respect to Pt heta in (5) needs to be computed with approximation such as contrastive divergence
(see Section 5.2).
We first study the last term of (43). Let
Qs ≡
T
X
(r[t−1] )> U h[t]
t=s
[s−1] >
= (r
) U h[s] + Qs+1 ,
(44)
(45)
for s ∈ [1, T ], where QT +1 ≡ 0, so that Q ≡ Q1 is the last term of (43). Taking the partial derivative
10
[s−1]
with respect to ri
, we obtain
∂Qs
[s−1]
∂ri
=
∂
(r[s−1] )> U h[s] +
[s−1]
X ∂rj[s] ∂Qs+1
∂ri
[s−1]
j
= Ui,: h[s] +
X
[s]
∂ri
[s]
rj (1 − rj ) ui,j
∂Qs+1
j
[s]
∂rj
[s]
(46)
(47)
,
∂rj
where Ui,: denotes the i-th row of U, ui,j denotes the (j, i)-th element of U, and the last equality
follows from (40)-(42). In vector-matrix notations, we can write
∇r[s−1] Qs = U h[s] + r[s] · (1 − r[s] ) · ∇r[s] Qs+1 ,
(48)
where · denotes elementwise multiplication. Because Qs is not a function of r[0] , . . . , r[s−2] , we have
∇r[s−1] Q = ∇r[s−1] Qs .
(49)
Therefore, the partial derivative of Q with respect to r[s−1] is given recursively as follows:
∇r[s−1] Q = U h[s] + r[s] · (1 − r[s] ) · ∇r[s] Q
(50)
for s = 1, . . . , T and
∇r[T ] Q = 0.
(51)
We now take the derivative of Q with respect to the parameters in θ, starting with U:
T
[t]
X X ∂r
∂Q
∂Q
dQ
k
+
=
[t]
dui,j
∂u
∂u
i,j
i,j
∂rk
t=0 k
=
T
X
[t]
[t]
[t−1]
rj (1 − rj ) ri
t=1
∂Q
[t]
∂rj
(52)
+
T
X
[t−1]
ri
[t]
hj ,
(53)
t=1
where the last equality follows from (40)-(42). In vector-matrix notations, we can write
∇U Q =
T
X
>
r[t−1] r[t] · (1 − r[t] ) · ∇r[t] Q + h[t] ,
(54)
t=1
where ∇r[t] Q is given by (50)-(51).
The gradient of Q with respect to other parameters can be derived as follows:
∇W Q =
T
X
>
x[t] r[t] · (1 − r[t] ) · ∇r[t] Q
(55)
r[t] · (1 − r[t] ) · ∇r[t] Q
(56)
· (1 − r[0] ) · ∇r[0] Q
(57)
t=0
∇bH Q =
T
X
t=1
[0]
∇binit Q = r
∇bV Q = 0
(58)
11
The gradients of Q can be used to show the following gradients of the energy:
∇U Eθ (x, h) = −
T
X
>
r[t−1] r[t] · (1 − r[t] ) · ∇r[t] Q + h[t]
(59)
t=1
∇W Eθ (x, h) = −
T
X
x[t] (h[t] )> −
t=0
∇bH Eθ (x, h) = −
T
X
t=1
[0]
∇binit Eθ (x, h) = −h
∇bV Eθ (x, h) = −
T
X
T
X
>
x[t] r[t] · (1 − r[t] ) · ∇r[t] Q
(60)
t=0
h[t] −
T
X
r[t] · (1 − r[t] ) · ∇r[t] Q
(61)
t=1
− r[0] · (1 − r[0] ) · ∇r[0] Q
(62)
x[t] ,
(63)
t=0
where ∇r[t] Q is given by (50)-(51). The gradient of the log-likelihood of the given time-series x now
follows from (72) in [28].
4.3.3
Extensions
The RTRBM has been extended in various ways. Mittleman et al. study a structured RTRBM, where
units are partitioned into blocks, and only the connections between particular blocks are allowed [25].
Lyu et al. replaces the RNN of RTRBM with the one with Long Short-Term Memory (LSTM) [22].
Schrauwen and Buesing replaces the RNN of RTRBM with an echo state network [36].
An RNN-RBM slightly generalizes RTRBM by relaxing the constraint of the RTRBM that r[t] must
be the expected value of h[t] [7]. Namely, an RNN-RBM is an RNN but gives an RBM as an output,
where the RNN and RBM do not share parameters, while an RTRBM shares parameters between an
RNN and an RBM.
5
Dynamic Boltzmann machines
BPTT is not desirable for online learning, where we update θ every time a new pattern x[t] is observed.
The per-step computational complexity of BPTT in online learning grows linearly with the length of
the preceding time-series. Such online learning, however, is needed for example when we cannot store
all observed patterns in memory or when we want to adapt to changing environment.
The dynamic Boltzmann machine (DyBM) is proposed as a time-series model that allows efficient
online learning [31, 32]. The per-step computational complexity of the learning rule of a DyBM is
independent of the length of the preceding time-series. In Section 5.1, we start by reviewing the
DyBM introduced in [31, 32] with relation of its learning rule to spike-timing dependent plasticity
(STDP).
In Section 5.2, we study the relaxation of some of the constraints that the DyBM has required in
[31, 32] in a way that it becomes more suitable for inference and learning [27]. The primary purpose of
these constraints in [31, 32] was to mimic a particular form of STDP. The relaxed DyBM generalizes
the original DyBM and allows us to interpret it as a form of logistic regression for time-series data.
In Section 5.3, we review DyBMs dealing with real-valued time-series [27, 8]. These DyBMs are
analogous to how Gaussian Boltzmann machines [23, 43, 13] deal with real-valued patterns as opposed
to Boltzmann machines [2, 14] for binary values. The Gaussian DyBM can be related to a vector
autoregressive (VAR) model [21]. Specifically, we show that a special case of the Gaussian DyBM is a
VAR model having additional variables that capture long term dependency of time-series. These additional variables correspond to DyBM’s eligibility traces, which represent how recently and frequently
spikes arrived from a neuron to another. We also review an extension of the Gaussian DyBM to deal
with time-series patterns in continuous space [17].
12
i
Wij[d]
j
Figure 6: A dynamic Boltzmann machine unfolded through time (Figure 1(c) from [31]).
Some of the models and learning algorithms in this section have been implemented in Python or
Java and open-sourced at https://github.com/ibm-research-tokyo/dybm.
5.1
Dynamic Boltzmann machines for binary-valued time-series
5.1.1
Finite dynamic Boltzmann machines3
The DyBM in [31, 32] is defined as a limit of a sequence of Boltzmann machines (DyBM-T ) consisting
of T layers as T tends to infinity (see Figure 6). Formally, the DyBM-T is defined as the CRBM (see
Section 3.1) having T layers of N visible units (T − 1 layers of input units and one layer of output
units) and no hidden units, so that its conditional energy is defined as
Eθ (x[t] | x[t−T +1,t−1] ) = −b> x[t] −
T
−1
X
(x[t−δ] )> W[δ] x[t] ,
(64)
δ=1
where the weight of the DyBM-T (W[1] , . . . , W[T −1] ) assumes a particular parametric form with a
finite number of parameters that are independent of T , which we discuss in the following.
The parametric form of the weight in the DyBM-T is motivated by observations from biological
neural networks [1] but leads to particularly simple, exact, and efficient learning rule. In biological
neural networks, STDP has been postulated and supported experimentally. In particular, the weight
from a pre-synaptic neuron to a post-synaptic neuron is strengthened, if the post-synaptic neuron fires
(generates a spike) shortly after the pre-synaptic neuron fires (i.e., long term potentiation or LTP).
This weight is weakened, if the post-synaptic neuron fires shortly before the pre-synaptic neuron fires
(i.e., long term depression or LTD). These dependency on the timing of spikes is missing in the Hebbian
rule for the Boltzmann machine (see (36) from [28]).
To have a learning rule with the characteristics of STDP with LTP and LTD, the DyBM-T assumes
[δ]
the weight of the form illustrated in Figure 7. For δ > 0, we define the weight, wi,j , as the sum of two
[δ]
[−δ]
weights, ŵi,j and wj,i :
[δ]
[δ]
[−δ]
wi,j = ŵi,j + ŵj,i ,
3 This
section closely follows [31].
13
(65)
weight
^
Wij[d]
Wij[d]
- dji
d
dij
^
Wji[-d]
Figure 7: The figure illustrates Equation (65) with particular forms of Equation (66) (Figure 2 from
[δ]
[31]). The horizontal axis represents δ, and the vertical axis represents the value of wi,j (solid curves),
[δ]
[−δ]
[δ]
ŵi,j (dashed curves), or ŵj,i (dotted curves). Notice that wi,j is defined for δ > 0 and is discontinuous
[δ]
[−δ]
at δ = d. On the other hand, ŵi,j and ŵj,i are defined for −∞ < δ < ∞ and discontinuous at δ = di,j
and δ = −dj,i , respectively, where recall that we assume di,j = dj,i = d in this paper.
where
[δ]
ŵi,j
0
ui,j λδ−d
=
−vi,j µ−δ
if δ = 0
if δ ≥ d
otherwise.
(66)
for λ, µ ∈ [0, 1). For simplicity, we assume a single decay rate λ for δ ≥ d and a single decay rate µ
for δ < d, as opposed to multiple ones in [31, 32]. For simplicity, we assume that the conduction delay
d is uniform for all connections, as opposed to variable conduction delay in [32]. See also [9, 29] for
ways to tune the values of the conduction delay.
[δ]
In Figure 7, the value of ŵi,j is high when δ = d, the conduction delay from i-th (pre-synaptic)
[0]
unit to the j-th (post-synaptic) unit. Namely, the post-synaptic neuron is likely to fire (i.e., xj = 1)
[−d]
immediately after the spike from the pre-synaptic unit arrives with the delay of d (i.e., xi
= 1). This
[δ]
likelihood is controlled by the LTP weight ui,j . The value of ŵi,j gradually decreases, as δ increases
from d. That is, the effect of the stimulus of the spike arrived from the i-th unit diminishes with time
[1].
[d−1]
[0]
The value of ŵi,j is low, suggesting that the post-synaptic unit is unlikely to fire (i.e., xj = 1)
immediately before the spike from the i-th (pre-synaptic) unit arrives. This unlikelihood is controlled
[δ]
by the LTD weight vi,j . As δ decreases from d − 1, the magnitude of ŵi,j gradually decreases [1].
[δ]
Here, δ can get smaller than 0, and ŵi,j with δ < 0 represents the weight between the spike of the
pre-synaptic neuron that is generated after the spike of the post-synaptic neuron.
14
Neural eligibility trace: 𝛾𝑖
𝑡
Synaptic eligibility trace: 𝛼𝑖
𝑡
FIFO queue
𝑥𝑖
𝑡
𝑥𝑖
𝑡−1
𝑥𝑖
𝑡−2
𝑥𝑖
𝑡−𝑑+1
𝑥𝑗
Pre-synaptic
neuron
𝑡
Post-synaptic
neuron
Figure 8: A connection from a (pre-synaptic) neuron i to a (post-synaptic) neuron j in a DyBM.
5.1.2
Dynamic Boltzmann machine as a limit of a sequence of finite dynamic Boltzmann
machines4
The DyBM is defined as a limit of the sequence of DyBM-T as T → ∞. Because each DyBM-T
is a CRBM, we can also define the limit of the sequence of the conditional probability defined by
DyBM-T , and this limit is considered as the conditional probability defined by the DyBM. Likewise,
the conditional energy of the DyBM is defined as the limit of the sequence of the conditional energy
of DyBM-T .
Specifically, as T → ∞, the conditional energy of DyBM-T in (64) converges to
Eθ (x[t] | x[<t] ) = −b> x[t] −
∞
X
(x[t−d] )> W[d] x[t] ,
(67)
d=1
where the convergence is due to the parametric form (66). This conditional energy in turn defines
the conditional distribution via (9), where we now have no hidden units. Although the conditional
energy (67) of the DyBM involves an infinite sum, it can be evaluated with a finite sum because of the
parametric form (66).
In fact, the DyBM can be understood as an artificial model of a spiking neural network where all
computation for inference and learning is performed locally at each synapse using only the information
available around the synapse. Specifically, a (pre-synaptic) neuron is connected to a (post-synaptic)
neuron via a first-in-first-out (FIFO) queue and a synapse (see Figure 8). At each discrete time t, a
[t]
[t]
neuron i either fires (xi = 1) or not (xi = 0). The spike travels along the FIFO queue and reaches
the synapse after conduction delay, d. In other words, the FIFO queue has the length of d − 1 and
stores, at time t, the spikes that have been generated by the pre-synaptic neuron from time t − d + 1
to time t − 1.
Each synapse in a DyBM stores a quantity called a synaptic eligibility trace. The value of the
synaptic eligibility increases when a spike arrives at the synapse from the FIFO queue; otherwise, it is
[t]
decreased by a constant factor. Specifically, at time t, the value of the synaptic eligibility trace, αi ,
that is stored at the synapse from a pre-synaptic neuron i is updated as follows:
[t]
[t−1]
αi = λ (αi
[t−d+1]
+ xi
),
(68)
where λ is a decay rate and satisfies 0 ≤ λ < 1. Figure 9 shows an example of how the value of
the synaptic eligibility trace changes depending on the spikes arrived at the synapse. Observe that
[t]
αi represents how recently and frequently spikes arrived from a pre-synaptic neuron i and can be
represented non-recursively as follows:
[t−1]
αi
=
t−d
X
s=−∞
15
[s]
λt−s−d xi .
(69)
Eligibility and spike values
2.5
2.0
1.5
1.0
0.5
0.00
5
10
15
Time (DyBM step)
20
Figure 9: The value of a synaptic or neural eligibility trace as a function of time. For a synaptic
eligibility trace at a synapse, the bars represent the spikes arrived from a FIFO queue at that synapse.
For a neural eligibility trace at a neuron, the bars represent the spikes generated by that neuron.
Each neuron in a DyBM stores a quantity called a neural eligibility trace5 . The value of the neural
eligibility increases when the neuron fires; otherwise, it is decreased by a constant factor. Specifically,
[t]
at time t, the value of the neural eligibility trace, γi , at a neuron i is updated as follows:
[t]
[t−1]
γi = µ (γi
[t]
+ xi ),
(70)
[t]
where µ is a decay rate and satisfies 0 ≤ µ < 1. Observe that γi represents how recently and frequently
the neuron i has fired and can be represented non-recursively as follows:
[t−1]
γi
=
t−1
X
[s]
µt−s xj
(71)
s=−∞
A neuron in a DyBM fires according to the probability distribution that depends on the energy of
the DyBM. A neuron is more likely to fire when the energy becomes lower if it fires than otherwise.
[t]
Let Eθ,j xj |x[<t] be the energy associated with a neuron j at time t, which can depend on whether
[t]
j fires at time t (i.e., xj ) as well as the preceding spiking activities of the neurons in the DyBM (i.e.,
x[<t] ). The firing probability of a neuron j is then given by
[t]
Pθ,j (xj |x[<t] )
=
[t]
exp − Eθ,j (xj |x[<t] )
X
exp − Eθ,j (x̃|x[<t] )
(72)
x̃∈{0,1}
[t]
[t]
for xj ∈ {0, 1}. Specifically, Eθ,j xj |x[<t] can be represented as follows:
[t]
[t]
[t]
[t]
LTP
LTD
Eθ,j xj |x[<t] = −bj xj + Eθ,j
xj |x[<t] + Eθ,j
xj |x[<t] ,
(73)
where bj is the bias parameter of a neuron j and represents how likely j spikes (j is more likely to fire
4 This
5 We
section closely follows [27].
assume a single neural eligibility trace, as opposed to multiple ones in [32], at each neuron.
16
if bj has a large positive value), and we define
N
X
[t]
[t−1] [t]
LTP
Eθ,j
xj |x[<t] ≡ −
ui,j αi
xj
(74)
i=1
N
N
X
X
[t]
[t−1] [t]
[t−1] [t]
LTD
Eθ,j
xj |x[<t] ≡
vi,j βi
xj +
vj,k γk
xj ,
i=1
(75)
k=1
[t−1]
where βi
represents how soon and frequently spikes will arrive at the synapse from the FIFO queues
from i to j:
[t−1]
βi
≡
t−1
X
[s]
µs−t xi .
(76)
s=t−d+1
[t−1]
[t−1]
Although βi
can also be represented in a recursive manner, recursively computed βi
is prone to
numerical instability.
In (74), the summation with respect to i is over all of the pre-synaptic neurons that are connected to
j. Here, ui,j is the weight parameter from i to j and represents the strength of Long Term Potentiation
(LTP). This weight parameter is thus referred to as LTP weight. A neuron j is more likely to fire
[t]
[t−1]
(xj = 1) when αi
is large for a pre-synaptic neuron i connected to j (spikes have recently arrived
at j from i) and the corresponding ui,j is positive and large (LTP from i to j is strong).
In (75), the summation with respect to i is over all of the pre-synaptic neurons that are connected
to j, and the summation with respect to k is over all of the post-synaptic neurons which j is connected
to. Here, vi,j represents the strength of Long Term Depression from i to j and referred to as LTD
weight. The neuron j is less likely to fire when βi is large for a pre-synaptic neuron i connected to j
(spikes will soon and frequently reach j from i) and the corresponding vi,j is positive and large (LTD
from i to j is strong). The second term in (75) represents that a pre-synaptic neuron j is less likely to
fire if a post-synaptic neuron has recently and frequently fired (γk is large), and the strength of this
LTD is given by vj,k . Notice that the timing of a spike is measured with respect to when the spike
reaches synapse, where the spike from a pre-synaptic neuron has the delay d, and the spike from a
post-synaptic neuron reaches immediately.
The learning rule of the DyBM has been derived in a way that it maximizes the log likelihood of
given time-series with respect to the probability distribution given by (72) [32]. Specifically, at time t,
the DyBM updates its parameters according to
[t]
[t]
bj ← bj + η xj − Eθ,j [Xj | x[<t] ]
(77)
[t−1]
[t]
[t]
ui,j ← ui,j + η αi
xj − Eθ,j [Xj | x[<t] ]
(78)
[t]
[t]
[t−1]
[t]
[t]
[t−1]
hXi i − xi
(79)
vi,j ← vi,j + η βi
Eθ,j [Xj | x[<t] ] − xj + η γj
[t]
for each of neurons i and j, where η is a learning rate, xj is the training data given to j at time t,
[t]
[t]
and Eθ,j [Xj | x[<t] ] denotes the expected value of xj (i.e., firing probability of a neuron j at time
t) according to the probability distribution given by (72). By following stochastic gradient methods
[6, 18, 10, 41, 33], the learning rate η may be adjusted over time t.
5.1.3
Relation to spike-timing dependent plasticity6
In spike-timing dependent plasticity (STDP), the amount of the change in the weight between two
neurons that fired together depends on the precise timings when the two neurons fired. STDP supplements the Hebbian rule [11] and has been experimentally confirmed in biological neural networks
[5].
6 This
section closely follows [27].
17
[t]
In (78), ui,j is increased (LTP gets stronger) when xj = 1 is given to j. Then j becomes more
[·]
likely to fire when spikes from i have recently and frequently arrived at j (i.e., αi is large). This
[t−1]
amount of the change in ui,j depends on αi
, exhibiting a key property of STDP. In particular, ui,j
is increased by a large amount if spikes from i have recently and frequently arrived at j.
According to the second term on the right-hand side of (79), vi,j is increased (LTD gets stronger)
[t]
when xj = 0 is given to a post-synaptic neuron j. Then j becomes less likely to fire when spikes from
[·]
i are expected to reach j soon (i.e., βi is large). This amount of the change in vi,j is large if there
are spikes in the FIFO queue from i to j and they are close to j. According to the last term of (79),
[t]
vi,j is increased when xi = 0 is given to the pre-synaptic i, and this amount of the change in vi,j is
proportional to γj (i.e., how frequently and recently the post-synaptic j has fired). This learning rule
of (79) thus exhibits some of the key properties of LTD with STDP.
[t]
In (77), bj is increased when xj = 1 is given to j, so that j becomes more likely to fire (in
accordance with the training data), but the amount of the change in bj is small if j is already likely
[t]
[t]
to fire (Eθ,j [Xj | x[<t] ] ≈ 1). This dependency on Eθ,j [Xj | x[<t] ] can be considered as a form of
homeostatic plasticity [42, 20].
Related work There has been a significant amount of the prior work towards understanding STDP
from the perspectives of machine learning [26, 4, 35]. For example, Nessler et al. show that STDP
can be understood as approximating the expectation maximization (EM) algorithm [26]. Nessler et
al. study a particularly structured (winner-take-all) network and its learning rule for maximizing the
log likelihood of given static patterns. On the other hand, the DyBM does not assume particular
structures in the network, and the learning rule having the properties of STDP applies for any synapse
in the network. Also, the learning rule of the DyBM maximizes the log likelihood of given time-series,
and its learning rule does not involve approximations beyond what is assumed in stochastic gradient
methods [6].
5.2
Giving flexibility to the DyBM7
It has been shown in [32] that the DyBM in Section 5.1 has the capability of associative memory and
anomaly detection for sequential patterns, but the applications of the DyBM have been limited to
simple tasks with relatively low dimensional time-series. In [27], we relax some of the constraints of
this DyBM in a way that it gives more flexibility that is useful for learning and inference.
Specifically, observe that the first term on the right-hand side of (75) can be rewritten with the
[t−1]
definition of βi
in (76) as follows:
N
X
[t−1]
vi,j βi
[t]
xj =
i=1
N
X
t−1
X
[s]
[t]
vi,j µs−t xi xj
(80)
i=1 s=t−d+1
=
N X
d−1
X
[δ]
[t−δ]
vi,j xi
[t]
xj ,
(81)
i=1 δ=1
[δ]
[δ]
where we let vi,j ≡ vi,j µ−δ . Here, vi,j represents how unlikely j fires at time t if i fired at time
[δ]
t − δ. The parametric form of vi,j ≡ vi,j µ−δ assumes that this LTD weight decays geometrically as
the interval, δ, between the two spikes increases.
[δ]
In the following, we relax this constraint on vi,j for δ = 1, . . . , d − 1 and assumes that these LTD
weights can take independent values. Then the energy of the DyBM with N neurons can be represented
7 This
section closely follows [27].
18
conveniently with matrix and vector operations:
Eθ (x[t] |x[<t] ) ≡
N
X
[t]
Eθ,j (xj |x[<t] )
(82)
j=1
[t−1] >
− b> x[t] − (αλ
) U x[t] +
d−1
X
(x[t−δ] )> V[δ] x[t] + (x[t] )> V γµ[t−1] ,
(83)
δ=1
where b ≡ (bj )j=1,...,N is a vector, U ≡ (ui,j )(i,j)∈{1,...,N }2 is a matrix, and other boldface letters are
[t−1]
defined analogously (a vector is lowercase and a matrix is uppercase). For eligibility traces (αλ
and
[t−1]
γµ ), we append the subscript to explicitly represent the dependency on the decay rate (λ and µ).
The functional form of the energy completely determines the dynamics of a DyBM, and relaxing its
constraints allows the DyBM to represent a wider class of dynamical systems.
Notice that the last term of (83) can be divided into two terms:
(x[t] )> V γµ[t−1] = (γµ[t−1] )> V x[t]
(84)
= (α[t−1]
)> V x[t] +
µ
d−1
X
(x[t−δ] )> V̂[δ] x[t] ,
(85)
δ=1
[t−1]
is the same as the vector of synaptic eligibility traces but with the decay rate µ, and
where αµ
V̂[δ] ≡ µ−δ V. Comparing (85) and (83), we find that, without loss of generality, the energy of the
DyBM can be represented with the following form:
[t]
Eθ (x |x
[<t]
d−1
L
X
X
[t−1] > [`]
>
[t−δ] >
[δ]
)=− b +
(x
) W +
(αλ` ) U
x[t] ,
δ=1
(86)
`=1
where we define W[δ] = −V[δ] − V̂[δ] . The energy in (86) reduces to the original energy in (73) when
W[δ] = −µ−δ V − µδ V> , U[1] = U, U[2] = −µd V> , λ1 = λ, λ2 = µ, and L = 2. With L > 2, one
can also incorporate multiple synaptic or neural eligibility traces with varying decay rates in [32].
Equivalently, we can represent the energy using neural eligibility traces, γµ` , instead of synaptic
eligibility traces, αλ` , as follows:
[t]
Eθ (x |x
[<t]
d−1
L
X
X
>
[t−δ] >
[δ]
[t−1] >
)=− b +
(x
) W +
(γµ` ) V` x[t] .
δ=1
5.2.1
(87)
`=1
Learning rule in vector-matrix notations
The learning rule corresponding to the representation with (86) is as follows:
b ← b + η (x[t] − Eθ [X [t] | x[<t] ])
W
[δ]
U
[`]
←W
←U
[δ]
[`]
+ηx
+
[t−δ]
[t−1]
η αλ`
(x
(x
[t]
[t]
− Eθ [X
− Eθ [X
[t]
[t]
(88)
|x
|x
[<t]
[<t]
>
])
>
])
(89)
(90)
for each δ and each `, where Eθ [X [t] | x[<t] ] is the conditional expectation with respect to
exp(−Eθ (x[t] | x[<t] ))
.
Pθ (x[t] | x[<t] ) = X
exp(−Eθ (x̃[t] | x[<t] ))
x̃[t]
19
(91)
Specifically,
Eθ [X [t] | x[<t] ] =
exp(m[t] )
1 + exp(m[t] )
(92)
with
[t]
m
>
≡b +
d−1
X
(x
[t−δ] >
) W
[δ]
+
L
X
δ=1
[t−1] >
(αλ`
) U[`] ,
(93)
`=1
where exponentiation and division of vectors are elementwise.
The form of (92) implies that the DyBM is a kind of a logit model, where the feature vector,
[t−1]
[t−1]
[t−d+1]
(x
, . . . , x[t−1] , αλ , αµ ), depends on the prior values, x[<t] , of the time-series. By applying
the learning rules given in (77)-(79) to given time-series, we can learn the parameters of the DyBM
or equivalently the parameters of the logit model (i.e., b, W[δ] for δ = 1, . . . , d − 1, and U[`] for
` = 1, . . . , L) in (92).
5.3
Dynamic Boltzmann machines for real-valued time-series
5.3.1
Gaussian dynamic Boltzmann machines8
In this section, we show how a DyBM can deal with real-valued time-series in the form of a Gaussian
[t]
DyBM [27, 8]. A Gaussian DyBM assumes that xj follows a Gaussian distribution for each j:
(j) [t]
pθ (xj |x[<t] )
[t]
[t] 2
xj − mj
exp −
=q
2 σj2
2
2 π σj
1
(94)
,
[t]
where mj is given by (93), and σj2 is a variance parameter. This Gaussian distribution is in contrast
to the Bernoulli distribution of the DyBM given by (72).
The conditional energy of the Gaussian DyBM can be represented as follows:
Eθ (x[t] | x[<t] ) =
[t]
[t]
N
X
(xj − mj )2
j=1
=
[t]
N
X
(xj − bj )2
j=1
(95)
2 σj2
2 σj2
−
N
N X
d−1 X
X
[t−δ]
xi
[δ]
[t]
wi,j xj −
N X
N
L X
X
[t−1]
[`]
[t]
αi,λ` ui,j xj + C,
(96)
`=1 i=1 j=1
δ=1 i=1 j=1
where C is the term that does not depend on x[t] . Because C is canceled out between the numerator
[δ]
and the denominator in (9), we omit it from the conditional energy. By letting Wσ be the matrix
[δ]
[`]
[`]
whose (i, j)) element is wi,j /σj2 and Uσ be the matrix whose (i, j)) element is ui,j /σj2 , the conditional
energy of the Gaussian DyBM can be represented as follows:
Eθ (x[t] | x[<t] ) =
[t]
N
X
(xj − bj )2
j=1
2 σj2
−
d−1
L
X
X
[t−1]
[t]
(x[t−δ] )> Wσ[δ] x[t] −
(αλ` )> U[`]
σ x .
δ=1
(97)
`=1
The conditional energy of the Gaussian DyBM may be compared against the energy of the Gaussian
Bernoulli restricted Boltzmann machine (see [19] or (181) from [28]).
8 This
section closely follows [27].
20
We now derive a learning rule for the Gaussian DyBM in a way that it maximizes the log-likelihood
of given time-series x:
X
log pθ (x[t] |x[<t] ) =
N
XX
t
t
[t]
log pi (xi |x[−∞,t−1] ),
(98)
i=1
where the summation over t is over all of the time steps of x, and the conditional independence between
[t]
[t]
xi and xj for i 6= j given x[<t] is the fundamental property of the DyBM as shown in [32].
The approach of stochastic gradient is to update the parameters of the Gaussian DyBM at each
step, t, according to the gradient of the conditional probability density of x[t] :
∇ log pθ (x[t] |x[<t] ) = −
N
X
1
i=1
2
[t]
∇ log σi2 + ∇
[t]
xi − mi )2
,
2 σi2
(99)
where the equality follow from (94). From (99) and (93), we can derive the derivative with respect to
each parameter as follows:
[t]
[t]
xj − µj [t]
∂
xj
log pθ (x[t] |x[−∞,t−1] ) =
∂bj
σj2
[t]
(100)
[t]
xj − µj [t−1]
∂
log pθ (x[t] |x[−∞,t−1] ) =
αi,j
∂ui,j
σj2
∂
[δ]
∂wi,j
[t]
log pθ (x[t] |x[−∞,t−1] ) =
(101)
[t]
xj − µj [t−δ]
xi
σj2
(102)
[t] 2
[t]
xj − µj
∂
1
log pθ (x[t] |x[−∞,t−1] ) = − +
∂σj
σj
σj3
,
(103)
where δ ∈ {1, . . . , d − 1}, ` ∈ {1, . . . , }, and i, j ∈ {1, . . . , N }.
These parameters are thus updated with learning rate η as follows:
x[t] − m[t]
σ2
2
[t]
x − m[t] − σ 2
σ ←σ+η
σ3
b←b+η
(104)
(105)
!>
W[δ] ← W[δ] + η x[t−δ]
[t−1]
U[`] ← U[`] + η αλ`
x[t] − m[t]
σ2
!>
x[t] − m[t]
σ2
(106)
(107)
where division and exponentiation of vectors are elementwise, and m[t] is given by (93).
The maximum likelihood estimator of x[t] by the Gaussian DyBM is given by m[t] in (93). The
Gaussian DyBM can thus be understood as a modification to the standard vector autoregressive (VAR)
model. Specifically, the last term in the right-hand side of (93) involves eligibility traces, which can
be understood as features of historical values, x(−∞,t−d] , and are added as new variables to the VAR
model. Because the value of the eligibility traces can depend on the infinite past, the Gaussian DyBM
can take into account the history beyond the lag d − 1.
21
5.3.2
Natural gradients9
In this section, we study a learning rule based on natural gradient for the Gaussian DyBM. Consider
a stochastic model that gives the probability density of a pattern x as pθ (x). With natural gradients
[3], the parameters, θ, of the stochastic model are updated as follows:
θt+1 = θt − η G−1 (θt ) ∇ log pθ (x)
(108)
at each step t, where η is the learning rate, and G(θ) denotes the Fisher information matrix:
Z
G(θ) ≡ pθ (x) ∇ log pθ (x) ∇ log pθ (x)> dx.
(109)
Due to the conditional independence in (98), it suffices to derive a natural gradient for each Gaussian
unit. Here, we consider the parametrization with mean m and variance v ≡ σ 2 . The probability density
function of a Gaussian distribution is represented with this parametrization as follows:
(x − m)2
1
exp −
.
(110)
p(x; m, v) = √
2v
2π v
The log likelihood of x is then given by
log p(x; m, v) = −
(x − m)2
1
1
− log v − log 2π.
2v
2
2
(111)
Hence, the gradient and the inverse Fisher information matrix in (108) are given as follows:
!
∇ log pθ (x) =
G−1 (θ) =
x−m
v
(x−m)2
1
− 2v
2
2v
−1
1
0
v
0 2v12
(112)
=
0
2v 2
v
0
(113)
,
The parameters θt ≡ (mt , vt ) are then updated as follows:
mt+1 = mt + η (x − mt )
(114)
2
vt+1 = vt + η (x − mt ) − vt .
(115)
[t]
In the context of a Gaussian DyBM, the mean is given by (93), where mj is linear with respect
[`]
to bj , wi,j , and ui,j . Also, the variance is given by σj2 . Hence, the natural gradient gives the learning
rules for these parameters as follows:
b ← b + η (x[t] − m[t] )
2
σ 2 ← σ 2 + η x[t] − m[t] − σ 2
(117)
W[δ] ← W[δ] + η x[t−δ] (x[t] − m[t] )>
(118)
[t−1]
η αλ`
(119)
U[`] ← U[`] +
(x[t] − m[t] )> ,
(116)
where the exponentiation of a vector is elementwise. We can compare (116)-(119) against what the
standard gradient gives in (104)-(107).
9 This
section closely follows [27].
22
5.3.3
Using nonlinear features in Gaussian DyBMs
The Gaussian DyBM is a linear model and has limited capability in modeling complex time-series. A
way to take into account non-linear features of time-series with a Gaussian DyBM is to apply nonlinear mapping to input time-series and feed the resulting non-linear features as additional input to
the Gaussian DyBM. An example of such non-linear mapping is an echo state network (ESN) [16].
An ESN maps an input sequence, x, into ψ recursively as follows:
(120)
ψ [t] = (1 − ρ) ψ [t−1] + ρ tanh Wrec ψ [t−1] + Win x[t] ,
where Wrec and Win are randomly chosen and fixed parameters10 , and ρ is a leak parameter satisfying
0 < ρ < 1. In (120), tanh is a hyperbolic tangent function but may be replaced with other nonlinear
functions such as a sigmoid function.
An eligibility trace may be considered as a linear counterpart of the nonlinear features created by
an ESN. Because these features are generated by mappings with fixed parameters and just given as
input to a Gaussian DyBM, the learning rules for the Gaussian DyBM stay unchanged. The nonlinear
DyBM in [8] uses an ESN in a slightly different manner.
5.4
Functional dynamic Boltzmann machines
We now review a functional DyBM, which models time-series of functions (patterns over a continuous
space Z) [17]. Recall that a Gaussian DyBM defines the conditional distribution of the next real-valued
vector given the preceding sequence of real-valued vectors. A functional DyBM defines the conditional
distribution of the next function (i.e., g [t] ) given the preceding sequence of partial observations of
[s]
preceding functions. At each time s, a set of points Z [s] ≡ (zi )i=1,...,Ns is observed, where Ns is the
number of points that are observed at s.
The functional DyBM assumes that the conditional distribution of g [t] (·) is given by a Gaussian
process, whose mean µ[t] (·) varies over time depending on preceding functions as follows:
µ[t] (z) = b(z) +
d−1 Z
X
δ=1
w[δ] (z, z 0 ) g [t−δ] (z 0 ) dz 0 +
Z
L Z
X
`=1
[t−1]
Z
u` (z, z 0 ) α`
(z 0 ) dz 0
(121)
for x ∈ Z, where b(·) is a functional bias, w[δ] (·, ·) and u` (·, ·) are functional weight for each δ and for
each `, and
[t−1]
α`
(·) =
t−d
X
λ`t−s−d g [s] (·)
(122)
s=−∞
is a functional eligibility trace for each `. The covariance kσ2 (·, ·) of the Gaussian process consists of
two components such that
kσ2 (z, z 0 ) = k(z, z 0 ) + σ 2 δ(z, z 0 ),
(123)
where k(·, ·) is a arbitrary kernel, δ(·, ·) is a delta function, and σ is a hyperparameter.
For tractability, Kajino proposes particular parametrization for the functional bias and functional
weight [17]. Let P = (p1 , . . . , pM ) be a set of arbitrarily selected M points in Z
b(z) = kσ2 (z, P ) b
[δ]
0
(124)
w (z, z ) = kσ2 (z, P ) W
0
u` (z, z ) = kσ2 (z, P ) U
10 The
spectral radius of Wrec is set smaller than 1.
23
0
(125)
kσ2 (P, z )
(126)
[δ]
[`]
kσ2 (P, z )
0
for each δ and each `, where
kσ2 (z, P ) ≡ (kσ2 (z, pi ))i=1,...,M
(127)
kσ2 (P, z 0 ) ≡ (kσ2 (pi , z 0 ))i=1,...,M
(128)
is a row vector, and
is a column vector.
Because g [t] is in the reproducing kernel Hilbert space with kernel kσ2 , substituting (124)-(126)
into (121) gives the following expression:
d−1
L
X
X
[t]
[t−1]
µθ (z) = kσ2 (z, P ) b +
W[δ] g [t−δ] (P ) +
U[`] α` (P ) ,
δ=1
(129)
`=1
where g [t−δ] (P ) is a column vector with i-th element being g [t−δ] (pi ), and the eligibility-trace vector
[t−1]
α` (P ) can be recursively updated as follows:
[t]
[t−1]
α` (P ) = λ` α`
(P ) + g [t−d+1] (P ) .
(130)
Here, we use θ ≡ (b, W[1] , . . . , W[d−1] , U[1] , . . . , U[L] ) to collectively denote the parameters.
While g [s] (pi ) for i ∈ [1, M ] is not observed, Kajino uses a maximum a posteriori (MAP) estimator
[s]
ĝ (pi ) in [17]:
[s]
[t]
ĝ [s] (pi ) = µθ (pi ) + k(pi , Z [t] ) kσ2 (Z [t] , Z [t] )−1 g [t] (Z [t] ) − µθ (Z [t] )) ,
[t]
[t]
(131)
[t]
where kσ2 (Z [t] , Z [t] ) is an Ns × Ns matrix with (i, j)-th element being kσ2 (zi , zj ), and µθ (Z [t] ) is a
column vector defined analogously to g [t] (Z [t] ).
The objective of learning a functional DyBM is to maximize the log likelihood of observed values.
The conditional probability density of the functional values of locations Z [t] at time t is given by
1
>
[t]
[t]
pθ (g [t] (Z [t] ) | g [<t] ) ∼ exp − g [t] (Z [t] ) − µθ (Z [t] ) kσ2 (Z [t] , Z [t] )−1 g [t] (Z [t] ) − µθ (Z [t] ) .
2
(132)
The objective is thus to maximize
X
f (θ) ≡
ft (θ),
(133)
t
where
ft (θ) ≡ log pθ (g [t] (Z [t] ) | g [<t] )
>
1
[t]
[t]
= − g [t] (Z [t] ) − µθ (Z [t] ) kσ2 (Z [t] , Z [t] )−1 g [t] (Z [t] ) − µθ (Z [t] ) + C,
2
(134)
(135)
where C is the term independent of θ.
The gradient of ft (θ) is given by
[t]
[t]
∇ft (θ) = ∇µθ (Z [t] )> kσ2 (Z [t] , Z [t] )−1 g [t] (Z [t] ) − µθ (Z [t] ) ,
24
(136)
𝐙𝛿
𝐡 𝑡−𝛿
𝐙1
𝐡 𝑡−1
𝐡𝑡
𝐕1
𝐕𝛿
𝐔𝛿
𝐱 𝑡−𝛿
𝐔1
𝐱 𝑡−1
𝐖1
𝐱𝑡
𝐛
𝐖𝛿
Figure 10: A dynamic Boltzmann machine with hidden units (modified Figure 1 from [30]).
where (129) gives
∂ [t] [t] >
µ (Z ) = kσ2 (pi , Z [t] )
∂bi θ
∂
[t]
µ (Z [t] )> = kσ2 (pi , Z [t] ) g [t−δ] (pj )
[δ] θ
∂wi,j
∂
[t]
[`]
∂ui,j
[t−1]
µθ (Z [t] )> = kσ2 (pi , Z [t] ) α`
(137)
(138)
(pj )
(139)
for each i, j, δ, `.
The gradient implies the following learning rule with stochastic gradient:
[t]
b ← b + η kσ2 (P, Z [t] ) kσ2 (Z [t] , Z [t] )−1 g [t] (Z [t] ) − µθ (Z [t] )
W
[δ]
←
U[`] ←
[t]
W + η kσ2 (P, Z ) kσ2 (Z , Z )
g (Z ) − µθ (Z [t] ) g [t−δ] (P )>
[t−1]
[t]
U[`] + η kσ2 (P, Z [t] ) kσ2 (Z [t] , Z [t] )−1 g [t] (Z [t] ) − µθ (Z [t] ) α` (P )> ,
[δ]
[t]
[t]
[t] −1
[t]
[t]
(140)
(141)
(142)
where η is a learning rate, and g [t−δ] (P ) is estimated with the MAP estimator ĝ [t−δ] (P ) in (131).
5.5
Dynamic Boltzmann machines with hidden units11
In this section, we study a DyBM with hidden units (see Figure 10). Each layer of this DyBM
corresponds to a time t − δ for 0 ≤ δ < ∞ and has two parts: visible and hidden. The visible part
x[t−δ] at the δ-th layer represents the values of the time-series at time t − δ. The hidden part h[t−δ]
represents the values of hidden units at time t−δ. Here, units within each layer do not have connections
to each other. We let x[<t] ≡ (x[s] )s<t and define h[<t] analogously.
The Boltzmann machine in Figure 10 has bias parameter b and weight parameter (U, V, W, Z).
Let θ ≡ (V, W, b) be the parameters connected to visible units x[t] (from the units in the past, x[s] or
h[s] for s < t) and φ ≡ (U, Z). The conditional energy of this Boltzmann machine is given as follows:
Eθ,φ (x[t] , h[t] | x[<t] , h[<t] ) = Eθ (x[t] | x[<t] , h[<t] ) + Eφ (h[t] | x[<t] , h[<t] ),
11 This
section closely follows [30].
25
(143)
where we define
Eθ (x[t] | x[<t] , h[<t] ) = −b> x[t] −
Eθ (h[t] | x[<t] , h[<t] ) = −b> h[t] −
∞
X
(x[t−δ] )> W[δ] x[t] −
∞
X
(h[t−δ] )> V[δ] x[t]
(144)
(h[t−δ] )> Z[δ] h[t] .
(145)
δ=1
∞
X
δ=1
∞
X
δ=1
δ=1
(x[t−δ] )> U[δ] h[t] −
and assume the following parametric form for δ ≥ d:
W[δ] = λδ−d W[d]
V
[δ]
Z
[δ]
U
[δ]
=λ
δ−d
=λ
δ−d
=λ
δ−d
(147)
[d]
(148)
[d]
(149)
V
Z
(146)
[d]
U ,
where λ is a decay rate satisfying 0 ≤ λ < 1. Then the conditional energy can be represented as
follows:
Eθ (x[t] | x[<t] , h[<t] )
= −b> x[t] −
d−1
X
(x[t−δ] )> W[δ] x[t] −
d−1
X
(h[t−δ] )> V[δ] x[t] − (α[t−1] )> W[d] x[t] − (β [t−1] )> V[d] x[t] ,
δ=1
δ=1
(150)
Eφ (h[s] | x[<s] , h[<s] )
=−
d−1
X
δ=1
(x[s−δ] )> U[δ] h[s] −
d−1
X
(h[s−δ] )> Z[δ] h[s] − (α[s−1] )> U[d] h[s] − (β [s−1] )> Z[d] h[s] .
(151)
δ=1
where α[t−1] corresponds to the eligibility trace in the DyBM in (69), and we define an eligibility trace
β [t−1] for the hidden part analogously:
α
[t−1]
≡
β [t−1] ≡
∞
X
δ=d
∞
X
λδ−d x[t−δ]
(152)
λδ−d h[t−δ] .
(153)
δ=d
The energy in (150)-(151) gives the conditional probability distribution over x[t] and h[t] given x[<t]
and h[<t] . Specifically, we have
exp(−Eθ (x[t] | x[<t] , h[<t] ))
Pθ (x[t] | x[<t] , h[<t] ) = X
exp(−Eθ (x̃[t] | x[<t] , h[<t] ))
(154)
x̃[t]
exp(−Eφ (h[s] | x[<s] , h[<s] ))
Pφ (h[s] | x[<s] , h[<s] ) = X
exp(−Eφ (h̃[s] | x[<s] , h[<s] ))
h̃[t]
for any binary vectors x[t] and h[t] .
26
(155)
5.5.1
Learning a dynamic Boltzmann machine with hidden units
The DyBM with hidden units gives the probability of a time-series, x ≡ (x[t] )t=`,...,u , by
Pθ,φ (x) =
X
Pφ (h̃ | x)
u
Y
Pθ (x[t] | x[<t] , h̃[<t] )
(156)
t=`
h̃
P
where h̃ denotes the summation over all of the possible values of hidden units from time t = ` to
t = u, and
Pφ (h̃ | x) ≡
u
Y
Pφ (h̃[s] | x[<s] , h̃[<s] ),
(157)
s=`
where we arbitrarily define x[s] = 0 and h̃[s] = 0 for s < `.
We seek to maximize the log likelihood of a given x by maximizing a lower bound given by Jensen’s
inequality:
log Pθ,φ (x) = log
X
Pφ (h̃ | x)
X
Pφ (h̃ | x) log
X
Pφ (h̃ | x)
u
Y
Pθ (x[t] | x[<t] , h̃[<t] )
(159)
u
X
log Pθ (x[t] | x[<t] , h̃[<t] )
(160)
t=`
h̃
=
(158)
t=`
h̃
=
Pθ (x[t] | x[<t] , h̃[<t] )
t=`
h̃
≥
u
Y
u X
X
Pφ (h̃[<t] | x[<t−1] ) log Pθ (x[t] | x[<t] , h̃[<t] )
t=` h̃[<t]
≡ Lθ,φ (x),
(161)
where the summation with respect to h̃[<t] is over all of the possible values of h̃[s] for s ≤ t − 1, and
[<t]
Pφ (h̃
|x
[<t−1]
)≡
t−1
Y
Pφ (h̃[s] | x[<s] , h̃[<s] ).
(162)
s=`
Learning weight to visible units The gradient of the lower bound with respect to θ is then given
by
∇θ Lθ,φ (x) =
u X
X
t=`
Pφ (h̃[<t] | x[<t−1] ) ∇θ log Pθ (x[t] | x[<t] , h̃[<t] ).
(163)
h̃[<t]
The right-hand side of (163) is a summation of expected gradients, which suggests a method of stochastic gradient. Namely, at each step t, we sample H [t−1] (ω) according to Pφ (h[t−1] | x[<t−1] , h[<t−1] )
and update θ on the basis of
∇θ log Pθ (x[t] | x[<t] , H [<t] (ω)).
(164)
This learning rule is equivalent to the one for the model where all of the units are visible, except
that the values for the hidden units are given by sampled values. Therefore, the learning rule for θ
27
follows directly from Section 5.1.2:
b ← b + η x[t] − Eθ X [t] | x[<t] , H [<t] (ω)
>
W[d] ← W[d] + η α[t−1] x[t] − Eθ X [t] | x[<t] , H [<t] (ω)
>
V[d] ← V[d] + η β [t−1] (ω) x[t] − Eθ X [t] | x[<t] , H [<t] (ω)
>
W[δ] ← W[δ] + η x[t−δ] x[t] − Eθ X [t] | x[<t] , H [<t] (ω)
>
V[δ] ← V[δ] + η H [t−δ] (ω) x[t] − Eθ X [t] | x[<t] , H [<t] (ω)
(165)
(166)
(167)
(168)
(169)
for 1 ≤ δ < d, where Eθ [X [t] | x[<t] , H [<t] (ω)] denotes the conditional expectation with respect to
Pθ (· | x[<t] , H [<t] (ω)), and we make explicit that β [s−1] is computed with sampled hidden values:
β [s−1] (ω) =
∞
X
λδ−d H [s−δ] (ω).
(170)
δ=d
Now we take the gradient of Lθ,φ (x) with respect to φ:
Learning weight to hidden units
∇φ Lθ,φ (x) =
u X
X
∇φ pφ (h̃[<t] | x[<t−1] ) log pθ (x[t] | x[<t] , h̃[<t] ),
(171)
t=` h̃[<t]
where
∇φ pφ (h̃[<t] | x[<t−1] ) = ∇φ
t−1
Y
pφ (h̃[s] | x[<s] , h̃[<s] )
(172)
s=`
=
t−1
X
∇φ log pφ (h̃[s] | x[<s] , h̃[<s] )
t−1
Y
0
0
0
pφ (h̃[s ] | x[<s ] , h̃[<s ] )
s0 =`
s=`
= pφ (h̃[<t] | x[<t−1] )
t−1
X
∇φ log pφ (h̃[s] | x[<s] , h̃[<s] ).
(173)
s=`
Plugging (173) into the right-hand side of (171), we obtain
∇φ Lθ,φ (x) =
u X
X
pφ (h̃[<t] | x[<t−1] ) log pθ (x[t] | x[<t] , h̃[<t] )
t=` h̃[<t]
t−1
X
∇φ log pφ (h̃[s] | x[<s] , h̃[<s] ).
s=`
(174)
Similar to (163), the expression of (174) suggests a method of stochastic gradient: at each time t, we
sample H [t−1] (ω) according to pφ (h[t−1] | x[<t−1] ) and update φ on the basis of the following stochastic
gradient:
log pθ (x[t] | x[<t] , H [<t] (ω)) Gt−1 ,
(175)
where
Gt−1 ≡
t−1
X
∇φ log pφ (H [s] (ω) | x[<s] , H [<s] (ω)).
s=`
28
(176)
The learning rules for U and Z are derived from (175)-(176) as follows:
U[d] ← U[d] + η log pθ (x[t] | x[<t] , h[<t] )
t−1
X
>
α[s−1] H [s] (ω) − Eφ H[s] | x[<s] , H [<s] (ω)
(177)
s=`
t−1
X
Z[d] ← Z[d] + η log pθ (x[t] | x[<t] , h[<t] )
>
β [s−1] (ω) H [s] (ω) − Eφ H [s] | x[<s] , H [<s] (ω)
(178)
s=`
U
[δ]
←U
[δ]
+ η log pθ (x
[t]
[<t]
|x
[<t]
,h
)
t−1
X
>
x[s−δ] H [s] (ω) − Eφ H[s] | x[<s] , H [<s] (ω)
(179)
s=`
Z[δ] ← Z[δ] + η log pθ (x[t] | x[<t] , h[<t] )
t−1
X
>
H [s−δ] (ω) H [s] (ω) − Eφ H [s] | x[<s] , H [<s] (ω)
s=`
(180)
for 1 ≤ δ < d, where Eφ [H [s] (ω) | x[<s] , H [<s] (ω)] denotes the conditional expectation with respect to
Pφ (· | x[<s] , H [<s] (ω)).
Computation of (174) involves mainly two interrelated inefficiencies. First, although (174) can be
approximately computed using sampled hidden values H [<t] (ω) in the same way as (163), the samples
cannot be reused after updating φ because it was sampled from the distribution with the previous
parameter. Second, since each summand of Gt−1 depends on φ, Gt−1 also has to be recomputed after
each update. Thus, the computational complexity of (175) grows linearly with respect to the length
of the time-series (i.e., t − `), in contrast to (164), whose complexity is independent of that length.
Observe in (175) that ∇φ Lθ,φ (x) consists of the products of log pθ (x[t] | x[<t] , H [<t] (ω)) and
∇φ log pφ (H [s] (ω) | x[<s] , H [<s] (ω)) for s < t. Without the dependency on log pθ (x[t] | x[<t] , H [<t] (ω)),
the parameter φ is updated in a way that H [s] (ω) is more likely to be generated (i.e., the learning rule
would be equivalent to that for visible units). Such an update rule is undesirable, because H [s] (ω) has
been sampled and is not necessarily what we want to sample again. The dependency on log pθ (x[t] |
x[<t] , H [<t] (ω)) suggests that φ is updated by a large amount if the sampled H [s] (ω) happens to make
the future values, x[t] for t > s, likely. Intuitively, weighting ∇φ log pφ (H [s] (ω) | x[<s] , H [<s] (ω)) by
log pθ (x[t] | x[<t] , H [<t] (ω)) for t > s is inevitable, because whether the particular values of hidden
units are good for the purpose of predicting future values will only be known after seeing future values.
Approximations
One could approximately compute (176) recursively:
Gt ← γ Gt−1 + (1 − γ)∇φ log pφ (H [t] (ω) | x[<t] , H [<t] (ω)),
(181)
where γ ∈ [0, 1) is a discount factor. The recursive update rule with γ < 1 puts exponentially small
weight γ t−s on ∇φ log pφ (H [s] (ω) | x[<s] , H [<s] (ω)) computed with an old value of φ (i.e., s t).
This recursively computed Gt is related to the momentum in gradient descent [33].
In (177)-(180), the value of Eφ [H [s] | x[<s] , H [<s] (ω)] is computed with the latest values of φ.
Let φ[t−1] be the value of φ immediately before step t. With the recursive computation of (181), the
29
learning rules of (177)-(180) are approximated with the following learning rules:
U[d] ← U[d] + η (1 − γ) log pθ (x[t] | x[<t] , H [<t] (ω))
t−1
X
>
γ t−1−s α[s−1] H [s] (ω) − Eφ[s−1] H [s] | x[<s] , H [<s] (ω)
(182)
s=`
Z[d] ← Z[d] + η (1 − γ) log pθ (x[t] | x[<t] , H [<t] (ω))
t−1
X
>
γ t−1−s β [s−1] H [s] (ω) − Eφ[s−1] H [s] | x[<s] , H [<s] (ω)
(183)
s=`
U[δ] ← U[δ] + η (1 − γ) log pθ (x[t] | x[<t] , H [<t] (ω))
t−1
X
>
γ t−1−s x[s−δ] H [s] (ω) − Eφ[s−1] H [s] | x[<s] , H [<s] (ω)
(184)
s=`
Z[δ] ← Z[δ] + η (1 − γ) log pθ (x[t] | x[<t] , H [<t] (ω))
t−1
X
>
γ t−1−s H [s−δ] (ω) H [s] (ω) − Eφ[s−1] H [s] | x[<s] , H [<s] (ω)
(185)
s=`
for 1 ≤ δ < d, where H [s] (ω) is a sample according to Pφ[s−1] (· | x[<s] , H [<s] (ω)) for each s. In
(182)-(182), the quantity such as
G0t−1 ≡
t−1
X
>
γ t−1−s α[s−1] H [s] (ω) − Eφ[s−1] H [s] | x[<s] , H [<s] (ω)
(186)
s=`
can be computed recursively as
>
G0t ← γ G0t−1 + (1 − γ) α[t−1] H [t] (ω) − Eφ[s−1] H [t] | x[<t] , H [<t] (ω)
.
(187)
In [30], we present an alternative approach of learning the DyBM with hidden units in a bidirectional
manner, where we consider a backward DyBM that shares the parameters of the (forward) DyBM. Our
key observation is that the parameters that are difficult to learn in the forward DyBM are relatively
easy to learn in the backward DyBM. By training both the forward DyBM and the backward DyBM,
we can effectively learn the parameters of the forward DyBM.
6
Conclusion
We have reviewed Boltzmann machines for time-series modeling. Such Boltzmann machines can be
used for prediction [8, 30, 17, 38, 7, 22] and anomaly detection based on observed time-series. They
may be also used to generate time-series such as human motion [40, 39, 38], music [22], and movies.
The use of Boltzmann machines is only one approach to modeling and learning time-series. Popular
time-series models include but not limited to recurrent neural networks [34], long short term memory
[15], autoregressive models, and hidden Markov models. As we have seen some of the examples, the
best time-series model for a particular application might be obtained by appropriately combining some
of existing time-series models.
Acknowledgments
This work was supported by JST CREST Grant Number JPMJCR1304, Japan. The author thanks
Diyuan Lu for pointing out several typographical errors in the original version.
30
References
[1] L. F. Abbott and S. B. Nelson. Synaptic plasticity: Taming the beast. Nature Neuroscience,
3:1178–1183, 2000.
[2] D. H. Ackley, G. E. Hinton, and T. J. Sejnowski. A learning algorithm for Boltzmann machines.
Cognitive Science, 9:147–169, 1985.
[3] S. Amari and H. Nagaoka. Methods of Information Geometry. Oxford University Press, 2000.
[4] Y. Bengio, T. Mesnard, A. Fischer, S. Zhang, and Y. Wu. STDP as presynaptic activity times
rate of change of postsynaptic activity. arXiv:1509.05936v2, 2016.
[5] G. Bi and M. Poo. Synaptic modifications in cultured hippocampal neurons: Dependence on spike
timing, synaptic strength, and postsynaptic cell type. Journal of Neuroscience, 18:10464–10472,
1998.
[6] L. Bottou. Online learning and stochastic approximations. In D. Saad, editor, On-Line Learning
in Neural Networks, chapter 2, pages 9–42. Cambridge University Press, 2009.
[7] N. Boulanger-Lewandowski, Y. Bengio, and P. Vincent. Modeling temporal dependencies in
high-dimensional sequences: Application to polyphonic music generation and transcription. In
Proceedings of the 29th International Conference on Machine Learning (ICML-12), pages 1159–
1166, 2012.
[8] S. Dasgupta and T. Osogami. Nonlinear dynamic Boltzmann machines for time-series prediction.
In The 31st AAAI Conference on Artificial Intelligence (AAAI-17), January 2017.
[9] S. Dasgupta, T. Yoshizumi, and T. Osogami. Regularized dynamic Boltzmann machine with delay
pruning for unsupervised learning of temporal sequences. In Proceedings of the 23rd International
Conference on Pattern Recognition, 2016.
[10] J. Duchi, E. Hazan, and Y. Singer. Adaptive subgradient methods for online learning and stochastic optimization. Journal of Machine Learning Research, 12:2121–2159, 2011.
[11] D. O. Hebb. The organization of behavior: A neuropsychological approach. Wiley, 1949.
[12] G. E. Hinton and A. D. Brown. Spiking Boltzmann machines. In Advances in Neural Information
Processing Systems 12, pages 122–128. November 1999.
[13] G. E. Hinton and R. Salakhutdinov. Reducing the dimensionality of data with neural networks.
Science, 313:504–507, 2006.
[14] G. E. Hinton and T. J. Sejnowski. Optimal perceptual inference. In Proc. IEEE Conference on
Computer Vision and Pattern Recognition, pages 448–453, June 1983.
[15] S. Hochreiter and J. Schmidhuber. Long short-term memory. Neural Computation, 9(8):1735–
1780, 1997.
[16] H. Jaeger and H. Haas. Harnessing nonlinearity: Predicting chaotic systems and saving energy in
wireless communication. Science, 304(5667):78–80, 2004.
[17] H. Kajino. A functional dynamic Boltzmann machine. In Proceedings of the International Joint
Conference on Artificial Intelligence (IJCAI-17), pages 1987–1993, 2017.
[18] D. P. Kingma and J. Ba. Adam: A method for stochastic optimization. In Proceedings of the
International Conference on Learning Representations (ICLR), arXiv:1412.6980, 2015.
31
[19] A. Krizhevsky. Learning multiple layers of features from tiny images. Master’s thesis, Computer
Science Department, University of Toronto, Toronto, Canada, 2009.
[20] A. Lazar, G. Pipa, and J. Triesch. SORN: A self-organizing recurrent neural network. Frontiers
in Computational Neurosci., 3:Article 23, 2009.
[21] H. Lütkepohl. New Introduction to Multiple Time Series Analysis. Springer-Verlag Berlin Heidelberg, 2005.
[22] Q. Lyu, Z. Wu, and J. Zhu. Polyphonic music modelling with LSTM-RTRBM. In Proceedings of
the 23rd ACM international conference on Multimedia, pages 991–994, October 2015.
[23] T. Marks and J. Movellan. Diffusion networks, products of experts, and factor analysis. In
Proceedings of the Third International Conference on Independent Component Analysis and Blind
Source Separation, 2001.
[24] R. Memisevic and G. E. Hinton. Unsupervised learning of image transformations. In Proceedings
of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR 2007), pages 1–8,
2007.
[25] R. Mittelman, B. Kuipers, S. Savarese, and H. Lee. Structured recurrent temporal restricted
Boltzmann machines. In Proc. 31st Annual International Conference on Machine Learning (ICML
2014), pages 1647–1655, June 2014.
[26] B. Nessler, M. Pfeiffer, L. Buesing, and W. Maass. Bayesian computation emerges in generic
cortical microcircuits through spike-timing-dependent plasticity. PLoS Computational Biology,
9(4):e1003037.
[27] T. Osogami. Learning binary or real-valued time-series via spike-timing dependent plasticity.
CoRR, abs/1612.04897 (presented at Computing with Spikes NIPS 2016 Workshop, Barcelona,
Spain, December 2016), 2016.
[28] T. Osogami. Boltzmann machines and energy-based models. Technical Report RT0979, IBM
Research - Tokyo, 2017.
[29] T. Osogami and S. Dasgupta. Learning the values of the hyperparameters of a dynamic Boltzmann
machine. IBM Journal of Research and Development, 61(4/5):to appear, 2017.
[30] T. Osogami, H. Kajino, and T. Sekiyama. Bidirectional learning for time-series models with
hidden units. In Proceedings of the 34th International Conference on Machine Learning (ICML
2017), pages 2711–2720, August 2017.
[31] T. Osogami and M. Otsuka. Learning dynamic Boltzmann machines with spike-timing dependent
plasticity. Technical Report RT0967, IBM Research, 2015.
[32] T. Osogami and M. Otsuka. Seven neurons memorizing sequences of alphabetical images via
spike-timing dependent plasticity. Scientific Reports, 5:14149, 2015.
[33] N. Qian. On the momentum term in gradient descent learning algorithms. Neural Networks: The
Official Journal of the International Neural Network Society, 12(1):145–151, 1999.
[34] D. E. Rumelhart, G. E. Hinton, and R. J. Williams. Learning internal representations by error
propagation. In D. E. Rumelhart, J. L. McClelland, and PDP Research Group, editors, Parallel
Distributed Processing: Explorations in the Microstructure of Cognition, vol. 1: Foundations,
chapter 8. MIT Press, 1986.
[35] B. Scellier and Y. Bengio. Equilibrium propagation: Bridging the gap between energy-based
models and backpropagation. arXiv:1602.05179v4, 2016.
32
[36] B. Schrauwen and L. Buesing. A hierarchy of recurrent networks for speech recognition. In NIPS
Workshop on Deep Learning for Speech Recognition and Related Applications, 2009.
[37] I. Sutskever and G. E. Hinton. Learning multilevel distributed representations for high-dimensional
sequences. In Proceedings of the Eleventh International Conference on Artificial Intelligence and
Statistics (AISTATS-07), volume 2, pages 548–555. Journal of Machine Learning Research - Proceedings Track, 2007.
[38] I. Sutskever, G. E. Hinton, and G. W. Taylor. The recurrent temporal restricted Boltzmann
machine. In Advances in Neural Information Processing Systems 21, pages 1601–1608. December
2008.
[39] G. W. Taylor and G. E. Hinton. Factored conditional restricted Boltzmann machines for modeling
motion style. In Proc. 26th Annual International Conference on Machine Learning (ICML 2009),
pages 1025–1032, June 2009.
[40] G. W. Taylor, G. E. Hinton, and S. T. Roweis. Modeling human motion using binary latent
variables. In P. B. Schölkopf, J. C. Platt, and T. Hoffman, editors, Advances in Neural Information
Processing Systems 19, pages 1345–1352. MIT Press, 2007.
[41] T. Tieleman and G. E. Hinton. Lecture 6.5—Rmsprop: Divide the gradient by a running average
of its recent magnitude. COURSERA: Neural Networks for Machine Learning, 2012.
[42] G. G. Turrigiano and S. B. Nelson. Homeostatic plasticity in the developing nervous system.
Nature Rev. Neurosci., 5:97107, 2004.
[43] M. Welling, M. Rosen-Zvi, and G. E. Hinton. Exponential family harmoniums with an application
to information retrieval. In Advances in Neural Information Processing Systems 17, pages 1481–
1488. 2004.
33
| 9cs.NE
|
Cascade and locally dissipative realizations of linear quantum systems
for pure Gaussian state covariance assignment ⋆
arXiv:1604.03182v1 [quant-ph] 12 Apr 2016
Shan Ma a , Matthew J. Woolley a , Ian R. Petersen a , Naoki Yamamoto b
a School
of Engineering and Information Technology, University of New South Wales at the Australian Defence Force Academy, Canberra
ACT 2600, Australia
b Department
of Applied Physics and Physico-Informatics, Keio University, Yokohama 223-8522, Japan
Abstract
This paper presents two realizations of linear quantum systems for covariance assignment corresponding to pure Gaussian states. The
first one is called a cascade realization; given any covariance matrix corresponding to a pure Gaussian state, we can construct a cascaded
quantum system generating that state. The second one is called a locally dissipative realization; given a covariance matrix corresponding
to a pure Gaussian state, if it satisfies certain conditions, we can construct a linear quantum system that has only local interactions with
its environment and achieves the assigned covariance matrix. Both realizations are illustrated by examples from quantum optics.
Key words: Linear quantum system, Cascade realization, Locally dissipative realization, Covariance assignment, Pure Gaussian state.
1 Introduction
For stochastic systems, many of the performance objectives
are expressed in terms of the variances (or covariances) of
the system states. In a large space structure, for example,
the vibration at certain points on the structure must be reduced to an acceptable level. This objective in fact involves
keeping the variances of some variables such as deflections
within prescribed bounds. One way to achieve this is to assign an appropriate matrix value to the covariance of the
state vector. This method, referred to as covariance assignment, has been extensively studied in a series of papers by
Skelton and colleagues, e.g., in [1–3]. For linear stochastic systems with white noises, the covariance matrix can be
computed by solving the Lyapunov equation for the system.
In this case, the covariance assignment problem reduces to
designing system matrices such that the corresponding Lyapunov equation has a prescribed solution.
Turning our attention to the quantum case, we find that a
⋆ This work was supported by the Australian Research Council
and JSPS Grant-in-Aid No. 40513289. The material in this paper
was partially presented at the 2014 IEEE Conference on Control Applications (CCA), Oct. 8-10, 2014, Antibes, France. Corresponding author S. Ma. Tel. +61 2 62688818.
Email addresses: [email protected] (Shan Ma),
[email protected] (Matthew J. Woolley),
[email protected] (Ian R. Petersen),
[email protected] (Naoki Yamamoto).
Preprint submitted to Automatica
covariance matrix plays an essential role as well in the field
of quantum information. In particular for a linear quantum
system, the importance of a covariance matrix stands out,
because it can fully characterize the entanglement property,
which is indeed crucial for conducting quantum information
processing [4,5]. Therefore it should be of great use to investigate the covariance assignment problem for linear quantum
systems. In fact, there are several such proposals; [6] studies
a quantum feedback control problem for covariance assignment, and [7–10] analyze systems that generate a pure Gaussian state. Note that, since a Gaussian state (with zero mean)
is uniquely determined by its covariance matrix, the aforementioned covariance assignment problem is also known as
the Gaussian state generation problem; thus, if a linear quantum system achieves a covariance matrix corresponding to a
target Gaussian state, we call that the system generates this
Gaussian state.
Let us especially focus on Refs. [7–10], which provide the
basis of this paper. As mentioned before, in those papers
pure Gaussian states are examined, which are a particularly
important subclass of Gaussian states such that the highest
performance of Gaussian quantum information processing
can be realized [4, 5, 11, 12]. Then they provided several
methods to construct a stable linear quantum system generating a given pure Gaussian state. Moreover, conditions
for generating an arbitrary pure entangled Gaussian state are
given there; surely these are important results, because such
a state serves as an essential resource for Gaussian quan-
8 January 2018
tum information processing tasks. Of course in the literature several methods for generating various pure entangled
Gaussian states have been proposed. For instance, [13] gives
a systematic method to generate an arbitrary pure entangled
Gaussian state; the idea is to construct a coherent process by
applying a sequence of prescribed unitary operations (composed of beam splitters and squeezers in optics case) to an
initial state. Thus this method is essentially a closed-system
approach. In contrast, the approach we take here is an opensystem one; that is, we aim to construct dissipative processes
such that the system is stable and uniquely driven into a desired target pure Gaussian state. This strategy is categorized
into the so-called reservoir engineering method [14–18]; in
general, this approach has a clear advantage that the system
has good robustness properties with respect to initial states
and evolution time.
the cubic-phase gate or photon counting on that extracted
Gaussian state, we can realize, e.g., entanglement distillation
and universal quantum computation [5]. On the other hand, a
generated internal Gaussian state is not necessarily extracted
to outside for the purpose of precision measurement in the
scenario of quantum metrology; for instance a spin squeezed
state of an atomic ensemble can be directly used for ultraprecise magnetometry [30].
Notation. For a matrix A = [A jk ] whose entries A jk are
complex numbers or operators, we define A⊤ = [Ak j ],
A† = [A∗k j ], where the superscript ∗ denotes either the
complex conjugate of a complex number or the adjoint
of an operator. diag[τ1 , · · · , τn ] denotes an n × n diagonal matrix with τ j , j = 1, 2, · · · , n, on its main diagonal. PN is a 2N × 2N permutation matrix defined by
PN [x1 x2 x3 x4 · · · x2N ]⊤ = [x1 x3 · · · x2N−1 x2 x4 · · · x2N ]⊤
for any column vector [x1 x2 x3 x4 · · · x2N ]⊤ .
Now we describe the problem considered in this paper. The
methods developed in [7–10] lead to infinitely many linear
quantum systems that uniquely generate a target pure Gaussian state. Some of these systems are easy to implement,
while others are not. Then a natural question is how to find
a linear quantum system that is simple to implement, while
still uniquely generates the desired pure Gaussian state.
2 Preliminaries
We consider a linear quantum system G of N modes. Each
mode is characterized by a pair of quadrature operators
{q̂ j , p̂ j }, j = 1, 2, · · · , N. Collecting them into an operatorvalued vector x̂ , [q̂1 · · · q̂N p̂1 · · · p̂N ]⊤ , we write the
canonical commutation relations as
In this paper, we provide two convenient realizations of a
linear quantum system generating a target pure Gaussian
state. The first one is a cascade realization, which is a typical
system structure found in the literature [19–21]. We show
that, given any covariance matrix corresponding to a pure
Gaussian state, we can construct a cascaded quantum system
uniquely generating that state. This cascaded system is a
series connection of several subsystems in which the output
of one is fed as the input to the next. A clear advantage of the
cascade realization is that those subsystems can be placed at
remote sites. Note that the cascade structure has also been
widely studied in the classical control literature [22–24].
h
x̂, x̂
⊤
i
⊤
, x̂x̂ − x̂x̂⊤ = iΣ,
⊤
Σ,
"
0 IN
−IN 0
#
.
(1)
Here we emphasize that the transpose operation ⊤, when applied to an operator-valued matrix (say, x̂x̂⊤ ), only exchanges
the indices of the matrix and leaves the entries unchanged.
⊤
Therefore x̂x̂⊤ 6= x̂x̂⊤ . Let Ĥ be the Hamiltonian of the
system, and let {ĉ j }, j = 1, 2, · · · , K, be Lindblad operators
that represent the interactions between the system and its
environment. For convenience, we collect all the Lindblad
h
i⊤
operators as an operator-valued vector L̂ = ĉ1 ĉ2 · · · ĉK
The second one is a locally dissipative realization, which
is motivated by the specific system structure found in, e.g.
[9,25–27]. Note that in these references the notion of quasilocality has been studied, but in this paper we focus on a
stronger notion, locality. Here “locally dissipative” means
that all the system-environment interactions act only on one
system component. Implementations of locally dissipative
systems should be considerably easier than that of systems
which have non-local interactions [28]. In this paper, we
show that, given a covariance matrix corresponding to a
pure Gaussian state, if it satisfies certain conditions, we can
construct a locally dissipative quantum system generating
that state.
and call L̂ the coupling vector. Suppose Ĥ is quadratic in x̂,
i.e., Ĥ = 21 x̂⊤ M x̂, with M = M ⊤ ∈ R2N×2N , and L̂ is linear
in x̂, i.e., L̂ = Cx̂, with C ∈ CK×2N , then the quantum system G can be described by the following quantum stochastic
differential equations (QSDEs)
Lastly we remark that the state generated by our method is
an internal one confined in the system (e.g. an intra-cavity
state in optics), rather than an external optical field state. This
means that, if we aim to perform some quantum information
processing with that Gaussian state, it must be extracted to
outside by for instance the method developed in [29]. In
particular by acting some non-Gaussian operations such as
h
i⊤
d x̂(t) = A x̂(t)dt + B d Â⊤ (t) d † (t) ,
(2)
dŶ (t) = C x̂(t)dt + d Â(t),
where A = Σ(M + Im(C†C)), B = iΣ[−C† C⊤ ], C = C [8],
⊤
[31, Chapter 6]. The input d Â(t) = d Â1 (t) · · · d ÂK (t)
represents K independent quantum stochastic processes,
with d  j (t), j = 1, 2, · · · , K, satisfying the following quan2
tum Itō rules:
(
d  j (t)d Â∗k (t) = δ jk dt,
d  j (t)d Âk (t) = d Â∗j (t)d Â∗k (t) = d Â∗j (t)d Âk (t) = 0,
corresponding to a given pure Gaussian state. Since a pure
Gaussian state (with zero mean) is uniquely specified by its
covariance matrix, so if a linear quantum system achieves a
covariance matrix corresponding to a pure Gaussian state, we
can simply say that such a linear quantum system uniquely
generates the pure Gaussian state. The problem can be expressed mathematically as:
(3)
where δ jk is the Kronecker δ -function. The output
⊤
dŶ (t) = dŶ1 (t) · · · dŶK (t) satisfies quantum Itō rules
similar to (3) [8, 31–35]. The quantum expectation of the
vector x̂ is denoted by hx̂i and the covariance matrix is given
by V = 12 h△x̂△x̂⊤ + (△x̂△x̂⊤ )⊤ i, where △x̂ = x̂ − hx̂i; see,
e.g., [7,8,11]. The time evolutions of the mean vector hx̂(t)i
and the covariance matrix V (t) can be derived from (2) by
using the quantum Itō rule. They are given by
dhx̂(t)i
= A hx̂(t)i,
dt
dV (t) = A V (t) + V (t)A ⊤ + 1 BB † .
dt
2
find
subject to
where V is the covariance matrix corresponding to the desired target pure Gaussian state. Here a matrix A is said to
be Hurwitz if all its eigenvalues have strictly negative real
parts. A system described by (2) is said to be asymptotically stable if the matrix A is a Hurwitz matrix. Recently, a
necessary and sufficient condition has been developed in [7]
for solving the pure Gaussian state covariance assignment
problem. The result is summarized as follows.
(4)
(5)
As in the classical case, a Gaussian state is completely characterized by the mean vector hx̂i and the covariance matrix V . Since the mean vector hx̂i contains no information
about noise and entanglement, we will restrict our attention to zero-mean Gaussian states (i.e., hx̂i = 0). A Gaussian
state is pure if and only if its covariance matrix V satisfies
det(V ) = 2−2N . In fact, when a Gaussian state is pure, its
covariance matrix V can always be factored as
1
V = SS⊤,
2
S=
"
Y
− 12
XY
− 21
0
Y
1
2
#
,
M = M ⊤ ∈ R2N×2N and C ∈ CK×2N
A is Hurwitz,
1
A V + V A ⊤ + BB † = 0,
2
Lemma 1 ( [7, 8]). Let V be the covariance matrix corresponding to a given N-mode pure Gaussian state. Assume
that V is expressed in the factored form (6). Then this pure
Gaussian state is uniquely generated by the linear quantum
system (2) if and only if
M=
(6)
where X = X ⊤ ∈ RN×N , Y = Y ⊤ ∈ RN×N and Y > 0 [11,
36]. For example, the N-mode vacuum state is a special
pure Gaussian state with X = 0 and Y = IN . It can be seen
from (6) that a pure Gaussian state is uniquely specified by
a complex, symmetric matrix Z , X + iY , which is referred
to as the graph corresponding to a pure Gaussian state [11].
Note also that the matrix S satisfies SΣS⊤ = Σ, which means
that S is a symplectic matrix. The symplectic nature of S
guarantees that the mapping x̂ 7−→ x̂′ , Sx̂ preserves the
canonical commutation relations (1), that is
"
XRX + Y RY − ΓY −1 X − XY −1 Γ⊤ −XR + ΓY −1
−RX + Y −1 Γ⊤
R
#
,
(7)
and
C = P⊤ [−Z IN ],
(8)
where R = R⊤ ∈ RN×N , Γ = −Γ⊤ ∈ RN×N , and P ∈ CN×K
are free matrices satisfying the following rank condition
rank
h
i
−1
P QP · · · QN−1 P = N, Q , −iRY + Y Γ. (9)
Remark 2. From (8), we see that the resulting coupling vector L̂ of the engineered system is L̂ = Cx̂ = P⊤ [−Z IN ]x̂ =
P⊤ ([ p̂1 · · · p̂N ]⊤ − Z [q̂1 · · · q̂N ]⊤ ). Therefore, all the components of L̂ are nullifiers for the desired target pure Gaussian state [11]. As a special example, one can engineer a
purely dissipative system (with Ĥ = 0) to generate a pure
Gaussian state. In this case, one could take R = Γ = 0N×N
and P = IN in Lemma 1. Then the resulting coupling vector
L̂ is the so-called nullifier vector for the desired target pure
Gaussian state.
i
i
h
i h
h
x̂′ , x̂′⊤ = Sx̂, (Sx̂)⊤ = S x̂, x̂⊤ S⊤ = S (iΣ) S⊤ = iΣ.
Note that if the system G is initially in a Gaussian state,
then the system G will always be Gaussian, with the mean
vector hx̂(t)i and the covariance matrix V (t) obeying (4)
and (5), respectively. We shall be particularly interested in
the steady-state covariance matrix V (∞).
Assume that the system G is initially in a Gaussian state.
The problem of pure Gaussian state covariance assignment
is to find a Hamiltonian Ĥ and a coupling vector L̂ such that
the corresponding linear quantum system described by (2)
is asymptotically stable and achieves the covariance matrix
Remark 3. Lemma 1 has a simple interpretation in terms
of symplectic transformations [8, 37]. As mentioned before,
vacuum states are a special class of pure Gaussian states.
3
quantum systems that uniquely generate a given pure Gaussian state. Based on this fact, we provide two feasible realizations of linear quantum systems for covariance assignment corresponding to pure Gaussian states, and this section
is devoted to the first one, the cascade realization.
The covariance matrix corresponding to the N-mode vacuum
state is V = 12 IN . By using physical realizability conditions,
it can be proved that the N-mode vacuum state can only be
generated by an N-mode passive linear quantum system [37].
The converse is also true. That is, an N-mode passive linear
quantum system, if it is asymptotically stable, must evolve
toward the N-mode vacuum state [38]. Recall that for a
passive linear quantum system, the Hamiltonian is always of
"
#
R̃ Γ̃
1 ⊤
, R̃ = R̃⊤ ∈ RN×N ,
the form Ĥ = 2 x̂ M̃ x̂, with M̃ =
Γ̃⊤ R̃
and Γ̃ = −Γ̃⊤ ∈ RN×N , and the coupling vector is always of
the form L̂ = C̃x̂, with C̃ = P̃⊤ [−iIN IN ], P̃ ∈ CN×K [21,39–
41]. Now we apply a symplectic transformation to x̂, that is,
we define x̂′ , Sx̂. Then, in terms of x̂′ , the Hamiltonian is
rewritten as Ĥ = 21 x̂′⊤ S−⊤ M̃S−1 x̂′ and the coupling vector
is rewritten as L̂ = C̃S−1 x̂′ . We also observe that the relation
between the covariance matrix V ′ of x̂′ and the covariance
matrix V of x̂ is given as follows:
3.1 The cascade realization
For convenience, we denote a linear quantum system G with
the Hamiltonian Ĥ and the coupling vector L̂ as G = (Ĥ, L̂).
Suppose we have two linear quantum systems G1 = (Ĥ1 , L̂1 )
and G2 = (Ĥ2 , L̂2 ). If we feed the output of the system
G1 into the input of the system G2 , we will obtain a cascaded quantum system G = G2 ✁ G1 , as shown in Fig. 1.
Based on the quantum theory of cascaded linear quantum
Â(t)
Ŷ (t)
systems [43], the Hamiltonian Ĥ and the coupling vector L̂
of the cascaded system G are, respectively, given by
Ĥ = Ĥ + Ĥ + 1 L̂† L̂ − L̂† L̂ ,
2
1
1 2
2 1
2i
L̂ = L̂2 + L̂1 .
If the passive linear quantum system is asymptotically stable,
then based on the result in [38], we have V → 12 IN , as t →
+∞. As a result, V ′ → 12 SS⊤, which gives the desired pure
Gaussian state. Combining the results above, we conclude
that for a given pure Gaussian state V = 12 SS⊤, a complete parametrization of the linear quantum system G that
uniquely generates this pure Gaussian state is given by
M = S−⊤ M̃S−1 ,
C = C̃S−1 ,
G2
G
Fig. 1. The cascade connection of two linear quantum systems:
G = G2 ✁ G1 .
1
V ′ = h△x̂′ △x̂′⊤ + (△x̂′ △x̂′⊤ )⊤ i
2
1
= Sh△x̂△x̂⊤ + (△x̂△x̂⊤ )⊤ iS⊤
2
= SV S⊤ .
G1
(12)
This result can be extended to the cascade connection of N
one-dimensional harmonic oscillators. Suppose we have N
one-dimensional harmonic oscillators G j with the Hamilto2×2 , ξ̂ , [q̂ p̂ ]⊤ , and
nian Ĥ j = 21 ξ̂ j⊤ M j ξ̂ j , M j = M ⊤
j
j j
j ∈R
the coupling vector L̂ j = C j ξ̂ j , C j ∈ CK×2 , j = 1, 2, · · · , N.
The system G is obtained by a cascade connection of these
harmonic oscillators, that is, G = GN ✁ · · · ✁ G2 ✁ G1 , as
shown in Fig. 2. By repeatedly using (12), the Hamiltonian
Ĥ and the coupling vector L̂ of the cascaded system G are
given by the following lemma.
(10)
(11)
where (M̃, C̃) form an asymptotically stable passive lin"
#
R̃ Γ̃
ear quantum system. Substituting M̃ =
and C̃ =
Γ̃⊤ R̃
P̃⊤ [−iIN , IN ] into (10), (11) and using some additional matrix transformations, we will obtain the formulas (7), (8),
respectively. This is the idea behind Lemma 1. The rank
constraint (9) indeed gives a sufficient and necessary stability condition for the original passive linear quantum system [40, 41]. As a result, it also guarantees the stability of
the linear quantum system G based on the linear transformation theory in the control field [42].
Â(t)
G1
G2
G
GN
Ŷ (t)
Fig. 2. The cascade connection of N one-dimensional harmonic
oscillators: G = GN ✁ · · · ✁ G2 ✁ G1 .
Lemma 4 ( [20]). Suppose that the system G is obtained
via a cascade connection of the aforementioned N onedimensional harmonic oscillators G j , j = 1, 2, · · · , N, that
is, G = GN ✁ · · ·✁ G2 ✁ G1 . Then the Hamiltonian Ĥ and the
coupling vector L̂ of the system G are, respectively, given by
3 The cascade realization
Ĥ = 1 x̂⊤ M x̂, M = P MP ⊤
N
N
2
L̂ = Cx̂, C = [C1 C2 · · · CN ] PN⊤ ,
As we have seen in Lemma 1, the matrices R, Γ and P
are free matrices, although they must satisfy the rank condition (9). By varying them we can obtain different linear
4
where M = [M jk ] j,k=1,··· ,N is a symmetric block matrix with
M j j = M j , M jk = Im(C†j Ck ) whenever j > k and M jk = M⊤
kj
whenever j < k.
The stability of A and the Lyapunov equation (13) guarantee that the cascaded system G constructed above is asymptotically stable and achieves the covariance matrix V . In
other words, the cascaded system G uniquely generates the
desired target pure Gaussian state.
It can be seen from Lemma 4 that due to the cascade feature, the Hamiltonian matrix M and the coupling matrix C
of the cascaded system G depend on each other in a complicated way. Nevertheless, given any pure Gaussian state, we
can always construct a cascade connection of several onedimensional harmonic oscillators such that this cascaded
quantum system is asymptotically stable and achieves the
covariance matrix corresponding to the desired target pure
Gaussian state. The result is stated as follows.
3.2 Example
Example 6. We consider the generation of two-mode
squeezed states [11]. Two-mode squeezed states are highly
symmetric entangled states, which are very useful in several
quantum information protocols such as quantum teleportation [44]. The covariance matrix V corresponding to a
two-mode squeezed state is
0
0
cosh(2α ) sinh(2α )
0
0
1
sinh(2α ) cosh(2α )
V=
,
2
0
0
cosh(2α ) − sinh(2α )
0
0
− sinh(2α ) cosh(2α )
(14)
Theorem 5. Any N-mode pure Gaussian state can be
uniquely generated by constructing a cascade of N onedimensional harmonic oscillators.
Proof. We prove this result by construction. Recall that for
an arbitrary N-mode pure Gaussian state, the corresponding
covariance matrix V has the factorization shown in (6). Using the matrices X and Y obtained from (6), we construct a
cascaded system G = GN ✁ · · · ✁ G2 ✁ G1 with the Hamiltonian Ĥ j and the coupling vector L̂ j , j = 1, 2, · · · , N, given by
Ĥ j = 0,
0(2 j−2)×2
where α is the squeezing parameter. Using the factoriza#
"
cosh(2α ) − sinh(2α )
.
tion (6), we have X = 0 and Y =
− sinh(2α ) cosh(2α )
Therefore, the graph corresponding to a two-mode squeezed
"
#
i cosh(2α ) −i sinh(2α )
state is given by Z = X + iY =
.
−i sinh(2α ) i cosh(2α )
1
.
L̂ j = C j ξ̂ j , C j = iY − 2 [−Z IN ] PN
I2
0(2N−2 j)×2
Next we provide two different cascade realizations. The first
one, Realization 1, is constructed based on a heuristic derivation, while the second one, Realization 2, is constructed
based on the proof of Theorem 5.
Using Lemma 4, we can calculate the Hamiltonian Ĥ =
1 ⊤
2 x̂ M x̂ and the coupling vector L̂ = Cx̂ for the cascaded
system G. We find that M = 0 and C = iY −1/2 [−Z IN ]. Then
it follows from the QSDE (2) that
Realization 1. In this cascade realization, the subsystems
G1 = (Ĥ1 , L̂1 ) and G2 = (Ĥ2 , L̂2 ) are, respectively, given by
A = Σ(M + Im(C†C))
"
#!
(X − iY )Y −1 (X + iY ) −(X − iY )Y −1
= Σ Im
−Y −1 (X + iY )
Y −1
"
#
2 Q1
1
⊤
Ĥ = ξ̂
ξ̂1 , L̂1 = [iQ2 1]ξ̂1 ,
1 2 1 Q1 2
"
#
2
Q
1
1
⊤
Ĥ2 = − ξ̂
ξ̂2 , L̂2 = [iQ2 1]ξ̂2 ,
2 2 Q1 2
= ΣΣ = −I2N ,
1
D , BB † = Σ Re(C†C)Σ⊤
2
"
#!
(X − iY )Y −1 (X + iY ) −(X − iY )Y −1
= Σ Re
Σ⊤
−Y −1 (X + iY )
Y −1
"
#
Y −1
Y −1 X
.
=
XY −1 XY −1 X + Y
(2α )
− sinh(2α ) and Q2 , sinh(2α ) −
where Q1 , sinh
cosh(2α )
cosh(2α ). It can be proved that the cascaded system
G = G2 ✁ G1 is asymptotically stable and achieves the
covariance matrix (14). The proof is similar to that of Theorem 5, and hence is omitted. Using the result in [45], a
corresponding quantum optical realization is provided in
Fig. 3. For each subsystem G j , j = 1, 2, the Hamiltonian
Ĥ j is realized by a nonlinear crystal pumped by a classical
field, and the coupling operator L̂ j is realized by implementing an auxiliary cavity. This auxiliary cavity interacts
2
Clearly, A is Hurwitz. Furthermore, it can be verified that
A V + V A ⊤ + D = 0.
(13)
5
^
with the subsystem via a cascade of a pumped crystal and
a beam splitter. It has a fast mode that can be adiabatically
eliminated.
Y(t)
eiπ
eiπ
Auxiliary
cavity
e
Auxiliary
cavity
^
G2
A 1(t)
Auxiliary
cavity
Auxiliary
cavity
G2
iπ
a^1
^
^
A(t)
G1
Y1(t)
G1
^
eiπ
a^ 2
A 2(t)
^
Y2 (t)
^
a^ 2
a1
Fig. 4. Another optical cascade realization of the two-mode linear
quantum system that uniquely generates a two-mode squeezed
state.
Fig. 3. An optical cascade realization of the two-mode linear
quantum system that uniquely generates a two-mode squeezed
state. The square with an arrow represents a pumped crystal. The
symbol eiπ with a square on it represents a phase shift π . Solid
(dark) rectangles denote perfectly reflecting mirrors, while unfilled
rectangles denote partially transmitting mirrors. The dark line “ ”
represents an optical beam splitter.
locally dissipative realization cannot generate all pure Gaussian states, but as shown later the class of stabilizable states
is fairly broad.
4.1 The locally dissipative realization
Realization 2. The second realization is constructed according to the method shown in the proof of Theorem 5.
By direct calculation, the subsystems G1 = (Ĥ1 , L̂1 ) and
G2 = (Ĥ2 , L̂2 ) are, respectively, given by
Ĥ = 0,
1
Ĥ2 = 0,
L̂1 =
L̂2 =
"
#
cosh(α ) i cosh(α )
− sinh(α ) i sinh(α )
"
#
− sinh(α ) i sinh(α )
cosh(α ) i cosh(α )
As we have noted in Section 2, the coupling vector L̂ is
an operator-valued vector that consists of K elements, i.e.,
L̂ = [ĉ1 ĉ2 · · · ĉK ]⊤ . Each element ĉ j , j = 1, 2, · · · , K, called
a Lindblad operator, represents an interaction between the
system and its environment. A Lindblad operator ĉ j is said
to be local if it acts only on one system mode. As an example, consider the system depicted in Fig. 5. The Lindblad
operator ĉ1 = q̂1 + p̂1 acts only on the first system mode, so
it is a local operator. On the other hand, the Lindblad operator ĉ2 = q̂1 + q̂2 acts on two system modes, so by definition
it is not a local operator. If all the Lindblad operators in L̂ are
local, then the system is called a locally dissipative quantum
system. A locally dissipative quantum system could be relatively easy to implement in practice. Therefore, we would
like to characterize the class of pure Gaussian states that
can be generated using locally dissipative quantum systems.
The result is given by the following theorem.
ξ̂1 ,
ξ̂2 .
Using the result in [45], a corresponding quantum optical
realization of such a cascaded quantum system G = G2 ✁ G1
is provided in Fig. 4. This cascaded system G has two crucial features. First, because Ĥ1 = Ĥ2 = 0, implementations
of the Hamiltonians involve no pumped crystals. Second,
the first component of the coupling vector L̂1 = [ĉ1,1 ĉ1,2 ]⊤
" #
h
i q̂
√
1
is ĉ1,1 = cosh(α ) i cosh(α )
= 2 cosh(α )â1 , where
p̂1
√
â1 = (q̂1 + i p̂1 )/ 2 denotes the annihilation operator of the
first mode. This operator ĉ1,1 represents the standard linear dissipation of a cavity mode into a continuum of field
modes outside of the cavity. A similar case also occurs in the
coupling vector L̂2 . As can be seen in Fig. 4, Realization 2
requires two pumped crystals, in contrast to the case of Realization 1, where four pumped crystals are used. From this
viewpoint, Realization 2, which is constructed based on our
result, has a clear advantage over Realization 1.
1
2
c^ 2= q^ 1 + q^ 2
c^ 1= q^ 1 + p^ 1
Fig. 5. An illustration of local Lindblad operators. ĉ1 is a local
Lindblad operator, while ĉ2 is not a local one.
Theorem 7. Let V be the covariance matrix corresponding
to a given N-mode pure Gaussian state. Assume that it is
expressed in the factored form (6). Then this pure Gaussian
state can be uniquely generated in an N-mode locally dissipative quantum system if and only if there exists an integer
ℓ, 1 ≤ ℓ ≤ N, such that
4 The locally dissipative realization
Z(ℓ, j) = Z( j,ℓ) = 0,
In this section, we describe the second realization of linear
quantum systems for covariance assignment corresponding
to pure Gaussian states. Unlike the cascade realization, the
∀ j 6= ℓ and 1 ≤ j ≤ N,
(15)
where Z(ℓ, j) denotes the (ℓ, j) element of the graph matrix
Z = X + iY for the pure Gaussian state.
6
Proof. We prove the sufficiency part by construction.
Equation h (15) implies that ithere exists a row vector ϒ = 01×(ℓ−1) τ1 01×(N−ℓ) with τ1 6= 0, such that
i
h
ϒZ = 01×(ℓ−1) τ2 01×(N−ℓ) , where τ2 = τ1 Z(ℓ,ℓ) . Using
the Gram-Schmidt method, we can create three N × N matrices U1 , U2 and Λ, where U1 is a unitary matrix with the
1/2 ⊤
first column being YY 1/2 ϒϒ⊤ ; U2 is a unitary matrix with
k
k
h
i⊤
the first column being √1N 1 1 · · · 1 and Λ is a purely
imaginary matrix Λ = i diag[α1 , · · · , αN ], with α j ∈ R,
j = 1, · · · , N, and α j 6= αk , ∀ j 6= k.
operator ĉk = Pk⊤ [−Z IN ] x̂ is local. Suppose that ĉk acts on
the ℓth mode of the system. Then we have
Let P = ϒ⊤ , R = −Y −1/2 Im(U1U2† ΛU2U1† )Y −1/2 and Γ =
Y 1/2 Re(U1U2† ΛU2U1† )Y 1/2 in (7), (8). Then it can be verified
that R = R⊤ , Γ = −Γ⊤ . Moreover, substituting Q = −iRY +
Y −1 Γ = Y −1/2U1U2† ΛU2U1†Y 1/2 into (9) yields
Substituting (17) into (16) gives
h
i
P QP · · · QN−1 P
h
= rank U2U1†Y 1/2 ϒ⊤ ΛU2U1†Y 1/2 ϒ⊤ · · ·
Pk⊤ [−Z IN ]
h
i
= 01×(ℓ−1) τ3 01×(N−ℓ) 01×(ℓ−1) τ4 01×(N−ℓ) ,
where τ3 and τ4 are complex numbers. It follows that
i
h
Pk⊤ Z = 01×(ℓ−1) −τ3 01×(N−ℓ) ,
i
h
Pk⊤ = 01×(ℓ−1) τ4 01×(N−ℓ) , τ4 6= 0.
Y 1/2 ϒ⊤
= rank √
N
=N.
1
1
.
..
1
(17)
h
i
Pk⊤ Z =τ4 Z(ℓ,1) Z(ℓ,2) · · · Z(ℓ,N)
h
i
= 01×(ℓ−1) −τ3 01×(N−ℓ) .
rank
ΛN−1U2U1†Y 1/2 ϒ⊤
(iα1 ) · · · (iα1 )N−1
(iα2 ) · · · (iα2 )N−1
..
..
. ···
.
(iαN ) · · · (iαN )N−1
(16)
Since τ4 6= 0, we have Z(ℓ, j) = 0, ∀ j 6= ℓ. Since Z = Z ⊤ , we
have Z(ℓ, j) = Z( j,ℓ) = 0, ∀ j 6= ℓ. That is, Equation (15) holds.
This completes the proof.
i
Remark 8. The basic idea of Theorem 7 is that for any
choice of P 6= 0, there always exist matrices R = R⊤ and
Γ = −Γ⊤ such that the rank condition (9) is satisfied. So we
can first specify a matrix P such that the coupling matrix C
in (8) has a local structure. After obtaining P, we determine
the other two matrices R and Γ to get a system Hamiltonian,
under the rank constraint (9). Generally, for a given nonzero
matrix P, we have infinite solutions (R, Γ) that satisfy the
rank condition (9). Different choices of (R, Γ) lead to different system Hamiltonians. The optimization problem over
these Hamiltonians is beyond the scope of this paper and is
not considered, but in the next subsection we will show a
specific recipe for determining those matrices (R, Γ).
Here we have used the full rank property of a Vandermonde
matrix. Hence the rank condition (9) is satisfied. Based on
Lemma 1, we now obtain an N-mode locally dissipative
quantum system that is asymptotically stable and achieves
the given covariance matrix. The coupling vector L̂ of the
system, which consists of only one Lindblad operator, is
given by
Remark 9. Suppose an N-mode pure Gaussian state is generated in an N-mode dissipative quantum system and the
ℓth mode is locally coupled to the environment. Then from
Equation (15), it is straightforward to see that the ℓth mode
is not entangled with the rest of the system modes when the
system achieves the steady state.
C = P⊤ [−Z IN ] = [−ϒZ ϒ]
h
i
= 01×(ℓ−1) −τ2 01×(N−ℓ) 01×(ℓ−1) τ1 01×(N−ℓ) ,
L̂ = Cx̂ = −τ2 q̂ℓ + τ1 p̂ℓ .
4.2 Examples
We see that L̂ acts only on the ℓth mode, and hence it is
local. The Hamiltonian Ĥ of the system can also be obtained
by directly substituting the matrices R and Γ above into (7).
This completes the sufficiency part of the proof.
Example 10. We consider the generation of canonical Gaussian cluster states, which serve as an essential
resource in quantum computation with continuous variables [4,11,12]. We mention that an interesting class of cluster states, called bilayer square-lattice continuous–variable
cluster states, has been proposed recently in [46]. This class
of cluster states has some practical advantages over canonical Gaussian cluster states for quantum computation [46].
For the sake of simplicity, we use canonical Gaussian cluster states to illustrate the developed theory. The covariance
Next we prove the necessity part. Suppose an N-mode pure
Gaussian state can be uniquely generated in an N-mode locally dissipative quantum system. Based on Lemma 1, there
exists a P ∈ CN×K , P 6= 0, such that the coupling vector
L̂ = P⊤ [−Z IN ] x̂ is local. Let Pk , 1 ≤ k ≤ K, be a nonzero
column in the matrix P. Then the corresponding Lindblad
7
(BS)
where Ĥ jk = (q̂ j p̂k − p̂ j q̂k ) = i(â j â∗k − â∗j âk ), where â j =
√
√
(q̂ j + i p̂ j )/ 2 and â∗j = (q̂ j − i p̂ j )/ 2, is the Hamiltonian
representing the coupling between the jth and kth optical
modes at a beam splitter. Also the coupling vector is given
by
matrix V corresponding to an N-mode canonical Gaussian
"
#
2α B
2α I
e
e
N
cluster state is given by V = 12
,
e2α B e−2α IN + e2α B2
where B = B⊤ ∈ RN×N and α is the squeezing parameter.
Note that in the limit α → ∞, the canonical Gaussian cluster state approximates the corresponding ideal cluster state.
Using (6), we obtain X = B and Y = e−2α IN . The graph
corresponding to a canonical Gaussian cluster state is given
by Z = X + iY = B + ie−2α IN .
√
L̂ = −( 2 + e−2α i)q̂4 + p̂4,
which acts only on the fourth mode and hence it is local.
Finally, using the result in [45], a corresponding optical realization of this linear quantum system is shown in Fig. 6.
Note that three pumped crystals are used; we conjecture that
this is the minimum number required for constructing a desired locally dissipative system.
Let us consider a simple case where
0
1
X =
0
0
10 0
01 0
,
10 0
√
00 2
Y = e−2α I4 .
(18)
a^ 2
a1
These matrices satisfy X4 j = 0 and Y4 j = 0 for all j 6= 4. Thus
by Theorem 7, the corresponding canonical Gaussian cluster state can be generated in a f our-mode locally dissipative
system. To construct such a system, let us take P = [0 0 0 1]⊤
in Lemma 1. The next step is to determine the other system
parameters R and Γ. For a practical implementation, one of
the basic requirements on the system is that, as mentioned
before, the system has as few pumped crystals as possible.
Motivated by the structure of the passive quantum systems
described in Remark 3, we choose
R = 04×4. As a result,
"
# the
2α
⊤
2α
−e (ΓX + XΓ ) e Γ
Hamiltonian matrix is M =
. The
04×4
e2α Γ⊤
(1, 2) block in M is a skew matrix e2α Γ. So if we can additionally take the (1, 1) block to be a diagonal matrix, then
the interaction Hamiltonian between the modes is passive
and can be simply realized by beam splitters. According to
this guideline, we now seek Γ such that −e2α (ΓX + XΓ⊤ )
is a diagonal matrix. By direct calculation, we obtain
γ1
a^ 4
0
iπ
e
^
A(t)
Fig. 6. The optical dissipative system that uniquely generates the
canonical Gaussian cluster state (18). The coupling vector L̂ acts
only on the fourth mode, and hence it is local.
Example 11. We next consider a canonical Gaussian cluster
state specified by the following matrices X and Y :
0
1
X =
0
0
1 0 0
0 1 0
,
1 0 1
0 1 0
Y = e−2α I4 .
(19)
The strength of Theorem 7 is that it readily tells us that this
canonical Gaussian cluster state cannot be generated in any
four-mode locally dissipative system. Nonetheless let us take
the same matrix P as before, i.e., P = [0 0 0 1]⊤ , and follow
the same guideline as discussed in Example 10. That is, we
set R = 04×4 and seek Γ such that ΓX + XΓ⊤ is a diagonal
matrix. Then, again by direct calculation, we find
where γ1 ∈ R and γ2 ∈ R. Substituting the matrices P, R and
Γ above into the rank condition (9), we obtain that if γ1 γ2 6= 0,
the resulting linear quantum system is asymptotically stable
and achieves the covariance matrix corresponding to (18).
The Hamiltonian of this linear quantum system is now determined as
0
γ1
0
γ2
−γ
γ1 + γ2 0
0
1
Γ=
.
0 − (γ1 + γ2 )
γ
0
1
0
−γ1 0
−γ2
Ĥ = −γ1 e2α q̂21 + γ1 e2α q̂23 + e2α γ1 (Ĥ12 + Ĥ23 )
√ (BS)
(BS)
(BS)
+ e2α γ2 (Ĥ14 + 2Ĥ24 + Ĥ34 ),
(BS)
Auxiliary
cavity
^
Y (t)
γ2
√
−γ
0
γ1
2γ2
1
Γ=
,
0
γ2
−γ1
0
√
−γ2 − 2γ2 −γ2 0
0
a^ 3
^
(BS)
8
Let us take γ1 = 1 and γ2 = 0. Then the corresponding system
Hamiltonian is given by
Ĥ = −e2α (q̂21 − q̂24 ) + e2α (Ĥ12
(BS)
(BS)
+ Ĥ23
Remark 12. The method in [9] is based on essentially
the same idea; given an N-mode pure Gaussian state with
graph Z = X + iY , instead of generating it directly, we enlarge the system by adding a single-mode auxiliary system and then specify the target state as X̃ = diag[X, λ ] and
Ỹ = diag[Y, 1]. By Theorem 7, this (N + 1)-mode target state
can be uniquely generated in an (N + 1)-mode locally dissipative system. The original N-mode pure Gaussian state is
then obtained as a reduced state of the target state.
(BS)
+ Ĥ34 ). (20)
It can be verified that the rank condition (9) is satisfied,
hence the system constructed here is asymptotically stable
and achieves the desired covariance matrix corresponding
to (19), though in this case the system needs to have the
following non-local interaction with its environment:
5 Conclusion
h
i
L̂ = 0 0 −1 −e−2α i 0 0 0 1 x̂ = −q̂3 − ie−2α q̂4 + p̂4 .
In this paper, we have provided two feasible realizations
of linear quantum systems for covariance assignment corresponding to pure Gaussian states: a cascade realization
and a locally dissipative realization. First, we have shown
that given any covariance matrix corresponding to a pure
Gaussian state, we can construct a cascaded quantum system that achieves the assigned covariance matrix. This cascaded quantum system is constructed as a cascade connection of several one-dimensional harmonic oscillators, without any direct interaction Hamiltonians between these oscillators. Second, we have given a complete characterization of
the class of pure Gaussian states that can be generated using
locally dissipative quantum systems. In particular, we have
shown a specific recipe for constructing a system having a
relatively simple Hamiltonian coupling between the system
modes. The results developed in this paper are potentially
useful for the preparation of pure Gaussian states. In the examples, we have provided realizations of (Ĥ, L̂) in quantum
optics using the result in [45]. The circuit figures shown in
the examples are not necessarily the simplest realizations in
quantum optics. Also, a system with (Ĥ, L̂) could be realized by other instances of linear quantum systems such as
atomic ensembles and optomechanical systems [18, 47].
An optical realization, which yet contains an abstract component corresponding to this non-local interaction, is depicted
in Fig. 7. A practical implementation of the non-local inter^
Y (t)
^
A(t)
a^ 1
a^ 3
a^ 4
a^ 2
Fig. 7. The optical linear quantum system that uniquely generates
the canonical Gaussian cluster state (19). The coupling vector L̂
acts on the third and fourth modes, and hence it is not local.
action depicted in Fig. 7 could be experimentally difficult.
Nonetheless this issue can be resolved by taking the following method: add an auxiliary system with a single mode
x̂A = [q̂A p̂A ]⊤ , and specify the target canonical Gaussian
cluster state as
0
1
X̃ = 0
0
1 00 0
0 1 0 0
1 0 1 0,
0 1 0 0
0 0 00 λ
Appendix
Ỹ = e−2α I5 .
Here we briefly review the synthesis theory of linear quantum systems in quantum optics developed in [45].
(21)
1. Realization of a quadratic Hamiltonian
Suppose a quadratic Hamiltonian is given by Ĥd = 21 ξ̂ ⊤ Md ξ̂ ,
where ξ̂ = [q̂ p̂]⊤ and Md = Md⊤ ∈ R2×2 . This Hamiltonian
can be realized by placing a crystal with a classical pump
inside an optical cavity, as shown in Fig. 8. Working in the
frame rotating at half the pump frequency, the Hamiltonian
is written as
Since X̃5 j = 0 and Ỹ5 j = 0 for all j 6= 5, by Theorem 7,
we can construct a five-mode locally dissipative system that
uniquely generates the above canonical Gaussian cluster
state (21). By choosing P = [0 0 0 0 1]⊤ and then taking a
similar procedure as in the case of Example 10, we can obtain such a desired locally dissipative quantum system. Now
we obtain an important observation: for an N-mode canonical Gaussian cluster state with the graph matrix X = B and
the squeezing matrix Y = e−2α IN , it is always possible to
generate this state in a locally dissipative quantum system
by adding a single-mode auxiliary system and specifying the
target state as X̃ = diag[B, λ ] and Ỹ = e−2α IN+1 .
i
Ĥr = △â∗ â + ε (â∗ )2 − ε ∗ â2
" 2
#
1 ⊤ △ − Im(ε ) Re(ε )
△
= ξ̂
ξ̂ − ,
2
2
Re(ε ) △ + Im(ε )
9
(22)
where △ = ωcav − ω p/2 is the detuning between the cavity
mode frequency and the half pump frequency. ε is a measure
of the effective pump intensity [34]. From (22), we see that
by choosing the values of △ and ε , one can make Ĥr =
△
Ĥd − △
2 . Note that the constant term − 2 does not affect
the dynamics of a linear quantum system, and hence can
be ignored. Therefore, the desired Hamiltonian Ĥd can be
realized in this scheme.
a^
a^ 1
(r2 , t 2)
a^ 2
a^ 4
( r1 , t 1 )
Â(t)
a^ 3
Ŷ(t)
Fig. 9. A beam-splitter-like interaction Hamiltonian can be realized
by placing a beam splitter for the two incoming modes â1 and â2 .
Fig. 8. A quadratic Hamiltonian can be realized by placing a
crystal with a classical pump inside an optical cavity.
the frame rotating at half the pump frequency, the interaction
Hamiltonian is written as
2. Realization of a beam-splitter-like interaction Hamiltonian
Ĥab =
where ε1 determines the effective pump intensity and ε2
determines the parameters of the beam splitter. Assume that
the coupling coefficient γ of the partially transmitting mirror
is large so that the mode b̂ is heavily damped and can be
adiabatically eliminated. Then after elimination of b̂, the
resulting coupling operator is given by
Suppose a Hamiltonian is given by Ĥd = hd â∗1 â2 + h∗d â∗2 â1 ,
where hd ∈ C. This Hamiltonian can be realized by implementing a beam splitter for the two incoming modes â1 and
â2 , as shown in Fig. 9. At the beam splitter, we have the
following transformations
" #
â3
â4
=
"
t2 r1
r2 t1
#" #
â1
â2
i
i
ε1 â∗ b̂∗ − ε1∗ âb̂ + ε2 â∗ b̂ − ε2∗ âb̂∗ ,
2
2
1
L̂r = √ (−ε2∗ â + ε1 â∗ ).
γ
,
From (24), we see that by choosing the values of ε1 , ε2 ,
and γ with γ being large, we can make L̂r = L̂d . That is, the
desired coupling operator L̂d can be realized in this scheme.
See [45] for details.
where â3 and â4 denote the outgoing modes, and r1 , t1 ∈ C
denote the (complex) reflectance and transmittance of the
beam splitter, respectively. Note that r1 , t1 , r2 and t2 satisfy
the following relations: |r2 | = |r1 |, |t2 | = |t1 |, |r1 |2 + |t1 |2 = 1,
r1∗t2 + r2t1∗ = 0, and r1∗t1 + r2t2∗ = 0 [48]. Let us parametrize
them as r1 = e−iφ sinθ , r2 = −eiφ sinθ , and t1 = t2 = cosθ .
Then the interaction Hamiltonian Ĥ (BS) for this beam splitter
is given by
Ĥ (BS) = iθ e−iφ â∗1 â2 − iθ eiφ â∗2 â1 .
(24)
Â(t)
ei
π
^
A
c
Y (t)
^
(23)
a^
From (23), we see that by choosing the values of θ and φ , one
can make Ĥ (BS) = Ĥd . That is, the desired beam-splitter-like
interaction Hamiltonian Ĥd can be realized in this scheme.
b
Fig. 10. Realization of a dissipative coupling operator L̂.
3. Realization of a dissipative coupling L̂
References
c1√
−ic2
2
To realize a coupling operator L̂d = c1 q̂+c2 p̂ =
â+
c1√
+ic2
â∗ , we consider the configuration shown in Fig. 10.
2
The configuration consists of a ring cavity with mode â and
an auxiliary ring cavity with mode b̂. The cavity modes â
and b̂ interact through a crystal pumped by a classical beam,
and a beam splitter. The frequency of the auxiliary cavity
mode b̂ is matched to half the pump frequency. Working in
[1] A. Hotz and R. E. Skelton, “Covariance control theory,” International
Journal of Control, vol. 46, no. 1, pp. 13–32, 1987.
[2] J. E. G. Collins and R. E. Skelton, “A theory of state covariance
assignment for discrete systems,” IEEE Transactions on Automatic
Control, vol. 32, no. 1, pp. 35–41, 1987.
[3] R. E. Skelton and M. Ikeda, “Covariance controllers for linear
continuous-time systems,” International Journal of Control, vol. 49,
no. 5, pp. 1773–1785, 1989.
10
[23] S. Huang, M. R. James, and Z. P. Jiang, “L∞ -bounded robust control
of nonlinear cascade systems,” Systems & Control Letters, vol. 54,
no. 3, pp. 215–224, 2005.
[4] S. L. Braunstein and A. K. Pati, Quantum Information with
Continuous Variables. Springer, 2003.
[5] C. Weedbrook, S. Pirandola, R. Garcı́a-Patrón, N. J. Cerf, T. C.
Ralph, J. H. Shapiro, and S. Lloyd, “Gaussian quantum information,”
Reviews of Modern Physics, vol. 84, no. 2, pp. 621–669, 2012.
[24] L. Liu and J. Huang, “Global robust stabilization of cascadeconnected systems with dynamic uncertainties without knowing the
control direction,” IEEE Transactions on Automatic Control, vol. 51,
no. 10, pp. 1693–1699, 2006.
[6] K. Ohki, S. Hara, and N. Yamamoto, “On quantum-classical
equivalence for linear systems control problems and its application
to quantum entanglement assignment,” in Proceedings of IEEE 50th
Annual Conference on Decision and Control (CDC), December 2011,
pp. 6260–6265.
[25] B. Kraus, H. P. Büchler, S. Diehl, A. Kantian, A. Micheli, and
P. Zoller, “Preparation of entangled states by quantum Markov
processes,” Physical Review A, vol. 78, no. 4, p. 042307, 2008.
[26] M. Rafiee, C. Lupo, H. Mokhtari, and S. Mancini, “Stationary and
uniform entanglement distribution in qubit networks with quasilocal
dissipation,” Physical Review A, vol. 85, p. 042320, 2012.
[7] K. Koga and N. Yamamoto, “Dissipation-induced pure Gaussian
state,” Physical Review A, vol. 85, no. 2, p. 022103, 2012.
[8] N. Yamamoto, “Pure Gaussian state generation via dissipation: a
quantum stochastic differential equation approach,” Philosophical
Transactions of the Royal Society A: Mathematical, Physical and
Engineering Sciences, vol. 370, no. 1979, pp. 5324–5337, 2012.
[27] F. Ticozzi and L. Viola, “Stabilizing entangled states with quasi-local
quantum dynamical semigroups,” Philosophical Transactions of the
Royal Society A: Mathematical, Physical and Engineering Sciences,
vol. 370, no. 1979, pp. 5259–5269, 2012.
[9] Y. Ikeda and N. Yamamoto, “Deterministic generation of Gaussian
pure states in a quasilocal dissipative system,” Physical Review A,
vol. 87, no. 3, p. 033802, 2013.
[28] H. A. Bachor and T. C. Ralph, A Guide to Experiments in Quantum
Optics. Wiley, 2004.
[29] T. Tufarelli, A. Ferraro, A. Serafini, S. Bose, and M. S. Kim,
“Coherently opening a high-Q cavity,” Physical Review Letters, vol.
112, no. 13, p. 133605, 2014.
[10] S. Ma, M. J. Woolley, I. R. Petersen, and N. Yamamoto,
“Preparation of pure Gaussian states via cascaded quantum systems,”
in Proceedings of IEEE Conference on Control Applications (CCA),
October 2014, pp. 1970–1975.
[30] G. Tóth and I. Apellaniz, “Quantum metrology from a quantum
information science perspective,” Journal of Physics A: Mathematical
and Theoretical, vol. 47, no. 42, p. 424006, 2014.
[11] N. C. Menicucci, S. T. Flammia, and P. van Loock, “Graphical
calculus for Gaussian pure states,” Physical Review A, vol. 83, no. 4,
p. 042335, 2011.
[31] H. M. Wiseman and G. J. Milburn, Quantum Measurement and
Control. Cambridge University Press, 2010.
[12] N. C. Menicucci, P. van Loock, M. Gu, C. Weedbrook, T. C. Ralph,
and M. A. Nielsen, “Universal quantum computation with continuousvariable cluster states,” Physical Review Letters, vol. 97, no. 11, p.
110501, 2006.
[32] R. L. Hudson and K. R. Parthasarathy, “Quantum Ito’s formula and
stochastic evolutions,” Communications in Mathematical Physics,
vol. 93, no. 3, pp. 301–323, 1984.
[13] G. Adesso, “Generic entanglement and standard form for N-mode
pure Gaussian states,” Physical Review Letters, vol. 97, p. 130502,
2006.
[33] V. P. Belavkin, “Quantum stochastic calculus and quantum nonlinear
filtering,” Journal of Multivariate Analysis, vol. 42, no. 2, pp. 171–
201, 1992.
[14] J. I. Cirac, A. S. Parkins, R. Blatt, and P. Zoller, “‘Dark’ squeezed
states of the motion of a trapped ion,” Physical Review Letters,
vol. 70, no. 5, pp. 556–559, 1993.
[34] C. W. Gardiner and P. Zoller, Quantum Noise: A Handbook of
Markovian and Non-Markovian Quantum Stochastic Methods with
Applications to Quantum Optics. Springer, 2000.
[15] J. F. Poyatos, J. I. Cirac, and P. Zoller, “Quantum reservoir
engineering with laser cooled trapped ions,” Physical Review Letters,
vol. 77, no. 23, pp. 4728–4731, 1996.
[35] L. Bouten, R. V. Handel, and M. R. James, “An introduction to
quantum filtering,” SIAM Journal on Control and Optimization,
vol. 46, no. 6, pp. 2199–2241, 2007.
[16] Y. D. Wang and A. A. Clerk, “Reservoir-engineered entanglement in
optomechanical systems,” Physical Review Letters, vol. 110, no. 25,
p. 253601, 2013.
[36] M. M. Wolf, G. Giedke, O. Krüger, R. F. Werner, and J. I. Cirac,
“Gaussian entanglement of formation,” Physical Review A, vol. 69,
no. 5, p. 052320, 2004.
[17] H. Krauter, C. A. Muschik, K. Jensen, W. Wasilewski, J. M. Petersen,
J. I. Cirac, and E. S. Polzik, “Entanglement generated by dissipation
and steady state entanglement of two macroscopic objects,” Physical
Review Letters, vol. 107, no. 8, p. 080503, 2011.
[37] O. Techakesari and H. I. Nurdin, “On the quasi-balanceable class
of linear quantum stochastic systems,” Systems & Control Letters,
vol. 78, pp. 25–31, 2015.
[38] H. I. Nurdin, “Structures and transformations for model reduction of
linear quantum stochastic systems,” IEEE Transactions on Automatic
Control, vol. 59, no. 9, pp. 2413–2425, 2014.
[18] M. J. Woolley and A. A. Clerk, “Two-mode squeezed states in
cavity optomechanics via engineering of a single reservoir,” Physical
Review A, vol. 89, no. 6, p. 063805, 2014.
[39] I. R. Petersen, “Low frequency approximation for a class of linear
quantum systems using cascade cavity realization,” Systems &
Control Letters, vol. 61, no. 1, pp. 173–179, 2012.
[19] C. W. Gardiner, “Driving a quantum system with the output field from
another driven quantum system,” Physical Review Letters, vol. 70,
no. 15, pp. 2269–2272, 1993.
[40] M. Guţă and N. Yamamoto, “System identification for passive linear
quantum systems,” IEEE Transactions on Automatic Control, vol. 61,
no. 4, pp. 921–936, 2016.
[20] H. I. Nurdin, “On synthesis of linear quantum stochastic systems by
pure cascading,” IEEE Transactions on Automatic Control, vol. 55,
no. 10, pp. 2439–2444, 2010.
[41] J. E. Gough and G. Zhang, “On realization theory of quantum linear
systems,” Automatica, vol. 59, pp. 139–151, 2015.
[21] I. R. Petersen, “Cascade cavity realization for a class of complex
transfer functions arising in coherent quantum feedback control,”
Automatica, vol. 47, no. 8, pp. 1757–1763, 2011.
[42] K. Zhou, J. C. Doyle, and K. Glover, Robust and Optimal Control.
Prentice Hall, 1996.
[43] J. Gough and M. R. James, “The series product and its application
to quantum feedforward and feedback networks,” IEEE Transactions
on Automatic Control, vol. 54, no. 11, pp. 2530–2544, 2009.
[22] P. Seibert and R. Suarez, “Global stabilization of nonlinear cascade
systems,” Systems & Control Letters, vol. 14, no. 4, pp. 347–352,
1990.
11
[44] S. Iida, M. Yukawa, H.Yonezawa, N. Yamamoto, and A. Furusawa,
“Experimental demonstration of coherent feedback control on optical
field squeezing,” IEEE Transactions on Automatic Control, vol. 57,
no. 8, pp. 2045–2050, 2012.
[45] H. I. Nurdin, M. R. James, and A. C. Doherty, “Network synthesis
of linear dynamical quantum stochastic systems,” SIAM Journal on
Control and Optimization, vol. 48, no. 4, pp. 2686–2718, 2009.
[46] R. N. Alexander, P. Wang, N. Sridhar, M. Chen, O. Pfister, and
N. C. Menicucci, “One-way quantum computing with arbitrarily
large time-frequency continuous-variable cluster states from a
single optical parametric oscillator,” 2015. [Online]. Available:
http://arxiv.org/abs/1509.00484
[47] C. A. Muschik, E. S. Polzik, and J. I. Cirac, “Dissipatively driven
entanglement of two macroscopic atomic ensembles,” Physical Review
A, vol. 83, no. 5, p. 052312, 2011.
[48] C. Gerry and P. Knight, Introductory Quantum Optics.
University Press, 2004.
Cambridge
12
| 3cs.SY
|
Network Synchronization with Nonlinear Dynamics and
Switching Interactions∗
arXiv:1401.6541v3 [cs.SY] 24 Aug 2015
Tao Yang†, Ziyang Meng‡, Guodong Shi§, Yiguang Hong¶and Karl Henrik Johanssonk
Abstract
This paper considers the synchronization problem for networks of coupled nonlinear dynamical systems under switching communication topologies. Two types of nonlinear agent
dynamics are considered. The first one is non-expansive dynamics (stable dynamics with
a convex Lyapunov function ϕ(·)) and the second one is dynamics that satisfies a global
Lipschitz condition. For the non-expansive case, we show that various forms of joint connectivity for communication graphs are sufficient for networks to achieve global asymptotic
ϕ-synchronization. We also show that ϕ-synchronization leads to state synchronization provided that certain additional conditions are satisfied. For the globally Lipschitz case, unlike
the non-expansive case, joint connectivity alone is not sufficient for achieving synchronization. A sufficient condition for reaching global exponential synchronization is established in
terms of the relationship between the global Lipschitz constant and the network parameters.
We also extend the results to leader-follower networks.
Keywords: Multi-agent systems, nonlinear agents, switching interactions, synchronization.
∗
This work has been supported in part by the Knut and Alice Wallenberg Foundation and the Swedish Research
Council.
†
T. Yang is with the Pacific Northwest National Laboratory, 902 Battelle Boulevard, Richland, WA 99352
USA (e-mail: [email protected]).
‡
Z. Meng is with the Institute for Information-Oriented Control, Technische Universität München, D-80290
Munich, Germany (e-mail: [email protected]).
§
G. Shi is with the College of Engineering and Computer Science, The Australian National University, Canberra
ACT 0200, Australia (e-mail: [email protected]).
¶
Y. Hong is with the Key Laboratory of Systems and Control, Institute of Systems Science, Chinese Academy
of Science, Beijing 100190, China (e-mail: [email protected]).
k
K. H. Johansson is with the ACCESS Linnaeus Centre, School of Electrical Engineering, Royal Institute of
Technology, Stockholm 10044, Sweden (e-mail: [email protected]).
1
1
Introduction
We consider the synchronization problem for a network of coupled nonlinear agents with agent
set V = {1, 2, . . . , N }. Their interactions (communications in the network) are described by
a time-varying directed graph Gσ(t) = (V, Eσ(t) ), with σ : [0, ∞) → P as a piecewise constant
signal, where P is a finite set of all possible graphs over V. The state of agent i ∈ V at time t is
denoted as xi (t) ∈ Rn and evolves according to
ẋi = f (t, xi ) +
X
aij (t)(xj − xi ),
(1)
j∈Ni (σ(t))
where f (t, xi ) : [0, ∞) × Rn → Rn is piecewise continuous in t and continuous in xi representing
the uncoupled inherent agent dynamics, Ni (σ(t)) is the set of agent i’s neighbors at time t, and
aij (t) > 0 is a piecewise continuous function marking the weight of edge (j, i) at time t.
Systems of the form (1) have attracted considerable attention. Most works focus on the case
where the communication graph Gσ(t) is fixed, e.g., [1–7]. It is shown that for the case where
f (t, xi ) satisfies a Lipschitz condition, synchronization is achieved for a connected graph provided
that the coupling strength is sufficiently large. However, for the case where the communication
graph is time-varying, the synchronization problem becomes much more challenging and existing
literature mainly focuses on a few special cases when f (t, xi ) is linear, e.g., the single-integrator
case [8–10], the double-integrator case [11], and the neutrally stable case [12, 13]. Other studies
assume some particular structures for the communication graph [14–16]. In particular, in [14],
the authors focus on the case where the adjacency matrices associated with all communication
graphs are simultaneously triangularizable. The authors of [15] consider switching communication graphs that are weakly connected and balanced at all times. A more general case where
the switching communication graph frequently has a directed spanning tree has been considered
in [16]. These special structures on the switching communication graph are rather restrictive
compared to joint connectivity where the communication can be lost at any time.
This paper aims to investigate whether joint connectivity for switching communication
graphs can render synchronization for the nonlinear dynamics (1). We distinguish two classes
depending on whether the nonlinear agent dynamics f (t, xi ) is expansive or not. For the nonexpansive case, we focus on the case where the nonlinear agent dynamics is stable with a convex
Lyapunov function ϕ(·). We show that various forms of joint connectivity for communication
graphs are sufficient for networks to achieve global asymptotic ϕ-synchronization, that is, the
2
function ϕ of the agent state converges to a common value. We also show that ϕ-synchronization
implies state synchronization provided that additional conditions are satisfied. For the expansive case, we focus on when the nonlinear agent dynamics is globally Lipschitz, and establish
a sufficient condition for networks to achieve global exponential synchronization in terms of a
relationship between the Lipschitz constant and the network parameters.
The remainder of this paper is organized as follows. Section 2 presents the problem definition
and main results. Section 3 provides technical proofs. In Section 4, we extend the results to
leader-follower networks. Finally, Section 5 concludes the paper.
2
Problem Definition and Main Results
2.1
Problem Set-up
Throughout the paper we make a standard dwell time assumption [17] on the switching signal
σ(t): there is a lower bound τD > 0 between two consecutive switching time instants of σ(t).
We also assume that there are constants 0 < a∗ ≤ a∗ such that a∗ ≤ aij (t) ≤ a∗ for all t ≥ 0.
We denote x = [xT1 , xT2 , . . . , xTN ]T ∈ RnN and assume that the initial time is t = t0 ≥ 0, and the
initial state x(t0 ) = (xT1 (t0 ), . . . , xTN (t0 ))T ∈ RnN . A digraph is strongly connected if it contains a
directed path from every node to every other node. The joint graph of Gσ(t) in the time interval
[t1 , t2 ) with t1 < t2 ≤ ∞ is denoted as G([t1 , t2 )) = ∪t∈[t1 ,t2 ) G(t) = (V, ∪t∈[t1 ,t2 ) Eσ(t) ). For the
communication graph, we introduce the following definition.
Definition 1 (i). Gσ(t) is uniformly jointly strongly connected if there exists a constant T > 0
such that G([t, t + T )) is strongly connected for any t ≥ 0.
(ii). Assume that Gσ(t) is undirected for all t ≥ 0. Gσ(t) is infinitely jointly connected if
G([t, ∞)) is connected for any t ≥ 0.
In this paper, we are interested in the following synchronization problems.
Definition 2 The multi-agent system (1) achieves global asymptotic ϕ-synchronization, where
ϕ : Rn → R is a continuously differentiable function, if for any initial state x(t0 ), there exists a
constant d⋆ (x(t0 )), such that limt→∞ ϕ(xi (t)) = d⋆ (x(t0 )) for any i ∈ V and any t0 ≥ 0.
3
Definition 3 (i) The multi-agent system (1) achieves global asymptotic synchronization if
limt→∞ (xi (t) − xj (t)) = 0 for any i, j ∈ V, any t0 ≥ 0 and any x(t0 ) ∈ RnN .
(ii) Multi-agent system (1) achieves global exponential synchronization if there exist γ ≥ 1
and λ > 0 such that
max
{i,j}∈V×V
kxi (t) − xj (t)k2 ≤ γe−λ(t−t0 )
max
{i,j}∈V×V
kxi (t0 ) − xj (t0 )k2 , t ≥ t0 ,
(2)
for any t0 ≥ 0 and any x(t0 ) ∈ RnN .
Remark 1 ϕ-synchronization is a type of output synchronization where the output of agent
i ∈ V is chosen to be ϕ(xi ). It is related to but different from χ-synchronization [18, 19] since ϕ
is a function of an individual agent state while χ is a function of all agent states.
2.2
Non-expansive Inherent Dynamics
In this section, we focus on when the nonlinear inherent agent dynamics is non-expansive as
indicated by the following assumption.
Assumption 1 ϕ : Rn → R is a continuously differentiable positive definite convex function
satisfying
(i). limkηk→∞ ϕ(η) = ∞;
(ii). h∇ϕ(η), f (t, η)i ≤ 0 for any η ∈ Rn and any t ≥ 0.
The following lemma shows how Assumption 1 enforces non-expansive dynamics.
Lemma 1 Let Assumption 1 hold. Along the multi-agent dynamics (1), maxi∈V ϕ(xi (t)) is
non-increasing for all t ≥ 0.
We now state main results for the non-expansive case.
Theorem 1 Let Assumption 1 hold. The multi-agent system (1) achieves global asymptotic
ϕ-synchronization if Gσ(t) is uniformly jointly strongly connected.
Theorem 2 Let Assumption 1 hold. Assume that Gσ(t) is undirected for all t ≥ t0 . The multiagent system (1) achieves global asymptotic ϕ-synchronization if Gσ(t) is infinitely jointly connected.
4
Remark 2 For the linear time-varying case f (t, x) = A(t)x, if there exists a matrix P = P T > 0
such that
P A(t) + AT (t)P ≤ 0,
∀t ≥ 0,
(3)
then ϕ(x) = xT P x for x ∈ Rn satisfies Assumption 1. For the linear time-invariant case
f (t, x) = Ax, the condition (3) is equivalent to that the matrix A is neutrally stable [13].
2.3
ϕ-synchronization vs. State Synchronization
The following result establishes conditions under which ϕ-synchronization may imply state synchronization.
Theorem 3 Let Gσ(t) ≡ G with G being a fixed, strongly connected digraph under which the
multi-agent system (1) achieves global asymptotic ϕ-synchronization for some positive definite
function ϕ : Rn → R. Let Assumption 1 hold. Moreover, assume that
(i). f (t, η) is bounded for any t ≥ 0 and any η ∈ Rn .
(ii). c1 kηk2 ≤ ϕ(η) ≤ c2 kηk2 for some 0 < c1 ≤ c2 ; and
(iii). ϕ(·) is strongly convex.
Then the multi-agent system (1) achieves global asymptotic synchronization.
2.4
Lipschitz Inherent Dynamics
We consider also the case when the nonlinear inherent agent dynamics is possibly expansive.
We focus on when the dynamics satisfies the following global Lipschitz condition.
Assumption 2 There exists a constant L > 0 such that
kf (t, η) − f (t, ζ)k ≤ Lkη − ζk,
∀η, ζ ∈ Rn , ∀t ≥ 0.
(4)
Our main result for this case is given below.
Theorem 4 Let Assumption 2 hold. Assume that Gσ(t) is uniformly jointly strongly connected.
Global exponential synchronization is achieved for the multi-agent system (1) if L < ρ∗ /2, where
ρ∗ is a constant depending on the network parameters.
5
Remark 3 Assumption 2 and its variants have been considered in the literature for fixed communication graphs, e.g., [1–7]. Compared with the existing literature, we here study a more
challenging case, where the communication graphs are time-varying. Unlike the fixed case where
the global Lipschitz condition is sufficient to guarantee synchronization, Theorem 4 established a
sufficient synchronization condition related to the Lipschitz constant and the network parameters.
3
Proofs of the Main Results
In this section, we provide proofs of the main results.
3.1
Proof of Lemma 1
Recall that the upper Dini derivative of a function h(t) : (a, b) → R at t is defined as D + h(t) =
lim sups→0+
h(t+s)−h(t)
.
s
The following lemma from [10, 20] is useful for the proof.
Lemma 2 Let Vi (t, x) : R × Rn → R (i = 1, . . . , N ) be continuously differentiable and V (t, x) =
maxi=1,...,N Vi (t, x). If I(t) = {i ∈ {1, 2, . . . , N } : V (t, x(t)) = Vi (t, x(t))} is the set of indices
where the maximum is reached at t, then D + V (t, x(t)) = maxi∈I(t) V̇i (t, x(t)).
Denote I(t) = {i ∈ V : maxi∈V ϕ(xi (t)) = ϕ(xi (t))}. We first note that the convexity property
of ϕ(·) implies that [21, pp.69]
h∇ϕ(η), ζ − ηi ≤ ϕ(ζ) − ϕ(η),
∀η, ζ ∈ Rn .
(5)
It then follows from Lemma 2, Assumption 1(ii) and (5) that
D + max ϕ(xi (t)) = max ∇ϕ(xi ), f (t, xi ) +
i∈V
i∈I(t)
≤ max
i∈I(t)
X
aij (t)(xj − xi )
j∈Ni (σ(t))
X
aij (t)(ϕ(xj ) − ϕ(xi )) ≤ 0,
j∈Ni (σ(t))
where the last inequality follows from ϕ(xj ) ≤ ϕ(xi ). This proves the lemma.
3.2
Proof of Theorem 1
It follows from Lemma 1 that for any initial state x(t0 ) ∈ RnN , there exists a constant d∗ =
d⋆ (x(t0 )) ≥ 0, such that limt→∞ maxi∈V ϕ(xi ) = d⋆ . We shall show that d⋆ is the required
constant in Definition 2 of ϕ-synchronization.
6
We first note that by Lemma 1 that for all i ∈ V, there exist constants 0 ≤ αi ≤ βi ≤ d⋆ ,
such that
lim inf ϕ(xi (t)) = αi ,
lim sup ϕ(xi (t)) = βi .
t→∞
t→∞
Also note that it follows from limt→∞ maxi∈V ϕ(xi (t)) = d⋆ that for any ε > 0, there exists
T1 (ε) > 0 such that
ϕ(xi (t)) ∈ [0, d⋆ + ε],
∀i ∈ V, ∀t ≥ T1 (ε).
(6)
The proof of Theorem 1 is based on a contradiction argument and relies on the following
lemma.
Lemma 3 Let Assumption 1 hold. Assume that Gσ(t) is uniformly jointly strongly connected.
If there exists an agent k0 ∈ V such that 0 ≤ αk0 < d⋆ , then there exists 0 < ρ̄ < 1 and t̄ such
that for all i ∈ V, ϕ(xi (t̄ + (N − 1)T0 )) ≤ ρ̄M0 + (1 − ρ̄)(d⋆ + ε), where
T0 , T + 2τD ,
(7)
with T given in Definition 1(i) and τD is the dwell time.
Proof: Let us first define M0 ,
αk0 +βk0
2
< d⋆ . Then there exists an infinite time sequence
t0 < t̃1 < . . . < t̃k < . . . with limk→∞ t̃k = ∞ such that ϕ(x(t̃k )) = M0 for all k = 1, 2, . . .. We
then pick up one t̃k , k = 1, 2, . . . such that it is greater than or equal to T1 (ε) and denote it as
t̃k0 .
We now prove the lemma by estimating an upper bound of the scalar function ϕ(xi ) agent
by agent. The proof is based on a generalization of the method proposed in the proof of [22,
Lemma 4.3] but with substantial differences on the agent dynamics and Lyapunov function.
Moreover, the convexity of ϕ(·) plays an important role.
Step 1. Focus on agent k0 . By using Assumption 1(ii), (5), and (6), we obtain that for all
t ≥ t̃k0 ,
d
ϕ(xk0 (t)) = ∇ϕ(xk0 ), f (t, xk0 ) +
dt
≤
X
X
ak0 j (t)(xj − xk0 )
j∈Nk0 (σ(t))
ak0 j (t) (ϕ(xj ) − ϕ(xk0 ))
j∈Nk0 (σ(t))
≤ a∗ (N − 1)(d⋆ + ε − ϕ(xk0 )).
7
(8)
It then follows that for all t ≥ t̃k0 ,
ϕ(xk0 (t)) ≤ e−λ1 (t−t̃k0 ) ϕ(xk0 (t̃k0 )) + 1 − e−λ1 (t−t̃k0 ) (d⋆ + ε),
(9)
where λ1 = a∗ (N − 1).
Step 2. Consider agent k1 6= k0 such that (k0 , k1 ) ∈ Eσ(t) for t ∈ [t̃k0 , t̃k0 + T0 ). The existence of
such an agent can be shown as follows. Since Gσ(t) is uniformly jointly strongly connected, it is
not hard to see that there exists an agent k1 6= k0 ∈ V and t1 ≥ t̃k0 such that (k0 , k1 ) ∈ Eσ(t) for
t ∈ [t1 , t1 + τD ) ⊆ [t̃k0 , t̃k0 + T0 ).
From (9), we obtain for all t ∈ [t̃k0 , t̃k0 + (N − 1)T0 ],
ϕ(xk0 (t)) ≤ κ0 , ρM0 + (1 − ρ)(d⋆ + ε),
∗ (N −1)2 T
0
where ρ = e−λ1 (N −1)T0 = e−a
(10)
.
We next estimate ϕ(xk1 (t)) by considering two different cases.
Case I: ϕ(xk1 (t)) > ϕ(xk0 (t)) for all t ∈ [t1 , t1 + τD ).
By using Assumption 1(ii), (5), (6), and (10), we obtain for all t ∈ [t1 , t1 + τD ),
d
ϕ(xk1 (t)) ≤
dt
X
j∈Nk1 (σ(t))\{k0 }
ak1 j (t) ϕ(xkj ) − ϕ(xk1 ) + ak1 k0 (t)(ϕ(xk0 ) − ϕ(xk1 ))
≤ a∗ (N − 2)(d⋆ + ε − ϕ(xk1 )) + a∗ (κ0 − ϕ(xk1 )) .
From the preceding relation, we obtain for t ∈ [t1 , t1 + τD ),
ϕ(xk1 (t)) ≤ e−λ2 (t−t1 ) ϕ(xk1 (t1 )) +
[a∗ (N − 2)(d⋆ + ε) + a∗ κ0 ] (1 − e−λ2 (t−t1 ) )
,
λ2
where λ2 = a∗ (N − 2) + a∗ . Therefore, we have
ϕ(xk1 (t1 + τD )) ≤ κ1 , µ(d⋆ + ε) + (1 − µ)κ0 ,
(11)
where
µ=
λ2 − a∗ (1 − e−λ2 τD )
.
λ2
(12)
By applying the same analysis as we obtained (9) to the agent k1 , we obtain for all t ≥ t1 + τD ,
i
h
ϕ(xk1 (t)) ≤ e−λ1 (t−(t1 +τD )) κ1 + 1 − e−λ1 (t−(t1 +τD )) (d⋆ + ε).
8
(13)
By combining the inequalities (10), (11) and (13), we obtain for all t ∈ [t1 + τD , t̃k0 + (N − 1)T0 ],
ϕ(xk1 (t)) ≤ ϕ1 M0 + (1 − ϕ1 )(d⋆ + ε),
(14)
where ϕ1 = (1 − µ)ρ2 .
Case II: There exists a time instant t̄1 ∈ [t1 , t1 + τD ) such that
ϕ(xk1 (t̄1 )) ≤ ϕ(xk0 (t̄1 )) ≤ κ0 .
(15)
By applying the similar analysis as we obtained (8) to the agent k1 , we obtain for all t ≥ t̃k0 ,
d
ϕ(xk1 (t)) ≤ a∗ (N − 1)(d⋆ + ε − ϕ(xk1 (t))).
dt
This leads to
ϕ(xk1 (t)) ≤ e−λ1 (t−t̄1 ) ϕ(xk1 (t̄1 )) + (1 − e−λ1 (t−t̄1 ) )(d⋆ + ε).
By combining the preceding relation, (10), and (15), and using 0 < ϕ1 = (1 − µ)ρ2 < ρ2 which
follows from 0 < µ < 1, we obtain for all t ∈ [t1 + τD , t̃k0 + (N − 1)T0 ],
ϕ(xk1 (t)) ≤ ρ2 M0 + (1 − ρ2 )(d⋆ + ε) < ϕ1 M0 + (1 − ϕ1 )(d⋆ + ε).
From the preceding relation and (14), it follows that for both cases, we have for all t ∈ [t1 +
τD , t̃k0 + (N − 1)T0 ],
ϕ(xk1 (t)) ≤ ϕ1 M0 + (1 − ϕ1 )(d⋆ + ε).
From the preceding relation, (10) and 0 < ϕ1 < ρ < 1, it follows that for all t ∈ [t1 + τD , t̃k0 +
(N − 1)T0 ],
ϕ(xj (t)) ≤ ϕ1 M0 + (1 − ϕ1 )(d⋆ + ε),
j ∈ {k0 , k1 }.
(16)
Step 3. Consider agent k2 ∈
/ {k0 , k1 } such that there exists an edge from the set {k0 , k1 } to the
agent k2 in Eσ(t) for t ∈ [t2 , t2 + τD ) ⊆ [t̃k0 + T0 , t̃k0 + 2T0 ). The existence of such an agent k2
and t2 follows similarly from the argument in Step 2.
Similarly, we can bound ϕ(xk2 (t)) by considering two different cases and obtain that for all
t ∈ [t2 + τD , t̃k0 + (N − 1)T0 ],
ϕ(xk2 (t)) ≤ ϕ2 M0 + (1 − ϕ2 )(d⋆ + ε),
where ϕ2 = ((1 − µ)ρ2 )2 .
9
(17)
By combining (16) and (17), and using 0 < ϕ2 < ϕ1 < 1, we obtain that for all t ∈
[t2 + τD , t̃k0 + (N − 1)T0 ],
ϕ(xj (t)) ≤ ϕ2 M0 + (1 − ϕ2 )(d⋆ + ε),
j ∈ {k0 , k1 , k2 }.
Step 4. By repeating the above process on time intervals [t̃k0 + 2T0 , t̃k0 + 3T0 ), . . . , [t̃k0 + (N −
2)T0 , t̃k0 + (N − 1)T0 ), we eventually obtain that for all i ∈ V,
ϕ(xi (t̃k0 + (N − 1)T0 )) ≤ ϕN −1 M0 + (1 − ϕN −1 )(d⋆ + ε).
where ϕN −1 = ((1 − µ)ρ2 )N −1 . The result of the lemma then follows by choosing ρ̄ = ϕN −1 and
t̄ = t̃k0 .
We are now ready to prove Theorem 1 by contradiction. Suppose that there exists an agent
k0 ∈ V such that 0 ≤ αk0 < d⋆ . It then follows from Lemma 3 that ϕ(xi (t̄ + (N − 1)T0 )) < d⋆ for
all i ∈ V, provided that ε <
ρ̄(d⋆ −M0 )
.
1−ρ̄
This contradicts the fact that limt→∞ maxi∈V ϕ(xi ) = d⋆ .
Thus, there does not exist an agent k0 ∈ V such that 0 ≤ αk0 < d⋆ . Hence, limt→∞ ϕ(xi (t)) = d⋆
for all i ∈ V.
3.3
Proof of Theorem 2
The proof relies on the following lemma.
Lemma 4 Let Assumption 1 hold. Assume that Gσ(t) is infinitely jointly connected. If there
exists an agent k0 ∈ V such that 0 ≤ αk0 < d⋆ , then there exist 0 < ρ̃ < 1 and t̃ such that
ϕ(xi (t̃ + τD )) ≤ ρ̃M0 + (1 − ρ̃)(d⋆ + ε),
∀i ∈ V.
Proof: The proof of Lemma 4 is similar to that of Lemma 3 and based on estimating an upper
bound for the scalar quantity ϕ(xi ) agent by agent. However, since Gσ(t) is infinitely jointly
connected, the method that we get the order of the agents based on the intervals induced by
the uniform bound T cannot be used here. We can however apply the strategy in [22] for the
analysis as shown below:
Step 1. In this step, we focus on agent k0 . Since Gσ(t) is infinitely jointly connected, we can
define
t̂1 ,
inf
{∃ i ∈ V | (k0 , i) ∈ Eσ(t) },
t∈[t̃k0 ,∞)
10
and the set
V1 , {i ∈ V | (k0 , i) ∈ Eσ(t̂1 ) }.
For t̃k0 ≤ t < t̂1 , agent k0 has no neighbor, it follows from Assumption 1(ii) that
d
dt ϕ(xk0 (t))
=
h∇ϕ(xk0 ), f (t, xk0 )i ≤ 0. Thus, ϕ(xk0 (t)) ≤ ϕ(xk0 (t̃k0 )) for t̃k0 ≤ t < t̂1 . By applying a similar
analysis as we obtained (10), we have for all t ∈ [t̂1 , t̂1 + τD ),
ϕ(xk0 (t)) ≤ κ̂0 , ρ̂M0 + (1 − ρ̂)(d⋆ + ε),
∗ (N −1)τ
D
where ρ̂ = e−a
.
Step 2. In this step, we focus on all k1 ∈ V1 . We then estimate ϕ(xk1 (t)) for all k1 ∈ V1 by
considering two different cases, i.e., Case I: If ϕ(xk1 (t)) > ϕ(xk0 (t)) for all t ∈ [t̂1 , t̂1 + τD ), and
Case II: If there exists a time instant t̄1 ∈ [t̂1 , t̂1 + τD ) such that ϕ(xk1 (t̄1 )) ≤ ϕ(xk0 (t̄1 )) ≤ κ̂0 .
By using the similar argument as the two-case analysis for agent k1 in the proof of Theorem 1,
we eventually obtain for all j ∈ k0 ∪ V1 ,
ϕ(xj (t̂1 + τD )) < ϕ̂1 M0 + (1 − ϕ̂1 )(d⋆ + ε),
(18)
where ϕ̂1 = (1 − µ)ρ̂2 with µ given by (12).
Step 3. We then view the set {k0 } ∪ V1 as a subsystem. Define t̂2 as the first time when there
is an edge between this subsystem and the remaining agents and V2 accordingly. By using the
similar analysis for agent k1 in Step 2, we can estimate the upper bound for all the agent in the
set {k0 } ∪ V1 ∪ V2 ,
Step 4. Since Gσ(t) is infinitely jointly connected, we can continue the above process until
V = {k0 } ∪ V1 ∪ · · · ∪ Vℓ for some ℓ ≤ N − 1. Eventually, we have
ϕ(xi (t̂ℓ + τD )) < ϕ̂N −1 M0 + (1 − ϕ̂N −1 )(d⋆ + ε),
∀i ∈ V.
The result of lemma then follows by choosing t̃ = t̂ℓ and ρ̃ = ϕ̂N −1 .
The remaining proof of Theorem 2 follows from a contradiction argument and Lemma 4 in
the same way as the proof of Theorem 1.
11
3.4
Proof of Theorem 3
If the multi-agent system (1) reaches asymptotic ϕ-synchronization, i.e., limt→∞ ϕ(xi (t)) = d⋆
for all i ∈ V, then for any ǫ > 0, there exists a Tǫ > 0 such that
d⋆ − ǫ ≤ ϕ(xi (t)) ≤ d⋆ + ǫ,
∀i ∈ V,
∀t ≥ Tǫ .
(19)
If d⋆ = 0 the desired conclusion holds trivially, i.e., limt→∞ xi (t) = 0 for all i ∈ V due to the
positive definiteness of ϕ(·). In the remainder of the proof we assume d⋆ > 0. We shall prove
the result by contradiction. Suppose that state synchronization is not achieved, then there exist
two agents i0 , j0 ∈ V such that lim supt→∞ kxi0 (t) − xj0 (t)k > 0. In other words, there exist an
infinite time sequence t1 < · · · < tk < . . . with limk→∞ tk = ∞, and a constant δ > 0 such that
p
kxi0 (tk ) − xj0 (tk )k = 2N −1 δ/c1 for all k ≥ 1, where c1 is given in condition (ii) of Theorem 3.
We divide the following analysis into three steps.
Step 1. In this step, we prove the following crucial claim.
Claim. For any tk , there are two agents i∗ , j∗ ∈ V with (i∗ , j∗ ) ∈ E such that kxi∗ (tk )−xj∗ (tk )k ≥
p
δ/c1 .
We establish this claim via a recursive analysis. If either (i0 , j0 ) ∈ E or (j0 , i0 ) ∈ E then
the result follows trivially. Otherwise we pick up another agent k0 satisfying that there is an
edge between k0 and {i0 , j0 }. This k0 always exists since G is strongly connected. Then either
p
p
kxk0 (tk ) − xi0 (tk )k ≥ 2N −2 δ/c1 or kxk0 (tk ) − xj0 (tk )k ≥ 2N −2 δ/c1 must hold. Thus, we have
again either established the claim, or we can continue to select another agent different from i0 ,
j0 , and k0 and repeat the argument. Since we have a finite number of agents, the desired claim
holds.
Furthermore, since there is a finite number of agent pairs, without loss of generality, we
assume that the given agent pair i∗ , j∗ does not vary for different tk (otherwise we can always
select an infinite subsequence of tk for the following discussions).
Step 2. In this step, we establish a lower bound of kxi∗ (t)−xj∗ (t)k2 for a small time interval after
a particular tk satisfying tk > Tǫ . From (19) and limkxk→∞ ϕ(x) = ∞ given in Assumption 1(i),
we see that xi (t) and ∇ϕ(xi (t) − xj (t)) are bounded for all i, j ∈ V and for all t ≥ Tǫ . It then
follows from condition (i) of Theorem 3 and (1) that
d
ϕ(xi∗ − xj∗ ) ≤
dt
∇ϕ(xi∗ − xj∗ ), f (t, xi∗ ) − f (t, xj∗ )i
12
+
∇ϕ(xi∗ − xj∗ ),
X
ai∗ k1 (t)(xk1 − xi∗ )i
X
aj∗ k2 (t)(xk2 − xj∗ )i ≤ L∗
k1 ∈Ni∗ (σ(t))
+
∇ϕ(xi∗ − xj∗ ),
k2 ∈Nj∗ (σ(t))
for all t ≥ tk and some L∗ > 0. Without loss of generality we assume that ǫ ≤ 1. Then L∗ will
be independent of ǫ.
By plugging in the fact that kxi∗ (tk ) − xj∗ (tk )k ≥
Theorem 3, we obtain that
kxi∗ (t) − xj∗ (t)k2 ≥
δ
,
2c2
p
δ/c1 and using the condition (ii) of
δ
t ∈ tk , tk +
.
2L∗
(20)
Step 3. We first note that the strong convexity of ϕ(·) implies that [21, pp.459] there exists an
m > 0 such that
h∇ϕ(η), ζ − ηi ≤ ϕ(ζ) − ϕ(η) −
m
kη − ζk2 , ∀η, ζ ∈ Rn .
2
(21)
By using Assumption 1(ii) and (21), we obtain for t ≥ Tǫ ,
d
m
ϕ(xj∗ (t)) ≤ aj∗ i∗ (t) ϕ(xi∗ ) − ϕ(xj∗ ) − kxi∗ − xj∗ k2
dt
2
X
aj∗ k (t) ϕ(xk ) − ϕ(xj∗ )
+
k∈Nj∗ (σ(t))\{i∗ }
m
≤ aj∗ i∗ (t) ϕ(xi∗ ) − ϕ(xj∗ ) − aj∗ i∗ (t)kxi∗ (t) − xj∗ (t)k2
2
X
aj∗ k (t) ϕ(xk ) − ϕ(xj∗ ) ,
+
(22)
k∈Nj∗ (σ(t))\{i∗ }
h
By using (19), (20), (22), and condition (ii) of Theorem 3, we obtain that for t ∈ tk , tk +
a∗ mδ
d
ϕ(xj∗ (t)) ≤ 2(N − 1)a∗ ǫ −
,
dt
4c2
δ
2L∗
i
,
which yields
δ
a∗ mδ
δ
∗
)) ≤ d⋆ + ǫ + 2(N − 1)a ǫ −
.
ϕ(xj∗ (tk +
2L∗
4c2 2L∗
It is then straightforward to see that
ϕ(xj∗ (tk +
δ
a∗ mδ2
)) < d⋆ −
2L∗
32c2 L∗
if we take
ǫ < min
n a mδ2
o
a∗ mδ
∗
.
,
32c2 L∗ 16(N − 1)a∗ c2
However, this contradicts the definition of ϕ-synchronization since tk is arbitrarily chosen. This
completes the proof and the desired conclusion holds.
13
3.5
Proof of Theorem 4
The proof is based on the convergence analysis of the scalar quantity
V (t, x(t)) =
max
{i,j}∈V×V
Vij (t, x(t)),
(23)
where
1
Vij (t, x(t)) = e−2L(t−t0 ) kxi (t) − xj (t)k2 , ∀{i, j} ∈ V × V.
2
(24)
Unlike the contradiction argument used for proof of Theorem 1, where the convergence rate is
unclear, here we explicitly characterize the convergence rate. The proof relies on the following
lemmas.
Lemma 5 Let Assumption 4 hold. Along the multi-agent dynamics (1), V (t, x(t)) is nonincreasing for all t ≥ 0.
Proof: This lemma establishes a critical non-expansive property along the multi-agent dynamics
(1) for the globally Lipschitz case. The proof follows from the same techniques as those for
proving Lemma 1 by investigating the Dini derivative of V (t, x(t)).
Let V 1 × V 2 be the set containing all the node pairs that reach the maximum at time t, i.e.,
V 1 (t) × V 2 (t) = {{i, j} ∈ V × V : Vij (t) = V (t)}. It is not hard to obtain that
D+ V =
max
{i,j}∈V 1 ×V 2
×
n
e−2L(t−t0 ) (xi − xj )T (f (t, xi ) − f (t, xj )) + e−2L(t−t0 ) (xi − xj )T
X
aik1 (t)(xk1 − xi ) − e−2L(t−t0 ) (xi − xj )T
X
ajk2 (t)(xk2 − xj )
k2 ∈Nj (σ(t))
k1 ∈Ni (σ(t))
o
−Le−2L(t−t0 ) kxi − xj k2
X
1
e−2L(t−t0 )
max
≤
2 {i,j}∈V 1 ×V 2
aik1 (t)(kxj − xk1 k2 − kxi − xj k2 )
k1 ∈Ni (σ(t))
X
+e−2L(t−t0 )
k2 ∈Nj (σ(t))
≤
max
X
{i,j}∈V 1 ×V 2
k1 ∈Ni (σ(t))
ajk2 (t)(kxi − xk2 k2 − kxi − xj k2 )
aik1 (t)(Vjk1 − Vij ) +
X
k2 ∈Nj (σ(t))
ajk2 (t)(Vik2
− Vij ) ≤ 0,
(25)
where the first equality follows from Lemma 2 and (1), the first inequality follows from (4) and
ab ≤
a2 +b2
2
and −ab ≤
a2 +b2
2
for all a, b ∈ Rn , and the second inequality follows from (24).
14
Lemma 6 Let Assumption 4 hold. Assume that Gσ(t) is uniformly jointly strongly connected.
Then there exists 0 < β̃ < 1 such that
Vij (N T0 , x(N T0 )) ≤ β̃V ∗ ,
∀{i, j} ∈ V × V,
where N = N − 1, T0 is given by (7) and V ∗ = V (t0 , x(t0 )).
Proof: The proof is based on the convergence analysis of Vij (t, x(t) for all agent pairs {i, j} ∈
V × V in several steps, which is similar to the proof of Lemma 3. Without loss of generality,
we assume that t0 = 0. We also sometimes denote V (t, x(t)) and Vij (t, x(t)) as V and Vij ,
respectively, for notational simplification.
Step 1. We begin by considering any agent i1 ∈ V. Since Gσ(t) is uniformly jointly strongly
connected, we know that i1 is the root and that there exists a time t1 and an agent i2 ∈ V \ {i1 }
such that (i1 , i2 ) ∈ E during t ∈ [t1 , t1 + τD ) ⊂ [0, T0 ].
We first note that it follows from (23) and Lemma 5 that for all t ∈ [0, N T0 ],
Vij (t, x(t)) ≤ V (t, x(t)) ≤ V ∗ , ∀{i, j} ∈ V × V.
(26)
Taking the derivative of Vij along the trajectories of (1), we obtain that for all t ∈ [t1 , t1 +τD ),
X
ai1 k1 (t)(xk1 − xi1 )
V̇i1 i2 = − Le−2L(t−t0 ) kxi1 − xi2 k2 + e−2L(t−t0 ) (xi1 − xi2 )T
k1 ∈Ni1 (σ(t))
X
−
ai2 k2 (t)(xk2 − xi2 ) + (f (t, xi1 ) − f (t, xi2 ))
k2 ∈Ni2 (σ(t))
≤
X
ai1 k (t)(Vi2 k1 − Vi1 i2 ) − ai2 i1 (t)Vi1 i2 +
k1 ∈Ni1 (σ(t))
X
ai2 k2 (t)(Vi1 k2 − Vi1 i2 )
k2 ∈Ni2 (σ(t))\{i1 }
≤ (N − 1)a∗ (V ∗ − Vi1 i2 ) − a∗ Vi1 i2 + (N − 2)a∗ (V ∗ − Vi1 i2 )
= − α(Vi1 i2 −
(2N − 3)a∗ ∗
V ),
α
where α = (2N − 3)a∗ + a∗ . The first inequality follows from (4) and (24), while the second
inequality follows from (26).
It thus follows that
Vi1 i2 (t1 + τD , x(t1 + τD )) ≤ α̂1 V ∗ ,
where α̂1 = 1 −
a∗
α (1
− e−ατD ) ∈ (0, 1).
15
(27)
Similarly we obtain that for all t ∈ [t1 +τD , N T0 ], V̇i1 i2 ≤ α(V ∗ −Vi1 i2 ), where α = 2(N −1)a∗ .
It then follows from (27) that
Vi1 i2 (t, x(t)) ≤ α∗1 V ∗ ,
∀t ∈ [t1 + τD , N T0 ],
(28)
where α∗1 = 1 − (1 − e−ατD ) aα∗ e−αN T0 ∈ (0, 1).
Step 2. Since Gσ(t) is uniformly jointly strongly connected, we know that that there exists a
time instant t2 and an arc from h ∈ V1 , {i1 , i2 } to i3 ∈ V \ V1 during [t2 , t2 + τD ) ⊂ [T0 , 2T0 ].
We then estimate an upper bound for Vhi3 by considering two different cases: h = i1 and
h = i2 . We eventually obtain that for all t ∈ [t2 + τD , N T0 ],
Vhi3 (t, x(t)) ≤ (1 − β∗2 )V ∗ ,
∀h ∈ V1 .
(29)
where
β∗ = (1 − e−ατD )
a∗ −αN T0
e
∈ (0, 1).
α
(30)
It then follows from (28), (29) and 1 − β∗2 > 1 − β∗ = α∗1 that for all t ∈ [2T0 , N T0 ],
Vi1 k (t, x(t)) ≤ (1 − β∗2 )V ∗ ,
∀k ∈ V2 \{i1 },
where V2 , {i1 , i2 , i3 }.
Step 3. By continuing the above process, we obtain that for all k ∈ V\{i1 },
Vi1 k (N T0 , x(N T0 )) ≤ (1 − β∗N )V ∗ .
(31)
Step 4. Since Gσ(t) is uniformly jointly strongly connected, (31) holds for any i1 ∈ V. By using
the same analysis, we eventually obtain that for all i, j ∈ V,
Vij (N T0 , x(N T0 )) ≤ (1 − β∗N )V ∗ .
Hence the result follows by choosing β̃ = 1 − β∗N with β∗ given by (30).
We are now ready to prove Theorem 4. By using Lemma 6 and (23), we obtain that
V (t, x(t)) ≤ β̃
⌊
t
⌋
N T0
V∗ ≤
1 −ρ∗ t ∗
e
V ,
β̃
where ⌊ NtT ⌋ denotes the largest integer that is not greater than
0
16
t
N T0
and ρ∗ =
1
N T0
ln β̃1 .
It then follows from (23) and (24) that
max
{i,j}∈V×V
kxi (t) − xj (t)k2 ≤
1 −(ρ∗ −2L)t
e
max kxi (0) − xj (0)k2 .
{i,j}∈V×V
β̃
Hence, global exponential synchronization is achieved with γ =
1
β̃
and λ = ρ∗ − 2L provided
that ρ∗ > 2L. This concludes the proof of the desired theorem.
4
Leader-follower Networks
Our focus so far has been on achieving synchronization for leaderless networks. In this section
we consider the synchronization problem for leader-follower networks. Suppose that there is an
additional agent, labeled as agent 0, which plays as a reference or leader for the agents in the
set V. In view of this we also call an agent in V a follower, and denote V̄ = V ∪ {0} as the overall
agent set. The overall communication in the network is described by a time-varying directed
graph Ḡσ(t) = (V̄, E¯σ(t) ). Here for the sake of simplicity we continue to use σ(·) to denote the
piecewise constant graph signal. We also make a standard dwell time assumption [17] on the
switching signal σ(t).
For the leader-follower communication graph, we introduce the following definition.
Definition 4 (i). Ḡσ(t) is leader connected if for any follower agent i ∈ V there is a directed
path from the leader 0 to follower agent i in Ḡσ(t) at time t. Moreover, Ḡσ(t) is jointly leader
connected in the time interval [t1 , t2 ) if the union graph Ḡ([t1 , t2 )) is leader connected.
(ii). Ḡσ(t) is uniformly jointly leader connected if there exists a constant T > 0 such that
the union graph Ḡ([t, t + T )) is leader connected for any t ≥ 0.
(iii). Ḡσ(t) is infinitely jointly leader connected if the union graph Ḡ([t, ∞)) is leader connected for any t ≥ 0.
For leader-follower networks, the evolutions of the follower state xi (t) and the leader state
y(t) are given by
P
ẋi = f (t, xi ) +
j∈Ni (σ(t)) aij (t)(xj − xi ) + bi (t)(y − xi ),
i ∈ V,
(32)
ẏ = f (y, t),
where the nonlinear function f (xi , t) and aij (t) follow from the same definitions as those of the
leaderless case in (1), and bi (t) > 0 is a piecewise continuous function marking the strength of
17
the edge (0, i), if any. Assume that there is a constant b∗ > 0 such that b∗ ≤ bi (t) for all t ≥ 0.
We also assume that the initial time is t = t0 ≥ 0 and denote the initial state for the leader as
y(t0 ) ∈ Rn .
For the leader-follower networks, we are interested in the following synchronization problems.
Definition 5 (i) The multi-agent system (32) achieves global asymptotic synchronization if
limt→∞ (xi (t) − y(t)) = 0 for any i ∈ V, any t0 ≥ 0, any x(t0 ) ∈ RnN , and y(t0 ) ∈ Rn .
(ii) Multi-agent system (32) achieves global exponential synchronization if there exist γ ≥ 1
and λ > 0 such that there exist γ ≥ 1 and λ > 0 such that
max kxi (t) − y(t)k2 ≤ γe−λ(t−t0 ) max kxi (t0 ) − y(t0 )k2 ,
i∈V
i∈V
t ≥ t0 ,
(33)
for any t0 ≥ 0, any x(t0 ) ∈ RnN , and y(t0 ) ∈ Rn .
4.1
Non-expansive Inherent Dynamics
In this section, we extend the results for the case when the agent dynamics is non-expansive to
leader-follower networks. We make the following assumption on the agent dynamics.
Assumption 3 ϕ∗ : Rn → R is a continuously differentiable positive definite function such that
the following conditions hold:
(i). limkηk→∞ ϕ∗ (η) = ∞;
(ii). h∇ϕ∗ (η − ζ), f (t, η) − f (t, ζ)i ≤ 0 for all η, ζ ∈ Rn and t ≥ 0.
Assumption 3 is similar to Assumption 1 however with possibly different functions ϕ∗ (·) and
Assumption 3(ii) holds in the relative coordinate.
The convexity property guarantees the non-expansive property along the leader-follower
multi-agent dynamics (32) as shown in the following lemma whose proof is similar to that of
Lemma 1 and thus omitted.
Lemma 7 Let Assumption 3 hold. Along the leader-follower multi-agent dynamics (32),
maxi∈V ϕ∗ (xi (t)) is non-increasing for all t ≥ 0..
We now state our results for the non-expansive case.
18
Theorem 5 Let Assumption 3 hold. The multi-agent system (32) achieves global asymptotic
synchronization if Ḡσ(t) is uniformly jointly leader connected.
Theorem 6 Let Assumption 3 hold. Assume that Gσ(t) is undirected for all t ≥ t0 . The multiagent system (32) achieves global asymptotic synchronization is achieved if Ḡσ(t) is infinitely
jointly leader connected.
The proofs of Theorems 5 and 6 are given in Appendices A and B, respectively, and based
on a generalization of the methods proposed in [9, 23] however the nonlinear agent dynamics
results in a different Lyapunov function maxi∈V ϕ∗ (x̄i (t)). The analysis are based on estimating
the scalar function ϕ∗ (x̄i (t)) agent by agent and thus yields an estimate of the convergence rate.
They are different from the contradiction arguments used in the proofs of Theorems 1 and 2
where the convergence rate is unclear.
Remark 4 If the scalar function ϕ∗ (·) satisfies an additional condition, i.e., there exist 0 <
c1 ≤ c2 , such that
c1 kηk2 ≤ ϕ∗ (η) ≤ c2 kηk2 , ∀η ∈ Rn ,
then Theorem 5 leads to global exponential synchronization.
Remark 5 For the leader-follower case, Theorems 5 and 6 show that global asymptotic synchronization is achieved while for the leaderless case, while global asymptotic ϕ-synchronization
is achieved as shown in Theorems 1 and 2. For the leader-follower case, the uniformly jointly
leader connected in Theorem 5 requires the leader to be a center node, while for the leaderless
case, the uniformly jointly strongly connected in Theorem 1 requires every node to be a center node. Thus, the connectivity condition of the leaderless case is stronger than that of the
leader-follower case.
4.2
Lipschitz Inherent Dynamics
In this section, we extend the result for the case when the agent dynamics is globally Lipschitz
to leader-follower networks.
Our main result for this case is given in the following theorem whose proof can be found in
Appendix C.
19
Theorem 7 Let Assumption 2 hold. Suppose that Ḡσ(t) is uniformly jointly leader connected.
The multi-agent system (32) achieves global exponential synchronization if L < ρ̂∗ /2, where ρ̂∗
is a constant depending on the network parameters.
5
Conclusions
In this paper, synchronization problems for networks with nonlinear inherent agent dynamics and
switching topologies have been investigated. Two types of nonlinear dynamics were considered:
non-expansive and globally Lipschitz. For the non-expansive case, we found that the convexity
of the Lyapunov function plays a crucial rule in the analysis and showed that the uniformly
joint strong connectivity is sufficient for achieving global asymptotic ϕ-synchronization. When
communication graphs are undirected, the infinitely joint connectivity is a sufficient synchronization condition. Moreover, we established conditions under which ϕ-synchronization implies
state synchronization. For the globally Lipschitz case, we found that joint connectivity alone is
not sufficient to achieve synchronization but established a sufficient synchronization condition.
The proposed condition reveals the relationship between the Lipschitz constant and the network
parameters. The results were also extended to leader-follower networks. An interesting future
direction is to study the synchronization problem for coupled non-identical nonlinear inherent
dynamics under general switching topologies.
References
[1] C. W. Wu and L. O. Chua, “Application of graph theory to the synchronization in an array
of coupled nonlinear oscillators,” IEEE Trans. Circuits Syst. I, Fundam. Theory Appl.,
vol. 42, no. 8, pp. 494–497, 1995.
[2] C. W. Wu, Synchronization in Complex Networks of Nonlinear Dynamical Systems.
Sin-
gapore: World Scientific: Singapore, 2007.
[3] V. Belykh, I. Belykh, and M. Hasler, “Connection graph stability method for synchronized
coupled chaotic systems,” Physica D, vol. 195, no. 1-2, pp. 159–187, 2004.
20
[4] P. DeLellis, M. D. Bernardo, and G. Russo, “On QUAD, Lipschitz, and contracting vector
fields for consensus and synchronization of networks,” IEEE Trans. Circuits Syst. I, Reg.
Papers, vol. 58, no. 3, pp. 576–583, 2011.
[5] W. Yu, G. Chen, and M. Cao, “Consensus in directed networks of agents with nonlinear
dynamics,” IEEE Trans. Autom. Control, vol. 56, no. 6, pp. 1436–1441, 2011.
[6] U. Münz, A. Papachristodoulou, and F. Allgöwer, “Consensus in multi-agent systems with
coupling delays and switching topology,” IEEE Trans. Autom. Control, vol. 56, no. 12, pp.
2976 –2982, 2011.
[7] H. Liu, M. Cao, and C. W. Wu, “Coupling strength allocation for synchronization in complex networks using spectral graph theory,” IEEE Trans. Circuits Syst. I, Reg. Papers,
vol. 61, no. 5, pp. 1520–1530, 2014.
[8] A. Jadbabaie, J. Lin, and A. S. Morse, “Coordination of groups of mobile autonomous
agents using nearest neighbor rules,” IEEE Trans. Autom. Control, vol. 48, no. 6, pp.
988–1001, 2003.
[9] L. Moreau, “Stability of multiagent systems with time-dependent communication links,”
IEEE Trans. Autom. Control, vol. 50, no. 2, pp. 169–182, 2005.
[10] Z. Lin, B. Francis, and M. Maggiore, “State agreement for continuous-time coupled nonlinear systems,” SIAM J. Contr. & Opt., vol. 46, no. 1, pp. 288–307, 2007.
[11] W. Ren and R. W. Beard, “Consensus seeking in multiagent systems under dynamically
changing interaction topologies,” IEEE Trans. Autom. Control, vol. 50, no. 5, pp. 655–661,
2005.
[12] L. Scardovi and R. Sepulchre, “Synchronization in networks of identical linear systems,”
Automatica, vol. 45, no. 11, pp. 2557–2562, 2009.
[13] Y. Su and J. Huang, “Stability of a class of linear switching systems with applications to
two consensus problem,” IEEE Trans. Autom. Control, vol. 57, no. 6, pp. 1420–1430, 2012.
[14] J. Zhao, D. J. Hill, and T. Liu, “Synchronization of complex dynamical networks with
switching topology: a switched system point of view,” Automatica, vol. 45, no. 11, pp.
2502–2511, 2009.
21
[15] J. Qin, H. Gao, and W. Zheng, “Exponential synchronization of complex networks of linear
systems and nonlinear oscillators: a unified analysis,” IEEE Trans. Neural Netw. Learn.
Syst., vol. 61, no. 2, pp. 499–511, 2014.
[16] G. Wen, Z. Duan, G. Chen, and W. Yu, “Consensus tracking of multi-agent systems with
Lipschitz-type node dynamics and switching topologies,” IEEE Trans. Circuits Syst. I, Reg.
Papers, vol. 61, no. 2, pp. 499–511, 2014.
[17] D. Liberzon and A. S. Morse, “Basic problem in stability and design of switched systems,”
IEEE Control Syst. Mag., vol. 19, no. 5, pp. 59–70, 1999.
[18] J. Cortés, “Distributed algorithms for reaching consensus on general functions,” Automatica, vol. 44, no. 3, pp. 726–737, 2008.
[19] X. Wang and Y. Hong, “Distributed finite-time χ-consensus algorithms for multi-agent
systems with variable coupling topology,” Journal of Systems Science and Complexity,
vol. 23, no. 2, pp. 209–218, 2010.
[20] J. Danskin, “The theory of max-min, with applications,” SIAM J. Appl. Math., vol. 14,
no. 6, pp. 641–664, 1996.
[21] S. Boyd and L. Vandenberghe, Convex Optimization. New York, NY: Cambridge University
Press, 2004.
[22] G. Shi, K. H. Johansson, and Y. Hong, “Reaching an optimal consensus: dynamical systems
that compute intersections of convex sets,” IEEE Trans. Autom. Control, vol. 58, no. 3, pp.
610–622, 2013.
[23] G. Shi and K. H. Johansson, “Robust consensus for continuous-time multi-agent dynamics,”
SIAM J. Contr. and Opt., vol. 51, no. 5, pp. 3673–3691, 2013.
A
Proof of Theorem 5
The proof of Theorems 5 relies on the following lemma whose proof is similar to that of Lemmas 3.
22
Lemma 8 Let Assumption 3 hold. Assume that Ḡσ(t) is uniformly jointly leader connected.
Then there exists 0 < ρ̄∗ < 1 such that
ϕ∗ (x̄i (t0 + T ∗ )) ≤ ρ̄∗ max ϕ∗ (x̄i (t0 )),
i∈V
∀i ∈ V,
where T ∗ = N T0 , with T0 given in (7).
Proof: Without loss of generality, we assume the initial time t0 = 0. Similar to the proof
of Lemma 3, we estimate ϕ∗ (x̄i (t)) agent by agent on the subintervals t ∈ [(j − 1)T0 , jT0 ] for
j = 1, . . . , N in several steps.
Step 1. In this step, we focus on an follower agent k1 ∈ V such that (0, k1 ) ∈ Ēσ(t) for t ∈
[t1 , t1 + τD ) ⊆ [0, T0 ]. The existence of such a follower agent k1 and t1 and follows from the fact
Ḡσ(t) is uniformly jointly leader connected. For convenience, we introduce x̄i = xi − y for all
i ∈ V as the relative state from the leader agent. From (32), we obtain the following dynamics.
x̄˙ i = f (t, xi ) − f (t, y) +
X
aij (t)(x̄j − x̄i ) − bi (t)x̄i .
(34)
j∈Ni (σ(t))
Similar to (8), we obtain that for all t ∈ [t1 , t1 + τD ],
d
∗
ϕ∗ (x̄k1 (t)) ≤ a (N − 1) max ϕ∗ (x̄i (0)) − ϕ∗ (x̄k1 ) − b∗ ϕ∗ (x̄k1 ).
i∈V
dt
We then obtain that
ϕ∗ (x̄k1 (t1 + τD )) ≤ δ̂1 max ϕ∗ (x̄k1 (0)),
i∈V
(35)
where
δ̂1 =
λ̂1 − b∗ (1 − e−λ̂1 τD )
λ̂1
,
(36)
with
λ̂1 = a∗ (N − 1) + b∗ .
(37)
For t ∈ [t1 + τD , T ∗ ], the edge (0, k1 ) may no longer exist. Nevertheless, we have for t ∈ [0, T ∗ ],
d
∗
ϕ∗ (x̄k1 (t)) ≤ a (N − 1) max ϕ∗ (x̄i (0)) − ϕ∗ (x̄k1 ) .
(38)
i∈V
dt
It then follows from (35) and (38) that for all t ∈ [T0 , T ∗ ],
ϕ∗ (x̄k1 (t)) ≤ δ1 max ϕ∗ (x̄k1 (0)),
i∈V
23
(39)
where δ1 = 1 − e−λ̄1 N T0 (1 − δ̂1 ) and λ̄1 = a∗ (N − 1).
Step 2. In this step, we analysis a follower agent k2 ∈ V \ {k1 } such that there either an edge
(0, k2 ) or an edge (k1 , k2 ) in Ēσ(t) for t ∈ [t2 , t2 + τD ) ⊆ [T0 , 2T0 ]. Again, the existence of k2 and
t2 due to the uniform joint leader connectivity. By going through the similar analysis as Step 2
of the proof for Lemma 3, we eventually obtain that for t ∈ [2T0 , T ∗ ],
ϕ∗ (x̄j (t)) < δ2 max ϕ∗ (x̄i (0)),
i∈V
j ∈ {k1 , k2 },
where δ2 = 1 − e−λ̄1 (N −1)T0 (1 − δ̂2 ), with
δ̂2 =
λ̄2 − a∗ (1 − δ1 )(1 − e−λ̄2 τD )
λ̄2
and λ̄2 = a∗ (N − 2) + a∗ .
Step 3. By applying the similar analysis on the subintervals [(ℓ − 1)T0 , ℓT0 ] for ℓ = 3, . . . , N , we
obtain that for all t ∈ [ℓT0 , T ∗ ] and for all j ∈ {k1 , . . . , kℓ },
ϕ∗ (x̄j (t)) < δℓ max ϕ∗ (x̄i (0)),
(40)
δℓ = 1 − e−λ̄1 (N −ℓ+1)T0 (1 − δ̂ℓ ),
(41)
i∈V
where
and
δ̂ℓ =
a∗ (N − 2) + a∗ − a∗ (1 − δℓ−1 )(1 − e−λ̄2 τD )
.
a∗ (N − 2) + a∗
(42)
It follows from (41) and (42) that δk ≤ δN for all k = 2, . . . , N . This together with (40) leads
to ϕ∗ (x̄j (T ∗ )) ≤ δN maxi∈V ϕ∗ (x̄i (0)) for all j ∈ V, where
δN = 1 − η N −1 e
(N−2)(N+1)
λ̄1 T0
2
b∗ (1 − e−λ̂1 τD )/(a∗ (N − 1) + b∗ ),
(43)
with
η=
e−λ̄1 (N +1)T0 (1 − e−λ̄2 τD )a∗
.
a∗ (N − 2) + a∗
The result of lemma then follows by choosing ρ̄∗ = δN .
We are ready to prove Theorem 5. Without loss of generality, we assume that t0 = 0, it then
follows from Lemma 8 that maxi∈V ϕ∗ (x̄i (T ∗ )) ≤ ρ̄∗ maxi∈V ϕ∗ (x̄i (0)). Thus, for s = 0, 1, . . ., we
have
max ϕ∗ (x̄i (sT ∗ )) ≤ ρ̄s∗ max ϕ∗ (x̄i (0)).
i∈V
i∈V
24
It then follows that
t
max ϕ∗ (x̄i (t)) ≤ ρ̄∗T ∗
−1
i∈V
max ϕ∗ (x̄i (0)),
i∈V
This together with the fact that ϕ∗ (·) is positive definite as given in Assumption 3 implies that
the multi-agent system (32) achieves global asymptotic synchronization.
B
Proof of Theorem 6
The proof of Theorems 6 relies on the following lemma whose proof is similar to that of Lemmas 3.
Lemma 9 Let Assumption 3 hold. Assume that Ḡσ(t) is infinitely jointly leader connected. Then
there exist 0 < ρ̃∗ < 1, Tp and t̃∗ such that
ϕ∗ (x̄i (t̃∗ + τD )) ≤ ρ̃∗ max ϕ∗ (x̄i (Tp )),
i∈V
∀i ∈ V.
Proof: Since Ḡσ(t) is infinitely jointly leader connected, there exist a sequence of time instants
0 = T0 < T1 < . . . < Tp < Tp+1 < . . .
(44)
Tp , tp1 < tp2 < . . . < tpN+1 , Tp+1
(45)
such that
for p = 0, 1, . . ., and Ḡ([tpℓ , tpℓ+1 )) is leader connected for ℓ = 1, . . . , N . Moreover, each edge
in Ḡ([tpℓ , tpℓ+1 )) exists for at least the dwell time τD during [tpℓ , tpℓ+1 ) for p = 0, 1, . . . and
ℓ = 1, . . . , N .
We shall estimate ϕ∗ (x̄i (t)) agent by agent on the subintervals [tpℓ , tpℓ+1 ], ℓ = 1, . . . , N for
the interval [Tp , Tp+1 ], p = 0, 1, . . ..
Step 1. In this step, we focus all the follower agent i ∈ V1 , where
V1 , {i ∈ V | (0, i) ∈ Ēσ(t1 ) }.
and
t1 ,
inf
t∈[tp1 ,tp2 )
{∃ i ∈ V | (0, i) ∈ E¯σ(t) }.
The existence of V1 and t1 due to the fact that Ḡ([tp1 , tp2 )) is leader connected. It then follows
from the similar analysis as we obtained (35) that
ϕ∗ (x̄k1 (t1 + τD )) ≤ δ̂1 max ϕ∗ (x̄i (Tp )),
i∈V
25
∀k1 ∈ V1 ,
(46)
where δ̂1 is given by (36).
Step 2. In this step, similar to Step 3 of the proof for Lemma 4, we view the set {0} ∪ V1 as
a subsystem. Define t2 as the first time when there is an edge between this subsystem and the
remaining follower agents and V2 accordingly.
By going through the similar analysis as Step 2 of the proof for Lemma 3, we eventually
obtain that for j ∈ V1 ∪ V2 ,
ϕ∗ (x̄j (t2 + τD )) < δ̃2 max ϕ∗ (x̄j (Tp )),
i∈V
where
δ̃2 =
λ̄2 − a∗ e−λ̄1 τD (1 − δ̃1 )(1 − e−λ̄2 τD )
,
λ̄2
(47)
with
δ̃1 = 1 − e−λ̄1 τD (1 − δ̂1 ),
(48)
and δ̂1 given by (36).
Step 3. Since Ḡσ(t) is infinitely jointly leader connected, we proceed the above analysis until
V = V1 ∪ . . . ∪ Vm0 for some m0 ≤ N such that
ϕ∗ (x̄j (tm0 + τD )) < δ̃m0 max ϕ∗ (x̄i (Tp )),
i∈V
∀j ∈ V,
with tm0 defined similarly to t1 and t2 , and
δ̃ℓ =
λ̄2 − a∗ e−λ̄1 τD (1 − δ̃ℓ−1 )(1 − eλ̄2 τD )
,
λ̄2
ℓ = 3, . . . , m0 .
From the preceding relation and (47), we obtain for ℓ = 2, . . . , m0 ,
a∗ e−λ̄1 τD (1 − e−λ̄2 τD )
1 − δ̃ℓ
=
, η̃ < 1.
a∗ (N − 2) + a∗
1 − δ̃ℓ−1
It is then easy to see that δ̃ℓ−1 < δ̃ℓ for all ℓ = 2, . . . , m0 . This together with (36), (48), and
m0 ≤ N implies that for all j ∈ V,
ϕ∗ (x̄j (tm0 + τD )) ≤ δ̃N max ϕ∗ (x̄i (Tp )),
i∈V
where
δ̃N = 1 − η̃ N −1 b∗ e−λ̄1 τD (1 − e−λ̂1 τD )/(a∗ (N − 1) + b∗ ),
with λ̂1 given by (37). The result then follows by choosing Tp as defined in (44) and (45),
t̃∗ = tm0 and ρ̃∗ = δ̃N .
26
We are ready to prove Theorem 5. Without loss of generality, we assume that t0 = 0, it
follows then follows from Lemma 7, Lemma 9, and the fact that t̃∗ + τD ≤ Tp+1 for t̃∗ = tm0 ,
which follows from the definition of Tp+1 given in (44) and (45), that for p = 0, 1, . . .
max ϕ∗ (x̄i (Tp+1 )) ≤ max ϕ∗ (x̄i (tm0 + τD )) ≤ ρ̃∗ max ϕ∗ (x̄i (Tp )).
i∈V
i∈V
i∈V
Thus, we have for s = 0, 1, . . ., maxi∈V ϕ∗ (x̄i (Ts )) ≤ ρ̃s∗ maxi∈V ϕ∗ (x̄i (0)). This together with the
fact that ϕ∗ (·) is positive definite as given in Assumption 3 implies that the multi-agent system
(32) achieves global asymptotic synchronization.
C
Proof of Theorem 7
The proof is similar to that of Theorem 4 however in the relative coordinate x̄i = xi − y whose
evolution is given by (34), Again, without loss of generality, we assume the initial time t0 = 0.
The proof is based on the convergence analysis of the nonnegative scalar
V (t, x̄(t)) = max Vi (t, x̄i (t)),
(49)
i∈V
where x̄(t) = [x̄T1 (t), x̄T2 (t), . . . , x̄TN (t)]T and
1
Vi (t, x̄i (t)) = e−2Lt kx̄i (t)k2 ,
2
∀i ∈ V.
(50)
Let us define I(t) = {i ∈ V : Vi (t, x̄i (t)) = V (t, x̄(t))}. Similar to Lemma 5, we obtain that
D + V (t, x̄(t)) ≤ 0 for all t ≥ 0 along the multi-agent dynamics (32). By combining the preceding
relation we have Vi (t, x̄i ) ≤ V (t, x̄) ≤ V (t0 , x̄(0)) , V∗ , for all t ≥ t0 and all i ∈ V. Following
the similar analysis as the proof of Theorem 5, we can show that
V (t, x̄(t)) ≤
1 −ρ̂∗ t
e
V∗ ,
δN
where δN is given by (43) and ρ̂∗ is given by (51),
ρ̂∗ =
1
1
ln
.
∗
T
δN
(51)
with T ∗ = (N − 1)T0 . It then follows that
max kx̄i (t)k2 ≤
i∈V
1 −(ρ̂∗ −2L)t
e
max kx̄i (0)k2 .
i∈V
δN
Hence, global exponential synchronization is achieved with γ =
that ρ̂∗ > 2L.
27
1
δN
and λ = ρ̂∗ − 2L provided
| 3cs.SY
|
Neural Networks for Joint Sentence Classification
in Medical Paper Abstracts
Franck Dernoncourt∗
MIT
[email protected]
Ji Young Lee∗
MIT
[email protected]
arXiv:1612.05251v1 [cs.CL] 15 Dec 2016
Abstract
Existing models based on artificial neural networks (ANNs) for sentence classification often do not incorporate the context in which sentences appear, and classify sentences individually. However, traditional sentence classification approaches
have been shown to greatly benefit from
jointly classifying subsequent sentences,
such as with conditional random fields. In
this work, we present an ANN architecture
that combines the effectiveness of typical
ANN models to classify sentences in isolation, with the strength of structured prediction. Our model achieves state-of-theart results on two different datasets for sequential sentence classification in medical
abstracts.
1
Introduction
Over 50 million scholarly articles have been published (Jinha, 2010), and the number of articles published every year keeps increasing (Druss
and Marcus, 2005; Larsen and Von Ins, 2010).
Approximately half of them are biomedical papers. While this repository of human knowledge
abounds with useful information that may unlock
new, promising research directions or provide conclusive evidence about phenomena, it has become
increasingly difficult to take advantage of all available information due to its sheer amount. Therefore, a technology that can assist a user to quickly
locate the information of interest is highly desired,
as it may reduce the time required to locate relevant information.
When researchers search for previous literature,
for example, they often skim through abstracts in
order to quickly check whether the papers match
∗
These authors contributed equally to this work.
Peter Szolovits
MIT
[email protected]
their criteria of interest. This process is easier
when abstracts are structured, i.e., the text in an
abstract is divided into semantic headings such as
objective, method, result, and conclusion. However, a significant portion of published paper abstracts is unstructured, which makes it more difficult to quickly access the information of interest.
Therefore, classifying each sentence of an abstract
to an appropriate heading can significantly reduce
time to locate the desired information.
We call this the sequential sentence classification task, in order to distinguish it from general
text classification or sentence classification that
does not have any context. Besides aiding humans,
this task may also be useful for automatic text
summarization, information extraction, and information retrieval.
In this paper, we present a system based on
ANNs for the sequential sentence classification
task. Our model makes use of both token and
character embeddings for classifying sentences,
and has a sequence optimization layer that is
learned jointly with other components of the
model. We evaluate our model on the NICTAPIBOSO dataset as well as a new dataset we compiled based on the PubMed database.
2
Related Work
Existing systems for sequential sentence classification are mostly based on naive Bayes
(NB) (Ruch et al., 2007; Huang et al., 2013),
support vector machines (SVMs) (McKnight and
Srinivasan, 2003; Yamamoto and Takagi, 2005;
Hirohata et al., 2008), Hidden Markov models
(HMMs) (Lin et al., 2006), and conditional random fields (CRFs) (Kim et al., 2011; Hassanzadeh et al., 2014; Hirohata et al., 2008). They
often require numerous hand-engineered features
based on lexical (bag-of-words, n-grams, dic-
tionaries, cue words), semantic (synonyms, hyponyms), structural (part-of-speech tags, headings), and sequential (sentence position, surrounding features) information.
On the other hand, recent approaches to natural language processing (NLP) based on artificial neural networks (ANNs) do not require manual features, as they are trained to automatically
learn features based on word as well as character embeddings. Moreover, ANN-based models
have achieved state-of-the-art results on various
NLP tasks. For short-text classification, many
ANN models use word embeddings (Socher et al.,
2013; Kim, 2014; Kalchbrenner et al., 2014), and
most recent works are based on character embeddings (Zhang et al., 2015; Conneau et al., 2016;
Xiao and Cho, 2016). Dos Santos and Gatti (2014)
use both word and character embeddings.
However, most existing works using ANNs for
short-text classification do not use any context.
This is in contrast with sequential sentence classification, where each sentence in a text is classified
taking into account its context. The context utilized for the classification could be the surrounding sentences or possibly the whole text. One exception is a recent work on dialog act classification (Lee and Dernoncourt, 2016), where each utterance in a dialog is classified into its dialog act,
but only the preceding utterances were used, as the
system was designed with real-time applications
in mind.
3
yj
y2
…
a1
yn-1
yn
an-1
an
…
aj
a2
Feed forward
s
bi-LSTM
concatanate
e1
…
…
…
…
concatenate
c
bi-LSTM
em
ei
e2
t
concatanate
Token
embeddings
…
…
c1
c2
cl-1
cl
Character embeddings
z1
z2
zl-1
zl
x
Figure 1: ANN model for sequential sentence classification. x: token, t: token embeddings (300), zi : ith character
of x, ci : character embeddings (25), c: character-based token embeddings (50), ei : hybrid token embeddings (350), s:
sentence vector (200), aj : sentence label vector (number of
classes), yj : sentence label. The numbers in parenthesis indicate the dimensions of the vectors. Token embeddings are
initialized with GloVe (Pennington et al., 2014) embeddings
pretrained on Wikipedia and Gigaword 5 (Parker et al., 2011).
Model
In the following, we denote scalars in italic lowercase (e.g., k, bf ), vectors in bold lowercase
(e.g., s, xi ), and matrices in italic uppercase
(e.g., Wf ) symbols. We use the colon notations
xi:j and vi:j to denote the sequences of scalars
(xi , xi+1 , . . . , xj ) and vectors (vi , vi+1 , . . . , vj ),
respectively.
3.1
y1
ANN model
Our ANN model (Figure 1) consists of three components: a hybrid token embedding layer, a sentence label prediction layer, and a label sequence
optimization layer.
3.1.1 Hybrid token embedding layer
The hybrid token embedding layer takes a token
as an input and outputs its vector representation
utilizing both the token embeddings and as well as
the character embeddings.
Token embeddings are a direct mapping VT (·)
from token to vector, which can be pre-trained on
large unlabeled datasets using programs such as
word2vec (Mikolov et al., 2013b; Mikolov et al.,
2013a; Mikolov et al., 2013c) or GloVe (Pennington et al., 2014). Character embeddings are also
defined in an analogous manner, as a direct mapping VC (·) from character to vector.
Let z1:` be the sequence of characters that comprise a token x. Each character zi is first mapped
to its embedding ci = VC (zi ), and the resulting
sequence c1:` is input to a bidirectional LSTM,
which outputs the character-based token embedding c.
The output e of the hybrid token embedding
layer for the token x is the concatenation of the
character-based token embedding c and the token
embedding t = VT (x).
3.1.2 Sentence label prediction layer
Let x1:m be the sequence of tokens in a given sentence, and e1:m be the corresponding embedding
output from the hybrid token embedding layer.
The sentence label prediction layer takes as input the sequence of vectors e1:m , and outputs a,
where the k th element of a, denoted a[k], reflects
the probability that the given sentence has label k.
To achieve this, the sequence e1:m is first input
to a bidirectional LSTM, which outputs the vector
representation s of the given sentence. The vector s is subsequently input to a feedforward neural
network with one hidden layer, which outputs the
corresponding probability vector a.
3.1.3 Label sequence optimization layer
The label sequence optimization layer takes the sequence of probability vectors a1:n from the label
prediction layer as input, and outputs a sequence
of labels y1:n , where yi is the label assigned to the
token xi .
In order to model dependencies between subsequent labels, we incorporate a matrix T that contains the transition probabilities between two subsequent labels; we define T [i, j] as the probability
that a token with label i is followed by a token with
the label j. The score of a label sequence y1:n is
defined as the sum of the probabilities of individual labels and the transition probabilities:
s(y1:n ) =
n
X
i=1
ai [yi ] +
n
X
T [yi−1 , yi ].
i=2
These scores can be turned into probabilities of the
label sequences by taking a softmax function over
all possible label sequences. During the training
phase, the objective is to maximize the log probability of the gold label sequence. In the testing
phase, given an input sequence of tokens, the corresponding sequence of predicted labels is chosen
as the one that maximizes the score.
4
4.1
Experiments
Datasets
We evaluate our model on the sentence classification task using the following two medical abstract
datasets, where each sentence of the abstract is annotated with one label. Table 1 presents statistics
on each dataset.
NICTA-PIBOSO This dataset was introduced
in (Kim et al., 2011) and was the basis of the
ALTA 2012 Shared Task (Amini et al., 2012).
PubMed 20k RCT We assembled this corpus
consisting of randomized controlled trials (RCTs)
from the PubMed database of biomedical literature, which provides a standard set of 5 sentence
labels: objectives, background, methods, results
and conclusions.
Dataset |C| |V |
Train
Validation
Test
PubMed 5 68k 15k (195k) 2.5k (33k) 2.5k (33k)
NICTA
6 17k 722 (8k)
77 (0.9k)
200 (2k)
Table 1: Dataset overview. |C| denotes the number of
classes, |V | the vocabulary size. For the train, validation and
test sets, we indicate the number of number of abstracts followed by the number of sentences in parentheses.
4.2
Training
The model is trained using stochastic gradient descent, updating all parameters, i.e., token embeddings, character embeddings, parameters of bidirectional LSTMs, and transition probabilities, at
each gradient step. For regularization, dropout
with a rate of 0.5 is applied to the characterenhanced token embeddings and before the label
prediction layer.
5
Results and Discussion
Table 2 compares our model against several baselines as well as the best performing model (Lui,
2012) in the ALTA 2012 Shared Task, in which
8 competing research teams participated to build
the most accurate classifier for the NICTAPIBOSO corpus.
The first baseline (LR) is a classifier based on
logistic regression using n-gram features extracted
from the current sentence: it does not use any information from the surrounding sentences. The
second baseline (Forward ANN) uses the model
presented in (Lee and Dernoncourt, 2016): it computes sentence embeddings for each sentence, then
classifies the current sentence given a few preceding sentence embeddings as well as the current
sentence embedding. The third baseline (CRF) is
a CRF that uses n-grams as features: each output variable of the CRF corresponds to a label for
a sentence, and the sequence the CRF considers
is the entire abstract. The CRF baseline therefore uses both preceding and succeeding sentences
when classifying the current sentence. Lastly, the
model presented in (Lui, 2012) developed a new
NICTA
71.6
75.1
81.2
82.0
82.7
Background
Conclusion
Methods
Objectives
Results
Total
Support
3621
4571
9897
2333
9713
30135
Table 3: Detailed results of our model on the PubMed 20k
RCT dataset.
End -0.24
0.14
-0.10
0.08
-0.23 -0.01 -0.20
Start -0.26 -0.08 -0.14
0.26
0.23
0.02
Objectives -0.03
0.13
-0.23
0.11
0.27
-0.04 -0.19
Background -0.02
0.08
-0.38
0.35
0.07
-0.06
0.04
Conclusion -0.24 -0.14
0.28
0.01
-0.10 -0.01
0.22
-0.13
0.20
-0.18 -0.20 -0.02 -0.02
< -0.3
En
rou
-0.1
Ob
sio
ckg
Ba
clu
0.1
-0.08
d
Results 0.25
nd
jec
tiv
es
Sta
rt
-0.24 -0.17
n
0.05
ds
0.25
Co
n
0.04
> 0.3
-0.02
Methods 0.17
tho
approach called feature stacking, which is a metalearner that combines multiple feature sets, and
is the best performing system on NICTA-PIBOSO
published in the literature.
The LR system performs honorably on PubMed
20k RCT (F1-score: 83.0), but quite poorly on
NICTA-PIBOSO (F1-score: 71.6): this suggests
that using the surrounding sentences may be more
important in NICTA-PIBOSO than in PubMed
20k RCT.
The Forward ANN system performs better than
the LR system, and worse than the CRF: this is unsurprising, as the Forward ANN system only uses
the information from the preceding sentences but
does not use any information from the succeeding
sentences, unlike the CRF.
Our model performs better than the CRF system and the (Lui, 2012) system. We hypothesize
that the following four factors give an edge to our
model.
No human-engineered features: Unlike most
other systems, our model does not rely on any
human-engineered features.
No n-grams: While other systems heavily rely
on n-grams, our model maps each token to a token embedding, and feeds it as an input to an
RNN. This helps combat data scarcity: for example, “chronic tendonitis” and “chronic tendinitis”
are two different bigrams, but their token embeddings should be very similar since they share the
same meaning.
Structured prediction: The labels for all sentences in an abstract are predicted jointly, which
improves the coherence between the predicted labels in a given abstract.
Joint learning: Our model learned the features
and token embeddings jointly with the sequence
optimization.
PubMed 20k RCT
Recall F1-score
88.2
79.1
92.9
93.2
96.2
94.9
48.1
59.6
93.1
93.9
89.8
89.9
Precision
71.8
93.5
93.7
78.2
94.8
90.0
lts
Table 2: F1-scores on the test set with several baselines, the
best published method (Lui, 2012) from the literature, and
our model. Since PubMed 20k was introduced in this work,
there is no previous best published method for this dataset.
The presented results for the ANN-based models are the F1scores on the test set of the run with the highest F1-score on
the validation set.
Label
Me
PubMed 20k
83.0
86.1
89.3
–
89.9
Re
su
Model
LR
Forward ANN
CRF
Best published
Our model
Figure 2: Transition matrix learned on PubMed 20k. The
rows represent the label of the previous sentence, the columns
represent the label of the current sentence.
Figure 2 presents an example of a transition matrix after the model has been trained on PubMed
20k RCT. We can see that it effectively reflects
transitions between different labels. For example,
it learned that the first sentence of an abstract is
most likely to be either discussing objective (0.23)
or background (0.26). By the same token, a sentence pertaining to the methods is typically followed by a sentence pertaining to the methods
(0.25) or the results (0.17).
Table 3 details the result of our model for each
label in PubMed 20k RCT: the main difficulty the
classifier has is distinguishing background sentences from objective sentences.
6
Conclusions
In this article we have presented an ANN architecture to classify sentences that appear in sequence.
We demonstrate that jointly predicting the classes
of all sentences in a given text improves the quality
of the predictions and yields better performance
than a CRF. Our model achieves state-of-the-art
results on two datasets for sentence classification
in medical abstracts.
References
[Amini et al.2012] Iman Amini, David Martinez, and
Diego Molla. 2012. Overview of the ALTA 2012
Shared Task. In Australasian Language Technology
Association Workshop 2012, volume 7, page 124.
[Kim2014] Yoon Kim. 2014. Convolutional neural
networks for sentence classification. In Proceedings of the 2014 Conference on Empirical Methods
in Natural Language Processing (EMNLP), pages
1746–1751. Association for Computational Linguistics (ACL).
[Conneau et al.2016] Alexis
Conneau,
Holger
Schwenk, Loı̈c Barrault, and Yann Lecun.
2016.
Very deep convolutional networks for
natural language processing.
arXiv preprint
arXiv:1606.01781.
[Larsen and Von Ins2010] Peder Olesen Larsen and
Markus Von Ins. 2010. The rate of growth in scientific publication and the decline in coverage provided by science citation index. Scientometrics,
84(3):575–603.
[dos Santos and Gatti2014] Cı́cero Nogueira dos Santos and Maira Gatti. 2014. Deep convolutional neural networks for sentiment analysis of short texts.
In International Conference on Computational Linguistics (COLING), pages 69–78.
[Lee and Dernoncourt2016] Ji Young Lee and Franck
Dernoncourt. 2016. Sequential short-text classification with recurrent and convolutional neural networks. In Human Language Technologies 2016:
The Conference of the North American Chapter
of the Association for Computational Linguistics,
NAACL HLT.
[Druss and Marcus2005] Benjamin G Druss and
Steven C Marcus. 2005. Growth and decentralization of the medical literature: implications for
evidence-based medicine. Journal of the Medical
Library Association, 93(4):499.
[Hassanzadeh et al.2014] Hamed Hassanzadeh, Tudor
Groza, and Jane Hunter. 2014. Identifying scientific artefacts in biomedical literature: The evidence
based medicine use case. Journal of biomedical informatics, 49:159–170.
[Hirohata et al.2008] Kenji Hirohata, Naoaki Okazaki,
Sophia Ananiadou, Mitsuru Ishizuka, and Manchester Interdisciplinary Biocentre. 2008. Identifying sections in scientific abstracts using conditional
random fields. In International Joint Conference
on Natural Language Processing (IJCNLP), pages
381–388.
[Huang et al.2013] Ke-Chun Huang, I-Jen Chiang,
Furen Xiao, Chun-Chih Liao, Charles Chih-Ho Liu,
and Jau-Min Wong. 2013. PICO element detection in medical text without metadata: Are first sentences enough? Journal of biomedical informatics,
46(5):940–946.
[Jinha2010] Arif E Jinha. 2010. Article 50 million: an
estimate of the number of scholarly articles in existence. Learned Publishing, 23(3):258–263.
[Lin et al.2006] Jimmy Lin, Damianos Karakos, Dina
Demner-Fushman, and Sanjeev Khudanpur. 2006.
Generative content models for structural analysis of
medical abstracts. BioNLP06 Linking Natural Language Processing and Biology: Towards Deeper Biological Literature Analysis, 6:65–72.
[Lui2012] Marco Lui. 2012. Feature stacking for sentence classification in evidence-based medicine. In
Australasian Language Technology Workshop 2012:
ALTA Shared Task, page 134.
[McKnight and Srinivasan2003] Larry McKnight and
Padmini Srinivasan. 2003. Categorization of sentence types in medical abstracts. In American Medical Informatics Association (AMIA).
[Mikolov et al.2013a] Tomas Mikolov, Kai Chen, Greg
Corrado, and Jeffrey Dean. 2013a. Efficient estimation of word representations in vector space. arXiv
preprint arXiv:1301.3781.
[Mikolov et al.2013b] Tomas Mikolov, Ilya Sutskever,
Kai Chen, Greg S Corrado, and Jeff Dean. 2013b.
Distributed representations of words and phrases
and their compositionality. In Advances in neural
information processing systems, pages 3111–3119.
[Mikolov et al.2013c] Tomas Mikolov, Wen-tau Yih,
and Geoffrey Zweig. 2013c. Linguistic regularities
in continuous space word representations. In HLTNAACL, pages 746–751.
[Kalchbrenner et al.2014] Nal Kalchbrenner, Edward
Grefenstette, and Phil Blunsom. 2014. A convolutional neural network for modelling sentences. In
Proceedings of the 52nd Annual Meeting of the Association for Computational Linguistics. Proceedings of the 52nd Annual Meeting of the Association
for Computational Linguistics.
[Parker et al.2011] Robert Parker, David Graff, Junbo
Kong, Ke Chen, and Kazuaki Maeda. 2011. English
Gigaword fifth edition. Technical report, Linguistic
Data Consortium, Philadelphia.
[Kim et al.2011] Su Nam Kim, David Martinez,
Lawrence Cavedon, and Lars Yencken. 2011.
Automatic classification of sentences to support
evidence based medicine. BioMed Central (BMC)
Bioinformatics, 12(2):1.
[Pennington et al.2014] Jeffrey Pennington, Richard
Socher, and Christopher D Manning. 2014. GloVe:
global vectors for word representation. Proceedings
of the Empiricial Methods in Natural Language Processing (EMNLP 2014), 12:1532–1543.
[Ruch et al.2007] Patrick Ruch, Celia Boyer, Christine Chichester, Imad Tbahriti, Antoine Geissbühler,
Paul Fabry, Julien Gobeill, Violaine Pillet, Dietrich
Rebholz-Schuhmann, Christian Lovis, et al. 2007.
Using argumentation to extract key sentences from
biomedical abstracts. International journal of medical informatics, 76(2):195–200.
[Socher et al.2013] Richard Socher, Alex Perelygin,
Jean Y Wu, Jason Chuang, Christopher D Manning,
Andrew Y Ng, and Christopher Potts. 2013. Recursive deep models for semantic compositionality over
a sentiment treebank. In Proceedings of the conference on empirical methods in natural language processing (EMNLP), volume 1631, page 1642. Citeseer.
[Xiao and Cho2016] Yijun Xiao and Kyunghyun Cho.
2016. Efficient character-level document classification by combining convolution and recurrent layers.
arXiv preprint arXiv:1602.00367.
[Yamamoto and Takagi2005] Yasunori Yamamoto and
Toshihisa Takagi. 2005. A sentence classification system for multi biomedical literature summarization. In 21st International Conference on Data
Engineering Workshops (ICDEW’05), pages 1163–
1163. IEEE.
[Zhang et al.2015] Xiang Zhang, Junbo Zhao, and Yann
LeCun. 2015. Character-level convolutional networks for text classification. In Advances in Neural
Information Processing Systems (NIPS), pages 649–
657.
| 9cs.NE
|
Quantum Predicative Programming
Anya Tafliovich and E.C.R. Hehner
arXiv:quant-ph/0602156v1 17 Feb 2006
University of Toronto
Abstract. The subject of this work is quantum predicative programming — the study of developing of programs intended for execution on
a quantum computer. We look at programming in the context of formal
methods of program development, or programming methodology. Our
work is based on probabilistic predicative programming, a recent generalisation of the well-established predicative programming. It supports
the style of program development in which each programming step is
proven correct as it is made. We inherit the advantages of the theory,
such as its generality, simple treatment of recursive programs, time and
space complexity, and communication. Our theory of quantum programming provides tools to write both classical and quantum specifications,
develop quantum programs that implement these specifications, and reason about their comparative time and space complexity all in the same
framework.
1
Introduction
Modern physics is dominated by concepts of quantum mechanics. Today, over
seventy years after its recognition by the scientific community, quantum mechanics provides the most accurate known description of nature’s behaviour.
Surprisingly, the idea of using the quantum mechanical nature of the world to
perform computational tasks is very new, less than thirty years old. Quantum
computation and quantum information is the study of information processing
and communication accomplished with quantum mechanical systems. In recent
years the field has grown immensely. Scientists from various fields of computer
science have discovered that thinking physically about computation yields new
and exciting results in computation and communication. There has been extensive research in the areas of quantum algorithms, quantum communication
and information, quantum cryptography, quantum error-correction, adiabatic
computation, measurement-based quantum computation, theoretical quantum
optics, and the very new quantum game theory. Experimental quantum information and communication has also been a fruitful field. Experimental quantum
optics, ion traps, solid state implementations and nuclear magnetic resonance
all add to the experimental successes of quantum computation.
The subject of this work is quantum programming — the study of developing
programs intended for execution on a quantum computer. We assume a model
of a quantum computer proposed by Knill [24]: a classical computer with access
to a quantum device that is capable of storing quantum bits, performing certain
2
operations and measurements on these bits, and reporting the results of the
measurements.
We look at programming in the context of formal methods of program development, or programming methodology. This is the field of computer science
concerned with applications of mathematics and logic to software engineering
tasks. In particular, the formal methods provide tools to formally express software specifications, prove correctness of implementations, and reason about various properties of specifications (e.g. implementability) and implementations (e.g.
time and space complexity). Today formal methods are successfully employed in
all stages of software development, such as requirements elicitation and analysis,
software design, and software implementation.
In this work the theory of quantum programming is based on probabilistic
predicative programming, a recent generalisation of the well-established predicative programming [19,20], which we deem to be the simplest and the most elegant
programming theory known today. It supports the style of program development
in which each programming step is proven correct as it is made. We inherit the
advantages of the theory, such as its generality, simple treatment of recursive
programs, and time and space complexity. Our theory of quantum programming provides tools to write both classical and quantum specifications, develop
quantum programs that implement these specifications, and reason about their
comparative time and space complexity all in the same framework.
The rest of this work is organised as follows. Section 2.1 is the introduction
to quantum computation. It assumes that the reader has some basic knowledge
of linear algebra and no knowledge of quantum computing. Section 2.2 contains the introduction to probabilistic predicative programming. The reader is
assumed to have some background in logic, but no background in programming
theory is necessary. The contribution of this work is section 3 which defines
the quantum system, introduces programming with the quantum system, and
several well-known problems, their classical and quantum solutions, and their
formal comparative time complexity analyses. Section 4 states conclusions and
outlines directions for future research.
1.1
Related work
Traditionally, quantum computation is presented in terms of quantum circuits.
Recently, there has been an attempt to depart from this convention for the same
reason that classical computation is generally not presented in terms of classical
circuits. As we develop more complex quantum algorithms, we will need ways to
express higher-level concepts with control structures in a readable fashion.
In 2000 Ömer [28] introduced the first quantum programming language QCL.
Following his work, Bettelli et. al. developed a quantum programming language
with syntax based on C++. These two works did not involve any verification
techniques.
Sanders and Zuliani in [29] introduced a quantum language qGCL, which is an
extension of pGCL [26], which in turn generalises Dijkstra’s guarded-command
language to include probabilism. Zuliani later extends this attempt at formal
3
program development and verification in [36], which discusses treatment of nondeterminism in quantum programs, and in [38], where the attempt is made to
build on Aharonov’s work to reason about mixed states computations. Zuliani
also provides tools to approach the task of compiling quantum programs in [37].
A large amount of work in the area was performed in the past two years.
In [4], [25], and [22] process algebraic approaches were explored. Tools developed in the field of category theory were successfully employed by [1], [2], [3],
[11], [30], and others to reason about quantum computation. In [7] and [8] a functional language with semantics in a form of a term rewrite system is introduced
and a notion of linearity and how it pertains to quantum systems are examined.
A functional language QML with design guided by its categorical semantics is
defined in [5]. Following on this work, [6] provides a sound and complete equational theory for QML. Weakest preconditions appropriate for quantum computation are introduced in [15]. This work is interesting, in part, because it diverts
from the standard approach of reducing quantum computation to probabilistic
one. It also provides semantics for the language of [30]. Other interesting work
by the same authors include reasoning about knowledge in quantum systems
([16]) and developing a formal model for distributed measurement-based quantum computation ([12]). A similar work is introduced in [17], where a language
CQP for modelling communication in quantum systems is defined. The latter
approaches have an advantage over process algebraic approaches mentioned earlier in that they explicitly allow a quantum state to be transmitted between
processes. Building of the work of [31], [33] defines a higher order quantum programming language based on a linear typed lambda calculus, which is similar to
the work of [34].
1.2
Our contribution
Our approach to quantum programming amenable to formal analysis is very
different from almost all of those described above. Work of [29], [36], [38] is the
only one which is similar to our work. The contribution of this paper is twofold.
Firstly, by building our theory on that in [20], we inherit the advantages it offers. The definitions of specification and program are simpler: a specification is
a boolean (or probabilistic) expression and a program is a specification. The
treatment of recursion is simple: there is no need for additional semantics of
loops. The treatment of termination simply follows from the introduction of a
time variable; if the final value of the time variable is ∞, then the program is
a non-terminating one. Correctness and time and space complexity are proved
in the same fashion; moreover, after proving them separately, we naturally obtain the conjunction. Secondly, the way Probabilistic Predicative Programming
is extended to Quantum Predicative Programming is simple and intuitive. The
use of Dirac-like notation makes it easy to write down specifications and develop
algorithms. The treatment of computation with mixed states does not require
any additional mechanisms. Quantum Predicative Programming fully preserves
Predicative Programming’s treatment of parallel programs and communication,
4
which provides for a natural extension to reason about quantum communication protocols, such as BB84 ([9]), distributed quantum algorithms, such as distributed Shor’s algorithm ([35]), as well as their time, space, and entanglement
complexity.
2
2.1
Preliminaries
Quantum Computation
In this section we introduce the basic concepts of quantum mechanics, as they
pertain to the quantum systems that we will consider for quantum computation.
The discussion of the underlying physical processes, spin- 21 -particles, etc. is not of
our interest. We are concerned with the model for quantum computation only. A
reader not familiar with quantum computing can consult [27] for a comprehensive
introduction to the field.
The Dirac notation, invented by Paul Dirac, is often used in quantum mechanics. In this notation a vector v (a column vector by convention) is written
inside a ket : |vi. The dual vector of |vi is hv|, written inside a bra. The inner products are bra-kets hv|wi. For n-dimensional vectors |ui and |vi and m-dimensional
vector |wi, the value of the inner product hu|vi is a scalar and the outer product
operator |vihw| corresponds to an m by n matrix. The Dirac notation clearly
distinguishes vectors from operators and scalars, and makes it possible to write
operators directly as combinations of bras and kets.
In quantum mechanics, the vector spaces of interest are the Hilbert spaces of
dimension 2n for some n ∈ N. A convenient orthonormal basis is what is called
a computational basis, in which we label 2n basis vectors using binary strings of
length n as follows: if s is an n-bit string which corresponds to the number xs ,
then |si is a 2n -bit (column) vector with 1 in position xs and 0 everywhere else.
The tensor product |ii ⊗ |ji can be written simply as |iji. An arbitrary vector
in a Hilbert space can be written as a weighted sum of the computational basis
vectors.
Postulate 1 (state space) Associated to any isolated physical system is a
Hilbert space, known as the state space of the system. The system is completely described by its state vector, which is a unit vector in the system’s
state space.
Postulate 2 (evolution) The evolution of a closed quantum system is described by a unitary transformation.
Postulate 3 (measurement) Quantum measurements are described by a collection {Mm } of measurement operators, which act on the state space of the
system being measured. The index m refers to the possible measurement
outcomes. If the state of the system immediately prior to the measurement
is described by a vector |ψi, then the probability of obtaining result m is
†
Mm |ψi, in which case the state of the system immediately after the
hψ|Mm
5
measurement is described by the vector √
operators satisfy the completeness equation
Mm |ψi
. The measurement
†
hψ|Mm
P Mm |ψi
†
m · Mm
Mm
= I.
An important special class of measurements is projective measurements, which
are equivalent to general measurements provided that we also have the ability
to perform unitary transformations.
A projective measurement is described by an observable M , which is a Hermitian operator on the state space of the
P system being measured. This observable
has a spectral decomposition M =
m · λm × Pm , where Pm is the projector
onto the eigenspace of M with eigenvalue λm , which corresponds to the outcome
of the measurement. The probability of measuring m is hψ|Pm |ψi, in which case
immediately after the measurement the system is found in the state √ Pm |ψi .
hψ|Pm |ψi
Given an orthonormal basis |vm i, 0 ≤ m < 2n , measurement with respect to
this basis
P is the corresponding projective measurement given by the observable
M = m · λm × Pm , where the projectors are Pm = |vm ihvm |.
Measurement with respect to the computational basis is the simplest and the
most commonly used class of measurements. In terms of the basis |mi, 0 ≤ m <
2n , the projectors are Pm = |mihm| and hψ|Pm |ψi = |ψm |2 . The state of the
system immediately after measuring m is |mi.
In the case of a single qubit, for example, measurement of the state α ×
|0i + β × |1i results in the outcome 0 with probability |α|2 and outcome 1 with
probability |β|2 . The state of the system immediately after the measurement is
|0i or |1i, respectively.
Suppose the result of the measurement is ignored and we continue the computation. In this case the system is said to be in a mixed state. A mixed state is
not the actual physical state of the system. Rather it describes our knowledge
of the state the system is in. In the above example, the mixed state is expressed
by the equation |ψi = |α|2 × {|0i} + |β|2 × {|1i}. The equation is meant to
say that |ψi is |0i with probability |α|2 and it is |1i with probability |β|2 . An
application of operation U to the mixed state results in another mixed state,
U (|α|2 × {|0i} + |β|2 × {|1i}) = |α|2 × {U |0i} + |β|2 × {U |1i}.
Postulate 4 (composite systems) The state space of a composite physical
system is the tensor product of the state spaces of the component systems.
If we have systems numbered 0 up to and excluding n, and each system i,
0 ≤ i < n, is prepared in the state |ψi i, then the joint state of the composite
system is |ψ0 i ⊗ |ψ1 i ⊗ . . . ⊗ |ψn−1 i.
While we can always describe a composite system given descriptions of the component systems, the reverse is not true. Indeed, given a state vector that describes a composite system, it may not be possible to factor it to obtain the
state vectors
systems. A well-known example is the state
√
√ of the component
|ψi = |00i/ 2 + |11i/ 2. Such state is called an entangled state.
6
2.2
Probabilistic Predicative Programming
This section introduces the programming theory of our choice, on which our work
on quantum programming is based — probabilistic predicative programming.
We briefly introduce parts of the theory necessary for understanding section 3 of
this work. For a course in predicative programming the reader is referred to [19].
Introduction to probabilistic predicative programming can be found in [20].
Predicative programming In predicative programing a specification is a
boolean expression. The variables in a specification represent the quantities of
interest, such as prestate (inputs), poststate (outputs), and computation time
and space. We use primed variables to describe outputs and unprimed variables
to describe inputs. For example, specification x′ = x + 1 in one integer variable x
states that the final value of x is its initial value plus 1. A computation satisfies
a specification if, given a prestate, it produces a poststate, such that the pair
makes the specification true. A specification is implementable if for each input
state there is at least one output state that satisfies the specification.
We use standard logical notation for writing specifications: ∧ (conjunction), ∨
(disjunction), ⇒ (logical implication), = (equality, boolean equivalence), 6= (nonequality, non-equivalence), and if then else. == and =⇒ are the same as = and
⇒, but with lower precedence. We use standard mathematical notation, such as
+ − ∗ / mod. We use lowercase letters for variables of interest and uppercase
letters for specifications.
In addition to the above, we use the following notations: σ (prestate), σ ′
(poststate), ok (σ ′ = σ), and x := e (x′ = e ∧ y ′ = y ∧ . . .). ok specifies that
the values of all variables are unchanged. In the assignment x := e, x is a state
variable (unprimed) and e is an expression (in unprimed variables) in the domain
of x.
If R and S are specifications in variables x, y, . . . , R′′ is obtained from R
by substituting all occurrences of primed variables x′ , y ′ , . . . with double-primed
variables x′′ , y ′′ , . . . , and S ′′ is obtained from S by substituting all occurrences
of unprimed variables x, y, . . . with double-primed variables x′′ , y ′′ , . . . , then the
sequential composition of R and S is defined by
R; S == ∃x′′ , y ′′ , . . . · R′′ ∧ S ′′
.
Various laws can be proven about sequential composition. One of the most
important ones is the substitution law, which states that for any expression e of
the prestate, state variable x, and specification P ,
x := e; P == (for x substitute e in P )
Specification S is refined by specification P if and only if S is satisfied whenever P is satisfied:
∀σ, σ ′ · S ⇐ P
7
Specifications S and P are equal if and only if they are satisfied simultaneously:
∀σ, σ ′ · S = P
Given a specification, we are allowed to implement an equivalent specification
or a stronger one.
Informally, a bunch is a collection of objects. It is different from a set, which
is a collection of objects in a package. Bunches are simpler than sets; they don’t
have a nesting structure. See [20] for an introduction to bunch theory. A bunch of
one element is the element itself. We use upper-case to denote arbitrary bunches
and lower-case to denote elements (an element is the same as a bunch of one
element). A, B denotes the union of bunches A and B. A : B denotes bunch
inclusion — bunch A is included in bunch B. We use notation x, ..y to mean
from (including) x to (excluding) y.
If x is a fresh (previously unused) name, D is a bunch, and b is an arbitrary
expression, then λx : D · b is a function of a variable (parameter) x with domain
D and body b. If f is a function, then ∆f denotes the domain of f . If x : ∆f ,
then f x (f applied to x) is the corresponding element in the range. A function of
n variables is a function of 1 variable, whose body is a function of n− 1 variables,
for n > 0. A predicate is function whose body is a boolean expression. A relation
is a function whose body is a predicate. A higher-order function is a function
whose parameter is a function.
A quantifier is a unary prefix operator that applies to functions. If p is a
predicate, then ∀p is the boolean result, obtained by first applying p to all the
elements in its domain and then taking the conjunction of those results. Taking
the disjunction
of the results produces ∃p. Similarly, if f is a numeric function,
P
then
f is the numeric result, obtained by first applying f to all the elements
in its domain and then taking the sum ofPthose results.
For example, applyingPthe quantifier
to the function λi : 0, ..2n · |ψi|2 , for
n
2
some functionP
ψ, yields:
λi : 0, ..2 · |ψi| , which for the sake of simplicity we
abbreviate to i : 0, ..2n ·|ψi|2 . In addition, we allow a few other simplifications.
For example, we can omit the domain of a variable if it is clear from the context.
P
We canP
also group variables from several quantifications.
For example,
i :
P
0, ..2n · j : 0, ..2n · 2−m−n can be abbreviated to
i, j : 0, ..2n · 2−m−n .
A program is an implemented specification. For simplicity we only take the
following to be implemented: ok, assignment, if then else, sequential composition, booleans, numbers, bunches, and functions.
Given a specification S, we proceed as follows. If S is a program, there is no
work to be done. If it is not, we build a program P , such that P refines S, i.e.
S ⇐ P . The refinement can proceed in steps: S ⇐ . . . ⇐ R ⇐ Q ⇐ P .
One of the best features of Hehner’s theory, is its simple treatment of recursion. In S ⇐ P it is possible for S to appear in P . No additional rules are
required to prove the refinement. For example, it is trivial to prove that
x ≥ 0 ⇒ x′ = 0 ⇐= if x = 0 then ok else (x := x − 1; x ≥ 0 ⇒ x′ = 0)
8
The specification says that if the initial value of x is non-negative, its final
value must be 0. The solution is: if the value of x is zero, do nothing, otherwise
decrement x and repeat.
How long does the computation take? To account for time we add a time
variable t. We use t to denote the time, at which the computation starts, and t′
to denote the time, at which the computation ends. In case of non-termination,
t′ = ∞. This is the only characteristic by which we distinguish terminating
programs from non-terminating ones. See [21] for a discussion on treatment of
termination. We choose to use a recursive time measure, in which we charge 1
time unit for each time P is called. We replace each call to P to include the time
increment as follows:
P ⇐= if x = 0 then ok else (x := x − 1; t := t + 1; P )
It is easy to see that t is incremented the same number of times that x is
decremented, i.e. t′ = t + x, if x ≥ 0, and t′ = ∞, otherwise. Just as above, we
can prove:
x ≥ 0 ∧ t′ = t + x ∨ x < 0 ∧ t′ = ∞
⇐= if x = 0 then ok
else (x := x − 1; t := t + 1; x ≥ 0 ∧ t′ = t + x ∨ x < 0 ∧ t′ = ∞)
Probabilistic predicative programming Probabilistic predicative programming was introduced in [19] and was further developed in [20]. It is a generalisation of predicative programming that allows reasoning about probability
distributions of values of variables of interest. Although in this work we apply
this reasoning to boolean and integer variables only, the theory does not change
if we want to work with real numbers: we replace summations with integrals.
A probability is a real number between 0 and 1, inclusive. A distribution is an
expression whose value is a probability and whose sum over all values of variables
is 1. For example, if n is a positive natural
then 2−n is a distribution,
Pvariable,
since for any n, 2−n is a probability, and
n · 2−n = 1. In two positive natural
variables m and n, 2−n−m is also a distribution. If a distribution of several
variables can be written as a product of distributions of the individual variables,
then the variables are independent. For example, m and n in the previous example
are independent. Given a distribution of several variables, we can sum out some
of the variables
to obtain a distribution of the rest of the variables. In our
P
example,
n · 2−n−m = 2−m , which is a distribution of m.
To generalise boolean specifications to probabilistic specifications, we use
1 and 0 for boolean true and f alse, respectively.1 If S is an implementable
deterministic specification and p is a distribution of the initial state x, y, ..., then
the distribution of the final state is
X
x, y, ... · S × p
1
Readers familiar with ⊤ and ⊥ notation can notice that we take the liberty to equate
⊤ = 1 and ⊥ = 0.
9
If R and S are specifications in variables x, y, . . . , R′′ is obtained from R by
substituting all occurrences of primed variables x′ , y ′ , . . . with double-primed
variables x′′ , y ′′ , . . . , and S ′′ is obtained from S by substituting all occurrences
of unprimed variables x, y, . . . with double-primed variables x′′ , y ′′ , . . . , then the
sequential composition of R and S is defined by
X
R; S ==
x′′ , y ′′ , . . . · R′′ × S ′′
If p is a probability and R and S are distributions, then
if p then R else S == p × R + (1 − p) × S
Various laws can be proven about sequential composition. One of the most
important ones, the substitution law, introduced earlier, applies to probabilistic
specifications as well.
To implement a probabilistic specification we use a pseudo-random number
generator. Since we cannot, even in theory, produce a real random number generator by means of traditional computing, we assume that a pseudo-random
number generator generates truly random numbers and we simply refer to it
as random number generator. For a positive natural variable n, we say that
rand n produces a random natural number uniformly distributed in 0, ..n. To
reason about the values supplied by the random number generator consistently,
we replace every occurrence of rand n with a fresh variable r whose value has
probability (r : 0, ..n)/n. If rand occurs in a context such as r = rand n, we
replace the equation by r : (0, ..n)/n. If rand occurs in the context of a loop, we
parametrise the introduced variables by the execution time.
Recall the earlier example. Let us change the program slightly by introducing
probabilism:
P ⇐= if x = 0 then ok else (x := x − rand 2; t := t + 1; P )
In the new program at each iteration x is either decremented by 1 or it is unchanged, with equal probability. Our intuition tells us that the revised program
should still work, except it should take longer. Let us prove it. We replace rand
with r : time → (0, 1) with rt having probability 1/2. Ignoring time we can
prove:
x ≥ 0 ⇒ x′ = 0
⇐= if x = 0 then ok else (x := x − rand 2; x ≥ 0 ⇒ x′ = 0)
As for the execution time, we can prove that it takes at least x time units to
complete:
t′ ≥ t + x
⇐= if x = 0 then ok else (x := x − rand 2; t := t + 1; t′ ≥ t + x)
How long should we expect to wait for the execution to complete? In other
words, what is the distribution of t′ ? Consider the following distribution of the
10
final states:
(0 = x′ = x = t′ − t) + (0 = x′ < x ≤ t′ − t) ×
n!
n
where
=
m
m! × (n − m)!
′
t −t−1
1
× t′ −t ,
2
x−1
We can prove that:
X
rt ·
1
×
2
if x = 0 then ok
else
x := x − rt; t := t + 1;
(0 = x′ = x = t′ − t) +
′
1
t −t−1
× t′ −t
2
x−1
′
1
t
−
t
−
1
== (0 = x′ = x = t′ = t) + (0 = x′ < x ≤ t′ − t) ×
× t′ −t
2
x−1
(0 = x′ < x ≤ t′ − t) ×
Now, since for positive x, t′ is distributed according to the negative binomial
distribution with parameters x and 12 , its mean value is
X
′
1
t −t−1
′
′
t · (t − t) × (0 = x = t − t) + (0 < x ≤ t − t) ×
× t′ −t
2
x−1
′
′
== 2 × x + t
Therefore, we should expect to wait 2 × x time units for the computation to
complete.
3
Quantum Predicative Programming
This section is the contribution of the paper. Here we define the quantum system, introduce programming with the quantum system and several well-known
problems, their classical and quantum solutions, and their formal comparative
time complexity analyses. The proofs of refinements are omitted for the sake of
brevity. The reader is referred to [32] for detailed proofs of some of the algorithms.
3.1
The quantum system
Let C be the set of all complex numbers with the absolute value operator | · |
∗
and the complex conjugate operator
n-qubit system is a
P . Then na state2 of an
n
function ψ : 0, ..2 → C, such that
x : 0, ..2 · |ψx| = 12 .
2
We should point out that this kind of function operations is referred to as lifting
11
If ψ and φ are two states of an n-qubit system, then their inner product,
hψ|φi : C, is defined by:
X
hψ|φi =
x : 0, ..2n · (ψx)∗ × (φx)
A basis of an n-qubit system is a collection of 2n quantum states b0,..2n , such
that ∀i, j : 0, ..2n · hbi |bj i = (i = j).
We adopt the following Dirac-like notation for the computational basis: if
x : 0, ..2n , then x denotes the corresponding n-bit binary encoding of x and
|xi : 0, ..2n → C is the following quantum state:
|xi = λi : 0, ..2n · (i = x)
If ψ is a state of an m-qubit system and φ is a state of an n-qubit system,
then ψ ⊗ φ, the tensor product of ψ and φ, is the following state of a composite
m + n-qubit system:
ψ ⊗ φ = λi : 0, ..2m+n · ψ(i div 2n ) × φ(i mod 2n )
We write ⊗n to mean tensored with itself n times.
An operation defined on a n-qubit quantum system is a higher-order function,
whose domain and range are maps from 0, ..2n to the complex numbers. An
identity operation on a state of an n-qubit system is defined by
I n = λψ : 0, ..2n → C · ψ
For a linear operation A, the adjoint of A, written A† , is the (unique) operation,
such that for any two states ψ and φ, hψ|Aφi = hA† ψ|φi. The unitary transformations that describe the evolution of a n-qubit quantum system are operations
U defined on the system, such that U † U = I n . In this setting, the tensor product
of operators is defined in the usual way. If ψ is a state of an m-qubit system, φ
is a state of an n-qubit system, and U and V are operations defined on m and
n-qubit systems, respectively, then the tensor product of U and V is defined on
an m + n qubit system by (U ⊗ V )(ψ ⊗ φ) = (U ψ) ⊗ (V φ).
Just as with tensor products of states, we write U ⊗n to mean operation U
tensored with itself n times.
Suppose we have a system of n qubits in state ψ and we measure it. Suppose
also that we have a variable r from the domain 0, ..2n , which we use to record the
result of the measurement, and variables x, y, . . ., which are not affected by the
measurement. Then the measurement corresponds to a probabilistic specification
that gives the probability distribution of ψ ′ and r′ (these depend on ψ and on
the type of measurement) and states that the variables x, y, . . . are unchanged.
For a general quantum measurement described by a collection MP= M0,..2n of
measurement operators, which satisfy the completeness equation
m : 0, ..2n ·
†
Mm Mm = I, the specification is measureM ψ r, where
′
ψ
M
r
× (σ ′ = σ)
measureM ψ r == hψ|Mr†′ Mr′ ψi × ψ ′ = q
†
hψ|Mr′ Mr′ ψi
12
where σ ′ = σ is an abbreviation of (x′ = x) × (y ′ = y) × . . . and means “all other
other variables are unchanged”. To obtain the distribution of, say, r′ we sum
out the rest of the variables as follows:
X
′
ψ
M
r
×(σ ′ = σ)
ψ ′ , x′ , y ′ , . . . · hψ|Mr†′ Mr′ ψi × ψ ′ = q
†
hψ|Mr′ Mr′ ψi
== hψ|Mr†′ Mr′ ψi
P
For projective measurements defined by an observable O =
m · λm × Pm ,
where Pm is the projector on the eigenspace of O with eigenvalue λm :
!
Pr′ ψ
′
measureO ψ r == hψ|Pr′ ψi × ψ = p
× (σ ′ = σ)
hψ|Pr′ ψi
Given an arbitrary orthonormal basis B = b0,..2n , measurement of ψ in basis B
is:
measureB ψ r == |hbr′ |ψi|2 × (ψ ′ = br′ ) × (σ ′ = σ)
Finally, the simplest and the most commonly used measurement in the computational basis is:
measure ψ r == |ψr′ |2 × (ψ ′ = |r’i) × (σ ′ = σ)
In this case the distribution of r′ is |ψr′ |2 and the distribution of the quantum
state is:
X
r′ · |ψr′ |2 × (ψ ′ = |r’i)
which is precisely the mixed quantum state that results from the measurement.
In order to develop quantum programs we need to add to our list of implemented things from section 2.2. We add variables of type quantum state as
above and we allow the following three kinds of operations on these variables.
If ψ is a state of an n-qubit quantum system, r is a natural variable, and M is
a collection of measurement operators that satisfy the completeness equation,
then:
1. ψ := |0i⊗n is a program
2. ψ := U ψ, where U is a unitary transformation on an n-qubit system, is a
program
3. measureM ψ r is a program
The special cases of measurements, described in section 2.1, are therefore also
allowed: for an observable O, measureO q r is a program; for an orthonormal
basis B, measureB q r is a program; finally, measure q r is a program.
13
3.2
Deutsch-Jozsa algorithm
Deutsch-Jozsa problem ([14]), an extension of Deutsch’s Problem ([13]), is an
example of the broad class of quantum algorithms that are based on quantum
Fourier transform ([23]). The task is: given a function f : 0, ..2n → 0, 1 , such
that f is either constant or balanced, determine which case it is. Without any
restrictions on the number of calls to f , we can write the specification (let us
call it S) as follows:
(f is constant ∨ f is balanced) =⇒ b′ = f is constant
where b is a boolean variable and the informally stated properties of f are defined
formally as follows:
f is constant == ∀i : 0, ..2n · f i = f 0
X
f is balanced ==
i : 0, ..2n · (−1)f i = 0
It is easy to show that
(f is constant ∨ f is balanced)
=⇒ (f is constant == ∀i : 0, ..2n−1 + 1 · f i = f 0)
In our setting, we need to implement the specification R defined as follows:
b′ == ∀i : 0, ..2n−1 + 1 · f i = f 0
The Hadamard transform, widely used in quantum algorithms, is defined on a
1-qubit system and in our setting is a higher-order function from 0, 1 → C to
0, 1 → C:
√
H = λψ : 0, 1 → C · i : 0, 1 · (ψ0 + (−1)i × ψ1)/ 2
The operation H ⊗n on a n-qubit system applies H to every qubit of the system.
Its action on a zero state of an n-qubit system is:
X
√
H ⊗n |0i⊗n =
x : 0, ..2n · |xi/ 2n
On a general state |xi, the action of H ⊗n is:
H ⊗n |xi =
X
√
y : 0, ..2n · (−1)x·y × |yi/ 2n
where x · y is the bitwise inner product of x and y modulo 2 (bitwise XOR).
Another important definition is that of the quantum analog of a classical
oracle f :
Uf = λψ : 0, 1 → C · x : 0, 1 · (−1)f x × ψx
14
The quantum solution, in one quantum variable ψ and an integer variable r
is:
ψ := |0i⊗n ; ψ := H ⊗n ψ; ψ := Uf ψ; ψ := H ⊗n ψ; measure ψ r; b := (r′ = 0)
Let us add to the specification a restriction on the number of calls to the oracle
by introducing a time variable. Suppose the new specification is:
(f is constant ∨ f is balanced =⇒ b′ = f is constant) ∧ (t′ = t + 1)
where we charge 1 unit of time for each call to the oracle and all other operations
are free. Clearly, the above quantum solution works. Classically the specification
is unimplementable. In fact, the strongest classically implementable specification
is
(f is constant ∨ f is balanced =⇒ b′ = f is constant) ∧ (t′ = t + 2n−1 + 1)
3.3
Grover’s search
Grover’s quantum search algorithm ([18]) is well-known for the quadratic speedup it offers in the solutions of NP-complete problems. The algorithm is optimal
up to a multiplicative constant ([10]). The task is: given a function f : 0, ..2n →
0, 1, find x : 0, ..2n , such that f x = 1. For simplicity we assume that there is only
a single solution, which we denote x1 , i.e. f x1 = 1 and f x = 0 for all x 6= x1 .
The proofs are not very different for a general case of more than one solutions.
As before, we use a general quantum oracle, defined by
Uf |xi = (−1)f x × |xi
In addition, we define the inversion about mean operator as follows:
M : (0, ..N → C) → (0, ..N → C)
X
M ψ == λx : 0, ..N · 2 ×
i : 0, ..N · ψi/N − ψx
where N = 2n .
Grover’s algorithm initialises the quantum system to an equally weighted
superposition of all basis states |xi, x : 0, ..N . It then repeatedly applies Uf
followed by M to the system. Finally, the state is measured. The probability of
error is determined by the number of iterations performed by the algorithm.
The algorithm is easily understood with the help of a geometric analysis of
the operators. Let α be the sum over all x, which are not solutions, and let β be
the solution:
X
1
×
x 6= x1 · |xi
α= √
N −1
β = |x1 i
15
Then the oracle Uf performs a reflection about the vector α in the plane defined
by α and β. In other words, Uf (a × α + b × β) = a × α − b × β. Similarly, the
inversion about mean operator is a reflection about the vector ψ in the plane
defined by α and β. Therefore, the result of Uf followed by M is a rotation in
this plane. The quantum solution, in a quantum variable ψ, natural variables r,
i, and k, and time variable t, is,
2
p
S == sin (2 × (t′ − t) + 1) × arcsin 1/N
× (r′ = x1 ) +
2
p
′
×
1 − sin (2 × (t − t) + 1) × arcsin 1/N
(r′ 6= x1 )/(N − 1)
== P ; measure ψ r
P ⇐= i := 0; ψ := |0i⊗n ; ψ := H ⊗n ψ; R
R ⇐= if i = k then ok
else (i := i + 1; t := t + 1; ψ := Uf ψ; ψ := M ψ; R)
Specification S carries a lot of useful information. For example, it tells us that
the probability of finding a solution after k iterations is
2
p
sin((2 × k + 1) × arcsin 1/N )
Or we might ask how many iterations should be performed to minimise the
probability of an error. Examining first and second derivatives,p
we find that the
above probability is minimised when t′ − t = (π × i)/(4 × arcsin 1/N ) − 1/2 for
integer i. Of course, the number of iterations performed must be a natural number. It is interesting to note that probability of error is periodic in the number of
iterations, but since we don’t gain anything by performing extra iterations, we
pick i = 1. Finally, assuming 1 ≪ N = 2n ,√we obtain
an elegant approximation
to the optimal number of iterations: π × 2n /4 , with the probability of error
approximately 1/2n .
3.4
Computing with Mixed States
As we have discussed in section 2.1, the state of a quantum system after a
measurement is traditionally described as a mixed state. An equation ψ =
{|0i}/2 + {|1i}/2 should be understood as follows: the state ψ is |0i with probability 1/2 and it is |1i with probability 1/2. In contrast to a pure state, a mixed
state does not describe a physical state of the system. Rather, it describes our
knowledge of in what state the system is.
In our framework, there is no need for an additional mechanism to compute
with mixed states. Indeed, a mixed state is not a system state, but a distribution
over system states, and all our programming notions apply to distributions.
The above mixed state is the following distribution over a quantum state ψ:
(ψ = |0i)/2 + (ψ = |1i)/2. This expression tells us, for each possible value in
16
the domain of ψ, the probability of ψ having that value. For example, ψ is
the state |0i with probability (|0i = |0i)/2 + (|0i = |1i)/2, which is 1/2; it is
|1i with probability (|1i = |0i)/2 + (|1i = |1i)/2, which is also 1/2; for any
scalars α and β, not equal to 0 or 1, ψ is α × |0i + β × |1i with probability
(α × |0i + β × |1i = |0i)/2 + (α × |0i + β × |1i = |1i)/2, which is 0. One way to
obtain this distribution is to measure an equally weighted superposition of |0i
and |1i:
√
√
ψ ′ = |0i/ 2 + |1i/ 2; measure ψ r
√
√
== ψ = |0i/ 2 + |1i/ 2; |ψr′ |2 × (ψ ′ = |r’i)
measure
′
==
X
sequential composition
√
√
′′ ′ 2
r , ψ · (ψ = |0i/ 2 + |1i/ 2) × |ψ r | × (ψ ′ = |r’i)
′′
′′
′′
one point law
√
√
== |(|0i/ 2 + |1i/ 2) r′ |2 × (ψ ′ = |r’i)
== (ψ ′ = |r’i)/2
Distribution of the quantum state is then:
X
r′ · (ψ ′ = |r’i)/2 == (ψ ′ = |0i)/2 + (ψ ′ = |1i)/2
as desired.
Similarly, there is no need to extend the application of unitary operators.
Consider the following toy program:
ψ := |0i; ψ := Hψ; measure ψ r; if r = 0 then ψ := Hψ else ok
In the second application of Hadamard the quantum state is mixed, but this is
not evident from the syntax of the program. It is only in the analysis of the final
quantum state that the notion of a mixed state is meaningful. The operator is
applied to a (pure) system state, though we are unsure what that state is.
ψ := |0i; ψ := Hψ; measure ψ r ;
if r = 0 then ψ := Hψ else ok
as before
′
== (ψ = |r’i)/2;
if r = 0 then ψ := Hψ else ok
sequential composition
X
==
r′′ , ψ ′′ · (ψ ′′ = |r”i)/2 ×
((r′′ = 0) × (ψ ′ = Hψ ′′ ) × (r′ = r′′ ) +
′
(r′′ = 1) × (ψ ′ = ψ ′′ ) × (r′ = r′′ ))
′
′
′
== ((ψ = H|0i) × (r = 0) + (ψ = |1i) × (r = 1)) /2
√
√
== (ψ ′ = |0i/ 2 + |1i/ 2) × (r′ = 0)/2 +
(ψ ′ = |1i) × (r′ = 1)/2
one point law
17
The distribution of the quantum state after the computation is:
X
√
√
r′ · (ψ ′ = |0i/ 2 + |1i/ 2) × (r′ = 0)/2 + (ψ ′ = |1i) × (r′ = 1)/2
√
√
== (ψ ′ = |0i/ 2 + |1i/ 2)/2 + (ψ ′ = |1i)/2
A lot of properties of measurements and mixed states can be proven from the
definitions of measurement and sequential composition. For example, the fact
that a measurement in the computational basis, performed immediately following
a measurement in the same basis, does not change the state of the system and
yields the same result as the first measurement with probability 1, is proven as
follows:
measure ψ r; measure ψ r
′ 2
′
′ 2
measure
′
== |ψ r | × (ψ = |r’i); |ψ r | × (ψ = |r’i)
sequential composition
X
==
ψ ′′ , r′′ · |ψ r′′ |2 × (ψ ′′ = |r”i) × |ψ ′′ r′ |2 × (ψ ′ = |r’i) one point law
== |ψ r′ |2 × (ψ ′ = |r’i)
measure
== measure ψ r
In case of a general quantum measurement, the proof is similar, but a little more
computationally involved.
4
Conclusion and Future Work
We have presented a new approach to developing, analysing, and proving correctness of quantum programs. Since we adopt Hehner’s theory as the basis for
our work, we inherit its advantageous features, such as simplicity, generality, and
elegance. Our work extends probabilistic predicative programming in the same
fashion that quantum computation extends probabilistic computation. We have
provided tools to write quantum as well as classical specifications, develop quantum and classical solutions for them, and analyse various properties of quantum
specifications and quantum programs, such as implementability, time and space
complexity, and probabilistic error analysis uniformly, all in the same framework.
Current research an research in the immediate future involve reasoning about
distributed quantum computation. Current work involves expressing quantum
teleportation, dense coding, and various games involving entanglement, in a way
that makes complexity analysis of these quantum algorithms simple and natural. These issues will be described in a forthcoming paper. We can easily express
teleportation as refinement of a specification φ′ = ψ, for distinct qubits φ and
ψ, in a well-known fashion. However, we are more interested in the possibilities
of simple proofs and analysis of programs involving communication, both via
quantum channels and exhibiting the LOCC (local operations, classical communication) paradigm. Future work involves formalising quantum cryptographic
protocols, such as BB84 [9], in our framework and provide formal analysis of
these protocols. This will naturally lead to formal analysis of distributed quantum algorithms (e.g. distributed Shor’s algorithm of [35]).
18
References
1. S. Abramsky. High-level methods for quantum computation and information. In
Proceedings of the 19th Annual IEEE Symposium on Logic in Computer Science,
2004.
2. S. Abramsky and B. Coecke. A categorical semantics of quantum protocols. In
LICS 2004, 2004.
3. S. Abramsky and R. Duncan. A categorical quantum logic. In QPL 2004, pages
3–20, 2004.
4. P. Adao and P. Mateus. A process algebra for reasoning about quantum security.
In QPL 2005, 2005.
5. T. Altenkirch and J. Grattage. A functional quantum programming language. In
Proceedings of the 20th Annual IEEE Symposium on Logic in Computer Science,
2005.
6. T. Altenkirch, J. Grattage, J. K. Vizzotto, and A. Sabry. An algebra of pure
quantum programming. In QPL 2005, 2005.
7. P. Arrighi and G. Dowek. Operational semantics for formal tensorial calculus. In
QPL 2004, pages 21–38, 2004.
8. P. Arrighi and G. Dowek. Linear-algebraic lambda-calculus. In QPL 2005, 2005.
9. C. H. Bennet and G. Brassard. Quantum cryptography: Public key distribution
and coin tossing. In IEEE Int. Conf. Computers, Systems and Signal Processing,
pages 175–179, 1984.
10. M. Boyer, G. Brassard, P. Høyer, and A. Tapp. Tight bounds on quantum searching. In Fortschritte der Physik, pages 493–506, 1998.
11. B. Coecke. The logic of entanglement. 2004. quant-ph/0402014.
12. V. Danos, E. D’Hondt, E. Kashefi, and P. Panangaden. Distributed measurementbased quantum computation. In QPL 2005, 2005.
13. D. Deutsch. Quantum theory, the Church-Turing principle and the universal quantum computer. In Proceedings of the Royal Society of London, pages 97–117, 1985.
14. D. Deutsch and R. Jozsa. Rapid solution of problems by quantum computation.
Proceedings of the Royal Society of London, 439:553–558, 1992.
15. E. D’Hondt and P. Panangaden. Quantum weakest precondition. In QPL 2004,
pages 75–90, 2004.
16. E. D’Hondt and P. Panangaden. Reasoning about quantum knowledge. 2005.
quant-ph/0507176.
17. S. J. Gay and R. Nagarajan. Communicating quantum processes. In Proceedings
of the 32nd ACM SIGACT-SIGPLAN Symposium on Principles of Programming
Languages, 2005.
18. L. K. Grover. A fast quantum mechanical algorithm for database search. In
Twenty-Eighth Annual ACM Symposium on Theory of Computing, pages 212–219,
1996.
19. E.C.R. Hehner. a Practical Theory of Programming. Springer, New York, second
edition, 2004. Available free at www.cs.utoronto.ca/~hehner/aPToP.
20. E.C.R. Hehner. Probabilistic predicative programming. In Mathematics of Program
Construction, 2004.
21. E.C.R. Hehner. Retrospective and prospective for unifying theories of programming. In Symposium on Unifying Theories of Programming, 2006.
22. P. Jorrand and M. Lalire. Toward a quantum process algebra. In Proceedings of
the 1st ACM Conference on Computing Frontiers, 2004.
19
23. R. Jozsa. Quantum algorithms and the fourier transform. Proceedings of the Royal
Society of London, pages 323–337, 1998.
24. E. Knill. Conventions for quantum pseudocode. Technical Report LAUR-96-2724,
Los Alamos National Laboratory, 1996.
25. M. Lalire and P. Jorrand. A process algebraic approach to concurrent and distributed quantum computation: operational semantics. In QPL 2004, pages 109–
126, 2004.
26. C. Morgan and A. McIver. pQCL: formal reasoning for random algorithms. South
African Computer Journal, 22:14–27, 1999.
27. M. A. Nielsen and I. L. Chuang. Quantum Computation and Quantum Information.
Cambridge University Press, 2000.
28. B. Ömer. Quantum programming in QCL. Master’s thesis, TU Vienna, 2000.
29. J. W. Sanders and P. Zuliani. Quantum programming. In Mathematics of Program
Construction, pages 80–99, 2000.
30. P. Selinger. Towards a quantum programming language. Mathematical Structures
in Computer Science, 2004.
31. P. Selinger. Towards a semantics for higher-order quantum computation. In QPL
2004, 2004.
32. A. Tafliovich. Quantum programming. Master’s thesis, University of Toronto,
2004.
33. B. Valiron. Quantum typing. In QPL 2004, pages 163–178, 2004.
34. A. van Tonder. A lambda calculus for quantum computation. SIAM Journal on
Computing, 33(5):1109–1135, 2004.
35. A. Yimsiriwattana and S. J. Lomonaco Jr. Distributed quantum computing: A
distributed shor algorithm. 2004. quant-ph/0403146.
36. P. Zuliani. Non-deterministic quantum programming. In QPL 2004, pages 179–
195, 2004.
37. P. Zuliani. Compiling quantum programs. Acta Informatica, 41(7-8):435–474, 2005.
38. P. Zuliani. Quantum programming with mixed states. In QPL 2005, 2005.
| 6cs.PL
|
Are we Done with Object Recognition? The iCub robot’s
Perspective.
arXiv:1709.09882v1 [cs.RO] 28 Sep 2017
Giulia Pasquale · Carlo Ciliberto · Francesca Odone · Lorenzo Rosasco ·
Lorenzo Natale
Abstract We report on an extensive study of the current benefits and limitations of deep learning approaches
to robot vision and introduce a novel dataset used for
our investigation. To avoid the biases in currently available datasets, we consider a human-robot interaction
setting to design a data-acquisition protocol for visual
object recognition on the iCub humanoid robot. Considering the performance of off-the-shelf models trained
on off-line large-scale image retrieval datasets, we show
the necessity for knowledge transfer. Indeed, we analyze different ways in which this last step can be done,
and identify the major bottlenecks in robotics scenarios.
By studying both object categorization and identification tasks, we highlight the key differences between object recognition in robotics and in image retrieval tasks,
for which the considered deep learning approaches have
been originally designed. In a nutshell, our results confirm also in the considered setting the remarkable improvements yield by deep learning, while pointing to
specific open challenges that need to be addressed for
seamless deployment in robotics.
Keywords Humanoid Robotics · iCub · Machine
Learning · Deep Learning · Object Recognition ·
Transfer Learning · Dataset · Invariance
G. Pasquale · C. Ciliberto · L. Rosasco · L. Natale
iCub Facility,
Laboratory for Computational and Statistical Learning,
Istituto Italiano di Tecnologia
Via Morego 30, 16163, Genova, IT
E-mail: [email protected]
G. Pasquale · F. Odone · L. Rosasco
Dipartimento di Informatica, Bioingegneria, Robotica e Ingegneria dei Sistemi, Universitá degli Studi di Genova
Via Dodecaneso 35, 16100, Genova, IT
1 Introduction
Artificial intelligence has recently progressed dramatically, largely thanks to the advance in deep learning.
Computational vision, specifically object classification,
is perhaps the most obvious example where deep learning has achieved so stunning results to raise the question of whether this problem is actually solved (Krizhevsky
et al 2012b; Simonyan and Zisserman 2015; Szegedy
et al 2015; He et al 2016). Should this be the case,
robotics would be a main field where the benefits could
have far reaching effect. Indeed, the lack of reliable visual skills is largely considered a main bottleneck for the
successful deployment of robotics systems in everyday
life (Kemp et al 2007).
With this perspective in mind, we have recently
started an effort to isolate and quantify the benefits and
limitations, if any, of deep learning approaches to object recognition in robotics (Pasquale et al 2015, 2016a).
Clearly, visual perception is only one of the possible sensory modalities enabling object recognition in robotics
(Montesano et al 2008; Chitta et al 2011; Dahiya et al
2010; Natale et al 2004; Gorges et al 2010; Sinapov
et al 2014a; Hosoda and Iwase 2010; Moldovan et al
2012) and indeed, the comparison with human intelligence suggests there is more than “just” vision to object
recognition (Metta et al 2006; Fitzpatrick et al 2008;
Pinto et al 2016). Nonetheless, current deep learning
based artificial vision systems perform so well, that it
seems natural to ask how far they can go, before further
perceptual cues/modalities are needed. To this end, in
this work we mainly focus on object recognition tasks
in robotics using only visual cues.
To investigate the effectiveness of deep learning approaches in robotic applications, we designed a dataset
tailored to reflect a prototypical visual “experience” of
a humanoid robot. Indeed, the remarkable performance
of deep learning methods in object recognition has been
2
primarily reported on computer vision benchmarks such
as (Griffin et al 2007; Everingham et al 2010, 2015;
Russakovsky et al 2015), which have been essentially
designed for image retrieval tasks, and hardly are representative of a robotics scenario. In fact, this is a motivation common to other recent works such as (Lai et al
2011; Oberlin et al 2015; Borji et al 2016), where new
datasets have been proposed.
Using the iCub robot (Metta et al 2010), we devised
a human-robot interaction framework to acquire a corresponding dataset, named iCWT (iCubWorld Transformations). This dataset is rich and easy to expand to
include more data and complex perceptual scenarios. It
includes several object categories with many instances
per category, hence allowing to test both categorization
and identification capabilities. Notably, the dataset is
segmented in different sets of views, with the purpose
of testing specifically robustness and invariance properties of recognition systems within a realistic robotic
scenario.
Provided with the iCWT dataset, we performed extensive empirical investigation using different state of
the art Convolutional Neural Networks (CNN) architectures (Krizhevsky et al 2012a; Simonyan and Zisserman 2015; Szegedy et al 2015; He et al 2016). We began
checking to which extent systems already trained (on
other data) could be directly tested on iCWT. While
obtaining results much better than chance, these systems do not perform accurately enough in this case as perhaps one could expect. We hence followed a recent trend based on the general idea of transfer learning (Schwarz et al 2015; Oquab et al 2014; Sünderhauf
et al 2015; Pasquale et al 2016b; Chatfield et al 2014;
Simonyan and Zisserman 2015), where the pretrained
networks are adapted (fine-tuned) to the data at hand.
Indeed, our results confirm the effectiveness of these
strategies obtaining substantial improvements.
While impressive, these methods did not quite provide the close to perfect accuracy one would wish for.
We hence proceeded taking a closer look at the results, starting from the question of whether the missing
gap could be imputed to lack of data. Indeed, CNNs
are known to need massive amount of data to work
and data-augmentation is often used to improve results. As we discuss in detail later in the paper, investigating this latter question highlighted some differences between iCWT and other datasets such as ImageNet (Russakovsky et al 2015). More generally, we
identified clear differences between the object recognition task in robotics with respect to scenarios typically
considered in learning and vision.
Along the way, our analysis allowed to test invariance properties of the considered deep learning net-
Giulia Pasquale et al.
works and quantify their merits not only for categorization but also for identification. The description and
discussion of our empirical findings is concluded with a
critical review of some of the main venue of improvements, from a pure machine learning perspective but
also taking extensive advantage of the robotic platform.
Indeed, bridging the gap in performance appears to be
an exciting avenue for future multidisciplinary research.
In the next subsection we discuss several related
works, while the rest of the paper is organized as follows: Sec. 2 introduces the iCWT dataset and its acquisition setting. In Sec. 3 we review the deep learning
methods considered for our empirical analysis, which
is reported in Sec. 4 for the categorization task and
in Sec. 5 for object identification. Sec. 6 concludes our
study with the review of possible directions of improvement for visual recognition in robotics.
1.1 Deep Learning for Robotics
Deep Learning methods are receiving growing attention
in robotics, and are being adopted for a variety of problems such as object recognition (Schwarz et al 2015;
Pinto et al 2016; Held et al 2016; Pasquale et al 2016a),
place recognition and mapping (Sünderhauf et al 2016,
2015), object affordances (Nguyen et al 2016), grasping
(Redmon and Angelova 2014; Pinto and Gupta 2015;
Levine et al 2016) and tactile perception (Baishya and
Bäuml 2016). We limit our discussion to the work on
object recognition, which is more relevant to the work
described in this paper.
In (Schwarz et al 2015) the authors demonstrate
transfer learning from pre-trained deep Convolutional
Neural Networks (CNNs) and propose a way to include
depth information from an RGB-D camera. The main
idea is to extract a feature vector from the CNN and
train a cascade of Support Vector Machines (SVMs) to
discriminate objects’ class, identity and position. Depth
data is encoded in a planar image using a colorization
scheme. The authors report performance increase on
the Washington RGB-D benchmark (Lai et al 2011).
The work in (Pinto et al 2016) shows how the robot
can use self-generated explorative actions (like pushing
and poking objects) to autonomously extract training
data and train a CNN.
In this paper we employ a less constrained setting
than (Schwarz et al 2015), in that CNNs are trained
with data acquired by the robot during natural interaction (which undergo therefore more challenging viewpoint transformations). We employ similar techniques
for transfer learning, but we consider a wider range of
architectures and fine tuning techniques. In contrast
to (Pinto et al 2016) the use of a human teacher (see
Are we Done with Object Recognition? The iCub robot’s Perspective.
Sec. 2 for details on the acquisition setup) gives us more
control on the object trasformations. Clearly, our work
could be extended by introducing self-supervision using
explorative actions similar to the ones in (Pinto et al
2016).
The work in (Held et al 2016) and our work in
(Pasquale et al 2016a) are similar to the work presented in this paper in that they investigate invariance
properties of CNNs and learning from few examples.
They focus, however, on instance recognition, whereas
in this paper we consider – thus significantly extending (Pasquale et al 2016a) – the problem of object categorization. In addition we perform a detailed investigation of various fine-tuning techniques and systematic
evaluation of the recognition performance for specific
object transformations.
1.2 Datasets for Visual Recognition in Robotics
In the literature, several datasets for visual recognition
in robotics have been proposed: COIL (Nene et al 1996),
ALOI (Geusebroek et al 2005), Washington RGB-D (Lai
et al 2011), KIT (Kasper et al 2012), SHORT-100 (RiveraRubio et al 2014), BigBIRD (Singh et al 2014), Rutgers Amazon Picking Challenge RGB-D Dataset (Rennie et al 2016). One of the main characteristic of these
datasets is to capture images of an object while it undergoes specific viewpoint transformations. However,
these datasets are usually acquired in strictly controlled
“turntable” settings, aiming to provide accurate 3D annotations (e.g., information about the object’s pose in
each image) as a ground truth for grasping or manipulation, rather than substantial variations in the objects’ appearance. As a consequence, they mainly contain changes in the 3D point of view, under-representing
other visual transformations, such as scaling, background
changes and so forth, which are fundamental to benchmark visual recognition tasks. For instance, (Bakry et al
2015) use the Washington RGB-D and the Pascal 3D+
(Xiang et al 2014) datasets in order to evaluate the invariance of deep learning methods, and their analysis
consequently focuses mainly to 3D viewpoint changes.
Moreover, since the object is typically positioned on
a turntable that is subsequently rotated, these dataset
do not show more natural object transformations that
often occur in practice. See (Leitner et al 2015) for a
review of the major limitations of current datasets.
The NORB dataset (LeCun et al 2004) was also acquired in a similar turntable setting, but it is one of the
first to be released specifically in support of the investigation of invariance properties of recognition methods,
and in fact the images in the dataset have been artificially perturbed in order to increase variability. Re-
3
cently, (Borji et al 2016), motivated by similar considerations, presented the iLab-20M dataset: specifically,
the authors aim to create a large-scale visual recognition benchmark which, beyond representing a high
number of object instances, provides also a sufficient
number of images per object, in order to study invariance properties of deep learning methods. Also iLab20M is acquired in a turntable setting, in order to collect annotations for each image in terms of the object’s
pose.
The iCWT dataset presented in this paper separates from most previous work in that objects are captured while undergoing “natural” transformations. Acquisition is performed in a “semi-controlled” setting
intended to be a benchmark reproducing typical uncertainties faced by the visual recognition system of a
robot during a real-world task. In the following, we discuss the iCWT dataset and the related acquisition setting in detail.
2 The iCubWorld Dataset
In this section we present a novel dataset for visual
recognition, iCubWorld Transformations (iCWT), which
is used for the empirical analysis in this work. iCWT is
the latest release of the iCubWorld1 project, whose goal
is to benchmark and improve visual recognition systems
for robotics. iCubWorld datasets (Ciliberto et al 2013;
Fanello et al 2013b; Pasquale et al 2015) are designed
to record a prototypical visual “experience” of a robot,
the humanoid iCub (Metta et al 2010), while it is performing vision-based tasks. To this end, we devised and
implemented a simple human-robot interaction application, during which we directly acquire images for the
dataset from the robot’s cameras.
There is a remarkable advantage in collecting iCubWorld directly from the robot platform. Indeed, the resulting dataset offers a natural testbed for visual recognition in robotics, which is as close as possible to the
real application. In particular, this ensures that performance measured off-line on iCubWorld can be expected to generalize well when the system is deployed
on the actual robot. Note that this aspect of iCubWorld is extremely relevant since visual biases make it
typically difficult to generalize prediction performances
across different datasets and applications, as already
well-known from previous work (Pinto et al 2008; Torralba and Efros 2011; Khosla et al 2012; Hoffman et al
2013; Rodner et al 2013; Model and Shamir 2015; Stamos et al 2015; Tommasi et al 2015) and shown empirically in Sec. 4.1 of this paper.
1
https://robotology.github.io/iCubWorld/
4
Giulia Pasquale et al.
Fig. 1 iCWT categories. For each category in iCWT, we report one example image for one instance. (Red) categories appear
also in the ILSVRC 2012 dataset. (Black) categories appear in ImageNet but not in ILSVRC (see supplementary material for
more details).
Currently, to acquire iCubWorld releases we did not
make extensive use of the robot’s physical capabilities
(e.g., manipulation, exploration, etc.). This was done
because current deep learning methods achieve already
remarkable performance by relying solely on visual cues
and our goal was to evaluate their accuracy in isolation on in a robotic setting. Indeed, while exploiting the
robot body could provide further dramatic advantages
to modeling and recognition (see for instance (Montesano et al 2008) and, more recently, (Moldovan et al
2012), (Leitner et al 2014) and (Dansereau et al 2016)),
it would also prevent us to correctly assess the contribution of the purely visual components.
possibly contribute to) iCubWorld. Indeed, there are
few datasets for visual recognition in the literature that
were acquired directly on a robotic platform (Ramisa
et al 2011; Meger and Little 2013; Oberlin et al 2015).
One common reason is that scaling up the dataset size
is extremely expensive and time consuming when using
a single robot. To this end, an appealing solution, proposed in the Million Object Challenge project (Oberlin
et al 2015), is to involve multiple robots to collect a
shared dataset of visual experiences.
Note that, while we used an initial subset of iCWT
in (Pasquale et al 2016a), in this paper we present the
dataset for the first time in its entirety.
2.1 Acquisition Setup
2.2 Dataset Overview
Data acquisition for iCWT followed a protocol similar to the one in (Fanello et al 2013a), which was
adopted for previous iCubWorld releases (Ciliberto et al
2013; Fanello et al 2013b; Pasquale et al 2015): a human “teacher” shows an object to the robot and pronounces the associated label to annotate subsequent
images. The robot exploits bottom-up visual cues to
track and collect images of the object while the human
actor moves it and shows it from different poses. In
this work we decided to adopt an object tracker using
depth information (Pasquale et al 2016b), in place of
one exploiting the detection of independent motion in
the scene (Ciliberto et al 2011, 2012) (which has been
used to collect previous iCubWorld releases). In fact,
as shown in (Pasquale et al 2016b), this allows us to
extract a more precise bounding box around the object
of interest.
We developed an application to scale this acquisition procedure to hundreds of objects, which were
collected during multiple interactive sessions. iCWT is
available on-line and we plan to make also this application publicly available in order for other laboratories to use the same protocol to collect their own (or
iCWT is the largest iCubWorld release so far, comprising 200 objects evenly organized into 20 categories that
can be typically found in a domestic environment. Fig. 1
reports a sample image for each category in iCWT: 11
categories (in red in the figure) are also in the ImageNet
Large-Scale Visual Recognition Challenge (ILSVRC)
2012 (Russakovsky et al 2015), i.e. we found semantically and visually similar classes among the 1000 of
the classification challenge. The remaining 9 categories
do not appear ILSVRC but belong (or are similar) to
a synset in the larger ImageNet dataset (Deng et al
2009). To provide a qualitative intuition of the semantic
variability within a given category, namely the different visual appearance of objects in the same category,
Fig. 2 shows a sample image from the 10 instances in
the mug category. We refer the reader to the supplementary material (Fig. 13) for example images of all
object instances in iCWT.
The peculiarity of iCWT, motivating its name, is
that each object is shown in multiple image sequences
while it undergoes specific visual transformations (such
as rotations, scaling or background changes). This is
done to test the invariance of visual models in robotics.
Are we Done with Object Recognition? The iCub robot’s Perspective.
5
Table 1 Summary of the iCubWorld - Transformations
dataset
Fig. 2 Semantic variability. Sample images for the different object instances in the mug category to provide a qualitative intuition of the semantic variability in iCubWorld. See
Fig. 13 in the supplementary material for more examples.
# Categories
# Obj. per
Category
# Days
Transformations
20
10
2
2D ROT, 3D ROT
SCALE, BKG
MIX
# Frames per
Session
150
300
the image. No change in the object orientation (no
2D or 3D rotation) was applied (Fig. 3, third row).
Background The human moved the object around the
robot, keeping approximately the same distance (scale)
and pose of the object with respect to the camera plane. During this acquisition process only the
background changes while the object appearance remains approximately the same (Fig. 3, fourth row).
Mix The human moved the object freely in front of the
robot, as a person would naturally do when showing
a new item to a child. In this sequence all nuisances
in all combinations can appear (Fig. 3, fifth row).
Fig. 3 Visual transformations. Excerpts from the sequences acquired for one mug, representing the object while
it undergoes specific visual transformations.
Indeed, since very few works study “real” (i.e. nonsynthetic) visual transformations (Goodfellow et al 2009)
in order to evaluate the invariance properties of visual
representations, we believe that iCWT could be a significant contribution in this direction. To our knowledge, iCWT is the first dataset to address invariancerelated questions in robotics and accounts a much wider
range of visual transformations with respect to previous
datasets.
For each object instance, we acquired 5 different image sequences, each while the human supervisor performed a different visual transformation to the object.
Fig. 3 reports excerpts of these sequences, that contain,
respectively:
2D Rotation The human rotated the object parallel
to the camera image plane. The same scale and position of the object were mantained (see Fig. 3, first
row).
3D Rotation Similarly to 2D rotations, the object was
kept at same position and scale. However, this time
the human applied a generic rotation to the object
(not parallel to the image plane). As a consequence
different “faces” of the object where shown to the
camera (Fig. 3, second row).
Scale The human moved the object towards the cameras and back, thus changing the object’s scale in
Each sequence is composed by approximately 150
images acquired at 8 frames per second in the time interval of 20s, except for the Mix one that lasted 40s
and comprises ∼ 300 images. As anticipated, the acquisition of the 200 objects was split in multiple sessions
performed in different days. The acquisition location
was always the same (with little uncontrolled changes
in the background across days). The lighting condition
was not artificially controlled, since we wanted to investigate its role as a further nuisance: to this end, we
acquired objects at different times of the day and in
different days, so that lighting conditions are slightly
changing across the 200 objects (but not within the
five sequences of an object, which were all acquired in
the span of few minutes). Moreover, we repeated the
acquisition of each object in two different days, so that
we ended up with 10 sequences per object, containing
5 visual transformations in 2 different light conditions.
The adopted iCub’s cameras resolution is 640 × 480.
We recorded the centroid and bounding box provided
by the tracker at each frame (for details about this
procedure see (Pasquale et al 2016b)). Both left and
right images were acquired, to allow for offline computation of the disparity map and potentially further
improvement of the object’s localization and segmentation. Tab. 1 summarizes the main characteristics of
iCubWorld-Transformations.
3 Methods
In this section we review the learning methods considered in this work. In particular we briefly introduce the
principal concepts behind deep Convolutional Neural
6
Giulia Pasquale et al.
Networks (CNNs), describe the architectures used in
our analysis and the algorithms adopted to train and
apply them. We refer the interested reader to (Goodfellow et al 2016) for an in-depth introduction to CNNs
and deep learning in general.
3.1 Deep Convolutional Neural Networks
Deep Convolutional Neural Networks (CNNs) are hierarchical models organized as the concatenation of multiple processing layers. This structure allows mapping
the input signal (such as images, speech, etc.) through
a series of subsequent representations to progressively
select the features that are most relevant to the considered task. The prototypical structure of a CNN (see
Fig. 4) performs at each layer (usually referred to as
convolution layer) the following set of operations:
• Convolutions: local convolution with respect to a
bank of (learned) linear filters.
• Spatial Downsampling for instance using strided
convolutions.
• Element-wise Non Linearity such as sigmoid functions (Bishop 2006) or, more recently Rectifying Linear Units (He et al 2015).
• Spatial Pooling to aggregate local responses to the
signal in a single component, for instance by taking
the maximum value observed (max-pooling).
Each CNN architecture is characterized by the specific choice and implementation of the operations above.
The filters, locally processing the input signal at each
layer, are typically learned by minimizing a desired loss
function, such as the overall classification accuracy in
supervised settings or the reconstruction error in unsupervised ones. The spatial downsampling and pooling operations make the representation more robust
to transformations of the input at increasingly larger
scales. The non-linearities selectively retain only features that are relevant to the task, while suppressing
the others.
Until recently, the most common strategy in image
classification settings was to follow convolution layers
(C) with a set of fully connected layers (FC), namely
a standard multi-layer Neural Network. In the last few
years however, architectures with fewer or even just one
FC layer have started to be preferred since they achieve
comparable or better performance while optimizing significantly fewer parameters (Szegedy et al 2015; He et al
2016). In classification settings, the final layer of a network is typically a softmax function that maps the CNN
output into individual class likelihood scores. The CNN
3
https://www.clarifai.com/technology
prediction is then obtained by choosing the class with
maximum score.
The modular structure of CNNs allows training their
parameters simultaneously for all layers (also known as
end-to-end learning) via back-propagation (LeCun et al
1989). Given the large number of parameters to be optimized (in the order of millions), CNNs typically need
large amounts of training data to achieve good performance. Various forms of data augmentation are often
needed to artificially increase the number of training examples (e.g. by synthetically modifying the images via
geometrical or illumination changes). To further mitigate the risk of overfitting, regularization techniques
such as weights L2 regularization, dropout (Hinton et al
2012; Srivastava et al 2014) or, more recently, batch
normalization (Ioffe and Szegedy 2015), have proved
helpful.
In this work we investigate the performance of modern CNNs on the robotic setting of iCubWorld. For this
analysis we selected recent architectures achieving the
highest accuracy on the ImageNet Large-Scale Visual
Recognition Challenge (ILSVRC) (Russakovsky et al
2015) between 2012 and 2015. We used their corresponding implementation trained on ILSVRC 2012 and
publicly available within the Caffe (Jia et al 2014)
framework. Below we summarize their structures:
CaffeNet4 A small variation of the AlexNet model,
winner of ILSVRC 2012 (Krizhevsky et al 2012a).
Concatenates 5C + 3FC layers comprising ∼ 60M
parameters.
VGG-165 Second classified at ILSVRC 2014 (Simonyan
and Zisserman 2015). It maintains the general structure of AlexNet (but with 13C + 3FC layers comprising ∼ 140M parameters). It is relatively expensive to train in terms of memory and time because
of the large FC layers.
GoogLeNet6 Winner of ILSVRC 2014 (Szegedy et al
2015). It diverges from previous architectures in that
it concatenates so called inception modules and uses
just one FC layer at the very end, reducing the parameters number to ∼ 4M for 22 layers.
ResNet-507 The name is short for residual networks,
which won the ILSVRC 2015 (He et al 2016), of
which ResNet-50 is a smaller version stacking 50
layers in ∼ 20M parameters.
4
https://github.com/BVLC/caffe/tree/master/models/
bvlc_reference_caffenet
5
http://www.robots.ox.ac.uk/$\sim$vgg/research/
very_deep/
6
https://github.com/BVLC/caffe/tree/master/models/
bvlc_googlenet
7
https://github.com/KaimingHe/
deep-residual-networks
Are we Done with Object Recognition? The iCub robot’s Perspective.
7
Fig. 4 Example of a Convolutional Neural Network (Sec. 3.1) and of the two knowledge transfer approaches considered in this
work (Sec. 3.2). (Blue pipeline) feature extraction: in this case the response of one of the layers is used as a feature vector
for a “shallow” predictor like RLSCs, SVMs (see Sec. 3.2.1), which is trained on the new task. (Red pipeline) fine-tuning:
in this case the network is trained end-to-end to the new task by replacing the final layer and using the original model as a
“warm-restart” (see Sec. 3.2.2). Network image from3 .
3.2 Transfer Learning Techniques
Deep learning methods typically require very large datasets
to be successfully trained. While this could in principle
prevent their applicability to problems where training
data is scarce, recent empirical evidence has shown that
knowledge acquired by training a network on a largescale problem can be “transferred” to the new domain.
Different approaches to implement this strategy have
been proposed in the literature. In this section we review two of the most well-established method, which
we empirically assess in our experiments, namely feature extraction and fine-tuning.
3.2.1 Feature Extraction
It has been observed that deeper layers of a CNN selectively respond to specific properties of the visual scene
that are typically characteristic of specific objects or
object parts. (Chatfield et al 2014; Yosinski et al 2014;
Donahue et al 2014; Zeiler and Fergus 2014). These responses can be interpreted as representations of the visual input responding only to features that are most relevant to the task at hand. Following this intuition, it has
been shown that CNNs trained on large-scale datasets
can be indeed used as “feature extractors” on smaller
datasets, leading to remarkable performance (Schwarz
et al 2015; Oquab et al 2014; Sünderhauf et al 2015;
Pasquale et al 2016b). This is typically done by training a standard classifier such as a Support Vector Machine (SVM) or a Regularized Least Squares Classifier
(RLSC) (Bishop 2006) on top of feature vectors obtained from one or multiple layer responses. This strategy, depicted in Fig. 4 (Blue pipeline), can be interpreted as changing the last layer of the CNN and train-
ing it on the new dataset while keeping all other parameters of the network fixed.
Implementation Details. In the empirical analysis
of this work we used feature extraction as a strategy
to transfer knowledge from CNNs trained on ImageNet
to iCWT. The specific layers used for our experiments
are reported in Tab. 2 for the four architectures considered in this paper. We used RLSC with a Gaussian
Kernel as classifier for the CNN-extracted features. In
particular we implemented the Nystrom sub-sampling
approach considered in (Rudi et al 2015, 2016), which
is computationally appealing for mid and large-scale
settings and indeed allowed to significantly speed up
our experimental evaluation. We refer the reader to the
supplementary material for details about the model selection and image preprocessing protocols adopted in
this work.
3.2.2 Fine-tuning
A CNN trained on a large-scale dataset can be used
to “warm-start” the learning process on other (potentially smaller) domains. This strategy, known as finetuning (Chatfield et al 2014; Simonyan and Zisserman
2015), consists in performing back-propagation on the
new training set by initializing the parameters of the
network to those previously learned (see Fig. 4 (Red
pipeline)). In these settings it is necessary to adapt the
final layer to the new task (e.g. by changing the number of units in order to account to the new number of
classes to discriminate).
A potential advantage of fine-tuning is that it adapts
the parameters of all networks’ layers to the new problem rather than only those in the final layers. How-
8
Giulia Pasquale et al.
Table 2 Feature extraction layers for the four architectures
considered in this work. We used the notation adopted in the
literature, in which the number identifies the layer number
and the label specifies its type (i.e. fully connected or pooling
layer).
Model
CaffeNet
GoogLeNet
VGG-16
ResNet-50
refer to the supplementary material of this work for
an analysis of using other fine-tuning parameters and a
discussion on the model selection protocols used in our
experiments.
Output Layer
fc6 or fc7
pool5/7x7 s1
fc6 or fc7
pool5
Table 3 Fine-tuning protocols for CaffeNet and GoogLeNet.
Base LR is the starting learning rate of all layers that are
initialized with the original model. The FC layers that are
learned from scratch are indicated using their names in Caffe
models (2nd row), specifying the starting learning rate used
for each of them. For the other parameters, we refer the reader
to Caffe documentation.
4 Results (Categorization)
In this section we present our empirical investigation of
Deep Learning methods in the robotic setting of iCubWorld.
4.1 Deep Learning and (the Need for) Knowledge
Transfer
Modern datasets for visual recognition comprise an extremely large number of images (e.g. 1 Million for the
Base LR
1e-3
0
1e-5
0
ILSVRC challenge) depicting objects in a wide range of
fc8: 1e-4
loss3/classifier: 1e-2
natural scenes. This extreme variability opens the quesLearned
fc8: 1e-2
fc7: 1e-4
loss1(2)/classifier: 1e-3
FC Layers
tion of whether datasets such as iCubWorld, which repfc6: 1e-4
loss1(2)/fc: 1e-3
resent a smaller “reality”, could be interpreted simply
fc7: 50
pool5/drop 7x7 s1: 60
Dropout (%)
fc6: 50
loss1(2)/drop fc: 80
as sub-domains of larger ones. If this was the case, deep
Solver
SGD
Adam
learning models trained on ImageNet would achieve
LR Decay Policy
Polynomial (exp 0.5)
No decay
high recognition performance on iCubWorld as well,
# Epochs
6
36
6
without any re-training or adaptation.
Batch Size
256
32
While this result would lead to the appealing scenario where a robot can simply download and use a viever, this flexibility comes at the price of a more insual classifier “off-the-shelf” (i.e. trained off-line), prevolved training process. Indeed, the performance of a
vious analysis has shown that most computer vision
fine-tuned network are extremely dependent on the choice datasets are essentially biased, ultimately preventing a
of the (many) hyper-parameters available. Fine-tuning
learning algorithm to generalize from one domain to the
has been recently used to adapt network models learned
other (Pinto et al 2008; Torralba and Efros 2011; Model
on ImageNet to robotic tasks (see, e.g., (Eitel et al 2015;
and Shamir 2015; Khosla et al 2012; Stamos et al 2015;
Pinto and Gupta 2016; Redmon and Angelova 2015a;
Tommasi et al 2015; Hoffman et al 2013; Rodner et al
Pasquale et al 2016a; Nguyen et al 2016)).
2013). Similarly, it was recently observed that conventional “non-deep” models trained on ImageNet perform
poorly when applied to robotic settings (Goehring et al
Implementation Details. In our experiments we per2014).
formed fine-tuning only for CaffeNet and GoogLeNet,
To address this question, we evaluated four off-thewhich are representative of most recent architectures
shelf CNNs (the ones reviewed in Sec. 3.1) for the task
and were significantly faster to fine-tune than VGG-16
of image classification on iCWT. For these experiments
and ResNet-50. We considered two main regimes for
we restricted the test set to the 11 categories of iCWT
fine-tuning a network: one updating only one or more
that appear also in the ILSVRC challenge (see Sec. 2
FC layers while keeping the others fixed, and another
and Fig. 1). As a reference, we compared these results
one more aggressively adapting all layers, comprising
with the average accuracy of the same networks on the
the C layers, to the training dataset. This was done by
corresponding 11 categories of the ImageNet dataset.
selecting different learning rates for the neurons in each
The test set for iCWT was composed, for each category,
layer, i.e., the size of gradient steps at each iteration
by the images of all 10 object instances, comprising all
of back-propagation. We refer to these two protocols as
5 transformations for one day and the left camera (unconservative and adaptive and report the corresponding
less differently specified we always used this camera for
parameters in Tab. 3 as a reference. In the experiments
the experiments), for a total of ∼ 9000 images per catediscussed in this paper we consider only these strategory. For ImageNet, we downloaded the images for the
gies since we did not observe significant performance
corresponding 11 categories (refer to the supplementary
differences following other choices of parameters. We
CaffeNet
adaptive conservative
GoogLeNet
adaptive conservative
Are we Done with Object Recognition? The iCub robot’s Perspective.
100
90
80
accuracy (%)
70
60
50
40
iCWT off the shelf
iCWT adapted
ImageNet
Random Choice
30
20
10
0
CaffeNet
GoogLeNet
VGG-16
ResNet-50
Fig. 5 Average classification accuracy of off-the-shelf networks (trained on ILSVRC) tested on iCubWorld (Dark Blue)
or on ImageNet itself (Gray). The test sets for the two
datasets are restricted to the 11 shared categories (see Sec. 2).
(Light Blue) reports the classification accuracy when the
same networks are “transferred” to iCubWorld (see Sec. 4.1
for details). The (Orange line) shows the recognition chance
of a random classifier.
material for the corresponding synsets), comprising on
average ∼ 1300 images per category.
Fig. 5 reports the average classification accuracy on
iCWT (Dark Blue) and ImageNet (Gray). It can be
immediately observed that there is a substantial drop
in performance of ∼ 50 − 60% when testing on iCWT
rather than ImageNet, suggesting that differences between the two datasets exist, and in particular that
iCubWorld is not a sub-domain of ImageNet. This drop
is likely due to biases in ImageNet (see also (Tommasi
et al 2015; Hoffman et al 2013; Rodner et al 2013)).
While an in-depth analysis of bias is outside the scope
of this work, in the supplementary material we provide
some further evidence of this effect. We care to point
out that for a fair comparison we have restricted the
output of the off-the-shelf networks to the 11 classes
considered in the experiment with iCWT (from the
original 1000-dimensional output vector provided by
the off-the-shelf networks). We refer again to the supplementary material for the experiments reporting the
same performance when considering the entire 1000dimensional prediction.
Knowledge Transfer. The drop in performance observed in Fig. 5 implies that it is necessary to train the
recognition system on the target domain. However, the
experiment also shows that all the networks performed
substantially better than chance (Orange line). This
suggests that the networks retained some knowledge
about the objects’ appearance. Therefore, rather than
training a novel architecture “from scratch” (which typ-
9
ically requires large amounts of training data, time and
computational resources), a viable alternative is to transfer such knowledge, essentially by “adapting” the networks trained on ImageNet to the new setting (see Sec. 3.2).
The idea of knowledge transfer has recently received
considerable attention from the Deep Learning community (Hoffman et al 2013; Agrawal et al 2014; Azizpour
et al 2015; Huh et al 2016) because it provides a robust
approach to apply Deep Learning methods to smaller
datasets.
In Fig. 5 we report the classification performance
achieved by systems where knowledge transfer has been
applied (Light Blue). For these experiments we followed
the protocol described in Sec. 3.2.1, where RLSC predictors are trained on feature vectors extracted from
the deeper layers of the CNNs. We created a separate
training set from iCWT, by choosing 9 instances for
each category for training while keeping the 10th instance for testing. We repeated these experiments for
10 trials in order to allow each instance of a category to
be used in the test set. Fig. 5 reports the average accuracy over these trials. We observe a sharp improvement
for all networks, which achieve a remarkable classification accuracy in the range of ∼ 70 − 90%. On ImageNet
however, the corresponding performance is still significantly higher, although such gap seems to be reduced
for more recent architectures.
What are the reasons for this gap? In the following we empirically address this question, showing that
the observed behavior is due to fundamental differences
between robot vision and image retrieval settings.
4.2 Do we need more data?
Deep learning methods require large amounts of data
in order to be trained effectively. This is particularly
true when training a network from “scratch”, but valid
also (albeit on a smaller scale) to knowledge transfer
techniques. Indeed, as mentioned in Sec. 3, a common
practice when training a CNN is to perform data augmentation, namely to artificially increase the dataset
size by applying synthetic transformations to the original images (e.g. rotations, reflections, crops, illumination changes, etc.).
From this perspective the robotic setting seems particularly favorable. Indeed, data augmentation can be
performed by simply acquiring more and more frames
depicting an object while viewpoint or illumination change
naturally. While acquiring more frames of an object is
particularly convenient, in robotics it is typically expensive to gather instead images depicting different objects instances, since the robot needs to directly observe
each of them. In this section we investigate the impact
10
Giulia Pasquale et al.
RLSC
FINE-TUNING
RLSC
100
100
100
90
90
90
90
80
80
80
80
70
70
70
70
60
60
60
60
50
50
40
30
20
10
40
CaffeNet adapt
CaffeNet cons
GoogLeNet adapt
GoogLeNet cons
50
30
150
#frames/seq
300
20
10
accuracy (%)
accuracy (%)
FINE TUNING
100
CaffeNet
GoogLeNet
VGG-16
ResNet-50
50
50
50
40
30
150
300
#frames/seq
Fig. 6 Recognition accuracy vs # frames. (Left) Accuracy of CaffeNet and GoogLeNet models fine-tuned according
to the conservative and adaptive strategies (see Sec. 3.2.2).
(Right) Accuracy of RLSC classifiers trained over features
representations extracted from the 4 architectures considered
in this work (see Sec. 3.2.1).
of these aspects in robot vision, taking iCubWorld as
a testbed for our analysis. Note that in the following
we use the term instance to refer to a specific object
belonging to a given category, while frame denotes a
single image depicting a scene or an object.
4.2.1 What do we gain by adding more frames?
We considered a 15-class categorization task on iCWT
to compare the performance of learning models trained
on an increasing number of examples. We created training sets of different size by sampling respectively N =
10, 50, 150 and 300 frames from each transformation sequence of an object. For each category (see the supplementary material for the list of categories) we used 7
objects for training, 2 for validation and 1 for test. Validation and test sets contained all images available for
the corresponding instances. In order to account for statistical variability, we repeated these experiments for 10
trials, each time leaving out one different instance for
testing. For this experiment we trained and tested on
one of the two available days in iCWT (and one -the
left- camera).
Fig. 6 reports the average classification accuracy
of different learning models as more training examples
are provided. Surprisingly, most architecture achieve remarkably high accuracy already when trained on the
smallest training set and show little or no improvement
when new data is available. This finding is in contrast
with our expectations, since increasing the dataset size
does not seem key to a significant improvement in performance.
Secondary observations:
• Fine-tuning and RLSC achieve comparable accuracy (both for CaffeNet and GoogLeNet).
20
40
CaffeNet adapt
CaffeNet cons
GoogleNet adapt
GoogleNet cons
1
3
5
#obj/cat
CaffeNet
GoogleNet
VGG-16
ResNet-50
30
7
20
1
3
5
7
#obj/cat
Fig. 7 Recognition accuracy vs # instances (number
of object instances available during training). (Left) Accuracy of CaffeNet and GoogLeNet models fine-tuned according
to the conservative and adaptive strategies (see Sec. 3.2.2).
(Right) Accuracy of RLSC classifiers trained over features extracted from the 4 architectures considered in this work (see
Sec. 3.2.1).
• We confirm the ILSVRC trends, with more recent
networks outperforming older versions but with VGG16 features being better than those of GoogLeNet
when using feature extraction with RLSC.
• Note that CaffeNet performs worse when training
data is scarce because of the high number of parameters to be learned in the 3 FC layers (see Sec. 3
and the supplementary material).
• To further support these findings, in the supplementary material we report results for the same experiment performed using less example instances per
category.
4.2.2 What do we gain by adding more
instances?
We evaluated the impact of adding new object instances
when training a recognition system. We consider this
process as increasing the semantic variability of the
training set, in contrast to increasing the geometric
variability by showing more frames of a known object
instance. We considered the same 15-class categorization task of Sec. 4.2.1 and created four training sets
containing frames sampled from an increasing number
of instances per category, namely 1, 3, 5 and 7. For all
dataset, training data was randomly sub-sampled from
all available transformation sequences to achieve identical total size of 900 images per category. Validation and
test sets contained all frames for the remaining 2 and 1
instance (per category). We repeated the experiments
for 10 trials.
Fig. 7 reports the accuracy of different models tested
on this problem. Results show that increasing the semantic variability dramatically improves the accuracy
of all models by a similar margin (around 15 − 20%).
Are we Done with Object Recognition? The iCub robot’s Perspective.
# instances
per category
∞
IMAGE RETRIEVAL
ImageNet
ROBOT VISION
iCubWorld
1
# frames
per
instance
∞
Fig. 8 Different training regimes from robot vision and image retrieval settings.
Moreover, from the observed trends we would expect
these model to achieve even higher accuracy if more
object instances were available.
Secondary observations:
• Fine-tuning the networks seems to lead to worse performance than RLSC on feature extraction both for
CaffeNet and GoogLeNet.
• Also in this setting we confirm ILSVRC results, with
more recent networks outperforming previous ones
but with VGG-16 outperforming GoogLeNet when
using feature extraction with RLSC.
• To further support our findings, in the supplementary material we report results for a similar experiment but discriminating only between 10 or 5 categories.
4.2.3 Robot Vision and Image Retrieval
In this section we considered some of the main challenges associated to robot vision. While both robot vision and image retrieval address the same visual recognition problem, they are cast within two very different
training regimes. We have identified here the two main
causes of such difference (depicted in Fig. 8): 1) semantic variability, namely the amount of different object
instances available for each category and 2) the amount
of frames per instance.
Typically, in image retrieval settings a large number of images depicting different object instances for
each category is available, but for each such instance
very few images are provided. As an example, ILSVRC
training set comprises ∼ 1000 instances for each of the
1000 object categories, while only 1 image per instance
is provided. On the opposite end of the spectrum, in
robot vision settings it is easy to gather large amounts
11
of example images for a single instance (since the robot
can move around and observe an object from multiple
points of view). However there is a remarkable limitation in the total number of categories and instances that
can be experienced.
Our analysis has shown that the limited availability
of semantic variation that naturally occurs in robotics
applications can dramatically affect the recognition accuracy of a learning system. Moreover, contrarily to
our expectations, we observed that this limitation cannot be alleviated by simply feeding more training data
(frames) to the network. Indeed, classifiers trained on
datasets of different size (but same semantic variability) achieved identical performance. This can be extremely problematic since in robotics the overhead of
collecting and acquiring examples of a novel instance is
extremely expensive, while collecting more views of an
object comes at a low cost. Unfortunately, we cannot
adopt “data augmentation” strategies to artificially increase the semantic variability of a dataset as it is done
for the case of view-point variability.
From our findings it appears clear that bridging the
gap between robot vision and image retrieval needs a
deeper analysis. To this end, in the following we deepen
our analysis on the performance of CNNs on iCWT,
with particular focus on the concept of invariance.
4.3 Invariance
In Sec. 4.2.1 we saw that models trained on few frames
achieve comparable or even better accuracy to those
trained larger datasets. These results suggest that the
corresponding networks could leverage on a robust data
representation and indeed in this section we investigate
to what extent such models are invariant to the visual
transformations in iCWT. Indeed, invariance, i.e. robustness to identity-preserving visual transformations
of a scene, is a highly desirable behavior when designing and training a visual recognition system, since it can
dramatically reduce the “complexity” of the learning
problem (Anselmi et al 2016, 2015), increasing the capability of generalizing the visual appearance of an object
category from a limited number of examples (ideally,
just one). This is even more appealing in robotics applications where, usually, the robot is required to learn
new objects on the fly, and recognize them reliably in
unpredictable conditions.
To test the invariance of the CNNs, we considered
again the 15-class categorization problem introduced
in Sec. 4.2.2, where we used 7 object instances per category for training, 2 for validation and 1 for testing.
However, we did not mix examples from all the five
available sequences (2D Rotation, 3D Rotation, Scale,
12
Giulia Pasquale et al.
test: MIX
100
test: 3D ROT
test: SCALE
test: BKG
100
100
100
90
90
90
90
80
80
80
80
70
70
70
70
70
60
60
60
60
60
50
50
50
50
50
40
40
40
40
40
30
30
30
30
30
20
MIX 2DR 3DR SC BKG ALL
20
MIX 2DR 3DR SC BKG ALL
20
MIX 2DR 3DR SC BKG ALL
20
MIX 2DR 3DR SC BKG ALL
20
MIX 2DR 3DR SC BKG ALL
90
80
accuracy (%)
test: 2D ROT
100
CaffeNet
GoogLeNet
VGG-16
ResNet-50
Fig. 9 Generalization performance across different visual transformations. Learning models were trained on one of
the 5 transformations present in iCWT (specified in the horizontal axis), and then tested on all transformations. Accuracy is
reported separately in each plot for each tested transformation. Shaded areas represent the improvement achieved by including
all frames of a sequence instead of subsampling ∼ 20 of them.
Background and Mix). Instead we performed training
(and validation) using only an individual transformation, and then tested the learned model on the others.
We repeated these experiments for 10 trials. We considered two “regimes”: one with including all images
from a sequence, and another one subsampling images
of a factor of 7 (as we did in Sec. 4.2.2, Fig. 7), since
we previously observed that, while including more instances in the training set is important, including more
frames is not.
As in Sec. 4.2.1 and 4.2.2, we limit our analysis to
only one of the two available days in iCWT. Indeed,
in this setting we aim to study the effect of isolated
viewpoint transformations, without considering other
nuisances due to, for example, the changes in the light
and setting conditions that can appear from one day to
another.
For this experiment we report only the accuracy of
approaches that rely on the application of RLSC on
top of feature representations extracted using off-theshelf models (see Sec. 3.2.1). This is done to isolate the
invariance properties of the networks trained on ImageNet. Note however that we performed these experiments also by fine-tuning the networks, as for previous
tests, observing similar results.
Fig. 9 reports the generalization capabilities of models trained on a single transformation and tested on a
different one. As a reference, we considered also a 6th training set (All ), obtained by randomly sampling
points from all other 5 training sets, keeping similar
size. This training set is analogous to those used in
the experiments in Sec. 4.2.2 and 4.2.1 but with size
reduced to be ∼ 150 ∗ 7 = 1050 images per category
(and to only 150 images per category in the subsampled regime). We report the accuracy of the subsam-
pled training sets with a bold line, and the improvement
achieved by considering all frames from the sequences
with a shaded area of the same color.
We first confirm that adding example frames for a
transformation does not significantly improves the networks’ invariance (even to the transformation itself).
We then notice that overall the models exhibit invariance to transformations that have been included in the
training set. It is worth noticing that in all cases the
best performance are obtained when training and testing are performed on the same transformation and when
all transformations are included in the training set (All).
This demonstrates that enriching the dataset to include
more transformations does not degrade performance.
Moreover, training on Mix achieves as good performance as All when tested on every specific transformation. This suggests that showing an object to the robot
in a natural way, with all transformations appearing in
random combinations, instead of systematically collecting sequences comprising individual transformations, is
a feasible approach to obtain predictors invariant to
these transformations.
It is interesting to note that rather different models trained on ImageNet exhibit the same invariance
pattern. In particular, performance drops substantially
when testing on a transformation that was not present
on the training set. While this can be expected for
transformations involving 3D rotations of the object,
it is quite surprising for affine ones, namely Scale, 2D
Rotation and Background to which the convolutional
structure of the CNN should be invariant “by design”,
or learned during the training on ImageNet. We will
further investigate this fact in Sec. 5.2, where we study
the invariance of CNN models in the context of object
identification rather than categorization.
Are we Done with Object Recognition? The iCub robot’s Perspective.
5.1 Knowledge Transfer: from Categorization to
Identification
In Sec. 4.1 we have seen how knowledge acquired on
ImageNet can be transferred to the iCubWorld domain
to successfully tackle categorization tasks. A natural
question is whether the same strategy would be similarly favorable for object identification. Indeed, in object identification settings there is no semantic variability (there is only one instance per class), but for categorization we have observed that semantic variability
is extremely useful to the learning process (Sec. 4.2.2),
while simply adding more images of a given instance is
almost redundant (Sec. 4.2.1).
We addressed this question by considering an object identification task on iCWT, where we compared
CNN models trained with an increasing number of examples. The setting is similar to the one used for categorization but for a 50-class object identification problem: we chose the 50 object instances from the book,
flower, glass, hairbrush and hairclip categories, which
do not appear in the ILSVRC dataset. We created four
training sets containing respectively 10, 50, 150 and 300
RLSC
FINE TUNING
accuracy (%)
5 Results (Object Identification)
So far in this work we focused on visual categorization
problems, namely to assign a given object instance to
the corresponding category. In this section we move instead to the task of object identification (or instance
recognition), which consists in labeling a given image
according to the object instance appearing in it. This
problem is indeed very relevant to robotics settings,
since reliable instance recognition skills would provide
the system with finer control over tasks such as manipulation, grasping, etc. Also, considering a domestic
scenario, the robot could be asked to use a specific object rather than some object in a category (e.g. “iCub,
bring me my phone” instead of “a phone”).
Therefore, in parallel to the experiments on categorization, in this work we assessed the performance of
Deep Learning methods on the task of object identification in robotics. Note that until recently, this problem
has been approached with methods based on keypoints
extraction and template matching (Lowe 2004; Philbin
et al 2008; Collet et al 2011b; Crowley and Zisserman
2014; Collet et al 2011a, 2009; Muja et al 2011), however it has been recently observed that appraoches relying on more holistic visual representations perform
typically better in low/mid-resolution settings such as
iCubWorld (Ciliberto et al 2013). Following this, in this
section we will focus only on the CNN architectures
considered in Sec. 3.
13
100
100
90
90
80
80
70
70
60
60
50
40
30
20
10
50
40
CaffeNet adapt
CaffeNet cons
GoogLeNet adapt
GoogLeNet cons
50
30
150
#frames/seq
300
20
10
CaffeNet
GoogLeNet
VGG-16
ResNet-50
50
150
300
#frames/seq
Fig. 10 Recognition accuracy vs # frames (Identification). (Left) Accuracy of CaffeNet and GoogleNet models
fine-tuned according to the conservative and adaptive strategies (see Sec. 3.2.2). (Right) Accuracy of RLSC classifiers
trained over features representations extracted from the 4 architectures considered in this work (see Sec. 3.2.1).
images per object, sampled randomly from the 4 transformation sequences 2D Rot, 3D Rot, Scale and Bkg
(see Sec. 2). From each dataset, 20% images were retained for model selection. The images from the Mix
sequence were used to test the classification accuracy of
the methods considered. As for the categorization experiment, only the images from a single day were used
for these experiments.
Fig. 10 reports the average accuracy of CNN architectures trained on a growing number of examples.
These results are in stark contrast with their counterpart on categorization (Sec. 4.2.1): we can clearly notice
that adding more training data dramatically improves
the recognition capabilities of all networks. This suggests that having access to more views of an instance
allows the learning systems to create a more nuanced
model of the object. Apparently, this richer model does
not provide any advantage in categorization settings,
but is extremely useful to recognize the same object
in new images. In line with this observations, we notice also that in this setting the adaptive fine-tuning
strategy significantly outperforms the competitors, suggesting that the representation learned for categorization on ImageNet, albeit already beneficial to the task,
can be further improved to adapt to the object identification setting. These findings confirm and extend
recent similar results from the instance retrieval literature (Babenko et al 2014; Gordo et al 2016).
5.2 Invariance
In Sec. 4.1, we have seen that the accuracy of an object
identification system is greatly improved when it is provided with more examples of an instance. This observation seems to contradict the assumption that CNN are
14
Giulia Pasquale et al.
100
90
100
ResNet-50
VGG-15
GoogLeNet
CaffeNet
test: 2D ROT
100
test: 3D ROT
100
test: SCALE
100
90
90
90
90
80
80
80
80
70
70
70
70
70
60
60
60
60
60
50
50
50
50
50
40
40
40
40
40
30
30
30
30
30
80
accuracy (%)
test: MIX
20
2DR
3DR
SC
BKG
20
MIX
3DR
SC
BKG
20
MIX
2DR
SC
BKG
20
MIX
2DR
3DR
BKG
20
MIX
test: BKG
2DR
3DR
SC
Fig. 11 Generalization performance across different visual transformations (Identification). RLSC are trained on
one of the 5 transformations present in iCWT (specified in the horizontal axis), and then tested on the other transformations.
Accuracy is reported separately in each plot for each tested transformation.
designed to be already invariant to view-point transformations. We addressed this question by considering a
learning scenario similar to the one adopted in Sec. 4.3
to investigate invariance in categorization settings, but
focused here on object identification. Specifically, we
considered the same 50-class identification task, but restricted the training set to contain only images from
a single transformation sequence of the 5 available in
iCubWorld (Sec. 2). The resulting models were tested
separately on the remaining transformations to assess
the ability of the learning systems to generalize to new
viewpoints. Similarly to Sec. 4.3 we considered only one
day and evaluated only learning methods based on feature extraction (plus RLSC), in order to evaluate viewpoint invariance of off-the-shelf networks.
Fig. 11 reports the average accuracy of models trained
a single transformation using respectively 10 or 150
frames per object and tested on another. We notice
that the performance varies remarkably from one tested
transformation to another, suggesting that the representations considered in these experiments are not invariant to transformations observed in iCWT. Interestingly, by comparing the two training regimes (i.e. large
Vs small training set) we notice that the remarkable improvement observed in Sec. 5 when adding more frames
is confirmed and extended also to the setting where
the system is trained on a single transformation at the
time. As expected, more recent networks significantly
outperform previous models, with ResNet-50 achieving
the highest accuracy by a large margin.
Similarly to what we observed in the categorization
setting, Fig. 11 reports quite low performance when
generalizing across transformations. In the following we
discuss possible strategies to address this problem.
5.2.1 Improving the Invariance
In this section we consider different fine-tuning strategies to improve the invariance properties of off-the-shelf
CNNs in identification settings. This analysis is motivated by the observation in Fig. 10 that adaptive finetuning performs significantly better than conservative
and feature extraction based strategies. Indeed, this
suggests that adapting the representation of the CNNs
to the identification task is helpful to recognition, possibly because the network is actually learning the transformations to which it needs to be invariant. To this
end, here we investigate whether fine-tuning a CNN
on the iCubWorld domain can indeed improve/adapt
the invariance properties of the network to the identification task. We have performed a preliminary set of
experiments related to this question in (Pasquale et al
2016a), where we considered a smaller set of transformations and fine-tuning strategies, and focused only on
the CaffeNet architecture.
We consider the same learning setting of the experiments reported in Fig. 11, but we evaluate learning
models that are first “pre-fine-tuned” on one of the following datasets to improve their invariance:
• iCubWorld identification (iCWT id). This dataset
contains images of all instances of the 5 object categories of iCWT that were not used in the previous
experiments, namely oven glove. squeezer, sprayer,
body lotion and soda bottle. Fine-tuning was performed with the adaptive strategy on the identification task. Training and validation sets were obtained
following the same protocol of Sec. 5.1.
• iCubWorld categorization (iCWT cat). The
same dataset as (iCW id) but fine-tuning was performed on the categorization task for the 5 objects
Are we Done with Object Recognition? The iCub robot’s Perspective.
accuracy (%)
100
80
100
off-the-shelf
iCWT cat
iCWT id
iCWT + ImNet
ImNet
test: 2D ROT
100
test: 3D ROT
100
test: SCALE
100
80
80
80
80
60
60
60
60
60
40
40
40
40
40
20
2DR
100
accuracy (%)
test: MIX
15
80
3DR
SC
BKG
test: MIX
20
MIX
100
off-the-shelf
iCWT cat
iCWT id
iCWT + ImNet
ImNet
3DR
SC
BKG
test: 2D ROT
20
MIX
100
2DR
SC
BKG
test: 3D ROT
20
MIX
100
2DR
3DR
BKG
test: SCALE
20
MIX
100
80
80
80
80
60
60
60
60
60
40
40
40
40
40
20
2DR
3DR
SC
BKG
20
MIX
3DR
SC
BKG
20
MIX
2DR
SC
BKG
20
MIX
2DR
3DR
BKG
20
MIX
test: BKG
2DR
3DR
SC
test: BKG
2DR
3DR
SC
Fig. 12 Same experiment setting as in Fig. 11, but using different image representations, provided by CaffeNet (top, Orange)
or GoogLeNet (bottom, Blue) network models fine-tuned according to different strategies (see Sec. 5.2.1).
considered. Again, the adaptive strategy was used.
Train and validation sets were obtained following
the same protocol as in Sec. 4.2.1.
• iCubWorld + ImageNet (iCWT + ImNet).
This dataset contains images of the 5 categories
above but sampled from both iCWT and ImageNet.
Note that most iCWT categories do not appear in
ILSVRC but are in synsets in the larger ImageNet
dataset (see the supplementary material for a list of
the corresponding synsets). All images in the synset
were used. Fine-tuning was performed on the 5-class
categorization task. This dataset is conceived to test
the possibility to have the CNN directly learn the
relation between ImageNet examples on which it has
been originally trained and iCWT images.
• ImageNet (ImNet). This dataset contains the 5
ImageNet synsets corresponding to the 5 object categories on which the CNNs will be later trained
and tested for invariance, namely book, flower, glass,
hairbrush and hairclip. Fine-tuning was performed
on the 5-class categorization task. This dataset allows the network to focus on data available on-line
about the task at hand, but different from the images that the robot will observe.
Fig. 12 reports the accuracy of RLSC classifiers when
trained on features extracted by the CNNs fine-tuned
on the datasets described above. Training was performed
separately for each transformation similarly to previ-
ous section, on the dataset containing only 10 example images per object instance. It can be noticed that
pre-fine-tuning on the identification task is particularly
advantageous, leading to a dramatic improvement over
the model trained on off-the-shelf features. In particular, comparing these results with those in Fig. 11, we
see that (iCWT id) trained on 10 examples per object
performs on par with the off-the-shelf method trained
on 150 images per object. Moreover, the performance
of (iCWT id) are in general more stable across different viewpoint transformations, suggesting that the
preliminary fine-tuning could have indeed allowed the
CNNs to become partially invariant to transformations
in iCubWorld. Interestingly the other CNNs do not provide similar improvements to identification and, in some
cases, they even have worse performance than the offthe-shelf based classifier.
We conclude with a note related to categorization.
Indeed, following these experiments we could wonder
whether a similar pre-fine-tuning strategy could improve invariance also in the categorization setting of
Sec. 4.3. However, empirical evidence showed that when
performing fine-tuning on the datasets described above,
the performance of the CNNs does not change significantly on the categorization task. We refer to the supplementary material for the analysis of this specific setting. We hypothesize this to be due to two possible reasons: 1) networks trained on the ILSVRC are already
16
highly optimized for categorization and there is no gain
in adapting the visual representation or 2), the negative
effect of the limited semantic variability in iCubWorld
with respect to ImageNet may “overcome” the potential benefits of increasing the invariance to viewpoint
transformations. This motivates some of the proposed
directions of future investigation that we discuss in the
following section.
6 Discussion and Future Work
In this work we studied the application of modern Deep
Learning methods to visual recognition tasks in robotics.
We challenged Deep Learning methods on an object
recognition task that was specifically designed to represent a prototypical visual recognition problem in a
real robotics application. Our experiments show that
Deep Learning leads to remarkable performance for the
visual tasks of object classification and instance recognition. Proper adoption of knowledge transfer strategies
– in particular mixing Deep Learning with conventional
“shallow” classifiers based on Kernel methods – plays
a fundamental role, in that they leverage on the visual
representation learned on large-scale datasets to achieve
high performance in the robotic domain.
However, a substantial gap still exists between the
performance that can be obtained on the two domains.
Our analysis shows that this performance gap is due
to scarce semantic variability, which characterizes the
robotic domain and is due to the intrinsic cost of acquiring training samples. Yet, adoption of robotic systems in real applications requires to push further these
requirements to achieve performance that approaches
100%. Indeed, in real-world robotics applications, failures due to object mis-detection or mis-classification
can potentially lead to dramatically more critical and
harmful consequences than in standard image retrieval
scenarios. Therefore, the error rate of these systems will
need to be as close as possible to zero to be considered
for production and deployment in environments shared
with humans.
In this section we consider possible directions for future research, aimed at mitigating the impact of such
differences when performing visual recognition in robotics.
The goal of our discussion is to consider them together
in the same picture to present how, in our opinion, the
robot vision problem could be addressed.
Improving invariance. The experiments in Sec. 4.3
show that models tested in this work are mostly locally
invariant, namely their representation is robust to small
view-point changes. While local invariance has been of
main interest to computer vision in the past (Lowe
Giulia Pasquale et al.
2004; Mikolajczyk and Schmid 2004), current research
on invariance is instead focused on designing learning
systems able to learn and encode representations that
are robust to more dramatic (more “global”) transformations of an object (Anselmi et al 2016, 2015). Clearly,
the more robust the visual representation is to viewpoint changes, small deformations, background variations etc., the less the corresponding model trained on
an off-line large-scale dataset such as ImageNet will be
biased to the specific training conditions. Research on
invariant representations could therefore lead to both
robust off-the-shelf robotics systems that do not need
knowledge transfer and also allow to train on-the-fly
classifier that can be trained from very few examples of
a novel object and reliably recognize it from very different poses.
Multi-task learning. In this work we followed a typical approach to categorization, where no similarity or
relation is assumed among different classes. However, in
visual recognition settings object categories often have
a number of visual features in common (e.g. wheels,
windows, legs, handles and so on). Incorporating information about similarities across object classes within
the learning model could significantly improve the overall accuracy of a recognition system, especially in presence of scarce training data. In particular, in the literature on multi-task learning there have been proposed
several approaches to model relations across tasks (i.e.
object categories) to directly enforce them on the learning problem when they are available a-priori (Crammer and Singer 2000; Micchelli and Pontil 2004; Evgeniou et al 2005; Joachims et al 2009; Fergus et al 2010;
Lozano and Sindhwani 2011) or learn them when unknown (Argyriou et al 2008; Jacob et al 2008; Minh and
Sindhwani 2011; Dinuzzo et al 2011; Sindhwani et al
2012; Ciliberto et al 2015a,b).
“Augmenting” the Semantic Variability. The simplest method for increasing semantic variability is to
share data acquired in parallel from different robotic
systems. Indeed, a similar strategy has been used to
learn hand-eye coordination in a robotic setting, by
training networks using data acquired in parallel by
several robots (Levine et al 2016). Similarly, the Million Object Challenge (Oberlin et al 2015) aims to acquire a large dataset of object models by sharing data
acquired from different laboratories owning a Baxter
robot8 . Along a similar direction we plan to extend
iCubWorld with the help of the community of the iCub
robot (at the time of writing more than 30 research
groups in the world). Beyond expanding the dataset in
8
http://www.rethinkrobotics.com/
Are we Done with Object Recognition? The iCub robot’s Perspective.
the direction of semantic variability, this would also allow to collect multiple acquisitions of the same object
in different conditions, in order to extend the analysis
presented in this work to nuisances like changes in the
light and setting. Indeed, this is another critical aspect
that we started to consider in the supplementary material of this work.
An alternative, or complementary approach is data
augmentation. This method is often adopted in image
retrieval settings to increase the viewpoint (and also illumination) variability during training by applying synthetic transformations to each image in the dataset
(e.g. by rotating, flipping, cropping or changing the hue
value). Visual augmentation to cope for semantic variability is a much more challenging problem, although
recent work on inverse graphics (Mansinghka et al 2013;
Kulkarni et al 2014, 2015) is a starting point in this direction.
Exploiting Contextual Information. Most work in
visual recognition considers frames independently and
rarely uses contextual information. A robot is typically
exposed to a continuous stream of visual information,
in which information is highly correlated. Exploiting
these correlations can be done in many ways ranging
from trivial solutions (like temporal averaging of the
classification results) to more complex ones that rely
on scene reconstruction and object tracking (Song et al
2015)).
Integrating 3D Information. In robotics, the problem of integrating 3D informationwith RGB has been
recently investigated within a Deep Learning framework (Schwarz et al 2015; Redmon and Angelova 2015b).
These results, although very promising, leverage only
on off-the-shelf representations learned on large-scaled
datasets of RGB-only images. In the future, the challenge will be to determine how to best encode and process 3D information within a CNN but also to acquire
large-scale datasets for directly training such RGBDbased systems.
Self-supervised learning. One way to reduce the cost
of supervision is to implement explorative strategies in
which the robot autonomously interacts with the environment to collect training samples. Training instances
in this scenario could be extracted autonomously by
detecting invariances in the data which correspond to
physical entities (e.g. coherent motion pattern (Wang
and Gupta 2015) or bottom-up salency cues). Strategies specific to the robotic domain could be devised by
integrating multiple sensory modalities (Sinapov et al
2014b; Higy et al 2016) and a repertoire of explorative
17
actions designed specifically to extract these cues (like
grasping or pushing objects (Montesano et al 2008; Fitzpatrick et al 2003; Hgman et al 2016; Pinto et al 2016)).
7 Conclusions
In this paper we have described a systematic study
on the benefits of deep learning methods to robot vision. For our tests we have devised a prototypical vision task for a humanoid robot in which human-robot
interaction is exploited to obtain realistic supervision
and train an object recognition system. We presented
the iCWT dataset and an in-depth investigation of the
performance of state-of-the-art deep learning methods
applied to our task. Our results confirm deep learning is a remarkable step forward. However, there is still
a lot that needs to be done to reach the level of robustness and reliability required by real applications.
We identified specific challenges and possible directions
of research to bridge this gap. We are confident that
the next few years will be rich of exciting progress in
robotics.
Acknowledgements The work described in this paper is
supported by the Center for Brains, Minds and Machines
(CBMM), funded by NSF STC award CCF-1231216; and by
FIRB project RBFR12M3AC, funded by the Italian Ministry of Education, University and Research. We gratefully
acknowledge NVIDIA Corporation for the donation of the
Tesla k40 GPU used for this research.
References
Agrawal P, Girshick R, Malik J (2014) Analyzing the performance of multilayer neural networks for object recognition. In: Proceedings of the European Conference on
Computer Vision (ECCV)
Anselmi F, Leibo JZ, Rosasco L, Mutch J, Tacchetti A, Poggio T (2015) Unsupervised learning of invariant representations. Theoretical Computer Science DOI 10.1016/
j.tcs.2015.06.048, URL http://www.sciencedirect.com/
science/article/pii/S0304397515005587
Anselmi F, Rosasco L, Poggio T (2016) On invariance and selectivity in representation learning. Information and Inference 5(2):134–158, URL http:
//imaiai.oxfordjournals.org/content/5/2/134.full.
pdf+html?sid=932cd117-4307-409b-9fc0-3acd9be20af0
Argyriou A, Evgeniou T, Pontil M (2008) Convex multi-task
feature learning. Machine Learning 73
Azizpour H, Razavian AS, Sullivan J, Maki A, Carlsson S (2015) From generic to specific deep representations for visual recognition. In: 2015 IEEE Conference on Computer Vision and Pattern Recognition Workshops (CVPRW), pp 36–45, DOI 10.1109/CVPRW.2015.
7301270
Babenko A, Slesarev A, Chigorin A, Lempitsky V (2014)
Neural Codes for Image Retrieval, Springer International Publishing, Cham, pp 584–599. DOI 10.
18
1007/978-3-319-10590-1 38, URL http://dx.doi.org/
10.1007/978-3-319-10590-1_38
Baishya S, Bäuml B (2016) Robust material classification
with a tactile skin using deep learning. In: IEEE International Conference on Intelligent Robots and Systems
Bakry A, Elhoseiny M, El-Gaaly T, Elgammal A (2015) Digging Deep into the Layers of CNNs: In Search of How
CNNs Achieve View Invariance. arXiv preprint pp 1–20,
URL http://arxiv.org/abs/1508.01983, 1508.01983
Bishop CM (2006) Pattern Recognition and Machine Learning (Information Science and Statistics). Springer-Verlag
New York, Inc., Secaucus, NJ, USA
Borji A, Izadi S, Itti L (2016) ilab-20m: A large-scale controlled object dataset to investigate deep learning. In:
The IEEE Conference on Computer Vision and Pattern
Recognition (CVPR)
Bottou L (2012) Stochastic Gradient Tricks, vol
7700, Springer, p 430445. URL https://www.
microsoft.com/en-us/research/publication/
stochastic-gradient-tricks/
Chatfield K, Simonyan K, Vedaldi A, Zisserman A (2014)
Return of the devil in the details: Delving deep into convolutional nets. In: British Machine Vision Conference,
1405.3531
Chitta S, Sturm J, Piccoli M, Burgard W (2011) Tactile
Sensing for Mobile Manipulation. IEEE Transactions on
Robotics 27(3):558–568, DOI 10.1109/TRO.2011.2134130
Ciliberto C, Pattacini U, Natale L, Nori F, Metta G (2011)
Reexamining lucas-kanade method for real-time independent motion detection: Application to the icub humanoid
robot. In: 2011 IEEE/RSJ International Conference on
Intelligent Robots and Systems, IEEE, pp 4154–4160
Ciliberto C, Fanello SR, Natale L, Metta G (2012) A heteroscedastic approach to independent motion detection
for actuated visual sensors. In: 2012 IEEE/RSJ International Conference on Intelligent Robots and Systems,
IEEE, pp 3907–3913
Ciliberto C, Fanello SR, Santoro M, Natale L, Metta G,
Rosasco L (2013) On the impact of learning hierarchical representations for visual recognition in robotics. In:
2013 IEEE/RSJ International Conference on Intelligent
Robots and Systems, pp 3759–3764, DOI 10.1109/IROS.
2013.6696893
Ciliberto C, Mroueh Y, Poggio T, Rosasco L (2015a) Convex
learning of multiple tasks and their structure. In: International Conference on Machine Learning
Ciliberto C, Rosasco L, Villa S (2015b) Learning multiple visual tasks while discovering their structure. In: Proceedings of the IEEE Conference on Computer Vision and
Pattern Recognition, pp 131–139
Collet A, Berenson D, Srinivasa SS, Ferguson D (2009) Object
recognition and full pose registration from a single image
for robotic manipulation. In: Robotics and Automation,
2009. ICRA’09. IEEE International Conference on, IEEE,
pp 48–55
Collet A, Martinez M, Srinivasa SS (2011a) The moped
framework: Object recognition and pose estimation for
manipulation. The International Journal of Robotics Research p 0278364911401765
Collet A, Srinivasay SS, Hebert M (2011b) Structure discovery in multi-modal data: a region-based approach. In:
Robotics and Automation (ICRA), 2011 IEEE International Conference on, IEEE, pp 5695–5702
Crammer K, Singer Y (2000) On the learnability and design of
output codes for multiclass problems. In: In Proceedings
of the Thirteenth Annual Conference on Computational
Giulia Pasquale et al.
Learning Theory, pp 35–46
Crowley E, Zisserman A (2014) The state of the art: Object retrieval in paintings using discriminative regions. In:
BMVC
Dahiya RS, Metta G, Valle M, Sandini G (2010) Tactile Sensing - From Humans to Humanoids. IEEE Transactions on
Robotics 26(1):1–20, DOI 10.1109/TRO.2009.2033627
Dansereau DG, Singh SP, Leitner J (2016) Interactive computational imaging for deformable object analysis. In: 2016
International Conference on Robotics and Automation
(ICRA), IEEE
Deng J, Dong W, Socher R, Li LJ, Li K, Fei-Fei L (2009)
ImageNet: A Large-Scale Hierarchical Image Database.
In: CVPR09
Dinuzzo F, Ong CS, Gehler P, Pillonetto G (2011) Learning output kernels with block coordinate descent. International Conference on Machine Learning
Donahue J, Jia Y, Vinyals O, Hoffman J, Zhang N, Tzeng E,
Darrell T (2014) Decaf: A deep convolutional activation
feature for generic visual recognition. In: Jebara T, Xing
EP (eds) Proceedings of the 31st International Conference
on Machine Learning (ICML-14), JMLR Workshop and
Conference Proceedings, pp 647–655, URL http://jmlr.
org/proceedings/papers/v32/donahue14.pdf
Eitel A, Springenberg JT, Spinello L, Riedmiller M, Burgard W (2015) Multimodal deep learning for robust rgbd object recognition. In: Intelligent Robots and Systems
(IROS), 2015 IEEE/RSJ International Conference on, pp
681–687, DOI 10.1109/IROS.2015.7353446
Everingham M, Van Gool L, Williams CKI, Winn J, Zisserman A (2010) The pascal visual object classes (voc)
challenge. International Journal of Computer Vision
88(2):303–338, DOI 10.1007/s11263-009-0275-4, URL
http://dx.doi.org/10.1007/s11263-009-0275-4
Everingham M, Eslami SMA, Van Gool L, Williams CKI,
Winn J, Zisserman A (2015) The pascal visual object
classes challenge: A retrospective. International Journal
of Computer Vision 111(1):98–136
Evgeniou T, Micchelli CA, Pontil M (2005) Learning multiple tasks with kernel methods. In: Journal of Machine
Learning Research, pp 615–637
Fanello SR, Ciliberto C, Natale L, Metta G (2013a) Weakly
supervised strategies for natural object recognition in
robotics. IEEE International Conference on Robotics
and Automation pp 4223–4229, DOI 10.1109/ICRA.2013.
6631174
Fanello SR, Ciliberto C, Santoro M, Natale L, Metta G,
Rosasco L, Odone F (2013b) iCub World: Friendly Robots
Help Building Good Vision Data-Sets. In: 2013 IEEE
Conference on Computer Vision and Pattern Recognition
Workshops, pp 700–705, DOI 10.1109/CVPRW.2013.106
Fergus R, Bernal H, Weiss Y, Torralba A (2010) Semantic
label sharing for learning with many categories. European
Conference on Computer Vision
Fitzpatrick P, Metta G, Natale L, Rao S, Sandini G (2003)
Learning About Objects Through Action – Initial Steps
Towards Artificial Cognition. vol 3, pp 3140–3145, DOI
10.1109/ROBOT.2003.1242036
Fitzpatrick P, Needham A, Natale L, Metta G (2008) Shared
challenges in object perception for robots and infants. Infant and Child Development 17(1):7–24
Geusebroek JM, Burghouts GJ, Smeulders AW (2005) The
amsterdam library of object images. International Journal of Computer Vision 61(1):103–112, DOI 10.1023/B:
VISI.0000042993.50813.60, URL http://dx.doi.org/10.
1023/B:VISI.0000042993.50813.60
Are we Done with Object Recognition? The iCub robot’s Perspective.
Goehring D, Hoffman J, Rodner E, Saenko K, Darrell T
(2014) Interactive adaptation of real-time object detectors. In: Proceedings - IEEE International Conference on
Robotics and Automation, IEEE, pp 1282–1289, DOI
10.1109/ICRA.2014.6907018
Goodfellow I, Lee H, Le QV, Saxe A, Ng AY (2009)
Measuring invariances in deep networks. In: Bengio Y, Schuurmans D, Lafferty JD, Williams CKI,
Culotta A (eds) Advances in Neural Information
Processing Systems 22, Curran Associates, Inc.,
pp 646–654, URL http://papers.nips.cc/paper/
3790-measuring-invariances-in-deep-networks.pdf
Goodfellow I, Bengio Y, Courville A (2016) Deep learning,
URL http://www.deeplearningbook.org, book in preparation for MIT Press
Gordo A, Almazán J, Revaud J, Larlus D (2016) Deep Image Retrieval: Learning Global Representations for Image Search, Springer International Publishing, Cham, pp
241–257. DOI 10.1007/978-3-319-46466-4 15, URL http:
//dx.doi.org/10.1007/978-3-319-46466-4_15
Gorges N, Navarro SE, Gger D, Wrn H (2010) Haptic object
recognition using passive joints and haptic key features.
In: 2010 IEEE International Conference on Robotics and
Automation, pp 2349–2355, DOI 10.1109/ROBOT.2010.
5509553
Griffin G, Holub A, Perona P (2007) Caltech-256 object
category dataset. Tech. Rep. 7694, California Institute
of Technology, URL http://authors.library.caltech.
edu/7694
He K, Zhang X, Ren S, Sun J (2015) Delving deep into rectifiers: Surpassing human-level performance on imagenet
classification. In: The IEEE International Conference on
Computer Vision (ICCV)
He K, Zhang X, Ren S, Sun J (2016) Deep residual learning
for image recognition. In: Computer Vision and Pattern
Recognition (CVPR), 2016 IEEE Conference on
Held D, Thrun S, Savarese S (2016) Robust single-view instance recognition. In: 2016 IEEE International Conference on Robotics and Automation (ICRA), pp 2152–2159,
DOI 10.1109/ICRA.2016.7487365
Herranz L, Jiang S, Li X (2016) Scene Recognition With
CNNs: Objects, Scales and Dataset Bias. In: Conference
on Computer Vision and Pattern Recognition, pp 571–
579, DOI 10.1109/CVPR.2016.68
Higy B, Ciliberto C, Rosasco L, Natale L (2016) Combining
sensory modalities and exploratory procedures to improve
haptic object recognition in robotics. In: IEEE-RAS International Conference on Humanoid Robots
Hinton GE, Srivastava N, Krizhevsky A, Sutskever I,
Salakhutdinov RR (2012) Improving neural networks by
preventing co-adaptation of feature detectors. ArXiv eprints 1207.0580
Hoffman J, Tzeng E, Donahue J, Jia Y, Saenko K, Darrell T
(2013) One-Shot Adaptation of Supervised Deep Convolutional Models. ArXiv e-prints 1312.6204
Hosoda K, Iwase T (2010) Robust haptic recognition by
anthropomorphic bionic hand through dynamic interaction. In: Intelligent Robots and Systems (IROS), 2010
IEEE/RSJ International Conference on, IEEE, pp 1236–
1241
Huh M, Agrawal P, Efros AA (2016) What makes ImageNet
good for transfer learning? ArXiv e-prints 1608.08614
Hgman V, Bjrkman M, Maki A, Kragic D (2016) A Sensorimotor Learning Framework for Object Categorization.
IEEE Transactions on Cognitive and Developmental Systems 8(1):15–25, DOI 10.1109/TAMD.2015.2463728
19
Ioffe S, Szegedy C (2015) Batch normalization: Accelerating deep network training by reducing internal covariate shift. vol 37, pp 448–456, URL http://www.jmlr.org/
proceedings/papers/v37/ioffe15
Jacob L, Bach F, Vert JP (2008) Clustered multi-task learning: a convex formulation. Advances in Neural Information Processing Systems
Jia Y, Shelhamer E, Donahue J, Karayev S, Long J, Girshick
R, Guadarrama S, Darrell T (2014) Caffe: Convolutional
architecture for fast feature embedding. In: Proceedings
of the ACM International Conference on Multimedia MM ’14, ACM Press, pp 675–678, DOI 10.1145/2647868.
2654889
Joachims T, Hofmann T, Yue Y, Yu CN (2009) Predicting
structured objects with support vector machines. Commun ACM 52(11):97–104, DOI 10.1145/1592761.1592783,
URL http://doi.acm.org/10.1145/1592761.1592783
Kasper A, Xue Z, Dillmann R (2012) The kit object
models database: An object model database for object recognition, localization and manipulation in service
robotics. The International Journal of Robotics Research
31(8):927–934
Kemp CC, Edsinger A, Torres-Jara E (2007) Challenges for
robot manipulation in human environments [grand challenges of robotics]. IEEE Robotics Automation Magazine
14(1):20–29, DOI 10.1109/MRA.2007.339604
Khosla A, Zhou T, Malisiewicz T, Efros AA, Torralba
A (2012) Undoing the Damage of Dataset Bias,
Springer Berlin Heidelberg, Berlin, Heidelberg, pp 158–
171. DOI 10.1007/978-3-642-33718-5 12, URL http://
dx.doi.org/10.1007/978-3-642-33718-5_12
Kingma D, Ba J (2015) Adam: A Method for Stochastic Optimization. In: 3rd International Conference for Learning
Representations (ICLR)
Krizhevsky A, Sutskever I, Hinton G (2012a) Imagenet classification with deep convolutional neural networks. In:
Advances in Neural Information Processing Systems, pp
1097–1105
Krizhevsky A, Sutskever I, Hinton GE (2012b) Imagenet classification with deep convolutional neural
networks. In: Pereira F, Burges CJC, Bottou L,
Weinberger KQ (eds) Advances in Neural Information Processing Systems 25, Curran Associates, Inc.,
pp 1097–1105, URL http://papers.nips.cc/paper/
4824-imagenet-classification-with-deep-convolutional-neural-net
pdf
Kulkarni TD, Mansinghka VK, Kohli P, Tenenbaum JB
(2014) Inverse graphics with probabilistic cad models.
arXiv preprint arXiv:14071339
Kulkarni TD, Whitney WF, Kohli P, Tenenbaum J (2015)
Deep convolutional inverse graphics network. In: Advances in Neural Information Processing Systems, pp
2539–2547
Lai K, Bo L, Ren X, Fox D (2011) A large-scale hierarchical multi-view RGB-D object dataset. In: 2011 IEEE
International Conference on Robotics and Automation,
IEEE, pp 1817–1824, DOI 10.1109/ICRA.2011.5980382,
URL
http://ieeexplore.ieee.org/articleDetails.
jsp?arnumber=5980382
LeCun Y, Boser B, Denker JS, Henderson D, Howard RE,
Hubbard W, Jackel LD (1989) Backpropagation applied
to handwritten zip code recognition. Neural computation
1(4):541–551
LeCun Y, Huang FJ, Bottou L (2004) Learning methods for
generic object recognition with invariance to pose and
lighting. In: Computer Vision and Pattern Recognition,
20
2004. CVPR 2004. Proceedings of the 2004 IEEE Computer Society Conference on, vol 2, pp II–97–104 Vol.2,
DOI 10.1109/CVPR.2004.1315150
Leitner J, Förster A, Schmidhuber J (2014) Improving robot
vision models for object detection through interaction. In:
2014 International Joint Conference on Neural Networks
(IJCNN), IEEE, pp 3355–3362
Leitner J, Dansereau DG, Shirazi S, Corke P (2015) The
need for more dynamic and active datasets. In: CVPR
Workshop on The Future of Datasets in Computer Vision, IEEE
Levine S, Pastor P, Krizhevsky A, Quillen D (2016)
Learning Hand-Eye Coordination for Robotic Grasping
with Deep Learning and Large-Scale Data Collection.
arXiv:160302199 [cs] URL http://arxiv.org/abs/1603.
02199, arXiv: 1603.02199
Lowe DG (2004) Distinctive image features from
scale-invariant
keypoints.
International
Journal of Computer Vision 60(2):91–110, DOI
10.1023/B:VISI.0000029664.99615.94,
URL
http:
//dx.doi.org/10.1023/B:VISI.0000029664.99615.94
Lozano A, Sindhwani V (2011) Block variable selection in
multivariate regression and high-dimensional causal inference. Advances in Neural Information Processing Systems
Mansinghka V, Kulkarni TD, Perov YN, Tenenbaum J (2013)
Approximate bayesian image interpretation using generative probabilistic graphics programs. In: Advances in Neural Information Processing Systems, pp 1520–1528
Meger D, Little JJ (2013) The UBC Visual Robot Survey: A Benchmark for Robot Category Recognition,
Springer International Publishing, Heidelberg, pp 979–
991. DOI 10.1007/978-3-319-00065-7 65, URL http://
dx.doi.org/10.1007/978-3-319-00065-7_65
Metta G, Sandini G, Natale L, Craighero L, Fadiga L (2006)
Understanding mirror neurons: a bio-robotic approach.
Interaction studies 7(2):197–232
Metta G, Natale L, Nori F, Sandini G, Vernon D, Fadiga L,
von Hofsten C, Rosander K, Lopes M, Santos-Victor J,
Bernardino A, Montesano L (2010) The icub humanoid
robot: an open-systems platform for research in cognitive development. Neural networks : the official journal of
the International Neural Network Society 23(8-9):1125–
34, DOI 10.1016/j.neunet.2010.08.010
Micchelli CA, Pontil M (2004) Kernels for multi-task learning.
Advances in Neural Information Processing Systems
Mikolajczyk K, Schmid C (2004) Scale & affine invariant
interest point detectors. International journal of computer vision 60(1):63–86, URL http://link.springer.
com/article/10.1023/B:VISI.0000027790.02288.f2
Minh HQ, Sindhwani V (2011) Vector-valued manifold regularization. International Conference on Machine Learning
Model I, Shamir L (2015) Comparison of Data Set Bias
in Object Recognition Benchmarks. IEEE Access
3:1953–1962, DOI 10.1109/ACCESS.2015.2491921, URL
http://ieeexplore.ieee.org/lpdocs/epic03/wrapper.
htm?arnumber=7299607
Moldovan B, Moreno P, van Otterlo M, Santos-Victor J,
De Raedt L (2012) Learning relational affordance models for robots in multi-object manipulation tasks. In:
Robotics and Automation (ICRA), 2012 IEEE International Conference on, IEEE, pp 4373–4378
Montesano L, Lopes M, Bernardino A, Santos-Victor J (2008)
Learning object affordances: From sensory–motor coordination to imitation. IEEE Transactions on Robotics
24(1):15–26
Giulia Pasquale et al.
Muja M, Rusu RB, Bradski G, Lowe DG (2011) Rein-a fast,
robust, scalable recognition infrastructure. In: Robotics
and Automation (ICRA), 2011 IEEE International Conference on, IEEE, pp 2939–2946
Natale L, Metta G, Sandini G (2004) Learning
haptic
representation
of
objects.
URL
http:
//www.lira.dist.unige.it/projects/mirror/docs/
thirdyear/papers/img04.pdf
Nene SA, Nayar SK, Murase H (1996) Columbia object image
library (coil-100. Tech. rep.
Nguyen A, Kanoulas D, Caldwell DG, Tsagarakis NG (2016)
Detecting Object Affordances with Convolutional Neural Networks. In: Intelligent Robots and Systems (IROS),
2016 IEEE/RJS International Conference on (to appear)
Oberlin J, Meier M, Kraska T, Tellex S (2015) Acquiring Object Experiences at Scale. In: AAAI-RSS Special Workshop on the 50th Anniversary of Shakey: The Role of AI
to Harmonize Robots and Humans, blue Sky Award.
Oquab M, Bottou L, Laptev I, Sivic J (2014) Learning and
transferring mid-level image representations using convolutional neural networks. In: The IEEE Conference on
Computer Vision and Pattern Recognition (CVPR)
Pasquale G, Ciliberto C, Odone F, Rosasco L, Natale L (2015)
Teaching iCub to recognize objects using deep Convolutional Neural Networks. vol 43, pp 21–25, URL http:
//www.jmlr.org/proceedings/papers/v43/pasquale15
Pasquale G, Ciliberto C, Rosasco L, Natale L (2016a) Object Identification from Few Examples by Improving the
Invariance of a Deep Convolutional Neural Network. In:
Intelligent Robots and Systems (IROS), 2016 IEEE/RJS
International Conference on (to appear)
Pasquale G, Mar T, Ciliberto C, Rosasco LA, Natale
L (2016b) Enabling depth-driven visual attention on the icub humanoid robot: Instructions for
use and new perspectives. Frontiers in Robotics
and AI 3(35), DOI 10.3389/frobt.2016.00035, URL
http://www.frontiersin.org/humanoid_robotics/10.
3389/frobt.2016.00035/abstract
Philbin J, Chum O, Isard M, Sivic J, Zisserman A (2008) Lost
in quantization: Improving particular object retrieval in
large scale image databases. In: Computer Vision and
Pattern Recognition, 2008. CVPR 2008. IEEE Conference on, IEEE, pp 1–8
Pinto L, Gupta A (2015) Supersizing self-supervision: Learning to grasp from 50k tries and 700 robot hours. arXiv
preprint arXiv:150906825 URL http://arxiv.org/abs/
1509.06825
Pinto L, Gupta A (2016) Supersizing self-supervision: Learning to grasp from 50k tries and 700 robot hours. In: 2016
IEEE International Conference on Robotics and Automation (ICRA), pp 3406–3413, DOI 10.1109/ICRA.2016.
7487517
Pinto L, Gandhi D, Han Y, Park YL, Gupta A (2016)
The Curious Robot: Learning Visual Representations via
Physical Interactions. arXiv:160401360 [cs] URL http:
//arxiv.org/abs/1604.01360, arXiv: 1604.01360
Pinto N, Cox DD, DiCarlo JJ (2008) Why is real-world
visual object recognition hard? PLoS computational biology 4(1):e27, DOI 10.1371/journal.pcbi.0040027, URL
http://journals.plos.org/ploscompbiol/article?id=
10.1371/journal.pcbi.0040027
Ramisa A, Aldavert D, Vasudevan S, Toledo R, Lopez de
Mantaras R (2011) The iiia30 mobile robot object recognition dataset. In: Robotica 2011
Redmon J, Angelova A (2014) Real-Time Grasp Detection
Using Convolutional Neural Networks. arXiv:14123128
Are we Done with Object Recognition? The iCub robot’s Perspective.
[cs] URL http://arxiv.org/abs/1412.3128, arXiv:
1412.3128
Redmon J, Angelova A (2015a) Real-time grasp detection using convolutional neural networks. In: 2015 IEEE International Conference on Robotics and Automation (ICRA),
pp 1316–1322, DOI 10.1109/ICRA.2015.7139361
Redmon J, Angelova A (2015b) Real-time grasp detection
using convolutional neural networks. In: IEEE International Conference on Robotics and Automation, IEEE,
pp 1316–1322, URL http://ieeexplore.ieee.org/xpls/
abs_all.jsp?arnumber=7139361
Rennie C, Shome R, Bekris KE, Ferreira De Souza A
(2016) A dataset for improved rgbd-based object detection and pose estimation for warehouse pick-and-place.
IEEE Robotics and Automation Letters (RA-L) [Also accepted to appear at the 2016 IEEE International Conference on Robotics and Automation (ICRA)] 1:1179
– 1185, URL http://www.cs.rutgers.edu/~kb572/pubs/
icra16_pose_estimation.pdf
Rivera-Rubio J, Idrees S, Alexiou I, Hadjilucas L, Bharath
AA (2014) Small hand-held object recognition test
(SHORT). In: 2014 IEEE Winter Conference on Applications of Computer Vision (WACV), pp 524–531, DOI
10.1109/WACV.2014.6836057
Rodner E, Hoffman J, Donahue J, Darrell T, Saenko K
(2013) Towards Adapting ImageNet to Reality: Scalable
Domain Adaptation with Implicit Low-rank Transformations. ArXiv e-prints 1308.4200
Rudi A, Camoriano R, Rosasco L (2015) Less is more: Nystrom computational regularization. In: Advances in Neural Information Processing Systems, pp 1648–1656
Rudi A, Camoriano R, Rosasco L (2016) Generalization properties of learning with random features. arXiv preprint
arXiv:160204474
Russakovsky O, Deng J, Su H, Krause J, Satheesh S, Ma
S, Huang Z, Karpathy A, Khosla A, Bernstein M, Berg
AC, Fei-Fei L (2015) Imagenet large scale visual recognition challenge. International Journal of Computer Vision
115(3):211–252, DOI 10.1007/s11263-015-0816-y, URL
http://dx.doi.org/10.1007/s11263-015-0816-y
Schwarz M, Schulz H, Behnke S (2015) Rgb-d object recognition and pose estimation based on pre-trained convolutional neural network features. In: 2015 IEEE International Conference on Robotics and Automation (ICRA),
pp 1329–1335, DOI 10.1109/ICRA.2015.7139363
Simonyan K, Zisserman A (2015) Very deep convolutional
networks for large-scale image recognition. In: International Conference on Learning Representations
Sinapov J, Schenck C, Staley K, Sukhoy V, Stoytchev A
(2014a) Grounding semantic categories in behavioral
interactions: Experiments with 100 objects. Robotics
and Autonomous Systems 62(5):632–645, DOI 10.1016/
j.robot.2012.10.007, URL http://linkinghub.elsevier.
com/retrieve/pii/S092188901200190X
Sinapov J, Schenck C, Stoytchev A (2014b) Learning relational object categories using behavioral exploration and
multimodal perception. In: IEEE International Conference on Robotics and Automation, IEEE, pp 5691–5698,
URL http://ieeexplore.ieee.org/xpls/abs_all.jsp?
arnumber=6907696
Sindhwani V, Lozano AC, Minh HQ (2012) Scalable matrixvalued kernel learning and high-dimensional nonlinear
causal inference. CoRR abs/1210.4792
Singh A, Sha J, Narayan KS, Achim T, Abbeel P (2014) Bigbird: A large-scale 3d database of object instances. In:
2014 IEEE International Conference on Robotics and Au-
21
tomation (ICRA), pp 509–516, DOI 10.1109/ICRA.2014.
6906903
Song S, Zhang L, Xiao J (2015) Robot In a Room: Toward Perfect Object Recognition in Closed Environments.
arXiv:150702703 [cs] URL http://arxiv.org/abs/1507.
02703, arXiv: 1507.02703
Srivastava N, Hinton G, Krizhevsky A, Sutskever I, Salakhutdinov R (2014) Dropout: A simple way to prevent neural
networks from overfitting. Journal of Machine Learning
Research 15:1929–1958, URL http://jmlr.org/papers/
v15/srivastava14a.html
Stamos D, Martelli S, Nabi M, McDonald A, Murino V, Pontil M (2015) Learning with dataset bias in latent subcategory models. In: Proceedings of the IEEE Computer Society Conference on Computer Vision and Pattern Recognition, IEEE, vol 07-12-June, pp 3650–3658, DOI 10.1109/
CVPR.2015.7298988, URL http://ieeexplore.ieee.
org/lpdocs/epic03/wrapper.htm?arnumber=7298988
Sünderhauf N, Shirazi S, Dayoub F, Upcroft B, Milford M
(2015) On the performance of convnet features for place
recognition. In: Intelligent Robots and Systems (IROS),
2015 IEEE/RSJ International Conference on, pp 4297–
4304, DOI 10.1109/IROS.2015.7353986
Sünderhauf N, Dayoub F, McMahon S, Talbot B, Schulz R,
Corke P, Wyeth G, Upcroft B, Milford M (2016) Place
categorization and semantic mapping on a mobile robot.
In: 2016 IEEE International Conference on Robotics and
Automation (ICRA), pp 5729–5736, DOI 10.1109/ICRA.
2016.7487796
Szegedy C, Liu W, Jia Y, Sermanet P, Reed S, Anguelov
D, Erhan D, Vanhoucke V, Rabinovich A (2015) Going
deeper with convolutions. In: The IEEE Conference on
Computer Vision and Pattern Recognition (CVPR)
Tommasi T, Patricia N, Caputo B, Tuytelaars T (2015) A
deeper look at dataset bias. In: Lecture Notes in Computer Science (including subseries Lecture Notes in Artificial Intelligence and Lecture Notes in Bioinformatics),
Springer International Publishing, vol 9358, pp 504–516,
DOI 10.1007/978-3-319-24947-6 42
Torralba A, Efros AA (2011) Unbiased look at dataset bias.
In: Proceedings of the IEEE Computer Society Conference on Computer Vision and Pattern Recognition, pp
1521–1528, DOI 10.1109/CVPR.2011.5995347
Wang X, Gupta A (2015) Unsupervised Learning of Visual
Representations using Videos. arXiv:150500687 [cs] URL
http://arxiv.org/abs/1505.00687, arXiv: 1505.00687
Xiang Y, Mottaghi R, Savarese S (2014) Beyond pascal: A
benchmark for 3d object detection in the wild. In: IEEE
Winter Conference on Applications of Computer Vision,
pp 75–82, DOI 10.1109/WACV.2014.6836101
Yosinski J, Clune J, Bengio Y, Lipson H (2014) How transferable are features in deep neural networks? In: Advances
in Neural Information Processing Systems, pp 3320–3328
Zeiler MD, Fergus R (2014) Visualizing and Understanding
Convolutional
Networks,
Springer
International
Publishing,
Cham,
pp
818–833.
URL
http:
DOI
10.1007/978-3-319-10590-1 53,
//dx.doi.org/10.1007/978-3-319-10590-1_53
22
Giulia Pasquale et al.
Are we Done with Object Recognition? The iCub robot’s Perspective. Supplementary Material
Giulia Pasquale, Carlo Ciliberto, Francesca Odone, Lorenzo Rosasco and Lorenzo Natale
A iCubWorld Transformations
Fig. 13 reports one image for each object instance in the
iCWT dataset. We report also the reference number of the
associated synset in ImageNet (Russakovsky et al 2015). Note
that not all categories in iCWT are part of the ILSVRC, but
they belong (or at least are similar) to synsets in the larger
ImageNet dataset (Deng et al 2009).
Table 4 Preprocessing operations executed on images before
feeding the networks.
Mean
Subtraction
Scaling
Crop
Extraction
CaffeNet
ResNet-50
image
mean image size
(256 × 256)
2 × 2 grid + center
mirrored
GoogLeNet
pixel
256 × 256
2 × 2 grid + center
mirrored
VGG-16
pixel
shorter side to
256, 384, 512
Model
B Off-the-shelf CNNs
In Sec. 4.1 we observed that direct application of off-the-shelf
CNNs to iCubWorld leads to poor performance. In this section we provide some additional details on the experiment
that we reported in Fig. 5.
B.1 Image Preprocessing
When applying deep CNNs to our setting, we first evaluated
the impact of considering coarser or finer regions around the
object in the images. Indeed, each image in iCWT is annotated with the coordinates of the object’s centroid and with
a bounding box provided by segmentation of the depth map
(see Sec. 2.1). We took advantage of this rough localization
in order to avoid considering full images for classification and
we compared different strategies to extract a crop from the
images. Specifically, we tried either to extract a square crop
of fixed radius, just by using the information on the centroid,
or to use the bounding box provided by the routine segmenting the depth map (i.e., the smallest rectangle enclosing the
object’s blob). In both cases we tried two sizes: using a radius
of 256 or 384 for the square crop, and leaving a margin of 30
or 60 pixels around the depth’s bounding box.
Then, since deep CNNs require a fixed sized input (227 ×
227 for CaffeNet and 224 × 224 for the other considered models), different preprocessing strategies can be applied in order
to feed an image to a network. We reproduced the operations that are specified at the reference page of each particular Caffe model. For convenience we report them in Tab. 4.
These, in general, consist in the subtraction of the mean image (or the mean pixel value) of the dataset on which the
model has been trained and in the extraction of a grid of crops
(eventually at multiple scales as for VGG-16) of the required
size. Usually, the final prediction is computed by aggregating
the predictions of more crops to gain robustness. However,
since in our case the image is in fact already a crop around
the object, we compared the advantage of this grid approach
against just considering the central crop (and a single scale
in the case of VGG-16).
Fig. 14 reports the results of Sec. 4.1, Fig. 5, but obtained
by testing off-the-shelf CNNs on iCWT with the different described strategies. It can be noticed that performing a finer
5 × 5 grid
at each scale
mirrored
localization of the object in the image (Blue and Green) provides better performance for all architectures. Also, considering more than the central crop provides a little advantage (for
VGG-16 this is even detrimental). This is reasonable because
we already localized the object and we are in fact considering
a portion of the image.
We finally chose to apply CNNs by extracting a square
region of 256 × 256 from iCWT images, and then considering
only the central crop (Light Green in Fig. 14). Hence, we
did not need to scale the image (for VGG-16 we considered
a single scale). This is the same performance reported, for
instance, in Fig. 5 (Dark Blue) and the strategy used in all
experiments. We chose to extract fixed-size regions, instead of
relying on the bounding box of the depth map, since we aimed
at evaluating, in the following of our analysis, also invariance
properties to geometric transformations (as, e.g, scaling). We
precise that when applying the networks to ImageNet instead,
we used the multi-crop strategy suggested for each of the
architectures (see Tab. 4).
B.2 1000-class Categorization Results
In Fig. 5 we reported the accuracy of the off-the-shelf CNNs
applied on iCWT and ImageNet. However, since the task was
reduced to a 11-class categorization setting, we considered
the prediction of the networks limited to the 11 corresponding labels rather than the full vector of 1000 scores normally
produced by a CNN trained on the ILSVRC. For completeness, in Fig. 15 we report the resulting accuracy of the same
CNNs but taking as prediction the class achieving maximum
score over all available 1000, even if not present in the considered test sets (the one of iCWT and the reduced ImageNet).
As can be noticed, performance drops even below chance on
iCWT (since now the “real” chance is 1/1000 rather 1/11).
B.3 Viewpoint Biases
In this section we follow-up the preliminary analysis reported
in Sec. 4.1 discussing the potential biases in ImageNet, pre-
Are we Done with Object Recognition? The iCub robot’s Perspective.
soda bottle
n03983396
cellphone
n02992529
mouse
n03793489
coffee mug
n03063599
pencil case
n03908618
perfume
n03916031
remote
n04074963
ring binder
n02840245
soap dispenser
n04254120
sunglasses
n04356056
wallet
n04548362
oven glove
n02885462
squeezer
n04293119
sprayer
n02754103
body lotion
n02862916
book
n02870526
flower
n11669921
glass
n03438257
hairbrush
n03475581
hair clip
n03476684
Fig. 13 Example images for the 200 objects in iCubWorld Transformations.
23
24
Giulia Pasquale et al.
100
90
80
accuracy (%)
70
60
bb disp 30 - multi crop
bb disp 30
bb disp 60 - multi crop
bb disp 60
256 x 256 - multi crop
256 x 256
384 x 384 - multi crop
384 x 384
50
40
30
20
10
0
CaffeNet
GoogLeNet
VGG-16
ResNet-50
Fig. 14 Average classification accuracy of off-the-shelf networks tested on iCWT segmenting the object according to
different strategies.
100
90
iCWT off the shelf
ImageNet
80
accuracy (%)
70
the 10 object instances of that category as rows of a matrix (the temporal dimension, or frame index, is reported in
the horizontal axis). In each row, the frame-by-frame predictions over the sequence are represented as vertical bars:
W hite if the prediction is correct, Red if it is wrong (and
Black if the sequence is terminated). It can be noted that,
since during the acquisition the operator was moving the object back and forth in front of the robot, starting close and
slowly moving backward (eventually re-approaching the robot
in some sequences), the CNN manages to recognize the object
only when this appears at a relatively large scale (white bars
mostly concentrated in the first half of rows). This qualitative
observation confirms recent studies (Herranz et al 2016) that
specifically highlight the problem of the scale bias in CNNs,
preventing models trained on object-centric datasets as ImageNet to generalize to scene-centric settings, where objects
mostly appear in a small part of the image (like in the second
part of our sequences).
In Fig. 17 we report the same for sequences containing 2D
Rotation. In this case, the operator was rotating the object
(always maintaining the same face visible to the robot) at a
constant speed. Indeed, it can be noticed that on many sequences there are periodic time intervals when the CNN does
(or does not) recognize the object, corresponding to configurations in which that category appears more or less frequently
in the ImageNet dataset. For example, we observed that soap
dispensers and mugs in ImageNet are mostly placed on tables, and consequently are never recognized when they are,
e.g. upside-down.
60
50
C Model (Pre) Selection
40
30
In this section we provide details regarding the parameter
selection of the learning methods adopted in this work and
described in Sec. 3.2.
20
10
0
CaffeNet
GoogLeNet
VGG-16
ResNet-50
Fig. 15 Average classification accuracy of off-the-shelf networks (trained on ILSVRC) tested on iCWT (Dark Blue) or
on ImageNet itself (Gray). The test sets for the two datasets
are restricted to the 11 shared categories (see Sec. 2).
C.1 Feature Extraction and RLSC
Here we specify the design choices that we made to implement
each of the steps described in Sec. 3.2.1.
When feeding images to the networks in order to extract
their representation, we applied the preprocessing pipeline
that was explained and motivated in Sec. B.1. For each archiventing off-the-shelf models to generalize well when tested on
tecture, the layers considered as output to extract such repreiCubWorld without applying knowledge transfer techniques.
sentations, were specified in Tab. 2. For CaffeNet and VGGTo this end, we report a series of excerpts showing the frame16 we indicated either fc6 or fc7 layers, since these networks
by-frame predictions of some of the CNNs considered when
have more than one FC layer that can provide 1-dimensional
tested on iCWT sequences. This qualitative analysis allows
output representations of the full image, the best choice deto better understand what frames of iCubWorld are actually
pending on the application. We then resorted to (Rudi et al
“harder” to recognize for the CNNs and compare them with
2015) for the Nystrom subsampling approach that we used
prototypical examples in ImageNet.
to implement the Gaussian RLSC. This algorithm in fact inWe report here the predictions obtained by using GoogLeNet, volves determining some hyperparameters, the most critical
being the number m of training examples to subsample to
but we made similar observations for the other architectures.
approximate the kernel matrix.
As in Fig. 5, the prediction was computed as the maximum
score among those of 11 the classes present in test set, rather
In order to evaluate the best output layer for CaffeNet
than the full vector of 1000. We restricted the analysis to a
and VGG-16, and also to assign a reasonable value to m,
subset of the test set considered in Fig. 5: specifically, to the
we considered one small and one large categorization tasks
5 best recognized categories and two specific visual transforon iCWT, representative of the smallest and larger task that
mations: Scale and 2D Rotation.
we expected to run in our analysis, and compared the RLCS
accuracy on them when varying these factors.
In Fig. 16 we report frame-by-frame predictions on sequences containing the Scale transformation separately for
We fixed a 15-class categorization problem, training on 7
object instances per category and validating on 2. For each
each category. In each plot, we represent the sequences of
Are we Done with Object Recognition? The iCub robot’s Perspective.
cellphone
mouse
1
2
3
4
5
6
7
8
9
10
mug
1
2
3
4
5
6
7
8
9
10
50
25
remote
1
2
3
4
5
6
7
8
9
10
100 150
soap dispenser
1
2
3
4
5
6
7
8
9
10
50 100 150
1
2
3
4
5
6
7
8
9
10
50 100 150 200
50 100 150 200
50 100 150
Fig. 16 Frame-by-frame predictions of GoogLeNet on image sequences containing the Scale transformation, reported for 5
categories in different plots. The sequences of the 10 object instances belonging to each category are represented as matrix
rows (the temporal dimension, or frame index, is reported in the horizontal axis)). In each row, the frame-by-frame predictions
over the sequence are represented as vertical bars: W hite if the prediction is correct, Red if it is wrong (and Black if the
sequence is terminated).
cellphone
mouse
1
2
3
4
5
6
7
8
9
10
mug
1
2
3
4
5
6
7
8
9
10
50 100 150
remote
1
2
3
4
5
6
7
8
9
10
soap dispenser
1
2
3
4
5
6
7
8
9
10
50 100 150 200
1
2
3
4
5
6
7
8
9
10
50 100 150 200
50 100 150 200
50 100 150
Fig. 17 Same as Fig. 16, but on image sequences containing 2D Rotation.
CaffeNet
100
GoogLeNet
100
VGG-16
100
ResNet-50
100
CaffeNet
100
GoogLeNet
100
VGG-16
100
90
90
90
90
90
90
90
90
80
80
80
80
80
80
80
80
70
70
70
70
70
70
70
70
60
60
60
60
60
60
60
60
50
50
50
50
50
50
50
40
40
40
40
40
30
20
325
fc6 other
fc6 same
fc7 other
fc7 same
5852
m
30
20
15811 325
pool5 other
pool5 same
5852
m
30
20
15811 325
fc6 other
fc6 same
fc7 other
fc7 same
5852
m
accuracy (%)
accuracy (%)
100
40
30
20
15811 325
40
pool5 other
pool5 same
5852
15811
m
Fig. 18 Classification accuracy reported by training RLSC
over image representations extracted by the four considered
architectures. A “large” categorization experiment is performed (with ∼√ N = 95000 training examples) and m is
varied between N and 15000 (horizontal axis). Performance
on the same training day (Light colors) and on another one
(Dark colors) is reported.
instance, we considered the sequences of all 5 visual transformations and one camera. We considered only one day, among
30
20
50357
fc6 other
fc6 same
fc7 other
fc7 same
1319
m
30
20
2534 50357
pool5 other
pool5 same
1319
m
30
20
2534 50357
ResNet-50
50
fc6 other
fc6 same
fc7 other
fc7 same
1319
m
40
30
20
2534 50357
pool5 other
pool5 same
1319
2534
m
Fig. 19 Similar to Fig. 18 but on a “small” categorization
experiment
(∼ N = 2500 training examples), varying m be√
tween N and N .
the two available in iCWT, for training, and tested on the
instance left out from the training set, in both days. For the
large task, we considered all frames per sequence, leading to
∼ 6300 training examples per class, whereas for the small task
we subsampled ∼ 4 frames per sequence, leading to ∼ 170 examples per class (similarly to what we did in Sec. 4.2.1.
26
Fig. 19 and 18 reports the average accuracy respectively
for the small and large experiment. We performed the two
experiments using the representations from the four considered architectures (reported separately), using either fc6 layer
(Gray) or fc7 (Pink) for CaffeNet and VGG-16. In each case,
we increased logarithmically
the value of m (horizontal axis)
√
starting from N , N being the size of the training set. For
the small experiment we stopped at N , whereas for the large
experiment we stopped at values where we observed very little performance improvements with respect to much longer
training times. The performance on the same day on which
we trained is reported in Light (Gray/Pink), whereas the performance on the other day is reported in Dark colors.
We first observed that fc6 features consistently performed
better than fc7, hence we decided to use this layer when extracting representations from off-the-shelf CaffeNet or VGG16 on images of iCWT. Then, we also observed that relatively
small values of m could provide good accuracies and therefore
we chose to fixed m = min(15000, N ) in all experiments.
It is worth noting that we performed a similar empirical analysis to the one reported here in order to assess the
best output layer to use to extract visual representations from
the models of Sec. 5.2.1 and D.3 that were previously finetuned on subsets of iCWT. In that case, we observed that fc7
features were providing the best performance and therefore
we used this layer in our experiments in those sections. This
can be explained considering that “more specialized” features
can be extracted from models that have been adapted to our
setting, while “more general” features of off-the-shelf models
provide better generalization performance on iCWT.
C.2 Fine-tuning
In this section we provide details on how we performed the
fine-tuning procedure and on the choice of the two strategies,
adaptive and conservative, reported in Tab. 3.
C.2.1 General Protocol
The image preprocessing pipeline was similar to the one reported in Sec. B.1: once a square 256 × 256 region around
the object’s centroid was extracted, we either subtracted the
mean image (CaffeNet) or pixel (GoogLeNet) of the training
set.
When training, the rest of processing was performed within
Caffe, specifically extracting a random 227 × 227 crop (randomly mirrored horizontally) before passing it to the network. At test phase, the prediction over one image was taken
by extracting the central crop instead of a random one (see
Sec. B.1).
The training set was shuffled before starting the finetuning process since we observed that similarity of images
within a batch negatively affected convergence of the backpropagation algorithm. When fine-tuning a network, we evaluated the model’s performance on a validation set every epoch
and we finally chose the epoch achieving highest validation
accuracy.
We left the batch size to the values specified in Caffe
(and reported in Tab. 3). The number of iterations – or, equivalently, of epochs – was fixed empirically, observing that in
general the performance was saturating after 6 epochs for all
models but for conservative fine-tuning of CaffeNet, that involves the learning of all parameters of the 3 FC layers and
Giulia Pasquale et al.
therefore takes more time to converge. In this case, we fixed
the number of epochs to ∼ 36.
C.2.2 Hyperparameters Choice
In this section we report on the empirical analysis that we
performed in order to understand the effect and the relative
importance of the hyperparameters involved in fine-tuning
deep architectures as CaffeNet or GoogLeNet in our scenario.
While model selection is per se an open problem when
dealing with deep networks, in our setting this issue is even
more complicated by the fact that we do not have a fixed
reference task on which to optimize the training (as e.g.,
can be the ILSVRC), but we instead plan to span over wide
range of tasks, comprising small or large training sets. Indeed,
the study presented in this work aims at comparing models
trained on different categories, objects, transformations, and
so forth.
We considered therefore the same two experiments that
we used to perform parameter selection for the RLSC approach (the “small” and “large” categorization tasks described
in Sec. C.1) and we fine-tuned CaffeNet and GoogLeNet on
them by varying the values of multiple hyperparameters. In
the following, we report this analysis separately for the two
architectures.
CaffeNet. We considered the parameters appearing in Tab. 3
and varied the value of each of them in the following way:
Base LR: the starting learning rate of the layers that are initialized with the parameters of the off-the-shelf model.
We tried 10−3 , 5 ∗ 10−4 , 10−4 , 5 ∗ 10−5 , 10−5 , 10−6 , 0.
Learned FC Layers: which fully-connected (FC) layers are from
scratch and their specific starting LR. We tried to learn
(i) only f c8 with starting LR set to 10−2 , or (ii) including also f c7 and (iii) finally also f c6. As an empirical
rule, every time we included one more layer to learn from
scratch, we decreased the starting LR of these layers of a
factor of 10 (hence 10−3 in (ii) and 10−4 in (iii)).
Dropout %: percentage of dropout in FC layers. We tried
50% (default Caffe value) or 65%.
Solver: the algorithm used for the stochastic gradient descent. We used the SGD solver (Bottou 2012) in Caffe.
LR Decay Policy: the decay rate of the learning rates. We
tried either polynomial decay with exponent 0.5 or −3,
or step decay decreasing the LR of a factor of 10 every 2
epochs.
We tried all possible combinations of values. For the parameters that are missing here, we kept their value as specified
in Caffe reference models (see also the previous section).
We observed that the dropout percentage had a small
influence and left it to the default value. We also observed
that the polynomial decay with smaller slope was consistently
better (of ∼ 5 − 10% accuracy). The most critical parameters were instead the base LR and the numbers of FC layers
learned from scratch. In Fig. 20 we report as an example the
accuracy obtained respectively when learning only f c8 (Left)
or f c6, f c7 and f c8 (Right). In each case, we decreased the
base LR of all other layers from 10−3 to 0 (horizontal axis).
Performance is reported, as in Fig. 19 and 18, for both the
same day of training (Light Gray) and a different one (Dark
Gray). While we tried all values of base learning rate on the
“small” experiment (Continuous Line), for the “large” experiment we chose one small, medium and large value (Dots).
For interpreting the results, it can be useful to recall that
the strategy that learns from scratch the three FC layers
Are we Done with Object Recognition? The iCub robot’s Perspective.
fc6-7-8 with starting LR = 10 -4
Adam
100
80
80
80
80
60
60
60
60
40
20
40
small - same
small - other
large - same
large - other
0
1e-3 5e-4 1e-4 5e-5 1e-5 1e-6
base LR
accuracy (%)
100
100
accuracy (%)
fc8 with starting LR = 10 -2
27
20
0
40
20
0
1e-3 5e-4 1e-4 5e-5 1e-5 1e-6
0
base LR
Fig. 20 Classification accuracy provided by fine-tuning CaffeNet according to different strategies: either learning from
scratch only f c8 (Left) or f c6, f c7 and f c8 (Right). Performance is reported for both the same day of training (Light
Gray) and a different one (Dark Gray). We tried multiple values of base LR on a “small” training set (Continuous Line),
and only one small, medium and large value (Dots) on a
“large” one.
100
40
small - same
small - other
large - same
large - other
0
1e-3 5e-4 1e-4 5e-5 1e-5 1e-6
base LR
SGD
20
0
0
1e-3 5e-4 1e-4 5e-5 1e-5 1e-6
0
base LR
Fig. 21 Classification accuracy provided by fine-tuning
GoogLeNet according to different strategies: either using the
Adam solver (Left) or SGD (Right). The rest of the figure is
similar to Fig. 20
LR Decay Policy: when using SGD, we used polynomial decay with exponent 0.5 or −3; when using Adam, we maintained the learning rate constant.
As for CaffeNet, we tried all combinations and left the
not mentioned parameters to default values in Caffe.
We first observed that, differently from CaffeNet, overall for this architecture the impact of learning or not the
FC layers from scratch was very small, with the three tested
strategies behaving very similarly. We chose the last one (iii),
that was slightly better than the others. We also observed a
little benefit from using higher dropout percentages.
One little more critical aspect was instead the choice of
the solver. To this end, in Fig. 21 we report the accuracy
obtained respectively when using the Adam solver (Left) or
SGD (Right). In the latter case we applied the polynomial
decay with smaller slope, since we observed again that it was
consistently better. Performance is reported, as in Fig. 20, for
both the same day of training (Light Gray) and a different
one (Dark Gray). We varied again the base LR from 10−3 to 0
(horizontal axis), trying all values on the “small” experiment
(Continuous Line) and one small, medium and large value for
the “large” experiment (Dots).
It can be observed that, while the SGD solver is more
robust to different choices of base LR, Adam provides slightly
better accuracies for mid-range values of base LR, both for
the small and the large experiment. We therefore opted for
GoogLeNet. We performed for GoogLeNet a similar analthis latter solver.
ysis to the one reported for CaffeNet.
Then, similar observations as for the CaffeNet architecWe recall that the considered GoogLeNet architecture is
ture led us to conclude that learning slowly (with a smaller
composed of “main” branch, ending with one FC layer (called
base LR) can be preferable when we add more frames to the
loss3/classifier), and two identical “auxiliary” branches, ending with two FC layers (called loss1(2)/fc and loss1(2)/classifier). training set, since the model is less prone to overfit the training condition (i.e. with base LR equal to 0 we gain perforWe refer to (Szegedy et al 2015) for a detailed description of
mance on both days when passing from the Continuos Line
the model. By considering this structure, we then explored
to the corresponding Dot). Larger base LR can be preferable
the following parameters:
instead when the training set is smaller. As reported in Tab. 3
Base LR: varied as for CaffeNet.
therefore, also for GoogLeNet we identified an adaptive stratLearned FC Layers: we always learned loss3/classifier from
egy with base LR equal to 0 and a conservative strategy with
−2
scratch with starting LR equal to 10 ; regarding the
base LR equal to 10−5 .
auxiliary branches, we tried (i) to cut them out, (ii) to
learn also loss1(2)/classifier from scratch with starting
LR equal to 10−2 , or, finally, (iii) to learn from scratch
D More Experiments on Categorization
also loss1(2)/fc, setting the starting LR of these layers
and of loss1(2)/classifier to 10−3 .
In this section we complement the experiments discussed in
Dropout %: we tried either default Caffe values (40% for
the paper with further analysis. These results are reported to
loss3/classifier and 70% for loss1(2)/fc) or a higher pershow that the behavior observed for the specific conditions
centage: 60% for loss3/classifier and 80% for loss1(2)/fc.
Solver: we tried either SGD or Adam (Kingma and Ba 2015)
considered in the paper (e.g. number of object classes considsolvers in Caffe.
ered for the tasks) holds also in other settings.
(Right) has much more free parameters than the one learning only the last FC layer (Left). Indeed, we notice that this
latter strategy is more robust to the small-scale scenario, for
any base LR (Continuos Line in the range 40-60% to the Left
and 20-40% to the Right).
We also observe that, when we add frames to the training
set, both strategies benefit from learning slowly (i.e., when
we pass from the Continuos Line to the corresponding Dot
we gain more with smaller base LRs). Indeed, in the largescale experiment learning f c6-7-8 with small base LR achieves
the best performance. On the contrary, learning only f c8
with large base LR almost does not benefit from seeing more
frames and even overfits the day of training.
Therefore, since in our analysis we aimed at spanning the
size of the training set in a wide range, we chose two representative strategies providing best performance respectively
in the small and large-scale experiment: learning f c6-7-8 with
small (actually 0) base LR, that we call the conservative strategy, to account for our larger-scale settings, and learning only
f c8 with large base LR, that we call the adaptive strategy, to
account for smaller-scale settings.
28
Giulia Pasquale et al.
FINE-TUNING
RLSC
100
100
90
90
90
90
80
80
80
80
70
70
70
70
60
60
60
60
50
50
40
40
CaffeNet adapt
CaffeNet cons
GoogLeNet adapt
GoogLeNet cons
30
20
10
50
30
150
300
50
150
300
Fig. 22 Recognition accuracy vs # frames (number of
object views available during training). Same experiment as
Fig. 6 but executed by training only on 3 example instances
per category.
90
90
80
80
70
70
60
60
50
50
40
20
40
CaffeNet adapt
CaffeNet cons
GoogleNet adapt
GoogleNet cons
30
1
3
CaffeNet
GoogleNet
VGG-16
ResNet-50
30
5
7
20
1
3
#obj/cat
3
CaffeNet
GoogleNet
VGG-16
ResNet-50
30
5
7
20
1
3
5
7
#obj/cat
Fig. 24 Recognition accuracy vs # instances (number
of object instances available during training). Same experiment as Fig. 7 executed for a 5-class categorization problem.
D.3 Improving Invariance of CNNs on Categorization
RLSC
100
1
40
#obj/cat
#frames/seq
FINE-TUNING
50
CaffeNet adapt
CaffeNet cons
GoogleNet adapt
GoogleNet cons
30
20
20
10
100
50
40
CaffeNet
GoogLeNet
VGG-16
ResNet-50
#frames/seq
accuracy (%)
RLSC
100
accuracy (%)
accuracy (%)
FINE TUNING
100
5
7
#obj/cat
Fig. 23 Recognition accuracy vs # instances (number
of object instances available during training). Same experiment as Fig. 7 executed for a 10-class categorization problem.
D.1 Increasing the Training Set Size
Fig. 22 reports the results of the same experiment shown
in Fig. 6, but performed including only 3 object instances
per category in the training sets. For the sake of completeness, we complement the information provided in Sec. 4.2.1
by precising that we considered a 15-class categorization task
comprising the following categories: cellphone, mouse, coffee
mug, pencil case, perf ume, remote, ring binder, soap dispenser, sunglasses, f lower, wallet, glass, hairbrush, hair
clip, book.
These results further confirm that, in this setting, adding
frames without increasing semantic variability cannot be used
as a viable strategy to improve categorization accuracy, that
remains definitely lower than the performance achieved using
7 example instances (Fig. 6).
D.2 Increasing the Semantic Variability
Fig. 23 and 24 report the results of an experiment analogous
to the one shown in Fig. 7, but for respectively 10 and 5 object
categories rather than 15. As can be noticed, the observations
reported in Sec. 4.2.2 apply also here, namely the accuracy
increases remarkably as more example instances per category
are made available (without saturating), supporting the claim
that semantic variability is indeed critical for categorization
even in settings that involve few categories.
As mentioned at the end of Sec. 5.2.1, here we verify whether
the approach adopted to improve the invariance of CNNs in
identification settings may also be adopted to improve performance for categorization.
We considered the 15-class categorization task of Sec. 4.3,
Fig. 9, performed by sampling ∼ 20 frames per sequence and
applying RLSC on top of representations extracted by off-theshelf CNNs. In this experiment we compared the performance
on this task achieved by applying RLSC on top of representations extracted by CNNs which had been previously finetuned. To this end, as in Sec. 5.2.1, we evaluated the impact
of different fine-tuning strategies .
Specifically, we considered exactly the same fine-tuned
models as in Sec 5.2.1 for the first three strategies (namely,
iCWT id, iCWT cat, iCWT + ImNet) and, for the last
one (ImNet), we fine-tuned over the 15 ImageNet synsets
corresponding to the 15 categories that will be later involved
in the categorization task. We report results in Fig. 25 using
the same notation as in Fig. 12: as anticipated, none of these
strategies is beneficial when considering a categorization task.
Indeed, we do not achieve any improvement in performance
with respect to the off-the-shelf baseline.
E Generalization Across Days
In this section we test the robustness of the considered visual
recognition systems with respect to variations such as changes
of illumination, background, etc., which are neither semantic
nor geometric. In the current release of iCWT, for each object
instance we have acquired image sequences in two different
days, in order to naturally have images with these kind of
variations. For the sake of brevity, in the main paper we did
not report on experiments related to this aspect, which is
however relevant.
When training for visual recognition on a given day and
testing on a second one, the possibly small contextual variation in the scene can cause the system to experience a degradation of its recognition accuracy. In the experiments reported below, we report this loss as the difference between
the average classification accuracy observed when testing a
model on the same day it was trained on and on the other
available day in iCWT.
In particular, we consider the models trained for the categorization experiment in Sec. 4.2.1 and Sec. 4.2.2, and for
Are we Done with Object Recognition? The iCub robot’s Perspective.
accuracy (%)
100
80
off-the-shelf
iCWT cat
iCWT id
iCWT + ImNet
ImNet
100
test: 2D ROT
100
test: 3D ROT
100
test: SCALE
test: BKG
100
80
80
80
80
60
60
60
60
60
40
40
40
40
40
20
MIX 2DR 3DR SC BKG MIX
20
MIX 2DR 3DR SC BKG MIX
20
MIX 2DR 3DR SC BKG MIX
20
MIX 2DR 3DR SC BKG MIX
20
MIX 2DR 3DR SC BKG MIX
100
accuracy (%)
test: MIX
29
80
test: MIX
off-the-shelf
iCWT cat
iCWT id
iCWT + ImNet
ImNet
100
test: 2D ROT
100
test: 3D ROT
100
test: SCALE
test: BKG
100
80
80
80
80
60
60
60
60
60
40
40
40
40
40
20
MIX 2DR 3DR SC BKG MIX
20
MIX 2DR 3DR SC BKG MIX
20
MIX 2DR 3DR SC BKG MIX
20
MIX 2DR 3DR SC BKG MIX
20
MIX 2DR 3DR SC BKG MIX
Fig. 25 Same experimental setting as in Fig. 9, but using different image representations, provided by CaffeNet (Top, Orange)
or GoogLeNet (Bottom, Blue), network models fine-tuned according to different strategies (see Sec. D.3).
RLSC
FINE-TUNING
RLSC
0
0
0
-5
-5
-5
-5
-10
-10
-10
-10
-15
-15
-15
-15
-20
-25
-30
10
accuracy (%)
accuracy (%)
FINE TUNING
0
-20
CaffeNet adapt
CaffeNet cons
GoogLeNet adapt
GoogLeNet cons
50
-25
150
#frames/seq
300
-30
10
-20
CaffeNet
GoogLeNet
VGG-16
ResNet-50
50
-20
CaffeNet adapt
CaffeNet cons
GoogleNet adapt
GoogleNet cons
-25
150
300
#frames/seq
-30
1
3
5
#obj/cat
CaffeNet
GoogleNet
VGG-16
ResNet-50
-25
7
-30
1
3
5
7
#obj/cat
Fig. 26 Categorization accuracy vs # frames - Generalization across days. Drop in performance (difference
between test accuracy) observed when testing the models
trained for the categorization task reported in Sec. 4.2.1 on
the same day of training and on a different one.
Fig. 27 Categorization accuracy vs # instances Generalization across days. Drop in performance (difference between test accuracy) observed when testing the models
trained for the categorization task reported in Sec. 4.2.2 on
the same day of training and on a different one.
identification in Sec. 5.1. In this section, we test these models
on a day different from the one used for training (that is, we
took the images of the object instances selected for testing in
the original experiment, but from the second available day).
Categorization Fig. 26 and Fig. 27 report the loss in performance observed in the categorization setting, respectively
for the case where we tested on the relevance of the training
set size (see Fig. 6) and of semantic variability (see Fig. 7).
Surprisingly, in both settings all networks exhibit a substantial drop in performance, of the order of more than 5% accuracy and in some cases even around 20%. Critically, the
adaptive fine-tuning strategy, which appeared more effective
when tested on the same day, is also the strategy that leads
to the biggest drop. This implies that the CNN is overfitting
the variability the training day and is therefore less capable of
generalizing to other days. Interestingly however, when using
less aggressive strategies such as the conservative fine-tuning
or the feature extraction based classifier, modern networks
such as GoogleNet and ResNet-50 seem in general quite robust.
Identification The dramatic loss in performance observed
for the adaptive fine-tuning in categorization could be in principle worrisome if we recall that, in the identification setting,
such strategy was adopted in Sec. 5.2.1 to improve the invariance properties of CNNs on a preliminary dataset. Interestingly however, in Fig. 28 we notice that, in the identification setting, the adaptive fine-tuning strategy seems to be not
30
Giulia Pasquale et al.
accuracy (%)
FINE TUNING
RLSC
0
0
-5
-5
-10
-10
-15
-15
-20
-25
-30
10
-20
CaffeNet adapt
CaffeNet cons
GoogLeNet adapt
GoogLeNet cons
50
-25
150
300
-30
10
CaffeNet
GoogLeNet
VGG-16
ResNet-50
50
150
300
#frames/seq
Fig. 28 Identification accuracy vs # frames - Generalization across days. Drop in performance (difference
between test accuracy) observed when testing the models
trained for the identification task reported in Sec. 5.1 on the
same day of training and on a different one.
overfitting the training day and, albeit experiencing a drop
in performance when generalizing to another day, such loss
is small and on par with the other strategies considered in
this work. Indeed, it can be noticed that all methods exhibit
a performance drop around ∼ 5% in accuracy.
| 1cs.CV
|
Age of Information in a Network of Preemptive
Servers
Roy D. Yates
arXiv:1803.07993v1 [cs.IT] 21 Mar 2018
WINLAB, Department of Electrical and Computer Engineering
Rutgers University
[email protected]
Abstract—A source submits status updates to a network for
delivery to a destination monitor. Updates follow a route through
a series of network nodes. Each node is a last-come-first-served
queue supporting preemption in service. We characterize the
average age of information at the input and output of each node
in the route induced by the updates passing through. For Poisson
arrivals to a line network of preemptive memoryless servers, we
show that average age accumulates through successive network
nodes.
I. I NTRODUCTION
The need for timely knowledge of the system state in
remote monitoring applications has led to the development and
analysis of status update age metrics. When randomly arriving
updates are queued in a service facility, early work [1], [2]
showed that it is generally in the self-interest of a source to
limit its offered load. In particular, the updating must balance
between too infrequent updates and overly frequent updates
that induce queueing delays.
This observation has prompted the study of age in lossy systems that discard updates to avoid building queues. These include the last-come-first served (LCFS) queue with preemption
either in waiting or service [1], [3], [4] and packet management
mechanisms that restrict the the number of queued packets [5]
or discard waiting packets as they become stale [6]. However,
these contributions consider only single hop communication
systems.
Recently, there have been effort to examine age in multihop network settings [7]–[9]. In particular, this work is
closely related to the Last-Generated-First-Served (LGFS)
multihop networks studied in [7], [8]. When update transmission times over network links are exponentially distributed,
sample path arguments were used to show that a preemptive
Last-Generated, First-Served (LGFS) policy results in smaller
age processes at all nodes of the network than any other causal
policy. However, these structural results do not facilitate the
explicit calculation of age.
In this work, we consider preemptive LCFS servers, a
special case of the LGFS discipline, in the multihop line network shown in Figure 1. For Poisson arrivals and memoryless
preemptive servers, we use a stochastic hybrid systems (SHS)
model of the age processes in the network to derive a simple
expression for the average age at each node.
This work was supported by NSF award CNS-1422988 and will be
presented at the 2018 IEEE Infocom Age of Information workshop.
This work was motivated by a simple question regarding
the line network with n = 2 nodes and service rates µ1 and
µ2 . The network traffic depends qualitatively on these service
rates µ1 and µ2 . For example if µ1 λ ≤ µ2 , packet dropping
will occur primarily at node 1 but traffic will pass quickly, and
with relatively little dropping, through node 2, On the other
hand, with µ1 λ ≥ µ2 there will be substantial dropping at
node 2, but little dropping at node 1. From the perspective
of average age at the monitor, is it better for dropping to
occur earlier (µ1 < µ2 ) or later (µ1 > µ2 ) in the network? A
reasonable hypothesis is that µ1 < µ2 wastes the resources of
the faster downstream server. In fact, we will see in Theorem 2
that average age at the monitor is insensitive to the ordering
of the servers.
In this work, the network model is described in Section II.
This is followed in Section III with a summary of results on
stochastic hybrid systems (SHS) for age analysis from [10] that
will be the basis for our age analysis. In Section IV, we apply
SHS to the 2-node tandem queue. We use a 4-state model that
describes the occupancy of each server in the network. While
this analysis may be instructive, it is shown in Section V
that the preemptive service facilitates an analysis using a
“fake updates” technique from [10] that reduces the discrete
state space to just one state. Some simulation experiments are
provided in Section VI and conclusions appear in Section VII.
II. S YSTEM M ODEL
We model the updating process as a source that submits
update packets as a rate λ Poisson process to a network.
As depicted in Figure 1, updates follow a route through n
nodes in a network to a monitor. At node i, an update has
an exponential (µi ) service time, independent of the arrival
process and service times at other nodes. However, we forgo
queueing in this network; each node is a ·/M/1/1 preemptive
server. Upon arrival at a server, an update immediately goes
into service and any update currently in service is preempted
and discarded.
This is a useful model when the time an update spends in
the head-of-line position in the network interface is dominated
by waiting for transmission. For example, in a congested
wireless network, the service time would be dominated by the
MAC access delay and the transmission time of the packet
is negligible. In this case, it would be feasible and desirable
Source
λ
µ1
µ2
µn
Monitor
Fig. 1. The n-node line network model. Node i is a rate µi ·/M/1/1 preemptive
server.
for the head-of-line update packet, i.e. the nominal update in
service, to be preempted by a more recent arrival.
Starting at time t = 0, the source submits status updates
at successive times U1 , U2 . . . such that update i submitted at
time Ui is delivered at time Ui + Ti . The Ti are dependent
random variables. Moreover, because of preemption, the line
network is lossy and Ti = ∞ for those updates that are
discarded. At time t, the most recent received update is
time-stamped U (t) = max{Ui |Ui + Ti ≤ t} and thus the
status update age, which we refer to as simply the age, is
∆(t) = t − U (t). The system performance is given by the
average status update age ∆ = limt→∞ E[∆(t)].
For the n-node network in Figure 1, we denote the average
age at the monitor by ∆(λ, µ1 , · · · , µn ). In the case of n = 1
node, we have a simple M/M/1/1 preemptive queue. This has
also been called the Last-Come First-Served with preemption
in service (LCFS-S) queue [10]. Using a graphical approach,
it was shown [3] that the time average of ∆(t) approaches
∆(λ, µ1 ) =
1
1
+
.
λ µ1
(1)
In this prior work, an end-to-end network is modeled as a
single service facility. The key analytical steps involve the the
system time T and the interarrival time Y of delivered packets
and the challenge is the computation of the correlation E[T Y ].
However, when the network has n ≥ 2 nodes, the graphical
method fails because the queueing of updates in the network
is non-trivial for several reasons:
• The line network is lossy as updates are discarded when
they are preempted. Even the calculation of the effective
arrival rate of updates at the monitor is challenging.
• The departure process at each node, even node 1, is not
memoryless.
• The arrivals at node 2 and subsequent nodes are not
fresh; instead they are aged by their passage through prior
nodes. The age of a packet arriving at node i may be
correlated with its service times at preceding nodes. That
is, the interarrival time of an update may be correlated
with its age.
• Updates that reach the monitor may be subject to a
survivor bias as each was lucky enough in its service
times to avoid being preempted.
III. S TOCHASTIC H YBRID S YSTEMS FOR AO I
Because of the complexity of the lossy queueing process in
the line network, we take a different non-graphical approach to
average age analysis. Following [10], we model the system as
a stochastic hybrid system (SHS) with hybrid state (q(t), x(t)).
In the SHS model, q(t) ∈ Q is discrete and typically describes
the Markov state of the queueing system while the row vector
x(t) ∈ Rn+1 is continuous and captures the evolution of a
collection of age-related processes.
In general, SHS is a powerful modeling framework with
many variations [11], [12]. In this work, we use a simplified
form of SHS for AoI analysis introduced in [10] in which x(t)
is a piecewise linear process. In the interest of completeness,
we now summarize the basics of this simplified SHS; further
details can be found in [10] and references therein.
In the graphical representation of the Markov chain q(t),
each state q ∈ Q is a node and each transition l ∈ L is
a directed edge (ql , ql0 ) with transition rate λ(l) δql ,q(t) . Note
that the Kronecker delta function δql ,q ensures that transition
l occurs only in state ql . For each state q, we define
L0q = {l ∈ L : ql0 = q},
Lq = {l ∈ L : ql = q}
(2)
as the respective sets of incoming and outgoing transitions.
For each transition l, there is transition reset mapping that can
induce a discontinuous jumps in the continuous state x(t).
For AoI analysis, we employ a linear mapping of the form
x0 = xAl . That is, transition l causes the system to jump
to discrete state ql0 and resets the continuous state from x to
x0 = xAl . Moreover, in each discrete state q(t) = q, the
continuous state evolves according to ẋ(t) = bq .
In using a piecewise linear SHS for AoI, the elements of
bq will be binary. We will see that the ones in bq correspond
to certain relevant components of x(t) that grow at unit rate
in state q while the zeros mark components of x(t) that are
irrelevant in state q to the age process and need not be tracked.
For tracking of the age process, the transition reset maps
(n+1)×(n+1)
are binary: Al ∈ {0, 1}
. The linear mappings Al
will depend on the specific network system and the indexing
scheme for updates in the system.
The transition rates λ(l) correspond to the transition rates
associated with the continuous-time Markov chain for the
discrete state q(t); but there are some differences. Unlike
an ordinary continuous-time Markov chain, the SHS may
include self-transitions in which the discrete state is unchanged
because a reset occurs in the continuous state. Furthermore,
for a given pair of states i, j ∈ Q, there may be multiple
transitions l and l0 in which the discrete state jumps from i to
j but the transition maps Al and Al0 are different.
It will be sufficient for average age analysis to define for
all q̂ ∈ Q = {0, 1, . . . , m},
πq̂ (t) = E δq̂,q(t) ,
(3a)
vq̂j (t) = E xj (t)δq̂,q(t) , j ∈ {0, . . . , n},
(3b)
and the vector functions
vq̂ (t) = [vq̂0 (t), . . . , vq̂n (t)] = E x(t)δq̂,q(t) .
(3c)
We note that πq̂ (t) denotes the discrete Markov state probabilities, i.e.,
πq̂ (t) = E δq̂,q(t) = P[q(t) = q̂].
(4)
Similarly, vq̂ (t) measures correlation between the age process
x(t) and the occupancy of the discrete state q(t).
A fundamental assumption for age analysis is that the
Markov chain q(t) is ergodic; otherwise, time-average age
analysis makes little sense. Under this assumption, the state
probability vector π(t) = [π0 (t) · · · πm (t)] always converges
to the unique stationary vector π̄ = [π̄0 · · · π̄m ] satisfying
X
X
λ(l) π̄ql , q̄ ∈ Q,
(5a)
π̄q̄
λ(l) =
1
(10)
1
X
π̄q̄ = 1.
5
0
(00)
When π(t) = π̄, it is shown in [10] that v(t) =
[v0 (t) · · · vm (t)] obeys the system of first order differential
equations such that for all q̄ ∈ Q,
X
X
λ(l) . (6)
λ(l) vql (t)Al − vq̄ (t)
v̇q̄ (t) = bq̄ π̄q̄ +
2
(01)
4
(5b)
q̄∈Q
6
3
l∈L0q̄
l∈Lq̄
2
Fig. 2. The SHS Markov chain for the line network with n = 2 nodes. The
transition rates and transition/reset maps for links l = 1, . . . , 8 are shown in
Table I.
ql → ql0
λ(l)
xAl
1
0→1
λ
[x0 0 0 ]
Depending on the reset maps Al , the differential equation (6)
may or may not be stable.
However,
when (6) is stable, v̇(t) →
0 and each vq̄ (t) = E x(t)δq̄,q(t) converges to a limit v̄q̄ as
t → ∞. In this case, it follows that
2
1→1
λ
[x0 0 0 ]
3
1→2
µ1
[x0 0 x1 ]
4
2→0
µ2
[x2 0 0 ]
E[x] ≡ lim E[x(t)]
t→∞
X
X
E x(t)δq̄,q(t) =
= lim
v̄q̄ .
5
2→3
λ
[x0 0 x2 ]
6
3→1
µ2
[x2 x1 0 ]
7
3→2
µ1
[x0 0 x1 ]
8
3→3
λ
[x0 0 x2 ]
t→∞
l∈Lq̄
q̄∈Q
(7)
q̄∈Q
Following the convention in [10] that x0 (t) = ∆(t) is the age
at the monitor, the
P average age of the process of interest is then
∆ = E[x0 ] = q̄∈Q v̄q̄0 . The following theorem provides a
simple way to calculate the average age in an ergodic queueing
system.
Theorem 1: [10, Theorem 4] If the discrete-state Markov
chain q(t) is ergodic with stationary distribution π̄ and we can
find a non-negative solution v̄ = [v̄0 · · · v̄m ] such that
X
X
v̄q̄
λ(l) = bq̄ π̄q̄ +
λ(l) v̄ql Al ,
q̄ ∈ Q,
(8a)
l∈Lq̄
l∈L0q̄
then the differential equation (6) is stable and the average age
of the AoI SHS is given by
X
∆=
v̄q̄0 .
(8b)
q̄∈Q
IV. AGE IN THE TANDEM Q UEUE WITH P REEMPTION
We now use Theorem 1 to evaluate the age ∆(λ, µ1 , µ2 )
at the monitor in the line network of Figure 1 with n = 2
intermediate nodes. This example will demonstrate how to use
Theorem 1 in a straightforward way to evaluate average age
in a network.
In the 2-node network, the set of discrete states is Q =
{00, 10, 01, 11} such that for q1 q2 ∈ Q, qi = 1 indicates that
node i is serving an update packet. It will also be convenient to
refer to the states Q = {0, 1, 2, 3} such that q = q1 +2q2 ∈ Q.
The continuous state is x(t) = [x0 (t), x1 (t), x2 (t)] such
that x0 (t) is the age at the monitor, and, when there is an
update in service at node i, xi (t) is the age of that update.
When node i is idle, xi (t) is irrelevant and we set xi (t) = 0.
8
7
l
l∈L0q̄
3
(11)
h 1A0 l0 i
v̄ql Al
000
[v̄00 0
0 ]
000
[v̄10 0
0 ]
001
[v̄10 0 v̄11 ]
000
[v̄22 0
000
[v̄20 0 v̄22 ]
010
[v̄32 v̄31 0 ]
001
[v̄30 0 v̄31 ]
000
001
[v̄30 0 v̄32 ]
h 01 00 00 i
h 01 00 00 i
h 00 00 00 i
h 11 00 00 i
h 00 00 10 i
h 11 00 00 i
h 01 00 00 i
0 ]
TABLE I
TABLE OF TRANSITIONS FOR THE M ARKOV CHAIN IN F IGURE 2.
In particular, in any state q in which node i is idle, we hold
xi (t) = 0 while in that state. Otherwise, if node i is serving
an update in a state q, then xi (t) increases at unit rate in that
state. It follows that in state q we set
[1 0 0] q = 0,
[1 1 0] q = 1,
(9)
bq =
[1 0 1] q = 2,
[1 1 1] q = 3.
It follows from (9) that in each q, the age at the monitor, x0 (t),
grows at unit rate. On the other hand, in states q ∈ {1, 3}, node
1 is serving an update whose age x1 (t) is growing at unit rate.
Similarly, in states q ∈ 2, 3, node 2 is serving an update whose
age x2 (t) is growing at unit rate.
The Markov chain for the occupancy of the network is
shown in Figure 2. The edges are labeled l ∈ {1, 2, . . . , 8}. For
each transition l, Table I lists the state transition pair ql → ql0 ,
the transition rate λ(l) , the transition mapping x0 = xAl , the
matrix Al and, to facilitate using Theorem 1, vql Al .
We now describe the transitions l. We note that the age x0 (t)
at the monitor changes only in those transitions in which an
update completes service at node 2 and is delivered to the
monitor. Corresponding to Table I, the transitions are:
1) In an idle network, a fresh update arrives at node 1.
2)
3)
4)
5)
6)
7)
8)
The age x00 = x0 at the monitor is unchanged, x01 = 0
because the arrival is fresh, and x02 = 0 because x2 is
irrelevant in state 1.
In state 1, a fresh update arrives and preempts the update
in service at node 1. The age x00 = x0 at the monitor
is unchanged, x01 = 0 because the arrival is fresh, and
x02 = 0 because x2 is irrelevant in state 1.
In state 1, the update at node 1 completes service and
moves to node 2. The age x00 = x0 at the monitor is
unchanged, x01 = 0 because x1 becomes irrelevant in
state 2 and x02 = x1 because the update now at node 2
is the update that was previously at node 1.
In state 2, the update at node 2 completes service and is
delivered to the monitor. The system moves to state 0.
The age at the monitor becomes x00 = x2 . In addition,
x01 = x02 = 0 since x1 and x2 are irrelevant in state 0.
The arrival of a fresh update at node 1 induces the 2 → 3
state transition. The age x00 = x0 is unchanged. At node
1, x01 = 0 because the update is fresh. At node 2, the
age x02 = x2 is unchanged.
The update at node 2 completes service and is delivered
to the monitor. The system moves to state 1. The age
at the monitor becomes x00 = x2 . In addition, x01 = x1
is unchanged because the update at node 1 remains in
place and x02 = 0 since x2 are irrelevant in state 1.
The update at node 1 completes service and moves to
node 2, preempting the update that had been in service
at node 2. The age x00 = x0 is unchanged, x01 = 0
because x1 becomes irrelevant in state 2, and x02 = x1
because the update now at node 2 is the update that was
previously at node 1.
A fresh update arrives at node 1, preempting the update
that had been in service at node 1. The age x00 = x0 is
unchanged, x01 = 0 because the new update at node 1 is
fresh, and x02 = x2 is unchanged because the update at
node 2 stays in place.
To use Theorem 1 to find the average age, we first find the
stationary probabilities π̄. It is straightforward to verify that
λ
µ1 µ2
,
π̄2 =
π̄0 ,
(10a)
(µ1 + λ)(µ2 + λ)
µ2
λ µ1 + µ2 + λ
λ2
π̄1 =
π̄0 , π̄3 =
π̄0 . (10b)
µ1
µ1 + µ2
µ2 (µ1 + µ2 )
π̄0 =
Second, we need to find a non-negative v̄ satisfying (8a).
Defining αi = λ + µi for i = 1, 2, and α3 = λ + µ1 + µ2 , it
follows from (8a) and Table I that
v̄0 λ = π̄0 b0 + µ2 [v̄22 0 0],
(11a)
v̄1 α1 = π̄1 b1 + λ[v̄00 0 0]
+ λ[v̄10 0 0] + µ2 [v̄32 v̄31 0],
(11b)
v̄2 α2 = π̄2 b2 + µ1 [v̄10 0 v̄11 ] + µ1 [v̄30 0 v̄31 ],
(11c)
v̄3 α3 = π̄3 b3 + λ[v̄20 0 v̄22 ] + λ[v̄30 0 v̄32 ].
(11d)
Since each v̄q has three components, there are twelve equations in (11). However, because x1 and x2 are irrelevant in
state q = 0, it follows from (11a) that v̄01 = v̄02 = 0.
Similarly, because x2 is irrelevant in state q = 1, and x1 is
irrelevant in state q = 2, it follows from (11b) and (11c) that
v̄12 = v̄21 = 0. Thus we are left with the eight equations
v̄00 λ = π̄0 + µ2 v̄22 ,
v̄10 µ1 = π̄1 + λv̄00 + µ2 v̄32 ,
(12a)
(12b)
v̄11 (λ + µ1 ) = π̄1 + µ2 v̄31 ,
(12c)
v̄20 (λ + µ2 ) = π̄2 + µ1 v̄10 + µ1 v̄30 ,
(12d)
v̄22 (λ + µ2 ) = π̄2 + µ1 v̄11 + µ1 v̄31 ,
(12e)
v̄30 (µ1 + µ2 ) = π̄3 + λv̄20 ,
v̄31 (λ + µ1 + µ2 ) = π̄3 ,
v̄32 (µ1 + µ2 ) = π̄3 + λv̄22 .
(12f)
(12g)
(12h)
With some algebra, it follows from (12) that
∆(λ, µ1 , µ2 ) = v̄00 + v̄10 + v̄20 + v̄30
3
1
1
1 X
=
+
+
π̄q
λ µ1
µ2 q=0
=
1
1
1
+
+
.
λ µ1
µ2
(13)
When we have 2 preemptive servers, (13) verifies that the
average age at the monitor is indeed insensitive to the ordering
of the servers. Moreover, in comparing (1) and (13), one can
see the simple pattern that the ith node contributes 1/µi to
the age at the monitor. To verify this observation for an nnode network, however, the SHS method we employed for
n = 2 nodes tracks the state of each node in the network and
thus requires the solution of (n + 1)2n equations. In the next
section, we show how to extend this simple pattern to n nodes
using an SHS that generates only n + 1 equations.
V. T HE n N ODE L INE N ETWORK : FAKE U PDATES
A unusual feature of the LCFS-S server at each node is that
tracking the idle/busy state of a node is not actually essential
because an arrival goes immediately into service whether or
not an update is in service. When node i is busy, xi (t) encodes
the age of the update in service. If that update completes
service at time t0 , it enters service at node i+1 and x0i+1 = xi ,
whether or not it preempts any update that may have been in
service at node i + 1. To avoid tracking the idle/busy state
at node i, when an update departs node i, we create a fake
update at node i, with the same timestamp (and age xi (t)) as
the update that just departed. If a new update from node i − 1
arrives at node i, it preempts the fake update and the fake
update causes no delay to an arriving real update. If the fake
update does depart node i, it will go into service at node i + 1,
but it will have the same age as the update it will preempt at
node i + 1.
In [10], fake updates are introduced in an SHS derivation
of the average age of an LCFS-S queue in which the SHS has
a single discrete state. In this section, we show how to extend
the fake updates approach to prove the following theorem.
Theorem 2: With Poisson arrivals of rate λ to an n-node
line network of LCFS-S servers with service rates µ1 , . . . , µn ,
the average age at the monitor is
n−1
n
1 X 1
.
∆(λ, µ1 , . . . , µn ) = +
λ i=1 µi
•
•
Transition l = 0 marks the arrival of a fresh update at
node 1. In the continuous state x, we set x01 = 0 but all
other components of x are unchanged.
In a transition l ∈ {1, . . . , n − 1} an update departs node
l and arrives at node l + 1. At node l + 1, x0l+1 = xl . At
node l, x0l = xl because a fake update begins service.
In transition n, an update departs node n and is delivered
to the monitor. The age at the monitor is reset to x00 = xn .
Each continuous-state age component xi (t), whether representing a real or fake update, ages at unit rate and thus
b0 = [1 1 · · · 1]. In the following, our notation will be
simplified by defining µ0 = λ. Applying Theorem 1 to the
transitions given in Table II, we obtain
v̄0
n
X
µi = b0 + µ0 [v̄00 0 v̄02 v̄03 · · · v̄0n ]
i=0
+ µ1 [v̄00 v̄01 v̄01 v̄03 · · · v̄0n ]
+ µ2 [v̄00 v̄01 v̄02 v̄02 · · · v̄0n ]
..
.
+ µn [v̄0n v̄01 v̄02 v̄03 · · · v̄0n ].
(14)
It then follows from (14) that
v̄01
n
X
µi = 1 + v̄01
i=0
n
X
µi ,
(15)
i=1
implying v̄01 = 1/λ. In addition, for k = 2, 3, . . . , n,
v̄0k
n
X
µi = 1 + v̄0k
i=0
k−2
X
µi + µk−1 v̄0,k−1 + v̄0k
i=0
n
X
µi . (16)
i=k
It follows from (16) that v̄0k = 1/µk−1 + v̄0,k−1 , implying
k−1
v̄0k
1 X 1
,
= +
λ i=1 µi
k = 1, . . . , n.
(17)
Finally, for component v̄00 of v0 , (14) implies
v̄00
n
X
i=0
µi = 1 + v̄00
n−1
X
µi + v̄0n µn .
(18)
i=0
This implies v̄00 = 1/µn + v̄0n . The claim then follows from
(17) and Theorem 1.
1
0
..
Proof: Using fake updates, the Markov chain has the
singleton state space Q = {0} with the trivial stationary
distribution π̄0 = 1. As shown in Figure 3, there is a 0 → 0
transition for each link l. The transitions l ∈ {0, 1, . . . , n} are
shown in Table II. In the table, we omit the ql → ql0 entry
since it is 0 → 0 for each l. Note that
•
0
n
. ···
2
3
Fig. 3. The SHS Markov chain for the line network with n nodes. The
transition rates and transition/reset maps for links l = 0, . . . , n are shown in
Table II.
λ(l)
xAl
λ [x0 0 x2 x3
µ1 [x0 x1 x1 x3
µ2 [x0 x1 x2 x2
..
..
.
.
n µn [xn x1 x2 x3
l
0
1
2
v̄ql Al
· · · xn ] [v̄00 0 v̄02 v̄03 · · · v̄0n ]
· · · xn ] [v̄00 v̄01 v̄01 v̄03 · · · v̄0n ]
· · · xn ] [v̄00 v̄01 v̄02 v̄02 · · · v̄0n ]
..
.
· · · xn ] [v̄0n v̄01 v̄02 v̄03 · · · v̄0n ]
TABLE II
TABLE OF TRANSITIONS FOR THE M ARKOV CHAIN IN F IGURE 3.
VI. N UMERICAL E XPERIMENTS
Because Theorem 2 provides such a simple characterization
of the average age in the preemptive line network, here we
examine age sample paths in order to get a better sense of how
age fluctuates across a sequence of preemptive servers. Our
focus is a line network with n = 3 nodes. In these experiments,
let ∆i (t) denote the instantaneous age at the output of node i
at time t.
We start with Figure 4 which depicts representative sample
paths of ∆i (t) for i = 1, 2, 3. These sample paths are the result
of 50 updates of a rate 1 Poisson process arriving at node
1 of a three-node network with service rates [µ1 µ2 µ3 ] =
[1 0.5 0.25]. These sample paths demonstrate that while the
average age grows predictably, preemption thins the update
arrival process at successive nodes. Because of this thinning,
higher average age at successive nodes is a consequence of
age growing in between less-frequent updates. We also see
some evidence of survivor bias; occasionally some updates
pass quickly through the network enabling the age at the
monitor to be reset to a relatively low value.
In Figure 5, we examine the reverse ordering of the servers;
[µ1 µ2 µ3 ] = [0.25 0.5 1] so that downstream servers become
progressively faster. In sample paths for this system, most
updates are dropped at the first node and the age processes at
nodes 2 and 3 mimic the age at node 1, but with some small
additional lag. In this sample path, there is no evidence of
survivor bias. However, it is apparent that in both Figures 4
and 5 there is a general trend that thinning of the updates
induces larger age fluctuations at nodes further down the line.
To see the convergence of the average age, we now examine
the running average age
Z t
˜ i (t) = 1
∆i (τ ) dτ
(19)
∆
t 0
15
20
∆ 1(t)
∆ 2(t)
15
∆ 3(t)
10
Age
Age
25
10
5
5
0
0
0
20
40
60
t
0
50
100
150
200
250
t
Fig. 4. Fifty arrivals to the three node line network with arrival rate λ = 1
and service rates µ1 = 1, µ2 = 0.5 and µ3 = 0.25.
˜ i (t), i = 1, 2, 3, with
Fig. 6. Sample paths of the running sample average ∆
arrival rate λ = 1 and service rates µ1 = 1, µ2 = 0.5 and µ3 = 0.25. The
bundles of sample paths converging to ages 2, 4, and 8 support Theorem 2.
15
∆ 1(t)
∆ 2(t)
10
Moreover, when each server in the network is a memoryless
preemptive server, the method of fake updates can greatly
simplify the age computation. We caution however that fake
updates appear to be useful only when servers support preemption in service. For example, when preemptions occur in
waiting, the discrete state must track whether a server is busy;
i.e. it must distinguish between real updates and fake updates.
In such cases, the system state will grow exponentially with
the number of network nodes.
Age
∆ 3(t)
5
0
0
10
20
30
40
50
t
Fig. 5. Fifty arrivals to the three node line network with arrival rate λ = 1
and service rates µ1 = 0.25, µ2 = 0.5 and µ3 = 1.
˜ i (t)
at each node i. In Figure 6, we plot 10 sample paths of ∆
for each node i. In each sample path, 200 updates arrive at
node 1 as a rate λ = 1 Poisson process. The nodes have service
rates µ1 = 1, µ2 = 0.5 and µ3 = 0.25. For each sample path,
˜ i (t) through the delivery time of the final received
we plot ∆
update at the monitor. From Theorem 2, we expect to see
i
2, i = 1,
X
1
1
˜
lim ∆i (t) = +
= 4 i = 2,
(20)
t→∞
λ
µk
k=1
8 i = 3.
In Figure 6, we see that the bundles of running average sample
paths clustered around ages 2, 4 and 8 are consistent with this
conclusion. Moreover, the increasing variation and apparently
slower convergence of the running average age in successive
servers are also consistent the larger fluctuations in age at
successive servers in the sample paths of Figures 4 and 5.
VII. C ONCLUSIONS
In this work, we use the SHS approach to analyze age in an
n-node line network of exponential servers. From the two node
example in Section IV, it should be apparent that SHS can be
employed to analyze a variety of other queues and simple
networks described by finite continuous-time Markov chains.
R EFERENCES
[1] S. Kaul, R. Yates, and M. Gruteser. Real-time status: How often should
one update? In Proc. IEEE INFOCOM Mini Conference, 2012.
[2] C. Kam, S. Kompella, and A. Ephremides. Effect of message transmission diversity on status age. In Proc. IEEE Int’l. Symp. Info. Theory,
pages 2411–2415, June 2014.
[3] S. Kaul, R. Yates, and M. Gruteser. Status updates through queues. In
Conf. on Information Sciences and Systems (CISS), March 2012.
[4] R. Yates and S. Kaul. Real-time status updating: Multiple sources. In
Proc. IEEE Int’l. Symp. Info. Theory, July 2012.
[5] M. Costa, M. Codreanu, and A. Ephremides. Age of information with
packet management. In Proc. IEEE Int’l. Symp. Info. Theory, pages
1583–1587, June 2014.
[6] C. Kam, S. Kompella, G. D. Nguyen, J.E. Wieselthier, and
A. Ephremides. Age of information with a packet deadline. In
Proc. IEEE Int’l. Symp. Info. Theory, pages 2564–2568, 2016.
[7] A. M. Bedewy, Y. Sun, and N. B. Shroff. Age-optimal information
updates in multihop networks. In Proc. IEEE Int’l. Symp. Info. Theory,
pages 576–580, June 2017.
[8] A. M. Bedewy, Y. Sun, and N. B. Shroff. The age of information in
multihop networks. CoRR, abs/1712.10061, 2017.
[9] R. Talak, S. Karaman, and E. Modiano. Minimizing age-of-information
in multi-hop wireless networks. In 55th Annual Allerton Conference
on Communication, Control, and Computing (Allerton), pages 486–493,
Oct 2017.
[10] R. D. Yates and S. K. Kaul. The age of information: Real-time status
updating by multiple sources. CoRR, abs/1608.08622, 2016. Submitted
to the IEEE Transactions on Information Theory.
[11] J.P. Hespanha. Modelling and analysis of stochastic hybrid systems. IEE
Proceedings-Control Theory and Applications, 153(5):520–535, 2006.
[12] A. R. Teel, A. Subbaraman, and A. Sferlazza. Stability analysis for
stochastic hybrid systems: A survey. Automatica, 50(10):2435–2456,
2014.
| 7cs.IT
|
Bulletin of the IEEE Technical Committee on Learning Technology, Volume 18, Number 2/3, 2016
6
Decision Trees for Helpdesk Advisor Graphs
S. Gezerlis and D. Kalles
Abstract—We use decision trees to build a helpdesk agent
reference network to facilitate the on-the-job advising of junior
or less experienced staff on how to better address
telecommunication customer fault reports. Such reports generate
field measurements and remote measurements which, when
coupled with location data and client attributes, and fused with
organization-level statistics, can produce models of how support
should be provided. Beyond decision support, these models can
help identify staff who can act as advisors, based on the quality,
consistency and predictability of dealing with complex
troubleshooting reports. Advisor staff models are then used to
guide less experienced staff in their decision making; thus, we
advocate the deployment of a simple mechanism which exploits
the availability of staff with a sound track record at the helpdesk
to act as dormant tutors.
Index Terms— customer relationship management; decision
trees; knowledge flow graph
C
I. INTRODUCTION
ustomer satisfaction is a key factor in making clients loyal
to a service provider [1]. In telecommunications, this is
closely linked to the time it takes to fix a problem and the way
a fault request is handled. Long-term customers are considered
more loyal to a brand and usually cost less to serve [2][3].
Big telecommunication organizations have adopted
Standard Operating Procedures (SOPs) and Key Performance
Indicators (KPIs) to help guide their policies in servicing
customer fault complaints.
SOPs are guidelines to be followed by engineers, to help
them decide how to solve a particular problem. SOPs are
usually combinations of rules and decision trees [4][5] and
constitute an integral part of staff training.
KPIs are numeric indices which attempt to capture aspects
of service quality (as perceived by the customer or by the
organization) and are usually expressed in terms of average
request handling speed, percentage of problems solved
remotely, or repeated complaint rates. Though a KPI may be
straightforward to define and to measure, relating it to soft
attributes of service provision is an elusive task. For example,
being able to conclude a job on time is an obvious target for a
maintenance engineer, and so is the ability to draw as few
organizational resources as possible for any particular task.
Though relating both to cost can also be straightforward,
Manuscript received June 25, 2016 (revised September 11, 2016).
Spyros Gezerlis was with Hellenic Telecommunications Agency, Greece.
He is now with Deutsche Telekom AG, Germany (e-mail:
[email protected], [email protected]).
Dimitris Kalles is with the Hellenic Open University, Greece (phone: +30
6932 231151; e-mail: [email protected]).
linking them to measures of customer satisfaction can be a
difficult exercise. For this reason, composite quality indices
are notoriously difficult to define and are subject to
painstaking secrecy and to simplifications [6].
SOPs and KPIs are usually deployed at two levels of
support: Remote (phone) Support, which handles all incoming
complaints and calls, and On-Site Service with field
technicians, who work on the landline network. The decision
to deploy On-site Support while already providing Remote
Support is sometimes a function of the Remote Support
technician's experience, workload, mental situation and agility
and either approach has to be judged based on the average cost
each of them incurs on the company, all of which are also
functions of time.
The element of cost is captured by the Operational Expense
(OPEX) formulation, where one assumes that a complaint
remotely resolved costs less, is resolved faster and still keeps
customer experience at a satisfactory level. With OPEXOS
standing for OPEX for On Site Support at 1, OPEXRS for
Remote resolution Service is OPEXOS/n , with typical values
of n > 10 (details are business confidential).
The complexity of making support-related decisions on
time, with pressure and with acceptable consistency with
organizational policies on customer support and cost
containment, cannot be overstated. As training is a key to
improvement, this paper puts forward an unconventional
approach, based on the on-the-fly identification of key
personnel who can temporary act as advisors to their peers in a
discreet, non-intrusive fashion.
The contribution of this article is two-fold. First, we use
decision trees to identify inconsistencies in the cost incurred
when dealing with customer complaints; we interpret such
inconsistencies as a measure of the ability of the company to
deliver an as-uniform-as-possible customer experience in
complaint troubleshooting. Then, we use these inconsistencies
to rank helpdesk agents who troubleshoot customer
complaints; this generates a knowledge hierarchy that is
dynamically updated based on identifying personnel who can
advise or who need to be advised.
The work serves as a bridge between the formality of logic
that is inherent in expressing SOPs and the inherent vagueness
as regards what is the "best" way to resolve a particular
customer complaint. It also advances the state of the practice
in the customer relationship management (CRM) genre [7],
where what little has been done until recently [8][9][10] as
regards customer complaints, mainly focuses on knowledge
about the value of the client (instead of complaint resolution).
Bulletin of the IEEE Technical Committee on Learning Technology, Volume 18, Number 2/3, 2016
II. DECISION TREES AND ADVISOR FLOW GRAPHS
For our case study, we selected a small group of Remote
Support agents (technicians), who operated under the
supervision of one of the co-authors, and recorded their
measurements as attributes for Internet and TV complaints
instances. Each instance corresponds to an Internet or IPTV
(Internet Protocol TV) over-xDSL complaint, where we record
synchronization between the Integrated Access Device and the
Central Office routers, in addition to other copper-related
metrics. For these agents, we also recorded whether they
handled the complaint at the helpdesk level or referred it to On
Site Support. One aims at maximizing the number of issues
resolved remotely, without having to resort to On Site
Support. Though this makes obvious sense from a financial
point of view, it is not trivial; sub-standard problem resolution
generates recurring complaints, which may or may not be
dealt by the same agent next time, thus raising costs (which is
seldom the case for the more expensive On Site Support).
The training set we use to build a decision tree, using the
C4.5 algorithm [4], looks like the one shown in Table I.
TABLE I
A SNAPSHOT OF A TRAINING SET
Agent
Product
Area
…
…
…
AGENT04 INTERNET Athens
AGENT02 INTERNET Patras
AGENT03 INTERNET Athens
…
…
…
Profile Sync
…
2
30
24
…
Max
…
…
2048 9138
26421 27786
2045 3408
…
…
Dist
State
…
2.7
1.8
3.1
…
…
OS
RS
OS
…
State is the class variable (showing the type of support, On-Site/Remote).
Product refers to the service a client has bought and for which service the
complaint is being lodged.
Area refers to the city where the fault occured.
Profile refers to the nominal speed of the connection for that client (not
the actual capacity of the line, but the service capacity; for example, a lowcost client might buy a 2 Mb/s service even though the line might
accommodate much larger speeds).
Sync refers to the actual synchronization speed of a client's line.
Max refers to the theoretical maximum attainable synchronization of a
client's line based on signal strength, attenuation, and a variety of physical
characteristics.
Dist refers to the copper cable length from the central circuit
infrastructure (measured in km).
The training set was made up of about 1,500 helpdesk
reports, recorded over a period of about 3 weeks, by
automatically logging data collected at distributed corporate
information systems, for 5 helpdesk agents. The integrated
data set has been generated by combining those data together.
A. Cost-oblivious decision trees
For the base-level experiment we treat individual agents as
attributes. We have used RWeka and R for analyzing the data
and building the models. Running a 10-fold cross-validation
delivered an accuracy of 69.67% of predicting whether the
complaint would be resolved at the on-site or at the remote
level. Fig. 1 shows a snapshot of the decision tree; the (dark)
multi-valued AGENT attribute appears at all nodes near the
tree root and, clearly, should be avoided in any (agent-
7
agnostic) SOP. Dropping that attribute along with the AREA
attribute (agents service requests from a central location which
may better resolved just by knowing singularity aspects of the
actual physical network) resulted in an accuracy of 60.65%.
Fig. 1. A decision tree using all data labels - agent as a decision node.
We then use the data sets of individual agents to develop
decision trees which capture how each agent treats a customer
complaint as well as a measure of treatment uniformity. The
measurement is based on partitioned data sets which are crosstested [11]. We partition the overall complaints data set into
groups, according to how complaints were assigned to agents
and examine whether the problem resolution practices of one
agent apply to customer complaints handled by another agent.
On one hand, this approach allows us to see how each agent's
decision model is (or is not) compatible with other agents'
models. On the other hand, however, it also allows us to
consider model differences compared not to a theoretical
maximum of 1 but to a practical maximum as calculated by
the re-classification error of each agent's model.
TABLE II
CORRECTLY CLASSIFIED INSTANCES - CROSS-TEST RESULTS
1
1
2
3
4
5
2
0.60
0.49
0.49
0.54
0.50
3
0.38
0.72
0.66
0.67
0.47
4
0.48
0.56
0.60
0.52
0.52
5
0.54
0.50
0.50
0.75
0.53
0.50
0.42
0.42
0.61
0.71
Cross-test results among all available agents are shown in
Table II. Therein, each <i;i> entry (on the main diagonal)
refers to the re-classification testing accuracy; note that these
numbers are not even close to 1 (suggesting an inherently hard
problem) and that the relative standard deviation (RSD) is
0.1049 (indicating small differences in how each agent reclassifies the training data generated by him/herself and
concurring with Fig. 1, where the agent attribute seems to
dominate the differences in how a problem is dealt with). Each
<i;j> entry describes an experiment where data for the i-th
agent is used to build a decision tree which is subsequently
tested on data for the j-th agent.
Bulletin of the IEEE Technical Committee on Learning Technology, Volume 18, Number 2/3, 2016
B. Cost-sensitive decision trees
A cost-matrix that captures the differences between
misclassifying customer complaints as requiring On Site
Support vs Remote Resolution Service is:
CM
0
⎡
=⎢
OPESX
RS
⎣
OPESX OS ⎤
⎥
0
⎦
1
2
3
4
5
0.1
0.5
0.5
0.1
0.2
3
0.1
0.3
0.3
0.1
0.1
4
0.1
0.4
0.4
0.2
0.1
1
2
3
4
5
5
0.1
0.5
0.5
0.1
0.1
0.2
0.6
0.6
0.1
0.1
The RSD along the main diagonal now stands at 0.7071 and
raises the question why agents fare so differently when asked
to classify the very data they were trained on, while also
highlighting substantial cost differences between agents, when
some of them are used to classify data used by other agents.
It might seem that an agent may be unfairly penalized for
handling a majority of serious complaints requesting on-site
maintenance. We note that complaints are assigned to agents
randomly (for example, we do not pick experienced agents to
deal with frustrated customers); as a result, one expects that,
on average, hard and easy problems are uniformly allocated.
Moreover, any performance review will no doubt also focus
on how specific complaints were dealt with; our technique at
this point mainly serves to highlight inconsistencies in dealing
with customer problems and not the actual cost of the
maintenance approach (as a result, the question of why two
particular agents performed a different classification on a
similar set of complaints is of relatively less importance).
So, incorporating the cost aspect into the decision tree
classifier allowed us to build agent-aligned models and show
how these scale up to new data sets. To stress how traditional
techniques provide rather just crude clues, note that traditional
KPI-based agent monitoring usually consists of at least 2
indices per agent. Table IV reports on the conventional cost
index for each agent (using actual values for two real KPI
indices; their names as well as the formula to calculate the
composite cost index are withheld due to commercial secrecy
but, as most KPI indices, they do bear some relation with the
KPI1
0.30
0.20
0.14
0.25
0.25
KPI2
AGENTCOST
0.44
0.42
0.39
0.46
0.44
0.72
0.69
0.69
0.68
0.69
Table V finally reviews our experiments (ordered by
increasing ability to tell agents apart in terms of performance).
TABLE V
COMPARISON OF TECHNIQUE RESOLUTION
Table
4
2
3
AVERAGE MISCLASSIFICATION COST
2
TABLE IV
TRADITIONAL KPIS AND COST EVALUATION
Agenti
TABLE III
1
effectiveness of complaint resolution, also taking into account
recurring complaints). We now observe an RSD of a mere
0.0219; this clearly suggests that traditional indices grossly
hide differences.
(1)
This is not a conventional misclassification cost matrix
[12]; we actually note the cost incurred to carry out some
maintenance activity, while we choose to ignore the cost of
that activity when the model correctly predicts the class label
(of the maintenance activity). So, when an On Site instance is
correctly classified by our model, we assign it a cost of 0,
though it incurs a cost of OPEXOS to resolve; instead, we
reserve the cost of OPEXOS only for those instances which our
model classifies as On Site though the class label in the data
set reads Remote. Cross-testing agents with the above cost
perspective is shown in Table III.
8
RSD
Cost using Traditional KPIs
Correctly Classified Instances only
Average Misclassification Cost
0.022
0.105
0.707
C. Advisor Flow Graph: just-in-time collaborative training
Since some agents' models seem to be better suited when
classifying hereto un-seen instances, it might be reasonable to
ask those agents to serve as advisors for their colleagues.
Though organizational level training involves the review of
past cases as well as helpdesk tactics (how to use one's
communication skills, for example), the helpdesk data analysis
models can serve as on-the-spot training; when an agent
addresses a troubleshooting report, it makes sense to offer
him/her an alternative approach to the same problem. It is
crucial to present that advice as a hint and not as a
recommendation to be followed; since the actual classification
problem is very difficult on its own, any advice that sounds
too firm might be easily resisted on grounds of just a
counterexample (and, as we have seen, these do occur).
To capture a picture of which agent can help a colleague,
we draw an advisor flow graph, where at source nodes we
place those agents with "better" models, whereas agents who
seem to act with increased misclassification costs are placed at
destination nodes. Table III is used to build the advisor flow
graph shown in Fig. 2, using Equation 2 for the corresponding
adjacency and weight matrix. The semantics of the advisor
flow graph are straightforward, with larger edge weights
reflecting larger quality differences in the corresponding
cross-tested models, according to Equation 2. For clarity
purposes, we have drawn the graph in three levels to denote
that the top level nodes are source-only nodes whereas the
bottom level nodes are destination-only nodes. For our
example, Agent 3 should consult Agents 1, 4 or 5, before
deciding, whereas Agent 2 might also consult Agent 3, should
the others be unavailable (a pop-up window describing an
alternative action and the justification could suffice).
Bulletin of the IEEE Technical Committee on Learning Technology, Volume 18, Number 2/3, 2016
⎧
⎪C i , j > C j ,i → w(i, j ) = C i , j − C j ,i
⎨
⎪
⎩C i , j < C j ,i → w( j , i ) = C i , j − C j ,i
(2)
9
probably be tackled on a trial-and-error basis.
Simply put, the contribution of this article is the use of
decision trees as a classification mechanism and as a
knowledge transfer mechanism for helpdesk technicians. But,
while it may be easy to build the advisor graph, deploying it
and monitoring how it evolves and how it actually improves
quality of service is a key part of our agenda to close the loop
between monitoring and acting, via peer-training.
ACKNOWLEDGMENT
The work was carried out while the first author was with the
Hellenic Telecommunications Organization (OTE S.A.),
Greece. We acknowledge the granting of permission to use
individual technicians' data. All data is confidential and the
property of OTE S.A.
Fig. 2. An advisor flow graph showing the flow of information.
III. CONCLUSIONS AND FUTURE WORK
We identify helpdesk agents capable of on-the-job advising
their colleagues, using cost-sensitive decision tree models of
how agents deal with customer complaints. These decision
trees pinpoint agent differences and help build a graph to
capture possible advice flows between agents. These advice
flows are, essentially, by-products of standard data mining
activities and can be used as non-intrusive on-the-job
recommendation, suitable not only for decision making but
also for reflection.
An advisor graph is a dynamic knowledge hierarchy [13]
and captures the implicit reputation enjoyed by members of a
community of practice, even though such knowledge may be
implicit (and volatile). This is relevant to leader selection via
clustering as performed in swarm-based optimization
problems [14], where leaders help balance quality and
diversity; in our case, this refers to organizational goals of
effectiveness and efficiency. An advisor graph can gracefully
scale up with the number of agents as its calculation can be
carried out off-line; subsequently, selecting who-adviseswhom is performed by selecting any of the available source
nodes for a particular destination node. It can also scale up
with the number of attributes recorded for each complaint; it is
not unnatural to enhance each record with some agent-specific
attributes so that agent similarities may be detected and
exploited across a variety of other criteria (note, however, that
recording data such as agent age, sex, education background,
etc. can ran contrary to law and/or organizational regulations,
which is why, for the context of our case study, we have
limited data collection to the business-specific domain only).
Knowledge volatility issues arise as the advisor graph
changes over time, due to the variability of the troubleshooting
activities or to the helpdesk agents. We expect that, for an
application, one might decide to analyze medium-sized agent
groups and a larger time window. Establishing the right
combination of group size and time window will likely be a
difficult problem and, for all practical purposed, should
REFERENCES
[1]
[2]
[3]
[4]
[5]
[6]
[7]
[8]
[9]
[10]
[11]
[12]
[13]
[14]
Gustafsson, A., Johnson, M. D. and Roos, I. 2005. The Effects of
Customer Satisfaction, Relationship Commitment Dimensions, and
Triggers on Customer Retention. Journal of Marketing 69, 210-218.
Ganesh, J., Arnold, M.J. and Reynolds, K.E. 2000. Understanding the
Customer Base of Service Providers: An Examination of the Differences
Between Switchers and Stayers. Journal of Marketing 64 (3), 65-87.
Hwang, H., Jung, T. and Suh, E.: An LTV model and customer
segmentation based on customer value: a case study on the wireless
telecommunication industry. Expert Systems with Applications 26 (2),
181-188.
Quinlan, J.R. 1993. C4.5: Programs for Machine Learning. Morgan
Kaufmann Publishers.
Muggleton, S. and De Raedt, L. 1994. Inductive Logic Programming:
Theory and methods. The Journal of Logic Programming 1920(Supplement 1), 629-679.
Reichheld, F.F. 2003. The One Number You Need to Grow. Harvard
Business Review (December).
Ngai, E.W.T., Xiu, L., and Chau, D.C.K. 2009. Application of data
mining techniques in customer relationship management: A literature
review and classification. Expert Systems with Applications 36 (2, Part
2), 2592-2602.
Khodakarami, F. and Chan, Y.E. 2014. Exploring the role of customer
relationship management (CRM) systems in customer knowledge
creation. Information & Management 51(1), 27-42.
Tseng, S.-M. 2016. Knowledge management capability, customer
relationship management, and service quality. Journal of Enterprise
Information Management 29(2), 202-221.
Soltani, Z. and Navimipour, N.J. 2016. Customer relationship
management mechanisms: A systematic review of the state of the art
literature and recommendations for future research. Computers in
Human Behavior 61, 667-688.
Hadzilacos, T., Kalles, D., Pierrakeas, C. and Xenos, M. 2006. On Small
Data Sets revealing Big Differences. Advances in Artificial Intelligence,
Antoniou, G., Potamias, G., Spyropoulos, C. and Plexousakis, D. (Eds.),
LNCS 3955, 512-515.
Turney, P.T. 1995. Cost-Sensitive Classification: Empirical Evaluation
of a Hybrid Genetic Decision Tree Induction Algorithm. Journal of
Artificial Intelligence Research 2, 369-409.
Tamargo, L.H., Garca, A.J, Falappa, M.A. and Simari, G.R. 2014. On
the revision of informant credibility orders. Artificial Intelligence 212,
36-58.
Al Moubayed, N., Petrovski, A. and McCall, J. 2011. Clustering-Based
Leaders' Selection in Multi-Objective Particle Swarm Optimisation.
Intelligent Data Engineering and Automated Learning, Yin, H., Wang,
W. and Rayward-Smith, V. (Eds.), LNCS 6936,100-107.
| 2cs.AI
|
Elicitation for Preferences Single Peaked on Trees
arXiv:1604.04403v1 [cs.GT] 15 Apr 2016
Palash Dey
Indian Institute of Science, Bangalore
[email protected]
Neeldhara Misra
Indian Institute of Technology, Gandhinagar
[email protected]
April 18, 2016
Abstract
In multiagent systems, we often have a set of agents each of which have a preference ordering
over a set of items and one would like to know these preference orderings for various tasks, for
example, data analysis, preference aggregation, voting etc. However, we often have a large
number of items which makes it impractical to ask the agents for their complete preference
ordering. In such scenarios, we usually elicit these agents’ preferences by asking (a hopefully
small number of) comparison queries — asking an agent to compare two items. Prior works on
preference elicitation focus on unrestricted domain and the domain of single peaked preferences
and show that the preferences in single peaked domain can be elicited by much less number of
queries compared to unrestricted domain. We extend this line of research and study preference
elicitation for single peaked preferences on trees which is a strict superset of the domain of single
peaked preferences. We show that the query complexity crucially depends on the number of
leaves, the path cover number, and the distance from path of the underlying single peaked tree,
whereas the other natural parameters like maximum degree, diameter, pathwidth do not play
any direct role in determining query complexity. We then investigate the query complexity for
finding a weak Condorcet winner for preferences single peaked on a tree and show that this
task has much less query complexity than preference elicitation. Here again we observe that the
number of leaves in the underlying single peaked tree and the path cover number of the tree
influence the query complexity of the problem.
1
Introduction
In multiagent systems, we often have scenarios where agents have to arrive at a consensus when
choosing between multiple options. Typically, the agents have preferences over a set of items, and
the problem of aggregating these preferences in a suitable manner is one of the most well-studied
problems in social choice theory [BCE+ 15]. There are many ways of expressing preferences over a
set of alternatives. One of the most comprehensive ways is to specify a complete ranking over the
set of alternatives. However, one of the downsides of this model is the fact that it can be expensive
to solicit a preference when there are a large number of alternatives, and many agents are involved.
Since asking agents to provide their complete rankings is impractical, a popular notion is one
of elicitation, where we ask agents simple comparison queries, such as if they prefer alternative X
over Y . This naturally gives rise to the problem of preference elicitation, where we hope to recover
the complete ranking (or possibly the most relevant part of the ranking) based on a small number
of queries.
The paradigm of voting is a popular general methodology for aggregating preferences, where
one devises “voting rules” for mapping a collection of preferences (which we refer to as votes) to
either a winning alternative or a consensus ranking. Keeping in line with the terminology used
in voting, we will refer to alternatives as candidates, and a collection of votes will be termed a
preference profile. In the context of a fixed voting rule, we may also want to query the voters up to
the point of determining the winner (or the aggregate ranking, as the case may be). Yet another
refinement in this setting is when we have prior information about how agents are likely to vote,
and we may want to determine which voters to query first, to be able to quickly rule out a large
number of alternatives, as explored by [CS02].
When our goal is to elicit preferences that have no prior structure, one can demonstrate scenarios where it is imperative to ask each agent (almost) as many queries as would be required to
determine an arbitrary ranking. However, in recent times, there has been considerable interest in
voting profiles that are endowed with additional structure. The motivation for this is two-fold.
The first is that in several application scenarios commonly considered, it is rare that votes are adhoc, demonstrating no patterns whatsoever. For example, the notion of single-peaked preferences,
which we will soon discuss at length, forms the basis of several studies in the analytical political sciences [HM97]. In his work on eliciting preferences that demonstrate the “single-peaked”
structure, [Con09] argues that the notion of single-peakedness is also a reasonable restriction in
applications outside of the domain of political elections.
The second motivation for studying restricted preferences is somewhat more technical, but is
just as compelling. To understand why structured preferences have received considerable attention
from social choice theorists, we must first take a brief detour into some of the foundational ideas
that have shaped the landscape of voting rules as we understand them today. As it turns out, the
axiomatic approach of social choice involves defining certain “properties” that formally capture the
quality of a voting rule. For example, we would not want a voting rule to be, informally speaking, a
dictatorship, which would essentially mean that it discards all but one voter’s input. Unfortunately,
a series of cornerstone results establish that it is impossible to devise voting rules which respect
some of the simplest desirable properties. Indeed, the classic works of Arrow [Arr50] and GibbardSatterthwaite [Gib73, Sat75] show that there is no straight forward way to simultaneously deal
with properties like voting paradoxes, strategy-proofness, nondictatorship, unanimity etc. We refer
to [Mou91] for a more elaborate discussion. Making the matter worse, many classical voting rules
turn out to be computationally intractable.
This brings us to the second reason for why structured preferences are an important consideration. The notion of single-peakedness that we mentioned earlier is an excellent illustration (we
refer the reader to Section 2 for the formal definition). Introduced by [Bla48], it not only captures
the essence of structure in political elections, but also turns out to be extremely conducive to many
natural theoretical considerations. To begin with, one can devise voting rules that are “nice” with
respect to several properties, when preferences are single-peaked. Further, they are structurally
elegant from the view of winner determination, since they always admit a weak Condorcet winner
— a candidate which is not defeated by any other candidate in pairwise election — thus working
around the Condorcet paradox which is otherwise a prominent concern in the general scenario. In
a landmark contribution, [BBHH15] show that several computational problems that are intractable
in the general setting become polynomially solvable when we consider single-peaked preferences.
A natural question at this point is if the problem of elicitation becomes any easier — that is,
if we can get away with fewer queries — by taking advantage of the structure provided by single-
peakedness. It turns out that the answer to this is in the affirmative, as shown in a detailed study
by [Con09]. The definition of single-peakedness involves an ordering over the candidates (called the
harmonious ordering by some authors). The work of [Con09] shows that O(mn) queries suffice,
assuming either that the harmonious ordering is given, or one of the votes is known. Dey and
Misra [DM16] show a query complexity bound of O(mn) for the domain of single crossing profiles
and a large number of voters.
We now return to the theme of structural restrictions on preferences. As it turns out, the
single peaked preference domain has subsequently been generalized to single peakedness on trees
(roughly speaking, these are profiles that are single peaked on every path) [Dem82, Tri89]. This is
a class that continues to exhibit many desirable properties of single peaked domains. For example,
there always exists a weak Condorcet winner and further, many voting rules that are intractable in
unrestricted domain are polynomial time computable if the underlying single peaked tree is “nice”
[YCE13, PE16]. We note the class of profiles that are single peaked on trees are substantially more
general than the class of single peaked preferences. Note that the latter is a special case since a
path is, in particular, a tree. Our work here addresses the issue of elicitation on profiles that are
single-peaked on trees, and can be seen as a significant generalization of the results in [Con09]. We
now detail the specifics of our contributions.
Parameter
Upper Bound
Lower Bound
Path width (w)
Maximum degree (∆)
Path cover number (k)
Number of leaves (`)
Distance from path (d)
Diameter (ω)
O(mn log m) Observation 1
O(mn log m) Observation 1
O(mn log k) Theorem 2
O(mn log `) Corollary 1
O(mn + nd log d) Theorem 3
O(mn log m) Observation 1
Ω(mn log m) even for w = 1, log m Corollary 3
Ω(mn log m) even for ∆ = 3, m − 1 Corollary 4
Ω(mn log k) Corollary 2
Ω(mn log `) Theorem 4
Ω(mn + nd log d) Theorem 6
Ω(mn log m) even for ω = 2, m/2 Corollary 5
Table 1: Summary of query complexity bounds for Preference Elicitation.
1.1
Motivation
1.2
Our Contributions
We study the query complexity for preference elicitation when the preference profile is single peaked
on a tree. We provide tight connections between various parameters of the underlying tree and the
query complexity for preference elicitation. Our broad goal is to provide a suitable generalization of
preference elicitation for single peaked profiles to profiles that are single peaked on trees. Therefore,
we consider various ways of quantifying the “closeness” of a tree to a path, and reflect on how these
measures might factor into the query complexity of an algorithm that is actively exploiting the
underlying tree structure.
We summarize our results for preference elicitation in Table 1, where the readers will note that
most of the parameters (except diameter) chosen are small constants (typically zero, one or two)
when the tree under consideration is a path. Observe that in some cases — such as the number
of leaves, or the path cover number — the dependence on the parameter is transparent (and we
recover the results of [Con09] as a special case), while in other cases, it is clear that the perspective
provides no additional mileage (the non-trivial results here are the matching lower bounds).
In terms of technique, our strategy is to “scoop out the paths from the tree” and use known
algorithms to efficiently elicit the preference on the parts of the trees that are paths. We then
efficiently merge this information across the board, and that aspect of the algorithm varies depending on the parameter considered. The lower bounds typically come from trees that provide large
“degrees of freedom” in reordering candidates, typically these are trees that don’t have too many
long paths (such as stars). The arguments are often subtle but intuitive.
We then study the query complexity for finding a weak Condorcet winner of a preference profile
which is single peaked on a tree. Here, we are able to show that a weak Condorcet winner can
be found with far fewer queries than the corresponding elicitation problem. In particular, we
establish that a weak Condorcet winner can be found using O(mn) many queries for profiles that
are single peaked on trees [Theorem 7], and we also show that this bound is the best that we can
hope for [Theorem 10]. We also consider the problem for the special case of single peaked profiles.
While [Con09] showed that Ω(mn) queries are necessary to determine the aggregate ranking, we
show that only O(n log m) queries suffice if we are just interested in (one of the) weak Condorcet
winners. Moreover, we show this bound is tight under the condition that the algorithm does not
interleave queries to different voters [Theorem 11] (our algorithm indeed satisfies this condition).
Finally, expressing the query complexity for determining a weak Condorcet winner in terms of
a measure of closeness to a path, we show an algorithm with query complexity O(nk log m) where
k is the path cover number of T [Theorem 12] or the number of leaves in T [Corollary 6]. We now
elaborate further on our specific contributions for preference elicitation.
– We design novel algorithms for preference elicitation for profiles which are single peaked on
a tree with ` leaves with query complexity O(mn log `) [Corollary 1]. Moreover, we prove
that there exists a tree T with ` leaves such that any preference elicitation algorithm for
profiles which are single peaked on tree T has query complexity Ω(mn log `) [Theorem 4]. We
show similar results for the parameter path cover number of the underlying tree [Theorem 2
and Corollary 2]. We provide a preference elicitation algorithm with query complexity O(mn+
nd log d) for single peaked profiles on trees which can be made into a path by deleting at most
d nodes [Theorem 3]. We show that our query complexity upper bound is tight up to constant
factors [Theorem 6]. These results show that the query complexity tightly depends on the
number of leaves, the path cover number, and the distance from path of the underlying tree.
– We then show that there exists a tree T with pathwidth one or log m [Corollary 3] or maximum degree is 3 or m − 1 [Corollary 4] or diameter is 2 or m/2 [Corollary 5] such that any
preference elicitation algorithm for single peaked profiles on the tree T has query complexity
Ω(mn log m). These results show that the query complexity of preference elicitation does not
directly depend on the parameters above.
We next study query complexity for finding a weak Condorcet winner for profiles which are
single peaked on trees and we have the following results.
– We show that a weak Condorcet winner can be found using O(mn) many queries for profiles
that are single peaked on trees [Theorem 7] which is better than the query complexity for
preference elicitation. Moreover, we prove that this bound is tight in the sense that any
algorithm for finding a weak Condorcet winner for profiles that are single peaked on stars has
query complexity Ω(mn) [Theorem 10].
– On the other hand, we can find a weak Condorcet winner using only O(n log m) many queries
for single peaked profiles [Theorem 8]. Moreover, we show this bound is tight under the
condition that the algorithm does not interleave queries to different voters [Theorem 11] (our
algorithm indeed satisfies this condition). For any arbitrary underlying single peaked tree
T , we provide an algorithm for finding a weak Condorcet winner with query complexity
O(nk log m) where k is the path cover number of T [Theorem 12] or the number of leaves in
T [Corollary 6].
To summarize, we remark that our results non-trivially generalize earlier works on query complexity for preference elicitation in [Con09]. We believe revisiting the preference elicitation problem
in the context of profiles that are single peaked on trees is timely, and that this work also provides
fresh algorithmic and structural insights on the domain of preferences that are single peaked on
trees.
1.3
Related Work
We have already mentioned the work of [Con09] addressing the question of eliciting preferences
in single-peaked profiles, which is the closest predecessor to our work. Before this, Conitzer and
Sandholm addressed the computational hardness for querying minimally for winner determination
[CS02]. They also prove that one would need to make Ω(mn log m) queries even to decide the winner
for many commonly used voting rules [CS05] which matches with the trivial O(mn log m) upper
bound for preference elicitation in unrestricted domain based on sorting. Ding and Lin study
preference elicitation under partial information setting and show interesting properties of what
they call a deciding set of queries [DL13]. Lu and Boutilier provide empirical study of preference
elicitation under probabilistic preference model [LB11b] and devise several novel heuristics which
often work well in practice [LB11a].
The rest of the paper is organized as follows: We introduce the basic terminologies and formally
define the problems in Section 2, present results for eliciting preference profile and finding a weak
Condorcet winner in Section 3 and Section 4 respectively, and conclude with future works in
Section 5.
2
Preliminaries
For a positive integer n, we denote the set {1, . . . , n} by [n]. Suppose we have a set V = {vi ∶ i ∈ [n]}
of n voters each of which has a preference ≻i , i ∈ [n], alternatively called vote, which is a complete
order over a set C = {cj ∶ j ∈ [m]} of m candidates. We denote the set of preferences over C by L(C).
The tuple (≻i )i∈[n] of the preferences of all the voters is called a profile. If not mentioned otherwise,
we use m and n to denote the number of candidates and the number of voters respectively. Let
≻ ∈ L(C) be any complete order over C. For a subset X ⊆ C of candidates, we denote the restriction
of an order ≻ to the subset of candidates X by ≻ (X ). We denote the restriction of a profile
P = (≻i )i∈[n] to X by P(X ) = (≻i (X ))i∈[n] . We say that a candidate x ∈ C is placed at the ith
position of a preference ≻ ∈ L(C) if x is preferred over all but exactly (i − 1) candidates in ≻. Given
an n voters profile P = (≻i )i∈[n] , a candidate x is called a Condorcet winner of P if for every other
candidate y, a strict majority of the voters prefer x over y; that is ∣{i ∈ [n] ∶ x ≻i y}∣ > n/2 for every
y ∈ C ∖ {x}. A candidate x is called a weak Condorcet winner of an n voters profile P = (≻i )i∈[n] if
there does not exist any other candidate y whom a strict majority of the voters prefer over x; that
is ∣{i ∈ [n] ∶ y ≻i x}∣ ≤ n/2 for every y ∈ C ∖ {x}. For an order ≻ ∈ L(C), we denote the “reverse order”
Ð
Ð
Ð
of ≻ by ←
≻ ; that is ←
≻ = {x←
≻ y ∶ x, y ∈ C and y ≻ x}. A preference ≻ ∈ L(C) over a set of candidates
C is called single peaked with respect to an order ≻′ ∈ L(C) if, for every candidates x, y ∈ C, we have
x ≻ y whenever we have either c ≻′ x ≻′ y or y ≻′ x ≻′ c, where c ∈ C is the candidate at the first
position of ≻. A profile P = (≻i )i∈[n] is called single peaked with respect to an order ≻′ ∈ L(C) if ≻i
is single peaked with respect to ≻′ for every i ∈ [n]. Notice that if a profile P is single peaked
Ð
with respect to an order ≻′ ∈ L(C), then P is also single peaked with respect to the order ←
≻ ′ . Given
a path Q = (x1 , x2 , . . . , x` ) from a vertex x1 to another vertex x` in a tree T , we define the order
induced by the path Q to be x1 ≻ x2 ≻ ⋯ ≻ x` . Given a tree T = (C, E) with the set of nodes as
the set of candidates C, a profile P is called single peaked on the tree T if P is single peaked on
the order induced by every path of the tree T ; that is for every two candidates x, y ∈ C, the profile
P(X ) is single peaked with respect to the order ≻ of the candidates X induced by the unique path
from x to y in T . We call the tree T the underlying single peaked tree. It is known (c.f. [Dem82])
that there always exists a weakly Condorcet winner for a profile P which is single peaked on a tree
T.
Trees.
The following definitions pertaining to the structural aspects of trees will be useful.
– The pathwidth of T is the minimum width of a path decomposition of T [Hei93].
– A set of disjoint paths Q = {Q1 = (X1 , E1 ), . . . , Qk = (Xk , Ek )} is said to cover a tree T = (X , E)
if X = ∪i∈[k] Xi , Ei ⊆ E, Xi ∩ Xj = ∅, Ei ∩ Ej = ∅ for every i, j ∈ [k] with i ≠ j. The path cover
number of T is the cardinality of the smallest set Q of disjoint paths that cover T .
– The distance of a tree T from a path is the smallest number of nodes whose removal makes
the tree a path.
– The diameter of a tree T is the number of edges in the longest path in T .
We also list some definitions of subclasses of trees (which are special types of trees, see also Figure 1).
– A tree is a star if there is a center vertex and every other vertex is a neighbor of this vertex.
– A tree is a subdivision of a star if it can be constructed by replacing each edge of a star by a
path.
– A subdivision of a star is called balanced if there exists an integer ` such that the distance of
every leaf node from the center is `.
– A tree is a caterpillar if there is a central path and every other vertex is at a distance of one
it.
– A tree is a complete binary tree rooted at r if every nonleaf node has exactly two children
and there exists an integer h, called the height of the tree, such that every leaf node is at a
distance of either h or h − 1 from the root node r.
Figure 1: Depicting classes of trees: (a) a path, (b) a star, (c) a (balanced) subdivision of a star,
(d) a caterpillar.
Problem Definitions and Known Results. Suppose we have a profile P with n voters and
m candidates. For any pair of distinct candidates x and y, and a voter ` ∈ [n], we introduce the
boolean-valued function Query(x ≻` y). The output of this function is true if the voter ` prefers
the candidate x over the candidate y and false otherwise. We now formally state the two problems
that we consider in this paper.
Definition 1 Preference Elicitation
Given a tree T = (C, E) and an oracle access to the function Query(⋅) for a profile P which is
single peaked on T , find P.
Definition 2 Weak Condorcet Winner
Given a tree T = (C, E) and an oracle access to the function Query(⋅) for a profile P which is
single peaked on T , find a weak Condorcet winner of P.
Suppose we have a set of candidates C = {c1 , . . . , cm }. We say that an algorithm A makes q
queries if there are exactly q distinct tuples (`, ci , cj ) ∈ [n] × C × C with i < j such that A calls
Query(ci ≻` cj ) or Query(cj ≻` ci ). We call the number of queries made by an algorithm A its
query complexity.
We state some known results that we will appeal to later. The first observation employs a
sorting algorithm like merge sort to elicit every vote with O(m log m) queries, while the second
follows from the linear-time merge subroutine of merge sort ([Cor09]).
Observation 1 There is a Preference Elicitation algorithm with query complexity
O(mn log m).
Observation 2 Suppose C1 , C2 ⊆ C form a partition of C and ≻ is a ranking of the candidates in
C. Then there is a polynomial time algorithm that finds ≻ given ≻ (C1 ) and ≻ (C2 ) with query
complexity O(∣C∣).
Theorem 1 [Con09] There is a Preference Elicitation algorithm with query complexity
O(mn) for single peaked profiles.
We now state the Weak Condorcet Winner problem, which asks for eliciting only up to the
point of determining a weak Condercet winner (recall that at least one such winner is guaranteed
to exist on profiles that are single-peaked on trees).
Definition 3 Weak Condorcet Winner
Given a tree T = (C, E) and an oracle access to the function Query(⋅) for a profile P which is
single peaked on T , find a weak Condorcet winner of P.
3
Results for Preference Elicitation
In this section, we present our results for Preference Elicitation for profiles that are single
peaked on trees. Recall that we would like to generalize Theorem 1 in a way to profiles that are
single peaked on trees. Since the usual single peaked profiles can be viewed as profiles single peaked
with respect to a path, we propose the following measures of how much a tree resembles a path.
– Leaves. Recall any tree has at least two leaves, and paths are the trees that have exactly two
leaves. We consider the class of trees that have ` leaves, and show an algorithm with a query
complexity of O(mn log `).
– Path Cover. Consider the notion of a path cover number of a tree, which is the smallest number
of disjoint paths that the tree can be partitioned into. Clearly, the path cover number of a
path is one; and for trees that can be covered with k paths, we show an algorithm with query
complexity O(mn log k).
– Distance from Paths. Let d be the size of the smallest set of vertices whose removal makes
the tree a path. Again, if the tree is a path, then the said set is simply the empty set. For
trees that are at a distance d from being a path (in the sense of vertex deletion), we provide
an algorithm with query complexity O(mn log d).
– Pathwidth and Maximum Degree. Finally, we note that paths are also trees that have pathwidth one, and maximum degree two. These perspectives turn out to be less useful: in
particular, there are trees where these parameters are constant, for which we show that elicitation is as hard as it would be on an arbitrary profile, and therefore the easy algorithm
from Observation 1 is actually the best that we can hope for.
For the first three perspectives that we employ, that seemingly capture an appropriate aspect
of paths and carry it forward to trees, the query complexities that we obtain are tight — we have
matching lower bounds in all cases. Also, while considering structural parameters, it is natural to
wonder if there is a class of trees that are incomparable with paths but effective for elicitation. Our
attempt in this direction is to consider trees of bounded diameter. However, again, we find that
this is not useful, as we have examples to show that there exist trees of diameter two that are as
hard to elicit as general profiles.
We remark at this point that all these parameters are polynomially computable for trees, making the algorithmic results viable. Also, for the parameters of pathwidth, maximum degree and
diameter, we show lower bounds on trees where these parameters are large (such as trees with pathwidth O(log m), maximum degree m − 1, and diameter m/2), which — roughly speaking — also
rules out the possibility of getting a good inverse dependence. As a concrete example, motivated by
the O(mn) algorithm for paths, which have diameter m, one might wonder if there is an algorithm
with query complexity O( mnloglogω m ). This possibility, in particular, is ruled out. We are now ready
to discuss the results in Table 1.
We next show a structural result about trees namely any tree with ` leaves can be partitioned
into ` paths. The idea is to fix some nonleaf node as root and iteratively find a path from low
depth nodes (depth of a node is its distance from root) to some leaf node which is disjoint from all
the paths chosen so far. We formalize this idea below.
Lemma 1 Let T = (X , E) be a tree with ` leaves. Then there is a polynomial time algorithm which
partitions T into ` disjoint paths Qi = (Xi , Ei ), i ∈ [`]; that is we have Xi ∩ Xj = ∅, Ei ∩ Ej = ∅ for
every i, j ∈ [`] with i ≠ j, X = ∪i∈[`] Xi , E = ∪i∈[`] Ei , and Qi is a path in T .
Proof: We first make the tree T rooted at any arbitrary nonleaf node r. We now partition the
tree T into paths iteratively as follows. Initially every node of the tree T is unmarked and the
set of paths Q we have is empty. Let Q1 = (X1 , E1 ) be the path in T from the root node to any
leaf node. We put Q1 into Q and mark all the nodes in Q1 . More generally, in the ith iteration
we pick an unmarked node u that is closest to the root r, breaking ties arbitrarily, and add any
path Qi in T from u to any leaf node in the subtree Tu rooted at u, and mark all the nodes in
Qi = (Xi , Ei ). Since u is unmarked, we claim that every node in Tu is unmarked. Indeed, otherwise
suppose there is a node w in Tu which is already marked. Then there exists two paths from r to w
one including u and another avoiding u since u is currently unmarked and w is already marked.
This contradicts the fact that T is a tree. Hence every node in Tu is unmarked. We continue until
all the leaf nodes are marked and return Q. Since T has ` leaves and in every iteration at least
one leaf node is marked (since every Qi contains a leaf node), the algorithm runs for at most `
iterations. Notice that, since the algorithm always picks a path consisting of unmarked vertices
only, the set of paths in Q are pairwise disjoint. We claim that Q forms a partition of T . Indeed,
otherwise there must be a node x in T that remains unmarked at the end. From the claim above,
we have all the nodes in the subtree Tx rooted at x unmarked which includes at least one leaf node
of T . This contradicts the fact that the algorithm terminates when all the leaf nodes are marked. ◻
Using Lemma 1, and the fact that any path can account for at most two leaves, we have that
the path cover number of a tree is same as the number of leaves up to a factor of two.
Lemma 2 Suppose the path cover number of a tree T with ` leaves is k. Then we have `/2 ≤ k ≤ `.
Proof: The inequality `/2 ≤ k follows from the fact that any path in T can involve at most two
leaves in T and there exists k paths covering all the leaf nodes. The inequality k ≤ ` follows from
the fact from Lemma 1 that any tree T with ` leaves can be partitioned into ` paths.
◻
3.1
Algorithmic Results
We now present our main algorithmic results. We begin with generalizing the result of Theorem 1
to any single peaked profiles on trees whose path cover number is at most k. The idea is to partition
the tree into k disjoint paths, use the algorithm from Theorem 1 on each paths to obtain an order
of the candidates on each path of the partition, and finally merge these suborders intelligently. We
now formalize this idea as follows.
Theorem 2 There is a Preference Elicitation algorithm with query complexity O(mn log k)
for profiles that are single peaked on trees with path cover number at most k.
Proof: Since the path cover number is at most k, We can partition the tree T = (C, E) into t disjoint
paths Pi = (Ci , Ei ), i ∈ [t], where t is at most k. We now show that we can elicit any preference
≻ which is single peaked on the tree T by making O(m log t) queries which in turn proves the
statement. We first find the preference ordering restricted to Ci using Theorem 1 by making
O(∣Ci ∣) queries for every i ∈ [t]. This step needs ∑i∈[t] O(∣Ci ∣) = O(m) queries since Ci , i ∈ [t] forms
a partition of C. We next merge the t orders ≻ (Ci ), i ∈ [t], to obtain the complete preference ≻ by
using a standard divide and conquer approach for t-way merging which makes O(m log t) queries
[HUA83]. Thus the query complexity of the algorithms is O(m + m log t) = O(m log k).
◻
Towards our results on leaves, we show that the path cover of a tree with ` leaves is at least
⌊`/2⌋ and at most `. The lower bound is easy since every path can account for at most two leaves.
The upper bound comes from a careful partitioning of the nodes following maximal paths from
leaves up to a root and using a careful marking scheme. We defer the details of this argument to
a full version. Using the fact that a tree with ` leaves can be partitioned into at most ` paths, we
can use the algorithm from Theorem 2 to obtain the following bound in terms of leaves.
Initially we have t orders to merge. We arbitrarily pair the t orders into ⌈t/2⌉ pairs with at most
one of them being singleton (when t is odd). By renaming, suppose the pairings are as follows:
(≻ (C2i−1 ), ≻ (C2i )), i ∈ [⌊t/2⌋]. Let us define Ci′ = C2i−1 ∪ C2i for every i ∈ [⌊t/2⌋] and C⌈′t/2⌉ = Ct if t is an
odd integer. We merge ≻ (C2i−1 ) and ≻ (C2i ) to get ≻ (Ci′ ) for every i ∈ [⌊t/2⌋] using Observation 2.
The number queries the algorithm makes in this iteration is ∑i∈[⌊t/2⌋] O(∣C2i−1 ∣ + ∣C2i ∣) = O(m) since
(Ci )i∈[t] forms a partition of C. At the end of the first iteration, we have ⌈t/2⌉ orders ≻ (Ci′ ), i ∈ [⌈t/2⌉]
to merge to get ≻. The algorithm repeats the step above O(log t) times to obtain ≻ and the query
complexity of each iteration is O(m).
Corollary 1 There is a Preference Elicitation algorithm with query complexity O(mn log `)
for profiles that are single peaked on trees with at most ` leaves.
Finally, if we are given a subset of vertices whose removal makes the given tree a path, then
we have an elicitation algorithm that makes O(mn + nd log d). As before, we can determine the
ordering among the candidates on the path with O(m − d) queries, and we determine the ordering
among the rest in O(d log d) queries using Observation 1, and finally merge using Observation 2.
This leads us to the following.
Theorem 3 There is a Preference Elicitation algorithm with query complexity O(mn +
nd log d) for profiles that are single peaked on trees with distance d from path.
Proof: Let X be the smallest set of nodes of a tree T = (C, E) such that T ∖ X , the subgraph of
T after removal of the nodes in X , is a path. We have ∣X ∣ ≤ d. For any preference ≻, we make
O(d log d) queries to find ≻ (X ) using Observation 1, make O(∣C ∖ X ∣) = O(m − d) queries to
find ≻ (C ∖ X ) using Theorem 1, and finally make O(m) queries to find ≻ by merging ≻ (X ) and
≻ (C ∖ X ) using Observation 2. This gives an overall query complexity of O(mn + nd log d).
◻
3.2
Lower Bounds
We now turn to the lower bounds. Our first result is based on a counting argument, showing that
the query complexity in terms of the number of leaves, given by Corollary 1, is tight up to constant
factors. Indeed, let us consider a subdivision of a star with ` leaves and let t denote the distance
from the center, so that we have a total of t` + 1 vertices. One can show that if the candidates are
written out in level order, the candidates that are distance i from the star can be ordered arbitrarily
within this ordering. This tells us that the number of possible preferences ≻ that are single peaked
on the tree T is at least (`!)t . We obtain the lower bound by using a decision tree argument,
wherein we are able to show that it is always possible for an oracle answering the comparison
queries to answer in such a way that the total space of possibilities decreases by at most a factor
of half. Since the tree must entertain at least (`!)t leaves to account for all possibilities, we obtain
the claimed lower bound.
Theorem 4 Let T = (C, E) be a balanced subdivision of a star with ` leaves. Then any Preference Elicitation algorithm for single peaked profiles on T has query complexity Ω(mn log `).
Proof: Suppose the number of candidates m be (t` + 1) for some integer t. Let c be the center
of T . We denote the shortest path distance between any two nodes x, y ∈ C in T by d(x, y). We
consider the partition (C0 , . . . , Ct ) of the set of candidates C where Ci = {x ∈ C ∶ d(x, c) = i}. We
claim that the preference ≻= π0 ≻ π1 ≻ ⋯ ≻ πt of the set of candidates C is single peaked on the
tree T where πi is any arbitrary order of the candidates in Ci for every 0 ≤ i ≤ t. Indeed; consider
any path Q = (X , E ′ ) in the tree T . Let y be the candidate closest to c among the candidates in
X ; that is y = argminx∈X d(x, c). Then clearly ≻ (X ) is single peaked with respect to the path Q
having peak at y. We have ∣Ci ∣ = ` for every i ∈ [t] and thus the number of possible preferences ≻
that are single peaked on the tree T is at least (`!)t .
Let A be any Preference Elicitation algorithm for single peaked profiles on the tree T .
We now describe our oracle to answer the queries that the algorithm A makes. For every voter
v, the oracle maintains the set Rv of possible preferences of the voter v which is consistent with
the answers to all the queries that the algorithm A have already made for the voter v. At the
beginning, we have ∣Rv ∣ ≥ (`!)t for every voter v as argued above. Whenever the oracle receives a
query on v for any two candidates x and y, it computes the numbers n1 and n2 of orders in Rv
which prefers x over y and y over x respectively; the oracle can compute the integers n1 and n2
since the oracle has infinite computational power. The oracle answers that the voter v prefers the
candidate x over y if and only if n1 ≥ n2 and updates the set Rv accordingly. Hence, whenever the
oracle is queried for a voter v, the size of the set Rv decreases by a factor of at most two. On the
other hand, we must have, from the correctness of the algorithm, Rv to be a singleton set when
the algorithm terminates for every voter v – otherwise there exists a voter v (whose corresponding
Rv is not singleton) for which there exist two possible preferences which are single peaked on the
tree T and are consistent with all the answers the oracle has given and thus the algorithm A
fails to output the preference of the voter v correctly. Hence every voter must be queried at least
Ω(log((`!)t )) = Ω(t` log `) = Ω(m log `) times.
◻
Since the path cover number of a subdivided star on ` leaves is at least `/2, we also have the
following.
Corollary 2 There exists a tree T with path cover number k such that any Preference Elicitation algorithm for single peaked profiles on T has query complexity Ω(mn log k).
Mimicking the level order argument above on a generic tree with ` leaves, and using the connection between path cover and leaves, we obtain lower bounds that are functions of (n, `) and (n, k),
as given below. This will be useful for our subsequent results.
Theorem 5 Let T = (C, E) be any arbitrary tree with ` leaves and path cover number k. Then
any Preference Elicitation algorithm for single peaked profiles on T has query complexity
Ω(n` log `) and Ω(nk log k).
Proof: Let X be the set of leaves in T . We choose any arbitrary nonleaf node r as the root of T .
We denote the shortest path distance between two candidates x, y ∈ C in the tree T by d(x, y). Let
t be the maximum distance of a node from r in T ; that is t = maxy∈C∖X d(r, x). We partition the
candidates in C ∖ X as (C0 , C1 , . . . , Ct ) where Ci = {y ∈ C ∖ X ∶ d(r, y) = i} for 0 ≤ i ≤ t. We claim
that the preference ≻= π0 ≻ π1 ≻ ⋯ ≻ πt ≻ π of the set of candidates C is single peaked on the tree
T where πi is any arbitrary order of the candidates in Ci for every 0 ≤ i ≤ t and π is an arbitrary
order of the candidates in C ∖ X . Indeed, otherwise consider any path Q = (Y, E ′ ) in the tree T .
Let y be the candidate closest to r among the candidates in Y; that is y = argminx∈Y d(x, r). Then
clearly ≻ (Y) is single peaked with respect to the path Q having peak at y. We have the number
of possible preferences ≻ that are single peaked on the tree T is at least ∣X ∣! = `!. Again using
the oracle same as in the proof of Theorem 4, we deduce that any Preference Elicitation
algorithm A for profiles that are single peaked on the tree T needs to make Ω(n` log `) queries.
The bound with respect to the path cover number now follows from Lemma 2.
◻
The following results can be obtained simply by applying Theorem 5 on particular graphs. For
instance, we use the fact that stars have (m − 1) leaves and have pathwidth one to obtain the first
part of Corollary 3, while appealing to complete binary trees that have O(m) leaves and pathwidth
O(log m) for the second part. These examples also work in the context of maximum degree, while
for diameter we use stars and caterpillars with a central path of length m/2.
We next consider the parameter pathwidth of the underlying single peaked tree. We immediately
get the following result for Preference Elicitation on trees with pathwidths one or log m from
Theorem 5 and the fact that the pathwidths of a star and a complete binary tree are one and log m
respectively.
Corollary 3 There exist two trees T and T ′ with pathwidths one and log m respectively such that
any Preference Elicitation algorithm for single peaked profiles on T and T ′ respectively has
query complexity Ω(mn log m).
Corollary 4 There exist two trees T and T ′ with maximum degree ∆ = 3 and m−1 respectively such
that any Preference Elicitation algorithm for single peaked profiles on T and T ′ respectively
has query complexity Ω(mn log m).
Proof: Using Theorem 5, we know that any Preference Elicitation algorithm for profiles
which are single peaked on a complete binary tree has query complexity Ω(mn log m) since a
complete binary tree has Ω(m) leaves. The result now follows from the fact that the maximum
degree ∆ of a node is three for any binary tree. The case of ∆ = m − 1 follows immediately from
Theorem 5 applied on stars.
◻
Corollary 5 There exists two trees T and T ′ with diameters ω = 2 and ω = m/2 respectively such
that any Preference Elicitation algorithm for profiles which are single peaked on T and T ′
respectively has query complexity Ω(mn log m).
Proof: The ω = 2 and ω = m/2 cases follow from Theorem 5 applied on star and caterpillar graphs
with a central path of length m/2 respectively.
◻
Our final result, which again follows from Theorem 5 applied of caterpillar graphs with a central
path of length m − d, shows that the bound in Theorem 3 is tight.
Theorem 6 For any integers m and d with 1 ≤ d ≤ m/4, there exists a tree T with distance d from
path such that any Preference Elicitation algorithm for profiles which are single peaked on T
has query complexity Ω(mn + nd log d).
Proof: Consider the caterpillar graph where the length of the central path Q is m − d; there exists
such a caterpillar graph since d ≤ m/4. Consider the order ≻= π ≻ σ of the set of candidates C
where π is an order of the candidates in Q which is single peaked on Q and σ is any order of the
candidates in C ∖ Q. Clearly, ≻ is single peaked on the tree T . Any elicitation algorithm A needs to
make Ω(m − d) queries involving only the candidates in Q to elicit π due to [Con09] and Ω(d log d)
queries to elicit σ due to sorting lower bound for every voter. This proves the statement.
◻
4
Weak Condorcet Winner
We now show that we can find a weak Condorcet winner of profiles that are single peaked on trees
using fewer queries than the number of queries needed to find the profile itself. First, note that if
a Condorcet winner is guaranteed to exist, then it can be found using O(mn) queries — we pit an
arbitrary pair of candidates x, y and use O(n) queries to determine if x defeats y. We push the
winning candidate forward and repeat the procedure, clearly requiring at most m rounds. Now,
if a profile is single peaked with respect to a tree, and there are an odd number of voters, then
we have a Condorcet winner and the procedure that we just described would work. Otherwise, we
simply find a Condorcet winner among the first (n − 1) voters. It can be shown that such a winner
is one of the weak Condorcet winners for the overall profile, and we therefore have the following
upper bound.
We begin with the following general observation.
Observation 3 Let P be a profile where a Condorcet winner is guaranteed to exist. Then we can
find the Condorcet winner of P by making O(mn) queries.
Proof: For any two candidates x, y ∈ C we find whether x defeats y or not by simply asking all
the voters to compare x and y; this takes O(n) queries. The algorithms maintains a set S of
candidates which are potential Condorcet winners. We initialize S to C. In each iteration we pick
any two candidates x, y ∈ S from S, remove x from S if x does not defeat y and vice versa using
O(n) query complexity until S is singleton. After at most m − 1 iterations, the set S will be
singleton and contain only Condorcet winner since we find a candidate which is not a Condorcet
winner in every iteration and thus the size of the set S decreases by at least one in every iteration.
This gives a query complexity bound of O(mn).
◻
Using Observation 3 we now develop a Weak Condorcet Winner algorithm with query
complexity O(mn) for profiles that are single peaked on trees.
Theorem 7 There is a Weak Condorcet Winner algorithm with query complexity O(mn) for
single peaked profiles on trees.
For the special case of single peaked profiles, we can do even better. Here we take advantage of
the fact that a “median candidate” [MCWG+ 95] is guaranteed to be a weak Condorcet winner. We
make O(log m) queries per vote to find the candidates placed at the first position of all the votes
using the algorithm in [Con09] and find a median candidate to show the following.
if a profile is single peaked (on a path), then there is a Weak Condorcet Winner algorithm
with query complexity O(n log m) as shown below. Let us define the frequency f (x) of a candidate
x ∈ C to be the number of votes where x is placed at the first position. Then we know that a median
candidate according to the single peaked ordering of the candidates along with their frequencies as
defined above is a weak Condorcet winner for single peaked profiles [MCWG+ 95].
Theorem 8 There is a Weak Condorcet Winner algorithm with query complexity O(n log m)
for single peaked profiles (on a path).
Proof: Let P be a profile that is single peaked with respect to an ordering ≻∈ L(C) of candidates.
Then we find, for every voter v, the candidate the voter v places at the first position using
O(log m) queries using the algorithm in [Con09] and return a median candidate.
◻
Proof: Let P = (≻i )i∈[n] be a profile which is single peaked on a tree T . If n is an odd integer, then
we know that there exists a Condorcet winner in P since no two candidates can tie and there always
exists at least one weak Condorcet winner in every single peaked profile on trees. Hence, if n is an
odd integer, then we use Observation 3 to find a weak Condorcet winner which is the Condorcet
winner too. Hence let us now assume that n is an even integer. Notice that P−1 = (≻2 , . . . , ≻n ) is
also single peaked on T and has an odd number of voters and thus has a Condorcet winner. We
use Observation 3 to find the Condorcet winner c of P−1 and output c as a weak Condorcet winner
of P. We claim that c is a weak Condorcet winner of P. Indeed otherwise there exists a candidate
x other than c who defeats c in P. Since n is an even integer, x must defeat c by a margin of at
least two (since all pairwise margins are even integers) in P. But then x also defeats c by a margin
of at least one in P−1 . This contradicts the fact that c is the Condorcet winner of P−1 .
◻
The next result uses Theorem 8 on paths in a path cover eliminating the case of even number of
voters by the idea of setting aside one voter that was used in Theorem 7.
Theorem 9 Let T be a tree with path cover number at most k. Then there is an algorithm for
Weak Condorcet Winner for profiles which are single peaked on T with query complexity
O(nk log m).
Recalling that the number of leaves bounds the path cover number, we have the following
consequence.
Corollary 6 Let T be a tree with ` leaves. Then there is an algorithm for Weak Condorcet
Winner for profiles which are single peaked on T with query complexity O(n` log m).
We now state the lower bounds pertaining to Weak Condorcet Winner. First, we show
that any algorithm for single peaked profiles on stars has query complexity Ω(mn), showing that
the bound of Theorem 7 is tight.
Theorem 10 Any Weak Condorcet Winner algorithm for single peaked profiles on stars must
have query complexity Ω(mn).
Proof: Let T be a star with center vertex c. We now design an oracle that will “force” any Weak
Condorcet Winner algorithm A for single peaked profiles on T to make Ω(mn) queries. For
every voter v, the oracle maintains a set of “marked” candidates which can not be placed at the
first position of the preference of v. Suppose the oracle receives a query to compare two candidates
x and y for a voter `. If the order between x and y for the voter ` follows from the answers the
oracle has already provided to all the queries for the voter `, then the oracle answers accordingly.
Otherwise it answers x ≻` y if y is unmarked and marks y; otherwise the oracle answers y ≻` x and
marks x. Notice that the oracle marks at most one unmarked candidate every time it is queried.
We now claim that there must be at least n/10 votes which have been queried at least m/4 times.
If not, then there exists n − n/10 = 9n/10 votes each of which has at least m − m/4 = 3m/4 candidates
unmarked. In such a scenario, there exists a constant N0 such that for every m, n > N0 , we have
at least two candidates x and y who are unmarked in at least (⌊n/2⌋ + 1) votes each. Now if the
algorithm outputs x, then we put y at the first position in at least (⌊n/2⌋ + 1) votes and at the
second position in the rest of the votes and this makes y the (unique) Condorcet winner. If the
algorithm does not output x, then we put x at the first position in at least (⌊n/2⌋ + 1) votes and
at the second position in the rest of the votes and this makes x the (unique) Condorcet winner.
Hence the algorithm fails to output correctly in both the cases contradicting the correctness
of the algorithm. Also the resulting profile is single peaked on T with center at y in the first
case and at x in the second case. Therefore the algorithm A must have query complexity Ω(mn). ◻
Our concluding result uses an intricate adversary argument, and shows that the query complexity for Weak Condorcet Winner for single peaked profiles in Theorem 8 is essentially optimal,
provided that the queries to different voters are not interleaved, as is the case with our algorithm.
Theorem 11 Any Weak Condorcet Winner algorithm for single peaked profiles which does
not interleave the queries to different voters has query complexity Ω(n log m).
Proof: Let a profile P be single peaked with respect to the ordering of the candidates
≻= c1 ≻ c2 ≻ ⋯ ≻ cm . The oracle maintains two indices ` and r for every voter such that any
candidate from {c` , c`+1 , . . . , cr } can be placed at the first position of the preference of the voter
v and still be consistent with all the answers provided by the oracle for v till now and single
peaked with respect to ≻ . The algorithm initializes ` to one and r to m for every voter. The
oracle answers any query in such a way that maximizes the new value of r − `. More specifically,
suppose the oracle receives a query to compare candidates ci and cj with i < j for a voter v.
If the ordering between ci and cj follows, by applying transitivity, from the answers to the
queries that have already been made so far for this voter, then the oracle answers accordingly.
Otherwise the oracle answers as follows. If i < `, then the oracle answers that cj is preferred over
ci ; else if j > r, then the oracle answers that ci is preferred over cj . Otherwise (that is when
` ≤ i < j ≤ r), if j − ` > r − i, then the oracle answers that ci is preferred over cj and changes r to
j; if j − ` ≤ r − i, then the oracle answers that cj is preferred over ci and changes ` to i. Hence
whenever the oracle answers a query for a voter v, the value of r − ` for that voter v decreases by
a factor of at most two. Suppose the election instance has an odd number of voters. Let V be
the set of voters. Now we claim that the first ⌊n/5⌋ voters must be queried (log m − 1) times each.
Suppose not, then consider the first voter v ′ that is queried less than (log m − 1) times. Then
there exist at least two candidates ct and ct+1 each of which can be placed at the first position of
the vote v ′ . The oracle fixes the candidates at the first positions of all the votes that have not
been queried till v ′ is queried (and there are at least ⌈4n/5⌉ such votes) in such a way that ⌊n/2⌋
voters in V ∖ {v ′ } places some candidate in the left of ct at the first positions and ⌊n/2⌋ voters in
V ∖ {v ′ } places some candidate in the right of ct+1 . If the algorithm outputs ct as the Condorcet
winner, then the oracle makes ct+1 the (unique) Condorcet winner by placing ct+1 at the top
position of v ′ , because ct+1 is the unique median in this case. If the algorithm does not output
ct as the Condorcet winner, then the oracle makes ct the (unique) Condorcet winner by placing
ct at the top position of v ′ , because ct is the unique median in this case. Hence the algorithm
fails to output correctly in both the cases thereby contradicting the correctness of the algorithm. ◻
From Theorem 7 and 8 we have the following result for any arbitrary tree.
Theorem 12 Let T be a tree with path cover number at most k. Then there is an algorithm
for Weak Condorcet Winner for profiles which are single peaked on T with query complexity
O(nk log m).
Proof: Let P be the input profile and Qi = (Xi , Ei ) i ∈ [t] be t(≤ k) disjoint paths that cover
the tree T . Here again, if the number of voters is even, then we remove any arbitrary voter and
the algorithm outputs the Condorcet winner of the rest of the votes. The correctness of this
step follows from the proof of Theorem 7. Hence we assume, without loss of generality, that
we have an odd number of voters. The algorithm proceeds in two stages. In the first stage, we
find the Condorcet winner wi of the profile P(Xi ) for every i ∈ [t] using Theorem 8. The query
complexity of this stage is O(n ∑i∈[t] log ∣Xi ∣) = O(nt log(m/t)). In the second stage, we find the
Condorcet winner w of the profile P({wi ∶ i ∈ [t]}) using Theorem 7 and output w. The query
complexity of the second stage is O(nt log t). Hence the overall query complexity of the algorithm
is O(nt log(m/t)) + O(nt log t) = O(nk log m).
◻
5
Conclusions and Future Work
We show algorithms for preference elicitation for profiles which are single peaked on trees. Moreover,
we prove that the query complexity of our algorithms are optimal up to constant factors. We also
show that we need to query fewer number of times than preference elicitation if we only want to
find any weak Condorcet winner.
In this work we do not assume any partial information about the preferences. However, in
many scenarios, we may have some knowledge about the profile. An interesting future direction
of research is to study how having some partial information helps us reduce query complexity of
preference elicitation.
Acknowledgement
Palash Dey wishes to gratefully acknowledge support from Google India for providing him with a
special fellowship for carrying out his doctoral work.
References
[Arr50]
Kenneth J Arrow. A difficulty in the concept of social welfare. The Journal of Political
Economy, pages 328–346, 1950.
[BBHH15]
Felix Brandt, Markus Brill, Edith Hemaspaandra, and Lane A Hemaspaandra. Bypassing combinatorial protections: Polynomial-time algorithms for single-peaked electorates. Journal of Artificial Intelligence Research (JAIR), pages 439–496, 2015.
[BCE+ 15]
Felix Brandt, Vincent Conitzer, Ulle Endriss, Jérôme Lang, and Ariel Procaccia.
Handbook of computational social choice, 2015.
[Bla48]
Duncan Black. On the rationale of group decision-making. The Journal of Political
Economy, pages 23–34, 1948.
[Con09]
Vincent Conitzer. Eliciting single-peaked preferences using comparison queries. Journal of Artificial Intelligence Research (JAIR), 35:161–191, 2009.
[Cor09]
Thomas H Cormen. Introduction to algorithms. MIT press, 2009.
[CS02]
Vincent Conitzer and Tuomas Sandholm. Vote elicitation: Complexity and strategyproofness. In Eighteenth National Conference on Artificial Intelligence (AAAI), pages
392–397, 2002.
[CS05]
Vincent Conitzer and Tuomas Sandholm. Communication complexity of common
voting rules. In Proceedings of the 6th ACM conference on Electronic Commerce
(EC), pages 78–87. ACM, 2005.
[Dem82]
Gabrielle Demange. Single-peaked orders on a tree. Mathematical Social Sciences,
3(4):389–396, 1982.
[DL13]
Ning Ding and Fangzhen Lin. Voting with partial information: what questions to
ask? In Proceedings of the 12th International Conference on Autonomous Agents
and Multi-agent Systems (AAMAS), pages 1237–1238. International Foundation for
Autonomous Agents and Multiagent Systems, 2013.
[DM16]
Palash Dey and Neeldhara Misra. Preference elicitation for single crossing domain.
In IJCAI 2016, Proceedings of the 25nd International Joint Conference on Artificial
Intelligence (IJCAI), 2016.
[Gib73]
Allan Gibbard. Manipulation of voting schemes: a general result. Econometrica:
Journal of the Econometric Society, pages 587–601, 1973.
[Hei93]
Katherine Heinrich. Path decomposition. Le matematiche, 47(2):241–258, 1993.
[HM97]
Mel vin J Hinich and Michael C Munger. Analytical politics. Cambridge University
Press, 1997.
[HUA83]
John E Hopcroft, Jeffrey David Ullman, and Alfred Vaino Aho. Data structures and
algorithms, volume 175. Addison-Wesley Boston, MA, USA:, 1983.
[LB11a]
Tyler Lu and Craig Boutilier. Robust approximation and incremental elicitation
in voting protocols. In IJCAI 2011, Proceedings of the 22nd International Joint
Conference on Artificial Intelligence (IJCAI), pages 287–293, 2011.
[LB11b]
Tyler Lu and Craig Boutilier. Vote elicitation with probabilistic preference models:
Empirical estimation and cost tradeoffs. In Algorithmic Decision Theory, pages 135–
149. Springer, 2011.
[MCWG+ 95] Andreu Mas-Colell, Michael Dennis Whinston, Jerry R Green, et al. Microeconomic
theory, volume 1. Oxford university press New York, 1995.
[Mou91]
Hervi Moulin. Axioms of cooperative decision making. Number 15. Cambridge University Press, 1991.
[PE16]
Dominik Peters and Edith Elkind. Preferences single-peaked on nice trees. 2016.
[Sat75]
Mark Allen Satterthwaite. Strategy-proofness and arrow’s conditions: Existence and
correspondence theorems for voting procedures and social welfare functions. Journal
of Economic Theory, 10(2):187–217, 1975.
[Tri89]
Michael A Trick. Recognizing single-peaked preferences on a tree. Mathematical
Social Sciences, 17(3):329–334, 1989.
[YCE13]
Lan Yu, Hau Chan, and Edith Elkind. Multiwinner elections under preferences that
are single-peaked on a tree. In Proceedings of the Twenty-Third International Joint
Conference on Artificial Intelligence (IJCAI), pages 425–431. AAAI Press, 2013.
| 8cs.DS
|
Scheduling Two Agents on a Single Machine:
A Parameterized Analysis of NP-hard Problems⋆
Danny Hermelin1 , Judith-Madeleine Kubitza2 , Dvir Shabtay1 ,
arXiv:1709.04161v1 [cs.DS] 13 Sep 2017
Nimrod Talmon3 , and Gerhard Woeginger4
1
Ben Gurion University of the Negev, Israel
[email protected], [email protected]
2
TU Berlin, Germany
[email protected]
3
Weizmann Institute of Science, Israel
[email protected]
4
Eindhoven University of Technology, The Netherlands
[email protected]
Abstract. Scheduling theory is an old and well-established area in combinatorial optimization, whereas
the much younger area of parameterized complexity has only recently gained the attention of the community. Our aim is to bring these two areas closer together by studying the parameterized complexity
of a class of single-machine two-agent scheduling problems. Our analysis focuses on the case where the
number of jobs belonging to the second agent is considerably smaller than the number of jobs belonging
to the first agent, and thus can be considered as a fixed parameter k. We study a variety of combinations
of scheduling criteria for the two agents, and for each such combination we pinpoint its parameterized
complexity with respect to the parameter k. The scheduling criteria that we analyze include the total
weighted completion time, the total weighted number of tardy jobs, and the total weighted number of
just-in-time jobs. Our analysis draws a borderline between tractable and intractable variants of these
problems.
⋆
Preliminary version of this work appeared in the proceedings of the International symposium on Parameterized
and Exact Computation (IPEC) 2015.
1
Introduction
Scheduling is a well-studied area in operations research that provides fertile grounds for several combinatorial
problems. In a typical scheduling problem, we are given a set of jobs that are to be scheduled on a set of
machines which is arranged according to a specific machine setting. The objective is to determine a schedule
which minimizes a predefined scheduling criterion such as the makespan, total weighted completion time,
and total weighted tardiness of the schedule. There are various machine settings including the single machine
setting, parallel machines, flow-shop and job-shop, and each scheduling problem may in addition have various
attributes and constraints. We refer the reader to e.g. [6,19,28] for an extensive introduction to the area of
scheduling, and for a detailed survey of classical results.
Many scheduling problems are NP-hard. Typically, such hard problems include a multitude of parameters,
and many NP-hardness proofs exploit the fact that these parameters can be arbitrary large in theory.
However, in many practical settings, one or more of these parameters will actually be quite small. For
example, the number of different items that can be processed in the shop might be limited, resulting in
a scheduling instance where only a limited number of different processing times appear. A limited set of
planned delivery dates resulting in a scheduling instance with a limited number of different due dates is
another example. It is therefore natural to ask whether NP-hard scheduling problems become tractable
when some of their parameters can be assumed to be comparatively small in practice. Luckily, a framework
for answering such questions has been recently developed by the computer science community - the theory
of parameterized complexity.
Parameterized complexity facilitates the analysis of computational problems in terms of various instance
parameters that may be independent of the total input length. In this way, problem instances are analyzed
not only according to the total input length n, but also according to an additional numerical parameter
k that may encode other aspects of the input. A problem is considered tractable if there is an algorithm
that optimally solves any instance in f (k) · nO(1) time, where f () is allowed to be any arbitrary computable
function which is independent of n, and the exponent in nO(1) is required to be independent of k. For
example, a running-time of 2O(k) · n3 is considered tractable in the parameterized setting, while nO(k) is not.
2
In this way we can model scenarios where certain problem parameters are typically much smaller than the
total input length, yet may not be small enough to be considered constant.
Parameterized complexity has enjoyed tremendous success since its first developments in the early 90s,
as can be exemplified by the various textbooks on the subjects [7,10,8,25]. However, there are currently very
few papers that attack scheduling problems from the parameterized perspective [5,9,21,31,32]. This is rather
disappointing since scheduling problems seem to be particulary adequate for parameterized analyses. For one,
scheduling problems which are NP-hard lack polynomial-time algorithms for finding optimal solutions, and
in several applications, approximate solutions can result in big revenue losses. This gives strong motivation
for computing exact solutions even if computing such solutions requires a lot of resources. Secondly, as
argued above, scheduling problems typically have an abundance of natural problem parameters that can be
comparatively small in practice. Thus, algorithms whose running times grow exponentially in such parameters
alone should be quite useful for practical purposes.
Our aim in this paper is to help close the gap between research in parameterized complexity and the
area of scheduling. We initiate a parameterized analysis on problems occurring in the setting of multi-agent
scheduling [2], a contemporary area which is nowadays at the cutting edge of scheduling research. We focus
on the most basic case where there are only two agents, and all jobs are to be processed on a single machine.
Furthermore, our parameterized analysis focuses on the scenario where the second agent has a significantly
smaller number of jobs than the other. The number of jobs belonging to this agent is thus taken as a
parameter, and is denoted by k throughout the paper. We preform an extensive parameterized analysis
for several two-agent single-machine scheduling problems with respect to this parameter, providing a clear
picture of the applicability of parameterized algorithmics to these problems.
1.1
Our contribution
In this work we investigate a variety of combinations for the objective functions of each agent. For each such
combination, we consider the problem where each agent has a bound on his objective function, and the goal
is to determine whether there exists a single-machine schedule that meets both bounds simultaneously. The
objective functions we consider are (see Section 2 for formal definitions):
3
1. Total weighted completion time, where jobs have weights, and the goal is to minimize the sum of weighted
completion time over the entire job set of the agent.
2. Total weighted number of tardy jobs, where jobs have weights and due-dates, and the objective is to
minimize the total weighted number of jobs that terminate after their due-date.
3. Total weighted number of just-in-time ( JIT) jobs, where jobs have weights and due-dates, and the goal
is to maximize the total weighted number of jobs that terminate precisely on their due date.
We consider several combinations of these scheduling criteria for which the corresponding problem is NPcomplete, and for each such combination we determine whether or not the corresponding scheduling problem
becomes fixed-parameter tractable with respect to k. There are also other subtleties that we consider, such
as the unit weight case or the unit processing-time case. The paper is organized according to the first agent’s
scheduling criteria. Thus, Section 3 deals with the case where the scheduling criterion of agent 1 is the total
weighted completion time, Section 4 focuses on problems where the first agent criterion is the total weighted
number of tardy jobs, and Section 5 is concerned with problems where the first agent criterion is the total
weighted number of JIT jobs.
1.2
Related work
The set of two-agent scheduling problems was first introduced by Baker and Smith [4] and Agnetis et al. [3].
For different combinations of the scheduling criteria, Baker and Smith focus on analyzing the problem of
finding a schedule that minimizes the weighted sum of the two criteria, while Agnetis et al. focus on analyzing
the problem of minimizing the first agent criterion while keeping the value of the second agent criterion not
greater than a given bound. Following these two fundamental papers, numerous researchers have studied
different combinations of multi-agent scheduling problems, see e.g. [14,17,20,23,34,33]. Detailed surveys of
these problems appear in Perez-Gonzalez and Framinan [27] and in a recent book by Agnetis et al. [2]. We
give further detail of the results that are more directly related to our work in the appropriate sections of the
remainder of the paper.
4
2
Preliminaries
In this section we introduce the notation and terminology that will be used throughout the paper. In
particular, we provide concrete definitions for the problems we study, as well as a very brief introduction to
the theory of parameterized complexity.
2.1
Scheduling notation and problem definitions
In all problems considered in this paper, the input consists of two sets of jobs that have to be processed non(1)
(1)
preemptively on a single machine. The first set J (1) = {J1 , ..., Jn } belongs to agent 1, while the second
(2)
(2)
set J (2) = {J1 , ..., Jk } belongs to agent 2. We assume that k ≤ n, and for practical purposes one should
(i)
think of k as much smaller than n. Let pj
(i)
be a positive integer denoting the processing time of job Jj .
(i)
(i)
Moreover, when relevant, let dj and wj be two positive integers representing the due date and the weight
(i)
(i)
(i)
(i)
(i)
of job Jj , respectively. A schedule σ of J (1) ∪ J (2) is a set of disjoint time intervals Ij = (Cj − pj , Cj ]
(i)
for j = 1, ..., n if i = 1 and j = 1, ..., k if i = 2, where Ij
(i)
is processed on the single machine. Note that Cj
(i)
represents the completion time of Jj . In case, one or
(i)
(i)
(i)
(i)
(i)
(i)
(i)
≤ 0, then job Jj
(i)
= Cj − dj
both of the agents jobs have due dates, we will use Lj
and we set Lmax = maxj Lj . If Lj
(i)
represents the time interval in σ where job Jj
(i)
to denote the lateness of job Jj ,
is an early job in σ, and otherwise it is tardy.
(i)
Accordingly, the set E (i) = {Jj ∈ J (i) |Lj ≤ 0} is the set of early jobs in σ that belongs to agent i, and the
(i)
set T (i) = {Jj
(i)
∈ J (i) |Lj > 0} is the set of tardy jobs that belong to agent i. We also use Eb(i) to denote
(i)
(i)
the set Eb(i) = {Jj ∈ J (i) |Lj = 0}.
The quality of a schedule is measured by two different criteria, one per each agent. We focus on problems
for which either one of the two agents criteria may be either one of the following three possibilities:
1. The weighted sum of completion times, denoted by
(i)
2. The weighted number of tardy jobs, where job Jj
(i)
indicator variable Uj
P
(i)
(i)
is said to be tardy if Cj
(i)
which indicates whether or not Jj
number of tardy jobs of agent i.
5
(i)
wj Cj .
is tardy, and
P
(i)
(i)
> dj . We use a binary
(i)
wj Uj
denotes the weighted
(i)
(i)
(i)
3. The weighted number of just-in-time (JIT) jobs, where job Jj is said to be just-in-time if Cj = dj . We
(i)
use a binary indicator variable Ej
(i)
which indicates whether or not Jj
is just-in-time, and
denotes the weighted number of just-in-time jobs of agent i.
P
(i)
(i)
wj Ej
Note that while first two criteria are minimization criteria, the latter is a maximization criterion. For
each possible combination of the criteria above, we consider the decision problem where we are given two
positive integer bounds A1 and A2 , one for each agent, and we need to find if there exists a job schedule
in which both bounds are met. In case the scheduling criterion is the sum of weighted completion times or
weighted number of tardy jobs, the bound Ai is regarded as an upper-bound, while for weighted number of
just-in-time jobs it is a lower-bound. We refer to such a job schedule, if it exists, as a feasible job schedule.
Using the standard three field notation in scheduling, we denote this set of problems by 1 C(1) , C(2) , where
C(i) ∈
nX
(i)
(i)
wj Cj ≤ Ai ,
X
(i)
(i)
wj Uj
≤ Ai ,
X
o
(i) (i)
wj Ej ≥ Ai ,
for i = 1, 2. We will sometimes consider special cases of these problems, and when doing so we use the middle
(1)
field to denote restrictions on our input. For example, the 1 pj
= 1,
P
(1)
(1)
wj Cj
≤ A1 ,
P
(2)
Cj
≤ A2
problem is the problem where the scheduling criterion for the first agent is the weighted sum of completion
times, for the second agent is the (unweighed) sum of completion time and agent 1 has jobs with unit
processing times.
2.2
Basic concepts in parameterized complexity theory
The main objective in parameterized complexity theory is to analyze the tractability of NP-hard problems
with respect to input parameters that are not necessarily related to the total input size. Thus, problem
instances are not only measured in terms of their input size n, but also in terms of an additional parameter k.
In this context, a problem is said to be tractable, or fixed-parameter tractable (FPT), if there is an algorithm
that solves each instance of size n and parameter k in f (k) · nO(1) time. Here, the function f () can be any
arbitrary computable (e.g. exponential) function so long as it depends only on k, and the exponent in nO(1) is
independent of k. The reader is referred to the excellent texts on the subject for more information [7,10,8,25].
2O(k)
In parameterized complexity, a running-time of 2O(k) · n3 is considered tractable, and even 22
· n100 is
considered tractable. Note that while the above definition might allow some quite large running-times, when
6
k is sufficiently smaller than n, any such run-time drastically outperforms more common algorithms with
2
running-times of nO(k) or nO(k ) , for example. Moreover, a running time of, say, 2O(k) · n3 , with moderate
constants in the exponent can be quite fast in practice. In any case, parameterized complexity provides the
most convenient form of analyzing the complexity an NP-hard problem with respect to the size of a given
parameter. In our context, the parameter of each instance will always be the number of jobs of agent 2.
Thus, we consider the setting where agent 1 has significantly more jobs to schedule, but nevertheless we still
wish to meet both agents criteria.
Note that if a problem is NP-hard already for constant values of its parameter, then a fixed-parameter
tractable algorithm for the problem will imply that P=NP. Thus, in our context, if we show that one of
the problems we consider is already NP-hard when agent 2 has a constant number of jobs, this excludes the
possibility that the problem has a fixed-parameter tractable algorithm under the assumption of P6=NP. For
showing such hardness results, we will use the classical NP-complete Partition problem [11], often used in
the context of scheduling problems:
Definition 1 (The Partition problem). Given a set X = {x1 , . . . , xm } of positive integers (encoded in
binary) with
P
3
xj ∈S1
xj =
Pm
j=1
P
xj = 2z, determine whether X can be partitioned into two sets S1 and S2 such that
xj ∈S2
xj = z.
Weighted Sum of Completion Times
In this section we study the 1
P
(1)
(1)
wj Cj
≤ A1 , C(2) problem where C(2) can be any of the three scheduling
criteria discussed in Section 2. We first show that all the three corresponding problems are unlikely to admit
a fixed-parameter algorithm, since they are all NP-complete even for k = 1 (i.e, agent 2 has a single job)
as we show in Theorem 1. This motivates us to study four special cases: We show that in case the jobs of
agent 1 all have unit weight, the problem becomes FPT when the criteria for agent 2 is either weighted sum
of completion times or weighted number of tardy jobs. However, when the criteria of agent 2 is the number
of just-in-time jobs, the problem remains intractable in this case as well. We also provide an FPT algorithm
for the case where both criteria are the weighted sum of completion times, and agent 1 has jobs with unit
processing times.
7
3.1
Intractability of the general problem with respect to k
P
The fact that the single agent 1
wj Cj problem is solvable in O(n log n) time (see Smith [30]) gives
us some hope that at least one of the 1
P
(1)
(1)
wj Cj
≤ A1 , C(2)
problems is tractable when k is small.
Unfortunately, in the following theorem we show that this is not the case for all three criteria, even if the
second agent has a single job of a unit weight. We will show this via a reduction from the NP-complete
Partition problem (see Definition 1).
Theorem 1. The 1
P
(2)
Cj
≤ A2 ,
P
(2)
Uj
P
(1)
(1)
wj Cj
≤ A2 , or
P
≤ A1 , C(2)
(2)
Ej
problem is NP-complete for k = 1, when C(2) is either
≥ A2 .
Proof. We provide a reduction from the NP-complete Partition problem defined above. Given an instance
(X, z) to the Partition problem, with X = {x1 , . . . , xm }, we construct the following two-agent scheduling
instance: Agent 1 will have n = m jobs and agent 2 will have a single job (i.e., k = 1). For j = 1, ..., n, we
(1)
set pj
(1)
= wj
(2)
= xj . Moreover, we set p1 = 1, and in case C(2) ∈ {
(2)
P
(2)
Uj
≤ A2 ,
P
(2)
Ej
≥ A2 }, we also set
d1 = z + 1. The bound on the total weighted completion time of agent 1 is set to A1 = z +
Pn
i=1
Pi
j=1
xi xj .
The bound of agent 2 depends on his scheduling criterion: If it is the sum of completion times, we set
A2 = z + 1, if it is the weighted number of tardy jobs we set A2 = 0, and if it is the weighted number of JIT
jobs we set A2 = 1.
Now suppose that X can be partitioned into two sets S1 and S2 with
P
xi ∈S1
xi =
P
xi ∈S2
xi = z.
We construct a schedule σ where we first schedule all agent’s 1 jobs corresponding to elements of S1 in an
(2)
arbitrary order, followed by job J1 , followed by all of the jobs of agents 1 corresponding to the elements
(2)
of S2 in an arbitrary order. Observe that job J1
completes in σ after
P
xi ∈S1
xi + 1 = z + 1 time units,
and so the bound of agent 2 is met in all three criteria C(2) . To see that the first agent bound is met as
(2)
well, observe that if we exclude J1
precisely
Pn
i=1
Pi
j=1
from σ then the total weighted completion time of agent 1 jobs is
(2)
xi xj . Adding job J1
increases the completion time of each of the first agent jobs that
(2)
correspond to elements of S2 by a unit. Thus, J1
contributes precisely
P
xi ∈S2
xi = z to the total weighted
completion time of agent 1 jobs, and so the first agent bound on the total weighted completion time is met
as well.
8
(1)
For the other direction, suppose there is a feasible schedule σ for any possible option of C(2) . Let J1
(2)
note the set of agent 1 jobs that are scheduled before job J1
(1)
in σ, and let J2
denote agent 1 remaining jobs.
Since agent 2 bound is satisfied in σ, in each of the three possible criteria it must be that
Moreover, note that the total weighted completion time of agent 1 jobs is
Since agent 1 bound is also met by σ, it must be that
of agent 1 jobs is 2z, we get that
(1)
and S2 = {xj : Jj
3.2
(1)
P
(1)
(1)
Jj
pj
(1)
∈J1
=
P
P
(1)
(1)
Jj
∈J2
(1)
(1)
(1)
pj
(1)
Jj ∈J2
pj
Pn
i=1
Pi
j=1
P
(1)
(1)
Jj
xi xj +
(1)
∈J1
P
pj
P
(1)
Cj
≤ A1 ,
P
(2)
≤ z.
(1)
(1)
Jj
(1)
∈J2
pj .
≤ z. Since the sum of all processing times
(1)
= z. Thus, setting S1 = {xj : Jj
(1)
∈ J1 }
∈ J2 } yields a solution to our Partition instance.
An FPT algorithm for the 1
de-
(2)
wj C j
⊓
⊔
problem
≤ A2
In stark contrast to the result in Theorem 1, we next show that, although being NP-complete (see Agnetis
et al. [3]), the 1
P
(1)
Cj
≤ A1 ,
P
(2)
(2)
wj Cj
≤ A2
problem is much easier to handle. We present an FPT
algorithm for this problem for parameter k, using the powerful result of Lenstra concerning mixed integer
linear programs [18]. To begin with, we will need the following lemma which can be easily derived by using
a simple pair-wise interchange argument (see also Agnetis et al. [3] that prove the same argument for the
less general case where both agents jobs have unit weights):
Lemma 1. If there is a feasible solution for the 1
P
(1)
Cj
≤ A1 ,
P
(2)
(2)
wj Cj
≤ A2
problem, then there
(1)
exists a feasible solution where the jobs of agent 1 are scheduled in a non-decreasing order of pj , i.e.,
according to the shortest processing time (SPT) rule.
Consider now the 1
P
(1)
Cj
≤ A1 ,
P
(2)
(2)
wj Cj
≤ A2 problem. Due to Lemma 1, we assume, without loss
(1)
(1)
(1)
of generality that the jobs of agent 1 are numbered according to the SPT rule such that p1 ≤ p2 ≤ ... ≤ pn .
By allowing an additional multiplicative factor of k! to the running time of our algorithm, we can focus on
a reduced subproblem where the ordering of the second agent jobs is predefined. Given a subproblem, we
renumber the second agent jobs according to this ordering. Then, to determine whether a feasible schedule
is actually possible for the given subproblem, we only need to figure out if it is possible to interleave the
two ordered sets of jobs together in a way that satisfies both agents bounds. Towards this aim, we formalize
any given subproblem as a mixed integer linear program (MILP) where the number of integer variables is
k. We then complete the proof by using the celebrated result of Lenstra [18] which states that determining
9
whether a given MILP has a feasible solution is fixed-parameter tractable with respect to the number of
integer variables.
(2)
To formulate a given subproblem as an MILP, we define an integer variable xj for each job Jj
represent-
(2)
ing the number of jobs belonging to agent 1 that are scheduled before Jj . Therefore, for each 1 ≤ j ≤ k,
we add the constraint that
0 ≤ xj ≤ n.
(1)
Moreover, for 1 ≤ j ≤ k − 1, we add the constraint that
xj ≤ xj+1 .
(2)
Lemma 2. The bound on the total completion time of the first agent jobs can be formulated by the following
constraint:
n
X
j=1
k
X
(2)
(1)
(n − j + 1) · pj + (n − xj ) · pj ≤ A1 .
(3)
j=1
Proof. Observe that the first term in the left-hand side of constraint (3) is precisely the total completion
time of the first agent jobs when no job of the second agent is scheduled at all. We now add the second
agent jobs according to the intended meaning of the variables x1 , . . . , xk . If there are xj jobs of agent 1
(2)
scheduled prior to Jj
(2)
in the presumed schedule, then Jj
(2)
causes an increase of pj
to the completion time
of n − xj jobs belonging to the first agent. Thus, adding all of agent 2 jobs causes an additional increase of
Pk
j=1 (n
(2)
− xj ) · pj
⊓
⊔
to the total completion time of the first agent jobs.
(2)
The encoding of the second agent bound is a bit more involved. Specifically, for each Jj , we introduce
a real-valued variable yj which we would like to be equal to the contribution of the first agent jobs to the
(2)
completion time of Jj . Note that by our intended meaning for variable xj , this is precisely
Pxj
i=1
(1)
pi .
However, we cannot encode this directly as a linear constraint. We therefore introduce n additional real(2)
valued variables corresponding to Jj , denoted as yij for i ∈ {1, . . . , n}, which are ensured to be non-negative
by adding the constraint that
yij ≥ 0
10
(4)
for i ∈ {1, . . . , n} and j ∈ {1, . . . , k}. The yij variables are used to provide upper-bounds to the “steps” in
the contribution of the first agent jobs as depicted in Fig. 1. Accordingly, we add the constraints
(1)
yij ≥ (xj − i + 1) · (pi
(1)
− pi−1 )
(5)
(1)
for each i ∈ {1, . . . , n} and j ∈ {1, . . . , k} (naturally, we set here p0 = 0). Furthermore, we add the constraint
that
yj ≥
n
X
(6)
yij
i=1
for any j ∈ {1, . . . , k} so that yj will equal its intended meaning.
p
(1)
yij ≥ (xj − i + 1) · (pi
(1)
− pi−1 )
(2)
(1) (1)
Jj
J1 J2
(2)
Fig. 1. The contribution of xj jobs that of agent 1 which are scheduled prior to Jj
of this job by
Px j
i=1
increases the completion time
(1)
pi . The variable yij is intended to capture the i’th “step” of this contribution.
Lemma 3. The bound on the total weighted completion time of the second agent jobs can be formulated by
the following constraint:
k
X
i=1
(2)
wi
·
i
X
j=1
k
X
(2)
(2)
wj · yj ≤ A2 .
pj +
(7)
j=1
Proof. Observe that the first term in the left-hand side of constraint (7) is the total weighted completion
(2)
(2)
time of the second agent ordered jobs J1 , . . . , Jk , assuming no jobs of agent 1 are scheduled. We argue
that the second term upper-bounds the contribution of agent 1 jobs. For this, it suffices to show that the
(2)
contribution of agent 1 jobs to the completion time of Jj , for each j ∈ {1, . . . , k}, is at most yj . We know
that this contribution is
Pxj
i=1
(1)
pi . Since we are concerned only with feasible solutions where the constraints
11
on variables yi , yi1 . . . , yin are met, we have
yj ≥
n
X
yij ≥
i=1
n
X
(1)
max{0, (xj − i + 1) · (pi
(1)
− pi−1 )} =
xj
X
(1)
(xj − i + 1) · (pi
(1)
− pi−1 )
i=1
i=1
(1)
(1)
(1)
(1)
= xj p1 + (xj − 1)(p2 − p1 ) + · · · + (p(1)
xj − pxj −1 ) =
xj
X
(1)
pi .
i=1
⊓
⊔
To summarize, due to Lemma 1, we can solve the 1
P
(1)
Cj
≤ A1 ,
P
(2)
(2)
wj Cj
≤ A2
problem by
solving O(k!) MILP formulations, each of which has only k integer variables. Correctness of each of these
formulations follows from Lemmas 2 and 3, and the analysis above. Using Lenstra’s result [18], we therefore
obtain the following theorem:
Theorem 2. 1
3.3
P
(1)
Cj
≤ A1 ,
P
(2)
(2)
wj Cj
(1)
An FPT algorithm for 1 pj
≤ A2
= 1,
P
is fixed-parameter tractable with respect to k.
(1)
(1)
wj C j
≤ A1 ,
P
(2)
(2)
wj C j
≤ A2
The problem of determining whether there exists a schedule where both agents meet their respective bounds
on their total weighted completion times was shown to be NP-complete even if the jobs of both agents all
have unit processing times [26]. Here we complement this result, as well as the result given in Theorem 1,
by showing that the problem is FPT with respect to k when the jobs of the first agent have unit processing
times. The following lemma is crucial for the construction of the FPT algorithm, and can be proven by a
simple pairwise interchange argument.
(1)
Lemma 4. If there is a feasible schedule for the 1 pj
= 1,
P
(1)
(1)
wj Cj
≤ A1 ,
P
(2)
(2)
wj Cj
≤ A2
problem,
then there is a feasible schedule for the problem where the first agent jobs are scheduled in a non-increasing
weight order.
(1)
Assume, without loss of generality, that w1
(1)
≥ · · · ≥ wn . Accordingly, based on Lemma 4, we can
(1)
restrict our search for a feasible schedule to those schedules in which job Jj
(1)
is scheduled before Jj+1 for
all j ∈ {1, . . . , n − 1}. As in the proof of Theorem 2, by allowing an additional multiplicative factor of k! to
the running time of our algorithm, we can focus on a reduced subproblem where the ordering of the second
12
agent jobs is also predefined. Given such an ordering, we renumber the second agent jobs according to the
ordering.
In what follows we prove that each subproblem is FPT with respect to k. The proof uses the same ideas
as the proof of Theorem 2 and thus is briefly presented. Here as well we define variables x1 , . . . , xk , but this
(2)
time xj represents the number of the first agent jobs that are scheduled after Jj . Accordingly, we have to
include the constraint in (1) and for 1 ≤ j ≤ k − 1 the following constraint as well:
xj ≥ xj+1 .
(8)
The bound on the weighted sum of completion times of the second agent jobs can be expressed by
k
i
k
X
X
X
(2)
(2)
(2)
wj (n − xj ) ≤ A2 ,
pj +
wi
(9)
j=1
j=1
i=1
where the first term in the left-hand side is the weighted sum of completion times of the second agent jobs if
they are scheduled one after the other at the beginning of the schedule, and the second term in the left-hand
side corresponds to the contribution of the first agent jobs to the weighted sum of completion times of the
second agent jobs.
To bound the total weighted completion time of the first agent jobs, we need again to introduce a set
of real-valued variables yj , yj1 , . . . , yjn for each j ∈ {1 . . . , k}. Here, variable yj is meant to encode the
(2)
contribution of Jj
(2)
pj
Pn
to the total weighted completion time of the first agent jobs. Note that this is precisely
(1)
i=n−xj +1
wi . We use the variables yij to encode lower-bounds on the steps of the sum
Pn
i=n−xj +1
(1)
wi ,
as done in the proof of Theorem 2. Accordingly, for j ∈ {1, . . . , k}, we include the set of constraints in (6).
Moreover, for i ∈ {1, . . . , n} and j ∈ {1, . . . , k}, we include the set of constraints in (4) and also the following
set of constraints
(1)
yij ≥ (xj − n + i)(wi
(1)
− wi+1 ),
(10)
(1)
where wn+1 = 0 by definition. Finally, we encode the bound on the total weighted completion time of the
first agent jobs by
n
X
j=1
k
X
(1)
(2)
jwj +
pj yj ≤ A1 ,
(11)
j=1
where the first term in the left-hand side is the weighted sum of completion time of the first agent jobs if
they are scheduled one after the other at the beginning of the schedule, and the second term in the left-hand
13
side corresponds to the contribution of the second agent jobs to the weighted sum of completion times of the
first agent jobs.
(1)
= 1,
Theorem 3. The 1 pj
with respect to k.
3.4
P
The 1
(1)
Cj
≤ A1 ,
P
P
(1)
(1)
≤ A1 ,
(2)
(2)
≤ A2
wj Cj
wj Uj
P
P
(1)
Cj
≤ A2
problem is fixed-parameter tractable
problem
P
Ng et al. [24] and Leung et al. [20] proved that the 1
We next prove that the more general 1
(2)
(2)
wj Cj
≤ A1 ,
(1)
Cj
P
(2)
≤ A1 ,
(2)
wj Uj
P
(2)
Uj
≤ A2
≤ A2
problem is NP-complete.
problem is FPT with respect to k.
Our proof depends on the following easy-to-prove lemma (recall the definitions of E (i) and T (i) in Section 2):
Lemma 5. If there is a feasible solution for an instance of the 1
P
(1)
Cj
≤ A1 ,
P
(2)
(2)
wj Uj
≤ A2
problem,
then for the same instance there exists a feasible solution in which (i) the jobs of agent 1 are scheduled in a
(1)
non-decreasing order of pj
(2)
decreasing order of dj
(i.e., according to the SPT rule); (ii) the jobs in E (2) are scheduled in a non-
(i.e., according to the EDD rule); and (iii) the jobs in T (2) are scheduled last in an
arbitrary order.
Consider now the 1
P
(1)
Cj
≤ A1 ,
P
(2)
(2)
≤ A2
wj Uj
problem and define a set of 2k subproblems
corresponding to the O(2k ) possible ways to partition set J (2) into E (2) and T (2) such that the condition
P
(2)
(2)
Jj ∈T (2)
wj
≤ A2 holds. Due to Lemma 5, each subproblem reduces to an instance of the 1
(2)
P
(1)
Cj
≤
A1 , Lmax ≤ 0 problem which includes the n jobs of agent 1 and only the O(k) early jobs of agent 2. In the
reduced subproblem, we need to find if it is possible to schedule the jobs in J (1) ∪ E (2) such that all jobs in
E (2) are indeed early (i.e., completed not later than its due date), and
1
P
(1)
Cj
(2)
P
(1)
Cj
≤ A1 . Thus, the fact that the
≤ A1 , Lmax ≤ 0 problem is solvable in O(n + k) = O(n) time (see Yuan et al. [34]) leads to the
following theorem:
Theorem 4. The 1
P
(1)
Cj
≤ A1 ,
P
(2)
(2)
wj Uj
≤ A2
14
problem is solvable in O(2k n) time.
3.5
P
Intractability of the 1
(1)
Cj
Consider an instance of the 1 C(1) ,
P
≤ A1 ,
(2)
Ej
P
(2)
Ej
≥ A2
≥ A2
problem
problem, with k = 1 and A2 = 1. If there is a feasible
solution for such an instance, then the single job of agent 2 is scheduled in a JIT mode, i.e., during time
(2)
(2)
(2)
interval (d1 − p1 , d1 ]. Thus, such an instance is equivalent to an instance of a 1 C(1)
problem with a
single non-availability interval (more commonly denoted by 1 C(1) , n − a ). This problem is known to be
NP-complete when C(1) =
following corollary:
Corollary 1. The 1
4
P
P
(1)
Cj
(1)
Cj
≤ A1 (see Adiri et al. [1] and Lee and Liman [16]). Thus, we have the
≤ A1 ,
P
(2)
Ej
≥ A2
problem is NP-complete even for k = 1.
Weighted Number of Tardy Jobs
In this section we study variants of our problem of the form 1
P
(1)
(1)
wj Uj
≤ A1 , C(2) . That is, variants
where the scheduling criteria of agent 1 is the weighted number of tardy jobs. Note that already the single
agent 1
P
wj Uj ≤ A problem is NP-complete, even when all due dates are equal (a resulting dating back
to Karp’s seminal NP-completeness paper [13]). Therefore, any variant of our problem when the jobs of the
first agent have weights is hard.
Corollary 2. The 1 dj = d,
P
(1)
(1)
wj Uj
≤ A1 , C(2)
problem is NP-complete for k ≥ 0.
Due to Theorem 2, we restrict our analysis below to the unweighed 1
P
(1)
Uj
≤ A1 , C(2)
problem. We
provide a fixed-parameter algorithm for the case where the criteria of agent 2 is also the (weighted) number
of tardy jobs. This algorithm is then extended to the case where the jobs of agent 1 may have arbitrary
weights, but both agents jobs have unit processing times. On the contrary, when the criteria for agent 2 is the
weighted number of JIT jobs, we show that the problem is intractable already for highly restrictive special
cases. We do not know whether the problem is fixed-parameter tractable when the criteria for agent 2 is the
total weighted completion time; in the case, we can only show an nO(k) -time algorithm.
15
An nO(k) -time algorithm for the 1
4.1
P
The question whether the 1
(1)
Uj
≤ A1 ,
P
P
(1)
Uj
≤ A1 ,
(2)
(2)
wj Cj
≤ A2
P
(2)
(2)
wj C j
problem
≤ A2
problem is fixed-parameter tractable or not
remains an open question. Nevertheless, we show below that the problem can be solved in much slower but
still non-trivial nO(k) time. This leads to the conclusion that the problem belongs to the parameterized class
XP (see e.g. [8] for a formal definition), and is solvable in polynomial time when k is upper bounded by a
constant. We begin with the following easy lemma.
Lemma 6. If there is a feasible solution for an instance of the 1
P
then for the same instance there exists a feasible solution in which (i)
(1)
Uj
P
≤ A1 ,
(1)
Uj
P
(2)
(2)
wj Cj
≤ A2 problem,
= A1 ; (ii) the jobs in E (1) ∪J (2)
are scheduled first followed by the jobs in T (1) that are scheduled last in an arbitrary order; and (iii) the jobs
(1)
in E (1) are scheduled according to the EDD rule (i.e., in non-decreasing order of dj ).
Following Lemma 6, we renumber the jobs in J (1) according to the EDD rule. Furthermore, we divide
the original problem into k! instances, each of which represent a different processing order of the jobs
in J (2) . Consider a given instance, and assume that the jobs in J (2) are numbered according to their
processing order. For each such instance, we consider all possible partitions of J (1) into k + 1 subsets,
(1)
J (1) = J0
(1)
(1)
∪ · · · ∪ Jk , and all possible sets of k + 1 integers {e0 , . . . , ek } with ei ≤ |Ji | for each
i ∈ {0, . . . , k + 1} and
Pk
i=0 ei
= n − A1 .
(1)
(1)
Note that there are O(n2k+2 ) such pairs ({J0 , . . . , Jk }, {e0 , . . . , ek }). Given such a pair, we are looking
for a restricted schedule that satisfies the following three conditions: (i) there are at most ei early jobs
(1)
among the jobs in Ji
(1)
Ji
2
for each i = 0, ..., k; (ii) job Ji+1
is scheduled right after the ei early jobs within
for i = 0, ..., k − 1; and (iii)
P
(2)
(2)
≤ A2 . Note that the instance corresponding to the particular
wj Cj
processing order of J (2) has a feasible schedule iff such a restricted schedule corresponding to some pair
(1)
(1)
({J0 , . . . , Jk }, {e0 , . . . , ek }) exists. Below we show how to compute a restricted schedule, if it exists, in
O(n log n) time. This will yield the following theorem:
Theorem 5. The 1
P
(1)
Uj
≤ A1 ,
(1)
P
(2)
(2)
wj Cj
problem is solvable in O(k!n2k+1 log n) time.
≤ A2
(1)
Consider some pair ({J0 , . . . , Jk }, {e0 , . . . , ek }) as above. Note that if our goal was only to find a
(1)
schedule for there are at most ei early jobs in Ji
for each i, then this translates to solving k disjoint
16
P
instances of the single agent 1
(1)
(1)
problem over each set of jobs Ji , where ni = |Ji |.
U j ≤ ni − e i
Each such instance can be solved in O(ni lg ni ) time by a slight modification of the classical algorithm of
Moore [22], which gives us a total of O(n log n) time for all k + 1 instances. Furthermore, Moore’s algorithm
computes the schedule with minimum makespan (i.e., final completion time) amongst all schedules with at
most ei tardy jobs. Thus, composing the k + 1 schedules into a single schedule for both agents by scheduling
(1)
(2)
Ji+1 after the final job scheduled in Ji , gives us a schedule which minimizes
P
(2)
(2)
wj Cj
over all schedules
which satisfy properties (i) and (ii) above. Thus, in O(n lg n) time we can determine whether there exists a
(1)
(1)
restricted schedule corresponding to ({J0 , . . . , Jk }, {e0 , . . . , ek }), and so Theorem 5 holds.
We mention that this algorithm can slightly be improved if the jobs of agent 2 have unit weights. In this
(2)
case, we know that it is optimal to order these jobs in a non-decreasing order of pj , i.e., according to the
shortest processing time (SPT) rule. Thus, we do not have to try out all possible orderings of J (2) , reducing
the time complexity of the algorithm above by a factor of O(k!).
Corollary 3. The 1
4.2
P
(1)
Uj
≤ A1 ,
P
An FPT algorithm for 1
P
We next show that the 1
(1)
Uj
P
(2)
Cj
(1)
Uj
≤ A1 ,
P
problem is solvable in O(n2k+1 log n) time.
≤ A2
≤ A1 ,
(2)
P
(2)
wj Uj
(2)
(2)
wj Uj
≤ A2
≤ A2
problem is FPT with respect to k. Our proof
depends on the following easy-to-prove lemma:
P
Lemma 7. If there is a feasible solution for an instance of the 1
(1)
Uj
≤ A1 ,
P
(2)
(2)
wj Uj
≤ A2
problem,
then for the same instance there exists a feasible solution in which the jobs in E (1) ∪ E (2) are scheduled first
(i)
according to the EDD rule (i.e., in non-decreasing order of dj ), followed by the jobs in T (1) ∪ T (2) that are
scheduled last in an arbitrary order.
Consider now an instance of the 1
P
(1)
Uj
≤ A1 ,
P
(2)
(2)
wj Uj
≤ A2
problem, and define a set of 2k
instances corresponding to the O(2k ) possible ways to partition set J (2) into E (2) and T (2) such that the
feasibility condition
an instance of the 1
P
(2)
(2)
Jj
P
∈T (2)
(1)
Uj
wj
≤ A2 holds in each instance. Due to Lemma 7, each of these instances is
(2)
≤ A1 , Lmax ≤ 0 problem in which we need to find a feasible schedule for agent
1 subject to scheduling the set of O(k) jobs in E (2) such that they are all early. Agnetis et al. showed that the
17
1
P
(1)
Uj
(2)
≤ A1 , Lmax ≤ 0 problem is solvable in O(n log n + k log k) = O(n log n) time [3]. Thus, we obtain
the following:
P
Theorem 6. The 1
(1)
Uj
P
≤ A1 ,
(1)
4.3
An FPT algorithm for 1 pj
P
The 1
(1)
(1)
wj Uj
≤ A1 ,
P
(2)
(2)
(2)
≤ A2
(2)
= 1,
(2)
wj Uj
= pj
≤ A2
wj Uj
problem is solvable in O(2k n log n) time.
P
(1)
(1)
wj Uj
≤ A1 ,
P
(2)
(2)
wj Uj
≤ A2
problem is NP-complete even for the case of unit processing
time [26]. Next we complement this result, as well as Theorem 2, by showing that this problem is FPT
with respect to k. First observe that Lemma 7 holds here as well. Thus, we again create 2k instances
(1)
from our 1 pj
(2)
= pj
= 1,
P
(1)
(1)
wj Uj
≤ A1 ,
P
(2)
(2)
wj Uj
≤ A2
instance, where in each instance the set
E (2) ⊆ J (2) may vary. In any given instance of these 2k instances we may assume that
P
(2)
(2)
Jj
∈T (2)
wj
(1)
Furthermore, again due to Lemma 7, each of these instances is in fact an instance of the 1 pj
1,
P
(1)
(1)
wj Uj
(2)
≤ A1 , Lmax ≤ 0
≤ A2 .
(2)
= pj
=
problem in which we need to find a feasible schedule for agent 1 subject
to scheduling the set of O(k) jobs in E (2) such that they are all early. This latter problem is solvable in
O(max{n log n, k}) = O(n log n) time [26]. Thus, we get:
(1)
Theorem 7. The 1 pj
(2)
= pj
= 1,
O(2k n log n) time.
4.4
Intractability of the 1
P
(1)
Uj
P
(1)
(1)
≤ A1 ,
wj Uj
≤ A1 ,
P
(2)
Ej
P
≥ A2
(2)
(2)
wj Uj
≤ A2
problem is solvable in
problem
(1)
= d,
(1)
= d,
Following the observation made in Section 3.5, we can conclude that an instance of the 1 dj
A1 ,
P
(2)
Ej
≥ A2
problem with k = 1 and A2 = 1 is equivalent to an instance of the 1 dj
P
P
(1)
≤
(1)
≤
Uj
Uj
A1 , n − a problem, which is known to be NP-complete (see Lee [15]). Thus, we have the following corollary:
(1)
Corollary 4. The 1 dj
5
= d,
P
(1)
Uj
≤ A1 ,
P
(2)
Ej
≥ A2
problem is NP-complete for k ≥ 1.
Weighted Number of Just-in-Time Jobs
We next consider problems of the form 1
P
(1)
(1)
wj Ej
≥ A1 , C(2) , i.e., problems where the criteria of agent
(i)
1 is the total weighted number of JIT jobs. Recall that a job Jj
18
is scheduled in JIT mode if it is scheduled
(i)
(i)
(i)
precisely at the time interval (dj − pj , dj ]. We will show that when either C(2) =
C(2) =
P
(2)
P
(2)
wj Cj
(2)
(2)
wj Ej
P
(2)
(2)
wj Uj
≤ A2 , or
≥ A2 , the problem is fixed-parameter tractable in k. We also show that when C(2) =
≤ A2 the problem is fixed-parameter tractable if the first agent jobs are unweighed, while the
more general case (where the first agent jobs are weighted) is left open.
5.1
The 1
P
The 1
general 1
(1)
Ej
P
P
(1)
Ej
P
≥ A1 ,
(1)
Ej
P
≥ A1 ,
(2)
Cj
≥ A1 ,
easy-to-prove lemma:
P
(2)
(2)
wj C j
problem
≤ A2
≤ A2 problem is known to be NP-complete [29]. We next prove that the more
(2)
(2)
wj Cj
≤ A2
problem is FPT with respect to k. As usual, we begin with an
P
Lemma 8. In any feasible schedule for the 1
(1)
Ej
≥ A1 ,
P
(2)
(2)
wj Cj
≤ A2
problem the jobs in E (1) are
scheduled in non-decreasing due date (EDD) order. Moreover, if there is a feasible solution for the problem,
P (1)
then there is a feasible solution in which |Eb(1) | =
Ej = A1 , and the jobs in set Tb (1) = J (1) \ Eb(1) are
scheduled last in an arbitrary order.
Following Lemma 8, we assume without loss of generality that the jobs in J (1) are indexed according to
(1)
the EDD rule, i.e., d1
(1)
≤ d2
(1)
≤ ... ≤ dn . We will show that for any fixed ordering of the jobs of agents
2, we can determine in polynomial time whether there exists a feasible schedule where the relative order of
agent 2 jobs is exactly this ordering. Since there are k! orderings of the jobs of agent 2, this will imply that
P
the 1
(1)
Ej
≥ A1 ,
P
(2)
(2)
wj Cj
≤ A2
(2)
problem is FPT with respect to k.
(2)
Consider any fixed ordering J1 , . . . , Jk
(1)
(1)
of the jobs of agent 2. For convenience purposes, let J0
(1)
(1)
Jn+1 be two dummy jobs of agent 1 with p0 = w0
(1)
any given feasible schedule σ with Ja
(1)
and Jb
(1)
(1)
(1)
and
(1)
= d0 = pn+1 = wn+1 = 0 and dn+1 = ∞. Consider
being two jobs of agent 1 that are scheduled in JIT mode,
(1)
(1)
(1)
(1)
with no other jobs of agent 1 scheduled in between them. Obviously, db − da − pb ≥ 0, as otherwise Ja
(1)
and Jb
cannot be both scheduled in JIT mode. For ℓ ∈ {0, . . . , k − 1}, let k(a, b, ℓ) denote the number of
(2)
(2)
jobs in a maximal consecutive subsequence of jobs in {Jℓ+1 , . . . , Jk } with total processing time not greater
(1)
(1)
(1)
than db − da − pb . Then the following lemma holds:
(1)
Lemma 9. Suppose there is a feasible schedule in which Ja
(1)
and Jb
are scheduled in JIT mode with
no other jobs of agent 1 scheduled between them, and there are exactly ℓ jobs belonging to agent 2 that are
19
(2)
(1)
(2)
scheduled prior to Ja . Then there exists a feasible schedule where the sequence of jobs Jℓ+1 , . . . , Jℓ+k(a,b,ℓ) is
(1)
scheduled right after the completion time of job Ja , and no other jobs are scheduled prior to the completion
(1)
of job Jb .
(1)
(1)
For a pair of jobs Ja , Jb
(2)
∈ J (1) and ℓ ∈ {0, . . . , k−1}, let w(a, b, ℓ) denote the minimum contribution of
(2)
the job sequence Jℓ+1 , . . . , Jℓ+k(a,b,ℓ) to the total weighted completion time of agent 2 in any feasible schedule
σ. If no such schedule exists, define w(a, b, ℓ) = ∞. According to Lemma 9 we can compute w(a, b, ℓ) by
using the following formula:
ℓ+k(a,b,ℓ)
w(a, b, ℓ) =
X
ℓ+k(a,b,ℓ)
(2) (2)
wj Cj
X
=
j=ℓ+1
(2)
wj
d(1)
a
j=ℓ+1
+
j
X
i=ℓ+1
(2)
pi
!
(12)
.
Now, for b, e ∈ {0, . . . , n + 1}, and ℓ ∈ {0, . . . , k − 1}, let W (b, e, ℓ) denote the minimum total
weighted completion time of the first ℓ jobs of agent 2, among all partial schedules on the job set
(1)
(1)
(2)
(2)
{J1 , . . . , Jb , J1 , . . . , Jℓ }, such that e = |Eb(1) | =
P
(1)
Ej
(1)
and Jb
∈ Eb(1) is the last job to be scheduled.
(Again, set W (b, e, ℓ) = ∞ if no such schedule exists.) Note that by definition, the completion time of such
(1)
a feasible partial schedule is exactly at time db . The value W (b, e, ℓ) can be computed using the following
recursion that considers all possible ways of appending partial schedules that have fewer jobs of agent 2 and
one less JIT job of agent 1:
(1)
(1)
′
′
W (a, e − 1, ℓ′ ) + w(a, b, ℓ′ ) : d(1)
b − da − pb ≥ 0, ℓ = ℓ + k(a, b, ℓ ).
W (b, e, ℓ) = min
0≤a≤b−1
0≤ℓ′ ≤ℓ
∞
: otherwise.
(13)
Our algorithm computes all possible W (b, e, ℓ) values for b ∈ {0, ..., n + 1}, e ∈ {0, ..., A1 }, and ℓ ∈
{0, ..., k}, using the recursion given in Equation 13. For this, it computes in a preprocessing step all values
w(a, b, ℓ) and k(a, b, ℓ), for 0 ≤ a < b ≤ n + 1 and 0 ≤ ℓ ≤ k. The base cases of the recursion are given by
W (0, 0, 0) = 0, and W (0, e, ℓ) = ∞ for e 6= 0 or ℓ 6= 0. We report that there exists a feasible schedule for our
1
P
(1)
Ej
≥ A1 ,
P
(2)
Cj
≤ A2
(2)
(2)
instance, restricted to the relative fixed ordering J1 , . . . , Jk
for the jobs
of agent 2, iff W (n + 1, A1 , k) ≤ A2 .
Correctness of our algorithm is immediate from the above discussion. Let us now analyze its time complexity. There are k! ways to order the jobs of agent 2. For each such order, we compute all values W (b, e, ℓ),
w(a, b, ℓ) and k(a, b, ℓ). All values k(a, b, ℓ) can be computed straightforwardly in O(n2 k 2 ) time. Using these
20
values, and Equation 12, all values w(a, b, ℓ) can also be computed in O(n2 k 2 ) time. Finally, using dynamic
programming along with Equation 13, computing all values W (b, e, ℓ) requires O(n3 k 2 ) time. Thus, for each
fixed ordering of J (2) , we spend a total of O(n3 k 2 ) time. All together, this gives us an O(k!k 2 n3 ) time
algorithm.
P
Theorem 8. The 1
5.2
P
The 1
(1)
(1)
Ej
(1)
wj Ej
≥ A1 ,
≥ A1 ,
P
P
(2)
(2)
≤ A2
(2)
(2)
≤ A2
wj Cj
wj Uj
problem is solvable in O(k!k 2 n3 ) time.
problem
P
We next show how to modify the ideas used in Section 5.1 so that they apply to the 1
P
A1 ,
(2)
(2)
wj Uj
≤ A2
problem. We begin with the following analog of Lemma 8:
Lemma 10. In any feasible schedule for an instance of the 1
P
(1)
(1)
wj Ej
≥ A1 ,
P
(2)
(2)
wj Uj
(1)
(1)
wj Ej
≥
≤ A2 problem,
the jobs in Eb(1) ∪ E (2) are scheduled in an earliest due date (EDD) order. Moreover, if there is a feasible
solution for the instance, then there is a feasible solution in which the jobs in set Tb (1) ∪ T (2) are scheduled
last in an arbitrary order.
Following Lemma 10, we assume that the jobs in J (1) are numbered according to the EDD rule such that
(1)
d1
(1)
≤ d2
(1)
≤ ... ≤ dn . We create O(2k ) instances of the problem, according to all possible candidate sets
for E (2) such that
(2)
A1 , Lmax ≤ 0
P
(2)
(2)
Jj ∈E
/ (2)
wj
≤ A2 holds. Each instance is in fact an instance of the 1
P
(1)
(1)
wj Ej
≥
problem which includes the n jobs of agent 1 and only the O(k) early jobs of agent 2. In
the reduced instance, we need to determine whether there exists a schedule where each job in E (2) is indeed
P
early (i.e., completed not later than its due date) and
this problem in polynomial time.
(2)
We begin by renumbering the jobs in E (2) such that d1
(1)
Section 5.1, we add two dummy jobs J0
(1)
(1)
(1)
≥ A1 . Below, we show how we can solve
wj Ej
(2)
≤ d2
(1)
and Jn+1 with d0
(2)
≤ ... ≤ dk′ , where k ′ = |E (2) |. As in
(1)
= p0
(1)
= w0
(1)
(1)
= pn+1 = wn+1 = 0 and
(1)
dn+1 = ∞. Furthermore, we again let k(a, b, ℓ) denote, for 0 ≤ a < b ≤ n and 0 ≤ ℓ < k ′ , the length
(2)
(2)
of the maximal consecutive subsequence of jobs in {Jℓ+1 , . . . , Jk′ } with total processing time not greater
(1)
(1)
(1)
(1)
than db − da − pb . Then Lemma 9 holds here as well, and we can assume that if Ja
(2)
(1)
and Jb
(2)
are both
scheduled in JIT mode, with no other agent 1 jobs between them, then Jℓ+1 , . . . , Jℓ+k(a,b,ℓ) is scheduled right
(1)
after the completion of Ja
(1)
(and no other jobs are scheduled prior to the completion of Jb ).
21
(1)
Since no jobs of agent 2 are allowed to be late, a partial schedule that includes jobs Ja
(1)
consecutive JIT jobs with ℓ′ early jobs of agent 2 scheduled before Ja
(1)
ℓ′ + k(a, b, ℓ′ ) early jobs of agent 2 scheduled before Jb
(1)
(1)
(1)
and Jb
as two
is a feasible partial schedule with
only if the following two conditions holds:
(1)
Condition 1 : db − da − pb ≥ 0; and
(1)
Condition 2 : da +
Pj
(2)
i=ℓ′ +1
pi
(2)
− dj
≤ 0 for all j = ℓ′ + 1, ..., l′ + | k(a, b, ℓ′ ) |.
Now, let W (b, ℓ) represent the maximum total weighted number of JIT jobs among all partial schedules
(1)
(1)
(2)
(2)
(2)
(2)
(1)
on job set {J1 , . . . , Jb , J1 , . . . , Jℓ } where all jobs {J1 , . . . , Jℓ } are early and Jb
∈ Eb(1) is the last
scheduled job. Note that as opposed to Section 5.1, here this value represents the criteria of agent 1. Each
value W (b, ℓ) can be computed with the following recursion that considers all possible ways of appending
partial schedules that have fewer early jobs of agent 2 and one less JIT job of agent 1:
W (b, ℓ) =
W (a, ℓ′ ) + w(1)
: conditions 1 and 2 hold and ℓ = ℓ′ + |k(a, b, ℓ′ )|.
b
max
(14)
0≤a≤b−1
0≤ℓ′ ≤ℓ−1
−∞
: otherwise.
The base cases for this recursion are given by W (0, 0) = 0, and W (0, ℓ) = −∞ for ℓ 6= 0.
Our algorithm reports that there exists a feasible solution to the instance of 1
A1 ,
P
(2)
(2)
wj Uj
≤ A2 problem iff for some set E (2) with
P
(2)
(2)
Jj ∈E
/ (2)
wj
P
(1)
(1)
wj Ej
≥
≤ A2 we have W (n+1, |E (2)|) ≥ A1 .
The running time of this algorithm can be bounded by O(2k k 2 n2 ), using a similar analysis to the one given
in Section 5.1. Thus, we obtain:
Theorem 9. The 1
5.3
The 1
P
P
(1)
(1)
(1)
wj Ej
(1)
wj Ej
≥ A1 ,
P
It is known that the 1
≥ A1 ,
(1)
P
(1)
wj Ej
P
(2)
(2)
wj Uj
(2)
(2)
wj Ej
≥ A1 ,
P
problem
≥ A2
(2)
problem is solvable in O(2k k 2 n2 ) time.
≤ A2
(2)
wj Ej
≥ A2
problem is NP-complete, and that it is
polynomial-time solvable if the weights of either one of the two agents are all equal [29]. Below we show that
(i) the problem is NP-complete even for the case of unit processing time; and that (ii) the general problem
(with arbitrary processing time) is FPT with respect to k.
(1)
Theorem 10. The 1 pj
(2)
= pj
= 1,
P
(1)
(1)
wj Ej
≥ A1 ,
22
P
(2)
(2)
wj Ej
≥ A2
problem is NP-complete.
Proof. Given an instance (X, z) to the NP-hard Partition problem (see Definition 1), we construct the
(2)
(1)
= pj
following instance for the 1 pj
= 1,
P
(2)
(1)
= wj
k = m, and for j = 1, ..., n we set wj
(1)
(1)
wj Ej
≥ A1 ,
(1)
= xj and dj
(1)
Moreover, we set A1 = A2 = z. Note that since jobs Jj
P
(2)
= dj
(2)
and Jj
(2)
(2)
wj Ej
≥ A2
problem. We set n =
(1)
= j (recall that here pj
(2)
= pj
= 1).
have the same due date of j for j = 1, ..., n,
only one of them can be completed in a JIT mode. Thus, the total gain for both agents is restricted to be
Pn
not more than
j=1
xj = 2z.
Suppose that X can be partitioned into two sets S1 and S2 with
(i)
each job in {Jj
(i)
P
xj ∈S1
xj =
P
xj ∈S2
xj = z. Schedule
| xj ∈ Si } during time interval (j − 1, j], and schedule the remaining jobs in an arbitrary
order. Then each job in {Jj
The fact that
P
xj ∈Si
(i)
| xj ∈ Si } is scheduled in JIT mode, and Eb(i) = {Jj | xj ∈ Si } for i = 1, 2.
xj = z, for i = 1, 2, implies that in σ we have
Thus, there is a feasible schedule for our constructed instance.
For the other direction, suppose there exists a schedule σ with
P
P
(i)
(i)
Jj ∈Eb(i)
(i)
(i)
wj Ej
wj
= z for each agent i.
≥ Ai = z for i = 1, 2. Then
since the total gain in the objective function of both agents is restricted to be not more than 2z, we have
that
P
(i)
(i)
wj Ej
(i)
= z for i = 1, 2. This means that by setting Si = {xj : Ej
solution for (X, z) with
P
xj ∈Si
Next, we show that the 1
= 1}, for i = 1, 2, we obtain a
⊓
⊔
xj = z for i = 1, 2.
P
(1)
(1)
wj Ej
≥ A1 ,
P
(2)
(2)
wj Ej
≥ A2
problem is FPT with respect to k. To
(i)
this end, it is convenient to view all jobs in J (1) ∪ J (2) as time intervals. For a job Jj ∈ J (1) ∪ J (2) , let the
(i)
time interval of Jj
(i)
(i)
(i)
(i)
(i)
be Ij = (dj − pj , dj ]. Then if Jj
is required to be scheduled in JIT mode, it has to
(i)
be scheduled within its time interval Ij . Thus, any pair of jobs Jj1 , Jj2 ∈ J (1) ∪ J (2) can be simultaneously
scheduled in JIT mode iff Ij1 ∩ Ij2 = ∅. This means that our goal now translates to finding a set of pairwise
disjoint time intervals, for which the total weight of set of intervals related to the jobs of each agent met its
bound.
We try out all possible candidates for Eb(2) ; that is, all subsets of agent 2 jobs J2 ⊆ J (2) that have
pairwise disjoint time intervals and
P
(2)
(2)
Jj ∈J2
wj
≥ A2 . For each such subset J2 , we compute the subset of
agent 1 jobs J1 ⊆ J (1) with time intervals that do not intersect any time interval of a job in J2 . That is,
(1)
J1 = {Jj
(1)
∈ J (1) : Ij
(2)
∩ Ii
(2)
= ∅ for all Ji
(1)
subset I1∗ of intervals in the set I1 = {Ij
∈ J2 }. We then compute a maximum weight pairwise disjoint
(1)
: Jj
∈ J1 }, and report that we have found a feasible schedule if
23
(1)
the total weight of the jobs in J1∗ = {Jj
(1)
: Ij
∈ I1∗ } is at least A1 . If no such subset of jobs J1∗ is found
for any possible candidate for Eb(2) , we report that our instance has no feasible schedule.
Correctness of our algorithm follows from the fact that we compute the optimal set Eb(1) for each possible
candidate for Eb(2) . There are O(2k ) candidates J2 for Eb(2) . For each candidate J2 , computing the set J1 can
be done in O(n log n + k log n) = O(n log n) time. Moreover, the set J1∗ can be computed in O(n log n) time
(e.g. using [12]). Thus, in total, our algorithm runs in O(2k n log n) time.
Theorem 11. The 1
6
P
(1)
(1)
wj Ej
≥ A1 ,
P
(2)
(2)
wj Ej
problem is solvable in O(2k n log n) time.
≥ A2
Conclusions and Open Problems
In this paper we initiated a parameterized analysis for two-agent single-scheduling problems, where the
parameter studied is the number k of jobs belonging to the second agent. We considered three possible
scheduling criteria – total weighted completion time, total weighted number of tardy jobs, and total weighted
number of JIT jobs – and all possible combinations of these criteria for each agent. Our analysis shows that
parameter k indeed provides various positive results in different settings, and is summarized in Table 1 below.
There are several directions to directly extend our work. First, one can find different parameters such
as the number of different processing times in the input, or the number of different due dates. Many such
parameterizations make perfect sense for practical applications. Second, one consider other scheduling criteria
not considered in this paper such as the maximal lateness. Below we list the three most important questions
that were left open directly from our work:
1. Determine the parameterized complexity of the 1
P
Uj
P
wj Ej
(1)
≤ A1 ,
its variants (unit weights, unit processing times, etc ...).
2. Determine the parameterized complexity of the 1
(1)
(1)
P
(2)
(2)
wj Cj
≥ A1 ,
P
(2)
≤ A2
(2)
wj Cj
problem, or any of
≤ A2
problem.
3. Can the running times of the algorithms presented in Sections 3.2 and 3.3 be improved? More specifically,
are there purely combinatorial algorithms for these problems, avoiding ILPs altogether?
24
P
≤ A2
P
= 1 (Th. 1),
Hard for wj
(1)
= 1 (Th. 2),
FPT for wj
(1)
= 1 (Th. 3).
(2)
(2)
wj C j
(2)
Hard for wj
P
(1)
(1)
wj C j
≤ A1
FPT for wj
FPT for pj
(2)
(1)
Hard in general (Cor. 2),
P
(1)
(1)
wj Uj
≤ A1
(1)
Open for wj
= 1.
(2)
(2)
wj Uj
P
≤ A2
(1)
(2)
≥ A2
= 1 (Th. 1),
Hard even when
= 1 (Th. 3).
wj = 1 (Cor. 1).
(i)
Hard in general (Cor. 2),
FPT for wj
(2)
wj Ej
= 1 (Th. 6),
Hard even when
(i)
(1)
wj = 1 and dj
= d (Cor. 4).
(i)
FPT for pj = 1 (Th. 7).
Open in general,
P
(1)
(1)
wj Ej
(1)
≥ A1 FPT when wj
= 1 (Th. 8).
FPT
(Th. 9)
FPT
(Th. 11)
Table 1. A table depicting most of the results presented in the paper. Rows correspond to the different scheduling
criteria for agent 1, while columns are associated with the different criteria for agent 2.
7
Acknowledgments
The research leading to these results has received funding from the People Programme (Marie Curie Actions)
of the European Union’s Seventh Framework Programme (FP7/2007-2013) under REA grant agreement
number 631163.11, and by the Israel Science Foundation (grant No. 1055/14). Nimrod Talmon was supported
by a postdoctoral fellowship from I-CORE ALGO.
References
1. I. Adiri, J. Bruno, E. Frostig, and A.H.G. Rinnooy Kan. Single machine flow-time scheduling with a single
breakdown. Acta Informatica, 26(7):679–696, 1989.
2. A. Agnetis, J.C. Billaut, S. Gawiejnowicz, D. Pacciarelli, and A. Soukhal. Multiagent Scheduling: Models and
Algorithms. Imprint: Springer, 2014.
3. A. Agnetis, P.B. Mirchandani, D. Pacciarelli, and A. Pacifici. Scheduling problems with two competing agents.
Operations Research, 52(2):229–242, 2004.
25
4. K.R. Baker and J.C. Smith. A multiple-criterion model for machine scheduling. Journal of Scheduling, 6(1):7–16,
2003.
5. H.L. Bodlaender and M.R. Fellows. W[2]-hardness of precedence constrained k-processor scheduling. Operations
Research Letters, 18(2):93–97, 1995.
6. P. Brucker. Scheduling algorithms. Springer Science, 2006.
7. M. Cygan, F.V. Fomin, L. Kowalik, D. Lokshtanov, D. Marx, M. Pilipczuk, and M. Pilipczuk. Parameterized
Algorithms. Springer, 2015.
8. R.G. Downey and M.R. Fellows. Parameterized Complexity. Springer, 1999.
9. M.R. Fellows and C. McCartin. On the parametric complexity of schedules to minimize tardy tasks. Theoretical
Computer Science, 298(2):317–324, 2003.
10. J. Flum and M. Grohe. Parameterized Complexity Theory. An EATCS Series: Texts in Theoretical Computer
Science. Springer, 1998.
11. M.R. Garey and D.S. Johnson. Computers and Intractability: A Guide to the Theory of NP-Completeness. 1979.
12. U. I. Gupta, D. T. Lee, and Joseph Y.-T. Leung. Efficient algorithms for interval graphs and circular-arc graphs.
Networks, 12(4):459–467, 1982.
13. R.M. Karp. Reducibility among combinatorial problems. In Raymond E. Miller and James W. Thatcher, editors,
Complexity of Computer Computations, pages 85–103. 1972.
14. M.Y. Kovalyov, A. Oulamara, and A. Soukhal. Two-agent scheduling on an unbounded serial batching machine.
In Combinatorial Optimization, pages 427–438. Springer, 2012.
15. C.Y. Lee. Machine scheduling with an availability constraint. Journal of Global Optimization, 9(3-4):395–416,
1996.
16. C.Y. Lee and S.D. Liman. Single machine flow-time scheduling with scheduled maintenance. Acta Informatica,
29(4):375–382, 1992.
17. K. Lee, B.C. Choi, J.Y.T. Leung, and M.L. Pinedo. Approximation algorithms for multi-agent scheduling to
minimize total weighted completion time. Information Processing Letters, 109(16):913–917, 2009.
18. H.L. Lenstra. Integer programming with a fixed number of variables. Mathematics of Operations Research,
8(4):538–548, 1983.
19. J.Y.T. Leung. Handbook of Scheduling: Algorithms, Models and Performance Analysis. Springer, 2015.
20. J.Y.T. Leung, M. Pinedo, and G. Wan. Competitive two-agent scheduling and its applications. Operations
Research, 58(2):458–469, 2010.
26
21. M. Mnich and A. Wiese. Scheduling meets fixed-parameter tractability. Mathematical Programming, 154(1):533–
562, 2015.
22. J.M. Moore. An n job, one machine sequencing algorithm for minimizing the number of late jobs. Management
Science, 15.
23. B. Mor and G. Mosheiov. Single machine batch scheduling with two competing agents to minimize total flowtime.
European Journal of Operational Research, 215(3):524–531, 2011.
24. C.T. Ng, T.C.E. Cheng, and J.J. Yuan. A note on the complexity of the problem of two-agents scheduling on a
single machine. Journal of Combinatorial Optimization, 12(4):387–394, 2006.
25. R. Niedermeier. Invitation to Fixed-Parameter Algorithms. Oxford Lecture Series in Mathematics and Its Applications. Oxford Univerity Press, 2006.
26. D. Oron, D. Shabtay, and G. Steiner. Single machine scheduling with two competing agents and equal job
processing times. European Journal of Operational Research, 244(1):86–99, 2015.
27. P. Perez-Gonzalez and J.M. Framinan. A common framework and taxonomy for multicriteria scheduling problems
with interfering and competing jobs: Multi-agent scheduling problems. European Journal of Operational Research,
235(1):1–16, 2014.
28. M.L. Pinedo. Scheduling: theory, algorithms, and systems. Springer Science & Business Media, 2012.
29. D. Shabtay, O. Dover, and M. Kaspi. Single-machine two-agent scheduling involving a just-in-time criterion.
International Journal of Production Research, 53(9):2590–2604, 2015.
30. W.E Smith. Various optimizers for single-stage production. Naval Research Logistics, 3:59–66, 1956.
31. R. van Bevern, M. Mnich, R. Niedermeier, and M. Weller. Interval scheduling and colorful independent sets.
Journal of Scheduling, 18(5):449–469, 2015.
32. R. van Bevern, R. Niedermeier, and O. Suchý. A parameterized complexity view on non-preemptively scheduling
interval-constrained jobs: few machines, small looseness, and small slack. CoRR, abs/1508.01657, 2015.
33. Y. Yin, S.R. Cheng, T.C.E. Cheng, D.J. Wang, and C.C. Wu. Just-in-time scheduling with two competing agents
on unrelated parallel machines. Omega, 63:41–47, 2016.
34. J.J. Yuan, W.P. Shang, and Q. Feng. A note on the scheduling with two families of jobs. Journal of Scheduling,
8(6):537–542, 2005.
27
| 8cs.DS
|
Constant Approximation for Capacitated k-Median with
(1 + )-Capacity Violation
arXiv:1603.02324v2 [cs.DS] 11 May 2016
Gökalp Demirci∗
Shi Li†
Abstract
We study the Capacitated k-Median problem for which existing constant-factor approximation algorithms are all pseudo-approximations that violate either the capacities or the upper
bound k on the number of open facilities. Using the natural LP relaxation for the problem, one
can only hope to get the violation factor down to 2. Li [SODA’16] introduced a novel LP to go
beyond the limit of 2 and gave a constant-factor approximation algorithm that opens (1 + )k
facilities.
We use the configuration LP of Li [SODA’16] to give a constant-factor approximation for
the Capacitated k-Median problem in a seemingly harder configuration: we violate only the
capacities by 1 + . This result settles the problem as far as pseudo-approximation algorithms
are concerned.
1
Introduction
In the capacitated k-median problem (CKM), we are given a set F of facilities together with their
capacities ui ∈ Z>0 for i ∈ F , a set C of clients, a metric d on F ∪ C, and a number k. We are
asked to open some of these facilities F 0 ⊆ F and give an assignment σ : C → F 0 connecting each
client to one of the open facilities so that the number of open facilities is not bigger than k, i.e.
|F 0 | ≤ k (cardinality constraint), and each facility i ∈ F 0 is connected to at most ui clients, i.e.
σ −1 (i) ≤ ui (capacity constraint). The goal is to minimize the sum of the connection costs, i.e.
P
j∈C d(σ(j), j).
Without the capacity constraint, i.e. ui = ∞ for all i ∈ F , this is the famous k-median problem
(KM). The first constant-factor approximation algorithm for KM is given by Charikar et al. [9],
guaranteeing a solution within 6 23 times the cost of the optimal solution. Then the approximation
ratio has been improved by a series of papers [13, 8, 3, 12, 17, 5]. The current best ratio for KM is
2.675 + due to Byrka et al. [5], which was obtained by improving a part of the algorithm given
by Li and Svensson [17].
On the other hand, we don’t have a true constant approximation for CKM. All known constantfactor results are pseudo-approximations which violate either the cardinality or the capacity constraint. Aardal et al. [1] gave an algorithm which finds a (7 + )-approximate solution to CKM
by opening at most 2k facilities, i.e. violating the cardinality constraint by a factor of 2. Guha
[11] gave an algorithm with approximation ratio 16 for the more relaxed uniform CKM, where all
∗
Department of Computer Science, University of Chicago, [email protected].
Department of Computer Science and Engineering, University at Buffalo, [email protected]. Supported in part
by NSF grant CCF-1566356.
†
1
capacities are the same, by connecting at most 4u clients to each facility, thus violating the capacity
constraint by 4. Li [14] gave a constant-factor algorithm for uniform CKM with capacity violation
of only 2 + by improving the algorithm in [9]. For non-uniform capacities, Chuzhoy and Rabani
[10] gave a 40-approximation for CKM by violating the capacities by a factor of 50 using a mixture
of primal-dual schema and lagrangian relaxations. Their algorithm is for a slightly relaxed version
of the problem called soft CKM where one is allowed to open multiple collocated copies of a facility
in F . The CKM definition we gave above is sometimes referred to as hard CKM as opposed to
this version. Recently, Byrka et al. [4] gave a constant-factor algorithm for hard CKM by keeping
capacity violation factor under 3 + .
All these algorithms for CKM use the basic LP relaxation for the problem which is known
to have an unbounded integrality gap even when we are allowed to violate either the capacity
or the cardinality constraint by 2 − . In this sense, results of [1] and [14] can be considered as
reaching the limits of the basic LP relaxation in terms of restricting the violation factor. In order
to go beyond these limits, Li [15] introduced a novel LP called the rectangle LP and presented a
constant-factor approximation algorithm for soft uniform CKM by opening (1 + )k facilities. This
was later generalized by the same author to non-uniform CKM [16], where he introduced an even
stronger LP relaxation called the configuration LP. Very recently, independently of the work in
this paper, Byrka et al. [6] used this configuration LP to give a similar algorithm for uniform CKM
violating the capacities by 1 + .
1.1
Our Result
In this paper, we use the configuration LP of [16] to give an O(1/5 )-approximation algorithm for
non-uniform hard CKM which respects the cardinality constraint and connects at most (1 + )ui
clients to any open facility i ∈ F . The running time of our algorithm is nO(1/) . Thus, with
this result, we now have settled the CKM problem from the view of pseudo-approximation algorithms: either (1 + )-cardinality violation or (1 + )-capacity violation is sufficient for a constant
approximation for CKM.
The known results for the CKM problem have suggested that designing algorithms with capacity
violation (satisfying the cardinality constraint) is harder than designing algorithms with cardinality
violation. Note, for example, that the best known cardinality violation factor for non-uniform CKM
among algorithms using only the basic LP relaxation (a factor of 2 in [1]) matches the smallest
possible cardinality violation factor dictated by the gap instance. In contrast, the best capacityviolation factor is 3 + due to [4], but the gap instance for the basic LP with the largest known
gap eliminates only the algorithms with capacity violation smaller than 2.
Furthermore, we can argue that, for algorithms based on the basic LP and the configuration
LP, a β-capacity violation can be converted to a β-cardinality violation, suggesting that allowing
capacity violation is more restrictive than allowing cardinality violation. Suppose we have an αapproximation algorithm for CKM that violates the capacity constraint by a factor of β, based on
the basic LP relaxation. Given a solution (x, y) to the basic LP for a given CKM instance I, we
construct a new instance I 0 by scaling k by a factor of β and scaling all capacities by a factor of
1/β (in a valid solution, we allow connections to be fractional, thus fractional capacities do not
cause issues). Then it is easy to see that (x, βy) is a valid LP solution to I 0 (with soft capacities).
A solution to I 0 that only violates the capacity constraint by a factor of β is a solution to I that
only violates the cardinality constraint by a factor of β. Thus, by considering the new instance, we
conclude that for algorithms based on the basic LP relaxation, violating the cardinality constraint
2
gives more power. The same argument can be made for algorithms based on the configuration
LP: one can show that a valid solution to the configuration LP for I yields a valid solution to
the configuration LP for I 0 . However, this reduction in the other direction does not work: due to
constraint (3), scaling y variables by a factor of 1/β does not yield a valid LP solution.
Our Techniques. Our algorithm uses the configuration LP introduced in [16] and the framework
of [16] that creates a two-level clustering of facilities. [16] considered the (1+)-cardinality violation
setting, which is more flexible in the sense that one has the much freedom to distribute the k extra
facilities. In our (1 + )-capacity violation setting, each facility i can provide an extra ui capacity;
however, these extra capacities are restricted by the locations of the facilities. In particular, we need
one more level of clustering to form so-called “groups” so that each group contains Ω(1/) fractional
open facility. Only with groups of Ω(1/) facilities, we can benefit from the extra capacities given
by the (1+)-capacity scaling. Our algorithm then constructs distributions of local solutions. Using
a dependent rounding procedure we can select a local solution from each distribution such that the
solution formed by the concatenation of local solutions has a small cost. This initial solution may
contain more than k facilities. We then remove some already-open facilities, and bound the cost
incurred due to the removal of open facilities. When we remove a facility, we are guaranteed that
there is a close group containing Ω(1/) open facilities and the extra capacities provided by these
facilities can compensate for the capacity of the removed facility.
Organization. The remaining part of the paper is organized as follows. In Sections 2 and 3, we
describe the configuration LP introduced in [16] and our three-level clustering procedure respectively. In Section 4, we show how to construct the distributions of local solutions. Then finally in
Section 5, we show how to obtain our final solution by combining the distributions we constructed.
2
The Basic LP and the Configuration LP
In this section, we give the configuration LP of [16] for CKM. We start with the following basic LP
relaxation:
P
min
s.t.
(Basic LP)
i∈F,j∈C d(i, j)xi,j
P
P
i∈F
i∈F
yi ≤ k;
xi,j = 1,
xi,j ≤ yi ,
P
(1)
∀j ∈ C;
∀i ∈ F, j ∈ C;
j∈C
(2)
xi,j ≤ ui yi ,
0 ≤ xi,j , yi ≤ 1,
(3)
∀i ∈ F ;
∀i ∈ F, j ∈ C.
(4)
(5)
In the LP, yi indicates whether a facility i ∈ F is open, and xi,j indicates whether client j ∈ C
is connected to facility i ∈ F . Constraint (1) is the cardinality constraint assuring that the number
of open facilities is no more than k. Constraint (2) says that every client must be fully connected
to facilities. Constraint (3) requires a facility to be open in order to connect clients. Constraint (4)
is the capacity constraint.
It is well known that the basic LP has unbounded integrality gap, even if we are allowed to
violate the cardinality constraint or the capacity constraint by a factor of 2 − . In the gap instance
for the capacity-violation setting, each facility has capacity u, k is 2u − 1, and the metric consists
of u isolated groups each of which has 2 facilities and 2u − 1 clients that are all collocated. In other
words, the distances within a group are all 0 but the distances between groups are nonzero. Any
integral solution for this instance has to have a group with at most one open facility. Therefore,
even with (2 − 2/u)-capacity-violation, we have to connect 1 client in this group to open facilities
3
in other groups. On the other hand, a fractional solution to the basic LP relaxation opens 2 − 1/u
facilities in each group and serves the demand of each group using only the facilities in that group.
Note that the gap instance disappears if we allow a capacity violation of 2.1
In order to overcome the gap in the cardinality-violation setting, Li [16] introduced a novel LP
for CKM called the configuration LP, which we formally state below. Let us fix a set B ⊆ F of
facilities. Let ` = Θ(1/) and `1 = Θ(`) be sufficiently large integers. Let S = {S ⊆ B : |S| ≤ `1 }
and Se = S ∪ {⊥}, where ⊥ stands for “any subset of B with size more than `1 ”; for convenience,
we also treat ⊥ as a set such that i ∈ ⊥ holds for every i ∈ B. For S ∈ S, let zSB indicate the event
B indicate the event that the number of open
that the set of open facilities in B is exactly S and z⊥
facilities in B is more than `1 .
B indicates the event that z B = 1 and i is open. (If i ∈ B but
For every S ∈ Se and i ∈ S, zS,i
S
B = z B ; we
i∈
/ S, then the event will not happen.) Notice that when i ∈ S 6= ⊥, we always have zS,i
S
e i ∈ S and client j ∈ C, z B indicates
keep both variables for notational purposes. For every S ∈ S,
S,i,j
B = 1 and j is connected to i. In an integral solution, all the above variables are
the event that zS,i
{0, 1} variables. The following constraints are valid. To help understand the constraints, it is good
B as z B · y and z B
B
to think of zS,i
i
S
S,i,j as zS · xi,j .
X
zSB = 1;
(6)
S∈Se
X
e
S∈S:i∈S
X
0
B
zS,i
= yi , ∀i ∈ B;
(7)
i∈S
X
B
zS,i,j
= xi,j ,∀i ∈ B, j ∈ C;
e
S∈S:i∈S
B
≤ zS,i,j
≤
B
zS,i
= zSB ,
X
(8)
j∈C
B
zS,i,j
≤ zSB ,
≤
zSB ,
e i ∈ S, j ∈ C;
∀S ∈ S,
(9)
(10)
e j ∈ C;
∀S ∈ S,
(11)
B
B
e i ∈ S;
zS,i,j
≤ ui zS,i
,∀S ∈ S,
X
B
zS,i
∀S ∈ S, i ∈ S;
i∈B
B
B
z⊥,i
≥ `1 z⊥
.
(12)
(13)
e Constraint (7) says that if i is open
Constraint (6) says that zSB = 1 for exactly one S ∈ S.
B
then there is exactly one S ∈ Se with zS,i = 1. Constraint (8) says that if j is connected to i then
there is exactly one S ∈ Se such that z B = 1. Constraint (9) is by the definition of variables.
S,i,j
Constraint (10) holds as we mentioned earlier. Constraint (11) says that if zSB = 1 then j can be
connected to at most 1 facility in S. Constraint (12) is the capacity constraint. Constraint (13)
B = 1, there are at least ` open facilities in B.
says that if z⊥
1
The configuration LP is obtained from the basic LP by adding the z variables and Constraints (6)
to (13) for every B ⊆ F . Since there are exponentially many subsets B ⊆ F , we don’t know how to
solve this LP efficiently. However, note that there are only polynomially many (nO(`1 ) ) z B variables
for a fixed B ⊆ F . Given a fractional solution (x, y) to the basic LP relaxation, we can construct the
values of z B variables and check their feasibility for Constraints (6) to (13) in polynomial time as
in [16]. Our rounding algorithm either constructs an integral solution with the desired properties,
or outputs a set B ⊆ F such that Constraints (6) to (13) are infeasible. In the latter case, we can
find a constraint in the configuration LP that (x, y) does not satisfy. Then we can run the ellipsoid
1
A similar instance can be given to show that the gap is still unbounded when the cardinality constraint is violated,
instead of the capacity constraint, by less than 2: let k = u + 1 and each group have 2 facilities and u + 1 clients.
4
method and the rounding algorithm in an iterative way (see, e.g., [7, 2]).
Notations From now
P on, we fix a solution ({xi,j : i ∈ F, j ∈ C} , {yi : i ∈ F }) to the basic LP.
We
cost of j, for every j ∈ C. Let Di :=
P define dav (j) := i∈F xi,j d(i, j) to be the connectionP
x
(d(i,
j)
+
d
(j))
for
every
i
∈
F
,
and
D
:=
for every S ⊆ F . We denote
av
S
j∈C i,j
i∈S DiP
P
the
value
of
the
solution
(x,
y)
by
LP
:=
x
d(i,
j)
=
i∈F,j∈C i,j
j∈C dav (j). Note that DF =
P
P
P
P
j∈C dav (j) Pi∈F xi,j = 2LP. For any set
i∈F,j∈C xi,j (d(i, j) + dav (j)) =
i∈F,j∈C xi,j d(i, j) +
F 0 ⊆ F of facilities and C 0 ⊆ C of clients, we shall let xF 0 ,C 0 := i∈F
P0 ,j∈C 0 xi,j ; we simply use
0
xi,C 0 for x{i},C 0 and xF 0 ,j for xF 0 ,{j} . For any F ⊆ F , let yF 0 :=
i∈F 0 yi . Let d(A, B) :=
mini∈A,j∈B d(i, j) denote the minimum distance between A and B, for any A, B ⊆ F ∪ C; we
simply use d(i, B) for d({i} , B).
Moving of Demands After the set of open facilities is decided, the optimum connection assignment from clients to facilities can be computed by solving the minimum cost b-matching problem.
Due to the integrality of the matching polytope, we may allow the connections to be fractional.
That is, if there is a good fractional assignment, then there is a good integral assignment. So we
can use the following framework to design and analyze the rounding algorithm. Initially there is
one unit of demand at each client j ∈ C. During the course of our algorithm, we move demands
fractionally within F ∪ C; moving α units of demand from i to j incurs a cost of αd(i, j). At the
end, all the demands are moved to F and each facility i ∈ F has at most (1 + O( 1` ))ui units of
demand. We open a facility if it has positive amount of demand. Our goal is to bound the total
moving cost by O(`5 )LP and the number of open facilities by k.
3
Representatives, Black Components, and Groups
Our algorithm starts with bundling facilities together with a three-phase process each of which
creates bigger and bigger clusters. At the end, we have a nicely formed network of sufficiently big
clusters of facilities. See Figure 1 for illustration of the three-phase clustering.
3.1
Representatives, Bundles and Initial Moving of Demands
In the first phase, we use a standard approach to facility location problems ([18, 19, 9, 16]) to
partition the facilities into bundles {Uv }v∈R , where each bundle Uv is associated with a center
v ∈ C that is called a representative and R ⊆ C is the set of representatives. Each bundle Uv has
a total opening at least 1/2.
Let R = ∅ initially. Repeat the following process until C becomes empty: we select the client
v ∈ C with the smallest dav (v) and add it to R; then we remove all clients j such that d(j, v) ≤
4dav (j) from C (thus, v itself is removed). We use v and its variants to index representatives, and
j and its variants to index general clients. The family {Uv : v ∈ R} is the Voronoi diagram of F
with R being the centers: let Uv = ∅ for every v ∈ R initially; for each location
S i ∈ F , we add i to
Uv for v ∈ R that is closest to i. For any subset V ⊆ R, we use U (V ) := v∈V Uv to denote the
union of Voronoi regions with centers V .
Lemma 3.1. The following statements hold:
(3.1a) for all v, v 0 ∈ R, v 6= v 0 , we have d(v, v 0 ) > 4 max {dav (v), dav (v 0 )}
(3.1b) for all j ∈ C, there exists v ∈ R, such that dav (v) ≤ dav (j) and d(v, j) ≤ 4dav (j);
(3.1c) yUv ≥ 1/2 for every v ∈ R;
5
representatives
facilities
(a). bundles {Uv }v∈R
black components
(b). black components J
and forest Υ∗J
groups
(c). groups G and forest ΥG
Figure 1: The three-phase clustering procedure. In the first phase (Figure (a)), we partition F into
bundles, centered at the set R of representatives. In the second phase (Figure (b)), we partition R
into a family J of black components and construct a degree-2 rooted forest over J . In the third
phase (Figure(c)), we partition J into a family G of groups; ΥG is formed from Υ∗J by contracting
each group into a single node.
(3.1d) for any v ∈ R, i ∈ Uv , and j ∈ C, we have d(i, v) ≤ d(i, j) + 4dav (j).
Proof. First consider Property (3.1a). Assume dav (v) ≤ dav (v 0 ). When we add v to R, we remove
all clients j satisfying d(v, j) ≤ 4dav (j) from C. If v 0 ∈ R, then it must have been d(v, v 0 ) > 4dav (v 0 ).
For Property (3.1b), just consider the iteration in which j is removed from C. The representative v
added to R in this iteration satisfy the property. Then consider Property
(3.1c). By Property
(3.1a),
P
P
we have B := {i ∈ F : d(i, v) ≤ 2dav (v)} ⊆ Uv . Since dav (v) = i∈F xi,v d(i, v) and i∈F xi,v = 1,
we have dav (v) ≥ (1 − xB,v )2dav (v), implying yUv ≥ yB ≥ xB,v ≥ 1/2, due to Constraint (3).
Then we consider Property (3.1d). By Property (3.1b), there is a client v 0 ∈ R such that
dav (v 0 ) ≤ dav (j) and d(v 0 , j) ≤ 4dav (j). Since d(i, v) ≤ d(i, v 0 ) as v 0 ∈ R and i was added to Uv , we
have d(i, v) ≤ d(i, v 0 ) ≤ d(i, j) + d(j, v 0 ) ≤ d(i, j) + 4dav (j).
The next lemma shows that moving demands from facilities to their corresponding representative
doesn’t cost much.
P
Lemma 3.2. For every v ∈ R, we have i∈Uv xi,C d(i, v) ≤ O(1)DUv .
Proof. By Property (3.1d), we have d(i, v) ≤ d(i, j) + 4dav (j) for every i ∈ Uv and j ∈ C. Thus,
X
X
X
xi,C d(i, v) ≤
xi,j d(i, j) + 4dav (j) ≤
4Di = 4DUv .
i∈Uv
i∈Uv ,j∈C
i∈Uv
Since {Uv : v ∈ R} forms a partition of F , we get the following corollary.
6
Corollary 3.3.
P
v∈R,i∈Uv
xi,C d(i, v) ≤ O(1)LP.
Initial Moving of Demands With this corollary, we now move all the demands from C to R.
First for every j ∈ C and i ∈ F , we move xi,j units of demand from j to i. The moving cost of this
step is exactly LP. After the step, all demands are at F and every i ∈ F has xi,C units of demand.
Then, for every v ∈ R and i ∈ Uv , we move the xi,C units of demand at i to v. The moving cost for
this step is O(1)LP. Thus, after the initial moving, all demands are at the set R of representatives:
a representative v has xUv ,C units of demand.
3.2
Black Components
In the second phase, we employ the minimum-spanning-tree construction of [16] to partition the
set R of representatives into a family J of so-called black components. There is a degree-2 rooted
forest Υ∗J over J with many good properties. For example, each non-root black component is not
far away from its parent, and each root black component of Υ∗J contains a total opening of Ω(`).
(For simplicity, we say the total opening at a representative v ∈ R is yUv , which is the total opening
at the bundle Uv .) The forest in [16] can have a large degree, while our algorithm requires the forest
to have degree 2. This property is guaranteed by using the left-child-right-sibling representation.
We now describe the framework of [16]. We run the classic Kruskal’s algorithm to find the
minimum spanning tree MST of the metric (R, d), and then color the edges in MST in black, grey
or white. In Kruskal’s algorithm, we maintain the set EMST of edges added to MST so far and
a partition
P of R. Initially, we have EMST = ∅ and P = {{v} : v ∈ R}. The length of an edge
R
e ∈ 2 is the distance between the two endpoints of e. We sort all edges in R2 in the ascending
order of their lengths, breaking ties arbitrarily. For each pair (v, v 0 ) in this order, if v and v 0 are not
in the same partition in P, we add the edge (v, v 0 ) to EMST and merge the two partitions containing
v and v 0 respectively.
We then color edges in EMST . For every v ∈ R, we say the weight of v is yUv ; so every representative v ∈ R has weight at least 1/2 by Property (3.1c). For a subset J ⊆ R of representatives, we
say J is big if the weight of J is at least `, i.e, yU (J) ≥ `; we say J is small otherwise. For any edge
e = (v, v 0 ) ∈ EMST , we consider the iteration in Kruskal’s algorithm in which the edge e is added to
MST. After the iteration we merged the partition Jv containing v and the partition Jv0 containing
v 0 into a new partition Jv ∪ Jv0 . If both Jv and Jv0 are small, then we call e a black edge. If Jv is
small and Jv0 is big, we call e a grey edge, directed from v to v 0 ; similarly, if Jv0 is small and Jv is
big, e is a grey edge directed from v 0 to v. If both Jv and Jv0 are big, we say e is a white edge. So,
we treat black and white edges as undirected edges and grey edges as directed edges.
We define a black component of MST to be a maximal set of vertices connected by black edges.
Let J be the set of all black components. Thus J indeed forms a partition of R. We contract
all the black edges in MST and remove all the white edges. The resulting graph is a forest ΥJ
of trees over black components in J . Each edge is a directed grey edge. Later in Lemma 3.4, we
show that the grey edges are directed towards the roots of the trees. For every component J ∈ J ,
we define L(J) := d(J, R \ J) to be the shortest distance between any representative in J and any
representative not in J.
A component J in the forest ΥJ may have many child-components. To make the forest binary,
we use the left-child-right-sibling binary-tree representation of trees. To be more specific, for every
component J 0 , we sort all its child-components J according to non-decreasing order of L(J). We add
a directed edge from the first child to J 0 and a directed edge between every two adjacent children
7
in the ordering, from the child appearing later in the ordering to the child appearing earlier. Let
Υ∗J be the new forest. Υ∗J naturally defines a new child-parent relationship between components.
Lemma 3.4. J and Υ∗J satisfy the following properties:
(3.4a) for every J ∈ J , there is a spanning tree over the representatives in J such that for every
edge (v, v 0 ) in the spanning tree we have d(v, v 0 ) ≤ L(J);
(3.4b) every root component J ∈ J of Υ∗J has yU (J) ≥ ` and every non-root component J ∈ J
has yU (J) < `;
(3.4c) every root component J ∈ J of Υ∗J has either yU (J) < 2` or |J| = 1;
(3.4d) for any non-root component J and its parent J 0 , we have L(J) ≥ L(J 0 );
(3.4e) for any non-root component J and its parent J 0 , we have d(J, J 0 ) ≤ O(`)L(J);
(3.4f) every component J has at most two children.
The rest of the section is dedicated to the proof of Lemma 3.4. We first prove some of the above
properties for the original forest ΥJ . We show that all black edges between the representatives
in J are considered before all the edges in J × (R \ J) in Kruskal’s algorithm. Assume otherwise.
Consider the first edge e in J × (R \ J) we considered. Before this iteration, J is not connected yet.
Then we add e to the minimum spanning tree; since J is a black component, e is gray or white. In
either case, the new partition J 0 formed by adding e will have weight more than `. This implies all
edges in J 0 × (R \ J 0 ) added later to the MST are not black. Moreover, J \ J 0 , J 0 \ J and J ∩ J 0 are
all non-empty. This contradicts the fact that J is a black component. Therefore, all black edges in
J has length at most L(J), implying Property (3.4a) .
Focus on a tree T in the initial forest ΥJ and any small black component J in T . All black
edges between the representatives in J are added to MST before any edge in J × (R \ J). The first
edge in J × (R \ J) added to MST is a grey edge directed from J to some other black component: it
is not white because J is small; it is not black since J is a black component. Thus, it is a grey edge
in T . Therefore, the growth of the tree T in Kruskal’s algorithm is as follows. The first grey edge
in T is added between two black components, one of them is big and the other is small. We define
the root of T to be the big component. At each time, we add a new small black component J to the
current T via a grey edge directed from J to T . (During this process, white edges incident to T may
be added.) So, the tree T is a rooted tree with grey edges, where all edges are directed towards the
root. So, Property (3.4b) holds. Moreover, the length of the grey edge between J and its parent J 0
is d(J, J 0 ) = L(J), which is stronger than Property (3.4e). Since d(J, J 0 ) ≥ d(J 0 , R \ J 0 ) = L(J 0 ),
we have Property (3.4d).
The root J of T is a big black component. Suppose it contains two or more representatives;
so it’s not a singleton. Consider the last black edge (v, v 0 ) added between Jv and Jv0 to make
J = Jv ∪ Jv0 . Since (v, v 0 ) is a black edge, both Jv and Jv0 are small, i.e. yU (Jv ) , yU (Jv0 ) < `.
Therefore, we have yU (J) = yU (Jv ) + yU (Jv0 ) < 2`, proving Property (3.4c).
Now, we move on to prove all the properties of the lemma for the final forest Υ∗J . We used the
left-child-right-sibling binary-tree representation of ΥJ to obtain Υ∗J . Thus , Property (3.4f) holds
for Υ∗J . Property (3.4a) is independent of the forest and thus still holds for Υ∗J . A component is
a root in ΥJ if and only if it is a root in Υ∗J . Thus, properties (3.4b) and (3.4c) are maintained
for Υ∗J . Since we sorted the children of a component according to L values before constructing the
left-child-right-sibling binary tree, Property (3.4d) holds for Υ∗J .
8
For every component J and its parent J 0 in the forest Υ∗J , we have L(J) = d(J, R \ J) =
d(J, J 00 ), where J 00 is the parent of J in the initial forest ΥJ . J 0 is either J 00 , or a child of J 00
in ΥJ . In the former case, we have d(J, J 0 ) = d(J, J 00 ) = L(J). In the latter case, we have that
d(J 0 , J 00 ) = L(J 0 ) ≤ L(J) = d(J, J 00 ). Due to Property (3.4a), we have a path connecting some
representative in J to some representative in J 0 , with internal vertices being representatives in J 00 ,
and all edges having length at most L(J). Moreover, there are at most 4` representatives in J 00 due
to Properties (3.4b), (3.4c), and (3.1c). Thus, we have d(J, J 0 ) ≤ O(`)d(J, J 00 ) = O(`)L(J). Thus,
Property (3.4e) holds for Υ∗J . This finishes the proof of Lemma 3.4.
3.3
Groups
In the third phase, we apply a simple greedy algorithm to the forest Υ∗J to partition the set J
of black components into a family G of groups, where each group G ∈ G contains many black
components that are connected in Υ∗J . By contracting each group G ∈ G, the forest Υ∗J over the
set J of black components becomes a forest ΥG over the set G of groups. Each group has a total
opening of Ω(`), unless it is a leaf-group in ΥG .
We partition the set J into groups using a technique similar to [4, 6]. For each rooted tree
T = (JT , ET ) in Υ∗J , we construct aPgroup G of black components as follows. Initially, let G contain
the root component of T . While J∈G yU (J) < ` and G 6= JT , repeat the following procedure.
Choose the component J ∈ JT \ G that is adjacent to G in T , with the smallest L-value, and add
J to G.
Thus, by the construction G is connected in T . After we have constructed the group G, we add
G to G. We remove all black components in G from T . Then, each T is broken into many rooted
trees; we apply the above procedure recursively for each rooted tree.
So, we have constructed a partition G for the set J of components. If for every G ∈ G, we
contract all components in G into a single node, then the rooted forest Υ∗J over J becomes a rooted
forest ΥG over the set G of groups. ΥG naturally defines a parent-child relationship over G. The
following lemma uses Properties (3.4a) to (3.4f) of J and the way we construct G.
Lemma 3.5. The following statements hold for the set G of groups and the rooted forest ΥG over
G:
(3.5a) any root group G ∈ G contains a single root component J ∈ J ;
P
(3.5b) if G ∈ G is not a root group, then J∈G yU (J) < 2`;
P
(3.5c) if G ∈ G is a non-leaf group, then J∈G yU (J) ≥ `;
(3.5d) let G ∈ G, G0 ∈ G be the
S parent of G, J ∈ G and v ∈ J, then the distance between v and
any representative in J 0 ∈G0 J 0 is at most O(`2 )L(J);
(3.5e) any group G has at most O(`) children.
Proof. For a root component J, we have yU (J) ≥ ` by Property (3.4b). Thus, any root group G
contains a single root component J, which is exactly Property (3.5a).
When constructing
the group G from the tree T = (JT , ET ), the terminating condition is
P
G = JT or J∈G yP
U (J) ≥ `. Thus, if G is not a leaf-group, then the condition G = JT does not
hold; thus we have J∈G yU (J) ≥ `, implying Property (3.5c).
By Property (3.4b), any non-root component J has yU (J) < `. Thus, if G is not a root group, the
terminating condition constructing G implies that G P
had total weight less than ` right before the
last black component was added to it. Then we have J∈G yU (J) < 2`, implying Property (3.5b).
9
Now, consider Property (3.5d). From Property (3.4d), it is easy to see that the group G
constructed from the tree T = (JT , ET ) has the following property: the L value of any component
in G is at most the L-value of any component in JT \ G. Let G be a non-root group and G0 be
its parent; let J ∈ G and J 0 ∈ G0 be black components. Thus, there is a path in Υ∗J from J to
J 0 , where components have L-values at most L(J). The edges in the path have length at most
O(`)L(J) by Property (3.4e). Moreover, Property (3.4a) implies that the representatives in each
component in the path are connected by edgesSof length at most L(J). Thus, we can find a path
from v to v 0 that go through representatives in J 00 ∈G∪G0 J 00 , and every edge in the path has length
at most O(`)L(J) = O(`)d(J, R\J). By Property (3.1c), (3.4b) and (3.4c), the total representatives
in the components contained in G (as well as in G0 ) is at most 4`. Thus, the distance between v
and v 0 is at most O(`2 )L(J), which is exactly Property (3.5d).
Finally, since the forest Υ∗J is binary and every group G ∈ G contains at most O(`) components,
we have that every group G contains at most O(`) children, implying Property (3.5e).
4
Constructing Local Solutions
In this section, we shall construct a local solution, or a distribution of local solutions, for a given
set V ⊆ R which is the union of some black components. A local solution for V contains a pair
U (V )
(S ⊆ U (V ), β ∈ R≥0 ), where S is the facilities we open in U (V ) and βi for each i ∈ U (V ) is
the amount of supply at i: the demand that can be satisfied by i. Thus βi = 0 if i ∈ U (V ) \ S.
We shall use the supplies at
PU (V ) to satisfy the xU (V ),C demands at V after the initial moving of
demands; thus, we require i∈U (V ) βi = xU (V ),C . There are two other main properties we need the
distribution to satisfy: (a) the expected size of S from the distribution is not too big, and (b) the
cost of matching the demands at V and the supplies at U (V ) is small.
We distinguish between concentrated black components and non-concentrated black components. Roughly speaking, a component J ∈ J is concentrated if in the fractional solution (x, y),
for most clients j ∈ C, j is either almost fully served by facilities in U (J), or almost fully served
by facilities in F \ U (J). We shall construct a distribution of local solutions for each concentrated
component J. We require Constraints (6) to (13) to be satisfied for B = U (J) (if not, we return
the set U (J) to the separation oracle) and let z B be the vector satisfying the constraints. Roughly
speaking, the z B -vector defines a distribution of
P local solutions for V . A local solution (S, β) is
good if S is not too big and the total demand i∈S βi satisfied by S is not too small. Then, our
algorithm randomly selects (S, β) from the distribution defined by z B , under the condition that
(S, β) is good. The fact that J is concentrated guarantees that the total mass of good local solutions
in the distribution is large; therefore the factors we lose due to the conditioning are small.
For non-concentrated components, we construct a single local solution (S, β), instead of a distribution of local solutions. Moreover, the construction is for the union V of some non-concentrated
components, instead of an individual component. The components that comprise V are close to
each other; by the fact that they are non-concentrated, we can move demands arbitrarily within
V , without incurring too much cost. Thus we can essentially treat the distances between representatives in V as 0. Then we are only concerned with two parameters for each facility i ∈ U (V ):
the distance from i to V and the capacity ui . Using a simple argument, the optimum fractional
local solution (that minimizes the cost of matching the demands and supplies) is almost integral:
it contains at most 2 fractionally open facilities. By fully opening the two fractional facilities, we
find an integral local solution with small number of open facilities.
10
The remaining part of this section is organized as follows. We first formally define concentrated
black components, and explain the importance of the definition. We then define the earth-moverdistance, which will be used to measure the cost of satisfying demands using supplies. The construction of local solutions for concentrated components and non-concentrated components will be
stated in Theorem 4.4 and Lemma 4.9 respectively.
4.1
Concentrated Black Components and Earth Mover Distance
The definition of concentrated black component is the same as that of [16], except that we choose
the parameter `2 differently.
P
Definition 4.1. Define πJ = j∈C xU (J),j (1 − xU (J),j ), for every black component J ∈ J . A black
component J ∈ J is said to be concentrated if πJ ≤ xU (J),C /`2 , and non-concentrated otherwise,
where `2 = Θ(`3 ) is large enough.
We use J C to denote the set of concentrated components and J N to denote the set of nonconcentrated components. The next lemma from [16] shows the importance of πJ . For the completeness of the paper, we include its proof here.
Lemma 4.2. For any J ∈ J , we have L(J)πJ ≤ O(1)DU (J) .
Proof. Let B = U (J). For every i ∈ B, j ∈ C, we have d(i, J) ≤ d(i, j) + 4dav (j) by Property (3.1d)
and the fact that i ∈ Uv for some v ∈ J. Thus,
X
X
L(J)π(J) = L(J)
xB,j (1 − xB,j ) = L(J)
xi,j xi0 ,j
j∈C,i∈B,i0 ∈F \B
j∈C
≤
X
i∈B,j∈C,i0 ∈F \B
=2
X
i∈B,j∈C
xi,j
xi,j xi0 ,j · 2d(i0 , J) ≤ 2
X
i∈B,j∈C
xi,j
X
xi0 ,j d(i0 , j) + d(j, i) + d(i, J)
i0 ∈F
X
dav (j) + d(j, i) + d(i, J) ≤ 2
xi,j 2d(i, j) + 5dav (j)
i∈B,j∈C
X
=2
(5Di ) = 10DB .
i∈B
The first inequality is by L(J) ≤ 2d(i0 , J) for any i0 ∈ F \ B = UR\J : d(i0 , R \ J) ≤ d(i0 , J) implies
L(J) = d(R \J, J) ≤ d(R \J, i0 )+d(i0 , J) ≤ 2d(i0 , J). The second inequality is by triangle inequality
and the third one is by d(i, J) ≤ d(i, j) + 4dav (j). All the equalities are by simple manipulations of
notations.
Recall that L(J) = d(J, R \ J) and xU (J),C is the total demand in J after the initial moving.
Thus, according to Lemma 4.2, if J is not concentrated, we can use DU (J) to charge the cost for
moving all the xU (J),C units of demand out of J, provided that the moving distance is not too
big compared to L(J). This gives us freedom for handling non-concentrated components. If J is
concentrated, the amount of demand that is moved out of J must be comparable to πJ ; this will
be guaranteed by the configuration LP.
In order to measure the moving cost of satisfying demands using supplies, we define the earth
mover distance:
11
Definition 4.3 (Earth Mover Distance). Given aP
set V ⊆ R P
with B = U (V ), a demand vector
V
B
α ∈ R≥0 and a supply vector β ∈ R≥0 such that
v∈V αv ≤
i∈B βi , the earth mover distance
P
from α to β is defined as EMDV (α, β) := inf f v∈V,i∈B f (v, i)d(v, i), where f is over all functions
from V × B to R≥0 such that
P
•
i∈B f (v, i) = αv for every v ∈ V ;
P
•
v∈V f (v, i) ≤ βi for every i ∈ B.
For some technical reason, we allow some fraction of a supply to be unmatched. From now on,
we shall use αv = xUv ,C to denote the amount of demand at v after the initial moving. For any set
V ⊆ R of representatives, we use α|V to denote the vector α restricted to the coordinates in V .
4.2
Distributions of Local Solutions for Concentrated Components
In this section, we construct distributions for components in J C , by proving:
Theorem 4.4. Let J ∈ J C and let B = U (J). Assume Constraints (6) to (13) are satisfied for
B. Then, we can find a distribution (φS,β )S⊆B,β∈RB of pairs (S, β), such that
≥0
(4.4a) sφ := E(S,β)∼φ |S| ∈ [yB , yB (1 + 2`πJ /xB,C )], and sφ = yB if yB > 2`,
and for every (S, β) in the support of φ, we have
(4.4b) |S| ∈ {bsφ c , dsφ e};
(4.4c) βi ≤ (1 + O(1/`))ui if i ∈ S and βi = 0 if i ∈ B \ S;
P
P
(4.4d)
i∈S βi = xB,C =
v∈J αv .
Moreover, the distribution φ satisfies
(4.4e) the support of φ has size at most nO(`) ;
(4.4f) E(S,β)∼φ EMDJ (α|J , β) ≤ O(`4 )DB .
To prove the theorem, we first construct a distribution ψ that satisfies most of the properties;
then we modify it to obtain the final distribution φ. Notice that a typical black component J has
yB ≤ 2`; however, when J is a root component containing a single representative, yB might be very
large. For now, let us just assume yB ≤ 2`. We deal with the case where |J| = 1 and yB > 2` at
the end of this section.
Since Constraints (6) to (13) are satisfied for B, we can use the z B variables satisfying these
constraints to construct a distribution ζ over pairs (χ ∈ [0, 1]B×C , µ ∈ [0, 1]B ), where µ indicates
the set of open facilities in B and χ indicates how the clients in J are connected to facilities in B.
Let S = {S ⊆ B : |S| ≤ `1 }Pand Se = S ∪ {⊥} as in Section 2. For simplicity, for any µ ∈ [0, 1]B ,
B×C , i ∈ B and j ∈ C, we shall use χ
we shallPuse µB to denote i∈B µP
i . For any χ ∈ [0, 1]
i,C to
P
P
denote j∈C χi,j , χB,j to denote i∈B χi,j , and χB,C to denote i∈B χi,C = j∈C χB,j .
The distribution ζ is defined as follows. Initially, let ζχ,µ = 0 for all χ ∈ [0, 1]B×C and µ ∈
[0, 1]B . For each S ∈ Se such that zSB > 0, increase ζχ,µ by zSB for the χ, µ satisfying χi,j =
B /z B , µ = z B /z B for every i ∈ B, j ∈ C. So, for every pair (χ, µ) in the support of ζ, we have
zS,i,j
i
S
S,i S
χ
≤
µ
,
χ
≤
ui µi for every i ∈ B, j ∈ C. Moreover, either µ is integral, or µB ≥ `1 . Since
i,j
i
i,C
P
B = 1, ζ is a distribution over pairs (χ, µ). It is not hard to see that E
z
(χ,µ)∼ζ χi,j = xi,j for
S∈Se S
every i ∈ B, j ∈ C and E(χ,µ)∼ζ µi = yi for every i ∈ B. The support of ζ has nO(`) size.
12
Definition 4.5. We say a pair (χ, µ) is good if
(4.5a) µB ≤ yB /(1 − 1/`);
(4.5b) χB,C ≥ (1 − 1/`)xB,C .
We are only interested in good pairs in the support of ζ. We show that the total probability of good pairs in the distribution ζ is large. Let Ξa denote the set of pairs (χ, µ) satisfying
Property (4.5a) and Ξb denote the set of pairs P
(χ, µ) satisfying Property (4.5b). Notice that
E(χ,µ)∼ζ µB = yB . By Markov inequality, we have (χ,µ)∈Ξa ζχ,µ ≥ 1/`. The proof of the following
lemma uses elementary mathematical tools.
P
Lemma 4.6.
(χ,µ)∈Ξ
/ b ζχ,µ ≤ `πJ /xB,C .
Proof. The idea is to use the property that J is concentrated. To get some intuition, consider the
case where πJ = 0. For every j ∈ C, either xB,j = 0 or xB,j = 1. Thus, all pairs (χ, µ) in the
support of ζ have χB,j = xB,j for every jP
∈ C; thus χB,C = xB,C .
Assume towards contradiction that
(χ,µ)∈Ξ
/ b ζχ,µ > `πJ /xB,C . We sort all pairs (χ, µ) in
the support of ζ according to descending order of χB,C . For any t ∈ [0, 1), and j ∈ C, define
gt,j ∈ [0, 1] as follows. Take the first pair (χ, µ) in the ordering such that the total ζ value of the
pairs (χ0 , µ0P
) before (χ, µ) in the ordering plus ζχ,µ is greater than t. Then, define gt,j = χB,j and
define gt = j∈C gt,j = χB,C .
Fix a client j ∈ C, we have
Z 1
Z 1
Z xB,j
(1 − 2t)dt =
1t<xB,j (1 − 2t)dt ≥
gt,j (1 − 2t)dt,
xB,j (1 − xB,j ) =
0
0
0
where 1t<xB,j is the indicator variable for the event that t < xB,j . The inequality comes from the
R1
R1
fact that 0 1t<xB,j dt = xB,j = 0 gt,j dt, gt,j ∈ [0, 1] for every t ∈ [0, 1), and 1 − 2t is a decreasing
function of t.
R1
πJ ≥ 0 gt (1−2t)dt. By our assumption that
P Summing up the inequality over all j ∈ C, we have
∗
(χ,µ)∈Ξ
/ b ζχ,µ > `πJ /xB,C , there exists a number t < 1 − `πJ /xB,C such that gt ≤ (1 − 1/`)xB,C
R1
for every t ∈ [t∗ , 1). As gt is a non-increasing function of g and 0 gt dt = xB,C , it is not hard
R1
to see that 0 gt (1 − 2t)dt is minimized when gt = (1 − 1/`)xB,C for every t ∈ [t∗ , 1) and gt =
xB,C −(1−1/`)xB,C (1−t∗ )
t∗
1/`+t∗ −t∗ /`
xB,C
t∗
=
Z
πJ ≥
0
t∗
for every t ∈ [0, t∗ ). We have
1/` + t∗ − t∗ /`
(1 − 2t)dt +
t∗
Z
1
t∗
!
(1 − 1/`)(1 − 2t)dt xB,C
1/` + t∗ − t∗ /` ∗
∗ 2
∗
∗ 2
xB,C
t − (t ) − (1 − 1/`) t − (t )
=
t∗
`πJ /xB,C
1
1 − t∗
= ∗ t∗ − (t∗ )2 xB,C =
xB,C >
xB,C = πJ ,
`t
`
`
P
leading to a contradiction. Thus, we have that (χ,µ)∈Ξ
/ b ζχ,µ ≤ `πJ /xB,C . This finishes the proof
of Lemma 4.6.
P
P
Overall, we have Q := (χ,µ) good ζχ,µ = (χ,µ)∈Ξa ∩Ξb ζχ,µ ≥ 1/` − `πJ /xB,C ≥ 1/` − 1/(2`) =
1/(2`), where the second inequality used the fact that πJ ≤ xB,C /(2`2 ) for J ∈ J C .
13
Now focus on each good pair (χ, µ) in the support of ζ. Since J ∈ J C and (χ, µ) ∈ Ξa ,
we have µB ≤ yB /(1 − 1/`) ≤ 2`/(1 − 1/`) < `1 (since we assumed yB ≤ 2`), if `1 is large
enough. So, µ ∈ {0, 1}B . Then, let S = {i ∈ B : µi = 1} be the set indicated by µ, and
βi = χi,C /(1 − 1/`) for every i ∈ B. For this (S, β), Property (4.4c) is satisfied, and we have
P
i∈B βi = χB,C /(1 − 1/`) ≥ xB,C . We then set ψS,β = ζχ,µ /Q. Thus, ψ indeed forms a distribution
over pairs (S, β). Moreover, the support of ζ has size nO(`) , so does the support of ψ. Thus
Property (4.4e) holds.
µB = yB . By Lemma 4.6,
Let sψP:= E(S,β)∼ψ |S| = E(χ,µ)∼ζ µB (χ, µ) good . Notice that E(χ,µ)∼ζ
). Since
we have (χ,µ)∈Ξ
/ b ζχ,µ ≤ `πJ /xB,C . Thus, E(χ,µ)∼ζ µB (χ, µ) ∈ Ξb ≤ yB /(1 − `πj /x
B,C
the condition
(χ,
µ)
∈
Ξ
requires
µ
to
be
upper
bounded
by
some
threshold,
E
µ
a
B
B (χ, µ) ∈
(χ,µ)∼ζ
Ξb ∩ Ξa can only be smaller. Thus, we have that sψ ≤ yB /(1 − `πJ /xB,C ) ≤ yB (1 + 2`πJ /xB,C ).
The proof of Property
(4.4f) for ψ is long and tedious.For simplicity, we use Ê[·] to denote
E(χ,µ)∼ζ · (χ, µ) good , and a = 1/(1 − 1/`) to denote the scaling factor we used to define β.
Indeed, we shall lose a factor O(`2 ) later and thus we shall prove Property (4.4f) for ψ with the
O(`2 ) term on the right:
Lemma 4.7. Ê[EMDJ (α|J , β)] ≤ O(`2 )DB , where β depends on χ as follows: βi = aχi,C for every
i ∈ B.
Proof. Focus on a good pair (χ, µ) and the β it defined: βi =P
aχi,C for every i ∈ B. We call
P α the
demand vector and β the supply vector. Since (χ, µ) is good, i∈B βi = aχB,C ≥ xB,C = v∈J αv .
Thus we can satisfy all the demands and EMD(α, β) is not ∞.
We satisfy the demands in two steps. In the first step, we give colorsPto the supplies and
demands;
each color is correspondent to a client j ∈ C. Notice that αv = j∈C xUv ,j and βi =
P
a j∈C χi,j . For every v ∈ J, j ∈ C, xUv ,j units of demand at v has color j; for every i ∈ B, j ∈ C,
aχi,j units of supply at i have color j. In this step, we match the supply and demand using the
following greedy rule: while for some j ∈ C, i, i0 ∈ B, there is unmatched demand of color j at v
and there is unmatched supply of color j at i, we match them as much as possible. The cost for
this step is at most the total cost of moving all supplies and demands of color j to j, i.e,
X
X
xi,j (d(v, i) + d(i, j)) + a
χi,j d(i, j)
v∈J,i∈Uv ,j∈C
≤
X
i∈B,j∈C
xi,C d(v, i) +
v∈J,i∈Uv
≤ O(1)DB +
X
(xi,j + aχi,j )d(i, j)
i∈B,j∈C
X
(xi,j + aχi,j )d(i, j),
by Lemma 3.2.
i∈B,j∈C
P
P
After this step, we have
j∈C max{xB,j − aχB,j , 0} ≤
j∈C max{xB,j − χB,j , 0} units of
unmatched demand.
In the second step, we match remaining demand and the supply. For every v ∈ J, i ∈ Uv , we
move the remaining supply at i to v. After this step, all the supplies and the demands are at J;
then we match them arbitrarily. The total cost is at most
X
X
aχi,C d(i, v) +
max{xB,j − χB,j , 0} × diam(J),
(14)
v∈J,i∈Uv
j∈C
where diam(J) is the diameter of J.
14
Notice that Ê[χi,j ] ≤ 2`xi,j since Pr(χ,µ)∼ζ [(χ, µ) good]
P ≥ 1/(2`) and E(χ,µ)∼ζ χi,j = xi,j . The
expected cost of the first step is at most O(1)DB + O(`) j∈C,i∈B
P xi,j d(i, j) = O(`)DB . Similarly,
the expected value of the first term of (14) is at most O(`) v∈J,i∈Uv xi,C d(i, v) ≤ O(`)DB by
Lemma 3.2.
Consider the second term of (14). Notice that Ê[max{xB,j − χB,j , 0}] ≤ xB,j . Also,
Ê[max{xB,j − χB,j , 0}] = Ê[max{(1 − χB,j ) − (1 − xB,j ), 0}]
≤ Ê[1 − χB,j ] ≤ 2`(1 − xB,j ).
So, Ê max{xB,j −χB,j , 0} ≤ min{xB,j , 2`(1−xB,j )} ≤ 3`xB,j (1−xB,j ): if xB,j ≥ 1−1/(2`) ≥ 2/3,
then we have 2`(1 − xB,j ) ≤ 2`(1 − xB,j ) · (3xB,j /2) = 3`xB,j (1 − xB,j ); if xB,j < 1 − 1/(2`), then
1 − xB,j > 1/(2`), implying xB,j ≤ 2`xB,j (1 − xB,j ).
P
Summing up the inequality over all clients j ∈ C, we have Ê
j∈C max{xB,C − χB,C , 0} ≤
O(`)πJ . So, the expected value of the second term of (14) is at most O(`)πJ ·diam(J) ≤ O(`2 )πJ L(J) ≤
O(`2 )DB , by Lemma 4.2. This finishes the proof of Lemma 4.7.
At this point, we may have sψ < yB . We can apply the following operation repeatedly. Take a
pair (S, β) with ψS,β > 0 and S ( B. We then shift some ψ-mass from the pair (S, β) to (B, β) so
as to increase sψ . Thus, we can assume Property (4.4a)
P holds for ψ.
Property (4.4d) may be unsatisfied: we only have i∈B βi ≥ xB,C for every (S, β)
P in the support
of ψ. To satisfy the property, we focus on each (S, β) in the support of ψ such that i∈B βi > xB,C .
0
By considering the
that achieves EMD(α|J , β), we can find a β 0 ∈ RB
≥0 such that βi ≤ βi
P matching
P
for every i ∈ B, i∈B βi0 = v∈J αv = xB,C , and EMD(α|J, β 0 ) = EMD(α|J, β). We then shift all
the ψ-mass at (S, β) to (S, β 0 ).
To sum up what we have so far, we have a distribution ψ over (S, β) pairs, that satisfies
Properties (4.4a), (4.4c), (4.4d), (4.4e) and Property (4.4f) with O(`4 ) replaced with O(`2 ). The
only Property that is missing is Property (4.4b); to satisfy the property, we shall apply the following
lemma to massage the distribution ψ.
Lemma 4.8. Given a distribution ψ over pairs (S ⊆ B, β ∈ [0, 1]B ) satisfying sψ := E(S,β)∼ψ |S| ≤
`1 , we can construct another distribution ψ 0 such that
0
≤ O(`2 )ψS,β for every pair (S, β);
(4.8a) ψS,β
(4.8b) every pair (S, β) in the support of ψ 0 has |S| ≤ dsψ e;
(4.8c) E(S,β)∼ψ0 max{|S|, bsψ c} ≤ sψ .
Property (4.8a) requires that the probability that a pair (S, β) happens in ψ 0 can not be too
large compared to the probability it happens in ψ. Property (4.8b) requires |S| ≤ dsψ e for every
(S, β) in the support of ψ 0 . Property (4.8c) corresponds to requiring |S| ≥ bsψ c: even if we count
the size of S as bsψ c if |S| ≤ bsψ c, the expected size is still going to be at most sψ .
Proof of Lemma 4.8. If sψ − bsψ c ≤
1 − 1/`, then
we shall throw0 away the pairs with |S| > bsψ c.
More formally, let Q = Pr(S,β)∼ψ |S| ≤ bsψ c and we define ψS,β
= ψS,β /Q if |S| ≤ bsψ c and
0
ψS,β = 0 if |S| ≥ bsψ c + 1. So, Property (4.8b) is satisfied. By Markov inequality, we have that
Q ≥ 1 − sψ /(bsψ c + 1) = (bsψ c − sψ + 1)/(bsψ c + 1) ≥ (1/`)/(bsψ c + 1) ≥ 1/(``1 + `) since
0
sψ ≤ `1 = O(`). Thus, ψS,β
≤ O(`2 )ψS,β for every pair (S, β), implying Property (4.8a). Every
pair (S, β) in the support of ψ 0 has |S| ≤ bsψ c and thus Property (4.8c) holds.
15
Now, consider the case where sψ − bsψ c > 1 − 1/`. In this case, sψ is a fractional number. Let
be the distribution obtained
from ψ by conditioning on pairs (S, β) with |S| ≤ dsψ e. By Markov
inequality, we have Pr(S,β)∼ψ |S| ≤ dsψ e ≥ 1 − sψ /(dsψ e + 1) ≥ 1 − sψ /(sψ + 1) ≥ 1/(`1 + 1) as
00 ≤ O(`)ψ
sψ ≤ `1 = O(`). So, ψS,β
S,β for every pair (S, β). Moreover, we have E(S,β)∼ψ 00 |S| ≤ sψ
since we conditioned on the event that |S| is upper-bounded by some-threshold; all pairs (S, β) in
the support of ψ 00 have |S| ≤ dsψ e.
Then we modify ψ 00 to obtain the final distribution ψ 0 . Notice that for a pair (S, β) with
|S| ≤ bsψ c, we have sψ − |S| ≤ sψ ≤ 2`1 (sψ − bsψ c). Thus,
ψ 00
X
(S,β):|S|≤bsψ c
00
ψS,β
(sψ − bsψ c) ≥
1
2`1
≥
1
2`1
X
(S,β):|S|≤bsψ c
X
(S,β):|S|=dsψ e
00
ψS,β
(sψ − |S|)
00
ψS,β
(dsψ e − sψ ),
where the second inequality is due to E(S,β)∼ψ00 |S| ≤ sψ .
00 . For every pair (S, β) such that |S| = ds e,
0
= ψS,β
For every pair (S, β) with |S| ≤ bsψ c, let ψS,β
ψ
P
0
00 /(2` ). Due to the above inequality, we have
0 (s − bs c) ≥
we define ψS,β
= ψS,β
ψ
1
ψ
ψ
(S,β):|S|≤bsψ c S,β
P
P
0
0 (|S| − s ), implying
max{|S|
−
s
,
bs
c − sψ } ≤ 0. Finally, we scale
ψ
ψ
ψ
ψ
ψ
(S,β) S,β
(S,β):|S|=dsψ e S,β
P
0
0
the ψ vector so that we have (S,β) ψS,β = 1; Properties (4.8b) and (4.8c) hold. The scaling factor
0
≤ O(`2 )ψS,β for every pair (S, β) and Property (4.8a)
is at most 2`1 = O(`). Overall, we have ψS,β
holds. This finishes the proof of Lemma 4.8.
With Lemma 4.8 we can finish the proof of Theorem 4.4 for the case yB ≤ 2`. We apply the
lemma to ψ to obtain the distribution ψ 0 . By Property (4.8a), Properties (4.4c), (4.4d) and (4.4e)
remain satisfied for ψ 0 ; Property (4.4f) also holds for ψ 0 , as we lost a factor of O(`2 ) on the expected
cost.
To obtain our final distribution φ, initially we let φS,β = 0 for every pair (S, β). For every
(S, β) in the support of ψ 0 , we apply the following procedure. If |S| ≥ bsψ c, then we increase φS,β
0 ; otherwise, take an arbitrary set S 0 ⊆ B such that S ⊆ S 0 and |S 0 | = bs c and increase
by ψS,β
ψ
φS 0 ,β by ψS,β . Due to Property (4.8b), every pair (S, β) in the support of φ has |S| ∈ {bsψ c , dsψ e}.
Property (4.8c) implies that sφ := E(S,β)∼φ |S| ≤ sψ ∈ [yB , (1 + 2`π)yB ]. If sφ < sψ , we increase
sφ using the following operation. Take an arbitrary pair (S, β) in the support of φ such that
|S| = bsψ c, let S 0 ⊇ S be a set such that S 0 ⊆ B and |S 0 | = dsψ e, we decrease φS,β and increase
φS 0 ,β . Eventually, we can guarantee sφ = sψ ; thus Properties (4.4a) and (4.4b) are satisfied. This
finishes the proof of Theorem 4.4 when yB ≤ 2`.
Now we handle the case where yB > 2`. By Properties (3.4b) and (3.4c), J is a root black
component that contains a single representative v and yUv =B > 2`. First we find a nearly integral
solution with at most 2` + 2 open facilities. Then we close two facilities serving the minimum
amount of demand and spread their demand among the remaining facilities. Since there is at least
2` open facilities remaining, we increase the amount of demand at any open facility by no more
than a factor of O(1/`).
x
Let u0i = yi,C
≤ ui . We may scale u0i by a factor of 1+O(1/`) during the course of the algorithm.
i
16
Consider the following LP with variables {λi }i∈B :
X
min
u0i λi d(i, v)
s.t.
(15)
i∈Uv
X
u0i λi = xB,C ;
i∈B
X
λi = yB ;
i∈B
λi ∈ [0, 1],
∀i ∈ B.
P
By setting λi = yi , we obtain a solution to LP(15) of value i∈Uv xi,C d(i, v) ≤ O(1)DUv , by
Lemma 3.2. So, the value of LP(15) is at most O(1)DUv . Fix on such an optimum vertex-point
solution λ of LP(15). Since there are only two non-box-constraints, λ has at most two fractional
λi . Moreover, as yUv ≥ 2`, there are at least 2` facilities in the support of λ.
We shall reduce the size of the support of λ by 2, by repeating the following
twice.
P procedure
∗
0
0
0
∗
∗
Consider the i in the support with the smallest λi ui∗ value. Let a := λi ui∗ / i∈Uv λi ui ≤ O( 1` ),
we then scale u0i P
by a factor of 1/(1 − a) ≤ 1 + O(1/`) for every i ∈ Uv \ i∗ and change λi∗ to 0.
So, we still have i∈Uv u0i λi = xUv ,C . The value of the objective function is scaled by a factor of at
most 1 + O(1/`).
Let S = {i ∈ Uv : λi > 0} and let βi = λi u0i for every
P i ∈ Uv . So, |S| ≤ yUv Properties (4.4c)
and (4.4d) are satisfied. Moreover, EMDJ (α|J , β) ≤ i∈Uv βi d(i, v) ≤ O(1)DUv since the value of
LP(15) is O(1)DUv and we have scaled each βi by at most a factor of 1 + O(1/`).
If we let φ contains the single pair (S, β) with probability 1, then all properties from (4.4c) to
(4.4f) are satisfied. To satisfy Properties (4.4a) and (4.4b), we can manually add facilities to S with
some probability, as we did before for the case yB ≤ 2`. This finishes the proof of Theorem 4.4.
4.3
Local Solutions for Unions of Non-Concentrated Components
In this section, we construct a local solution for the union V of some close non-concentrated black
components.
S
Lemma 4.9. Let J 0 ⊆ J N be a set of non-concentrated black components, V = J∈J 0 J and
B = U (V ). Assume there exists v ∗ ∈ R such that d(v, v ∗ ) ≤ O(`2 )L(J) for every J ∈ J 0 and
v ∈ J. Then, we can find a pair (S ⊆ B, β ⊆ RB
≥0 ) such that
(4.9a) |S| ∈ dyB e , dyB e + 1 ;
(4.9b) βi ≤ ui if i ∈ S and βi = 0 if i ∈ B \ S;
P
P
(4.9c)
i∈S βi = xB,C =
v∈V αv ;
(4.9d) EMDV (α|V , β) ≤ O(`2 `2 )DB .
Proof. We shall use an algorithm similar to the one we used for handling the case where yB > 2` in
x
Section 4.2. Again, for simplicity, we let u0i = yi,C
≤ ui to be the “effective capacity” of i. Consider
i
the following LP with variables {λi }i∈B :
X
min
u0i λi d(i, v) + `2 L(J) s.t.
(16)
J∈J 0 ,v∈J,i∈Uv
X
i∈B
u0i λi = xB,C ;
X
λi = yB ;
i∈B
17
λi ∈ [0, 1],
∀i ∈ B.
By setting λi = yi , we obtain a valid solution to the LP with the objective value
X
xi,C d(i, v) + `2 L(J)
J∈J 0 ,v∈J,i∈Uv
≤
X
xi,C d(i, v) + `2
X
J∈J 0
v∈V,i∈Uv
2
≤ O(1)DB + ` `2
X
xU (J),C L(J) ≤
X
v∈V
O(1)DUv + `2
X
`2 πJ L(J)
J∈J 0
2
O(1)DU (J) = O(` `2 )DB ,
(17)
J∈J 0
by Lemma 3.2 and Lemma 4.2. So, the value of LP(16) is at most O(`2 `2 )DB .
Fix such an optimum vertex-point solution λ of LP (16). Since there are only two non-boxconstraints, every vertex-point λ of the polytope has at most two fractional λi .
Let S = {i ∈ B : λi > 0} and let βi = λi u0i for every i ∈ B. So, Properties (4.9a), (4.9b) and
(4.9c) are satisfied.
Now we prove Property (4.9d). To compute EMDV (α|V , β), we move all demands in α|V and
all supplies in β to v ∗ . The cost is
X
X
αv d(v, v ∗ ) +
βi d(i, v ∗ )
v∈V
≤
X
i∈B
xUv ,C O(`2 )L(J) +
J∈J 0 ,v∈J
X
J∈J 0 ,v∈J,i∈Uv
βi d(i, v) + O(`2 )L(J) ≤ O(`2 `2 )DB ,
where the O(`2 `2 )DB for the first term was proved in (17) and the bound O(`2 `2 )DB for the
second term is due to the fact that γ is an optimum solution to LP(16). This finishes the proof of
Lemma 4.9.
5
Rounding Algorithm
In this section we describe our rounding algorithm. We start by giving the intuition behind the
algorithm. For each concentrated component J ∈ J , we construct a distribution of
S local solutions
using Theorem 4.4. We shall construct a partition V N of the representatives in J∈J N J so that
each V ∈ V N is the union of some nearby components in J N . For each set V ∈ V N , we apply
Lemma 4.9 to construct a local solution. If we independently and randomly choose a local solution
from every distribution we constructed, then we can move all the demands to the open facilities at
a small cost, by Property (4.4f) and Property (4.9d).
However, we may open more than k facilities, even in expectation. Noticing that the fractional
solution opens yB facilities in a set B, the extra number of facilities come from two places. In
Property (4.4a) of Theorem 4.4, we may open in expectation yB · 2`πJ /xB,C more facilities in
B than yB . Then in Property (4.9a) of Lemma 4.9, we may open dyB e or dyB e + 1 facilities in
B. To reduce the number of open facilities to k, we shall shut down (or remove) some alreadyopen facilities and move the demands satisfied by these facilities to the survived open facilities:
a concentrated component J ∈ J C is responsible for removing yB · 2`πJ /xB,C < 1 facilities in
expectation; a set V ∈ V N is responsible for removing up to 2 facilities. Lemma 4.2 allows us to
bound the cost of moving demands caused by the removal, provided that the moving distance is
not too big. To respect the capacity constraint up to a factor of 1 + , we are only allowed to scale
18
concentrated components
non-concentrated components
groups
sets in V N
(a)
(b)
Figure 2: Figure (a) gives the forest Υ∗J over J and the set G of groups (denoted by empty
polygons). Figure (b) gives V N : each set V ∈ V N is the union of components in a solid polygon.
the supplies of the survived open facilities by a factor of 1 + O(1/`). Both requirements will be
satisfied by the forest structure over groups and the fact that each non-leaf group contains Ω(`)
fractional opening (Property (3.5c)). Due to the forest structure and Property (3.5c), we always
have enough open facilities locally that can support the removing of facilities.
In order to guarantee that we always open k facilities, we need to use a dependent rounding
procedure for opening and removing facilities. As in many of previous algorithms, we incorporate
the randomized rounding procedure into random selections of vertex points of polytopes respecting
marginal probabilities. In many cases, a randomized selection procedure can be derandomized since
there is an explicit linear objective we shall optimize.
We now formally describe our rounding algorithm. For every group G ∈ G, we use ΛG to
denote the set of child-groups of G. We construct a partition JC of J C as follows. For each root
group G ∈ G, we add G ∩ J C to JC if it is not empty. For each non-leaf group G ∈ G, we add
S
0
C
C
N
N in the same way,
G0 ∈ΛG (G ∩ J ) to J , if it is not empty. We construct the partition J for J
N
0
N
except that
components in J N . We also define
S we consider
S a set V as follows: for every J ∈ J ,
N
N
we add J∈J 0 J to V ; thus, V forms a partition for J∈J N J. See Figure 2 for the definition of
V N.
In Section 5.1, we describe the procedure for opening a set S ∗ of facilities, whose cardinality
may be larger than k. Then in Section 5.2, we define the procedure remove, which removes one
open facility. We wrap up the algorithm in Section 5.3.
5.1
Constructing Initial Set S ∗ of Open Facilities
In this section, we open a set S ∗ of facilities, whose cardinality may be larger than k, and construct
a supply vector β ∗ ∈ RF≥0 such that βi∗ = 0 if i ∈
/ S ∗ . (S ∗ , β ∗ ) will be the concatenation of all local
solutions we constructed.
It is easy to construct local solutions forSnon-concentrated components. For each set J 0 ∈ JN
of components and its correspondent V = J∈J 0 J ∈ V N , we apply Lemma 4.9 to obtain a local
U (V )
solution S ⊆ U (V ), β ∈ R≥0 . Then, we add S to S ∗ and let βi∗ = βi for every i ∈ U (V ). Notice
that J 0 either contains a single root black component J, or contains all the non-concentrated black
components in the child-groups of some group G. In the former case, the diameter of J isSat most
O(`)L(J) by Property (3.4a); in the latter case, we let v ∗ be an arbitrary representative in J 0 ∈G J 0
19
and then any representative v ∈ J, J ∈ J 0 has d(v, v ∗ ) ≤ O(`2 )L(J) by Property (3.5d). Thus, all
the properties in Lemma 4.9 are satisfied.
For concentrated components, we only obtain distributions of local solutions by applying Theorem 4.4. For every J ∈ J C , we check if Constraints (6) to (13) are satisfied for B = U (J). If
not, we return a separation plane for the fractional
solution; otherwise we apply Theorem 4.4 to
each component J to obtain a distribution φJS,β S⊆U (J),β∈RU (J) . To produce local solutions for
≥0
concentrated components, we shall use a dependent rounding procedure that respects the marginal
probabilities. As mentioned earlier, we shall define a polytope and the procedure randomly selects
a vertex point of the polytope.
We let sJ := sφJ := E(S,β)∼φJ |S| be the expectation of |S| according to distribution φJ . For
notational convenience, we shall use a ≈ b to denote a ∈ bbc , dbe . Consider the following polytope
J }
2
P defined by variables {ψS,β
J∈J C ,S,β and {qJ }J∈J C .
J
ψS,β
, pJ ∈ [0, 1] ∀J ∈ J C , S, β;
(18)
X
(19)
S,β
J
= 1,
ψS,β
∀J ∈ J C ;
S,β
X X
J∈J C
J∈J 0
X
S,β
X X
J∈J 0
X
S,β
J
|S| − qJ ≈
ψS,β
qJ ≤ 1,
∀J 0 ∈ JC ;
J
|S| − qJ ≈ yU (J) , ∀J ∈ J C ;
ψS,β
X
yU (J) ,
J∈J 0
∀J 0 ∈ JC ;
X
J
yU (J) .
|S| − qJ ≈
ψS,β
(20)
(21)
(22)
(23)
J∈J C
In the above LP, ψ J is the indicator vector for local solutions for J and qJ indicates whether J
is responsible for removing one facility; if qJ = 1, we shall call remove(J) later. Up to changing of
variables, any vertex point of P is defined by two laminar families of tight constraints and thus P
is integral:
Lemma 5.1. P is integral.
Proof. To avoid negative coefficients, we shall let qJ0 = 1 − qJ and focus on ψ and q 0 variables.
Consider the set of tight constraints that define a vertex point. The tight constraints from (18),
(19) and (20) define a matroid base polytope for a laminar matroid.
For each JP∈ J , every pair (S, β) in the support of φJ has |S| ≈ sJ . Thus, Constraint (21) is
J (|S| − bs c) + q 0 ≈ y
equivalent to S,β ψS,β
J
U (J) + 1 − bsJ c. This is true since Constraint (19) holds
J
J
and ψ is a distribution. We do the same transformation for Constraints (22) and (23). It is easy
to see that the tight constraints from (21), (22) and (23) also define the matroid-base-polytope for
a laminar matroid.
Thus, by the classic matroid theory, the set of tight constraints define an integral solution; thus
P is integral.
∗J = φJ
∗
C
We set ψS,β
S,β and qJ = sJ − yU (J) for every J ∈ J and (S, β). Then,
Lemma 5.2. (ψ ∗ , q ∗ ) is a point in polytope P.
2
For every J ∈ J C , we only consider the pairs (S, β) in the support of φJ ; thus the total number of variables is
.
O(`)
n
20
Proof. Notice that sJ is the expected size of S according to the distribution φJ , while yU (J) is the
budget for the number of open facilities open in U (J). So qJ∗ is the expected number of facilities that
∗
∗
go beyond the budget. P
It is easy to seeP
that Constraints (18) and (19)
P hold for ψ and q . For every
0
C
∗
J ∈ J , we have that J∈J 0 qJ ≤ 2` J∈J 0 yU (J) πJ /xU (J),C ≤ 2` J∈J 0 yU (J) /`2 ≤ 2` × O(`2 )/`2
due to Properties (3.5e) and (4.4a). This at most P
1 if `2 = Θ(`3 ) is large enough. (If J 0 contains
∗
a root component J which has yU (J) > 2` then
J∈J 0 qJ = 0.) Thus, Constraint (20) holds.
P
∗J
∗
∗
∗ ∗
S,β ψS,β |S| − qJ = sJ − qJ = yU (J) . So, Constraints (21), (22) and (23) hold. So, (ψ , p ) is a
point in P.
J ] = ψ ∗J = φJ
We randomly select a vertex point (ψ, q) of P such that E[ψS,β
S,β
S,β for every
C
∗
C
J ∈ J , (S, β), and E[qJ ] = qJ = sJ − yU (J) for every J ∈ J . Since ψ is integral, for every J ∈ J ,
U (J)
J = 1; we add S to S ∗ and let
there is a unique local solution S ⊆ U (J), β ∈ R≥0
such that ψS,β
∗
βi = βi for every i ∈ U (J).
This finishes the definition of the initial S ∗ and β ∗ . Let α∗ = α (recall that αv = xUv ,C is the
demand at v after the initial moving, for every v ∈ R) be the initial demand vector. Later we shall
remove facilities from S ∗ and update α∗ and β ∗ . S ∗ , α∗ , β ∗ satisfy the following properties, which
will be maintained as the rounding algorithm proceeds.
P
P
∗
C
N
∗
(5.3a)
v∈V βv for every V ∈ J ∪ V ;
v∈V αv =
P
∗
(5.3b)
v∈R αv = |C|.
P
Property
(5.3a) is due to Properties (4.4d) and (4.9c). Property (5.3b) holds since v∈R αv∗ =
P
v∈R xUv ,C = xF,C = |C|.
5.2
The remove procedure
In this section, we define the procedure remove that removes facilities from S ∗ and updates α∗ and
β ∗ . The procedure takes a set V ∈ J C ∪ V N as input. If V is a root black component, then we let
G = {V } be the root group containing V ; if V is a non-root concentrated component, let G be the
parent group of the group containing V ; otherwise V is the union of non-concentrated
components
S
in all child-groups of some group, and we let G be this group. Let V 0 = J 0 ∈G J 0 . Before calling
remove(V ), we require the following properties to hold:
(5.4a) |S ∗ ∩ U (V )| ≥ 1;
(5.4b) S ∗ ∩ U (V 0 ) ≥ ` − 6.
While maintaining Properties (5.3a) and (5.3b), the procedure remove(V ) will
(5.5a) remove from S ∗ exactly one open facility, which is in U (V ∪ V 0 ),
(5.5b) not change α∗ |R\(V ∪V 0 ) and β ∗ |F \U (V ∪V 0 ) ,
(5.5c) increase αv∗ by at most a factor of 1 + O(1/`) for every v ∈ V ∪ V 0 and increase βi∗ by at
most a factor of 1 + O(1/`) for every i ∈ U (V ∪ V 0 ).
Moreover,
(5.5d) the moving cost for converting the old α∗ to the new α∗ is at most O(`2 )βi∗∗ L(J) for some
black component J ⊆ V and facility i∗ ∈ U (J);
(5.5e) for every V 00 ∈ J C ∪ V N , EMDV 00 α∗ |V 00 , β ∗ |U (V 00 ) will be increased by at most a factor of
1 + O(1/`).
21
Before formally describe the procedure remove(V ), we first highlight some key ideas. Assume
V is not a root component. We choose an arbitrary
facility i ∈ S ∗ ∩ U (V ). Notice that there are
P
∗
0
∗
∗
Ω(`) facilities in S ∩ U (V ). If the βi ≤ v0 ∈V 0 αv0 /`, then we can shut down i and send the
demands that should be sent to i to V 0 . We only need to increase the supplies in U (V 0 ) by a factor
of 1 + O(1/`). Otherwise, we shall shut down the facility i0 ∈ S ∗ ∩ U (V 0 ) with the smallest βi∗0 value.
Since there are at least Ω(`) facilities in U (V 0 ), we can satisfy the βi∗0 units of unsatisfied demands
using other facilities in S ∗ ∩ U (V 0 ). For this i0 , we have βi∗0 ≤ O(1)βi∗ . Thus, the total amount
of demands that will be moved is comparable to βi∗ . In either case, the cost of redistributing the
demands is not too big. When V is a root component, we shall shut down the facility i0 ∈ S ∗ ∩U (V )
with the smallest βi∗0 value.
We now formally describe the procedure remove(V ). We first consider the case that V is not a
root component. So, V is either a non-root component in J C , or the union of all non-concentrated
components in all child-groups of the group G. In this case, V ∩ V 0 = ∅. Let i ∈ U (V ) ∩ S ∗ be an
arbitrary facility; due to Property (5.4a), we can find such an i. Let J ⊆ V be the component that
contains i.
P
If a := βi∗ / v0 ∈V 0 αv∗0 ≤ 1` , then we shall shutdown i. Consider the matching f : V × U (V ) →
R≥0 between α∗ |V and β ∗ |U (V ) that achieves EMDV α∗ |V , β ∗ |U (V ) . (Due to Property (5.3a), the
total supply equals the total demand.) For every v ∈ V , we shall move fP
(v, i) units of demand
from v to V 0 . The total amount of demand moved from V to V 0 is exactly v∈V fv,i = βi∗ . Every
v 0 ∈ V 0 will receive aαv∗0 units of demand. We update α∗ to be the new demand vector: decrease αv∗
by f (v, i) for every v ∈ V and scale αv∗0 by a factor of (1 + a) for every v 0 ∈ V 0 . By Property (3.5d),
the cost of moving the demands is at most βi∗ O(`2 )L(J); thus, Property (5.5d) holds.
We remove i from S ∗ , change βi∗ to 0, and for every i0 ∈ U (V 0 ), scale βi∗0 by a factor of (1 + a).
For this new α∗ and β ∗ vector, EMDV (α∗ , β ∗ ) will not increase. α∗ |V 00 and β ∗ |U (V 00 ) are scaled by
a factor of 1 + a ≤ 1 + 1/` for every V 00 ∈ J C ∪ V N such that V 00 ⊆ V 0 . Thus Properties (5.5c)
and (5.5e) are satisfied. Properties (5.5a) and (5.5b) are trivially true. Moreover, we maintained
Properties (5.3a) and (5.3b).
Now consider the case a > 1/`. In this case, we shall remove the facility i0 ∈ S ∗ ∩ U (V 0 ) with
we have |S ∗ ∩ U (V 0 )| ≥ ` − 6 before we run remove(V )
the smallest βi∗0 value from S ∗ . Notice that
P
due to Property (5.4b). Let a0 := βi0 / i00 ∈U (V 0 ) βi∗00 ; so, we have a0 ≤ 1/(` − 6). To remove the
facility i0 , we consider the function f that achieves EMDV 0 (α∗ |V 0 , β ∗ |U (V
P0 ) ). We shall redistribute
0
0
0
0
the demands in V so that the new demand at v ∈ V will be (1 + a ) i00 ∈U (V 0 )\{i0 } f (v 0 , i00 ). We
remove i0 from S ∗ , change βi∗0 to 0 and scale up βi∗00 for all other i00 ∈ U (V 0 ) \ {i0 } by (1 + a0 ). Then,
the total cost for redistributing the demands in this procedure will be at most βi∗0 O(`)L(J), due to
Property (3.4d). This is at most O(`)βi∗ L(J) since a > 1/` and a0 ≤ 1/(` − 6). So, Properties (5.5a)
to (5.5e) are satisfied and Properties (5.3a) and (5.3b) are maintained.
The case where V is a root component can be handled in a similar way. In this case, we have
G = {V } and V 0 = V . By Property (5.4b), there are at least ` − 6 facilities in S ∗ ∩ U (V ). Then
we can remove the facility i ∈ U (V ) ∩ S ∗ with the smallest βi∗ . Using the same argument as above,
we can guarantee Properties (5.5a) to (5.5e) and maintain Properties (5.3a) and (5.3b).
5.3
Obtaining the Final Solution
To obtain our final set S ∗ of facilities, we call the remove procedures in some order. We consider
each group G using the top-to-bottom order. That is, before we consider a group G, we have
22
already considered its parent group. If G is a root group, then it contains a single root component
J. If J ∈ J N , repeat the the following procedure twice: if there is some facility in S ∗ ∩ U (J) then
we call remove(J). If J ∈ J C and
S qJ = 1 then we call remove(J). Now if G is a non-leaf group,
then do the following. Let V = G0 ∈ΛG ,J∈G0 ∩J N J. Repeat the following procedure twice: if there
is some facility in S ∗ ∩ U (V ) then we call remove(V ). For every G0 ∈ ΛG and J ∈ G0 ∩ J C such
that qJ = 1 we call remove(J).
Lemma 5.6. After the above procedure, we have |S ∗ | ≤ yF ≤ k.
Proof. We first show that whenever we call remove(V ), Properties (5.4a) and (5.4b) hold. For any
concentrated component J with qJ = 1, we have called remove(J). Notice that if qJ = 1, then
initially we have |S ∗ ∩ U (J)| ≥ 1 due to Constraint (21). Due to the top-down order of considering
components, and Property (5.5a), we have never removed a facility in S ∗ ∩ U (J) before calling
remove(J). Thus, Property (5.4a) holds. For V ∈ V N , we check if |S ∗ ∩ U (V )| ≥ 1 before we call
remove(V ) and thus Property (5.4a) holds.
S
Property (5.4b). For any non-leaf group G, initially, we have S ∗ ∩ J∈G U (J) ≥
PNow consider
J∈G yU (J) ≥ ` where the first inequality is due to Property (4.9a) and Constraint (22) and the
second is due to Property (3.5c). We may remove a facility from the set when we call remove(V ) for
V satisfying one of the following conditions: (a) V is a concentrated component in G or in a child
group of G, (b) V is the union of the non-concentrated components in the child-groups of G or (c)
V contains the non-concentrated components in G. For case (a), we removed at most 2 facilities
due to Constraint (20). For each
S (b) and (c), we
P remove at most 2 facilities. Thus, we shall remove
at most 6 facilities from S ∗ ∩ J∈G U (J) ≥ J∈G yU (J) . Thus, Property (5.4b) holds.
Thus, every call of remove is successful. For a concentrated component
J with
qJ = 1, we called
N
∗
remove(J) once. For each V ∈ V , initially we have |S ∩ U (V )| ∈ yU (V ) , yU (V ) + 1 . Before
calling remove(V ), we have never removed a facility from S ∗ ∩ U (V ). Thus, the number of times
we call remove(V ) is at least the initial value of |S ∗ ∩ U (V )| minus yU (V ) . Overall, the number of
P
P
P
J −q
+ V ∈V N yU (V ) <
ψ
facilities in S ∗ after the removing procedure is at most J∈J C
J
S,β S,β
P
P
J∈J C yU (J) +1+ V ∈V N yU (V ) = yF +1 ≤ k+1, where the first inequality is due to Constraint (23).
Since |S ∗ | is an integer, we have that |S ∗ | ≤ k.
By Properties (5.5b) and (5.5c), and Constraint (20), our final βi∗ is at most 1 + O(1/`) times
the initial βi∗ for every i ∈ V . Finally we have βi∗ ≤ (1 + O(1/`))ui for every i ∈ F . Thus, the
capacity constraint is violated by a factor of 1 + if we set ` to be large enough.
It remains to bound the expected cost of the solution S ∗ ; this is done by bounding the cost for
transferring the original α∗ to the final α∗ , as well as the cost for matching our final α∗ and β ∗ .
We first focus on the transferring cost. By Property (5.5e), when we call remove(V ), the
transferring cost is at most O(`2 )βi∗∗ L(J) for some black component J ⊆ V and i∗ . Notice that
βi∗∗ is scaled by at most a factor of (1 + O(1/`)), we always have βi∗∗ ≤ (1 + O(1/`))αU (J),C . So,
the cost is at most O(`2 )xU (J),C L(J). If V is the union of some non-concentrated components,
then this quantity is at most O(`2 )`2 πJ L(J) ≤ O(`2 `2 )DU (J) ≤ O(`2 `2 )DU (V ) . We call remove(V )
at most twice, thus the contribution of V to the transferring cost is at most O(`2 `2 )DU (V ) . If V
is a concentrated component J, then the quantity might be large. However, the probability we
call remove(J) is E[qJ ] = qJ∗ = sJ − yU (J) ≤ 2`yU (J) πJ /xU (J),C if yU (J) ≤ 2` and it is 0 otherwise
(by Property (4.4a)). So, the expected contribution of this V to the transferring cost is at most
23
O(`2 )xU (J),C L(J) × 2`yU (J) πJ /xU (J),C ≤ O(`4 )πJ L(J) ≤ O(`4 )DU (J) by Lemma 4.2. Thus, overall,
the expected transferring cost is at most O(`5 )DF = O(`5 )LP.
Then we consider
the matching cost. Since we maintained Property (5.3a), the matching cost
P
is bounded by V ∈J C ∪V N EMDV (α∗ |V , β ∗ |U (V ) ). Due to Property (5.5e), this quantity has only
increased by a factor of 1 + O(1/`) during the course
of removing facilities.
For the initial α∗ and
P
P
∗
4
β , the expectation of this quantity is at most J∈J C O(` )DU (J) + V ∈V N O(`2 `2 )DU (V ) due to
Properties (4.4f) and (4.9d). This is at most O(`5 )DF = O(`5 )LP.
We have found a set S ∗ of at most k facilities and a vector β ∗ ∈ RF≥0 such that βi∗ = 0 for every
i∈
/ S ∗ and βi∗ ≤ (1 + O(1/`))ui . If we set ` = Θ(1/) to be large enough, then βi∗ ≤ (1 + )ui . The
cost for matching the α-demand vector and the β ∗ vector is at most O(`5 )LP = O(1/5 )LP. Thus,
we obtained a O(1/5 )-approximation for CKM with (1 + )-capacity violation.
References
[1] Karen Aardal, Pieter L. van den Berg, Dion Gijswijt, and Shanfei Li. Approximation algorithms for hard capacitated k-facility location problems. European Journal of Operational
Research, 242(2):358 – 368, 2015.
[2] Hyung-Chan An, Mohit Singh, and Ola Svensson. LP-based algorithms for capacitated facility
location. In Proceedings of the 55th Annual IEEE Symposium on Foundations of Computer
Science, FOCS 2014.
[3] V. Arya, N. Garg, R. Khandekar, A. Meyerson, K. Munagala, and V. Pandit. Local search
heuristic for k-median and facility location problems. In Proceedings of the thirty-third annual
ACM symposium on Theory of computing, STOC ’01, pages 21–29, New York, NY, USA, 2001.
ACM.
[4] Jarosaw Byrka, Krzysztof Fleszar, Bartosz Rybicki, and Joachim Spoerhase. Bi-factor approximation algorithms for hard capacitated k-median problems. In Proceedings of the 26th Annual
ACM-SIAM Symposium on Discrete Algorithms (SODA 2015).
[5] Jarosaw Byrka, Thomas Pensyl, Bartosz Rybicki, Aravind Srinivasan, and Khoa Trinh. An
improved approximation for k-median, and positive correlation in budgeted optimization. In
Proceedings of the 26th Annual ACM-SIAM Symposium on Discrete Algorithms (SODA 2015).
[6] Jarosaw Byrka, Bartosz Rybicki, and Sumedha Uniyal. An approximation algorithm for uniform capacitated k-median problem with 1 + capacity violation, 2015. arXiv:1511.07494.
[7] Robert D. Carr, Lisa K. Fleischer, Vitus J. Leung, and Cynthia A. Phillips. Strengthening
integrality gaps for capacitated network design and covering problems. In Proceedings of the
Eleventh Annual ACM-SIAM Symposium on Discrete Algorithms, SODA ’00, pages 106–115,
Philadelphia, PA, USA, 2000. Society for Industrial and Applied Mathematics.
[8] M. Charikar and S. Guha. Improved combinatorial algorithms for the facility location and
k-median problems. In In Proceedings of the 40th Annual IEEE Symposium on Foundations
of Computer Science, pages 378–388, 1999.
24
[9] M. Charikar, S. Guha, . Tardos, and D. B. Shmoys. A constant-factor approximation algorithm
for the k-median problem (extended abstract). In Proceedings of the thirty-first annual ACM
symposium on Theory of computing, STOC ’99, pages 1–10, New York, NY, USA, 1999. ACM.
[10] Julia Chuzhoy and Yuval Rabani. Approximating k-median with non-uniform capacities. In
SODA 05, pages 952–958, 2005.
[11] Sudipto Guha. Approximation Algorithms for Facility Location Problems. PhD thesis, Stanford, CA, USA, 2000.
[12] K. Jain, M. Mahdian, and A. Saberi. A new greedy approach for facility location problems. In
Proceedings of the thiry-fourth annual ACM symposium on Theory of computing, STOC ’02,
pages 731–740, New York, NY, USA, 2002. ACM.
[13] K Jain and V. V. Vazirani. Approximation algorithms for metric facility location and k-median
problems using the primal-dual schema and Lagrangian relaxation. J. ACM, 48(2):274–296,
2001.
[14] Shanfei Li. An improved approximation algorithm for the hard uniform capacitated k-median
problem. In APPROX ’14/RANDOM ’14: Proceedings of the 17th International Workshop on
Combinatorial Optimization Problems and the 18th International Workshop on Randomization
and Computation, APPROX ’14/RANDOM ’14, 2014.
[15] Shi Li. On uniform capacitated k-median beyond the natural LP relaxation. In Proceedings
of the 26th Annual ACM-SIAM Symposium on Discrete Algorithms (SODA 2015).
[16] Shi Li. Approximating capacitated k -median with (1 + )k open facilities. In Proceedings of the
27th Annual ACM-SIAM Symposium on Discrete Algorithms (SODA 2016), pages 786–796,
2016.
[17] Shi Li and Ola Svensson. Approximating k-median via pseudo-approximation. In Proceedings
of the Forty-fifth Annual ACM Symposium on Theory of Computing, STOC ’13, pages 901–910,
New York, NY, USA, 2013. ACM.
[18] J. Lin and J. S. Vitter. -approximations with minimum packing constraint violation (extended abstract). In Proceedings of the 24th Annual ACM Symposium on Theory of Computing
(STOC), Victoria, British Columbia, Canada, pages 771–782, 1992.
[19] D. B. Shmoys, . Tardos, and K. Aardal. Approximation algorithms for facility location problems (extended abstract). In STOC ’97: Proceedings of the twenty-ninth annual ACM symposium on Theory of computing, pages 265–274, New York, NY, USA, 1997. ACM.
25
| 8cs.DS
|
arXiv:1604.01724v3 [math.OA] 17 May 2016
SOME REMARKS ABOUT THE WEAK CONTAINMENT PROPERTY FOR
GROUPOIDS AND SEMIGROUPS
CLAIRE ANANTHARAMAN-DELAROCHE
Abstract. A locally compact groupoid is said to have the weak containment property if its
full C ∗ -algebra coincide with its reduced one. Although it is now known that this property is
strictly weaker than amenability, we show that the two properties are the same under a mild
exactness assumption. Then we apply our result to get informations about the corresponding
weak containment property for some semigroups.
Introduction
The notion of amenability for locally compact groups takes many forms and is well understood
(see [39] for instance). Amenability was introduced in the measured setting for discrete group
actions and countable equivalence relations by Zimmer [53, 51, 52] at the end of the seventies.
Soon after, Renault extended this notion to general measured groupoids and to locally compact
groupoids [42]. This was followed by further studies, for example in [2, 3] for group actions. A
detailed general study is provided in the monograph [1]. In particular it has long been known
[42, 44, 1] that every amenable groupoid has the weak containment property, in the sense that its
full and reduced C ∗ -algebras coincide. Yet, at that time the converse was left open: is a locally
compact groupoid amenable when it has the weak containment property? For locally compact
groups, this is well known to be true, due to a theorem of Hulanicki [17]. More generally, this is
true for any transitive locally compact groupoid [7] (that is, a groupoid acting transitively on its
set of units). It was proved recently [29] that the weak containment property for an action of an
exact discrete group on a compact space implies the amenability of the action, leading to believe
in a positive answer to the question in full generality. However, in 2015 Willett exhibited [50] a
nice simple example of an étale groupoid having the weak containment property without being
amenable.
In this paper we show that, nevertheless, the answer to the above question is quite often positive.
It suffices that the groupoid satisfies a weak form of exactness that we call inner exactness (see
Definition 2.6). This covers many examples. Every minimal locally compact groupoid (i.e., such
that the only invariant closed subsets of the unit space are the empty set and the whole set of units)
is inner exact. In particular all locally compact groups are inner exact. For every continuous action
2010 Mathematics Subject Classification. Primary 43A07 ; Secondary 46L55, 54H20, 22A22, 20M30,20M18.
Key words and phrases. Groupoids, semigroups, weak containment, amenability.
1
2
CLAIRE ANANTHARAMAN-DELAROCHE
of an exact locally compact group on a locally compact space, the corresponding transformation
groupoid is inner exact. However it has long been known that there exist groupoids that are
not inner exact. The first example (for a non Hausdorff locally compact groupoid) was given by
Skandalis in [44]. Later, in order to provide counterexamples to the Baum-Connes conjecture for
groupoids, Higson, Lafforgue and Skandalis [15] have exhibited other (Hausdorff) locally compact
groupoids, in fact bundles of groups, that are not inner exact. The above mentioned example of
Willet is of the same kind.
The relations between weak containment and amenability are partially clarified in this paper
as follows (see Theorem 2.10).
Theorem A. Let G be an inner exact locally compact groupoid. Then G has the weak containment
property if and only if it is measurewise amenable.
Corollary. Let G be a locally compact groupoid. The two following conditions are equivalent:
(i) G is measurewise amenable;
(ii) for every G-C ∗ -algebra A, the full crossed product A ⋊ G coincide with the reduced crossed
product A ⋊r G.
Measurewise amenability is a notion which is slightly weaker than (topological) amenability
but is the same in many usual cases, for instance for étale groupoids. The difference is explained
in the main part of the text (see Definition 2.3 and Remark 2.4). So, for an étale inner exact
groupoid, the weak containment property is equivalent to amenability. This answers a question
raised by Willett in [50, §3].
The previous theorem has the following generalization that covers interesting examples.
Theorem B. Let G be a locally compact groupoid which is equivalent to an inner exact groupoid.
Then G has the weak containment property if and only if it is measurewise amenable.
After having established these results in Section 2, we turn in Section 3 to the case of discrete
semigroups. We limit ourself to semigroups not too far from the case of discrete groups, namely
inverse semigroups (defined in Section 3.1) and sub-semigroups of groups. We give partial answers
to a recurrent question concerning semigroups: what is the right definition of amenability for a
semigroup? To the semigroups that we consider are attached a full C ∗ -algebra and a reduced
C ∗ -algebra, generalizing the classical case of groups. There are three obvious candidates for the
notion of amenability, and it is natural to wonder what are the relations between them:
(1) left amenability, that is, there exists a left invariant mean on the semigroup;
(2) weak containment property, that is, the full and reduced C ∗ -algebras of the semigroup are
the same;
(3) nuclearity of the reduced C ∗ -algebra of the semigroup.
This problem has been addressed in many papers (see [38], [10], [34], [22], [9], [30], [26], [27], [12],
to cite a few of them). Of course, these three properties are equivalent for a discrete group.
Let us consider first the case of an inverse semigroup (see Section 3.1), that we denote by S.
A very useful feature of such a semigroup is that its full and reduced C ∗ -algebras are described
SOME REMARKS ABOUT THE WEAK CONTAINMENT PROPERTY
3
via the groupoid GS canonically associated to it [40]. As a consequence, we see that (3) ⇒ (2) in
this case: if the reduced C ∗ -algebra Cr∗ (S) is nuclear, then S has the weak containment property,
since GS is amenable.
It is the only general fact that can be stated. The example given by Willett, once reinterpreted
in the setting of inverse semigroups, allows us to show that (2) 6⇒ (3) in general, for inverse
semigroups (see Example 3.11). This answers a question raised in [12, Remark 3.7]. This example
is a Clifford inverse semigroup, that is an inverse semigroup which is a disjoint union S = ⊔e∈E Se
of groups where the set E of idempotents is contained in the center of S. This gives an example of
Clifford semigroup which has the weak containment property, although not all groups Se , e ∈ E,
are amenable. This answers a question raised in [38].
We observe that the notion of left amenability is not interesting, except when S has not a zero
element, i.e., an element 0 such that 0s = 0 = s0 for every s ∈ S. Indeed, any inverse semigroup
with a zero is left amenable, since the Dirac measure at zero is a left invariant mean. Even if S
has no zero, the left amenability of S does not imply the weak containment property, and a fortiori
the nuclearity of Cr∗ (S) (see Example 3.11).
However, using Theorem A we present a class of inverse semigroups for which Conditions (2)
and (3) are equivalent:
Theorem C. Let S be a strongly E ∗ -unitary inverse semigroup with an idempotent pure morphism
into an exact group G (see Definition 3.3). Then S has the weak containment property if and only
if the reduced C ∗ -algebra Cr∗ (S) is nuclear (Theorem 3.12 (2)).
Next, we consider the case of a pair (P, G) where P is a sub-semigroup of a group G containing
the unit e. As pointed out in [26, 27, 36], a handy tool in order to study the C ∗ -algebras of P is its
left inverse hull S(P ). It is an inverse semigroup with nice properties (Propositions 3.1 and 3.6).
Following Xin Li [26, 27], we define the full C ∗ -algebra of P to be the full C ∗ -algebra of GS(P ) .
This extends the definition given by Nica in [34] for quasi-lattice ordered groups (Definition 3.15).
On the other hand, Cr∗ (P ) is a quotient of the reduced C ∗ -algebra of GS(P ) .
We first observe that the left amenability of P always implies the nuclearity of Cr∗ (P ) (see
Proposition 3.17). It is not true in general that the weak containment property implies the left
amenability of P as shown by Nica in [34]. He considered the free group G = Fn on n generators
a1 , . . . , an and P = Pn is the semigroup generated by a1 , . . . , an . Using the uniqueness property
of the Cuntz algebra On , Nica proved that Pn has the weak containment property although it is
Cuntz-Toeplitz C ∗ -algebra, that is, the C ∗ -algebra
not left amenable. Note that Cr∗ (Pn ) is the P
generated by n isometries s1 , . . . sn such that 1≤i≤n si s∗i 1. The weak containment property is
equivalent to the uniqueness of the Cuntz-Toeplitz C ∗ -algebra. Moreover, Cr∗ (Pn ) is an extension
of On by the algebra of compact operators and therefore is nuclear.
Whether the weak containment property for P implies the nuclearity of Cr∗ (P ) is an old problem
that was raised by several authors, for instance by Laca and Raeburn [22, Remark 6.9], and more
recently by Xin Li [27, §9]. Using Theorem A, we give the following partial answer (Theorem
3.18).
4
CLAIRE ANANTHARAMAN-DELAROCHE
Theorem D. Let P be a subsemigroup of an exact group G, with e ∈ P . Then the weak containment property implies that Cr∗ (P ) is nuclear.
This result follows from the fact that S(P ) satisfies the assumptions of Theorem C.
In [26], Xin Li introduced the independence property for P , which can be rephrased by saying
that the quotient map from Cr∗ (GS(P ) ) onto Cr∗ (P ) is injective. In this case (which occurs for
instance for quasi-lattice ordered groups), the nuclearity of Cr∗ (P ) implies the weak containment
property for GS(P ) and thus for P .
In order to facilitate the reading of the paper, in the first section we recall the main facts about
groupoids and their C ∗ -algebras that will be used in the sequel, and we fix the notation. We
emphasize that the locally compact spaces will always be Hausdorff (unless explicitly mentioned)
and second countable. Locally compact groupoids will always come equipped with a Haar system
and Hilbert spaces will be separable.
1. Preliminaries
1.1. Groupoids. We assume that the reader is familiar with the basic definitions about groupoids.
For details we refer to [42], [40]. Let us recall some notation and terminology. A groupoid consists
of a set G, a subset G (0) called the set of units, two maps r, s : G → G (0) called respectively the
range and source maps, a composition law (γ1 , γ2 ) ∈ G (2) 7→ γ1 γ2 ∈ G, where
G (2) = {(γ1 , γ2 ) ∈ G × G : s(γ1 ) = r(γ2 )},
and an inverse map γ 7→ γ −1 . These operations satisfy obvious rules, such as the facts that
the composition law (i.e., product) is associative, that the elements of G (0) act as units (i.e.,
r(γ)γ = γ = γs(γ)), that γγ −1 = r(γ), γ −1 γ = s(γ), and so on (see [42, Definition 1.1]). For
x ∈ G (0) we set G x = r −1 (x) and Gx = s−1 (x). Usually, X will denote the set of units of G.
A locally compact groupoid is a groupoid G equipped with a locally compact topology such that
the structure maps are continuous, where G (2) has the topology induced by G × G and G (0) has the
topology induced by G. We assume that the range (and therefore the source) map is open, which
is a necessary condition for the existence of a Haar system. We denote by Cc (G) the algebra of
continuous complex valued functions with compact support on G.
Definition 1.1. Let G be a locally compact groupoid. A Haar system on G is a family λ = (λx )x∈X
of measures on G, indexed by the set X = G (0) of units, satisfying the following conditions:
• Support: λx has exactly G x as support, for every x ∈ X;
R
• Continuity: for every f ∈ Cc (G), the function x 7→ λ(f )(x) = G x f dλx is continuous;
• Invariance: for γ ∈ G and f ∈ Cc (G), we have
Z
Z
s(γ)
f (γ1 ) dλr(γ) (γ1 ).
f (γγ1 ) dλ (γ1 ) =
G s(γ)
G r(γ)
Examples 1.2. (a) Transformation groupoid. Let G be a locally compact group acting continuously to the right on a locally compact space X. The topological product space X × G has
SOME REMARKS ABOUT THE WEAK CONTAINMENT PROPERTY
5
a natural groupoid structure with X as space of units. The range and source maps are given
respectively by r(x, g) = x and s(x, g) = xg. The product is given by (x, g)(xg, h) = (x, gh) and
the inverse by (x, g)−1 = (xg, g−1 ). We denote by X ⋊ G this groupoid. A Haar system λ is given
by λx = δx × λ̃ where λ̃ is a left Haar measure on G. Similarly, one defines G ⋉ X for a left action
of G.
(b) Groupoid group bundle. It is a locally compact groupoid such that the range and source
maps are equal. By [44, Lemma 1.3], one can choose, for x ∈ G (0) , a left Haar measure λx on the
group G x = Gx in such a way that (λx )x∈X forms a Haar system on G. An explicit example will
be given in Section 1.3.
(c) Etale groupoids. A locally compact groupoid is called étale when its range (and therefore
its source) map is a local homeomorphism from G onto G (0) . Then G x and Gx are discrete and
G (0) is open in G. Moreover the family of counting measures λx on G x forms a Haar system (see
[42, Proposition 2.8]). Groupoids associated with actions, or more generally with partial actions
(that we define now), of discrete groups are étale.
(d) Partial transformation groupoid. A partial action of a discrete group G on a locally compact
space X is a family (βg )g∈G of partial homeomorphisms of X between open subsets, such that
βe = Id X and βg βh ≤ βgh for g, h ∈ G, meaning that βgh extends βg βh . Then
G ⋉ X = (g, x) : g ∈ G, x ∈ Xg−1 ⊂ G × X
with the topology induced from the product topology, where Xg−1 is the domain of βg , is an étale
groupoid. The range and source maps of G ⋊ X are given respectively by r(g, x) = βg (x) and
s(g, x) = x. The product is defined by (g, x)(h, y) = (gh, y) when x = βh (y), and the inverse is
given by (g, x)−1 = (g−1 , βg (x)).
As already said, in the sequel, the locally compact groupoids are implicitly supposed to be Hausdorff, second countable, and equipped with a Haar system λ. In the three above examples, λ will
be the mentioned Haar system.
1.2. Representations of a locally compact groupoid. Let (G, λ) be a locally compact groupoid
with a Haar system λ. We set X = G (0) . The space Cc (G) is an involutive algebra with respect to
the following operations for f, g ∈ Cc (G):
Z
(f ∗ g)(γ) = f (γ1 )g(γ1−1 γ)dλr(γ) (γ1 )
(1)
f ∗ (γ) = f (γ −1 ).
(2)
We define a norm on Cc (G) by
Z
Z
x
−1
x
kf kI = max sup |f (γ)| dλ (γ), sup
f (γ ) dλ (γ) .
x∈X
x∈X
Definition 1.3. A representation of Cc (G) is a ∗-homomorphism π from Cc (G) into the C ∗ -algebra
B(H) of bounded operators of a Hilbert space H such that kπ(f )k ≤ kf kI for every f ∈ Cc (G).
6
CLAIRE ANANTHARAMAN-DELAROCHE
Example 1.4. Let x ∈ X = G(0) . We denote by λx the image of λx be the inverse map γ 7→ γ −1 .
Let πx : Cc (G) → B L2 (Gx , λx ) be defined by
Z
f (γγ1−1 )ξ(γ1 ) dλx (γ1 )
(πx (f )ξ)(γ) =
Gx
for f ∈ Cc (G) and ξ ∈
L2 (G
x , λx ).
Then πx is a representation of Cc (G).
More generally, let µ be a (Radon) measure on X. We denote by ν = µ ◦ λ the measure on G
defined by the formula
Z
Z
Z
f (γ) dλx (γ) dµ(x).
f dν =
G
X
Gx
ν −1
Let
be the image of ν under the inverse map. For f ∈ Cc (G) and ξ ∈ L2 (G, ν −1 ) we define
the operator Indµ (f ) by the formula
Z
f (γγ1−1 )ξ(γ1 ) dλs(γ) (γ1 ).
Indµ (f )ξ (γ) =
Gs(γ)
Then Indµ is a representation of Cc (G), called the induced representation associated with µ. We
have Indδx = πx .
The full C ∗ -algebra C ∗ (G) of G is the completion of Cc (G) with respect to the norm
kf k = sup kπ(f )k
where π runs over all representations of Cc (G). The reduced C ∗ -algebra Cr∗ (G) is the completion
of Cc (G) with respect to the norm
kf k = sup kπx (f )k.
x∈X
Obviously, the identity map of Cc (G) extends to a surjective homomorphism from C ∗ (G) onto
Cr∗ (G).
Remark 1.5. Assume that G is a locally compact group. Let us observe that the involution
on Cc (G) that we introduced is not the usual one in group theory. If ∆ denotes the modular
function of G, usually the involution is defined by f ⋆ (γ) = f (γ −1 )∆(γ −1 ). The map f 7→ f˜ where
f˜(γ) = f (γ)∆(γ)−1/2 is an isomorphism of involutive algebra between the ∗-algebra Cc (G) with
the involution ∗ introduced in (2) and the usual one with the involution ⋆ . The full and reduced
C ∗ -algebras defined above are then canonically identified respectively with the classical full and
reduced C ∗ -algebras of the group G.
Similarly, the full and reduced C ∗ -algebras of a transformation groupoid X ⋊ G are identified
with the full crossed product C0 (X) ⋊ G and the reduced crossed product C0 (X) ⋊r G respectively,
where C0 (X) is the C ∗ -algebra of complex valued functions on X vanishing to 0 at infinity.
A familiar result in group theory relates in a bijective and natural way the non-degenerate
representations of the full group C ∗ -algebra and the unitary representations of the group. A
similar result holds for groupoids. Its statement requires some preparation.
SOME REMARKS ABOUT THE WEAK CONTAINMENT PROPERTY
7
Let (G, λ) be a locally compact groupoid and let µ be a (Radon) measure on X = G (0) . We set
ν = µ ◦ λ. We say that µ is quasi-invariant if ν is equivalent to ν −1 . In this case, we denote by
∆ the Radon-Nikodým derivative dν/ dν −1 . A groupoid (G, λ) equipped with a quasi-invariant
measure µ is called a measured groupoid.
Definition 1.6. A unitary representation of G is a triple (µ, H, U ) where
(i) µ is a quasi-invariant measure on X;
(ii) H = ({Hx : x ∈ X}, E) is a measurable field of Hilbert spaces over X (where E is a
fundamental sequence of measurable vector fields);
(iii) U is a measurable action of G on H by isometries, that is, for every γ ∈ G we have an
isometric isomorphism U (γ) : Hs(γ) → Hr(γ) such that
(a) for x ∈ X, U (x) is the identity map of Hx ;
(b) for (γ, γ1 ) ∈ G (2) , U (γγ1 ) = U (γ)U (γ1 );
(c) for ξ, η ∈ E, the function γ 7→ hξ ◦ r(γ), U (γ)η ◦ s(γ)ir(γ) is measurable.
We denote by H = L2 (X, H, µ) the Hilbert space of square integrable sections of H, and for
f ∈ Cc (G) we define the operator πU (f ) on H by the formula
Z Z
f (γ)∆(γ)−1/2 hξ ◦ r(γ), U (γ)η ◦ s(γ)ir(γ) dλx (γ) dµ(x).
hξ, πU (f )ηi =
X
Gx
Then f 7→ πU (f ) is a representation of Cc (G), called the integrated form of (µ, H, U ) (or simply U ).
A crucial result, due to J. Renault, asserts that every representation of Cc (G) can be disintegrated.
Here, the fact that the groupoid is assumed to be second countable is needed.
Theorem 1.7. ([43, Proposition 4.2]) Let π be a non-degenerate representation of Cc (G) on a
Hilbert space H. There is a unitary representation (µ, H, U ) of G such that π is unitary equivalent
to the integrated form πU of (µ, H, U ). We say that π disintegrates over µ.
Example 1.8. The left regular
representation of G over a quasi-invariant measure µ is (µ, H =
L2 (λ), L) where L2 (λ) = L2 (G x , λx ) : x ∈ X , E = Cc (G) , and
L(γ) : L2 (G s(γ) , λs(γ) ) → L2 (G r(γ) , λr(γ) )
is given, for ξ ∈ L2 (G s(γ) ), γ1 ∈ G r(γ) , by
L(γ)ξ (γ1 ) = ξ(γ −1 γ1 ).
Note that L2 (X, H, µ) = L2 (G, µ ◦ λ) and that, for f ∈ Cc (G), ξ ∈ L2 (G, µ ◦ λ) we have
Z
f (γ)∆(γ)−1/2 L(γ)ξ ◦ s(γ) dλx (γ).
πL (f )ξ x =
Gx
It is well known (and easy to see) that the map W : L2 (G, ν) → L2 (G, ν −1 ) defined by the
formula W ξ = ∆1/2 ξ is an isometric isomorphism which implements a unitary equivalence between
πL and Indµ .
8
CLAIRE ANANTHARAMAN-DELAROCHE
We denote by Cr∗ (G, µ) the norm closure of πL (Cc (G)) in B(L2 (G, µ ◦ λ)).
For f ∈ Cc (G) we define the seminorm sup kπ(f )k where π ranges over all representations of
Cc (G) that disintegrate over µ. We denote by C ∗ (G, µ) the C ∗ -algebra obtained by separation
and completion of Cc (G) with respect to this seminorma. Note that we have canonical surjective
homomorphisms from C ∗ (G) onto C ∗ (G, µ) and from C ∗ (G, µ) onto Cr∗ (G, µ).
These C ∗ -algebras will play a crucial role in the rest of the paper and are related to reductions
of the groupoid G. We give some details in the next section.
1.3. About reductions of a groupoid. Let G be a groupoid and Y a subset of X = G (0) . We
set G(Y ) = r −1 (Y ) ∩ s−1 (Y ). Then G(Y ) is a subgroupoid of G called the reduction of G by Y .
When Y is reduced to a single element x, then G(x) = r −1 (x) ∩ s−1 (x) is a group called the
isotropy group of G at x.
Let now G be a locally compact groupoid with Haar system and let Y be a locally compact
subset of X which is G-invariant, meaning that for γ ∈ G, we have r(γ) ∈ Y if and only if
s(γ) ∈ Y . Then G(Y ) is a locally compact groupoid whose Haar system is obtained by restriction
of the Haar system of G.
Let µ be a quasi-invariant measure on X and let F be its support. It is a closed G-invariant
subset of X. Then Indµ is a faithful representation of Cr∗ (G(F )) (see [19, Corollary 2.4] for
instance). It follows that Cr∗ (G, µ) is canonically identified with Cr∗ (G(F )).
Besides, the representations of Cc (G(F )) contains all the representations of Cc (G) that disintegrate over µ. It follows there is a canonical surjective map q ′ from C ∗ (G(F )) onto C ∗ (G, µ).
Let us consider the general situation where a closed G-invariant subset F of X is given and set
U = X \ F . It is well known that the inclusion ιU : Cc (G(U )) → Cc (G) extends to injective homomorphisms from C ∗ (G(U )) into C ∗ (G) and from Cr∗ (G(U )) into Cr∗ (G). Similarly, the restriction
map pF : Cc (G) → Cc (G(F )) extends to surjective homomorphisms from C ∗ (G) onto C ∗ (G(F ))
and from Cr∗ (G) onto Cr∗ (G(F )). Moreover the sequence
0 → C ∗ (G(U )) → C ∗ (G) → C ∗ (G(F )) → 0
is exact. For these facts, we refer to [42, page 102], [16, Section 2.4], or to [41, Proposition 2.4.2]
for a detailed proof.
On the other hand, the corresponding sequence with respect to the reduced C ∗ -algebras is not
always exact, as shown for a non-Hausdorff groupoid by Skandalis in the Appendix of [44].
Example 1.9. Another interesting class of examples was provided in [15] by Higson, Lafforgue
and Skandalis. There, the authors consider a residually finite group Γ and an decreasing sequence
b = N ∪ {∞}
Γ ⊃ N0 ⊃ Γ1 · · · ⊃ Nk ⊃ · · · of finite index normal subgroups with ∩k Nk = {e}. Let N
b
be the Alexandroff compactification of N. We set N∞ = {e} and, for k ∈ N, we denote by
b × Γ with respect to the
qk : Γ → Γ/Nk the quotient homomorphism. Let G be the quotient of N
aFor simplicity, we do not include λ in the notation, since the Haar system is always implicitely given.
SOME REMARKS ABOUT THE WEAK CONTAINMENT PROPERTY
9
equivalence relation
(k, t) ∼ (l, u) if k = l and qk (t) = qk (u).
Equipped with the quotient topology, G has a natural structure of (Hausdorff) étale locally comb the range and source maps are given by
pact groupoid group bundle: its space of units is N,
r([k, t]) = s([k, t]) = qk (t), where [k, t] = (k, qk (t)) is the equivalence class of (k, t). The fibre G(k)
of the bundle is the quotient group Γ/Nk if k ∈ N and Γ if k = ∞. We call this groupoid an
HLS-groupoid. A basic result of [15] is that the sequence
0 −→ Cr∗ (G(N)) −→ Cr∗ (G) −→ Cr∗ (G(∞)) −→ 0
is not exact whenever Γ has Kazdhan’s property (T) (it is not even exact in K-theory!).
1.4. Crossed products. For the definition of actions of groupoids on C ∗ -algebras we refer to
[20]. Let us recall a few facts.
Definition 1.10. Let X be a locally compact space. A C0 (X)-algebra is a C ∗ -algebra A equipped
with a homomorphism ρ from C0 (X) into the centre of the multiplier algebra of A, which is
non-degenerate in the sense that there exists an approximate unit (uλ ) of C0 (X) such that
limλ ρ(uλ )a = a for every a ∈ A.
Given f ∈ C0 (X) and a ∈ A, for simplicity we will write f a instead of ρ(f )a.
Let U be an open subset of X and F = X \U . We view C0 (U ) as an ideal of C0 (X) and we denote
by C0 (U )A the closed linear span of {f a : f ∈ C0 (U ), a ∈ A}. It is a closed ideal of A and in fact,
we have C0 (U )A = {f a : f ∈ C0 (U ), a ∈ A} (see [5, Corollaire 1.9]). We set AF = A/C0 (U )A and
whenever F = {x} we write Cx (X) instead of C0 (X \ {x}) and Ax instead of A{x} . We denote by
ex : A → Ax the quotient map and for a
Q∈ A we set a(x) = ex (a). Recall that kak = supx∈X ka(x)k
(so that a 7→ (a(x))x∈X from A into x∈X Ax is injective) and that x 7→ ka(x)k is upper semicontinuous (see [47, 5]). Then, (A, {ex : A → Ax }x∈X , X) is an upper semi-continuous field of
C ∗ -algebras.
Let A and B be two C0 (X)-algebras. A morphism α : A → B of C0 (X)-algebras is a morphism
of C ∗ -algebras which is C0 (X)-linear, that is, α(f a) = f α(a) for f ∈ C0 (X) and a ∈ A. For x ∈ X,
in this case α factors through a morphism αx : Ax → Bx such that αx (a(x)) = α(a)(x).
Let Y, X be locally compact spaces and f : Y → X a continuous map. To any C0 (X)-algebra
A is associated a C0 (Y )-algebra f ∗ A = C0 (Y ) ⊗ A F where F = {(y, f (y)) : y ∈ Y } ⊂ Y × X.
For y ∈ Y , we have f ∗ A)y = Af (y) (see [25, 20]).
Definition 1.11. ([25, 20]) Let (G, λ) be a locally compact groupoid with a Haar system and
X = G (0) . An action of G on a C ∗ -algebra A is given by a structure of C0 (X)-algebra on A and
an isomorphism α : s∗ A → r ∗ A of C0 (G)-algebras such that for every (γ1 , γ2 ) ∈ G (2) we have
αγ1 γ2 = αγ1 αγ2 , where αγ : As(γ) → Ar(γ) is the isomorphism deduced from α by factorization.
When A is equipped with such an action, we say that A is a G-C ∗ -algebra.
10
CLAIRE ANANTHARAMAN-DELAROCHE
Let A be a G-C ∗ -algebra. We set Cc (r ∗ (A)) = Cc (G)r ∗ (A). It is the space of the continuous
sections with compact support of the upper semi-continuous field of C ∗ -algebras defined by the
C0 (G)-algebra r ∗ A. Then, Cc (G)r ∗ (A) is a ∗-algebra with respect to the following operations:
Z
f (γ1 )αγ1 g(γ1−1 γ) dλr(γ) (γ1 )
(f ∗ g)(γ) =
G r(γ)
and
f ∗ (γ) = αγ f (γ −1 )∗
(see [32, Proposition 4.4]). We define a norm on Cc (r ∗ (A)) by
Z
Z
x
−1
x
kf kI = max sup
kf (γ)k dλ (γ), sup
f (γ ) dλ (γ) .
x∈X
Gx
x∈X
Gx
The full crossed product A ⋊α G is the enveloping C ∗ -algebra of the Banach ∗-algebra obtained
by completion of Cc (r ∗ (A)) with respect to k·kI .
Let us now define the reduced crossed product. For x ∈ X let us consider the Hilbert Ax module L2 (Gx , λx )⊗Ax , defined as the completion of the space Cc (Gx ; Ax ) of continuous compactly
supported functions from Gx into Ax , with respect to the Ax -valued inner product
Z
ξ(γ)∗ η(γ) dλx (γ).
hξ, ηi =
Gx
For f ∈ Cc (r ∗ (A)), ξ ∈ Cc (Gx ; Ax ) and γ ∈ Gx , we set
Z
α−1
f (γγ1−1 ) ξ(γ1 ) dλx (γ1 ).
πx (f )ξ (γ) =
γ
Gx
Then πx (f ) extends to a bounded operator with adjoint acting on the Hilbert Ax -module L2 (Gx , λx )⊗
Ax . In this way we get a representation of the ∗-algebra Cc (r ∗ (A)). The reduced crossed product
A ⋊α,r G is the completion of Cc (r ∗ (A)) with respect to the norm kf k = supx∈X kπx (f )k (see
[20])b.
Remarks 1.12. (a) We note that if Y is a locally compact G-invariant subset of X = G (0) , then
C0 (Y ) has a natural structure of G-C ∗ -algebra. Moreover, Cc (r ∗ (C0 (Y ))) = Cc (G(Y )) and C0 (Y )⋊G
and C0 (Y ) ⋊r G are canonically isomorphic to C ∗ (G(Y )) and Cr∗ (G(Y )) respectively.
(b) Let B be a C ∗ -algebra and set A = B ⊗ C0 (X). Since C0 (X) is a G-C ∗ -algebra, we see that
A = B ⊗ C0 (X) is a G-C ∗ -algebra, the action being trivial on B. Moreover, A ⋊ G and A ⋊r G are
canonically isomorphic to B ⊗max C ∗ (G) and B ⊗ Cr∗ (G) respectively.
bπ is what is denoted Λ in [20] except that the authors consider C (s∗ (A)) instead of C (r ∗ (A)). This explains
x
x
c
c
why our formula is not exactly the same.
SOME REMARKS ABOUT THE WEAK CONTAINMENT PROPERTY
11
2. Amenability and weak containment
The reference for this section is [1]. The notion of amenable locally compact groupoid has many
equivalent definitions. We will recall two of them. Before, let us recall a notation: given a locally
s(γ) , then γµ is the measure on G r(γ) defined by
compact
groupoid
R
R G, γ ∈ G and µ a measure on G
G r(γ) f dγµ = G s(γ) f (γγ1 ) dµ(γ1 ).
Definition 2.1. ([1, Definitions 2.2.2, 2.2.8]) We say that G is amenable if there exists a net (mi ),
where mi = (mxi )x∈G (0) is a family of probability measures mxi on G x , such that
R
(i) each mi is continuous in the sense that for all f ∈ Cc (G), the function x 7→ f dmxi is
continuous;
s(γ)
r(γ)
(ii) limi γmi − mi
= 0 uniformly on the compact subsets of G.
1
We say that (mi )i is an approximate invariant continuous mean on G. Note that if G is amenable
and if Y is a locally compact G-invariant subset of X, then the groupoid G(Y ) is amenable.
Proposition 2.2. ([1, Proposition 2.2.13]) Let (G, λ) be a locally compact groupoid with Haar
system. Then G is amenable if and only if there exists a net (gi ) of non-negative functions in
Cc (G) such that
R
(a) gi dλx ≤ 1 for every x ∈ G (0) ;
R
(b) limi gi dλx = 1 uniformly on the compact subsets of G (0) ;
R
(c) limi gi (γ −1 γ1 ) − gi (γ1 ) dλr(γ) (γ1 ) = 0 uniformly on the compact subsets of G.
We will also need the notion of measurewise amenability.
Definition 2.3. ([1, Proposition 3.2.14]) Let (G, λ) be a locally compact groupoid.
(i) Let µ be a quasi-invariant measure on X. We say that the measured groupoid (G, λ, µ) is
amenable if there exists a net (gi ) of (µ ◦ λ)-measurable non-negative functions on G such
that R
(a) gi dλx = 1 for a.e. x ∈ G (0) ;
R
(b) limi gi (γ −1 γ1 ) − gi (γ1 ) dλr(γ) (γ1 ) = 0 in the weak*-topology of L∞ (G, µ ◦ λ).
(ii) We say that (G, λ) is measurewise amenable if (G, λ, µ) is an amenable measured groupoid
for every quasi-invariant measure µ.
Remark 2.4. An amenable groupoid is measurewise amenable. The converse is true for groupoids
that have countable orbits (see [1, Theorem 3.3.7]). This is in particular the case for étale
groupoids and for locally compact groups.
Theorem 2.5. Let (G, λ) be a locally compact groupoid. Consider the following conditions:
(a) (G, λ) is measurewise amenable;
(b) for every quasi-invariant measure µ, the canonical surjection from C ∗ (G, µ) onto Cr∗ (G, µ)
is injective;
12
CLAIRE ANANTHARAMAN-DELAROCHE
(c) Cr∗ (G) is nuclear;
(d) for every G-C ∗ -algebra A, the canonical surjection from A ⋊ G onto A ⋊r G is injective;
(e) the canonical surjection from C ∗ (G) onto Cr∗ (G) is injective.
Then, we have (a) ⇔ (b) ⇒ (c) and (a) ⇒ (d) ⇒ (e). Moreover, if the isotropy groups G(x) are
discrete for every x ∈ X, then (c) ⇒ (a).
The equivalence between (a) and (b) is proved in [1, Theorem 6.1.4]. That (a) ⇒ (c) is contained
in [1, Corollary 6.2.14] as well as the fact that (c) ⇒ (a) when the isotropy is discrete. For the
proof of (a) ⇒ (d) see [44, Theorem 3.6] or [1, Proposition 6.1.10].
Definition 2.6. We say that an action of a locally compact groupoid (G, λ) on a C ∗ -algebra A
is inner exact if for every G-invariant closed ideal I of A the sequence
0 −→ I ⋊r G −→ A ⋊r G −→ (A/I) ⋊r G −→ 0
is exact.
We say that G is inner exact if the canonical action of G on C0 (G (0) ) is inner exact, i.e., if for
every invariant closed subset F of G (0) , the sequence
0 −→ Cr∗ (G(U )) −→ Cr∗ (G) −→ Cr∗ (G(F )) −→ 0
is exact.
The term “inner” in the above definitions aims to highlight that we only consider short sequences with respect to the specific given action of the groupoidc. A possible definition of exactness for a groupoid is the following one. Other candidates are considered in [4].
Definition 2.7. We say that a groupoid (G, λ) is exact in the sense of Kirchberg-Wassermann
(or KW-exact in short) if every action of (G, λ) on any C ∗ -algebra A is inner exact.
This notion was studied by Kirchberg and Wassermann for locally compact groups. They
proved in particular that this property is equivalent to the exactness of Cr∗ (G) for a discrete
group G.
Examples 2.8. (a) Every minimal groupoid is inner exact. In particular, every locally compact
group is inner exact.
(b) Every KW-exact groupoid is inner exact.
(c) Let G be a locally compact KW-exact group acting to the right on a locally compact space
X. Then the transformation groupoid G = X ⋊ G is KW-exact. Indeed let α be an action of G
on a C0 (X)-algebra A. Then G acts on A by (βg a)(x) = α(x,g) (a(xg)) and it is straightforward
to check that A ⋊β,r G is canonically isomorphic to A ⋊α,r G. Moreover, this identification is
functorial.
cThis notion is used in another context in [8].
SOME REMARKS ABOUT THE WEAK CONTAINMENT PROPERTY
13
In fact the groupoid is exact in a very strong sense. Indeed, G acts amenably on a compact
space Y d and therefore it acts amenably on Y × X by (y, x)g = (yg, xg) (see [1, Proposition
2.2.9]). Then G = X ⋊ G acts amenably on Y × X by (y, x)(x, g) = (yg, xg) and the momentum
map (y, x) 7→ x ∈ G (0) is proper. These facts imply that G is KW-exact. The proof is the same as
the proof showing that a group acting amenably on a compact space is KW-exact (see for instance
[3, Theorem 7.2]). More details on the notion of exactness for groupoids are given in [4].
(d) Let (βg )g∈G be a partial action of a discrete group G on a locally compact space X. The
associated groupoid G = G ⋊ X is KW-exact whenever G is exact. For the purpose of this paper,
we only give the easier proof of its inner exactness. Let F be a closed G-invariant subset of X
and set U = X \ F . The partial action of G on X induces a partial action (βgF )g∈G on F : the
domain of βgF is F ∩ Xg−1 where Xg−1 is the domain of βg and βgF acts by restriction. Similarly,
one defines a partial action (βgU )g∈G on U . The partial transformation groupoids G ⋉ F and
G ⋉ U are respectvely isomorphic to G(F ) and G(U ). Now we observe that G acts partially on the
C ∗ -algebras C0 (X), C0 (F ) and C0 (U ). Moreover by [28, Proposition 2.2], the groupoid C ∗ -algebra
Cr∗ (G ⋉ X) is canonically isomorphic to the reduced crossed product C0 (X) ⋊r G associated with
the partial action of G on C0 (X). We make the same observation with the two other partial
actions. Finally, we conclude by using the fact, proved in [13, Theorem 22.9], that the following
sequence
0 −→ C0 (U ) ⋊r G −→ C0 (X) ⋊r G −→ C0 (F ) ⋊r G −→ 0
is exact.
(e) The case of a groupoid group bundle is considered in the next proposition.
Proposition 2.9. Let G be a groupoid group bundle over a locally compact space X. Let us he
following conditions:
(i) G is inner exact;
(ii) for every x ∈ X the sequence
0 −→ Cr∗ (G(X \ {x})) −→ Cr∗ (G) −→ Cr∗ (G(x)) −→ 0
is exact;
(iii) Cr∗ (G) is a continuous field of C ∗ -algebras over X with fibres Cr∗ (G(x)).
Then we have (i) ⇒ (ii) ⇔ (iii).
Proof. (i) ⇒ (ii) is obvious.
Let us prove the equivalence between (ii) and (iii). For x ∈ X let πx be the canonical surjective
map from Cr∗ (G) onto Cr∗ (G(x)). Then (Cr∗ (G), {πx : Cr∗ (G) → Cr∗ (G(x))}x∈X , X) is a field of
C ∗ -algebras on X, which is lower semi-continuous in the sense that x 7→ kπx (a)k is lower semicontinuous for every a ∈ Cr∗ (G) (see for instance [23, Theorem 5.5]). On the other hand, Cr∗ (G)
is a C0 (X)-algebra. Indeed, for f ∈ C0 (X), g ∈ Cc (G) and γ ∈ G, we set (f g)(γ) = f ◦ r(γ)g(γ).
The map g 7→ f g extends continuously in order to define a structure of C0 (X)-algebra on Cr∗ (G).
dIn the non discrete case this fact is proved in a recent preprint of Brodzki, Cave and Li [6].
14
CLAIRE ANANTHARAMAN-DELAROCHE
We have Cc (G(X \ {x})) = Cx (X)Cc (G) and by continuity we get Cr∗ (G(X \ {x})) = Cx (X)Cr∗ (G).
Note that for f ∈ C0 (X), a ∈ Cr∗ (G) and x ∈ X, we have πx (f a) = f (x)πx (a). It follows from
[21, Lemma 2.3] that the function x 7→ kπx (a)k is upper semi-continuous at x0 for all a ∈ Cr∗ (G)
if and only if the kernel of πx0 is Cr∗ (G(X \ {x0 })).
Theorem 2.10. Let (G, λ) be a locally compact groupoid. The two following conditions are equivalent:
(1) C ∗ (G) = Cr∗ (G) and G is inner exact;
(2) the groupoid G is measurewise amenable.
Proof. Theorem 2.5 gives us (2) ⇒ (1). For the converse we have to show that, for every quasiinvariant measure µ, the canonical surjective map qµ : C ∗ (G, µ) → Cr∗ (G, µ) is injective, still using
Theorem 2.5. Let F be the support of µ and set U = X \F . The following diagram is commutative
0
/ C ∗ (G(U ))
/ C ∗ (G)
/ C ∗ (G(F ))
qF
q
0
/ C ∗ (G(U ))
r
/ C ∗ (G)
r
/0
/ C ∗ (G(F ))
r
/0
where the two lines are exact and q is injective. Then an elementary diagram chasing shows
that qF is injective. Recall that Cr∗ (G(F )) is canonically identified to Cr∗ (G, µ) (see Section 1.3).
Observe that qF factorizes through C ∗ (G, µ) as qF = qµ ◦ q ′ where q ′ : C ∗ (G(F )) → C ∗ (G, µ) is
surjective. It follows that qµ is injective.
The following theorem is a converse to [1, Proposition 6.1.10]. It is already known when G has
discrete isotropy since Condition (b) implies that C ∗ (G) = Cr∗ (G) and Cr∗ (G) nuclear.
Corollary 2.11. Let (G, λ) be a locally compact groupoid. The following conditions are equivalent:
(a) G is measurewise amenable;
(b) A ⋊ G = A ⋊r G for every G-C ∗ -algebra A.
Proof. We recalled in Theorem 2.5 that (a) ⇒ (b). Conversely, (b) immediately implies that G
has the weak containment property and is KW-exact and so (a) holds.
The following corollary is immediate since amenability and measurewise amenability coincide
for an étale groupoid.
Corollary 2.12. Let (G, λ) be an étale groupoid which is inner exact (for instance KW-exact or
minimal). Then G has the weak containment property if and only if it is amenable.
For the notion of (topological) equivalence between locally compact groupoids we refer to [1,
Definition 2.2.15].
Corollary 2.13. Let (G, λ) be a locally compact groupoid which is equivalent to an inner exact
groupoid. Then G has the weak containment property if and only if it is measurewise amenable.
SOME REMARKS ABOUT THE WEAK CONTAINMENT PROPERTY
15
Proof. Assume that G has the weak containment property and is equivalent to an inner exact
groupoid G ′ . Then by [49, Theorem 17], the groupoid G ′ has the weak containment property and
therefore is measurewise amenable. To conclude, we use the fact that measurewise amenability is
preserved under equivalence [1, Theorem 3.2.16].
In [50], Willett considered an HLS-groupoid constructed from a well-chosen sequence of normal
subgroups with finite index in the free group F2 with two generators. This étale groupoid has the
weak containment property although it is not amenable.
3. Applications to semigroup C ∗ -algebras
3.1. Semigroups. We will consider two kinds of semigroups: inverse semigroups and sub-semigroups of a group.
An inverse semigroup S is a semigroup such that for every s ∈ S there exists a unique element
s∗ such that ss∗ s = s and s∗ ss∗ = s∗ . Our references for this notion are [40, 24]. Note that
groups are inverse semigroups with exactly one idempotent. The set ES of idempotents of S
plays a crucial role. It is an abelian sub-semigroup of S. On S one defines the equivalence
relation s ∼ t if there exists an idempotent e such that se = te. The quotient S/σ is a group,
σ
called the maximal group homomorphism image of S, since every homomorphism from S into a
group G factors through S/σ. This group S/σ is trivial when S has a zero element 0, which is a
frequent situation.
By an abuse of notation, σ will also denote the quotient map from S onto S/σ. If S has a zero,
we denote by S × the set S \ {0}. When S does not have a zero, we set S × = S.
Given a set X, we denote by IS(X) the inverse semigroup of partial bijections of X. Its zero
element 0 is the application with empty domain. The Wagner-Preston theorem [40, Proposition
2.1.3] identifies any inverse semigroup S with a sub-semigroup of IS(S).
Let G be a group and P a sub-semigroup of G containing the unit e. The left inverse hull S(P )
of P is the inverse sub-semigroup of IS(P ) generated by the injection ℓp : x 7→ px. It has a unit,
namely ℓe . Observe that ℓ∗p is the map px 7→ x defined on pP . Every element of S(P ) is of the
form s = ℓ∗p1 ℓq1 · · · ℓ∗pn ℓqn with pi , qi ∈ P and n ≥ 1. Let us recall some important properties of
S(P ).
Proposition 3.1. Let (P, G) as above. Then
(1) 0 6∈ S(P ) if and only if P P −1 is a subgroup de G;
−1
(2) The application ψ : S(P )× → G sending s = ℓ∗p1 ℓq1 · · · ℓ∗pn ℓqn to p−1
1 q1 · · · pn qn is well
×
.
defined. It satisfies ψ(st) = ψ(s)ψ(t) if st 6= 0 and we have ψ −1 (e) = ES(P
)
Proof. P P −1 is a sub-group de G if and only if pP ∩ qP 6= ∅ for every p, q ∈ P (i.e., P is left
reversible). Then, assertion (1) is Lemma 3.4.1 of [36].
(2) is proved in [36, Proposition 3.2.11].
16
CLAIRE ANANTHARAMAN-DELAROCHE
Recall that on an inverse semigroup S, a partial order is defined as follows: s ≤ t if there exists
an idempotent e such that s = te (see [24, page 21] for instance).
Definition 3.2. An inverse semigroup S is said to be E-unitary if ES is the kernel of σ : S → S/σ
(equivalently, every element greater than an idempotent is an idempotent). When S has a zero,
this means that S = ES .
Definition 3.3. Let S be an inverse semigroup. A morphism (or grading) is an application ψ
from S × into a group G such that ψ(st) = ψ(s)ψ(t) if st 6= 0. If in addition ψ −1 (e) = ES× , we
say that ψ is an idempotent pure morphism. When such an application ψ from S × into a group
G exists, the inverse semigroup S is called strongly E ∗ -unitary.
Note that when S is without zero, S is strongly E ∗ -unitary if and only if it is E-unitary.
Remark 3.4. Let (P, G) with G = P P −1 . Then the map τ : S(P )/σ → G such that τ ◦ σ = ψ
is an isomorphism. Indeed ψ is surjective, so τ is also surjective. Assume that τ (σ(x)) = e, with
x ∈ S(P ). Since ψ is idempotent pure, we see that x is an idempotent and therefore σ(x) is the
unit of S(P )/σ.
In [35], Nica has introduced the Toeplitz inverse semigroup S(G, P ) which is the inverse subsemigroup of IS(P ) generated by the maps αg : g−1 P ∩ P → P ∩ gP , g ∈ G, where αg (x) = gx if
x ∈ g −1 P ∩ P . For p ∈ P we have αp = ℓp . Therefore we have S(P ) ⊂ S(G, P ).
Definition 3.5. We say that (P, G) satisfies the Toeplitz condition if S(P ) = S(G, P ).
We will give in Section 3.4 another characterization of the Toeplitz condition, along with
examples.
Proposition 3.6. Assume that (P, G) satisfies the Toeplitz condition. Let ψ : S(P )× → G as
defined in Proposition 3.1. Then, αg 6= 0 if and only if g is in the image of ψ. In this case we
have ψ(αg ) = g, and αg is the greatest element of ψ −1 (g).
For the proof, see [37, Proposition 4.1] or [35, Lemma 3.2].
3.2. Groupoid associated with an inverse semigroup. Let S be an inverse semigroup. We
recall the construction of the associated groupoid GS that is described in detail in [40]. We denote
by X the space of non-zero maps χ from ES into {0, 1} such that χ(ef ) = χ(e)χ(f ) and χ(0) = 0
whenever S has a zero. Equipped with the topology induced from the product space {0, 1}E , the
space X, called the spectrum of S, is locally compact and totally disconnected. Note that when
S is a monoid (i.e., has a unit element 1) then χ is non-zero if and only if χ(1) = 1, and therefore
X is compact.
The semigroup S acts on X as follows. The domain (open and compact) of t ∈ S is Dt∗ t =
{χ ∈ X : χ(t∗ t) = 1} and we set θt (χ)(e) = χ(t∗ et). We define on Ξ = {(t, χ) ∈ S × X : χ ∈ Dt∗ t }
the equivalence relation (t, χ) ∼ (t1 , χ1 ) if χ = χ1 and there exists e ∈ ES with χ(e) = 1 and
te = t1 e. Then GS is the quotient of Ξ with respect to this equivalence relation, equipped with
SOME REMARKS ABOUT THE WEAK CONTAINMENT PROPERTY
17
the quotient topology. The range of the class [t, χ] of (t, χ) is θt (χ) and its source is χ. The
composition law is given by [u, χ][v, χ′ ] = [uv, χ′ ] if θv (χ′ ) = χ (see [40] or [11] for details). In
general, GS is not Hausdorff. But for the inverse semigroups we are interested in, like S(P ), we
will see that the quotient topology is Hausdorff.
Proposition 3.7. Let S be a strongly E ∗ -unitary inverse semigroup, and let ψ : S × → G be an
idempotent pure morphism. Then there is a partial action of G on the spectrum X of S such
that the groupoid GS is topologically isomorphic to the groupoid G ⋉ X associated with the partial
action. In particular, GS is Hausdorff and étale. Moreover, X is compact when S has a unit.
Proof. This result is described in [31]. The partial action of G on the spectrum X of S is defined
by setting Xg−1 = ∪t∈ψ−1 (g) Dt∗ t (which can be empty). For χ ∈ Xg−1 we set βg (χ) = θt (χ) where
t ∈ ψ −1 (g) is such that χ ∈ Dt∗ t . This does not depend on the choice of t as shown in [31, Lemma
3.1]. Moreover, by [31, Theorem 3.2], the groupoid GS is canonically isomorphic to G ⋉ X.e
Proposition 3.8. Let S be a strongly E ∗ -unitary inverse semigroup. We assume that there is an
idempotent pure morphism ψ : S × → G such that if g 6= e is in the image of ψ, then ψ −1 (g) has
a greatest element αg . Then the groupoid GS is equivalent f to a transformation groupoid G ⋉ Y
for an action of G on an Hausdorff locally compact space Y .
Proof. Let t ∈ S × such that ψ(t) = g. Since t ≤ αg we have Dt∗ t ⊂ Dα∗g αg and therefore
Xg−1 = Dα∗g αg is a closed subset of X. Moreover, for χ ∈ Xg−1 we have βg (χ) = θαg (χ). It
follows that the cocycle c : G ⋉ X → G sending (g, x) to g is injective and closed. Injectivity
means that the map γ ∈ G ⋉ X 7→ (c(γ), s(γ)) is injective. The cocycle is said to be closed if
γ 7→ (r(γ), c(γ), s(γ)) from G ⋉ X into X × G × X is closed. Since the cocycle c is injective
and closed, there exists a locally compact space Y endowed with a continuous action of G such
that the transformation groupoid G ⋉ Y is equivalent to G ⋉ X = GS . When S has a unit, X is
compact and the equivalence is given by a groupoid isomorphism j from G ⋉ X onto a reduction
of G ⋉ Y ) (see [19, Theorem 1.8] and [45, Theorem 6.2]).
Corollary 3.9. Let P a sub-semigroup of a group G containing the unit.
(i) The groupoid GS(P ) is defined by a partial action of G on a compact space.
(ii) If (P, G) satisfies the Toeplitz condition, then GS(P ) is equivalent to a transformation
groupoid defined by an action of G on a locally compact space.
Proof. (i) By Proposition 3.1 there is an idempotent pure morphism ψ : S(P )× → G and we use
Proposition 3.7.
(ii) follows from Propositions 3.6 and 3.8.
eIn [31], the proofs are carried out assuming that S is E-unitary (i.e., without zero) but they immediately extend
to our setting.
fFor this notion of equivalence of groupoids we refer to [1, Definition 2.2.15].
18
CLAIRE ANANTHARAMAN-DELAROCHE
3.3. Weak containment for inverse semigroups. Let S be an inverse semigroup. Let us
recall the definition of the full and reduced C ∗ -algebras of S (for more details, see [40, §2.1]).
Given f, g ∈ ℓ1 (S), we set
X
(f ⋆ g)(t) =
f (u)g(v), f ∗ (t) = f (t∗ ).
uv=t
Then ℓ1 (S) is Banach ∗-algebra, and the full C ∗ -algebra C ∗ (S) of S is defined as the enveloping C ∗ algebra of ℓ1 (S). It is the universal C ∗ -algebra for the representations of S by partial isometries.
The left regular representation π2 : S → B(ℓ2 (S)) is defined by
π2 (t)δu = δtu if (t∗ t)u = u,
π2 (t)δu = 0 otherwise.
The extension of π2 to ℓ1 (S) is faithful. The reduced C ∗ -algebra Cr∗ (S) of S is the sub-C ∗ -algebra
of B(ℓ2 (S)) generated by π2 (S). We still denote by π2 : C ∗ (S) → Cr∗ (S) the extension of the left
regular representation to C ∗ (S).
When S has a zero, we have π2 (0)δ0 = δ0 and π2 (0)δt = 0 if t 6= 0. It follows that Cδ0 is an
ideal in C ∗ (S) that it is preferable to get rid of. So we set C0∗ (S) = C ∗ (S)/Cδ0 and similarly
∗
∗ (S) = C ∗ (S)/π (Cδ ). We denote by π
Cr,0
2
0
2,0 the canonical surjective homomorphism from C0 (S)
r
∗ (S) (see [36]).
onto Cr,0
∗ (S) are canonically isomorphic to
As shown in [40] and [19], the C ∗ -algebras C0∗ (S) and Cr,0
g
∗
∗
∗
S ) and Cr (GS ) respectively. This is the reason for having introduced C0 (S) and Cr,0 (S).
∗ (S) is nuclear if and only C ∗ (S) is so, and that π is injective if and only if it is the
Note that Cr,0
2
r
case for π2,0 .
C ∗ (G
Definition 3.10. We say that S has the weak containment property if π2 (or equivalently π2,0 )
is an isomorphism.
Observe that S has the weak containment property if and only if the groupoid GS has this
property.
Recall that a semigroup S is left amenable if there exists a left invariant mean on ℓ∞ (S). An
inverse semigroup with zero is of course left amenable since the Dirac measure at zero is a left
invariant mean.
If Cr∗ (S) is nuclear, the groupoid GS is amenable and therefore S has the weak containment
property. What about the converse?
The following example shows that left amenability, even in the absence of zero, does not imply
the weak containment property. It also shows that the weak containment property is strictly
weaker than the nuclearity of Cr∗ (S).
g More precisely in [40, 19], the authors consider the C ∗ -algebras C ∗ (S) and C ∗ (S), but their definition of G
S
r
(0)
is also slightly different because for the space X = GS they do not require that the maps χ from ES into {0, 1}
satisfy χ(0) = 0. Their proof also works in our setting.
SOME REMARKS ABOUT THE WEAK CONTAINMENT PROPERTY
19
Example 3.11. Let Γ be a residually finite group
and (Nk )k≥0 oa decreasing sequence as in
n
b t ∈ Γ . Formally, S = G, the HLS
Example 1.9, whose notation we keep. Let S = qk (t) : k ∈ N,
groupoid defined in Example 1.9 but we view S as an inverse semigroup in the following way.
The product is given by qm (t).qn (u) = qm∧n (tu) where m ∧ n isnthe smallest ofothe two elements
b that we identify
m, n. We set qm (t)∗ = qm (t−1 ). The set ES of idempotents is qm (e) : m ∈ N
b The product of two idempotents is given by m.n = m ∧ n. The spectrum X is the set
with
N.
n
o
b where χm (n) = 1 if and only if m ≤ n. It is homeomorphic to the compact space
χm : m ∈ N
b The groupoid GS associated with S is the space of equivalence classes of pairs qm (t), χk
N.
with k ≤ m, where qm (t), χk ∼ qn (u), χ′k if and only if k = k′ and qk (t) = qk (u). The map
sending the class of qm (t), χk to qk (t) is an isomorphism of topological groupoids from GS onto
the HLS-groupoid G.
The maximal group homomorphism image of S is the finite group Γ/N0 and σ : S → S/σ =
Γ/N0 is qm (t) 7→ q0 (t). It is not idempotent pure. Note that S has a zero if and only if N0 = Γ,
the zero being then q0 (e).
Let us observe that S = ⊔k∈Nb Γ/Nk is a Clifford semigroup.
Since S/σ is amenable, this semigroup S is left amenable by a result of Duncan and Namioka
(see [40, Proposition A.0.5]). If Γ = F2 and the sequence (Nk )k is the one defined by Willett in [50],
then S has the weak containment property but Cr∗ (S) is not nuclear. Moreover, S is a Clifford
semigroup for which not every subgroup is amenable, although it has the weak containment
property.
On the other hand, if we realize F2 as a finite index subgroup of SL(2, Z), and choose Nk to be
the intersection with F2 of the kernel of the reduction map SL(2, Z) → SL(2, Z/2k Z), then the
corresponding HLS-groupoid has not the weak containment property, as observed in [50, Remarks
2.9]. Hence, the left amenability of S does not imply its weak containment property in general.
The next theorem gives in particular a sufficient condition for the equivalence between the weak
containment property and the nuclearity of the reduced C ∗ -algebra.
Theorem 3.12. Let S be a strongly E ∗ -unitary inverse semigroup, and let ψ : S × → G be an
idempotent pure morphism.
(1) Assume that G is amenable. Then we have C ∗ (S) = Cr∗ (S) and Cr∗ (S) is nuclear.
(2) Assume that G is exact. Then, the weak containment property of S is equivalent to the
nuclearity of Cr∗ (S).
Proof. By Proposition 3.7, the groupoid GS is associated to a partial action of G on the spectrum
of S. If G is amenable, then GS is amenable (see [46]) and therefore the statement of (1) holds.
If G is exact, the groupoid GS is inner exact (example 2.8 (d)). Then, by Corollary 2.12, GS
has the weak containment property if and only if it is amenable, and therefore if and only if
∗ (S) = C ∗ (G ) is nuclear, since G is étale.
Cr,0
S
S
r
20
CLAIRE ANANTHARAMAN-DELAROCHE
Example 3.13. Let E = (E 0 , E 1 , r, s) be a directed graph and SE the associated graph inverse
semigroup (see [30, Section 4]). Let F be the free group generated by the set E 1 of edges. Then
there is an idempotent pure morphism ψ : SE× → F satisfying the assumption of Proposition 3.7
(2). It has the weak containment property (see [30, Theorem 4.3]) although F is not amenable if
the cardinal of E 1 is ≥ 2.
3.4. Weak containment for semigroups embedded in groups. In this section we consider
a discrete group G and a sub-semigroup P which contains the unit e. We denote by λ the left
regular representation of G, and for p ∈ P we denote by Vp : ℓ2 (P ) → ℓ2 (P ) the isometry given
by Vp δq = δpq . The reduced C ∗ -algebra Cr∗ (P ) of P is the C ∗ -algebra generated by the isometries
Vp , p ∈ P .
The right definition of the full C ∗ -algebra of P is more speculative. The universal C ∗ -algebra
generated by elements vp , p ∈ P , such that vp∗ vp = 1 and vp vq = vpq for every p, q ∈ P , is too big.
For instance Murphy proved that, for the commutative semigroup N2 , this universal C ∗ -algebra
is not nuclear [33]. A reasonable definition for the full C ∗ -algebra of P was introduced by Xin
Li in [26, Definition 2.2] and a variant in [26, Definition 3.2]. It is this variant (denoted Cs∗ (P )
in [26]) that we adopt as the definition of the full C ∗ -algebra of P in the sequel, and we denote
it C ∗ (P ). By [36, Proposition 3.3.1], C ∗ (P ) can be defined as C0∗ (S(P )). Let us recall (see [36,
Lemma 3.2.2]) that the inverse semigroup S(P ) is canonically isomorphic to the inverse semigroup
of partial isometries
V (P ) = Vp∗1 Vq1 · · · Vp∗n Vqn : n ∈ N, pi , qi ∈ P .
∗ (S(P )) → C ∗ (P ) such that
Let us also recall that there is a surjective homomorphism h : Cr,0
r
h(π2 (ℓp )) = Vp for p ∈ P (see [36, Lemma 3.2.12]). Therefore we have the following situation
π2,0
h
∗
C ∗ (P ) = C0∗ (S(P )) ≡ C ∗ (GS(P ) ) −։ Cr,0
(S(P )) ≡ Cr∗ (GS(P ) ) −։ Cr∗ (P ).
Definition 3.14. We say that P has the weak containment property if C ∗ (P ) = Cr∗ (P ).
Note that the weak containment property of P implies the weak containment property of S(P ).
Definition 3.15. Let (P, G) as above and assume in addition that P ∩ P −1 = {e}. Then we
define on G a partial order by setting x ≤ y if x−1 y ∈ P . We say that (P, G) is a quasi-lattice
ordered group if for every g ∈ G, we have either P ∩ gP = ∅ or P ∩ gP = rP for some r ∈ P
(equivalently, every pair of elements in G having a common upper bound has a least common
upper bound (see [9, Lemma 7] for more on this)).
The Toeplitz condition for (P, G) is equivalent to the following property: for every g ∈ G such
that EP λg EP 6= 0, there exist p1 , . . . , pn , q1 , . . . , qn ∈ P such that EP λg EP = Vp∗1 Vq1 . . . Vp∗n Vqn
(see for instance the proof of [37, Proposition 4.1]).
Quasi-lattice ordered semigroups and semigroups (P, G) such that G = P −1 P satisfy the
Toeplitz condition (see [27, §8]). Xin Li has also introduced an important condition for P , he
called independence ([26, Definition 2.26]). We will not describe it here. We only note that when
SOME REMARKS ABOUT THE WEAK CONTAINMENT PROPERTY
21
P is contained in a group, this condition is equivalent to the injectivity of h (see [36, Theorem
3.2.14]) and that it is satisfied for quasi-lattice ordered groups (see [26, Lemma 28]).
Proposition 3.16. Let P be a sub-semigroup of a group G with e ∈ P . Assume that P satisfies
the independence condition. Then the nuclearity of Cr∗ (P ) implies the weak containment property
for P , i.e., C ∗ (P ) = Cr∗ (P ).
Proof. Assume that Cr∗ (P ) is nuclear. Since Cr∗ (P ) = Cr∗ (GS(P ) ), we see that the groupoid GS(P )
is amenable. It follows that S(P ) (and thus P ) has the weak containment property.
Proposition 3.17. Let P be a sub-semigroup of a group G with e ∈ P .
(1) If G is amenable, then Cr∗ (P ) is nuclear.
(2) If P is left amenable, then P P −1 is an amenable subgroup of G and Cr∗ (P ) is nuclear.
Proof. Assume that G is amenable. By Proposition 3.1 and Proposition 3.7 the groupoid GS(P ) is
amenable since it is isomorphic to the groupoid defined be a partial action of G. It follows that
Cr∗ (GS(P ) ) is nuclear as well as its quotient Cr∗ (P ).
Suppose now that P is left amenable. Then it is left reversible (i.e., pP ∩qP 6= ∅ for all p, q ∈ P )
and therefore G′ = P P −1 is an amenable subgroup of G (see [39, Propositions 1.23, 1.27]). To
see that Cr∗ (GS(P ) ) is nuclear, we apply the first part of the proof.
Theorem 3.18. Let P be a sub-semigroup of an exact group G, containing the unit e. The two
following conditions are equivalent:
(1) C ∗ (P ) = Cr∗ (P );
(2) C ∗ -algebra Cr∗ (P ) is nuclear and P satisfies the independence condition.
Proof. The weak containment property for P implies the weak containment property for GS(P )
and also the independence. So the assertion (1) ⇒ (2) follows from Corollaries 3.9 (i) and 2.12.
The converse is Proposition 3.16.
4. Some questions
(1) The only example of an étale groupoid which satisfies the weak containment property
although it is not amenable was provided by Willett. It is a bundle of groups. At the opposite,
are there principal groupoids (i.e., such that the units are the only elements with the same source
and range) which satisfy the weak containment property without being amenable ? It would also
be interesting to know whether there are examples which are transformation groupoids G ⋉ X
where G is a discrete group acting on a locally compact space X. To this respect, is it true that
G is exact when the groupoid G ⋉ X has the weak containment property with X compact? This
question which seems quite difficult was already raised in [50].
For the boundary compact set X = ∂G = βG \ G, equipped with the natural action of G the
answer is positive. Indeed, the weak containment property for G ⋉ ∂G implies that the sequence
0 −→ Cr∗ (G ⋉ G) −→ Cr∗ (G ⋉ βG) −→ Cr∗ (G ⋉ ∂G) −→ 0
22
CLAIRE ANANTHARAMAN-DELAROCHE
is exact. Roe and Willett proved in [48] that this exactness property implies that G has Yu’s
property A and thus is exact.
(2) Find an example of a pair (P, G) such that Cr∗ (P ) is nuclear but P has not the weak
containment property.
(3) Find a characterization of the weak containment property for a locally compact groupoid.
Added remark. After having read this paper, Kang Li has found a nice proof of the above RoeWillett result. I thank him for allowing me to reproduce his proof here.
Let G be a discrete group and let ∂F G be its Furstenberg boundary [14]. We fix an element
x0 ∈ ∂F G and we denote by ι the canonical G-equivariant inclusion from C(∂F G) into C(βG) =
ℓ∞ (G) such that ι(f )(s) = f (sx0 ) for s ∈ G. Let us recall that C(∂F G) is G-injective (see [18]). It
follows that the identity map IdC(∂F G) extends to a G-equivariant unital contraction from C(βG)
onto C(∂F G). Therefore, there exists a G-equivariant conditional expectation E from C(βG) onto
C(∂F G).
Let us denote by Id G ⋉ ι (resp. Id G ⋉r ι) the canonical homomorphism from C ∗ (G ⋉ ∂F G)
into C ∗ (G ⋉ βG) (resp. from Cr∗ (G ⋉ ∂F G) into Cr∗ (G ⋉ βG)). The map Id G ⋉r ι is known to be
injective. Here, Id G ⋉ι is also injective. Indeed, the G-equivariant conditional expectation E yields
a completely positive map Id G ⋉ E from C ∗ (G ⋉ βG) onto C ∗ (G ⋉ ∂F G) and (Id G ⋉ E) ◦ (Id G ⋉ ι)
is the identity map of C ∗ (G ⋉ ∂F G). Therefore Id G ⋉ ι is injective.
Assume now that the groupoid G ⋉ ∂G has the weak containment property. In the following
commutative diagram
0
/ C ∗ (G ⋉ G)
/ C ∗ (G ⋉ βG)
/ C ∗ (G ⋉ ∂G)
/0
≃
0
/ C ∗ (G ⋉ G)
r
/ C ∗ (G ⋉ βG)
r
/ C ∗ (G ⋉ ∂G)
r
/0
the first line is an exact sequence and the canonical map from Cr∗ (G ⋉ G) into Cr∗ (G ⋉ βG) is
injective. This implies that the groupoid G ⋉ βG has the weak containment property.
From the commutativity of the diagram
C ∗ (G ⋉ ∂F G)
Id G ⋉ι
/ C ∗ (G ⋉ βG)
Id G ⋉r ι
≃
Cr∗ (G ⋉ ∂F G)
/ C ∗ (G ⋉ βG)
r
we deduce that G ⋉ ∂F G has also the weak containment property. Since the action of G on the
Furstenberg boundary is minimal, it follows from Corollary 2.12 that the groupoid G ⋉ ∂F G is
amenable, and therefore G is exact.
SOME REMARKS ABOUT THE WEAK CONTAINMENT PROPERTY
23
References
[1] Claire Anantharaman-Delaroche and Jean Renault. Amenable groupoids, volume 36 of Monographies de
L’Enseignement Mathématique. L’Enseignement Mathématique, Geneva, 2000. With a foreword by Georges
Skandalis and Appendix B by E. Germain.
[2] Claire Anantharaman-Delaroche. Systèmes dynamiques non commutatifs et moyennabilité. Math. Ann.,
279(2):297–315, 1987.
[3] Claire Anantharaman-Delaroche. Amenability and exactness for dynamical systems and their C ∗ -algebras.
Trans. Amer. Math. Soc., 354(10):4153–4178 (electronic), 2002.
[4] Claire Anantharaman-Delaroche. Exact groupoids, In preparation.
[5] Étienne Blanchard. Déformations de C ∗ -algèbres de Hopf. Bull. Soc. Math. France, 124(1):141–215, 1996.
[6] Jacek Brodzki, Chris Cave, and Kang Li. Exactness of locally compact groups. arXiv:math.GR/1603.0182v1,
2016.
[7] Madalina Buneci. C ∗ -algebras associated to the transitive groupoids. An. Univ. Craiova Ser. Mat. Inform.,
28:79–92, 2001.
[8] Alcides Buss, Ruy Exel, and Ralf Meyer. Reduced C ∗ -algebras of Fell bundles over inverse semgroups.
arXiv:math.OA/1512.05570v1, 2015.
[9] John Crisp and Marcelo Laca. On the Toeplitz algebras of right-angled and finite-type Artin groups. J. Aust.
Math. Soc., 72(2):223–245, 2002.
[10] J. Duncan and A. L. T. Paterson. C ∗ -algebras of inverse semigroups. Proc. Edinburgh Math. Soc. (2), 28(1):41–
58, 1985.
[11] Ruy Exel. Inverse semigroups and combinatorial C ∗ -algebras. Bull. Braz. Math. Soc. (N.S.), 39(2):191–313,
2008.
[12] Ruy Exel and Starling Charles. Amenable actions of inverse semigroups. arXiv:math.OA/1411.256v1, 2014.
[13] Ruy Exel. Partial Dynamical Systems, Fell Bundles and Applications. arXiv:math.OA/1511.04565v1, 2015.
[14] Harry Furstenberg. Boundary theory and stochastic processes on homogeneous spaces. Harmonic analysis on
homogeneous spaces. Proc. Sympos. Pure Math., Vol. XXVI, 193–229. Amer. Math. Soc., Providence, R.I.,
1973.
[15] N. Higson, V. Lafforgue, and G. Skandalis. Counterexamples to the Baum-Connes conjecture. Geom. Funct.
Anal., 12(2):330–354, 2002.
[16] Michel Hilsum and Georges Skandalis. Morphismes K-orientés d’espaces de feuilles et fonctorialité en théorie
de Kasparov (d’après une conjecture d’A. Connes). Ann. Sci. École Norm. Sup. (4), 20(3):325–390, 1987.
[17] A. Hulanicki. Means and Følner condition on locally compact groups. Studia Math., 27:87–104, 1966.
[18] Mehrdad Kalantar and Matthew Kennedy. Boundaries of reduced C ∗ -algebras of discrete groups.
arXiv:mathOA/1405.4359v3, 2014.
[19] Mahmood Khoshkam and Georges Skandalis. Regular representation of groupoid C ∗ -algebras and applications
to inverse semigroups. J. Reine Angew. Math., 546:47–72, 2002.
[20] Mahmood Khoshkam and Georges Skandalis. Crossed products of C ∗ -algebras by groupoids and inverse semigroups. J. Oper. Theory, 51(2):255–279, 2004.
[21] Eberhard Kirchberg and Simon Wassermann. Operations on continuous bundles of C ∗ -algebras. Math. Ann.,
303(4):677–697, 1995.
[22] Marcelo Laca and Iain Raeburn. Semigroup crossed products and the Toeplitz algebras of nonabelian groups.
J. Funct. Anal., 139(2):415–440, 1996.
[23] N. P. Landsman and B. Ramazan. Quantization of Poisson algebras associated to Lie algebroids. In Groupoids
in analysis, geometry, and physics (Boulder, CO, 1999), volume 282 of Contemp. Math., pages 159–192. Amer.
Math. Soc., Providence, RI, 2001.
[24] Mark V. Lawson. Inverse semigroups. World Scientific Publishing Co., Inc., River Edge, NJ, 1998. The theory
of partial symmetries.
[25] Pierre-Yves Le Gall. Théorie de Kasparov équivariante et groupoı̈des. I. K-Theory, 16(4):361–390, 1999.
24
CLAIRE ANANTHARAMAN-DELAROCHE
[26] Xin Li. Semigroup C∗ -algebras and amenability of semigroups. J. Funct. Anal., 262(10):4302–4340, 2012.
[27] Xin Li. Nuclearity of semigroup C ∗ -algebras and the connection to amenability. Adv. Math., 244:626–662, 2013.
[28] Xin Li. Partial transformation groupoids attached to graphs and semigroups. arXiv:math.OA/1603.09165v1,
2016.
[29] Masayoshi Matsumura. A characterization of amenability of group actions on C ∗ -algebras. J. Operator Theory,
72(1):41–47, 2014.
[30] David Milan. C ∗ -algebras of inverse semigroups: amenability and weak containment. J. Operator Theory,
63(2):317–332, 2010.
[31] David Milan and Benjamin Steinberg. On inverse semigroup c∗ -algebras and crossed products. Groups, Geometry and Dynamics, 8(2):485–512, 2014.
[32] Paul S. Muhly and Dana P. Williams. Renault’s equivalence theorem for groupoid crossed products., volume 3
of NYJM Monographs. Albany, NY: State University of New York, University at Albany, 2008.
[33] Gerard J. Murphy. C ∗ -algebras generated by commuting isometries. Rocky Mountain J. Math., 26(1):237–267,
1996.
[34] Alexandru Nica. C ∗ -algebras generated by isometries and Wiener-Hopf operators. J. Operator Theory,
27(1):17–52, 1992.
[35] Alexandru Nica. On a groupoid construction for actions of certain inverse semigroups. Internat. J. Math.,
5(3):349–372, 1994.
[36] Magnus Dahler Norling. Inverse semigroup C ∗ -algebras associated with left cancellative semigroups. Proc.
Edinb. Math. Soc. (2), 57(2):533–564, 2014.
[37] Magnus Dahler Norling. The K-theory of some reduced inverse semigroup C ∗ -algebras. Math. Scand., 117:186–
202, 2015.
[38] Alan L. T. Paterson. Weak containment and Clifford semigroups. Proc. Roy. Soc. Edinburgh Sect. A, 81(12):23–30, 1978.
[39] Alan L. T. Paterson. Amenability, volume 29 of Mathematical Surveys and Monographs. American Mathematical Society, Providence, RI, 1988.
[40] Alan L. T. Paterson. Groupoids, inverse semigroups, and their operator algebras, volume 170 of Progress in
Mathematics. Birkhäuser Boston Inc., Boston, MA, 1999.
[41] Birant Ramazan. Quantification par déformation des variétés de Lie-Poisson. PhD thesis, University of Orléans,
1998.
[42] Jean Renault. A groupoid approach to C ∗ -algebras, volume 793 of Lecture Notes in Mathematics. Springer,
Berlin, 1980.
[43] Jean Renault. Représentation des produits croisés d’algèbres de groupoı̈des. J. Operator Theory, 18(1):67–97,
1987.
[44] Jean Renault. The ideal structure of groupoid crossed product C ∗ -algebras. J. Operator Theory, 25(1):3–36,
1991. With an appendix by Georges Skandalis.
[45] Jean Renault and S. Sundar. Groupoids associated to Ore semigroup actions. J. Operator Theory, 73(2):491–
514, 2015.
[46] Jean Renault and Dana Williams. Amenability of groupoids arising from partial semigroup actions and topological higher rank graphs. arXiv:math.OA/1501.03027v2, 2015.
[47] Marc A. Rieffel. Continuous fields of C ∗ -algebras coming from group cocycles and actions. Math. Ann.,
283(4):631–643, 1989.
[48] John Roe and Rufus Willett. Ghostbusting and property A. J. Funct. Anal., 266(3):1674–1684, 2014.
[49] Aidan Sims and Dana P. Williams. Renault’s equivalence theorem for reduced groupoid C ∗ -algebras. J. Operator Theory, 68(1):223–239, 2012.
[50] Rufus Willett. A non-amenable groupoid whose maximal and reduced C ∗ -algebras are the same.
arXiv:math.OA/1504.05615v2, 2015.
[51] Robert J. Zimmer. Hyperfinite factors and amenable ergodic actions. Invent. Math., 41(1):23–31, 1977.
SOME REMARKS ABOUT THE WEAK CONTAINMENT PROPERTY
25
[52] Robert J. Zimmer. On the von Neumann algebra of an ergodic group action. Proc. Amer. Math. Soc., 66(2):289–
293, 1977.
[53] Robert J. Zimmer. Amenable ergodic group actions and an application to Poisson boundaries of random walks.
J. Functional Analysis, 27(3):350–372, 1978.
Université d’Orléans et CNRS (UMR 7349 et FR2964),
B. P. 6759, F-45067 Orléans Cedex 2
E-mail address: [email protected]
| 4math.GR
|
arXiv:1602.03427v2 [math.ST] 31 May 2016
Aggregation of supports along the Lasso
path
Pierre C. Bellec
ENSAE-CREST
1
Abstract: In linear regression with fixed design, we propose two procedures that aggregate a data-driven collection of supports. The collection
is a subset of the 2p possible supports and both its cardinality and its elements can depend on the data. The procedures satisfy oracle inequalities
with no assumption on the design matrix. Then we use these procedures
to aggregate the supports that appear on the regularization path of the
Lasso in order to construct an estimator that mimics the best Lasso
estimator. If the restricted eigenvalue condition on the design matrix is
satisfied, then this estimator achieves optimal prediction bounds. Finally,
we discuss the computational cost of these procedures.
1. Introduction
Let n, p be two positive integers. We consider the mean estimation problem
Yi = µi + ξi ,
i = 1, ..., n,
where µ = (µ1 , ..., µn )T ∈ Rn is unknown, ξ = (ξ1 , ..., ξn )T is a subgaussian
vector, that is,
E[exp(v T ξ)] ≤ exp
σ 2 |v|22
2
for all v ∈ Rn ,
(1.1)
where σ > 0 is the noise level and | · |2 is the Euclidean norm in Rn . We
only observe y = (Y1 , ..., Yn )T and wish to estimate µ. A design matrix X of
size n × p is given and p may be larger than n. We do not require that the
model is well-specified, i.e., that there exists β ∗ ∈ Rp such that µ = Xβ ∗ .
Our goal is to find an estimator µ̂ such that the prediction loss kµ̂ − µk2 is
small, where k · k2 is the empirical loss defined by
kuk2 =
1
n
1 2
1X
u2 ,
|u|2 =
n
n i=1 i
u = (u1 , ..., un )T ∈ Rn .
Accepted for presentation at Conference on Learning Theory (COLT) 2016.
1
Bellec/Aggregation of supports along the Lasso path
2
In a high-dimensional setting where p > n, the Lasso is known to achieve
good prediction performance. For any tuning parameter λ > 0, define the
l
Lasso estimate β̂ λ as any solution of the convex minimization problem
l
β̂ λ ∈ argmin
β∈R p
1
|y − Xβ|22 + λ |β|1 ,
2n
(1.2)
where |β|1 = nj=1 |βj | is the ℓ1 -norm. If XT X/n = Ip×p where Ip×p is
the identity matrix
p of size p, then an optimal choice of the tuning parameter is λuniv ∼ σ log(p)/n, up to a numerical constant. If the Restricted
Eigenvalue condition holds
p(cf. Definition 1 below), then the universal tuning parameter λuniv ∼ σ log(p)/n leads to good prediction performance
[Bickel et al., 2009]. However, if the columns of X are correlated and the
Restricted Eigenvalue condition is not satisfied, the question of the optimal
choice of the tuning parameter λ is still unanswered, even if the noise level
σ 2 is known. Empirical and theoretical studies [van de Geer and Lederer,
2013, Hebiri and Lederer, 2013, Dalalyan et al., 2014] have shown that if
the columns of X are correlated, the Lasso estimate with a tuning parameter substantially smaller than the universal parameter leads to a prediction
performance which is substantially better than that of the Lasso estimate
with the universal parameter. To summarize, these papers raise the following
question:
P
Problem 1 (Data-driven selection of the tuning parameter). Find a datal
driven quantity λ̂ such that the prediction loss kµ − Xβ̂λ̂ k2 is small with high
probability.
In this paper, we focus on a different problem, namely:
Problem 2 (Lasso Aggregation). Construct an estimator µ̂ that mimics
the prediction performance of the best Lasso estimator, that is, construct an
estimator µ̂ such that with high probability,
kµ̂ − µk2 ≤ C min
λ>0
l
l
kXβ̂ λ − µk2 + ∆(β̂ λ ) ,
l
(1.3)
where C ≥ 1 is a constant and ∆(β̂ λ ) is a small quantity.
1 and 2 have the same goal, that is, to achieve a small prediction loss with
high probability. In 1, the goal is to select a Lasso estimate that has small
prediction loss. In 2, we look for an estimator µ̂ such that the prediction
performance of µ̂ is almost as good as the prediction performance of any
l
Lasso estimate. The estimator µ̂ may be of a different form than β̂ λ̂ for
some data-driven parameter λ̂.
Bellec/Aggregation of supports along the Lasso path
3
Our motivation to consider 2 instead of 1 is the following. Let µ1 , ..., µM
be deterministic vectors Rn . If the goal is to mimic the best approximation
of µ among µ1 , ..., µM , it is well known in the literature on aggregation
problems that an estimator of the form µ̂ = f k̂ for some data-driven integer
k̂ is suboptimal (cf. Theorem 2.1 in Rigollet and Tsybakov [2012], Section 2
of Juditsky et al. [2008] and Proposition 6.1 in Gerchinovitz [2011]). Thus,
an optimal procedure cannot be valued in the discrete set {µ1 , ..., µM }. Optimal procedures for this problem are valued in the convex hull of the set
{µ1 , ..., µM }. Examples are the Exponential Weights procedures proposed in
Leung and Barron [2006], Dalalyan and Salmon [2012] or the Q-aggregation
procedure of Dai et al. [2014].
Although a lot of progress has been made for various aggregation problems, to our knowledge no previous work deals with the problem of aggrel
gation of nonlinear estimators such as the collection (Xβ̂ λ )λ>0 based on the
sample. In the setting of the present paper, the observation y and the Lasso
estimates are not independent: no data-split is performed and the same data
is used to construct the Lasso estimators and to aggregate them.
We will show that aggregation of nonlinear estimators of the form Xβ̂ is
possible, for any nonlinear estimators β̂ and without any assumption on X.
For instance, an estimator µ̂ that achieves (1.3) with
σ 2 |β|0
ep
∆(β) ≃
log
n
|β|0 ∨ 1
is given in Section 3. Here, |β|0 denotes the number of nonzero coefficients
of β and a ∨ b = max(a, b).
Given a design matrix X, we call support any subset T of {1, ..., p}. The
cardinality of T is denoted by |T | and for β ∈ Rp , supp(β) is the set of
indices k = 1, ..., p such that βk 6= 0. Given a support T , we denote by ΠT
the square matrix of size n which is the orthogonal projection on the linear
span of the columns of X whose indices belong to T . Denote by P({1, ..., p})
the set of all subsets of {1, ..., p}. We will consider the following problem.
Problem 3 (Aggregation of a data-driven collection of supports). Let F̂ be a
data-driven collection of supports, that is, an estimator valued in P({1, ..., p}).
Construct an estimator µ̂ such that with high probability,
kµ̂ − µk2 ≤ min
T ∈F̂
kΠT µ − µk2 + ∆(T ) ,
where ∆(·) is a function that takes small values.
(1.4)
Bellec/Aggregation of supports along the Lasso path
4
The set F̂ is a family of supports. Let us emphasize that both its cardinality and its elements can depend on the data y. Note that for any support
T , ΠT µ = Xβ ∗T where β ∗T minimizes |Xβ − µ|22 subject to βk = 0 for all
k 6∈ T . In Section 3, we construct an estimator µ̂ that satisfies (1.4) with
∆(T ) ≃ σ 2 |T | log(p/|T |)/n for all nonempty supports T . In the literature on
aggregation problems, one is given a collection of estimators {µ̂1 , ..., µ̂M }
where M ≥ 1 is a deterministic integer and the goal is to mimic the best
estimator in this collection, cf. Tsybakov [2014] and the references therein.
A novelty of the present paper is to consider aggregation of a collection of
estimators, where the cardinality of the collection depends on the data.
The main contributions of the present paper are the following.
• In Section 2, we propose an estimator µ̂qF̂ ,σ̂2 that satisfies the oracle
inequality (1.4) with ∆(T ) ≃ σ̂ 2 |T | log(p/|T |)/n for all nonempty supports T , where σ̂ 2 is an estimator of the noise level. This estimator
solves 3. We explain in Corollary 1 how Section 2 can be used to construct a procedure that aggregates nonlinear estimators of the form
Xβ̂.
• Section 3 is devoted to 2. Using the result from Section 2, we construct
an estimator µ̂ that satisfies (1.3) with ∆(β) ≃ σ 2 |β|0 log(p/|β|0 ). The
computational complexity of the procedure is the sum of the complexity of the regularization path of the Lasso and the complexity of a
convex quadratic program.
The proofs can be found in the appendix.
2. Aggregation of a data-driven family of supports
Throughout this section, let F̂ be a data-driven collection of supports and
let σ̂ 2 ≥ 0 be a real valued estimator. Let M̂ be the cardinality of F̂ , and
let (T̂j )j=1,...,M̂ be supports such that
F̂ = {T̂1 , ..., TˆM̂ }.
(2.1)
For all supports T ⊂ {1, ..., p}, define the weights [Rigollet and Tsybakov,
2012]
!
!−1
p
e − e−p
|T |
πT := Hp
e
,
Hp :=
.
|T |
e−1
Note that by construction, the constant Hp is greater than 1 and T ∈P({1,...,p}) πT =
1 where P({1, ..., p}) is the set of all subsets of {1, ..., p}. Given a support T ,
P
Bellec/Aggregation of supports along the Lasso path
5
the Least Squares estimator on the linear span of the covariates indexed by
T is ΠT y.
We will consider two estimators of µ based on F̂ and σ̂ 2 . The first estimator is defined as follows. Define the criterion
Critσ̂2 (T ) = |y − ΠT y|22 + 18σ̂ 2 log
1
.
πT
We have
1
1
≤ + 2|T | log(ep/|T |)
(2.2)
πT
2
for any support T . The lower bound is a direct consequence of Hp > 1 and
the upper bound is proved in [Rigollet and Tsybakov, 2012, (5.4)]. As (2.2)
holds, the above criterion is of the same nature as Cp , AIC, BIC and their
variants, cf. Birgé and Massart [2001]. Define the estimator
|T | ≤ log
ΠT̂
F̂ ,σ̂ 2
(y)
T̂F̂ ,σ̂2 ∈ argmin Critσ̂2 (T ).
where
(2.3)
T ∈F̂
The estimator (2.3) is the orthogonal projection of y onto the linear span of
the columns of X whose indices are in T̂F̂ ,σ̂2 . If F̂ is not data-dependent, the
procedure ΠT̂ 2 (y) is close to the one studied in Birgé and Massart [2001].
F̂ ,σ̂
We now define a second estimator valued in the convex hull of (ΠT y)T ∈F̂ .
Let M̂ be the cardinality of F̂ , and let (T̂j )j=1,...,M̂ be supports such that
(2.1) holds. For any j = 1, ..., M̂ , let µ̂j = ΠT̂j y. Define a simplex in RM̂ as
follows:
n
ΛM̂ = θ ∈ RM̂ ,
M̂
X
j=1
For any θ ∈ RM̂ , define µ̂θ =
o
θj = 1, ∀j = 1 . . . M̂ , θj ≥ 0 .
PM̂
HF̂ ,σ̂2 (θ) := |µ̂θ − y|22 +
j=1 θj µ̂j .
For all θ ∈ ΛM̂ , let
M̂
X
1
1
.
θj log
penQ (θ) + 26σ̂ 2
2
π
T̂j
j=1
(2.4)
where
penQ (θ) :=
M̂
X
j=1
θj |µ̂j − µ̂θ |22 .
(2.5)
The penalty (2.5) is inspired by recent works on the Q-aggregation procedure [Dai et al., 2012], and it was used to derive sharp oracle inequalities for aggregation of linear estimators [Dai et al., 2014, Bellec, 2014a]
Bellec/Aggregation of supports along the Lasso path
6
and density estimators [Bellec, 2014b]. The penalty pushes µ̂θ towards the
P
1
is another penalty that
points {µ̂1 , ..., µ̂M̂ }. Finally, the term M̂
j=1 θj log π
T̂j
pushes the coordinate θj to 0 if the size of the support T̂j is large.
Define the estimator µ̂qF̂ ,σ̂2 as any minimizer of the function HF̂ ,σ̂2 defined
in (2.4):
(2.6)
θ̂ ∈ argmin HF̂ ,σ̂2 (θ).
µ̂qF̂ ,σ̂2 := µ̂θ̂ ,
θ∈ΛM̂
Theorem 1. Let n, p be positive integers and let σ > 0. Let µ ∈ Rn and X
be any matrix of size n × p. Let F̂ be any data-driven collection of subsets of
{1, ..., p}. Assume that the noise ξ satisfies (1.1). Let σ̂ 2 be any real valued
estimator and let δ := P(σ̂ 2 < σ 2 ). Then for all x > 0, the estimator µ̂qF̂ ,σ̂2
defined in (2.6) satisfies with probability greater than 1 − δ − 2 exp(−x),
kµ̂qF̂ ,σ̂2 −µk2
σ̂ 2
ep
≤ min kΠT µ − µk +
24 + 96|T | log
n
|T | ∨ 1
T ∈F̂
Furthermore, the estimator ΠT̂
F̂ ,σ̂
!
22σ 2 x
.
n
(2.7)
(y)
satisfies
with
probability
greater
than
2
2
+
1 − δ − 2 exp(−x),
ep
σ̂ 2
26 + 104|T | log
kΠT̂ 2 (y)−µk ≤ min 3kΠT µ − µk +
F̂ ,σ̂
n
|T | ∨ 1
T ∈F̂
2
2
!
+
(2.8)
In previously studied aggregation problems, one is given a collection of
estimators {µ̂1 , ..., µ̂M } where M ≥ 1 is a deterministic integer and the goal
is to construct an estimator µ̂ such that with high probability,
kµ̂ − µk2 ≤
min kµ̂j − µk2 + ∆n (M ),
j=1,...,M
where ∆n (M ) is a small error term that increases with M , cf. Tsybakov
[2014] and the references therein. Theorem 1 is of a different nature for
several reasons. First, the set F̂ is random, its cardinality can depend on
the observed data y. Second, the error term that appears inside the minimum
of (2.7) does not depend on the cardinality of F̂ .
The estimator µ̂qF̂ ,σ̂2 of Theorem 1 with σ̂ 2 = σ 2 and F̂ being the set of all
subsets of {1, ..., p} was previously studied as the Exponential Screening estimator [Rigollet and Tsybakov, 2011] or as the Sparsity Pattern Aggregate
[Rigollet and Tsybakov, 2012]. In this special case, F̂ is deterministic and
contains all the 2p possible supports. Because of this exponential number
28σ 2 x
.
n
Bellec/Aggregation of supports along the Lasso path
7
of supports, computing the sparsity pattern aggregate in practice is hard.
An MCMC algorithm is developed in Rigollet and Tsybakov [2011] to compute an approximate solution of the sparsity pattern aggregate, but to our
knowledge there is no theoretical guarantee that this MCMC algorithm will
converge to a good approximation in polynomial time. The Sparsity Pattern
Aggregate satisfies (2.7) with σ̂ 2 = σ 2 and F̂ = P({1, ..., p}). This sharp
oracle inequality yields the minimax rate over all ℓq balls for all 0 < q ≤ 1,
under no assumption on the design matrix X [Dai et al., 2014, Tsybakov,
2014].
To construct the estimator µ̂qF̂ ,σ̂2 , one has to solve the optimization problem (2.6). This is a convex quadratic program of size |F̂ | with a simplex constraint. The complexity of computing µ̂qF̂ ,σ̂2 is polynomial in the cardinality
of F̂ . Thus, if F̂ is small then it is possible to construct µ̂qF̂ ,σ̂2 efficiently.
As the cardinality of F̂ decreases, the prediction performance of the estimator µ̂qF̂ ,σ̂2 becomes worse, but computing µ̂qF̂ ,σ̂2 becomes easier.
Problem 4. Construct a data-driven set of supports F̂ such that with high
probability, there exists a support T ∈ F̂ for which, simultaneously, the bias
kΠT µ − µk2 and the size |T | are small.
If we can construct such a set F̂ , by (2.7) the prediction loss of the estimator µ̂qF̂ ,σ̂2 will be small. Note that Theorem 1 needs no assumption on
the data-driven set F̂ and the design matrix X.
In the following Corollary, we perform aggregation of a family of nonlinear
estimators of the form (Xβ̂ k )j∈J for some set J. All estimators in the family
share the same design matrix X and this matrix is deterministic.
Corollary 1. Let n, p be positive integers and let σ > 0. Let µ ∈ Rn
and X be any matrix of size n × p. Let F̂ be any data-driven collection of
subsets of {1, ..., p}. Assume that the noise ξ satisfies (1.1). Let (β̂ j )j∈Jˆ be
a family of estimators valued in Rp . Both the cardinality of the family and
its elements can depend on the data. Let σ̂ 2 be any real valued estimator
and let δ := P(σ̂ 2 < σ 2 ). Define F̂ = {supp(β̂ j ), j ∈ Jˆ} and let µ̂qF̂ ,σ̂2 be
the estimator (2.6). Then for all x > 0, the estimator µ̂qF̂ ,σ̂2 satisfies with
probability greater than 1 − δ − 2 exp(−x),
kµ̂qF̂ ,σ̂2 −µk2
σ̂ 2
≤ min kXβ̂ j − µk +
n
j∈Jˆ
2
24 + 96|β̂ j |0 log
ep
|β̂ j |0 ∨ 1
!!!
+
22σ 2 x
.
n
Using (2.8), a similar result can be readily obtained for the estimator
Bellec/Aggregation of supports along the Lasso path
ΠT̂
F̂ ,σ̂ 2
8
(y) with the leading constant 3.
3. Aggregation of supports along the Lasso path
Let us recall some properties of the Lasso path [Efron et al., 2004]. For a
given observation y, there exists a positive integer K and a finite sequence
λ0 > λ1 > ... > λK = 0
l
such that β̂ λ = 0 for all λ > λ0 , and such that
∀λ ∈ (λk+1 , λk ),
l
l
supp(β̂ λ ) = supp(β̂ λk ).
Thus, there is a finite number of supports on the Lasso path. In this section,
l
we study the estimator of Theorem 1 in the special case F̂ = {supp(β̂ λk ), k =
0, ..., K}, that is, we aggregate all the supports that appear on the Lasso
path.
Theorem 2. Let n, p be positive integers and let σ > 0. Let µ ∈ Rn and X
be any matrix of size n × p. Assume that the noise ξ satisfies (1.1). Let σ̂ 2 be
any real valued estimator and let δ := P(σ̂ 2 < σ 2 ). Let λ0 > ... > λK be the
l
knots of the Lasso path. Let F̂ = {supp(β̂ λj ), j = 0, ..., K} be the family of all
supports that appear on the Lasso path and let µ̂qF̂ ,σ̂2 be the estimator (2.6).
Then for all x > 0, the estimator µ̂qF̂ ,σ̂2 satisfies with probability greater than
1 − δ − 2 exp(−x),
kµ̂qF̂ ,σ̂2 −µk2 ≤ min
λ>0
l
kXβ̂ λ
σ̂ 2
− µk2 +
n
24 +
l
96|β̂ λ |0 log
ep
l
|β̂ λ |0
∨1
!!!
22σ 2 x
,
n
(3.1)
+
l
where for all λ > 0, β̂ λ is the Lasso estimator (1.2).
Using (2.8), a similar result can be readily obtained for the estimator
ΠT̂ 2 (y) with the leading constant 3.
F̂ ,σ̂
The computational complexity of the procedure of Theorem 2 is polynomial in the number of knots of the Lasso path. This will be further discussed
in Section 4. In the rest of this section, we assume that σ̂ 2 = σ 2 and δ = 0.
We will come back to the estimation of the noise level in Section 5 below.
Interestingly, Theorem 2 does not need any assumption on the design
matrix X. The estimators µ̂qF̂ ,σ̂2 and ΠT̂ 2 (y) have a good performance as
F̂ ,σ̂
l
soon as for some possibly unknown λ > 0, both the support of β̂ λ and the
l
loss kXβ̂ λ − µk2 are small.
Bellec/Aggregation of supports along the Lasso path
9
3.1. Prediction guarantees under the restricted eigenvalue
condition
The goal of this section is to study the prediction performance of the procedure defined in Theorem 2 under the Restricted Eigenvalue condition on
the design matrix X.
Definition 1. For any s ∈ {1, ..., p} and c0 > 0, condition RE(s, c0 ) is
satisfied if
κ(s, c0 ) :=
min
min
T ⊂{1,...,p}:|T |≤s δ∈R p :|δ T c |1 ≤c0 |δ T |1
|Xδ|2
√
> 0.
n|δ T |2
The following result is a reformulation of Bickel et al. [2009, Theorem
6.2].
Theorem 3 (Bickel et al. [2009]). Let X be such that the diagonal elements
of XT X/n are all equal to 1. Assume that µ = Xβ∗ and let s := |β ∗ |0 .
Assume that ξ ∼ N (0, σ 2 In×n ) and that condition RE(s, 3) is satisfied. Let
−x0 on which
x0 > 0. There is an event Ω(x0 ) of probability greater than 1−e
p
the Lasso estimator (1.2) with tuning parameter λx0 = σ 8(x0 + log p)/n
satisfies simultaneously
64φmax
s,
κ2 (s, 3)
128σ 2 s(x0 + log p)
− β ∗ )k2 ≤
,
κ2 (s, 3)n
l
|β̂ λx |0 ≤
0
l
kX(β̂ λx0
(3.2)
(3.3)
where φmax is the largest eigenvalue of the matrix XT X/n.
Thus, if the restricted eigenvalue condition
p is satisfied, the Lasso estimator with the universal parameter λx0 = σ 8(x0 + log p)/n enjoys simultaneously an ℓ0 norm of the same order as the true sparsity (cf. (3.2)), and a
prediction loss of order s log(p)/n (cf. (3.3)).
Theorem 4 below is a direct consequence of Theorem 2 and the bounds
(3.2)-(3.3).
Theorem 4. Let n, p be positive integers and let σ > 0. Let µ ∈ Rn and
X be any matrix of size n × p. Let F̂ be any data-driven subset of {1, ..., p}.
Assume that µ = Xβ∗ and let s := |β ∗ |0 . Assume that ξ ∼ N (0, σ 2 In×n )
and that condition RE(s, 3) is satisfied.
l
Let λ0 > ... > λK be the knots of the Lasso path. Let F̂ = {supp(β̂ λj ), j =
0, ..., K} be the family of all supports that appear on the Lasso path and let
Bellec/Aggregation of supports along the Lasso path
10
µ̂qF̂ ,σ2 be the estimator (2.6) with σ̂ 2 = σ 2 . Then for all x > 0, the estimator
µ̂qF̂ ,σ2 satisfies with probability greater than 1 − 3 exp(−x),
kµ̂qF̂ ,σ2 − Xβ ∗ k2 ≤
128σ 2 sx
22σ 2 x
(128 + 48φmax )σ 2 s log p 24σ 2
+
+
+
.
κ2 (s, 3)n
n
κ2 (s, 3)n
n
(3.4)
Furthermore,
Ekµ̂qF̂ ,σ2 − Xβ ∗ k2 ≤
(128 + 48φmax )σ 2 s log p
384σ 2 s
90σ 2
+
+
.
κ2 (s, 3)n
κ2 (s, 3)n
n
(3.5)
Using (2.8), a similar result can be readily obtained for the estimator
ΠT̂ 2 (y) with different constants.
F̂ ,σ̂
Proof of Theorem 4. By Theorem 2 with δ = 0, there is an event Ωagg (x) of
probability greater than 1 − 2e−x such that on Ωagg (x) we have
kµ̂qF̂ ,σ2 −Xβ∗ k2
≤
σ
l
kX(β̂ λx −β∗ )k2 +
2
n
24 +
l
96|β̂ λx |0 log
ep
l
|β̂ λ |0 ∨ 1
!!
+
22σ 2 x
.
n
Let Ω(x) be the event defined in Theorem 3. Using the simple inequality
l
log(p/(|β̂ λ |0 ∨ 1)) ≤ log p, and the bounds (3.2)-(3.3), we obtain that (3.4)
holds on the event Ωagg (x) ∩ Ω(x). By the union bound, the event Ωagg (x) ∩
Ω(x) has probability greater than 1 − 3e−x . Finally, (3.5) is obtained from
(3.4) by integration.
The procedure studied in Theorem 4 aggregates the supports along the
Lasso path using the procedure (2.6). A similar result holds for the estimator
ΠT̂ 2 (y) with a leading constant equal to 3. Theorem 4 has the following
F̂ ,σ̂
implications.
First, if x > 0 is fixed, the prediction performance (3.4) of the estimator
µ̂qF̂ ,σ2 is similar to that of the Lasso with the universal tuning parameter λx ,
up to a multiplicative factor that only involves numerical constants and the
quantity φmax . As soon as φmax (the operator norm of XT X/n) is bounded
from above by a constant, the estimator studied in Theorem 4 enjoys the
best known prediction guarantees.
Second, Theorem 4 implies that the estimator µ̂qF̂ ,σ2 satisfies the prediction bound (3.4) simultaneously for all confidence levels. That is, (3.4) holds
for all x > 0 with probability greater than 1 − 3e−x , in contrast with the
Lasso estimator with the universal parameter λx0 which depends on a fixed
confidence level 1 − e−x0 . The Lasso estimator with the universal parameter
Bellec/Aggregation of supports along the Lasso path
11
λx0 satisfies the prediction bound (3.3) only for the confidence level 1 − e−x0 ,
but to our knowledge it is not known whether the Lasso estimator with the
universal parameter λx0 satisfies a similar bound for different confidence levels than 1−e−x0 . In this regard, the estimator studied in Theorem 4 provides
a strict improvement compared to the Lasso with the universal parameter.
Third, the estimator µ̂qF̂ ,σ2 of Theorem 4 satisfies the bound (3.5), that is,
a prediction bound in expectation. Again, to our knowledge, it is not known
whether the Lasso estimator with the universal parameter satisfies a similar
bound in expectation.
Assuming that the bound (3.3) is tight and putting computational issues
aside, the prediction performance of the procedure µ̂qF̂ ,σ2 of Theorem 4 is
substantially better than the performance of the Lasso with the universal
parameter, as soon as φmax is bounded from above by a constant.
An upper bound similar to (3.2) is given in [Belloni et al., 2014, Theorem
3 and Remark 3]. Namely, Belloni et al. [2014] prove that the square-root
Lasso estimator with the universal tuning parameter β̂ satisfies |β̂|0 ≤ Cs
with high probability, where s is the sparsity of the true parameter and C is
a constant that depends on the sparse eigenvalues of the matrix XT X/n, cf.
[Belloni et al., 2014, Condition P]. This upper bound can be used instead
of (3.2) to prove results similar to (3.4) where φmax is replaced by a smaller
constant that depends on the sparse eigenvalues of XT X/n.
4. Computational complexity of the Lasso path and µ̂qF̂ ,σ̂2
Computing the estimator µ̂qF̂ ,σ̂2 of Theorem 2 is done in two steps:
1. Compute the full Lasso path and let F̂ = {supp(λ0 ), ..., supp(λK )} be
all the supports that appear on the Lasso path, where λ0 , ..., λK are
the knots of the Lasso path.
2. Compute µ̂qF̂ ,σ̂2 as a solution of the quadratic program (2.6), where F̂
is defined by Step 1.
(We assume that the complexity of computing σ̂ 2 is negligible compared to
the complexity of Step 1 and Step 2 above). The time complexity of Step 2
is the complexity of a convex quadratic program of size |F̂ | ≤ K, where K is
the number of knots on the Lasso path. Thus, the global cost of computing
the estimator µ̂qF̂ ,σ̂2 of Theorem 2 is polynomial in K.
There exist efficient algorithms to compute the entire Lasso path [Efron et al.,
2004]. However, Mairal and Yu [2012] proved that for some values of X and y,
the regularization path of the Lasso contains more than 3p /2 knots. Hence,
Bellec/Aggregation of supports along the Lasso path
12
for some design matrix X and some observation y, an exact computation
of the full Lasso path is not realizable in polynomial time. In order to fix
this computational issue, Mairal and Yu [2012] propose an algorithm that
computes an approximate regularization path for the Lasso. For some fixed
√
ǫ > 0, this algorithm is guaranteed to terminate with less than O(1/ ǫ)
knots and the points on the approximate path have a duality gap smaller
than ǫ. This approximation algorithm can be used instead of computing the
exact Lasso path. That is, one may compute the estimator µ̂qF̂ ,σ̂2 where F̂
is the collection of supports that appear on the approximate path computed
by the algorithm of Mairal and Yu [2012].
Another solution to avoid computational issues is as follows. Let M be
a positive integer. Instead of computing the Lasso path, one may consider
a grid of tuning parameters λ1 , ..., λM > 0 and aggregate the supports of
l
l
corresponding Lasso estimates β̂ λ1 , ...β̂ λM . The advantage of this approach
l
is twofold. First, for all j = 1, ..., M the Lasso estimate β̂ λj can be computed
by standard convex optimization solvers. Second, the time complexity of the
procedure is guaranteed to be polynomial in M and p. For any x > 0, by
Corollary 1, this procedure satisfies, with probability greater than 1 − 3e−x
kµ̂qF̂ ,σ̂2 −µk2
≤
min
j=1,...,M
kXβ̂ l
λj
22σ 2 x
σ̂ 2
ep
l
+
− µk2 +
.
24 + 96|β̂ λj |0 log l
n
n
|β̂ λ |0 ∨ 1
j
This oracle inequality is not a strong as (3.1). However, if at least one of
l
the Lasso estimates {β̂ λj , j = 1, ..., M } enjoys a small prediction loss and a
small ℓ0 norm, then the prediction loss of µ̂qF̂ ,σ̂2 is also small.
5. A fully data-driven procedure using the Square-Root Lasso
This section proposes a fully data-driven procedure, based on the SquareRoot Lasso. The choice of grid comes from the empirical and theoretical
observations that for a correlated design matrix, there exists a tuning parameter smaller than the universal parameter which enjoys better prediction
performance than the universal parameter [van de Geer and Lederer, 2013,
Hebiri and Lederer, 2013, Dalalyan et al., 2014].
p
1. Let λmax = 2 log(p/0.01)/n be the universal parameter of the SquareRoot Lasso [Belloni et al., 2014] with confidence level 0.01.
2. Let λmin be a conservatively small value of the tuning parameter.
3. Let M be an integer.
Bellec/Aggregation of supports along the Lasso path
13
4. Consider the geometric grid {λ1 , ..., λM } such that
λj = λmin
λmax
λmin
((j−1)/M −1)
,
j = 1, ..., M.
sq
sq
5. Compute the Square-Root Lasso estimators β̂ λ1 , ...β̂ λM with parameters λ1 , ..., λM (it is possible to perform this computation simultaneously for all λ1 , ..., λM , cf. Pham et al. [2014] and the references
therein).
sq
6. Let F̂ = {supp(β̂ λj ), j = 1, ..., M } be the supports of the computed
Square-Root Lasso estimators.
7. Let σ̂ 2 be the variance estimated by the Square-Root Lasso with the
universal parameter λmax .
8. For this choice of σ̂ 2 and F̂ , return the estimator µ̂qF̂ ,σ̂2 or the estimator
ΠT̂ 2 (y) .
F̂ ,σ̂
This estimator µ̂qF̂ ,σ̂2 returned by this procedure enjoys the theoretical guarantee
kµ̂qF̂ ,σ̂2 −µk2 ≤
sq
min kXβ̂ λj − µk2 +
j=1,...,M
σ̂ 2
n
24 + 96|β̂ sq |0 log
λj
ep
sq
|β̂ λj |0
∨1
+
with probability greater than 1 − 3e−x . A similar guarantee with leading
constant 3 can be obtained for the estimator ΠT̂ 2 (y) using (2.8).
F̂ ,σ̂
6. Concluding remarks
We have presented two procedures (2.3) and (2.6) that aggregates a datadriven collection of supports F̂ . These procedures satisfy the oracle inequalities given in Theorem 1 above, which is the main result of the paper. Sections 3 and 4 study the situation where F̂ is the collection of supports that
appear along the Lasso path. These procedures may be used for other datadriven collections F̂ as well.
These procedures allow one to perform a trade-off between prediction
performance and computational cost. If F̂ contains all the 2p supports, these
procedures achieve optimal prediction guarantees with no assumption on the
design matrix X, but can not be realized in polynomial time. On the other
hand, if the cardinality of F̂ is small (say, polynomial in n and p), then it
is possible to compute the estimators (2.3) and (2.6) in polynomial time. In
view of (1.3), one should look for a data-driven set F̂ with the following
properties.
22σ 2 x
n
Bellec/Aggregation of supports along the Lasso path
14
1. The set F̂ is small so that the estimators (2.6) and (2.3) can be computed rapidly,
2. The set F̂ contains a support T such that |T | and kπT µ − µk2 are
simultaneously small, so that the procedures (2.6) and (2.3) enjoy good
prediction performance.
A natural choice for F̂ is the collection of supports that appear along the
Lasso path. This choice of F̂ was studied in Sections 3 and 4. Another natural
choice is to aggregate the supports of several hard-thresholded Lasso estimators, since the hard-thresholded Lasso is sign-consistent under weak conditions on the design [Meinshausen and Yu, 2009, Definition 5 and Corollary
2]. Further research will investigate other means to construct a data-driven
collection F̂ such that the above two properties are satisfied.
Acknoledgements
We would like to thank Alexandre Tsybakov for helpful comments during
the writing of this manuscript.
Appendix A: Proof of Theorem 1
For any matrix A ∈ Rn×n , define the operator norm of A and the Frobenius
norm of A by
|||A|||2 := sup |Au|2 ,
|u|22 =1
kAkF =
q
Tr(AT A),
respectively.
Proof of (2.7). For all S, T ⊂ {1, ..., p}, define the event
2
ΩS,T = Z(S, T ) ≤ 4σ |S| + 22σ
2
1
log
+x
πS πS
,
where
1
(A.1)
Z(S, T ) = 2ξ T (ΠS y − ΠT µ) − |ΠS y − ΠT y|22 .
2
Define the event V := {σ̂ 2 ≥ σ 2 }. On the event A := V ∩ (∩S,T ⊂{1,...,p}ΩS,T ),
we have simultaneously for all supports S, T
Z(S, T ) − 26σ̂ 2 log
1
1
1
− 22σ 2 log
≤ 22σ 2 x + 4σ 2 |S| − 4σ 2 log
≤ 22σ 2 x
πS
πT
πS
Bellec/Aggregation of supports along the Lasso path
where we have used that log
A we have
|µ̂qF̂ ,σ̂2
−
µ|22
1
πS
≥ |S|, cf. (2.2). By Lemma 1, on the event
≤ min |ΠT µ −
T ∈F̂
15
µ|22
1
+ (26σ̂ + 22σ ) log
πT
2
2
+ 22σ 2 x.
To obtain (2.7), we use (2.2) and the fact that on the event V, 26σ̂ 2 + 22σ 2 ≤
48σ̂ 2 .
It remains to bound from below the probability of the event A. Denote
by B c the complement of any event B. We proceed with the union bound as
follows,
X
P(ΩcS,T ).
P(Ac ) ≤ P(V c ) +
S,T ⊂{1,...,p}
By definition, δ = P(V c ) and for any S, T ⊂ {1, ..., p}, Lemma 2 with t =
P
x + log πS1πT yields that P(ΩcS,T ) ≤ πS πT 2 exp(−x). As S,T ⊂{1,...,p} πS πT =
P
( S⊂{1,...,p} πS )2 = 1, we have established that
P(Ac ) ≤ δ + 2 exp(−x).
The proof of (2.8) is close to the argument used in Birgé and Massart
[2001], cf. [Giraud, 2015, Section 2.3] for a recent reference on model selection.
The novelty of the present paper is to consider a data-driven collection of
estimators.
Proof of (2.8). Let Λ̂ = 18σ̂ 2 and let T̂ = T̂F̂ ,σ̂2 for notational simplicity. By
definition of ΠT̂ 2 (y) = ΠT̂ y, for all T ∈ F̂ we have Critσ̂2 (T̂ ) ≤ Critσ̂2 (T )
F̂ ,σ̂
which can be rewritten as
|ΠT̂
F̂ ,σ̂ 2
(y) − µ|22 + Λ̂ log
1
1
≤ |ΠT y − µ|22 + Λ̂ log
+ 2ξ T (ΠT̂ y − ΠT y),
πT̂
πT
1
≤ |ΠT µ − µ|22 + Λ̂ log
+ 2ξ T ΠT̂ ξ + 2ξ T (ΠT̂ µ − ΠT µ) − |ΠT ξ|22 .
πT
(A.2)
Define the event V := {σ̂ 2 ≥ σ 2 }. For all S, T ⊂ {1, ..., p}, define
W (S) = 2ξ T ΠS ξ − 10σ 2 log
1
,
πS
W ′ (S, T ) = 2ξ T (ΠS µ − ΠT µ) − 8σ 2 log
1
1
− |ΠS µ − ΠT µ|22 .
πS πT
4
Bellec/Aggregation of supports along the Lasso path
16
With this notation, using the simple inequality −|ΠT ξ|22 ≤ 0, (A.2) implies
that on the event V,
|ΠT̂
F̂ ,σ̂ 2
(y) − µ|22 ≤ |ΠT µ − µ|22 + Λ̂ log
1
1
1
+ 8σ 2 log
+ W (T̂ ) + W ′ (T̂ , T ) + |ΠT̂ µ − ΠT µ|22 ,
πT
πT
4
Using that |ΠT̂ µ−ΠT µ|22 ≤ 2|ΠT̂ µ−µ|22 +2|µ−ΠT µ|22 and that |ΠT̂ µ−µ|22 ≤
|ΠT̂ y − µ|22 , we obtain
1
3
1
1
|ΠT̂ 2 (y) − µ|22 ≤ |ΠT µ − µ|22 + Λ̂ log
+ 8σ 2 log
+ W (T̂ ) + W ′ (T̂ , T ).
F̂ ,σ̂
2
2
πT
πT
For all S, T ⊂ {1, ..., p}, define the events
ΩS := {W (S) ≤ 6σ 2 x},
ΩS,T := {W ′ (S, T ) ≤ 8σ 2 x}.
On the event V ∩ (∩S⊂{1,...,p}ΩS )∩ (∩S,T ⊂{1,...,p}ΩS,T ), (2.8) holds. It remains
to bound from below the probability of this event.
For any fixed S ⊂ {1, ..., p}, using (2.2) and (B.4) with t = x + log π1S we
have P(ΩcS ) ≤ πS e−x .
Let S, T ⊂ {1, ..., p} be fixed. By using (B.3) with v = 2(ΠS µ − ΠT µ))
and t = x + log πS1πT , we have that on an event of probability greater than
1 − πS πT e−x ,
q
T
2ξ (ΠS µ−ΠT µ) ≤ 2σ 2(x + log(1/πS πT ))|ΠS µ−ΠT µ|2 ≤ 8σ
2
1
1
x + log
+ |ΠS µ−ΠT µ|22 .
πS πT
4
Thus, P(ΩcS,T ) ≤ πS πT e−x .
As in the proof of (2.7), the union bound completes the proof.
Appendix B: Technical Lemmas
Lemma 1. For any estimator σ̂ 2 , let θ̂ be a minimizer of (2.4). Then,
almost surely,
|µ̂θ̂ −
µ|22
≤
min
k=1,...,M̂
where
W := max
S,T ∈F̂
|ΠT̂k µ −
µ|22
1
+ (26σ̂ + 22σ ) log
πT̂k
2
2
!
1
1
Z(S, T ) − 26σ̂ log
− 22σ 2 log
πS
πT
and Z(·, ·) is defined in (A.1).
2
+ W, (B.1)
Bellec/Aggregation of supports along the Lasso path
17
Proof of Lemma 1. Let Λ̂ = 26σ̂ 2 . The function HF̂ ,σ̂2 is convex and differentiable, it can be rewritten as
M̂
X
1
1
1
∀θ ∈ Λ , HF̂ ,σ̂2 (θ) = |µ̂θ |22 +|y|22 +
θj −2yT µ̂j + |µ̂j |22 + Λ̂ log
2
2
πT̂j
j=1
M̂
!
.
By simple algebra, for any θ ′ ∈ RM̂ ,
T
′
∇HF̂ ,σ̂2 (θ̂) θ =
µ̂Tθ̂ µ̂θ′
+
T
M̂
X
′
θj
j=1
∇HF̂ ,σ̂2 (θ̂) (−θ̂) = −|µ̂θ̂ −
µ|22
+
1
1
−2y µ̂j + |µ̂j |22 + Λ̂ log
2
πT̂j
T
|µ|22
+
M̂
X
j=1
θ̂j
!
, (B.2)
1
1
2ξ µ̂j − |µ̂j |22 − Λ̂ log
2
πT̂j
T
By summing the last display and equality (B.2) applied to θ ′ = ek , we get
∇HF̂ ,σ̂2 (θ̂)T (ek − θ̂) = −|µ̂θ̂ − µ|22 + |µ̂k − µ|22 + Λ̂ log
M̂
X
1
πT̂k
"
#
1
1
.
θ̂j 2ξ T (µ̂j − µ̂k ) − |µ̂j − µ̂k |22 − Λ̂ log
+
2
πT̂j
j=1
Since µ̂k = ΠT̂k y is a Least Squares estimator over the linear span of the
covariates in T̂k , we have |µ̂k − y|22 ≤ |ΠT̂k µ − y|22 which can be rewritten as
|µ̂k − µ|22 ≤ |ΠT̂k µ − µ|22 + 2ξ T (µ̂k − ΠT̂k µ).
We thus have
∇HF̂ ,σ̂2 (θ̂)T (ek − θ̂) ≤ −|µ̂θ̂ − µ|22 + |ΠT̂k µ − µ|22 + (Λ̂ + 22σ 2 ) log
M̂
X
"
1
πT̂k
#
1
1
1
θ̂j 2ξ (µ̂j − ΠT̂k µ) − |µ̂j − µ̂k |22 − Λ̂ log
− 22σ 2 log
+
.
2
πT̂j
πT̂k
j=1
T
For all k = 1, ..., M̂ , [Boyd and Vandenberghe, 2009, Section 4.2.3] yields
∇HF̂ ,σ̂2 (θ̂)T (ek − θ̂) ≥ 0. Furthermore, a linear function over the simplex is
maximized at a vertex, so almost surely we obtain (B.1).
!
,
Bellec/Aggregation of supports along the Lasso path
18
Lemma 2. Let t > 0. For any supports S, T ⊂ {1, ..., p}, the quantity
Z(S, T ) defined in (A.1) satisfies with probability greater than 1 − 2 exp(−t),
Z(S, T ) ≤ 4σ 2 |S| + 22σ 2 t.
Proof of Lemma 2. Let D = ΠS − ΠT . Then almost surely,
1
1
Z(S, T ) = 2ξ T ΠS ξ + ξ T (2Dµ − D2 µ) − |Dµ|22 − |Dξ|22 .
2
2
It is clear that −|Dξ|22 ≤ 0. As ξ satisfies (1.1), a Chernoff bound yields that
for all v ∈ Rn ,
√
P ξ T v > σ|v|2 2t ≤ exp(−t).
(B.3)
It is clear that |||D|||2 ≤ 2. We apply this concentration inequality to v =
2Dµ − D2 µ to get that with probability greater than 1 − exp(−t),
√
√
ξ T (2Dµ − D 2 µ) ≤ σ|2Dµ − D2 µ|2 2t ≤ σ|||2In − D|||2 |Dµ|2 2t,
√
1
≤ σ4|Dµ|2 2t ≤ 16σ 2 t + |Dµ|22 .
2
Finally, let r ≤ |S| be the rank of ΠS . The matrix ΠS is an orthogonal projector. Hence kΠS k2F = r and |||ΠS |||2 ≤ 1, so that applying the concentration
inequality from Hsu et al. [2012] yields that with probability greater than
1 − exp(−t),
√
(B.4)
2ξ T ΠS ξ ≤ 2σ 2 (r + 2 rt + 2t) ≤ 4σ 2 r + 6σ 2 t ≤ 4σ 2 |S| + 6σ 2 t.
A union bound completes the proof.
References
Pierre C. Bellec.
Optimal bounds for aggregation of affine
estimators.
arXiv:1410.0346, Submitted, 2014a.
URL
http://arxiv.org/abs/1410.0346.
Pierre C. Bellec.
Optimal exponential bounds for aggregation
of density estimators.
Accepted in Bernoulli, 2014b.
URL
http://arxiv.org/abs/1405.3907.
Alexandre Belloni, Victor Chernozhukov, and Lie Wang.
Pivotal estimation via square-root Lasso in nonparametric regression.
Ann. Statist., 42(2):757–788, 2014.
ISSN 0090-5364.
.
URL
http://dx.doi.org/10.1214/14-AOS1204.
Bellec/Aggregation of supports along the Lasso path
19
Peter J. Bickel, Ya’acov Ritov, and Alexandre B. Tsybakov. Simultaneous
analysis of lasso and Dantzig selector. Ann. Statist., 37(4):1705–1732,
2009. ISSN 0090-5364. . URL http://dx.doi.org/10.1214/08-AOS620.
Lucien Birgé and Pascal Massart. Gaussian model selection. J. Eur.
Math. Soc. (JEMS), 3(3):203–268, 2001. ISSN 1435-9855. . URL
http://dx.doi.org/10.1007/s100970100031.
Stephen Boyd and Lieven Vandenberghe. Convex optimization. Cambridge
university press, 2009.
Dong Dai, Philippe Rigollet, and Tong Zhang. Deviation optimal learning
using greedy Q-aggregation. Ann. Statist., 40(3):1878–1905, 2012. ISSN
0090-5364. . URL http://dx.doi.org/10.1214/12-AOS1025.
Dong Dai, Philippe Rigollet, Lucy Xia, and Tong Zhang. Aggregation
of affine estimators. Electron. J. Statist., 8(1):302–327, 2014. . URL
http://dx.doi.org/10.1214/14-ejs886.
Arnak S. Dalalyan and Joseph Salmon. Sharp oracle inequalities for aggregation of affine estimators. Ann. Statist., 40(4):2327–2355, 2012. ISSN
0090-5364. . URL http://dx.doi.org/10.1214/12-AOS1038.
Arnak S Dalalyan, Mohamed Hebiri, and Johannes Lederer. On the prediction performance of the lasso. arXiv preprint arXiv:1402.1700, 2014.
Bradley Efron, Trevor Hastie, Iain Johnstone, and Robert Tibshirani. Least
angle regression. Ann. Statist., 32(2):407–499, 2004. ISSN 0090-5364. .
URL http://dx.doi.org/10.1214/009053604000000067. With discussion, and a rejoinder by the authors.
Sébastien Gerchinovitz. Prediction of individual sequences and prediction in
the statistical framework: some links around sparse regression and aggregation techniques. PhD thesis, Université Paris Sud-Paris XI, 2011. URL
https://tel.archives-ouvertes.fr/tel-00653550.
Christophe Giraud. Introduction to high-dimensional statistics, volume 139
of Monographs on Statistics and Applied Probability. CRC Press, Boca
Raton, FL, 2015. ISBN 978-1-4822-3794-8.
Mohamed Hebiri and Johannes Lederer. How correlations influence lasso
prediction. IEEE Trans. Inform. Theory, 59(3):1846–1854, March 2013. .
URL http://dx.doi.org/10.1109/tit.2012.2227680.
Daniel Hsu, Sham M. Kakade, and Tong Zhang. A tail inequality
for quadratic forms of subgaussian random vectors. Electron. Commun. Probab., 17:no. 52, 6, 2012.
ISSN 1083-589X.
.
URL
http://dx.doi.org/10.1214/ECP.v17-2079.
A. Juditsky, P. Rigollet, and A. B. Tsybakov. Learning by mirror averaging. Ann. Statist., 36(5):2183–2206, 2008. ISSN 0090-5364. . URL
http://dx.doi.org/10.1214/07-AOS546.
Bellec/Aggregation of supports along the Lasso path
20
Gilbert Leung and Andrew R. Barron.
Information theory
and mixing least-squares regressions.
IEEE Trans. Inform.
Theory, 52(8):3396–3410, 2006.
ISSN 0018-9448.
.
URL
http://dx.doi.org/10.1109/TIT.2006.878172.
Julien Mairal and Bin Yu. Complexity analysis of the lasso regularization
path. arXiv preprint arXiv:1205.0079, 2012.
Nicolai Meinshausen and Bin Yu. Lasso-type recovery of sparse representations for high-dimensional data. Ann. Statist., 37(1):246–270, 2009. ISSN
0090-5364. . URL http://dx.doi.org/10.1214/07-AOS582.
Vu Pham, Laurent El Ghaoui, and Arturo Fernandez. Robust sketching
for multiple square-root lasso problems. arXiv preprint arXiv:1411.0024,
2014.
Philippe Rigollet and Alexandre Tsybakov. Exponential screening and optimal rates of sparse estimation. Ann. Statist., 39(2):731–771, 2011. ISSN
0090-5364. . URL http://dx.doi.org/10.1214/10-AOS854.
Philippe Rigollet and Alexandre B. Tsybakov. Sparse estimation by exponential weighting. Statist. Sci., 27(4):558–575, 2012. ISSN 0883-4237. .
URL http://dx.doi.org/10.1214/12-STS393.
A.B. Tsybakov. Aggregation and minimax optimality in high-dimensional estimation. In Proceedings of the International Congress of Mathematicians,
Seoul, 2014. To appear.
Sara van de Geer and Johannes Lederer. The Lasso, correlated design,
and improved oracle inequalities. In From probability to statistics and
back: high-dimensional models and processes, volume 9 of Inst. Math. Stat.
(IMS) Collect., pages 303–316. Inst. Math. Statist., Beachwood, OH, 2013.
. URL http://dx.doi.org/10.1214/12-IMSCOLL922.
| 10math.ST
|
Submitted to the Annals of Statistics
arXiv: arXiv:1703.06222
A UNIFIED TREATMENT OF MULTIPLE TESTING
WITH PRIOR KNOWLEDGE USING THE P-FILTER
arXiv:1703.06222v3 [stat.ME] 12 Sep 2017
By Aaditya Ramdas∗ , Rina F. Barber† ,
Martin J. Wainwright∗ , and Michael I. Jordan∗
University of California, Berkeley∗ and University of Chicago†
A significant literature studies ways of employing prior knowledge to improve power and precision of multiple testing procedures.
Some common forms of prior knowledge may include (a) a priori beliefs about which hypotheses are null, modeled by non-uniform prior
weights; (b) differing importances of hypotheses, modeled by differing penalties for false discoveries; (c) multiple arbitrary partitions
of the hypotheses into known (possibly overlapping) groups, indicating (dis)similarity of hypotheses; and (d) knowledge of independence,
positive or arbitrary dependence between hypotheses or groups, allowing for more aggressive or conservative procedures. We present a
unified algorithmic framework called p-filter for global null testing
and false discovery rate (FDR) control that allows the scientist to
incorporate all four types of prior knowledge (a)–(d) simultaneously,
recovering a wide variety of common algorithms as special cases.
1. Introduction. Multiple hypothesis testing is both a classical and
highly active research area: it dates back (at least) to an initially unpublished 1953 manuscript by Tukey entitled “the problem of multiple comparisons” [40]. Given a large set of null hypotheses, multiple testing commonly
deals with deciding which subset to reject, while guaranteeing some notion
of control on the number of false rejections. It is of practical importance to
incorporate different forms of prior knowledge into existing multiple testing procedures; such prior knowledge can yield improvements in power and
precision, and can also provide more interpretable results. The focus of this
paper is on methods that control the False Discovery Rate (FDR) or test
the global null (GN) hypothesis while incorporating any number of the following considerations: (a) using prior weights, (b) using penalty weights,
(c) partitioning the hypotheses into groups, (d) incorporating knowledge of
the dependence structure within the data, including options such as estimating and adapting to the unknown number of nulls under independence,
or reshaping rejection thresholds to preserve error control guarantees in the
MSC 2010 subject classifications: Primary 62J15, 60G10; secondary 62F03
Keywords and phrases: multiple testing, false discovery rate, prior knowledge, Simes,
Benjamini-Hochberg-Yekutieli, adaptivity, group FDR
1
2
RAMDAS, BARBER, WAINWRIGHT AND JORDAN
presence of arbitrary dependence. It is a challenge to incorporate all of these
forms of structure while maintaining internal consistency (coherence and
consonance) among the pattern of rejections and acceptances, and most existing work has managed to simultaneously employ only one or two of the
four considerations (a), (b), (c), (d). We present a general unified framework
for integrating all four, while performing a GN test or controlling FDR. This
framework allows scientists to mix and match techniques, and use multiple
different forms of prior knowledge simultaneously. Our framework simplifies the analysis of existing procedures, and generalize the conditions under
which they are known to work. We now outline our contributions.
Outline. Rather than presenting the most general version of our framework
at the outset, we introduce the framework gradually, beginning with simple
weighting schemes and then introducing more tools and notation as needed
to encompass more complex schemes. Since all types of prior information
considered in this paper have already been motivated in applied settings,
our focus is on the conceptual and mathematical aspects.
• Assumptions. In Section 2, we formalize the range of settings studied,
in terms of the assumptions on the underlying p-values that provide the
data for our multiple testing procedures—marginally, the p-values are
assumed to be super-uniform, and jointly, we consider the three settings
of (a) independence, (b) positive dependence, or (c) arbitrary dependence.
• Generalized BHY procedures. We discuss the use of reshaping functions, as introduced by Blanchard and Roquain [9], to guard against arbitrary dependence. We use these to describe a set of procedures that
generalize the classical Benjamini-Hochberg (BH) procedure [3] and the
Benjamini-Yekutieli (BY) procedure [7]; we thus refer to them as “generalized BHY procedures.” We develop a general super-uniformity lemma
that is at the core of FDR control proofs. As a first application, we show
that it yields a short proof of FDR control for generalized BHY procedures
under all three settings of dependence.
• Prior- and/or penalty-weighted BHY. In Section 4, we discuss
two natural ways to incorporate weights—prior-weighted p-values and
penalty-weighted FDR—and we describe a procedure that uses both prior
and penalty weights simultaneously. Using the super-uniformity lemma,
we show that this procedure controls (weighted) FDR under all three
settings of dependence. This immediately implies old and new results for
algorithms studied by Genovese et al. [16], Benjamini and Hochberg [4],
and Blanchard and Roquain [9].
• Adaptive, prior and/or penalty-weighted BH. In Section 5, we
present so-called “adaptive” procedures for independent p-values, which
UNIFIED MULTIPLE TESTING WITH PRIOR KNOWLEDGE
3
improve power by incorporating an estimate of the proportion of nulls, as
first proposed by Storey et al. [37, 38]. We then propose a novel adaptive
procedure, that can employ both prior and penalty weights, as well as
null proportion estimates. Using a new inverse-binomial lemma, we succinctly prove that our adaptive, doubly-weighted BH procedure controls
(weighted) FDR under independence.
• Grouped, adaptive, doubly-weighted procedures. In Section 6, we
tackle the problem of controlling the group-level FDR for a single userdefined partition of the hypotheses, when the group-level p-values may
or may not be formed from the elementwise p-values. We particularly
discuss one way to achieve this based on the Simes p-value, first showing
that the prior-weighted Simes test studied by Hochberg and Liberman
[19] is valid even under positive dependence, and can be reshaped to work
under arbitrary dependence. Then, using a novel group super-uniformity
lemma, we demonstrate how to use our novel weighted Simes tests in
conjunction with doubly-weighted adaptive BH procedures to control the
(weighted) group FDR.
• A unifying multilayer, grouped, adaptive, doubly-weighted procedure. In Section 7, we generalize the previous section to allow the use
of multiple arbitrary user-defined partitions of hypotheses that are possibly non-hierarchically arranged relative to each other. In the process,
we develop an algorithmic framework called p-filter, extending earlier
work of Barber and Ramdas [1]. p-filter allows the use of any or all of
the following features and extensions: different prior and penalty weights
for every group in every partition; null-proportion adaptivity (for all partitions whose groups are known to be independent); incomplete partitions that contain a “leftover set”; overlapping groups within a partition;
and reshaping functions to guard against arbitrary dependence. This algorithmic framework guarantees internal consistency, meaning that the
rejected hypotheses and groups are always in mutual agreement. We combine the super-uniformity lemma with novel proof techniques to prove that
p-filter guarantees (weighted) group FDR control for all partitions simulataneously, under all three forms of dependence.
Sometimes, we reprove existing results in the literature, usually for the
purposes of illustrating a novel proof technique; in such cases, we label our
results as propositions. In other cases, when results are new and not found
elsewhere in the literature, we label them as theorems. Table 1 summarizes
our contributions relative to some related work. Given space limitations,
we only list a few related references, and defer more detailed discussions of
related work to the point in the paper when those settings are discussed.
4
Form of incorporated structure
prior work
this paper
Benjamini and Hochberg [3]
Benjamini and Yekutieli [7]
Proposition 1
prior weights (FDR)
Genovese et al. [16]
Proposition 2
penalty weights (FDR)
Benjamini and Hochberg [4]
Finos and Salmaso [13]
Proposition 2
prior and/or penalty weights (FDR)
Blanchard and Roquain [9]
null proportion adaptivity (FDR)
prior and/or penalty weighted null proportion adaptivity (FDR)
Storey et al. [38]
—
Proposition 2
Proposition 3
Theorem 5.1
none (Global Null)
Simes [33]
Benjamini and Yekutieli [7]
Proposition 4
prior weights (Global Null)
Hochberg and Liberman [19]
Proposition 4
single partition into groups (FDR)
Barber and Ramdas [1]
Proposition 5
multiple arbitrary partitions (FDR)
multiple possibly incomplete arbitrary partitions
. of possibly overlapping groups + prior and/or penalty
. weights + group null proportion adaptivity (FDR)
Barber and Ramdas [1]
—
Theorem 7.1
Theorem 7.1
Table 1
This table summarizes our contributions relative to other known results regarding step-up procedures for multiple testing The traffic
lights
,
and
represent results that were proved under independence (easy), positive dependence (intermediate), and arbitrary
dependence (hardest). In the top box,
indicates that under uniformity, the FDR bound was proved with inequality instead of strict
equality. In the bottom box,
indicates that null-proportion adaptivity was not provided, and also the group-level p-values were
constrained to be Simes p-values.
RAMDAS, BARBER, WAINWRIGHT AND JORDAN
none (False Discovery Rate)
UNIFIED MULTIPLE TESTING WITH PRIOR KNOWLEDGE
5
2. Background. We begin by presenting background on multiple testing and FDR control, overviewing common assumptions in the literature,
and providing a discussion of the Benjamini-Hochberg procedure. Before
doing so, it is helpful to introduce some basic definitions and notation. For
a pair of vectors x, y ∈ [0, 1]n , we use the notation x y to mean that x ≤ y
in the orthant ordering, i.e., xi ≤ yi for all i ∈ {1, . . . , n}.
Definition 1 (Nondecreasing sets and functions). A set D ⊆ [0, 1]n is
said to be nondecreasing if x ∈ D implies y ∈ D for all y x. We say that a
function f : [0, 1]n 7→ [0, ∞) is nonincreasing, if x y implies f (x) ≥ f (y).
Finally, in order to handle the ratio “ 00 ” that often arises in FDR control
results, we adopt the “dotfraction” notation
a
if a 6= 0, b 6= 0,
b,
0,
if a = b = 0,
a
(2.1)
··· : =
b
0,
if a = 0, b 6= 0,
undefined
if a 6= 0, b = 0.
Dotfractions behave like fractions whenever the denominator is nonzero. We
formally derive properties of dotfractions in detail in Appendix G.
2.1. False Discovery Rate (FDR). In the standard multiple testing setup,
we are given a set of n different p-values, denoted by the random vector
P ∈ [0, 1]n . Each p-value corresponds to a different null hypothesis, and we
let H0 ⊆ [n] denote the subset of true null hypotheses. Our goal is to reject
some subset of the null hypotheses—or in other words, to select some subset
of {1, . . . , n} as our discoveries—while at the same retaining control over
the number of false discoveries. Consider any algorithm that, based on the
observed vector P of p-values, chooses a subset of hypotheses to reject; this
b ). In their seminal work, Benjamini
random subset is denoted by Sb = S(P
and Hochberg [3] proposed to measure performance of such algorithms by a
quantity called the False Discovery Rate (FDR), and gave a simple procedure to provably control it.1
The FDR is defined in terms of the False Discovery Proportion, given by
b
|H0 ∩S|
b is number of false discoveries—that null
FDP : = ············,
where |H0 ∩ S|
b
|S|
b is the
hypotheses that are true, and are incorrectly rejected—whereas |S|
1
As a historical footnote, the same procedure was proposed by Eklund [11], albeit as a
heuristic without any formal guarantees [12, 32].
6
RAMDAS, BARBER, WAINWRIGHT AND JORDAN
total number of discoveries, meaning null hypotheses that are rejected. The
FDR is the expectation of this random ratio:
n
o
P
"
#
b
b
|H0 ∩ S|
i∈H0 1 i ∈ S
o .
FDR : = E ················· = E ··································
(2.2)
P n
b
|S|
b
1
i
∈
S
i
Our goal is to construct procedures that control FDR at some target level
α ∈ (0, 1), that is to guarantee that FDR ≤ α.
2.2. P-value dependence assumptions. We assume that the marginal distribution of each null p-value is super-uniform, meaning that it is stochastically dominated by the uniform distribution. More precisely, for any index
i ∈ H0 , we assume that
(2.3)
Pr{Pi ≤ t} ≤ t
for all t ∈ [0, 1].
Of course, uniformly distributed p-values trivially satisfy this condition. We
use the phrase under uniformity to describe the situation in which the null
p-values are marginally exactly uniform. If this phrase is not employed, it is
understood that the null p-values are marginally super-uniform.
Regarding assumptions on the joint distribution of p-values, three possible
kinds of dependence will be considered in this paper: independence, positive
dependence or arbitrary dependence. In the independent setting, all p-values
are assumed to be mutually independent. The second possibility is that of
“positive dependence” as formalized by Positive Regression Dependence on
a Subset condition, or PRDS for short:
Definition 2 (PRDS). We say that the vector P satisfies PRDS if
for any null index i ∈ H0 and nondecreasing set D ⊆ [0, 1]n , the function
t 7→ Pr{P ∈ D | Pi ≤ t} is nondecreasing over t ∈ (0, 1].
The original positive regression dependence assumption as introduced
by Lehmann [24] as well as the PRDS assumption first made by Benjamini
and Yekutieli [7] both had Pi = t instead of Pi ≤ t in the definition, but one
can prove that both conditions are essentially equivalent.
The PRDS condition holds trivially if the p-values are independent, but
also allows for some amount of “positive” dependence. Let us consider a
simple example to provide some intuition. Let Z = (Z1 , . . . , Zn ) be a multivariate Gaussian vector with covariance matrix Σ; the null components
correspond to Gaussian variables with zero mean. Letting Φ be the CDF
of a standard Gaussian, the vector P = (Φ(Z1 ), . . . , Φ(Zn )) is PRDS on
UNIFIED MULTIPLE TESTING WITH PRIOR KNOWLEDGE
7
Pi for every index i if and only if all entries of the covariance matrix Σ
are non-negative. See Benjamini and Yekutieli [7] for additional examples of
this type. It should be noted the PRDS assumption is closely related to the
assumption of multivariate total positivity of order 2 (MTP2), as studied
by Karlin and Rinott [22]. Since MTP2 implies PRDS, all the results in this
paper that hold under PRDS also immediately also hold under MTP2.
3. Generalized Benjamini-Hochberg-Yekutieli procedures. We
begin by discussing the classical Benjamini-Hochberg procedure. Given a
vector P ∈ [0, 1]n , we let P(1) ≤ P(2) ≤ · · · ≤ P(n) denote the associated vector of order statistics. Given these order statistics, the Benjamini-Hochberg
procedure (BH) [3] with target FDR level α rejects the smallest b
k p-values,
where
α·k
b
b
k = kα (P ) : = max k : P(k) ≤
(3.1)
,
n
with that convention that b
k = 0 if the set
is, after choosing
n is empty. That
o
α·b
k
b
b
this value k, the set of rejections is S = i : Pi ≤ n , which by definition
of b
k, must contain exactly b
k many hypotheses. Benjamini and Hochberg [3]
prove that this procedure controls the FDR at the desired level, that is,
b under the assumption that the p-values
FDR ≤ α for this rejection set S,
are mutually independent. This result was shown to hold also in the positive
dependence setting by Benjamini and Yekutieli [7].
When no assumptions are made about the joint distribution of p-values,
however, it is conceivable that FDR-controlling procedures must be guarded
while proclaiming a discovery. Benjamini and Yekutieli [7] originally proposed running the BH procedure at the modified threshold Pnα 1 in order
i=1 i
to maintain FDR control under arbitrary dependence. We now introduce a
family of generalizations of this modification, which we will refer to as “generalized BY procedures” or simply “BY procedures” throughout the paper.
In order to do so, we first need the notion of a reshaping function β, as
introduced by Blanchard and Roquain [9].
Definition 3 (Reshaping). For any fixed probability measure ν on
[0, ∞), we define the reshaping function β = βν as
Z
βν (k) =
k
x dν(x).
0
8
RAMDAS, BARBER, WAINWRIGHT AND JORDAN
Fixing a reshaping function β, then, the generalized Benjamini-Yekutieli
(BY) procedure with target FDR level α makes
α · β(k)
b
b
(3.2)
k = kα (P ) : = max k : P(k) ≤
n
b
rejections, again with that convention
n that k = b0 oif the set is empty. The
k)
rejection set is now given by Sb = i : Pi ≤ α·β(
, which again contains
n
exactly b
k many hypotheses. Noting that β(k) = βν (k) ≤ k for any probability measure ν, thus requiring that the k lowest p-values fall below a lower
(more strict) threshold, we can immediately see that for any choice of β,
this generalized BY procedure is more conservative than the BH procedure
(if run at the same target FDR level α).
As an example, we can see that Benjamini and Yekutieli [7]’s original
proposal, i.e. running the BH procedure at the modified threshold Pnα 1 ,
i=1 i
is equivalent to running the BY procedure with the reshaping function
(3.3)
k
βBY (k) = Pn
1.
i=1 i
We now summarize some known properties of the BH and BY procedures.
Proposition 1. The BH procedure (3.1) and generalized BY procedures (3.2) have the following properties:
0
(a) Under independence and uniformity, the BH method has FDR = α |Hn | .
0
(b) Under positive dependence, the BH method has FDR ≤ α |Hn | .
(c) Under arbitrary dependence, the generalized BY procedure with any
0
reshaping function β yields FDR ≤ α |Hn | .
Statement (a) was proven with inequality by Benjamini and Hochberg
[3]. Benjamini and Yekutieli [7] proved statement (a) with equality, as well
as proving statement (b), and statement (c) for the specific reshaping function (3.3). For an arbitrary reshaping function β, statement (c) was proven
by Blanchard and Roquain [9].
3.1. The role of reshaping. Naturally, there is no universally optimal
choice of ν, and different choices of base probability measure ν (and hence
β) lead to different behaviors of the associated procedures. A typical choice
of the measure ν is a discrete distribution with support on {1, . . . , n}; such a
choice is sensible since the number of rejections is always an integer. Indeed,
UNIFIED MULTIPLE TESTING WITH PRIOR KNOWLEDGE
9
as noted by Blanchard and Roquain [9], any continuous measure ν can be
replaced by the discrete measure
(
ν((r − 1, r])
if r < n
0
ν (r) =
ν((n − 1, ∞)) if r = n
to yield a reshaping function βν 0 that is strictly larger than βν on the relevant
integer range, leading to a more powerful procedure.
In contrast with the discrete distributions considered in past work, this
paper also considers the case of continuous measures ν. Continuous measures are of interest in the context of the penalty-weighted procedures to
be considered in the sequel, where the function β corresponds to the total
weight assigned to subsets of the hypotheses. As opposed to the count, this
total weight can be fractional, so that one can no longer reduce to the case
of discrete measures ν, as we did above.
Many different examples of reshaping functions, and their connections to
other formulations of multiple testing methods, can be found in the literature, see e.g. Blanchard and Roquain [9], Sarkar [30, 31].
3.2. Alternative derivation of BH and BY procedures. There is an alternate way to motivate the BH and BY procedures, which will be useful for
designing algorithms later in this paper. Let FDP(t) denote the FDP of a
procedure that rejects all p-values less than or equal to t, and note that
P
b
|H0 ∩ S|
i∈H0 1 {Pi ≤ t}
P
(3.4) FDP(t) = ················· = ··································
b
|S|
i 1 {Pi ≤ t}
|H0 | · t
n·t
[
P
P
≈ ···························
≤ ···························
=: FDP(t),
i 1 {Pi ≤ t}
i 1 {Pi ≤ t}
P
where the approximation
{Pi ≤ t} ≈ |H0 | · t follows from the obi∈H0 1
P
0
servation that E
i∈H0 1 {Pi ≤ t} ≤ |H | · t—that is, we expect roughly
0
|H | · t many of the null p-values to be ≤ t. In order to guard against unex[
pected dependence structures, we would reshape thePdenominator of FDP(t)
so as to undercount rejections, replacing it by β ( i 1 {Pi ≤ t}). We then
choose the largest threshold t such that this estimated FDP is at most α.
Concretely, we define b
t=b
tα (P ) as
(3.5)
[
b
t : = max {t : FDP(t)
≤ α},
t∈[0,1]
10
RAMDAS, BARBER, WAINWRIGHT AND JORDAN
and we reject all hypotheses corresponding to p-values smaller than b
t. It is
α·β(b
k)
α·b
k
b
b
not hard to verify that t ≥ n (for BH) or t ≥ n (for BY), and that the
threshold rule (3.5) corresponds to rejecting the same set of hypotheses as
the original rule (3.1) or (3.2) (for the BH or BY procedures, respectively).
3.3. A super-uniformity lemma for FDR control. For a set x of n elements, we let x−i := {x1 , . . . , xi−1 , xi+1 , . . . , xn } denote the set of n − 1
elements that leaves out the i-th element. For a vector x ∈ Rn , we use
x
e−i := (x1 , . . . , xi−1 , 0, xi+1 , . . . , xn ) ∈ Rn to denote a vector with the
i-th coordinate set to zero. Of course, x
e−i and x−i contain the same information and can be reconstructed from each other.
Definition 4 (LOOP). A function f : [0, 1]n → [0, ∞) is said to satisfy
the leave-one-out property (LOOP) if for any null index i ∈ H0 and any
x ∈ [0, 1]n , we have
if xi ≤ f (x), then f (e
x−i ) = f (x),
−i
(3.6)
f (e
x ) > 0 and
−i
if xi > f (x), then xi > f (e
x ).
We will refer to LOOP in situations where x is the vector of p-values
P , and f (P ) represents a threshold. When f satisfies LOOP, even though
threshold f (Pe−i ) may differ significantly from f (P ), the p-value Pi will
either lie below both thresholds, or above both thresholds—in other words,
from the perspective of Pi , the threshold might as well have been f (Pe−i )
instead of f (P ).
To develop some intuition for the lemma that follows, we note that our
super-uniformity assumption (2.3) on null p-values can be reformulated as:
1 {Pi ≤ t}
0
(3.7)
For any i ∈ H , E ···················· ≤ 1 for any fixed t ∈ [0, 1].
t
Of course, if Pi is uniform then the above inequality holds with equality.
The following lemma guarantees that property (3.7) continues to hold for
certain random thresholds f (P ). Recall that the term “nonincreasing” is
interpreted coordinatewise, with respect to the orthant ordering (Def. 1).
Lemma 1 (Super-uniformity lemma).
Let i ∈ H0 be a null hypothesis.
(a) For any nonincreasing function f : [0, 1]n → [0, ∞), if Pi is independent of P −i , then we have
1 {Pi ≤ f (P )}
−i
≤ 1.
E ···························· P
f (P )
UNIFIED MULTIPLE TESTING WITH PRIOR KNOWLEDGE
11
Furthermore, if we additionally assume that f has range [0, 1] and
satisfies LOOP (3.6), and that Pi is uniformly distributed, then the
inequality is replaced with equality:
1 {Pi ≤ f (P )}
−i
E ···························· P
= 1.
f (P )
(b) For any nonincreasing function f : [0, 1]n → [0, ∞), if P is PRDS with
respect to Pi , then
1 {Pi ≤ f (P )}
E ···························· ≤ 1.
f (P )
(c) For any constant c ≥ 0, any function f : [0, 1]n → [0, ∞), and any
reshaping function β, under arbitrary dependence of the p-values,
1 {Pi ≤ c · β(f (P ))}
E ········································ ≤ 1.
c · f (P )
(d) For any constant c ≥ 0, any functions f1 , . . . , fm : [0, 1]n → [0, ∞),
and any reshaping functions β1 , . . . , βm , under arbitrary dependence
of the p-values,
Qm
1 {Pi ≤ c · `=1 β` (f` (P ))}
Qm
E ·····················································
≤ 1.
c·
`=1
f` (P )
The proofs of statement (a) with equality, and of statement (d), are given
in Appendix A. Statement (a) with inequality is recovered as a special case
of statement (b), which was proved by Blanchard and Roquain [9], who also
proved (c). The more general statement (d), with more than one reshaping
function present in the bound, will be required in the proof of a novel group
super-uniformity Lemma 3.
This super-uniformity lemma is central to succinct proofs of the results of
this paper. It can also be extended to allow for multiple reshaping functions
to be used simultaneously—see Corollary ?? below. To give a first hint of
the power of this lemma, we begin by providing a short proof that the BH
procedure controls FDR.
Proof of Proposition 1. We prove statement (b) first. By the definition of FDR, we have
o
o
n
n
P
α·b
k
α·b
k
0
1
P
≤
X α
(i)
i
n
i∈H0 1 Pi ≤ n
≤ α |H | ,
· E ··························
FDR = E ········································· =
α·b
k
n
n
b
k
i∈H0
n
12
RAMDAS, BARBER, WAINWRIGHT AND JORDAN
)
where step (i) follows by applying Lemma 1(b) with f (P ) = α·k(P
n . Turning
to statement (c), we repeat the same argument, but now incorporating the
reshaping function β:
o
n
o
P
n
α·β(b
k)
α·β(b
k)
0
1
P
≤
1
P
≤
X
(ii)
i
0
i
α
n
i∈H
n
≤ α |H | ,
FDR = E ············································· =
·E ·······························
α·b
k
n
n
b
k
0
b
i∈H
n
where step (ii) follows from Lemma 1(c) with c = α/n, f1 (P ) = b
k(P ) and
β1 = β.
In order to prove statement (a) with equality by applying Lemma 1(a),
it is straightforward to see that f (P ) satisfies the LOOP condition (3.6)
(for completeness, this statement is proved in a more general setting in
Proposition 2).
We note that leave-one-out style arguments, where FDR control over the
entire set of rejected hypotheses Sb is proved by considering each Pi individually, are not new (see, for example, Sarkar [30]); however, most past works
have either considered specific functions f , or have proved an inequality in
their version of statement (a). Our proof is a generalization of arguments
made in the special case of the BH procedure by Heesen and Janssen [18].
4. Using both prior and penalty weights. We now consider a setting where there are weights associated with each hypothesis. Prior work has
focused on two natural interpretations for such weights—as priors on the nullity of hypotheses, or as penalties for wrongly rejecting nulls. Accordingly,
we introduce a sequence of positive penalty weights {ui }ni=1 and
P positive
priorP
weights {wi }ni=1 , along with the normalization conditions ni=1 ui = n
and ni=1 wi = n. These two weight sequences can be interpreted as follows:
(U). The penalty weights ui encode unequal importances of hypotheses:
hypotheses with ui > 1 indicate that they are of more interest to
the scientist than hypotheses with ui < 1, and hence we count these
differently when measuring number of (false or total) discoveries.
(W). The prior weights wi encode prior evidence against the null: a large
wi > 1 implies that a prior belief that the i-th hypothesis is more likely
non-null, and a small wi < 1 indicates the opposite.
Such weights might be learned from data collected in an older study on
different subjects. We treat the weights as constants, but it is straightforward
to extend our results to a Bayesian setting in which both H0 and the weights
are random, but the null p-values are distributed (super)uniformly even after
conditioning on these weights.
UNIFIED MULTIPLE TESTING WITH PRIOR KNOWLEDGE
13
In order to model the setting (U), we define a modified criterion, following
Benjamini and Hochberg [4], called the weighted FDR :
n
o
P
b
i∈H0 ui 1 i ∈ S
n
o .
FDRu : = E ······································
(4.1)
P
b
u
1
i
∈
S
i i
In order to model setting (W), on the other hand, one does not need to alter
the definition of the FDR, since these weights simply allow us to re-prioritize
which hypotheses to reject, but does not alter our measure of error. We now
present a doubly-weighted version of the BH and BY procedures, that can
incorporate both sets of weights simultaneously.
4.1. Doubly-weighted BH and BY procedures. Following Blanchard and
Roquain [9], we define the prior+penalty weighted BH procedure (BHuw ) as
follows. We set Qi : = Pi /wi , and then reject all hypotheses having Qi ≤ b
t,
where the threshold b
t=b
tα,u,w (P ) is given by
n
o
t·n
[
[
P
(4.2a) b
t := max t : FDP(t)
≤ α , where FDP(t)
:= ······························.
t∈[0,1]
i ui 1(Qi ≤ t)
It is useful to rewrite this procedure as follows. Let u(i) represent the
weights corresponding to the order statistics Q(i) of the weighted p-values
P
Q1 , . . . , Qn , and define the sum of the first k weights to be U(k) : = ki=1 u(i) .
Then define
α · U(k)
b
b
k = kα,u,w (P ) : = max k : Q(k) ≤
(4.2b)
n
b
and reject those
n hypotheses
ocorresponding to the k smallest weighted pα·U(k)
b
values, Sb = i : Qi ≤
. Then, the BHuw procedure can be equivan
lently described by rejecting the first b
k p-values, corresponding to a total
penalty weight of U(bk) . If we wish to use a reshaping function β to guard
against arbitrary dependence among the p-values, then we can instead use a
doubly weighted BY procedure, denoted as BYuw , where we instead define
n
o
t·n
[
[
P
(4.3a) b
t := max t : FDP(t)
≤ α where FDP(t)
:= ······································,
β
t∈[0,1]
i ui 1(Qi ≤ t)
which corresponds to
(4.3b)
α · β(U(k) )
b
.
k=b
kα,u,w (P ) : = max k : Q(k) ≤
n
14
RAMDAS, BARBER, WAINWRIGHT AND JORDAN
(Implicitly, the choice of reshaping function β is considered to be fixed, and
is suppressed in our notation for the procedure.)
With this setup, we have the following guarantees on the weighted FDR (4.1):
Proposition 2. The doubly-weighted BH and BY procedures, given in (4.2b)
and (4.3b), have the following properties:
(a) Under independence and uniformity,
if maxi wi ≤ 1/α, the BHuw proα P
cedure yields FDRu = n
ui wi .
i∈H0
P
(b) Under positive dependence, the BHuw procedure yields FDRu ≤ αn
ui wi .
i∈H0
(c) Under arbitrary dependence, the
BYuw procedure with any reshaping
α P
function β yields FDRu ≤ n
ui wi .
i∈H0
Before we present the proof, let us discuss some special cases.
• When both sets of weights are used, Blanchard and Roquain [9] proved
Proposition 2(b,c) and also Proposition 2(a) with inequality. Our proof
of Proposition 2(b,c) mimics theirs, and is presented in our notation for
completeness.
• When only penalty weights are used, and p-values are independent, one
may set wi = 1 for all i in the BHuw procedure (4.2b) to recover the
penalty-weighted BH procedure proposed by Benjamini and Hochberg [4]
who proved Proposition 2(a), albeit with an inequality.
• When only prior weights are used, Genovese et al. [16] proposed a priorweighted BH procedure that controls the unweighted FDR under independence. Their result can be recovered by setting ui = 1 for all i in the
BHuw procedure (4.2b), and invoking Proposition 2(a).
• When all the weights equal one, the BHuw and BYuw reduce to the
unweighted BH and BY procedures, and Proposition 2 then reduces to
Proposition 1.
Proof of Proposition 2. We first prove statements (b) and (c). For
conciseness, to unite the proofs of the two statements, define a function γ
to be the identity if we are in the setting of (b), or γ = β (the reshaping
function) if we are in the setting of (c). Note that we can write the weighted
UNIFIED MULTIPLE TESTING WITH PRIOR KNOWLEDGE
FDR from equation (4.1) as
15
α·γ(U(k)
b )
u
1
Q
≤
i
n
i∈H0 i
FDRu = E ························································
U(bk)
α·γ(U(k)
b )
X
n
1 Qi ≤
=
ui · E ····································· .
U(bk)
i∈H0
P
Recalling that Qi = Pi /wi by definition, we multiply and divide by αwi /n
so as to obtain
P
α·γ(U(k)
b )
ui wi
1
P
≤
w
i
i
X αui wi
n
(i)
i∈H0
(4.4) FDRu =
· E ········································· ≤ α
.
α·U b
n
n
wi n(k)
i∈H0
Here inequality (i) follows by the observation that the function P 7→ U(bk(P ))
is nonincreasing in P , and then for each null hypothesis i ∈ H0 , we can
apply Lemma 1(b) with f (P ) = wi αU(bk(P )) /n, or Lemma 1(c) with f1 (P ) =
U(bk(P )) , β1 = β, and c = wi α/n.
Next we turn to the proof of statement (a), for which we begin by noting
that if maxi wi ≤ 1/α, then the range of f is [0, 1]. Next, we show that
α·U
f (P ) := wi n(k) satisfies the LOOP condition (3.6) with respect to index i;
if this holds, then under independence and uniformity, inequality (i) holds
with equality by Lemma 1(a). To prove LOOP, first, it is easy to observe that
f (Pe−i ) > 0 since b
k(Pe−i ) > 0 and ui , wi are strictly positive. Next, it is also
straightforward that if hypothesis i was already rejected by the procedure,
then setting Pi to zero does not change b
k. In order to apply Lemma 1, the
b
last condition we need to verify is that assuming Qi > α ·
U(k(
b P
e −i ))
n
U(k(P
b ))
n
holds (that
is, if Qi is not rejected), then Qi > α ·
also holds. Indeed, if Qi has
rank a within the order statistics of Q, then we are guaranteed that
U(a)
U(b)
and Q(b) > α
for all b > a,
n
n
since Qi is not rejected. Then, on setting Pi to 0, the number of rejections b
k(Pe−i ) must be bounded by a, since the other p-values are unchanged
U
and Q(b) > α n(b) for all b > a continues to hold by condition (4.5). Then
U(bk(Pe−i )) ≤ U(n) , and so we may then infer from condition (4.5) that Qi >
(4.5)
α
U(k(
b P
e −i ))
n
Qi = Q(a) > α
will also hold since the right-hand side is at most α
U(a)
n .
16
RAMDAS, BARBER, WAINWRIGHT AND JORDAN
5. Null proportion adaptivity with prior and/or penalty weights.
Recall that the original unweighted BH procedure provided FDR control at
level α|H0 |/n, which can be significantly smaller than the target α when
|H0 | n, leading to a possible loss in power compared to a procedure that
fully utilized its FDR budget of α. When the p-values are known to be independent, Storey and collaborators [37, 38] proposed a simple method to
estimate the number of nulls. Incorporating such an estimate into the BH
procedure leads to a procedure that has possibly higher power than BH.
We refer to this method as Storey’s adaptive BH procedure (St-BH). For
the rest of this section, we assume that the p-values are independent, and
we demonstrate how to combine the benefits of adaptivity with prior and
penalty weighting by analyzing a novel doubly-weighted St-BH procedure.
5.1. Storey’s adaptive BH method. Fix a user-defined constant λ ∈ [α, 1),
and define:
P
1 + j 1 {Pj > λ}
π
b0 = π
b0 (P ) : =
(5.1)
.
n(1 − λ)
The quantity π
b0 is a conservative estimate of the proportion of nulls, since
0
for any λ, we have E [b
π0 ] ≥ |Hn | . In analogy with the BH procedure (3.4),
the St-BH procedure chooses
π
b nt
[
[
P 0
(5.2) b
tα (P ) : = max {t : FDP(t)
≤ α}, where FDP(t)
: = ···························,
t∈[0,1]
i 1 {Pi ≤ t}
and then rejects all p-values smaller than b
tα (P ). This procedure comes with
the following guarantee.
Proposition 3. Under independence, the St-BH procedure (5.2) guarantees that FDR ≤ α.
Storey and coauthors [37, 38] establish the above result using martingale
arguments, and Blanchard and Roquain [10] gave an alternate proof using
Lemma 1. We do not prove this proposition here, since it follows as a special
case of Theorem 5.1 below. To proceed with a discussion of this result, we
need the following notation. Define the leave-one-out estimate of the nullproportion as
P
1 + j6=i 1 {Pj > λ}
−i
−i
π
b0 = π
b0 (Pe ) : =
,
n(1 − λ)
UNIFIED MULTIPLE TESTING WITH PRIOR KNOWLEDGE
17
which is obtained if the i-th p-value is set to zero, and then π
b is then estimated as in (5.1).
For future reference, we note that using simple properties of the binomials,
it is easy to show that the estimate π
b0−i satisfies the bound
1
1
n
(5.3)
E −i = E −i ≤
.
0|
|H
π
b0
π
b0
It turns out that the bound (5.3) is the only property of the estimator π
b0
that we require to prove FDR control. (In fact, the bound (5.3) is satisfied
by several other estimators of π0 as well—for examples, see Sarkar [30].
Blanchard and Roquain [10] propose and empirically compare several such
estimates of π0 , to find that Storey’s method had the best power amongst
all considered alternatives.)
5.2. A doubly-weighted adaptive BH procedure. In this section, we demonstrate how to simultaneously incorporate both prior weights w and penalty
weights u into Storey’s adaptive estimator. We introduce the shorthand
|u · w|∞ : = maxj uj wj , and define the St-BHuw procedure as follows:
1. Define the estimated null proportion as
P
|u · w|∞ + j uj wj 1 {Pj > λ}
π
b0 : =
n(1 − λ)
2. Estimate the false discovery proportion by
π
b0 · n · t
[
P
FDP(t)
: = ························································
u
1
{P
i ≤ min{wi · t, λ}}
i i
3. Choose the threshold for rejection as
(5.4)
[
b
t=b
tα,u,w (P ) : = max{t ∈ [0, λ] : FDP(t)
≤ α};
4. Reject all hypotheses such that Pi ≤ min{wi b
t, λ}.
In words, we apply the prior+penalty weighted BHuw procedure, while
also incorporating a weighted null proportion estimator.
The St-BHuw procedure has the following guarantee:
Theorem 5.1. Under independence, the St-BHuw procedure (5.4) controls FDRu at level α.
18
RAMDAS, BARBER, WAINWRIGHT AND JORDAN
Notice that if we set the weights to unity, we recover exactly the St-BH
procedure, and Theorem 5.1 reduces to Proposition 3. Also, if we set π
b0 =
λ = 1, we recover exactly the BHuw procedure. To prepare for the proof of
this theorem, we define the analogous leave-one-out estimate of the weighted
null proportion as
P
|u · w|∞ + j6=i uj wj 1 {Pj > λ}
−i
π
b0 : =
.
n(1 − λ)
As before, π
b0−i is a function of only P −i , and this observation is crucial in
the proof of Theorem 5.1. The other important property required by the
proof is the following inequality:
n
1
(5.5)
.
E −i ≤ P
π
b0
j∈H0 uj wj
To establish this, we need a more sophisticated result about weighted binomials, which we now present. (The bound (5.3) can be recovered as a special
case of this lemma, by setting all weights to one.)
Lemma 2 (Inverse binomial lemma).
Given a vector a ∈ [0, 1]m , constant
i.i.d.
b ∈ [0, 1], P
and Bernoulli variables Zi ∼ Bernoulli(b), the weighted sum
Z := 1 + m
i=1 ai Zi satisfies
1
1
1
Pm
P
(5.6)
≤E
≤
1 + b i=1 ai
Z
b(1 + m
i=1 ai )
P
Since E [Z] = 1 + b m
i=1 ai , the lower bound on E [1/Z] follows by Jensen’s
inequality. We include this bound to provide context for the upper bound
on E [1/Z]. The detailed proof can be found in Appendix B.
In order to see that required property (5.5) follows from Lemma 2, define
X
uj wj
Z := 1 +
aj 1 {Pj > λ} with aj =
, b = (1 − λ), m = |H0 | − 1.
|u
·
w|
∞
0
j∈H ,j6=i
Since Z ≤
n(1−λ) −i
b0 ,
|u·w|∞ π
applying Lemma 2 guarantees that
|u · w|∞
1
|u · w|∞
P
E
≤E
≤
.
−i
Z
(1
−
λ)(|u
·
w|
n(1 − λ)b
π0
∞+
j∈H0 ,j6=i uj wj )
Some simple algebra then leads to property (5.5). We are now ready to
prove Theorem 5.1.
UNIFIED MULTIPLE TESTING WITH PRIOR KNOWLEDGE
19
Proof of Theorem 5.1. By the definition of FDRu , we have
"P
#
b
u
1
P
≤
min{w
·
t
,
λ}
0
i
i
i
i∈H
FDRu = E ································································
P
b
i ui 1 Pi ≤ min{wi · t, λ}
"
#
(i) X
αui 1 Pi ≤ min{wi · b
t, λ}
E ·····················································
≤
π
b0 · b
t·n
0
i∈H
"
#
1 Pi ≤ min{wi · b
t, λ}
(ii) X αui wi
E ·············································· ,
=
n
π
b0 · wi · b
t
0
i∈H
π
b ·n·b
t
0
P
where step (i) uses the fact that ·········································
≤ α, by construction
b
i ui 1{Pi ≤min{wi ·t,λ}}
of the procedure; and step (ii) follows by multiplying and dividing by wi in
b0 whenever the numerator is one, we have
term i. Moreover, since π
b0−i = π
#
"
X αui wi
1 Pi ≤ min{wi · b
t, λ}
· E ··············································
FDRu ≤
n
t
π
b0−i · wi · b
0
i∈H
" "
#
#
1 Pi ≤ min{wi · b
t, λ}
1
(iii) X αui wi
=
· E E ·············································· P −i
n
wi · b
t
π
b0−i
i∈H0
#
" "
#
(iv) X αu w
1 P i ≤ wi · b
t
1
i i
−i
· E E ····························· P
.
≤
−i
n
wi · b
t
π
b
0
0
i∈H
where step (iii) follows since conditioning on P −i fully determines π
b0−i , and
step (iv) follows from the elementary bound min{wi · b
t, λ} ≤ wi · b
t. Noting
−i
−i
b
that for each P , the function P
i 7→ t(Pi , P ) is nonincreasing in Pi ,
1{Pi ≤wi ·b
t}
P −i ≤ 1, and hence that
applying Lemma 1 ensures that E ···················
b
wi ·t
X αui wi
1
FDRu ≤
· E −i
n
π
b0
0
i∈H
(v)
≤ α,
where inequality (v) follows from property (5.5), concluding the proof.
20
RAMDAS, BARBER, WAINWRIGHT AND JORDAN
6. Combining one layer of groups, with prior and/or penalty
weights and null proportion adaptivity. We now demonstrate how
to incorporate prior structural knowledge, where the hypotheses are separated into a partition of (non-overlapping) known groups, into the procedures introduced so far. In the next section, we will generalize some of these
ideas to handle multiple arbitrary possibly-incomplete partitions consisting
of possibly-overlapping groups.
When provided with group-level p-values (meaning a p-value corresponding to a test of whether the group is entirely null), one may just trivially run
a BHY procedure on these p-values to control the group FDR. This section
will tackle the setting where the group-level p-values are formed by combining individual-level p-values using the Simes procedure, and we would like to
control the group-level FDR, possibly in the presence of prior and/or penalty
weights. Our main technical tool will be a group-level super-uniformity
lemma (Lemma 3), in analogy with the elementwise super-uniformity lemma
(Lemma 1).
6.1. Formal Setup. Suppose that we have partitioned our hypotheses
into G groups of size n1 , n2 , . . . , nG , with n = n1 + · · · + nG and that we have
access to weights at the group level: prior weights u1 , . . . , uG , and penalty
weights w1 , . . . , wG . (The Simes p-values for each group may themselves
be formed using a weighted procedure, but these weights will be discussed
separately later on.)
To summarize, our data takes the form
(6.1)
P1 , . . . , P n 1
|
{z
}
Group A1 , weights u1 , w1
, . . . , Pn1 +···+nG−1 +1 , . . . , Pn ,
|
{z
}
Group AG , weights uG , wG
and we wish to select a subset of these groups, Sbgrp ⊆ [G], so that the
proportion of null groups (groups consisting entirely of null hypotheses) is
not too high. We define the set of null groups as
0
Hgrp
= g ∈ [G] : Ag ⊆ H0 ,
that is, any group Ag that contains only null hypotheses. The weighted group
FDR is then given by
n
o
P
b
u
1
g
∈
S
0
g
grp
g∈Hgrp
n
o .
FDRgrp
u : = E ················································
P
bgrp
u
1
g
∈
S
g
g
In order to test each group Ag for rejection, we need to compute a p-value
Pggrp for this group—in order to be a true p-value, we need to ensure that
UNIFIED MULTIPLE TESTING WITH PRIOR KNOWLEDGE
21
0 , that is whenever g
Pggrp is (super)uniformly distributed whenever g ∈ Hgrp
is a null group.
Many statistics have been proposed that combine elementwise p-values
P [17].
When the p-values
Pare independent, some options include Fisher’s −2 i ln Pi
and Rosenthal’s i Φ−1 (Pi ), where Φ is the Gaussian CDF (originally proposed by Stouffer et al. [39]). When there are very few non-nulls, the Bonferroni correction is known to be more powerful, and it also works
P under
arbitrary dependence, as does Rüschendorf’s [29, 41] proposal of 2 i Pi /n,
and Rüger’s [28, 25] proposal of P(k) · n/k for a fixed k.
In our setting, the group-level p-value can be arbitrary, either built by
combining the individual p-values of the hypotheses within the group, or
constructed from new independent data. No matter how the group-level pvalues are constructed, one may simply apply a BHY procedure, appropriately weighted or reshaped or adapted, to the group-level p-values to control
the group-level FDR. One example of special interest, closely related to the
BHY procedure, is the setting where the group p-values are formed from
elementwise p-values using the Simes procedure, which we next discuss.
6.2. Generalized Simes tests for the global null. Simes [33] proposed an
improvement to the Bonferroni procedure for global null testing at level α.
We first calculate the Simes p-value using a reshaping function βe if required:2
P(k) · n
,
e
1≤k≤n β(k)
Simes(P ) = min
and we reject HGN if Simes(P ) ≤ α. The connection to the BHY procedure is
quite transparent: note that Simes(P ) ≤ α if and only if the BHY procedure
makes at least one rejection at level α. It is well known that the Simes pvalue, Simes(P ), really is a bonafide p-value, a result that we will recover as
a special case of Proposition 4.
6.3. The prior-weighted Simesw test for the global null. The Simes test [33]
was extended by Hochberg and Liberman [19] to incorporate prior weights
(1)
under independence. As before, we define weighted p-values Qi : = Pi /wi
for each hypothesis, and then calculate the generalized Simesw p-value for
the group as
(6.2)
Q(k) · n
.
1≤k≤n
k
Simesw (P ) : = min
2
Here and henceforth, the tilde in βe will be used to signified a reshaping function for
calculating a Simes p-value within a single group, and we will continue the use of notation
β, without the tilde, when comparing these p-values across multiple groups.
22
RAMDAS, BARBER, WAINWRIGHT AND JORDAN
The global null hypothesis for the group Ag , i.e. the hypothesis that Ag ⊆ H0
consists entirely of nulls, is then rejected at the level α if Simesw (P ) ≤ α.
In a more general setting where the individual p-values Pi within the
group Ag may be arbitrarily dependent, we can instead consider the reshaped
weighted Simes p-value, given by
(6.3)
Q(k) · n
e
1≤k≤n β(k)
rSimesw (P ) : = min
for a reshaping function βe (recall Definition 3).
The following result states that the (weighted and/or reshaped) Simes
p-value really is a bonafide p-value.
Proposition 4. Under the global null hypothesis, the weighted Simes
p-value has the following properties:
(a) Under independence and uniformity, if maxi wi ≤ 1/α, Simesw (P ) is
exactly uniformly distributed.
(b) Under positive dependence, Simesw (P ) is super-uniformly distributed.
(c) Under arbitrary dependence, the reshaped Simes p-value rSimesw (P )
is super-uniformly distributed.
While statement (a) was first proven by Hochberg and Liberman [19], and
statement (c) under unit weights by Hommel [20], all the above statements
are straightforward consequences of the properties of the weighted BH and
BY procedures. For completeness, we prove this proposition below.
Proof of Proposition 4. First we consider the Simes p-value without
reshaping, in the setting of positive dependence. Examining the definition
of the BHuw procedure with the same prior weights wi and with uniform
penalty weights ui = 1, we see that the weighted Simes p-value, Simesw , is
the minimum threshold α for which P passes the BHw procedure:
Simesw (P ) = min {α ∈ [0, 1] : BHw makes at least one rejection at level α} .
In other words, we have the equivalence
(6.4)
Simesw (P ) ≤ t ⇔ b
kt,u,w (P ) ≥ 1
for any t ∈ [0, 1], recalling from (4.2b) that b
kt,u,w is the number of rejections
of the BHu,w procedure at the target FDR level t. Now, note that under the
global null, if we run BHuw at some level t, any discovery made by BHuw is
UNIFIED MULTIPLE TESTING WITH PRIOR KNOWLEDGE
23
a false discovery. Hence the FDP is equal to one whenever there
n is at least
o
one discovery (and zero otherwise), and we have FDR = Pr b
kt,u,w > 0 .
Our previous result, Proposition 2(b), proves that FDR ≤ t for the BHuw
procedure. Hence under the global null, we have
n
o
Pr{Simesw (P ) ≤ t} = Pr b
kt,u,w > 0 ≤ t
Under independence and uniformity, Proposition 2(a) also implies that this
statement holds with equality.
Next, we turn to the setting of arbitrary dependence, where we now use
the reshaped weighted Simes p-value, rSimesw (P ). In this setting, we can
similarly see that
rSimesw (P ) = min {α ∈ [0, 1] : BYw makes at least one rejection at level α} ,
where now we compare against the generalized BY procedure that uses the
same reshaping function βe as we use in our reshaped Simes test. Applying
Proposition 2(c), we again show that Pr{rSimesw (P ) ≤ t} is equal to the
FDR of the corresponding BY procedure, and therefore, Proposition 2(c)
implies that this quantity is ≤ t.
We now demonstrate how to employ the weighted Simes p-value to control
group FDR, first under independence between groups for clarity, and later
under more general assumptions. For this purpose, the following group-level
super-uniformity lemma will prove useful.
6.4. A group-level super-uniformity lemma. In analogy to the superuniformity Lemma 1, we present the following lemma, which contains analogous bounds under the settings of independent or positively dependent base
p-values (in which case the group p-value is constructed with a Simes pvalue), and in the setting of arbitrarily dependent base p-values (in which
case the group p-value can be constructed by any method—Simes, Fisher,
or others—as long as it is a valid p-value.)
0 be a null
Lemma 3 (Group-level super-uniformity lemma). Let g ∈ Hgrp
0
group, that is, Ag ⊆ H . Let PAg denote the p-values in this group, PAg =
(Pj )j∈Ag , and let P−Ag denote the remaining p-values, P−Ag = (Pj )j6∈Ag .
(a) If f : [0, 1]n → [0, ∞) is a nonincreasing function, and the base pvalues P1 , . . . , Pn are independent, then
1 Simesw (PAg ) ≤ f (P )
E ················································· P−Ag ≤ 1.
f (P )
24
RAMDAS, BARBER, WAINWRIGHT AND JORDAN
(b) If f : [0, 1]n → [0, ∞) is a nonincreasing function, and the base pvalues P1 , . . . , Pn are positively dependent, then
1 Simesw (PAg ) ≤ f (P )
E ················································· ≤ 1.
f (P )
(c) If the base p-values P1 , . . . , Pn are arbitrarily dependent, then for any
constant c > 0, any reshaping function β, and any function f : [0, 1]n →
[0, ∞), we have
1 T (PAg ) ≤ cβ(f (P ))
E ··············································· ≤ 1,
cf (P )
where T : [0, 1]|Ag | → [0, 1] is any valid group p-value, i.e. any function with the property that T (PAg ) is superuniform whenever g is a
null group. (For example, we may take T (PAg ) = rSimesw (PAg ), the
e
reshaped weighted Simes p-value formed using reshaping function β.)
(d) Let g1 , . . . , gk be a set of k possibly overlapping null groups, meaning
Ag1 , . . . , Agk ⊆ H0 , and S1 , . . . , Sk represent the corresponding Simes’
p-values. If f : [0, 1]n → [0, ∞) is a nonincreasing function, and the
base p-values P1 , . . . , Pn are positively dependent, then
1 {Simes(S1 , . . . , Sk ) ≤ f (P )}
E ·························································· ≤ 1.
f (P )
The proof of this lemma relies on Lemma 1, and can be found in Appendix C.
We remark thatSstatement (d) is different from statement (b) applied to the
null group g = ki=1 gi ; indeed, in statement (d), the arguments to the Simes’
procedure are themselves Simes’ p-values, and not the original base p-values.
If desired, statement (d) can be further bootstrapped to apply to the root
of an entire tree of null groups, where each internal node stores the Simes’
p-value calculated on its children.
6.5. Group FDR control for a single partition. With the tools developed
so far in place, it is now simple to construct a procedure that controls FDR
at the group level, given a single partition of our n hypotheses into groups
A1 , . . . , AG : we first form some summary p-value Pggrp for each group g =
1, . . . , G, and then run either a BH procedure or a BY procedure on the
resulting list P1grp , . . . , PGgrp , according to the dependence assumptions.
First we construct the group p-values. For each group, we may choose
to use a weighted Simes p-value to collapse the group into a single summary p-value, reducing the overall problem to a standard multiple testing
UNIFIED MULTIPLE TESTING WITH PRIOR KNOWLEDGE
25
problem on G p-values. To avoid conflicting notation with the group-level
(0)
(0)
weights w1 , . . . , wG , let w1 , . . . , wn denote the weights on the n individual
hypotheses for computing these group-level Simes p-values.
(0)
Referring to the notation of setup (6.1), define Qi : = Pi /wi and compute
the Simesw p-values for each group, Simesw (PA1 ), . . . , Simesw (PAG ), where
(6.5)
Pggrp
= Simesw (PAg ) =
min
Qg,(k) ·
1≤k≤|Ag |
P
j∈Ag
k
(0)
wj
,
where Qg,(k) is the k-th smallest p-value in group Ag . If we choose to use a
e then the weighted rSimes p-value is computed as
reshaping function β,
P
(0)
Qg,(k) · j∈Ag wj
grp
(6.6)
Pg = rSimesw (PAg ) = min
1≤k≤|Ag |
βeg (k)
for each group, where the subscript g in the reshaping function, i.e. βeg (k),
indicates that for each group we are free to choose different βe functions
(though there may be little reason to exercise this freedom).
P
(0)
Note that, in both definitions, the numerator uses j∈Ag wj instead of
(0)
|Ag | because, depending on how we normalize our weights wi , we may
have that the sum of all weights over all groups is n, but the sum of the
weights within the group Ag need not be |Ag |; hence the adjustment with
the appropriate normalization. As argued earlier, Simesw (PAg ) is a bonafide
p-value, i.e. is superuniform, under the “group intersection hypothesis” that
Ag is composed only of nulls, whenever p-values in Ag are independent
or positively dependent; under arbitrary dependence, the reshaped p-value
rSimesw (PAg ) is now the one that is guaranteed to be superuniform.
Finally, if we do not wish to use a Simes p-value, we are free to set
Pggrp = T (PAg )
for any function T : [0, 1]|Ag | → [0, 1] with the property that it is a valid
p-value under the group null, that is,
0
g ∈ Hgrp
⇒ T (PAg ) is superuniform.
This includes the reshaped Simes p-value as a special case, i.e. T (PAg ) =
rSimesw (PAg ), and also encompasses the Fisher and Rosenthal group p-value
constructions. (In principle, the group-level p-values could also be computed
from the raw data, which does not alter any of the results proved here, as
0 ).)
long as Pggrp is still superuniform for any null group g ∈ Hgrp
26
RAMDAS, BARBER, WAINWRIGHT AND JORDAN
Next, we state the results for FDR control in this group testing setting. As
expected, under independence we are free to use adaptivity, while under arbitrary dependence, we are obligated to use reshaping. Under independence,
the null proportion estimate (for fixed λgrp ∈ (0, 1)) is naturally defined as
(6.7)
π
bgrp :=
|u · w|∞ +
PG
g=1
ug wg 1 Pggrp > λgrp
G(1 − λgrp )
.
Proposition 5 (Group-level FDR control). Recalling setup (6.1):
(a) Under independence of the group p-values, the St-BHuw procedure applied to {Pggrp } achieves FDRgrp
u ≤ α.
(b) Under positive dependence of the base p-values in P , the BHuw procedure applied to Simes’ p-values {Simes(PAg )} achieves FDRgrp
u ≤ α.
(c) Under arbitrary dependence of group p-values, the BYuw procedure applied to {Pggrp } achieves FDRgrp
u ≤ α.
We do not provide an explicit proof of this proposition, since the proofs
for FDRgrp
u control directly mimic the corresponding proofs of FDRu control
for the same procedures applied to the base p-values in P , with the single
alteration that the appropriate statements of Lemma 3 are invoked in place
of their corresponding statements in Lemma 1. Further, the proof of the
above proposition follows as a special case of the proof of Theorem 7.1 for
the multilayer procedure p-filter to be introduced in the next section.
As observed in Barber and Ramdas [1] for the unweighted case, we can
view the group FDR procedure described in this subsection as an interpolation between the weighted Simes test, and the weighted St-BH procedure.
This is seen by considering two extremes: taking one single group of size n
recovers the Simes test of the global null (across all n hypotheses), while
taking instead n groups of size one recovers the BH or BY procedures for
testing individual hypotheses. Of course, while the above algorithm controls
group-level FDR, it has no guarantee for the elementwise FDR when the
hypotheses are considered individually rather than in groups. Our general
multilayer method, introduced in the next section, gives this type of simultaneous guarantee under various forms of dependence.
7. FDR control and internal consistency for multiple partitions.
We now turn to the general case in which we allow multiple partitions of the
set of hypotheses, as in Barber and Ramdas [1], while continuing to support
prior weights, penalty weights, group-level adaptivity, reshaping under arbitrary dependence, as well as other generalizations of overlapping groups and
UNIFIED MULTIPLE TESTING WITH PRIOR KNOWLEDGE
27
incomplete partitions. The presented framework will specialize, in the case
of a single partition, to all the procedures discussed in previous sections.
Our formal setup is as follows. We have M partitions of interest, with the
mth partition having Gm many groups:
(m)
(m)
A1 , . . . , AGm ⊆ [n] for m = 1, . . . , M .
The partitions and the groups within each partition can be specified in any
order. The order in which the partitions are described makes no difference
to the final output; similarly, the ordering within each partition’s groups is
also arbitrary.
As in Section 6, define the null set for the mth partition as containing
those groups that are entirely filled with null hypotheses:
o
n
0
0
.
Hm
= g ∈ [Gm ] : A(m)
⊆
H
g
As before, there is a natural notion of group-level FDR—namely, a falsely
discovered group is a group that is entirely composed of true nulls, for which
our algorithm has incorrectly proclaimed at least one discovery within that
group (i.e. effectively has rejected the global null hypothesis within that
group). Each group in each partition is associated with a summary p-value
Pgm , which we assume to be derived from the base p-values (Pi : i ∈ Am
g ),
but may more generally be calculated from the raw data or from another
source without altering our results.
For interpretability, it is arguably of interest to maintain “internal consistency” among rejected hypotheses and groups. For a single partition into
non-overlapping groups, we say that a set of rejections at the individual level
and group level(s) are internally consistent if:
(7.1)
We only reject groups containing at least one rejected hypothesis,
and we only reject a hypothesis if its group is also rejected.
We will require the rejections to be internally consistent simultaneously for
all partitions, and later extend this notion in two ways to partitions with
leftover sets and overlapping groups (weak and strong internal consistency).
The above notion of internal consistency is an extension, to the multilayer
setting, of the requirement of both coherence and consonance as defined in
the classical paper by Gabriel [14], and explored in depth in the FWER
literature by Sonnemann and Finner [34, 35, 36], and Romano et al. [27].
Even in the case of just one layer of groups on top of the individual
hypotheses, handling consistency while controlling both group-level and
individual-level FDR is not trivial. For example, a sequential procedure, of
28
RAMDAS, BARBER, WAINWRIGHT AND JORDAN
first rejecting groups at a target FDR level α1 , and then rejecting individual
hypotheses within rejected groups at level α2 , may neither succeed in controlling the FDR at the individual level (due to not accounting for selection
bias), nor succeed in being internally consistent (because some groups may
get selected in the first round, but none of the individuals within that group
may get selected in the second round). Further, such a method is not easily
generalized to non-hierarchical partitions. Similarly, a parallel procedure of
choosing which groups to reject independently of which individuals to reject
may also fail to be internally consistent. A naive solution to this issue is
intersection—i.e., rejecting those hypotheses whose groups are rejected at
every layer—but this may also fail to control FDR at both levels (see Barber
and Ramdas [1] for examples of this phenomenon).
Our algorithm can guarantee FDR control at the hypothesis and group
levels, that holds simultaneously for multiple arbitrary non-hierarchical partitions of the hypotheses, while also incorporating weights and adaptivity,
all while maintaining the internal consistency of rejected hypotheses. We
now introduce the framework formally.
7.1. The p-filter framework. The algorithmic framework presented in
this section generalizes the p-filter framework by Barber and Ramdas
[1]. Roughly speaking, the algorithm is a multivariate extension of classical
step-up procedures, based on the following sequence of steps:
• Select all hypotheses in each layer whose p-values are smaller than
some initial layer-specific threshold.
• Reject an elementary hypothesis if it is contained in a selected group
in every layer.
• In each layer, reject a group hypothesis if it contains a rejected elementary hypothesis. Then, estimate the group-FDP in each layer.
• Lower the initial thresholds at each layer, and repeat the steps above,
until the group-FDP is below the desired level for all partitions.
We use the same name “p-filter” for this extended algorithm, while noting
six important ways in which this paper adds to the flexibility of the earlier
framework.
1. Weights. The m-th partition is associated with two sets of Gm many
positive weights, one for each group g in that partition:
(m)
Penalties {u(m)
g } and priors {wg }, such that
Gm
X
(m)
u(m)
= Gm .
g wg
g=1
These can be used to take differing group sizes into account.
UNIFIED MULTIPLE TESTING WITH PRIOR KNOWLEDGE
29
2. Reshaping. If the p-values within or across layers are arbitarily dependent, we use βm to reshape thresholds in layer m (possibly different for the M layers). If Simes p-values are used to form group
(0)
level p-values, then we may place weights wi on the base p-values
P1 , . . . , Pn for the purpose of calculating a weighted Simes p-value,
); if additionally the p-values within the group are
Pgm = Simesw (PAm
g
arbitrarily dependent, we would then choose reshaping functions βemg ,
potentially with different functions chosen for each group, for comput).
ing Pgm = rSimesw (PAm
g
3. Adaptivity. For any partition whose group-level p-values are known
to be independent (i.e., independence between groups, not necessarily
within groups), we can incorporate Storey’s null-proportion adaptivity.
Every layer m has a user-defined constant λm ∈ (0, 1), with which we
define a weighted null proportion estimator for partition m:
(7.2)
π
bm :=
|u(m) · w(m) |∞ +
P
g
(m)
(m)
ug wg
1 Pgm > λm
Gm (1 − λm )
.
The use of null-proportion adaptivity in any one layer may improve
the power in all layers, since more groups being discovered in one
layer then allows more individual hypotheses to be discovered, and
hence more groups to be discovered in other layers.
4. Incomplete partitions. We allow the partitions to be incomplete—
we let the m-th partition’s “leftover” set L(m) ⊂ [n] represent all elements from the m-th partition that do not belong to any group in the
m-th partition:
[
L(m) = [n]\ A(m)
and `m = |L(m) |.
g
g
This gives additional flexibility to the user who may not want to assign
some hypotheses to any groups. It is important to note that L(m) is not
just another group—this set is not counted when calculating the grouplevel FDR in layer m, meaning that discoveries within this leftover set
do not alter the FDR at layer m, and hypotheses in this leftover set
have no internal consistency constraints imposed by layer m.
5. Overlapping groups. We allow the groups in any partition to overlap. Any elementary hypothesis need not be part of just a single
g ∈ [Gm ]; we let gm (i) denote the set of groups to which Pi belongs:
gm (i) = {g ∈ [Gm ] : Pi ∈ A(m)
g }.
Two natural extensions of internal consistency are as follows.
30
RAMDAS, BARBER, WAINWRIGHT AND JORDAN
• Weak internal consistency: we reject Hi if and only if in every
partition, the following holds: either there is at least one rejected
group containing i, or i ∈ L(m) , the leftover set, meaning that i
does not belong to any group at this layer.
• Strong internal consistency: we reject Hi if and only if in every partition, the following holds: either every group that contains
i is rejected, or i ∈ L(m) .
We remark these are not the only two notions of internal consistency
that can be handled by our framework: any monotone notion of internal consistency still works, meaning that decreasing the p-values can
only increase the number of rejections at all levels.
6. Arbitrary group p-values. The new p-filter algorithm no longer
necessarily uses Simes p-values at the group layers. In other words,
each group-level p-value at each layer can be arbitrary; it can be formed
by combining the elementwise p-values in any way (or, as mentioned
before, might be constructed from the raw data). Depending on how
these p-values are constructed, we can appropriately use adaptivity
(under independence, if Simes p-values are used), or reshaping (under dependence, or if we choose to use Fisher, Rosenthal, or other
combination methods), as needed.
p-filter will reject some subset Sb ⊆ [n] of hypotheses and a subset
b
Sm ⊆ [Gm ] of groups in each partition m, with the exact choice of these
rejection sets defined in the next subsection. Given Sb and the Sbm ’s, we then
define the weighted FDR for the mth partition as
n
o
P
(m)
b
u
1
g
∈
S
0
m
g∈Hm g
n
o .
FDR(m)
= E ··················································
u
P
(m)
b
g∈[Gm ] ug 1 g ∈ Sm
The rejections made by p-filter will be internally consistent, and satisfy
FDR(m)
≤ αm simultaneously for all m = 1, . . . , M.
u
To describe a situation where allowing for non-hierarchical layers, and
for more than two layers, may be useful, imagine that we can write the pvalues down in a rectangular grid where the p-values within a row have some
interpretable meaning (say, hypotheses corresponding to the same point in
space), and similarly the p-values within a column have a different meaning
(say, hypotheses corresponding to the same point in time). Then in addition
to controlling the overall FDR, we may also want to control the “spatial”
FDR using a row partition where each group is a row, and “temporal”
UNIFIED MULTIPLE TESTING WITH PRIOR KNOWLEDGE
31
FDR using a column partition where each group is a column, or even the
“spatio-temporal” FDR for using a rectangular partition when each group
is a block in space-time. Refer to Figure 1 for visualization, and see Barber
and Ramdas [1] for further discussion, numerical simulations, as well as a
neuroscience application.
Fig 1. Consider 16 hypotheses written as a 4 × 4 grid, with four partitions: ele-
mentary, rows, columns, blocks. On the top row is the underlying truth, with the
leftmost panel showing the hypothesis-level non-nulls, and the other three panels
showing which groups in each partition are hence identified as non-null. On the
bottom row is an example of a set of discoveries, with the leftmost panel showing
the hypothesis-level rejections, and the other three panels showing which groups are
correspondingly rejected (light-grey for correct rejections, black for false rejections).
The false discovery proportions in each partition are 0.2, 0, 0.33, 0.5 respectively.
7.2. Deriving the p-filter algorithm. To run the p-filter algorithm,
we need to search for rejection thresholds for each layer. These thresholds
will be parametrized by weighted discovery counts km ∈ [0, Gm ] for each layer
m = 1, . . . , M . The reader is cautioned that each km will not necessarily be
an integer but instead be a real number corresponding loosely to the total
(m)
rejected penalty weight. If the weights ug are all set to equal 1, then km
indeed corresponds to the number of groups in layer m that are rejected.
Given ~k := (k1 , . . . , kM ), we first perform an initial screening on each layer
separately:
(m)
wg αm km
init ~
m
b
(7.3a)
Sm (k) = g ∈ [Gm ] : Pg ≤ min{ πbm Gm , λm } .
32
RAMDAS, BARBER, WAINWRIGHT AND JORDAN
If we choose to use reshaping, we would instead define
(m)
wg αm βm (km )
init ~
m
b
, λm } .
(7.3b)
Sm (k) = g ∈ [Gm ] : Pg ≤ min{
π
bm Gm
(The null proportion estimate π
bm can be defined as in (7.2) if we choose to
use adaptivity; if not, we simply set π
bm = 1 and λm = 1, so that the same
equations can be used in either setting.)
We remark that while we are not obliged to do so, one option for group(m)
level p-values is to set Pgm = Simesw(1) (Ag ); for convenience, we now recompute these p-values in the notation of the present section. Recall that
(1)
Qi = Pi /wi , and that
(m) P
(1)
Qg,(i) j∈A(m) wj
g
(m)
Simesw(1) (Ag ) : = min
,
(m)
i
1≤i≤|Ag |
(m)
(m)
where Qg,(i) is the i-th smallest p-value in Ag . If we choose to use reshaped
Simes p-values, we would instead compute
(m) P
(1)
Qg,(i) j∈A(m) wj
g
(m)
Simesw(1) (Ag ) : = min
,
(m)
βemg (i)
1≤i≤|Ag |
where βemg is the reshaping function chosen for group g in layer m.
Also recall that if the groups in the m-th layer are independent, we may
use null-proportion adaptivity by setting π
bm as in (7.2), while otherwise we
would set π
bm ≡ 1. We then define the total set of rejections, either as
M
\
[
b ~k) = Sbweak (~k) =
∪ L(m) ,
(7.4a)
S(
A(m)
g
m=1
init (~
g∈Sbm
k)
if we wish to enforce weak internal consistency, or alternately as
M
\
[
b ~k) = Sbstrong (~k) =
[n]\
(7.4b)
S(
Am
g
m=1
init (~
g∈[Gm ]\Sbm
k)
for strong internal consistency. To interpret these expressions, we see that
under weak consistency (7.4a), hypothesis i is rejected if, for every layer m,
either i is in the leftover set L(m) , or i belongs to at least one of the groups
(m)
Ag rejected in this layer—we can rewrite this condition as
init ~
Sbweak (~k) = {Pi : ∀m, either Pi ∈ L(m) , or ∃ g ∈ g(m, i), A(m)
∈ Sbm
(k)}.
g
UNIFIED MULTIPLE TESTING WITH PRIOR KNOWLEDGE
33
For strong consistency (7.4b), hypothesis i is rejected if, for every layer m,
either i is in the leftover set L(m) , or i belongs to only rejected groups—that
is, if i does not belong to any groups Am
g which were not rejected—which
we can rewrite as
init ~
Sbstrong (~k) = {Pi : ∀m, either Pi ∈ L(m) , or ∀ g ∈ g(m, i), A(m)
∈ Sbm
(k)}.
g
Finally, we redefine the set of groups in layer m which are rejected in
order to enforce internal consistency:
n
o
init ~
b ~k) 6= ∅ and g ∈ Sbm
(7.6)
Sbm (~k) = g ∈ [Gm ] : A(m)
∩ S(
(k) ,
g
b ~k) = Sbweak (~k) or S(
b ~k) = Sbstrong (~k), as desired. Examining
where either S(
these definitions, readers may verify that the appropriate (weak or strong)
notion of internal consistency (7.1) is satisfied.
Of course, these definitions depend on the initial choice of the vector ~k.
Since we would like to make a large number of discoveries, we would like to
use a ~k that is as large as possible, while at the same time controlling the
layer-specific false discovery rates,
n
o
(m)
ug 1 g ∈ Sbm (~k)
~
n
o .
FDP(m)
u (k) : = ·························································
P
(m)
~
b
1
g
∈
S
(
k)
u
m
g∈[Gm ] g
P
g∈H0m
Now, define the data-dependent set of feasible vectors ~k = (k1 , . . . , kM ) as
X
b = ~k ∈ [0, G1 ] × · · · × [0, GM ] :
(7.7) K
u(m)
≥
k
for
all
m
,
m
g
g∈Sbm (~k)
b on input parameters such as
where we suppress the implicit dependence of K
(m)
(m)
αm , λm , {wg }, {ug }. In particular, if the penalty weights are all equal to
one, then the consistency condition defining the “feasible” ~k’s is equivalent
to requiring that |Sbm (~k)| ≥ km for all m = 1, . . . , M —i.e., the numbers of
rejections in each layer at the vector ~k are elementwise ≥ ~k. This is analogous
to the BH procedure (3.1) and the generalized BY procedure (3.2), which
each find a value k such that there are at least k many rejections at the
αβ(k)
threshold αk
(for BY). The above condition can be viewed
n (for BH) or
n
as a generalization, to the multi-partition setting, of the “self-consistency”
condition described by Blanchard and Roquain [9].
Also, it is worthy of note that the p-filter algorithm in Barber and Ramdas [1] was derived in terms of thresholds ~t instead of number of rejections ~k,
34
RAMDAS, BARBER, WAINWRIGHT AND JORDAN
[ m (~t) ≤ αm ,
and there the corresponding feasibleity condition was that FDP
[ m (~t) is an empirical-Bayes type estimate of the FDP. Indeed, if
where FDP
we avoid π
bm , βm , wm , um for simplicity, then associating b
tm to αm b
km /Gm
and comparing our derivation to the one in Barber and Ramdas [1], we can
see that the “self-consistency” viewpoint and the “empirical-Bayes” viewpoint are equivalent and lead to the same algorithms. However, when dealing
with arbitrary dependence, the proofs are vastly simpler in terms of ~k than
in terms of ~t, explaining the switch in choice of notation in this paper.
As with the BH and BY procedures, we then choose the largest feasible
thresholds km , given by:
n
o
b
b
(7.8)
km = max km : ∃k1 , . . . , km−1 , km+1 , . . . , kM s.t. ~k ∈ K
This choice defines our algorithm: the p-filter algorithm rejects the hybb
potheses S(
k1 , . . . , b
kM ), as defined in (7.4a) or (7.4b), with rejections at
layer m given by Sbm (b
k1 , . . . , b
kM ) as defined in (7.6). Next, we summarize
the theoretical guarantees of p-filter.
7.3. Theoretical guarantees. The following proposition states that the
b actually has a well-defined “maximum” corner.
set of feasible vectors K
b be defined as in equaProposition 6. Let the set of feasible vectors K
tion (7.7), and let the partition-specific maximum feasible vector b
km be defined as in equation (7.8). Then we have
(7.9)
b.
(b
k1 , . . . , b
kM ) ∈ K
Barber and Ramdas [1] proved this result in the setting of the original
p-filter algorithm; for completeness, we prove it in this more general setting in Appendix D.
The vector (b
k1 , . . . , b
kM ) is not just feasible from the perspective of selfb but it is also feasible from the perspective of
consistency as captured by K,
FDR control. Specifically, the next theorem proves that—assuming for now
bb
that we can find (b
k1 , . . . , b
kM )—selecting the set S(
k1 , . . . , b
kM ) guarantees
(m)
simultaneous control of FDRu for all M partitions.
Theorem 7.1. Any procedure that finds (b
k1 , . . . , b
kM ) from definition (7.8)
satisfies the following properties for all m = 1, . . . , M simultaneously:
(a) If the base p-values are independent, and all group p-values are given
), then employing adaptivity (defining π
bm as
by Pgm = Simesw (PAm
g
(m)
in (7.2)) guarantees that FDRu
≤ αm .
UNIFIED MULTIPLE TESTING WITH PRIOR KNOWLEDGE
35
(b) If the base p-values are +dependent and all group p-values are given
Pgm
(m)
P
(m)
FDRu
0
ug
(m)
wg
(m)
wg
), then
≤ αm g∈HmGm
by
= Simesw (PAm
≤ αm .
g
(c) When all p-values are arbitrarily dependent, and are constructed arbitrarily (under the assumption that Pgm is superuniform for any null
0 , meaning it is a valid p-value), then using reshaping as
group g ∈ Hm
(m)
P
0
ug
(m)
in (7.3b) guarantees that FDRu ≤ αm g∈HmGm
≤ αm .
(d) In the setting of part (c), if additionally the groups at layer m are
, for each g ∈
is independent from P−Am
independent (that is, PAm
g
g
[Gm ]), then using reshaping as in (7.3b) and adaptivity for layer m as
(m)
in (7.2), guarantees that FDRu ≤ αm .
To remark on the difference between parts (c) and (d), what these two
results guarantee is that if we use adaptivity for some set Madapt ⊂ [M ] of
layers, and do not use adaptivity (i.e. set π
bm = 1) for the remaining layers,
then FDR control is maintained across all layers as long as, for each m ∈
Madapt , the layer-wise independence statement holds—PAm
is independent
g
m
from P−Ag , for each g ∈ [Gm ]. If this condition fails for some m ∈ Madapt ,
the FDR control in other layers will in fact not be affected.
The proof, given in Appendix E, integrates the various ingredients required to handle weights, null proportions, and groups as detailed in earlier
sections. The proof includes some new ideas of independent interest, specific
to handling overlapping groups with dependent p-values. One application
of statement (d) is when the base p-values are independent, there are no
overlapping groups, and group p-values are formed using a Fisher, Rosenthal, or other combinations of the base p-values. Recently, Katsevich and
Sabatti [23] prove that in case (d), the FDR is controlled even without using
reshaping, albeit at a constant factor larger than the target level.
In practice, if we have accurate side information about group structures
that the rejected hypotheses likely respect, then we may significantly improve our precision, achieving a lower FDR than the theoretical bound,
without affecting our power much. However, inaccurate side information
may significantly lower our power, since each p-value would have additional
misguided constraints to meet. These aspects were explored in simulations
by Barber and Ramdas [1].
Special cases. The setting with a single partition (M = 1) recovers all
other algorithms discussed in this paper. When we have the finest partition
with n groups containing one hypothesis each, p-filter reduces exactly to
the generalized BHY procedures discussed in Section 3, which of course include BH and BY, and their prior- and/or penalty-weighted variants BHYuw
36
RAMDAS, BARBER, WAINWRIGHT AND JORDAN
described in Section 4. When we employ adaptivity, we recover the St-BH
method and its weighted variant in Section 5. When we instantiate p-filter
with the coarsest partitions with a single group containing all n hypotheses, we recover exactly the generalized Simes test, and its weighted variants
from Section 6. When instantiated to a single partition with G groups of hypotheses, we recover exactly the test that controls group-FDR, as discussed
at the end of Section 6. All the theorems and propositions in this paper are
essentially deduced as special cases of Theorem 7.1.
Algorithm 1 The p-filter for multi-layer FDR control
Input: M possibly incomplete partitions of possibly overlapping groups of indices [n];
A vector of base p-values P ∈ [0, 1]n ;
Group p-values Pgm for each group g = 1, . . . , Gm in layers m = 1, . . . , M ;
M target FDR levels {αm };
(m)
(m)
M sets of prior weights and/or penalty weights {wg , ug };
M thresholds for adaptive null proportion estimation {λm };
M reshaping functions {βm }, if desired.
Initialize: Set km = Gm , and π
bm as in definition (7.2).
repeat
for m = 1, . . . , M do
Update the mth vector: defining Sbm (~k) as in equation (7.6) (using weak or strong
consistency, as desired), let
X
0
0
∈ [0, Gm ] :
km ← max km
u(m)
≥ km
g
bm (k1 ,...,km−1 ,k0 ,km+1 ,...,kM )
g∈S
m
end for
until the vectors k1 , . . . , kM are all unchanged for one full cycle.
Output: Adaptive vector b
k = (k1 , . . . , km ).
7.4. An efficient implementation. Although one can employ a bruteforce grid search to find (b
k1 , . . . , b
kM ), the p-filter algorithm presented in
Algorithm 1 is able to find this vector efficiently using a coordinate-descent
style procedure, and is a strict generalization of the algorithm by the same
name in Barber and Ramdas [1]. Code for this procedure is publicly available
at https://www.stat.uchicago.edu/∼rina/pfilter.html. The following proposition provides a correctness guarantee for Algorithm 1:
Proposition 7. The output of Algorithm 1 is the maximum feasible
corner (b
k1 , . . . , b
km ) defined in equations (7.8) and (7.9).
This result was proved by Barber and Ramdas [1] in the setting of the original p-filter algorithm, where the km ’s take only integer values; here, the
UNIFIED MULTIPLE TESTING WITH PRIOR KNOWLEDGE
37
algorithm is slightly more subtle, with real-valued km ’s due to the presence
(m)
of penalty weights ug . The proof of the proposition for this more general
setting is given in Appendix D.
8. Discussion. This paper provides simple proofs for many existing
multiple testing procedure—while also generalizing the conditions under
which some of them work—including the global null test of Simes [33], the
weighted Simes test of Hochberg and Liberman [19], the FDR controlling
procedures of Benjamini and Hochberg [3] and Benjamini and Yekutieli [7],
the weighted FDR controlling procedure of Benjamini and Hochberg [4], the
p-value weighting procedure of Genovese et al. [16], and the null-proportion
adaptivity procedure of Storey [37]. We also described several new procedures, including prior+penalty weighted versions of BH and St-BH, and an
adaptive weighted Simes test, and showed how to unify these concepts in
the grouped setting. Finally, our p-filter algorithm unifies and generalizes
all of the above procedures, building on the original p-filter framework
introduced by Barber and Ramdas [1].
The procedures that we have analyzed and generalized do not fully cover
the huge literature on FDR controlling procedures. For example, all procedures in this paper are step-up procedures, and much work has also been
done on alternative styles of procedures, like step-down, step-up-down and
multi-step methods. For example, Benjamini and Liu [6] propose step-down
procedures that control FDR under independence. Later, procedures by Benjamini and Liu [5] and Romano and Shaikh [26] provably control FDR under
arbitrary dependence, with Gavrilov et al. [15] extending them to adaptive
control under independence. Two-step adaptive procedures have been analyzed in Benjamini et al. [8] under independence, and by Blanchard and
Roquain [10] under dependence. Different methods of incorporating weights
into such procedures have also been studied, e.g. a different notion of the
weighted Simes p-value proposed by Benjamini and Hochberg [4]. Hu et al.
[21] and Benjamini and Bogomolov [2] also propose ways to take a single
partition of groups into account, while Yekutieli [42] discusses hierarchical
testing. The super-uniformity lemmas (Lemma 1 and, in the grouped setting, Lemma 3), can be used to quickly prove FDR control for many of these
procedures, and may be a useful tool for exploring potential extensions of
multiple testing procedures into broader settings.
Acknowledgments. We thank Wenyu Chen for helping implement the
new p-filter algorithm, and Etienne Roquain for clarifying a point about
doubly-weighted procedures. We thank Eugene Katsevich, Aditya Guntuboyina, Sujayam Saha, Larry Wasserman and Fanny Yang for relevant discus-
38
RAMDAS, BARBER, WAINWRIGHT AND JORDAN
sions. The authors are also thankful to audience members at the Statistics departments of Stanford, Wharton, UC Davis, the St. Louis Workshop
on Higher Order Asymptotics and Post-Selection Inference, and the NIPS
Workshop on Adaptive Data Analysis, whose questions partly shaped this
work. This work was supported in part by the Office of Naval Research
under grant number W911NF-16-1-0368, the Air Force Office of Scientific
Resesarch under grant number AFOSR-FA9550-14-1-0016, by NSF award
DMS-1654076, and by an Alfred P. Sloan fellowship.
APPENDIX A: PROOF OF SUPER-UNIFORMITY LEMMA 1
Statement (b) of Lemma 1 follows directly from Blanchard and Roquain
[9], and independently reproved by Barber and Ramdas [1]. Statement (a)
with inequality (but not with equality) follows as a special case of (b), since
independence is a special case of positive dependence, and the distribution
of a null Pi does not change on conditioning on an independent set of pvalues. Statement (c) was proved also by Blanchard and Roquain [9]. We
now prove the statements (a), (d).
Statement (a). We prove the first part of Lemma 1, under the assumptions
that the function P 7→ f (P ) satisfies the leave-one-out property with respect
to index i, and that Pi is uniformly distributed and is independent of the
remaining p-values. Since Pr{Pi = 0} = 0, we ignore this possibility in the
following calculations. Since f satisfies the LOOP condition, we have
n
o
−i
e
1 Pi ≤ f (P )
1 {Pi ≤ f (P )}
····························· =
.
f (P )
f (Pe−i )
This can be seen by separately considering what happens when the numerator on the left-hand side is zero or one.
Since P −i determines f (Pe−i ), it immediately follows that
o
n
−i )
e
1
P
≤
f
(
P
i
1 {Pi ≤ f (P )}
E ····························· P −i = E
P −i
f (P )
f (Pe−i )
n
o
Pr Pi ≤ f (Pe−i ) f (Pe−i )
=
f (Pe−i )
= 1,
where the last step follows since f has range [0, 1],
n and Pi is uniformly diso
tributed and is independent of Pe−i — therefore, Pr Pi ≤ f (Pe−i ) f (Pe−i ) =
UNIFIED MULTIPLE TESTING WITH PRIOR KNOWLEDGE
39
f (Pe−i ). This concludes the proof of the super-uniformity lemma under independence and uniformity.
Statement (d). For each ` = 1, . . . , m, let ν` be a probability measure on
Rk
[0, ∞) chosen such that β` (k) = βν` (k) = x=0 x dν` (x), as in the definition
of a reshaping function (Definition 3). Let X` ∼ ν` be drawn independently
for each ` = 1, . . . , m, and let ν be the
Qmprobability measure on [0, ∞) corresponding to the distribution of Z = `=1 X` . Then
!
Z f` (P )
m
m
Y
Y
c·
β` (f` (P )) = c ·
x` dν` (x` )
`=1
x` =0
`=1
Z
∞
=c·
∞
m
Y
xm =0
`=1
Z
···
x1 =0
"
=c·E
m
Y
!
x` · 1 {x` ≤ f` (P )}
dνm (xm ) . . . dν1 (x1 )
#
(X` · 1 {X` ≤ f` (P )})
`=1
= c · E [Z · 1 {X1 ≤ f1 (P ), . . . , Xm ≤ fm (P )}]
)#
"
(
m
Y
f` (P )
≤c·E Z ·1 Z ≤
`=1
Z
=c·
Qm
`=1 f` (P )
z dν(z) = c · βν
z=0
m
Y
!
f` (P ) .
`=1
Therefore,
Q
Q
f` (P ))}
1 {Pi ≤ c · βν ( m
β` (f` (P ))}
1 {Pi ≤ c · m
`=1
`=1
Q
Q
≤ 1,
≤ E ·······················································
E ······················································
c· m
c· m
`=1 f` (P )
`=1 f` (P )
where the last step holds by Lemma 1(c).
APPENDIX B: PROOF OF INVERSE-BINOMIAL LEMMA 2
The
bound follows immediately from Jensen’s inequality, since E [Z] =
Plower
n
1 + b i=1 ai . We split the argument for the upper bound into three cases.
Case 1: integer weights. First, suppose that all the weights ai are integers,
that is, ai ∈ {0, 1} for all i. In this case, we have Z ∼ 1 + Binomial(k, b),
where k is the number of weights ai that are equal to 1. A simple calculation
40
RAMDAS, BARBER, WAINWRIGHT AND JORDAN
shows that
X
k
1
k z
1
=
b (1 − b)k−z
E
1 + Binomial(k, b)
1+z z
z=0
k
X
1
k + 1 z+1
=
b (1 − b)(k+1)−(z+1)
b(1 + k)
z+1
z=0
1
=
· Pr{Binomial(k + 1, b) ≤ k}
b(1 + k)
1
1
P
≤
=
.
b(1 + k)
b(1 + i ai )
Case 2: one non-integer weight. Suppose that exactly one of the weights ai
is a non-integer. Without loss of generality we can take a1 = · · · = ak = 1,
ak+1 = c, ak+2 = · · · = an = 0, for some k ∈ {0, . . . , n − 1} and some
c ∈ (0, 1). Let A = Z1 + · · · + Zk+1 ∼ Binomial(k + 1, b), and Y = Zk+1 ∼
A
Bernoulli(b). Note that Pr{Y = 1 | A} = 1+k
. Then
1
1
=E
E
Z
1 + A − (1 − c)Y
1
=E E
A
1 + A − (1 − c)Y
1
1
=E
· Pr{Y = 0 | A} +
· Pr{Y = 1 | A}
1+A
c+A
1
1
1
=E
+
−
· Pr{Y = 1 | A}
1+A
c+A 1+A
1
1−c
A
=E
+
·
1 + A (c + A)(1 + A) 1 + k
1
1−c
1+k
≤E
+
·
,
1 + A (c + 1 + k)(1 + A) 1 + k
where the inequality holds since
Simplifying, we get
A
c+A
≤
1+k
1+k+c
because 0 ≤ A ≤ k + 1.
1
1
2+k
1
2+k
1
1
P
E
≤E
·
≤
·
=
=
,
Z
1+A 1+k+c
b(2 + k) 1 + k + c
b(1 + k + c)
b(1 + i ai )
where the inequality uses the fact that E
calculated in Case 1.
h
1
1+Binomial(k+1,b)
i
≤
1
b(2+k)
as
UNIFIED MULTIPLE TESTING WITH PRIOR KNOWLEDGE
41
Case 3: general case. Now suppose
P that there are at least two non-integer
weights, 0 < ai ≤ aj < 1. Let C = `6=i,j a` Z` , then Z = 1+C +ai Zi +aj Zj .
Let α = min{ai , 1 − aj } > 0. Then
C = b2 ·
1
1
1
1
+b(1−b)·
+b(1−b)·
+(1−b)2 ·
1 + C + ai + aj
1 + C + ai
1 + C + aj
C
1
1
1
1
≤ b2 ·
+b(1−b)·
+b(1−b)·
+(1−b)2 · ,
1 + C + ai + aj
1 + C + (ai − α)
1 + C + (aj + α)
C
E
1
Z
where the inequality follows from a simple calculation using the assumption
that α ≤ ai ≤ aj ≤ 1 − α. Now, define a new vector of weights ã where
e = 1 + P ã` Z` ,
ãi = ai − α, ãj = aj + α and ã` = a` if ` ∈
/ {i,hj}.i Defining Z
`
the above calculation proves that E Z1 ≤ E 1e (by marginalizing over C).
Z
P
P
Note that
i ãi , but ãi has (at least) one fewer non-integer
i ai =
weight. Repeating this process inductively, we see that we can reduce to the
case where there is at most one non-integer weight (i.e., Case 1 or Case 2).
This proves the lemma.
APPENDIX C: PROOF OF GROUP SUPER-UNIFORMITY LEMMA 3
The proof of Lemma 3(c) is straightforward, by applying Lemma 1(c).
More precisely, define an augmented vector P 0 = (P1 , . . . , Pn , T (PAg )) ∈
[0, 1]n+1 , and define a function f 0 (P 0 ) = f (P1 , . . . , Pn ) = f (P ). Since T (PAg )
0
is assumed to be superuniform (since g ∈ Hgrp
is a null group), this means
0
that Pn+1 = T (PAg ) is superuniform, i.e. index n + 1 is a null p-value, in the
augmented vector of p-values P 0 . Then applying Lemma 1(c), with P 0 and
f 0 in place of P and f , and with index i = n + 1, yields the desired bound.
Lemma 3(a) is simply a special case of Lemma 3(b) since independence is
a special case of positive dependence, and conditioning on an independent
set of p-values P−Ag doesn’t change the distribution of PAg .
For Lemma 3(b), our proof strategy will be to reduce this statement
into a form where Lemma 1(b) becomes applicable. (Note that we cannot
simply take the approach of our proof of Lemma 3(c), because if we define
an augmented vector of p-values P 0 = P1 , . . . , Pn , Simesw (PAg ) , we do not
know if this vector is positively dependent—specifically, whether P 0 is PRDS
0
on entry Pn+1
= Simesw (PAg ).)
With this aim in mind, let b
kg ∈ {0, . . . , ng } be the number of discoveries
made by the BHw procedure when run on the p-values within group g at
level f (P ). Then, using the connection between the Simes test and the BH
42
RAMDAS, BARBER, WAINWRIGHT AND JORDAN
procedure, we may write
n
o
P
wi b
kg f (P )
n
o b
1
P
≤
i
kg
i∈Ag
ng
1 {Pg ≤ f (P )} = 1 b
kg > 0 = ····· = ··················································,
b
b
kg
kg
since for the BHw procedure at level f (P ), the ith p-value Pi will be rejected
if and only if Pi ≤
wi b
kg f (P )
.
ng
Hence, we may conclude that
n
o
n
o
wi b
kg f (P )
eg (P )
1
P
≤
1
P
≤
f
X
i
i
1 {Pg ≤ f (P )}
1
i∈Ag
ng
····························· = ·················································· =
wi ································,
f (P )
n
b
g
kg f (P )
feg (P )
i∈Ag
P
wb
k f (P )
where we have defined feg (P ) := i ng g . Taking expectations on both sides
and applying Lemma 1(b) immediately proves Lemma 3(b). (Specifically, we
know that P 7→ b
kg is a non-increasing function of P , and P 7→ f (P ) is also
assumed to be non-increasing; therefore, feg is also non-increasing in P .)
Given that Lemma 3(b) is proved, the proof of Lemma 3(d) follows exactly
the same argument as above, except that in the very last equation, Pi is
replaced by Si , and Lemma 3(b) is invoked in place of Lemma 1(b).
APPENDIX D: PROOF OF “MAXIMUM-CORNER” PROPOSITION 6
(m)
(m)
(m)
(m)
For each m, by definition of b
km , there is some k , . . . , k
,k
,...,k
1
m−1
m+1
M
such that
(D.1)
(m)
(m)
(m)
(m)
b.
(k1 , . . . , km−1 , b
km , km+1 , . . . , kM ) ∈ K
(m)
Thus, for each m0 6= m, b
km0 ≥ km0 by definition of b
km0 . Then
(m)
(m)
b
b b
b
b (m) , . . . , k (m) , b
bb
S(k
1
m−1 km , km+1 , . . . , kM ) ⊆ S(k1 , . . . , km−1 , km , km+1 , . . . , kM ) ,
b 1 , . . . , kM ) is a nondecreasing function of (k1 , . . . , kM ), and this
because S(k
immediately implies
(m)
(m)
(m)
(m)
Sbm (k1 , . . . , km−1 , b
km , km+1 , . . . , kM ) ⊆ Sbm (b
k1 , . . . , b
km−1 , b
km , b
km+1 , . . . , b
kM ).
Therefore, for each layer m,
X
u(m)
≥
g
g∈Sbm (b
k1 ,...,b
km )
X
(m)
g∈Sbm (k1
u(m)
≥b
km ,
g
(m)
(m)
(m)
,...,km−1 ,b
km ,km+1 ,...,kM )
where the second inequality holds by observation (D.1), and by definition of
b as the set of feasible vectors. Since this holds for all m, this proves that
K
b
(b
k1 , . . . , b
kM ) is itself a feasible vector, and hence (b
k1 , . . . , b
kM ) ∈ K.
43
UNIFIED MULTIPLE TESTING WITH PRIOR KNOWLEDGE
APPENDIX E: PROOF OF THEOREM 7.1
To be able to handle all four cases of the theorem, we define a function γm to
be the identity if we are not using reshaping (theorem statements (a,b)), or
γm = βm if we are using reshaping (theorem statements (c,d)). We also let
π
bm = 1 and λm = 1 if we are not using adaptivity (as in theorem statements
(b,c)), or let π
bm be defined as in (7.2) for theorem statements (a,d) where
adaptivity is used.
Fix any partition m. Since Pr{Pi = 0} = 0 for any i ∈ H0 by assumption,
we assume that Pi 6= 0 for any i ∈ H0 without further mention; this assump0 ,
tion then implies that if g ∈ Sbm (b
k1 , . . . , b
kM ) for some null group g ∈ Hm
b
we must have km > 0. We can then calculate
n
o
P
(m)
b
bm (b
1
g
∈
S
k
,
.
.
.
,
k
)
0 ug
1
M
g∈H
m
b
b
n
o
FDP(m)
u (k1 , . . . , kM ) = ··········································································
P
(m)
b
b
b
u
1
g
∈
S
(
k
,
.
.
.
,
k
)
m 1
M
g∈[Gm ] g
n
o
P
(m)
1 g ∈ Sbm (b
k1 , . . . , b
kM )
0 ug
g∈Hm
≤ ·········································································,
b
km
o
n
P
(m)
init (b
b
bm
k
,
.
.
.
,
k
)
u
1
g
∈
S
0
g
1
M
g∈Hm
≤ ···········································································,
b
km
n
o
P
(m)
km )
m ≤ min{w (m) αm γm (b
u
1
P
,
λ
}
0
g
g
m
g
g∈Hm
π
bm Gm
= ·······························································································,
b
km
b the
where the first inequality follows by definition (7.7) of the feasible set K,
init
second follows since Sbm (~k) ⊆ Sbm (~k) for any ~k by definition, and the last
init (~
step uses the definition of Sbm
k) in (7.3a) (without reshaping, for theorem
statements (a,b)) or (7.3b) (with reshaping, for theorem statements (c,d)).
(m)
Multiplying the numerator and denominator of each term by
taking expectations on both sides, we deduce that
αm wg
Gm
, and
(E.1)
FDR(m)
≤
u
αm
Gm
(m)
wg αm γm (b
km )
m
, λm }
X
π
bm Gm
1 Pg ≤ min{
(m) (m)
.
ug wg E ······································································
(m)
0
g∈Hm
wg
αm b
km
Gm
With these calculations in place, we now prove the four statements of the
theorem.
44
RAMDAS, BARBER, WAINWRIGHT AND JORDAN
(m)
Theorem statement (a). Define the function fg
that maps the vector P
Note that
is a nonincreasing function of P , since b
km is
to
a nonincreasing function of P by definition of our procedure, while π
bm is a
nondecreasing function of P .
Returning to (E.1), we then see
(m)
wg αm b
km
m
1 Pg ≤ min{ πbm Gm , λm }
αm X (m) (m)
FDR(m)
≤
u
w
E
·······························································
u
g
g
(m)
Gm
km
wg αm b
0
g∈Hm
π
bm πbm Gm
n
o
m ≤ f (m) (P )
1
P
X
g
m
g
αm
(m)
=
u(m)
g wg E 1 Pg ≤ λm · ·······································
(m)
Gm
π
bm fg (P )
0
g∈Hm
o
n
(m)
m
αm X (m) (m) 1 Pg ≤ fg (P )
(E.2)
ug wg E ·······································
,
≤
−g (m)
Gm
π
b
f
(P
)
0
m
g
g∈Hm
(m)
km
wg αm b
π
bm Gm .
(m)
fg
where
(E.3)
−g
π
bm
:=
|u(m) w(m) |∞ +
P
h6=g
n
o
(m) (m)
(m)
uh wh 1 Ph > λm
n(1 − λm )
.
−g
Note that, on the event Pgm ≤ λm , we have π
bm = π
bm
, which allows the
inequality (E.2) above. Returning to (E.2), we now condition on P−Am
for
g
each group g:
o
n
m ≤ f (m) (P )
1
P
X
g
g
αm
(m)
u(m)
FDR(m)
≤
g wg E ·······································
u
−g (m)
Gm
π
bm fg (P )
0
g∈Hm
o
n
m ≤ f (m) (P )
1
P
X
g
g
αm
(m)
u(m)
P−Am
=
g wg E E ·······································
g
(m)
−g
Gm
π
bm fg (P )
0
g∈Hm
o
n
m ≤ f (m) (P )
1
P
X
g
g
1
(i) αm
(m)
=
u(m)
P−Am
g wg E
−g E ·······································
g
(m)
Gm
π
b
m
fg (P )
0
g∈Hm
αm X (m) (m)
1
≤
ug wg E −g ,
Gm
π
bm
g∈H0
m
−g
where equality (i) holds because π
bm
is a function of only the p-values outside
of group g, i.e. of P−Am
,
while
the
last
inequality holds by Lemma 3(a).
g
45
UNIFIED MULTIPLE TESTING WITH PRIOR KNOWLEDGE
Finally, observe that independence between the different groups of partition m implies that the indicator variables 1 {Phm > λm } are independent
Bernoulli’s with probabilities ≤ 1 − λm of success. Thus, as a consequence
of Lemma 2, we can prove (analogous to property (5.5)) that
1
Gm
(E.4)
E −g ≤ P (m) (m) .
π
bm
uh wh
0
h∈Hm
Plugging this back into our bounds on FDR, we finally obtain
αm X (m) (m)
1
(m)
FDRu ≤
ug wg E −g
Gm
π
bm
0
g∈Hm
≤
Gm
αm X (m) (m)
ug wg
P
(m) (m)
Gm
uh wh
0
g∈Hm
0
h∈Hm
≤ αm .
Theorem statement (b). The proof of statement (b) follows the same steps
as for (a), but without the need to condition on P−Am
, since we do not
g
w
(m)
(m)
α b
k
(m)
is a
use adaptivity. Define the function fg (P ) = g Gmm m . Then fg
nonincreasing function of P , since b
km is a nonincreasing function of P .
Returning to (E.1), as in the proof of statement (a), we calculate
o
n
m ≤ f (m) (P )
1
P
X
g
g
α
m
(m)
.
FDR(m)
≤
u(m)
u
g wg E ·······································
(m)
Gm
f
(P
)
0
g
g∈Hm
" n
o#
(m)
1 Pgm ≤fg
(P )
By Lemma 3(b), we know that E ·····························
≤ 1, and therefore
(m)
fg
FDR(m)
≤
u
(P )
αm X (m) (m)
ug wg ,
Gm
0
g∈Hm
as desired.
Theorem statement (c).
We now turn to proving the method under reshap(m)
w
α
(m)
b
ing. Define
= km , and define constant cg = gGm m . Returning
to (E.1), as before, we calculate
n
o
(m)
m ≤ c(m) · β
f
(P
)
1
P
X
g
m g
g
αm
(m)
.
FDR(m)
≤
u(m)
u
g wg E ····························································
(m)
(m)
Gm
cg · fg (P )
0
g∈Hm
(m)
fg (P )
46
RAMDAS, BARBER, WAINWRIGHT AND JORDAN
"
o #
n
(m)
(m)
1 Pgm ≤cg ·βm fg (P )
≤ 1 since Pgm is
By Lemma 3(c), we know that E ·············································
(m) (m)
cg
·fg
(P )
0 . Therefore,
assumed to be superuniform for any null group g ∈ Hm
αm X (m) (m)
FDR(m)
≤
ug wg .
u
Gm
0
g∈Hm
Theorem statement (d). The proof of part (d) combines the calculations of
part (a) (where adaptivity is used) with part (c) (where reshaping is used).
(m)
w
α
(m)
(m)
−g
bm
is defined as in (E.3) from
Define fg = b
km and cg = g−g m , where π
π
bm Gm
(m)
cg
part (a). Note that
is no longer a constant, but nonetheless, proceeding
as in part (a), we can calculate
n
o
(m)
m ≤ c(m) · β
1
P
f
(P
)
X
g
g
m
g
αm
(m)
.
FDR(m)
≤
u(m)
u
g wg E ····························································
(m)
(m)
−g
Gm
π
bm · cg · fg (P )
0
g∈Hm
Next we condition on the p-values outside the group Am
g :
n
o
(m)
m ≤ c(m) · β
f
(P
)
1
P
X
g
g
m
g
αm
(m)
FDR(m)
≤
P−Am
u(m)
u
g wg E E ····························································
g
(m)
(m)
−g
Gm
π
b
·
c
·
f
(P
)
0
m
g
g
g∈Hm
n
o
(m)
(m)
m ≤c
1
P
·
β
f
(P
)
X
g
g
m
g
αm
(m) 1
,
=
u(m)
P−Am
g wg E
−g · E ····························································
g
(m)
(m)
Gm
π
b
m
c
·
f
(P
)
g
g
g∈H0
m
−g
where the last step holds since π
bm
is a function of P−Am
.
g
Finally, we will apply Lemma 3(c) to show that each of these conditional
expected values is ≤ 1. Of course, the subtlety here is that we must condition
(m)
on P−Am
. To do so, note that, after fixing P−Am
, the function fg (P ) can
g
g
be regarded as a function of only the remaining unknowns (i.e. of PAm
), and
g
(m)
is still non-increasing; the value cg is now a constant; and Pgm = Tgm (PAm
)
g
m
is indeed superuniform since, due to the independence of PAm
from
P
−Ag ,
g
its distribution has not changed. Therefore, we can apply Lemma 3(c) (with
in place of P , while P−Am
is treated as constant),
the random vector PAm
g
# g
" n
o
(m)
1 Pgm ≤cg
(m)
·βm fg
(P )
to see that E ·············································
(m) (m)
cg
·fg
FDR(m)
u
(P )
P−Am
≤ 1, and therefore,
g
αm X (m) (m)
1
≤
ug wg E −g .
Gm
π
bm
0
g∈Hm
47
UNIFIED MULTIPLE TESTING WITH PRIOR KNOWLEDGE
−g
Finally, we need to bound π
bm
. As in the proof of part (a), we see that the indicator variables 1 {Phm > λm } are independent, since Phm = Thm (PAm
), and
h
the sets of p-values PAm
are
assumed
to
be
independent
from
each
other.
h
m
Furthermore, since Th (PAm
) is assumed to be a valid p-value, i.e. superunih
0
form for any h ∈ Hm , this means that the variable 1 {Phm > λm } is Bernoulli
with chance ≥ 1 − λm of success. Therefore, the bound (E.4) calculated in
the proof of part (a) holds here as well, and so
FDR(m)
≤
u
αm X (m) (m)
Gm
ug wg
= αm .
P
(m) (m)
Gm
uh wh
0
g∈Hm
0
h∈Hm
This concludes the proof of all four parts of Theorem 7.1.
APPENDIX F: PROOF OF PROPOSITION 7
(s)
(s)
First we introduce some notation: let (k1 , . . . , kM ) be the vector after
(s)
the sth pass through the algorithm. We prove that km ≥ b
km for all m, s,
(0)
b
by induction. At initialization, km = Gm ≥ km for all m. Now suppose that
(s−1)
(s)
km
≥b
km for all m; we now show that km ≥ b
km for all m.
To do this, consider the m-th layer of the s-th pass through the algorithm.
(s)
(s)
(s−1) (s−1)
(s−1)
Before this stage, we have vectors k1 , . . . , km−1 , km , km+1 , . . . , kM ,
(s)
and we now update km . Applying induction also to this inner loop, and
(s)
assuming that km0 ≥ b
km0 for all m0 = 1, . . . , m − 1, we can now prove that
(s)
km ≥ b
km . By definition of the algorithm,
(F.1)
(s)
km
0
= 0 max
km
:
km ∈{0,1,...,Gm }
X
bm (k(s) ,...,k(s) ,k0 ,k(s−1) ,...,k(s−1) )
g∈S
1
m−1 m m+1
M
0
u(m)
≥ km
g
.
(s−1)
(s)
km0 for all m0 = 1, . . . , m − 1, and km0
≥b
km0 for all m0 =
Since km0 ≥ b
m + 1, . . . , M , we have
X
X
u(m)
u(m)
≥
g
g
(s)
(s)
(s−1)
(s−1)
g∈Sbm (k1 ,...,km−1 ,b
km ,km+1 ,...,kM
)
g∈Sbm (b
k1 ,...,b
km−1 ,b
km ,b
km+1 ,...,b
kM )
since Sbm (~k) is a nondecreasing function of ~k by definition. The right-hand
side of this expression is in turn ≥ b
km by definition of (b
k1 , . . . , b
kM ) being a
b
feasible vector. Therefore, km is in the feasible set for Eq. (F.1), and so we
(s)
must have km ≥ b
km . By induction, this is then true for all s, m, as desired.
48
RAMDAS, BARBER, WAINWRIGHT AND JORDAN
(s)
(s)
Now suppose that the algorithm stabilizes at (k1 , . . . , kM ), after s full
passes. After completing the mth layer of the last pass through the algo(s)
(s) (s−1)
(s−1)
rithm, we had vectors k1 , . . . , km , km+1 , . . . , kM ; however, since the al(s−1)
(s)
gorithm stops after the sth pass, this means that km0
= km0 for all m0 .
(s)
Using this observation in the definition of km , we see that
X
(s)
u(m)
≥ km
.
g
(s)
(s)
(s)
(s)
(s)
g∈Sbm (k1 ,...,km−1 ,km ,km+1 ,...,kM )
(s)
(s)
(s)
b and so km
This means that (k1 , . . . , kM ) ∈ K,
≤ b
km for all m by the
definition of b
k1 , . . . , b
km and Proposition 6. But by the induction above, we
(s)
also know that km ≥ b
km for all m at any iteration s; and this completes
the proof.
APPENDIX G: PROPERTIES OF DOTFRACTIONS
In this section we verify that “dotfractions” satisfy many of the same
properties as ordinary fractions, and thus the notation ··ab can be safely used
throughout the proofs of our main results. In all the following, the property
will be shown to hold assuming that all dotfractions appearing in its equation or inequality are well defined. Hence, throughout, we assume that the
various properties are only used if all of the dotfractions in the expression
are defined—that is, we may use these properties only if we never have ··ab
with a 6= 0 and b = 0. As a side note, observe that in the paper, we always
use ··ab when a, b ≥ 0 only.
1. Comparing two fractions:
a
b
c
c
If a ≥ b ≥ 0 and c ≥ 0, then ··· ≥ ··, and ··· ≤ ··.
c
c
a
b
To prove the first bound, if c > 0 then this reduces to ac ≥ cb , while if
c = 0 then we must have a = b = 0 (since, otherwise, ··ac and ··cb would
be undefined) and so ··ac = ··cb = 0. To prove the second bound, if b > 0,
then this reduces to ac ≤ cb , while if b = 0 then we must have c = 0
(since, otherwise, ··cb would be undefined), in which case ··ac = ··cb = 0.
2. Comparing against a scalar:
(G.1)
b
If c ≥ 0 and a ≥ ·· then ac ≥ b.
c
To prove this, if c 6= 0 then we have a ≥ cb , while if c = 0 then we must
have b = 0 (so that ··cb is not undefined), and so ac ≥ b is trivially true
as both sides equal zero.
(G.2)
UNIFIED MULTIPLE TESTING WITH PRIOR KNOWLEDGE
49
3. Adding numerators:
(G.3)
For any a, b, c,
a b
a+b
··· + ·· = ··········.
c c
c
To prove this, if c 6= 0 then this reduces to ac + cb = a+b
c , while if
c = 0 then we must have a = b = 0 (otherwise the dotfractions are
undefined), and so the left- and right-hand sides both equal zero.
4. Multiplying fractions:
(G.4)
For any a, b, c, d,
a c
ac
··· · ··· = ·····.
b d
bd
To prove this, if b, d 6= 0 then this reduces to ab · dc = ac
bd , while if b = 0
or d = 0, then either a = 0 or c = 0 (otherwise ··ab or ··dc would be
undefined), and so the left- and right-hand sides both equal zero.
5. Cancelling nonzero factors :
(G.5)
If c 6= 0 then for any a, b,
ac
a
····· = ···.
bc
b
To see why, we simply apply (G.4) with d = c (noting that, with the
assumption c 6= 0, we have ··cc = 1).
6. Multiplying by a scalar:
(G.6)
For any a, b, c,
a
ac
c · ··· = ·····.
b
b
To see why, if b 6= 0 then this reduces to c · ab = ac
b , while if b = 0 then
we must have a = 0 so that ··ab is not undefined, and so the left- and
right-hand sides are both zero.
While the above properties all carry over from fractions to dotfractions,
there are some settings where familiar manipulations with fractions may no
ac
longer be correct. For example, ··ab =
6 ····
bc when a, b 6= 0 while c = 0. Relatedly,
we cannot add fractions in the usual way, i.e. ··ab + ··dc may not be equal to
ad+bc
ad
··········;
this fails because implicitly we would be assuming that ··ab = ····
bd
bd and
c
bc
··d = ····
in
order
to
make
the
two
denominators
the
same,
which
may
fail if
bd
d = 0 or if b = 0.
REFERENCES
[1] Rina Foygel Barber and Aaditya Ramdas. The p-filter: multilayer false discovery
rate control for grouped hypotheses. Journal of the Royal Statistical Society: Series
B (Statistical Methodology), 2016.
[2] Yoav Benjamini and Marina Bogomolov. Selective inference on multiple families of
hypotheses. Journal of the Royal Statistical Society: Series B (Statistical Methodology), 76(1):297–318, 2014.
50
RAMDAS, BARBER, WAINWRIGHT AND JORDAN
[3] Yoav Benjamini and Yosef Hochberg. Controlling the false discovery rate: a practical
and powerful approach to multiple testing. Journal of the Royal Statistical Society:
Series B (Statistical Methodology), 57(1):289–300, 1995.
[4] Yoav Benjamini and Yosef Hochberg. Multiple hypotheses testing with weights. Scandinavian Journal of Statistics, 24(3):407–418, 1997.
[5] Yoav Benjamini and Wei Liu. A distribution-free multiple test procedure that controls
the false discovery rate. Tel Aviv, Tel Aviv University, 1999.
[6] Yoav Benjamini and Wei Liu. A step-down multiple hypotheses testing procedure that
controls the false discovery rate under independence. Journal of Statistical Planning
and Inference, 82(1):163–170, 1999.
[7] Yoav Benjamini and Daniel Yekutieli. The control of the false discovery rate in
multiple testing under dependency. The Annals of statistics, 29(4):1165–1188, 2001.
[8] Yoav Benjamini, Abba Krieger, and Daniel Yekutieli. Adaptive linear step-up procedures that control the false discovery rate. Biometrika, 93:491–507, 2006.
[9] Gilles Blanchard and Etienne Roquain. Two simple sufficient conditions for FDR
control. Electronic Journal of Statistics, 2:963–992, 2008.
[10] Gilles Blanchard and Étienne Roquain. Adaptive false discovery rate control under
independence and dependence. Journal of Machine Learning Research, 10(Dec):2837–
2871, 2009.
[11] Gunnar Eklund. Massignifikansproblemet. In Unpublished seminar papers, Uppsala
University Institute of Statistics, volume 1963, 1961.
[12] Gunnar Eklund and Paul Seeger. Massignifikansanalys. Statistisk tidskrift, 5:355–365,
1965.
[13] Livio Finos and Luigi Salmaso. FDR-and FWE-controlling methods using data-driven
weights. Journal of Statistical Planning and Inference, 137(12):3859–3870, 2007.
[14] Ruben Gabriel. Simultaneous test procedures–some theory of multiple comparisons.
The Annals of Mathematical Statistics, pages 224–250, 1969.
[15] Yulia Gavrilov, Yoav Benjamini, and Sanat K Sarkar. An adaptive step-down procedure with proven FDR control under independence. The Annals of Statistics, pages
619–629, 2009.
[16] Christopher Genovese, Kathryn Roeder, and Larry Wasserman. False discovery control with p-value weighting. Biometrika, 93(3):509–524, 2006.
[17] Nicholas Heard and Patrick Rubin-Delanchy. Choosing between methods for combining p-values. arXiv preprint arXiv:1707.06897, 2017.
[18] Philipp Heesen and Arnold Janssen. Dynamic adaptive multiple tests with finite
sample FDR control. Journal of Statistical Planning and Inference, 168:38–51, 2016.
[19] Yosef Hochberg and Uri Liberman. An extended Simes’ test. Statistics & Probability
Letters, 21(2):101–105, 1994.
[20] G Hommel. Tests of the overall hypothesis for arbitrary dependence structures.
Biometrische Zeitschrift, 25(5):423–430, 1983.
[21] James Hu, Hongyu Zhao, and Harrison Zhou. False discovery rate control with groups.
Journal of the American Statistical Association, 105(491), 2010.
[22] Samuel Karlin and Yosef Rinott. Classes of orderings of measures and related correlation inequalities. I. multivariate totally positive distributions. Journal of Multivariate
Analysis, 10(4):467–498, 1980.
[23] Eugene Katsevich and Chiara Sabatti. Multilayer knockoff filter: Controlled variable
selection at multiple resolutions. arXiv preprint arXiv:1706.09375, 2017.
[24] Erich Leo Lehmann. Some concepts of dependence. The Annals of Mathematical
Statistics, pages 1137–1153, 1966.
[25] D Morgenstern. Berechnung des maximalen signifikanzniveaus des testes lehneh0 ab,
UNIFIED MULTIPLE TESTING WITH PRIOR KNOWLEDGE
51
wennk untern gegebenen tests zur ablehnung führen. Metrika, 27(1):285–286, 1980.
[26] Joseph Romano and Azeem Shaikh. On stepdown control of the false discovery
proportion. In Optimality, pages 33–50. Institute of Mathematical Statistics, 2006.
[27] Joseph Romano, Azeem Shaikh, and Michael Wolf. Consonance and the closure
method in multiple testing. The International Journal of Biostatistics, 7(1):1–25,
2011.
[28] B Rüger. Das maximale signifikanzniveau des tests:lehneh o ab, wennk untern gegebenen tests zur ablehnung führen. Metrika, 25(1):171–178, 1978.
[29] Ludger Rüschendorf. Random variables with maximum sums. Advances in Applied
Probability, pages 623–632, 1982.
[30] Sanat Sarkar. On methods controlling the false discovery rate. Sankhyā: The Indian
Journal of Statistics, Series A (2008-), pages 135–168, 2008.
[31] Sanat Sarkar. Two-stage stepup procedures controlling FDR. Journal of Statistical
Planning and Inference, 138(4):1072–1084, 2008.
[32] Paul Seeger. A note on a method for the analysis of significances en masse. Technometrics, 10(3):586–593, 1968.
[33] John Simes. An improved bonferroni procedure for multiple tests of significance.
Biometrika, 73(3):751–754, 1986.
[34] Eckart Sonnemann. Allgemeine Lösungen multipler Testprobleme. Universität Bern.
Institut für Mathematische Statistik und Versicherungslehre, 1982.
[35] Eckart Sonnemann. Allgemeine lösungen multipler testprobleme. edv in medizin
und biologie 1982, 13 (4): 120–128. english version of the original article with minor
corrections by finner h: General solutions to multiple testing problems. Biom J, 50
(5):641–656, 2008.
[36] Eckart Sonnemann and Helmut Finner. Vollständigkeitssätze für multiple testprobleme. In Multiple Hypothesenprüfung/Multiple Hypotheses Testing, pages 121–135.
Springer, 1988.
[37] John Storey. A direct approach to false discovery rates. Journal of the Royal Statistical Society: Series B (Statistical Methodology), 64(3):479–498, 2002.
[38] John Storey, Jonathan Taylor, and David Siegmund. Strong control, conservative
point estimation and simultaneous conservative consistency of false discovery rates:
a unified approach. Journal of the Royal Statistical Society: Series B (Statistical
Methodology), 66(1):187–205, 2004.
[39] Samuel A Stouffer, Edward A Suchman, Leland C DeVinney, Shirley A Star, and
Robin M Williams Jr. The american soldier: Adjustment during army life.(studies in
social psychology in world war ii), vol. 1. 1949.
[40] John Wilder Tukey. The Problem of Multiple Comparisons: Introduction and Parts
A, B, and C. Princeton University, 1953.
[41] Vladimir Vovk. Combining p-values via averaging. arXiv preprint arXiv:1212.4966,
2012.
[42] Daniel Yekutieli. Hierarchical false discovery rate–controlling methodology. Journal
of the American Statistical Association, 103(481):309–316, 2008.
Aaditya Ramdas
Departments of Statistics and EECS
University of California, Berkeley
E-mail: [email protected]
Rina Foygel Barber
Department of Statistics
University of Chicago
E-mail: [email protected]
Martin J. Wainwright
Departments of Statistics and EECS
University of California, Berkeley
E-mail: [email protected]
Michael I. Jordan
Departments of Statistics and EECS
University of California, Berkeley
E-mail: [email protected]
| 10math.ST
|
Cooperative Robust Estimation with Local Performance Guarantees
arXiv:1603.05307v1 [cs.SY] 16 Mar 2016
M. Zamani
Abstract— The paper considers the problem of cooperative
estimation for a linear uncertain plant observed by a network of
communicating sensors. We take a novel approach by treating
the filtering problem from the view point of local sensors while
the network interconnections are accounted for via an uncertain
signals modelling of estimation performance of other nodes.
That is, the information communicated between the nodes is
treated as the true plant information subject to perturbations,
and each node is endowed with certain believes about these
perturbations during the filter design. The proposed distributed
filter achieves a suboptimal H∞ consensus performance. Furthermore, local performance of each estimator is also assessed
given additional constraints on the performance of the other
nodes. These conditions are shown to be useful in tuning the
desired estimation performance of the sensor network.
I. I NTRODUCTION
The research on cooperative filtering and estimation of
networked systems has gained much momentum during
the past decade, aiming at developing efficient estimation
algorithms for large assemblies of networked sensors [1]–
[5]. The mentioned references reflect the common trend in
the literature, where the main objective is to accomplish a
globally optimal or suboptimal estimation performance of the
network. Usually, the performance of individual sensors is
not considered in such problems. This observation motivates
the question about a relationship between the estimation
performance of the individual filters within a distributed
estimation network and the performance of the overall network. This paper considers this problem within the specific
framework of distributed H∞ consensus estimation [4], [5].
Our approach also targets the global convergence problem
however not with a brute-force decoupling of the global
solution. Rather we define a local objective function in
terms of uncertain signals capturing the performance of the
other nodes. This leads to decoupling of the distributed filter
while implicitly maintaining a meaningful connection to the
network. This way, the local objective function abstracts the
dependence on other nodes, eliminating the need to consider
their exact models or their raw measurements. Nevertheless,
the convergence of the local filter is dependent on the rest of
the network and we provide conditions that render the H∞
convergence of the network of the filters. Furthermore, by
asserting further conditions on the individual performances
of the other nodes a guaranteed H∞ performance of the
This work was supported by the Australian Research Council under
Discovery Projects funding scheme (project DP120102152). This work was
done while the first author was with the School of Engineering and IT,
UNSW Canberra, Canberra, Australia.
M. Zamani is with the DST Group Australia and V. Ugrinovskii is with
School of Engineering and IT, UNSW Canberra, Canberra, Australia, Email:
{m.zammani,v.ugrinovskii}@gmail.com
V. Ugrinovskii
individual filters is established. These conditions express that
if all the neighbours maintain a certain level of accuracy then
the local filter also guarantees a nominated H∞ performance.
To establish the above relationship, here we analyze the
distributed filter network consisting of estimators solving an
auxiliary optimal filtering problem at every node. In that
sense, our approach bears some resemblance with decentralized control where each controller is constructed to regulate a
local subsystem. The mentioned auxiliary filtering problem
originates in [6], [7]; it was shown in [8] to yield interconnected consensus-type filters that exchange information
between the network nodes although the parameters for each
filter could be computed online in the decentralized fashion.
The new element of this paper compared with [8] is how
the neighboring information is interpreted by each node.
In [8], each node was considered to be agnostic about the
amount of energy in the error between the true state of the
plant and the neighbours’ estimates of that state. In contrast,
here we consider a model where each node perceives a
relationship between the energy in the neighbours’ error and
the accuracy of its own filter. We give a detailed discussion
of this idea later in the paper; for now we only note that
technically our model adds a constraint on the energy in the
error inputs arising in the auxiliary minimum energy problems. Such a constraint has the form of an Integral Quadratic
Constraint previously used in robust decentralized control
problems and filtering problems; e.g., see [9], [10]. However,
unlike those problems, the parameters of the constraints used
here play the role of tunable parameters which are adjusted
according to the desired local and global performance. They
also serve as indicators of sensitivity of the individual filters
to the neighbours’ performance.
The main result of this paper are sufficient conditions on
the network parameters that ensure H∞ performance of the
network consisting of the proposed minimum energy filters.
As mentioned, not only global disturbance attenuation is
guaranteed by these conditions, but also certain local H∞
properties of the node filters are established. We show that
these conditions admit the form of a convex semidefinite
program, which enables constructing a filter network yielding
a suboptimal disturbance attenuation.
Notation: Rn is the Euclidean space of vectors, k · k is
the Euclidean norm, and for any positive semidefinite matrix
X, X = X ′ ≥ 0, kakX , (a′ Xa)1/2 . For 0 < T ≤ ∞,
L2 [0, T ) denotes the Lebesgue space of vector-valued signals
square-integrable on [0, T ). diag[X1 , . . . , XN ] denotes the
block diagonal matrix with X1 , . . . , XN as its diagonal
blocks, and ⊗ is the Kronecker product of matrices. λmin (Z)
is the smallest eigenvalue of a symmetric matrix Z.
II. P ROBLEM F ORMULATION
AND
P RELIMINARIES
A. The plant and the distributed estimator
Consider a linear system
ẋ = Ax + Bw,
n
x(0) = x0 ,
(1)
m
where x ∈ R and w ∈ R are, respectively, the state
and the unknown modeling disturbance input; the latter is
assumed to be L2 integrable on [0, ∞). The matrices A ∈
Rn×n and B ∈ Rn×m are known, however the initial state
x0 is unknown and is considered to be part of the uncertainty
about the system (1).
The main objective of the paper is to determine conditions
under which the plant state x(t) can be estimated by a
network of filters each using its plant measurement
yi = Ci x + Di vi ,
(2)
where i = 1, 2, . . . , N indicates the measurement taken at
node i of the network. Each measurement yi ∈ Rpi is
imperfect, it is subject to a measurement disturbance vi
taking values in Rmi that also belongs to the space L2 [0, ∞)
by assumption. The coefficients of each measured output
are matrices of the matching dimensions, Ci ∈ Rpi ×n ,
Di ∈ Rpi ×mi , with Ei , Di Di′ > 0.
In addition to its direct measurements of the plant, each
node receives information from other nodes of the network,
of the form
cij = Wij x̂j + Fij ǫij ,
(3)
where x̂j is the estimate of state x at the neighbouring
node j. The signal ǫij with values in Rmij represents the
communication errors or uncertainty in the communication
channel, ǫij ∈ L2 [0, ∞). We assume that Gij , Fij Fij′ > 0.
The network graph describing communications between
the filtering nodes is assumed to be directed, its node and
edge sets are denoted V = {1, . . . , N } and E ⊆ V × V,
respectively. The neighborhood of node i, i.e., the set of
nodes which send information to node i, is denoted by Ni =
{j : (i, j) ∈ E} and its cardinality is denoted li . The Laplace
matrix of the network graph is denoted L [11].
Following [1], [2], [4] and many other papers on distributed estimation, we consider a class of consensus-based
interconnected filters each processing the direct measurements yi and neighbours’ information cij by means of a
Luenberger-type observer of the form
X
x̂˙ i = Ax̂i + Li (yi − Ci x̂i ) +
Kij (cij − Wij x̂i ), (4)
j∈Ni
x̂i (0) = ξi .
The estimation problem in this paper is to determine
coefficients Li , Kij (which can be time-varying) that ensure
convergence of the network to trajectories of the plant, and
also guarantee an acceptable attenuation of the detrimental
effects of disturbances on the estimation error. Formally,
these properties are formulated as follows. Given a positive
semidefinite matrix P ∈ RnN ×nN and a collection of
positive semidefinite matrices Xi ∈ Rn×n , and constants γ 2
and γ̄i2 , i ∈ V, we wish to determine a collection of filters
of the form (4) that guarantee the following properties:
P1. In the absence of disturbances w, vi and ǫij , j ∈ Ni ,
i = 1, . . . , N the estimation error of the filter ei (t) =
x̂i (t) − x(t) converges to zero asymptotically.
P2. In the presence of disturbances, the network of filters (4)
attains the type of H∞ disturbance attenuation property
Z ∞
N
X
kx(0) − ξi k2Xi + N kwk22
kek2P dt ≤ γ 2
0
i=1
!
N
X
X
2
2
kvi k2 +
kǫij k2 , (5)
+
i=1
j∈Ni
where e = [(x̂1 − x) , . . . , (x̂N − x)′ ]′ and k.k22 is the
L2 norm.
P3. Provided the neighbours of node i contribute a sufficient
effort (this will be quantitatively defined later) to assist
i, it is also guaranteed that at that node
Z ∞
h
kei k2 ds ≤ γ̄i2 βi + kx0 − ξi k2Xi
0
Z ∞
X
i
kwk2 + kvi k2 +
+
kǫij k2 ds ; (6)
′
0
j∈Ni
βi > 0 is a constant which will be determined later.
These properties formalize the desired attributes of a
distributed filter that we want to achieve. In particular, property P2 specifies the desired global disturbance attenuation
performance across the sensor network using a network of
decoupled filter equations (4). Note that decoupled equations governing the gains in filters (4) will be provided
later. Furthermore, property P3 articulates the desired local disturbance attenuation provided that there is sufficient
contribution from the neighbours. The sufficient contribution
condition is quantitatively defined later in the paper.
We remark that properties P1, P2 jointly generalize the
property of H∞ consensus introduced in [4]; also see [5],
[12]. For example, let P = (L + L⊤ ) ⊗ P0 where P0 =
P0′ ≥ 0, and L⊤ is the Laplacian matrix of the graph
obtained from the network graph by reversing its edges.
This choice of P results in the left hand side of (5) being
equal toR thePweighted
H∞ disagreement cost between the
P
∞
2
kx̂
nodes, 0
i − x̂j kP0 ds [4], [5], [13]. More
i
j∈Ni
generally, letting P = (L + L⊤ ) ⊗ P0 + diag[P1 . . . PN ],
Pi = Pi′ > 0, reduces P1, P2 to the property of strong robust
synchronization introduced in [12]. In addition, property
P3 describes H∞ attenuation properties of individual node
filters. Including such property into analysis constitutes the
main difference between the problem posed above and the
previous work in the area of distributed estimation.
B. Representation of the neighboring information
To make performance analysis of individual filters possible, let us introduce the mismatch between the disturbancefree information contained in the signal cij and the corresponding true version of this information,
ηij = Wij (x̂j − x) = −Wij ej ∈ Rpij ,
j ∈ Ni .
(7)
PSfrag replacements
With these signals the information received by sensor i can
be represented as
cij = Wij x + ηij + Fij ǫij ,
j ∈ Ni .
j∈Ni
+Li Di vi +
Kij (ηij + Fij ǫij ),
(9)
j∈Ni
ηki = Wki ei ,
i ∈ Nk .
In this system, the signals ηij , j ∈ Ni , play the role
of exogenous disturbances and each signal ηki = Wki ei
represents the output used by agent k for whom i is the
neighbour, i.e., i ∈ Nk . This interpretation allowed us to
construct in [8] minimum energy filters of the form (4) with
the property that for any initial condition x0 , arbitrary L2 integrable disturbances w, vi , ǫij , ηij and an arbitrary T > 0
Z Th
Z T
kwk2 + kvi k2
kei k2Ri dt ≤ γ 2 kx0 − ξi k2Xi +
0
0
i
X
2
2
+
(kǫij k + kηij kZ −1 ) dt . (10)
j∈Ni
ei
ηki
+
(8)
Equation (8) can be regarded as an additional measurement
of the plant affected by disturbances ǫij and ηij .
Treating the signals ηij , for the purpose of filter derivation,
as the disturbances additional to w, vi and ǫij has an effect
of decoupling node i from its neighbours. Indeed, consider
the error dynamics of the filter (4) at node i,
X
ėi = A − Li Ci −
Kij Wij ei − Bw
X
Fij ǫij
w
vi
ij
Here, γ 2 and Ri = Ri′ > 0 are a positive constant and
matrices whose existence is determined by certain LMI
′
conditions in [8]. Also, Zij = Zij
> 0 are given matrices;
in [8] they were associated with the confidence of node i
about performance of node j.
Condition (10) provides an H∞ type bound on the energy
in the filter estimation errors at node i expressed in terms
of the energy of the disturbances affecting that node, and
is similar to (6) in property P3. The important difference
between (10) and (6) is that the former condition includes
the energy in the signals ηij that depend on the neighbours’
accuracy. Also, according to (10), the same level of disturbance attenuation γ 2 is stated for all nodes. Our goal is
to revisit the design of the filters (4) to obtain a possibly
sharper H∞ property for at least some of the local filters,
and for other filters, to provide a means for assessing their
local performance and sensitivity to the neighbours’ errors.
Owing to the relation ηij = −Wij ej , from the viewpoint
of node i, the error dynamics of the network can be seen as
an interconnection of two systems, representing, respectively,
i’s own error dynamics and the errors dynamics of the rest
of the system; see Figure 1. Motivated by (10), we propose
the following condition to formally capture the sensitivity of
each node to the accuracy of its neighbours’ filters:
For every i, there exist positive definite symmetric matrices
Z̄ij and constants dij ≥ 0, j ∈ Ni , such that for all t ∈
Fki ǫki
ηij
Fig. 1.
Other
subsystems
+
vj , ǫkj j 6= i
w
A two-block representation of the error dynamics system.
[0, ∞) and w, vi , ǫij ∈ L2 [0, t),
Z t
Z t
2
kηij kZ̄ −1 dt ≤
(kei k2 + kwk2 )dt + dij ,
0
ij
(11)
0
∀j ∈ Ni , i = 1, . . . , N.
As a generalized form of the property (10), condition (11)
reflects how the neighbours’ accuracy influences the local
disturbance attenuation property (6) at every node. Therefore
in what follows, we will use condition (11) to establish (6),
i.e., (11) is the quantitative characteristic of the neighbours’
effort mentioned in P3.
To demonstrate the role of (11) more vividly, take for
example Z̄ij = z̄ij I, with a scalar z̄ij > 0 and suppose (6)
holds provided (11) is satisfied with a very small z̄ij . Since
according to (6), the energy in ei is bounded, (11) suggests
that node i can only tolerate relatively ‘small’ mismatch
inputs ηij to be able guarantee (6). However, a small energy
in ηij can only be accomplished by the corresponding
neighbour j. This suggests that the eigenvalues of Z̄ij may
be indicative of sensitivity of the local filters to fidelity of its
neighbours’ estimates. In Section IV we will show that the
matrices Z̄ij can be computed jointly with the attenuation
levels γ 2 , γ̄i2 . This provides the means for performance
tuning of the local filters.
Similarly, the constants dij in (11) describe the bound
on the estimation error energy that node i is prepared to
tolerate from its neighbour j, in response to (hypothetically)
estimating the perfectly known plant (w = 0) with the
utmost precision (ei = 0). Indeed, in this hypothetical case,
condition (11) reduces to a bound on the energy in the
mismatch disturbance signal ηij of the neighbour j.
C. Distributed estimation problem
We are now in a position to present a formal definition of
the distributed estimation problem described in Section II-A.
Problem 1: Determine a collection of filters of the form
(4) and matrices Z̄ij ∈ Rpij ×pij , j ∈ Ni , i ∈ V, and
constants γ 2 and γ̄i2 such that the following conditions hold:
(i) Given a positive semidefinite matrix P ∈ RnN ×nN ,
the network of filters (4) achieves properties P1 and
P2 with this P and the found γ 2 .
(ii) The following implication holds with the found Z̄ij and
γ̄i2 : If signals ηij (t), j ∈ Ni , satisfy (11), then the filter
(4) guarantees the satisfaction of condition (6), i.e., P3
is satisfied.
We stress that the global performance properties P1 and P2
of the proposed distributed filter will be proved without using
condition (11). The IQC (11) will only be used to guarantee
certain local performance of each node i subject to acceptable performance of its neighbours. The latter development
will be analogous to how IQCs were used in the derivation
of decentralized robust controllers to quantify the uncertainty
arising from system interconnections; e.g, see [9], [14].
However, different from decentralized controllers in those
references, our aim is to maintain coupling between the
filters, to ensure cooperation between them.
III. D ISTRIBUTED
MINIMUM ENERGY FILTERING WITH
LOCAL PERFORMANCE GUARANTEES
In this section, our main results are presented. As was
explained in Section II-B, our goal is to obtain a converging
(in the H∞ sense) distributed filter which provides global
estimation performance described in item (i) of Problem 1,
and also characterize quantitatively the connection between
local H∞ properties of the filters and their sensitivity to
estimation accuracy of their neighbours.
To solve Problem 1, we first introduce an auxiliary robust
minimum energy filtering problem involving a modified version of the standard minimum energy cost [6]. This cost func′
′
. . . ηij
]′ ,
tional, depending on the signals w and ηi , [ηij
1
li
affecting the measurements yi |[0,t] and {cij |[0,t] j ∈ Ni }
available at node i is as follows:
J¯i,t (x, w, ηi )
Z
1 t
1
kwk2 + kyi − Ci xt,x k2E −1
= kxt,x (0) − ξi k2Xi +
i
2
2 0
X
t,x
2
+
kcij − Wij x − ηij kG−1
ij
j∈Ni
−2
t,x
2
−γ kx − x̂i kRi ds;
(12)
Compared to the standard minimum energy cost functional, it includes the additional weighted penalty on the
tracking error at node i; see the last term in (12). It
was shown in [8] that the inclusion of this term enforces
a guaranteed H∞ -type performance of the filter while a
minimum energy estimate is sought; cf. [15]. The weight
matrix of this term, Ri = Ri′ > 0, Ri ∈ Rn×n was
regarded as parameters of the filter, and a process of selecting
those matrices to optimize γ 2 was proposed in [8]. However,
different from [8], the cost (12) does not include a direct
quadratic penalty on ηij . Instead, our derivation of the local
filters will impose the constraint (11) on the mismatch signals
ηij , j ∈ Ni .
With these modifications, the auxiliary robust minimumenergy filtering problem consists of determining a set of the
unknowns x, w, ηi compatible with the measurements yi and
the communications cij and minimizing the energy cost (12)
subject to the constraint (11):
¯
inf
inf Ji,t (x, w, ηi ) .
(13)
inf
x
w∈L2 [0,t] ηi ∈Ξi,t
Here Ξi,t denotes the class of vector signals ηi obtained by
stacking up all ηij , j ∈ Ni , satisfying (11). Originated from
the minimum energy filtering [6] and least square fitting,
this problem will lead to the ‘most likely’ minimum-energy
trajectory x∗i,t (·) compatible with the data at node i, y|[0,t] ,
cij |[0,t] [6]. The subscripts i, t at x∗i,t (·) are to highlight that
the trajectory x∗i,t (·) is consistent with the data collected on
the interval [0, t] at node i. By definition, the end point of this
trajectory is the minimum-energy estimate of the state x(t),
given the measurement data y|[0,t] , cij |[0,t] : x̂i (t) , x∗i,t (t).
To solve the constrained optimization problem (13), we
apply the method of S-procedure [14]. In fact, since the cost
J¯i,t (x, w, ηi ) itself depends on x̂i (·), this requires us to solve
a family of minimum energy filtering problems, in which x̂i
is replaced with an arbitrary signal x̄i . Then we take the
fixed point of the mapping x̄i (t) → x∗i,t (t) generated by this
family of minimum energy filtering problems, as x̂i (t). Due
to lack of space, we omit the details and proceed assuming
that x̂i (t) is such a fixed point.
Let τi ∈ Rn be a vector τi = [τi1 . . . τiN ]′ such that
τij > 0 if j ∈ Ni , and τij = 0 otherwise. Then define
X τij Z t
τi
kηij k2Z̄ −1
(x, w, ηi ) = J¯i,t (x, w, ηi ) +
J¯i,t
ij
2 0
j∈Ni
− kxt,x − x̂i k2 − kwk2 − kyi − Ci xt,x k2
X
−
kcir − Wir xt,x − ηir k2 ds
(14)
r∈Ni
and for fixed t and x̂i (·), x, consider the unconstrained
optimization problem
V̄iτi (x, t) =
inf
w,ηi ∈L2 [0,t]
τi
(x, w, ηi ).
J¯i,t
(15)
For each t, x, the optimization problem (15) is a standard
optimal tracking problem with a fixed terminal condition
x(t) = x, which has a unique solution under the condition
X
τij < 1.
(16)
j∈Ni
We now establish a relationship between this problem and
the constrained inner optimization problem in (13).
Let Ti (t, x) , {τi : (16) holds and V̄iτi (x, t) > −∞}.
Also for convenience, define a vector di ∈ Rn whose jth
component is dij if j ∈ Ni and is 0 otherwise.
Lemma 1: For every x̂i (·), x ∈ Rn , if the corresponding
set Ti (t, x) is nonempty, then the value of the inner optimization problem in (13) is finite,
τ ′ di
.
V̄iτi (x, t) − i
inf J¯i,t (x, w, ηi ) ≥ sup
w∈L2 [0,t],
2
τi ∈Ti (t,x)
ηi ∈Ξi,t
(17)
From Lemma 1, a lower bound on the value of the
problem (13) follows:
inf
x
inf
w∈L2 [0,t],
ηi ∈Ξi,t
≥ inf
x
J¯i,t (x, w, ηi )
τ ′ di
V̄iτi (x, t) − i
.
2
τi ∈Ti (t,x)
sup
(18)
IV. D ESIGN OF
We now consider the following optimization problem
inf V̄iτi (x, t) = inf
inf
x w,ηi ∈L2
x
τi
(x, w, ηi ).
J¯i,t
(19)
A solution to this problem involves the differential Riccati
equation
Q̇τi i = Qτi i A′ + AQτi i − Qτi i Ci′ Ei−1 Ci
X
+
Wij′ Ūij−1 Wij − γ −2 Ri − W̄i Qτi i + Si , (20)
j∈Ni
τi
Qi (0) =
Xi−1 ,
τij−1 Z̄ij ,
τij )In , Ūij , Gij +
Si = 1 −
where W̄i = (
j∈Ni
−1
P
BB ′ .
j∈Ni τij
Lemma 2: Given fixed τi ∈ Ti (t, x) and T > 0. Suppose
the differential Riccati equation (20) has a symmetric nonsingular solution Qτi i = Qτi i (t) on the interval [0, T ]. Then the
following filter computes recursively the minimizer x̂τi i (t) of
the optimization problem (19) on the interval [0, T ],
x̂˙ τi i = Ax̂τi i + Qτi i Ci′ Ei−1 (yi − Ci x̂τi i )
X
τi
−1
′
Wij Ūij (cij − Wij x̂i ) ,
(21)
+
P
j∈Ni
x̂τi i (0)
= ξi .
The value of the optimization problem (19) is finite and for
x̂i = x̂τi i is given by
Z
X
1 t
τi
τi 2
τi 2
ρ̄i,t ,
kcij −Wij x̂i kŪ −1 ds.
kyi −Ci x̂i kE −1 +
ij
i
2 0
j∈Ni
Let
τi : (16) holds and the DRE (20)
T̄i (T ) ,
has a bounded positive definite .
solution on [0, T ].
Lemma 3: For all T > 0, T̄i (T ) ⊆
T
Ti (t, x).
t∈[0,T ], x∈Rn
The following theorem summarizes the above discussion.
2
Theorem 1: Given constants γ 2 and γT
i and matrices Z̄ij ,
j ∈ Ni , suppose the set T̄i (+∞) = T >0 T̄i (T ) is not
empty. Then for any ηij for which condition (11) holds,
the filter (21) computes recursively the process x̂τi i (t) which
satisfies condition (6) with βi = τi′ di .
Compared with the distributed minimum energy filter
in [8], we have now obtained a family of suboptimal minimum energy filters for each node parametrized by τi ∈
T̄i (+∞). To be able to apply Theorem 1, it is necessary
to have a method for computing at least one such vector τi
for every node i. In the next section, we will present an algorithm that accomplishes this task. In addition, this algorithm
obtains the matrices Z̄ij and constants γi2 consistent with the
found γ 2 , thus providing a complete solution to Problem 1.
A ROBUST DISTRIBUTED ESTIMATOR
The algorithm to compute a solution to Problem 1 utilizes
a collection of linear matrix inequalities (LMIs) including
the condition (16) and the following matrix inequalities:
P
′
−2
A Ȳi + Ȳi A + γ̄i
τij I
+
j∈Ni
P
′ −1
− Ci E i Ci +
Wij′ Υij Wij
j∈Ni
B ′ Ȳi
Ȳi = Ȳi′ > 0,
Υij = Υ′ij > 0,
Ȳi B
< 0,
P
τij − 1 I
j∈Ni
Υij < G−1
ij ,
j ∈ Ni ,
τij > 0,
i = 1, . . . , N,
Θ̄ > 0.
(22)
Here, the symmetric matrix Θ̄ is composed as follows. Its
diagonal blocks Θ̄ii are defined as
θ̄ii θ̄ii . . . θ̄ii
0
ii ′
1
li
G−1
0
ij1 . . .
.
.
.
.
..
.
.
.
−1
ii ′
(θl )
0 . . . Gij
(θ1 )
Θ̄ii =
...
i
θ̄0ii =
θ̄kii
=
X
li
,
Wij′ Υij Wij + γ̄i−2 +
j∈Ni
Wij′ k Υijk ,
X
j∈Ni
k = 1, . . . , li .
τij I − γ −2 Pii ,
Also, its off-diagonal blocks Θ̄ij , i, j = 1, . . . , N , j 6= i, are
#
"
−2
Ψ
−
γ
P
0
ij
ij
n×M
j
, i < j,
Θ̄ij =
0Mi ×Mj
0n×Mi
′
Θ̄ji ,
i > j,
where
Ψij =
′
′
−Wij′ Υij Wij − Wji Υji Wji ,
−Wij Υij Wij ,
′
−Wji Υji Wji ,
0
j
j
j
j
∈
∈
6∈
6∈
Ni , i
Ni , i
Ni , i
Ni , i
∈
6∈
∈
6∈
Nj ;
Nj ;
Nj ;
Nj .
The LMIs (22), (16) represent a linear constraint on the
variables Ȳi = Ȳi′ > 0, γ̄i−2 , Υij , τij > 0 (j ∈ Ni ,
i = 1, . . . , N ), and γ −2 . Since γ 2 represents the disturbance
attenuation level in the distributed filter, a suitable set of
filter parameters is of interest which minimizes this variable.
This can be numerically achieved by solving the convex
optimization problem
sup γ −2
∗2
subject to (22), (16).
(23)
Let γ be the value of the supremum in (23).
Theorem 2: Let the pair (A, B) be stabilizable. Given a
positive semidefinite weighting matrix P = P ′ ∈ RnN ×nN ,
suppose γ 2 > γ ∗ 2 , τij , Υij , γ̄i−2 I and Ȳi , j ∈ Ni ,
i = 1, . . . , N , are a feasible collection of matrices and
scalars that satisfy the constraints of the convex optimization
problem (23). Then each Riccati equation (20) with Ri =
(γ/γ̄i )2 has a positive definite bounded solution on [0, ∞).
Furthermore, the corresponding filtering algorithm (21), (20)
verifies claims (i) and (ii) of Problem 1.
TABLE I
S OLUTIONS TO THE PROBLEM (23)
Simulation 1: Z̄ij > 0
Simulation 2: Z̄ij > 0.1I
γ 2 = 0.2500
γ 2 = 0.3116
simulation indicates that robustness of the estimators with
respect to accuracy of their neighbours can be improved by
moderately increasing γ 2 and γ̄i2 .
Node
γ̄i2
minj λmin (Z̄ij )
γ̄i2
minj λmin (Z̄ij )
VI. CONCLUSIONS
1
2
3
4
5
0.2643
0.0185
0.0181
0.1313
0.0176
2.6219 × 10−4
0.0250
0.0158
2.7548 × 10−4
0.0263
0.6288
0.0260
0.0395
0.2904
0.0265
0.1074
0.3416
0.1788
0.1000
0.2682
In this paper we proposed a distributed filtering algorithm
by utilizing an H∞ minimum-energy filtering approach to
the design of constituent filters. The algorithm employs a
decoupled computation of the individual filter coefficients.
This is achieved by considering the estimation error of
neighbouring agents as additional exogenous disturbances
weighted according to the nodes’ confidence in their neighbours’ estimates. The conditions are obtained under which
the proposed filter to provides guaranteed internal stability
and desired disturbance attenuation of the network error
dynamics. In addition each local filter guarantees certain
disturbance attenuation when assisted by the neighbours.
We have also provided a simulation example that confirms
convergence of the proposed filter in the case a system has
undetectable pairs (A, Ci ) at some of the nodes. Tuning of
the filter is discussed to reduce the dependence of the local
filters from neighbours accurate estimates.
As Theorem 2 shows, solving the SDP problem (23)
allows us to determine the suboptimal γ 2 as well as the
local disturbance attenuation levels γ̄i2 that characterize local
performance of the node filters (see (6)) as well as the matrices Z̄ij in condition (11) consistent with that performance.
Then sensitivity of performance of the obtained local filters
to the neighbours’ accuracy can be assessed using, e.g., the
eigenvalues of Z̄ij , as explained in Section II-B. This process
is illustrated in the example presented next.
V. I LLUSTRATIVE E XAMPLE
R EFERENCES
In this section, a simulated network of five sensor nodes
is considered that are to estimate a three-dimensional plant.
The plant’s state matrix and the input matrix are
#
" #
"
A=
−3.2
10
0
1
−1
1
0 −14.87 0
,
B=
0.4
0.4
0.4
.
(24)
The matrix A corresponds to one of the regimes of the
controlled Chua electronic circuit considered in [13].
The network consists of five nodes, its connectivity is described by the set of directed edges E =
{(1, 3), (2, 3), (3, 1), (3, 2), (3, 4), (4, 3), (4, 5), (5, 4)}. The
matrices Ci were taken from [13] to be C1 = C4 =
0.001 × [3.1923 − 4.6597 1] and C2 = C3 = C5 =
[−0.8986 0.1312 − 1.9703]. Note that none of the pairs
(A, Ci ) are observable, with (A, C1 ) and A, C4 ) being not
detectable. Also following [13], all communication matrices
are taken to be Wij = I3×3 if (i, j) ∈ E. Also, we let
Di = 0.025I1×3 and Fij = 0.5I3×3 .
For the above system two distributed filter designs were
compared. Both filters were designed to achieve a suboptimal
H∞ consensus performance, that is, in (5) we selected P =
(L+L⊤ )⊗I, cf. [4], [5]. First, the optimization problem (23)
was solved with the above parameters. Next, an additional
constraint Z̄ij > 0.1I was imposed. The computed levels
of local H∞ attenuation γ̄i2 and the minimum eigenvalues
of the computed matrices Z̄ij with which the Property P3
is guaranteed by Theorem 2 are shown in Table I. One
can see that in the first case, the filters at nodes 1 and
4 have much larger constants γ̄i2 and substantially smaller
values of eigenvalues of matrices Z̄ij . Together these features
indicate that these filters are significantly more sensitive to
accuracy of their neighbours. This is not unexpected given
that the pairs (A, C1 ), (A, C4 ) are not detectable. The second
[1] R. Olfati-Saber. Distributed Kalman filter with embedded consensus
filters. In Proc. 44th IEEE CDC and 2005 ECC, 8179–8184, 2005.
[2] M. V. Subbotin and R. S. Smith. Design of distributed decentralized
estimators for formations with fixed and stochastic communication
topologies. Automatica, 45(11):2491 – 2501, 2009.
[3] T. R. Nelson and R. A. Freeman. Decentralized H∞ filtering in a
multi-agent system. In Proc. American Contr. Conf., pages 5755–
5760, St. Louis, MO, 2009.
[4] V. Ugrinovskii. Distributed robust filtering with H∞ consensus of
estimates. Automatica, 47(1):1 – 13, 2011.
[5] V. Ugrinovskii and C. Langbort. Distributed H∞ consensus-based
estimation of uncertain systems via dissipativity theory. IET Control
Theory & App., 5(12):1458–1469, 2011.
[6] R. E. Mortensen. Maximum-likelihood recursive nonlinear filtering.
J. Opt. Theory Appl., 2(6):386–394, 1968.
[7] O. Hijab. Minimum Energy Estimation. PhD Dissertation, UC
Berkeley, 1980.
[8] M. Zamani and V. Ugrinovskii. Minimum-energy distributed filtering.
In Proc 53rd IEEE CDC, Los Angeles, CA, 2014. arXiv:1409.5292.
[9] L. Li, V. Ugrinovskii, and R. Orsi. Decentralized robust control
of uncertain Markov jump parameter systems via output feedback.
Automatica, 43(11):1932–1944, 2007.
[10] S. O. R. Moheimani, A. V. Savkin, and I. R. Petersen. Robust filtering,
prediction, smoothing and observability of uncertain systems. IEEE
Transactions on Circuits and Systems. Part 1, Fundamental Theory
and Applications, 45(4):446–457, 1998.
[11] J. Clark and D. A. Holton. A first look at graph theory. World
Scientific, Singapore, 1991.
[12] V. Ugrinovskii. Gain-scheduled synchronization of parameter varying
systems via relative H∞ consensus with application to synchronization of uncertain bilinear systems. Automatica, 50(11):2880–2887,
2014. arXiv:1406.5622 [cs.SY].
[13] V. Ugrinovskii. Distributed robust estimation over randomly switching
networks using H∞ consensus. Automatica, 49(1):160–168, 2013.
[14] I. R. Petersen, V. Ugrinovskii, and A. V. Savkin. Robust Control
Design using H ∞ Methods. Springer-Verlag, 2000.
[15] William M. McEneaney. Robust/H∞ filtering for nonlinear systems.
Syst. Contr. Lett., 33(5):315 – 325, 1998.
[16] M. Athans and P. L. Falb. Optimal control: an introduction to the
theory and its applications. McGraw-Hill, 1966.
| 3cs.SY
|
Published as a conference paper at ICLR 2015
FAST C ONVOLUTIONAL N ETS W ITH fbfft :
A GPU P ERFORMANCE E VALUATION
arXiv:1412.7580v3 [cs.LG] 10 Apr 2015
Nicolas Vasilache, Jeff Johnson, Michael Mathieu,
Soumith Chintala, Serkan Piantino & Yann LeCun
Facebook AI Research
770 Broadway, New York, NY 10003, USA
{ntv,jhj,myrhev,soumith,spiantino,yann}@fb.com
A BSTRACT
We examine the performance profile of Convolutional Neural Network (CNN)
training on the current generation of NVIDIA Graphics Processing Units (GPUs).
We introduce two new Fast Fourier Transform convolution implementations: one
based on NVIDIA’s cuFFT library, and another based on a Facebook authored
FFT implementation, fbfft, that provides significant speedups over cuFFT (over
1.5×) for whole CNNs. Both of these convolution implementations are available in open source, and are faster than NVIDIA’s cuDNN implementation for
many common convolutional layers (up to 23.5× for a synthetic kernel configuration). We discuss different performance regimes of convolutions, comparing areas
where straightforward time domain convolutions outperform Fourier frequency
domain convolutions. Details on algorithmic applications of NVIDIA GPU hardware specifics in the implementation of fbfft are also provided.
1
I NTRODUCTION
Deep convolutional neural networks (CNNs) have emerged as one of the most promising techniques
to tackle large scale learning problems, whether in image and face recognition, audio and speech
processing or natural language understanding. A convolutional layer within these networks provides useful properties such as translation equivariance of activations. A limiting factor for use of
convolutional nets on large data sets was, until recently, their computational expense.
Krizhevsky et al. (2012) demonstrated that training of large CNNs with millions of weights and
massive data sets is tractable when graphics processing units (GPUs) are properly put to use. Since
then, renewed interest in CNNs insufflated a fresh breath in various frameworks and implementations, including Torch (Collobert et al. (2011a)), Theano (Bergstra et al. (2010)), cuda-convnet
(Krizhevsky (2014)) and Caffe (Jia et al. (2014)). Many of these frameworks are based around
codes for NVIDIA GPUs using CUDA (Garland et al. (2008)).
We discuss our contributions to convolution performance on these GPUs, namely using Fast Fourier
Transform (FFT) implementations within the Torch framework. We summarize the theory behind
training convolutional layers both in the time and frequency domain in Section 2. We then detail
our implementations. The first is based on NVIDIA’s cuFFT and cuBLAS libraries (Section 3). We
evaluate our relative performance to NVIDIA’s cuDNN library (Chetlur et al. (2014)) on over 8, 000
different configurations (Section 4). We significantly outperform cuDNN and other time domain
convolution implementations for a wide range of problem sizes.
Our second implementation is motivated by limitations in using a black box library such as cuFFT in
our application domain, which we describe. In reaction, we implemented a from-scratch opensource implementation of batched 1-D FFT and batched 2-D FFT, called Facebook FFT (fbfft),
which achieves over 1.5× speedup over cuFFT for the sizes of interest in our application domain.
This implementation achieves GPU efficiency ratios of over 75% in certain cases. We describe an ongoing effort to further improve the performance of our solution based on algorithmic tiling (Section
6) before we conclude. Our implementation is released as part of the fbcuda and fbcunn opensource libraries at http://github.com/facebook.
1
Published as a conference paper at ICLR 2015
2
C ONVOLUTION
Discrete convolution and cross-correlation are used in CNNs. We quickly summarize these and their
implementation, with a formulation mirroring Mathieu et al. (2013). Forward propagation (fprop)
inputs are a set f of input feature planes xi , i ∈ f . These are cross-correlated1 with f ′ × f different
filter kernel weights w(j,i) , j ∈ f ′ , i ∈ f , producing output feature planes yj , j ∈ f ′ . Each input
and output feature can be part of a minibatch S, so we have x(s,i) and y(s,j) , i ∈ f, j ∈ f ′ , s ∈ S:
y(s,j) =
X
x(s,i) ⋆ w(j,i)
i∈f
The feature planes f are reduced (summed) pointwise. For back-propagation (bprop), the gradient
of the loss with respect to outputs are convolved with the kernels:
X ∂L
∂L
=
∗ w(j,i)
∂x(s,i)
∂y(s,j)
′
j∈f
Reduction is over f ′ here. Finally, the kernel weights are updated using the gradient of the loss with
respect to the weights (accGrad):
X ∂L
∂L
=
⋆ x(s,i)
∂w(j,i)
∂y(s,j)
s∈S
Reduction is over S here. For purposes of this paper, we use set symbols interchangeably to refer to
their size: each input plane is a 2-D matrix of size h × w, and each filter kernel is a 2-D matrix of
size kh × kw 2 . The output planes y(s,i) are of size (h − kh + 1) × (w − kw + 1), and implement
valid-only convolution, as per MATLAB terminology. Input zero padding and input mirror padding
around the margins of the input (ph , pw ) can be optionally added.3
A popular convolution implementation is to unroll the data until the computation is in the form of a
large matrix multiplication (Chellapilla et al. (2006)). This is the strategy followed by many implementors, since matrix multiplication is a well-tuned linear algebra primitive available on virtually
any platform. While it is possible to provide instances of direct calculation that are faster than matrix
unrolling (e.g., for large S, Krizhevsky (2014)), it is challenging to provide an implementation that
is faster for more than just a small subset of possible convolution problems.
Introducing strides in this form of convolution (i.e., performing the convolution at every dh , dw -th
offset) is a popular way to reduce the computational cost at the expense of precision. The memory
accesses required are very similar but with fewer reuse opportunities. On the other hand, by the
convolution theorem, a convolution of two discrete signals can be performed with lower asymptotic
complexity by performing the multiplication in the frequency domain. Applied to the forward pass,
it becomes:
y(s,j) =
X
i∈f
x(s,i) ⋆ w(j,i) =
X
i∈f
F −1 F (x(s,i) ) ◦ F (w(j,i) )∗
where ∗ denotes complex conjugation and ◦ is the pointwise product. The discrete Fourier basis
used is the largest of the two components convolved and the output.4 Linearity of the DFT allows
one to perform the sum above in the Fourier domain if desired. Applying the FFT then yields
a O(Sf f ′ n2 + (Sf + f f ′ + Sf ′ )n2 log n) procedure in lieu of the original O(Sf f ′ n2 k 2 ), n =
h = w, k = kh = kw . Similar transformations apply for the other two passes. We call this
a frequency domain convolution, in contrast to time domain convolution via direct computation.
1
Torch practice is that the forward pass is cross-correlation, hence the ⋆.
2-D can be extended to n-D, n ≥ 1.
3
Input size (h + ph ) × (w + pw ), output size (h + ph − kh + 1) × (w + pw − kw + 1).
4
(h × w)-dimensional or even bigger for performance (Section 3.2).
2
2
Published as a conference paper at ICLR 2015
Strided convolutions via FFT can be implemented efficiently to obtain good performance Brosch &
Tam (2015). We do not consider those in this paper.
3
CU FFT
C ONVOLUTION I MPLEMENTATION
In this section we discuss implementation strategies using the NVIDIA cuFFT libraries and their
efficiency.
3.1
FFT C ONVOLUTION D ETAILS
We described the general formulation for the three types of convolutions in section 2. Here, we
borrow the Torch naming convention: input for x(s,i) ; weight for w(j,i) ; output for y(s,j) ; gradOutput
for ∂L/∂y(s,j) ; gradInput for ∂L/∂x(s,i) ; and gradWeight for ∂L/∂w(j,i) . All are stored as singleprecision floating point 4-D tensors in row-major layout, and are stored in memory using the socalled BDHW format. This is explicit in the expression InS×f ×h×w , with input image row data as
the innermost or most varying dimension.
Table 1 describes the in-order operations for FFT computation of the forward pass, using the
F F T 2D and IF F T 2D operators and Cgemm matrix multiplication. Similar implementations follow for the other two passes. The G prefix denotes gradients. The F suffix denotes C-valued frequency domain tensors; the rest are over R. The T suffix denotes transposed tensors.
Table 1: Implementation detail for forward propagation
INPUT
OUTPUT
F F T 2D
InFS×f ×(h+ph )×(⌊ w+pw ⌋+1)
F F T 2D
−−−−−→
InS×f ×h×w
2
W eif ′ ×f ×kh ×kw
−−−−−→
W eiFf ′ ×f ×(h+ph )×(⌊ w+pw ⌋+1)
InFS×f ×(h+ph )×(⌊ w+pw ⌋+1)
−−−−−−→
T rans2D
InF T(h+ph )×(⌊ w+pw ⌋+1)×S×f
W eiFf ′ ×f ×(h+ph )×(⌊ w+pw ⌋+1)
2
(
InF T(h+ph )×(⌊ w+pw ⌋+1)×S×f
2
∗
W eiF T(h+p
)×(⌊ w+pw ⌋+1)×f ′ ×f
−−−−−−→
2
h
T rans2D
Cgemm
−−−→
2
2
W eiF T(h+ph )×(⌊ w+pw ⌋+1)×f ′ ×f
2
OutF T(h+ph )×(⌊ w+pw ⌋+1)×S×f ′
2
2
OutF T(h+ph )×(⌊ w+pw ⌋+1)×S×f ′
2
OutFS×f ′ ×(h+ph )×(⌊ w+pw ⌋+1)
2
T rans2D
−−−−−−→
OutFS×f ′ ×(h+ph )×(⌊ w+pw ⌋+1)
IF F T 2D
OutS×f ′ ×(h−kh +1)×(w−kw +1)
−−−−−−→
2
Exact tensor dimensions are also given above. By taking advantage of the Hermitian symmetry
property of the 2-D DFT for R-valued inputs we only store about half the complex entries; the
w
remaining can be obtained by complex conjugation. This results in array sizes such as ⌊ w+p
2 ⌋ + 1.
We also perform interpolation by zero-padding, which serves multiple purposes. First, it is necessary
to handle boundary conditions.5 Second, it is required to interpolate all operands over the same
Fourier basis.6 Finally, padding has an impact on the FFT algorithm used in practice, as well as on
the floating point operation count of non-FFT operations (Section 3.2).
Following the conversion into frequency domain, we perform transpositions to prepare the tensors
for Cgemm matrix multiplication library calls. The transposition converts the BDHW layout into
HWBD. The transposition is currently out-of-place and implemented using the Cgeam routine; we
are also considering our own, in-place transposition routine. Cgemm library calls are performed
on transposed tensors in the frequency domain. Casting the operation as a Cgemm call allows us
to benefit from the heavily tuned cuBLAS routine. Eventually, we transpose the result back into
the BDHW format and perform a 2-D inverse FFT. At this point, the resulting real tensor, always
5
6
In this case, we typically have ph = ⌊ k2h ⌋ and pw = ⌊ k2w ⌋.
All tensors are zero-padded to (h + ph ) × (w + pw ) before F F T 2D.
3
Published as a conference paper at ICLR 2015
(h + ph) × (w + pw ), is clipped to the appropriate final size: (h − kh + 1) × (w − kw + 1) for fprop,
h × w for bprop, kh × kw for accGrad.
3.2
CU FFT
D ESIGN S PACE
We now discuss implementation aspects we explored. Multiple factors influence the computational
efficiency of FFTs: transform size n, n’s prime factor decomposition, and whether batched or iterated single transforms are applied. In the deep learning domain, it is commonplace to deal with
small sizes, n 6= 2k . If n has undesirable properties, efficiency can drop by an order of magnitude.7
cuFFT implements FFTs with the ubiquitous Cooley-Tukey algorithm (Cooley & Tukey (1965))
which takes advantage of trigonometric equalities to recursively decompose and reuse computations. This is further discussed in the Supplement. Decomposition is built on specialized kernels of
fixed sizes which correspond to the prime factor decomposition of n. cuFFT implements specialized
building blocks for radix sizes 2, 3, 5, 7, and for sizes n where 4|n, it can use more efficient kernels
exploiting the conjugate symmetry property. When n does not admit a prime factor decomposition
using those radices only, the expensive Bluestein algorithm is used (Bluestein (1970)). Because our
results are used in the time domain, we can in fact zero-pad the image and kernel to perform the FFT
at any larger size that may be handled more efficiently. Exploiting more efficient, larger sizes should
be balanced against the extra cost introduced in the subsequent transposition and matrix multiplication steps. Table 4’s last case is one in which the best tradeoff is not easily guessed. cuFFT also has
batched mode optimizations when multiple FFTs of the same size are being performed.
3.3
CU BLAS
D ESIGN S PACE
The cuBLAS library also comes with different implementations for batched and single operation
modes. We had the choice between 3 implementation options:
• for larger batches over small matrices, the cublasCgemmBatched library call;
• for smaller batches over larger matrices, multiple cublasCgemm calls from the host;
• for intermediate batch and matrix sizes, devices of compute capability 3.5 and higher support dynamic parallelism which allows CUDA kernels to launch other kernels. This can be
beneficial for many launches over small matrices.
Note that the discussion above applies to multiplications after transposition. So the matrix size is
either S × f , S × f ′ or f × f ′ and the number of such matrices is h × w. Vendor libraries are usually
optimized for throughput and not latency, so we expect it to be more efficient for larger sizes along
critical dimensions (i.e., image size for the batch case and S × f , S × f ′ or f × f ′ for the multiple
kernel case). Due to build system limitations we were not able to experiment with the dynamic
parallelism strategy; we leave this for future work.
At the system level, we use CUDA streams and buffering of all CUDA resources and intermediate
buffers to remove synchronization points across convolutions. We are mindful of memory consumption; to address this we keep one single buffered copy of each type of tensor involved. This behavior
is tailored for a bulk synchronous execution of layers on a GPU and is not adapted for multiple
asynchronous convolutions on the same GPU. The buffers are automatically expanded as required
and reused as much as possible.
3.4
AUTOTUNING
We combine the above implementation with a simple autotuning strategy. We devise a strategy
selection mechanism that runs once for each problem size and caches the fastest strategy out of a
few dozen for later reuse. The autotuning strategy explores different possible Fourier basis sizes that
can be decomposed in powers for which cuFFT has an efficient implementation. In other words, for
an FFT dimension of size n, we explore the sizes i ∈ [n, 2⌊log2 n⌋] where i = 2a 3b 5c 7d . When the
input size is a power of 2, the search space is reduced to a single point. In addition to Fourier basis
sizes, we weigh in various cuBLAS calls and asynchronous modes.
7
http://docs.nvidia.com/cuda/cufft/index.html#accuracy-and-performance
4
Published as a conference paper at ICLR 2015
4
4.1
CU FFT
C ONVOLUTION P ERFORMANCE
P ERFORMANCE
VERSUS CU DNN:
8,232 CONFIGURATIONS
We compare our cuFFT convolution results against NVIDIA’s cuDNN 1.0 library (Chetlur et al.
(2014)), which contains one of the fastest, general purpose convolution methods for the GPU, using
matrix unrolling. It has decent performance for many problem sizes thanks to heavy autotuning of
cuBLAS codes for different problems. It is a strong baseline for this reason.
Image CNNs to date have for the most part used square input images and filters, though rectangular filters are valid for other problems (notably text CNNs, Collobert et al. (2011b)). Thus, we
restrict ourselves to a 5-D problem domain {S, f, f ′ , n(= h = w), k(= kh = kw )}. Much of this
space is not used in practice. Some areas are perhaps over-emphasized (large S, small k) due to
current engineering concerns. We evaluate cuDNN vs cuFFT-based convolution for Table 2’s 8, 232
configurations.8
Table 2: Configuration elements evaluated
DIMENSION
SIZES EVALUATED
Minibatch size (S)
Input filters (f )
Output filters (f ′ )
Kernel h/w (k = kh = kw )
Output h/w (y = h − kh + 1 = w − kw + 1)
1, 16, 64, 128
1, 4, 16, 64, 96, 128, 256
1, 4, 16, 64, 96, 128, 256
3, 5, 7, 9, 11, 13
1, 2, 4, 8, 16, 32, 64
Figures 1-6 are performance summaries of cuFFT convolution versus cuDNN on a NVIDIA Tesla
K40m, averaged across all three passes. The y-axis problem size corresponds to the minibatch size
multiplied by number of input and output planes (Sf f ′); each one of these is a pass reduction
dimension. Many possible combinations of S, f, f ′ may map to the same problem size. cuDNN performance varies to a greater degree than cuFFT across passes. This is due to the asymmetry of
convolution sizes in each pass, and the fact that a larger convolution kernel (as seen with gradient
accumulation) is essentially free in the Fourier domain. Averaging the three passes together provides a proxy for overall performance. The x-axis corresponds to output height/width. For deeper
layers in image CNNs, output size will decrease while f, f ′ will increase, so depth corresponds to
moving from the upper right to the lower left of the graph. Black areas in the chart are due to failed
cuFFT runs, due to memory pressure or undetermined potential cuFFT 6.5 issues.
FFT convolutions make large kernel sizes inexpensive, which make the performance of all three
passes roughly equal (Table 4). On the other hand, zero-padding kh × kw to h × w penalizes smaller
kernels compared to cuDNN. For 3 × 3 kernels (Figure 1), cuFFT performance is poor compared
to cuDNN. The overhead of multiple kernel launches, streaming memory in and out multiple times,
and zero-padding to the input size often outweigh the algorithmic advantage of FFT. However, for
the largest problem sizes, 3 × 3 convolution via FFT can still be advantageous, with top speed 1.84×
faster than cuDNN. 5 × 5 kernels (Figure 2) show an increasing dominance of the FFT strategy, with
top speed 5.33× faster. The tendency is confirmed for larger kernel sizes: at 13 × 13, maximum
speedup is 23.54× over cuDNN.
8
Parameterized on output rather than input size h, w because the implied h = y + kh − 1, w = y + kw − 1
will be valid for any choice of kh , kw .
5
Published as a conference paper at ICLR 2015
1/16x
16x
1
96
512
4096
12288
32768
49152
65536
98304
131072
147456
196608
262144
393216
524288
589824
786432
1048576
1179648
1572864
2097152
3145728
4194304
8388608
output size
output size
64
32
16
problem size
64
32
16
8
64
32
16
8
4
2
1
64
32
16
8
problem size
Figure 4: 9 × 9 kernel (K40m)
1
96
512
4096
12288
32768
49152
65536
98304
131072
147456
196608
262144
393216
524288
589824
786432
1048576
1179648
1572864
2097152
3145728
4194304
8388608
4
1
96
512
4096
12288
32768
49152
65536
98304
131072
147456
196608
262144
393216
524288
589824
786432
1048576
1179648
1572864
2097152
3145728
4194304
8388608
output size
Figure 3: 7 × 7 kernel (K40m)
2
1
96
512
4096
12288
32768
49152
65536
98304
131072
147456
196608
262144
393216
524288
589824
786432
1048576
1179648
1572864
2097152
3145728
4194304
8388608
problem size
output size
4
2
1
64
32
16
8
4
2
problem size
Figure 2: 5 × 5 kernel (K40m)
1
96
512
4096
12288
32768
49152
65536
98304
131072
147456
196608
262144
393216
524288
589824
786432
1048576
1179648
1572864
2097152
3145728
4194304
8388608
1
8
output size
Figure 1: 3 × 3 kernel (K40m)
1
4
2
1
problem size
64
32
16
8
4
2
1
1
96
512
4096
12288
32768
49152
65536
98304
131072
147456
196608
262144
393216
524288
589824
786432
1048576
1179648
1572864
2097152
3145728
4194304
8388608
problem size
speedup
output size
Figure 5: 11 × 11 kernel (K40m)
Figure 6: 13 × 13 kernel (K40m)
6
Published as a conference paper at ICLR 2015
4.2
CNN P ERFORMANCE
In table 3, we show performance for real CNNs, AlexNet (Krizhevsky et al. (2012)) and OverFeat
fast (Sermanet et al. (2014)), comparing against cuDNN and cuda-convnet2 (ccn2) kernels in Torch.
The first layer uses cuDNN for the cuFFT runs because it is strided, but all other layers use cuFFT.
The timings include all convolutional layers of the network.
Table 3: AlexNet and OverFeat fast performance (K40, ms)
NETWORK
KERNEL
FPROP
BPROP
ACCGRAD
TOTAL
AlexNet
cuFFT
cuDNN
ccn2
94.34
147.32
99.03
96.69
167.79
104.59
93.20
153.96
103.29
284.23
469.07
306.91
OverFeat fast
cuFFT
cuDNN
ccn2
375.65
459.06
433.11
460.48
634.26
398.87
397.85
508.02
450.82
1233.98
1601.35
1282.80
Table 4 shows the performance of the cuDNN and our cuFFT convolution implementation for some
representative layer sizes, assuming all the data is present on the GPU. Our speedups range from
1.4× to 14.5× over cuDNN. Unsurprisingly, larger h, w, smaller S, f, f ′ , kh , kw all contribute to
reduced efficiency with the FFT. More surprisingly, we experience noticeable speedups on small
3 × 3 kernels as long as the input tensor remains of small size. The optimal FFT sizes that autotuning finds are reported in columns 2 and 3; note L5 padding being found by the autotuner.
Column 7 has the trillion equivalent time-domain reductions per second (single-precision floating
point multiply-adds) achieved by our implementation on a NVIDIA Tesla K40m on CUDA 6.5.
This number represents the throughput a time-domain kernel needs to achieve in order to match our
implementation; it is computed as (Sf f ′ kh kw (h − kh + 1)(w − kw + 1))/time. This is a metric
to compare relative efficiency across problem and padding sizes. In the cases L2, L3 and L4, a time
domain convolution would need to exceed the K40m peak of 4.29 Tflop/sec in order to match our
throughput.
5
fbfft I MPLEMENTATION
This section presumes familiarity with GPU architecture. Refer to the Supplement for details.
When designing high-performance libraries, multiple objectives must be balanced against each
other: memory latency/bandwidth tradeoffs, maximizing locality without sacrificing too much parallelism, good instruction mix, register usage and mapping strategy of computation and data to
memories and compute elements. A key principle is to design a set of leaf kernels with well-tuned
in-register performance and reduce the larger problem to a combination of these kernels by data and
loop tiling (Irigoin & Triolet (1988)) and recursive decompositions (Gunnels et al. (2001)). Since
vendors have to sustain high performance for a large class of application domains, there exist parameter configurations for which a carefully tuned approach significantly outperforms vendor-tuned
libraries (Shin et al. (2010)). For common deep learning use, convolutional layers consist of many
batched small 2-D convolutions. These are tiny relative to DSP and HPC standards and put us in
a regime where (a) we fall outside of the highly tuned regime, (b) feature dimensions are often
smaller than GPU warp sizes and can often fit exclusively in registers rather than in shared memory
(SMEM), and (c) we are very sensitive to latencies. We determined that it is possible to obtain better
efficiency than the existing batched cuFFT mode for CNNs.
5.1
L IMITATIONS
OF CU FFT
Because the cuFFT library is a black box, zero-padding9 has to be explicitly embedded in the input
and output arrays. The consequence is that one may need to allocate a duplicate, larger memory
9
This is different from the FFTW compatibility padding mode for in-place transforms.
7
Published as a conference paper at ICLR 2015
Table 4: Representative layer performance (S = 128, K40m)
LAYER
h + ph
L1
fprop
bprop
accGrad
L2
fprop
bprop
accGrad
L3
fprop
bprop
accGrad
L4
fprop
bprop
accGrad
L5
fprop
bprop
accGrad
Params:
128
128
128
Params:
64
64
64
Params:
32
32
32
Params:
16
16
16
Params:
13
13
13
w + pw
cuDNN
cuFFT
SPEEDUP
f = 3, f ′ = 96, h = w = 128, kh = kw = 11
128
125.11 ms 80.98 ms 1.54×
128
153.39 ms 66.49 ms 2.30×
128
155.07 ms 69.63 ms 2.22×
f = 64, f ′ = 64, h = w = 64, kh = kw = 9
64
354.83 ms 46.44 ms 7.64×
64
579.37 ms 46.25 ms 12.5×
64
416.34 ms 47.03 ms 8.85×
f = 128, f ′ = 128, h = w = 32, kh = kw = 9
32
130.89 ms 17.77 ms 7.36×
32
245.57 ms 16.97 ms 14.5×
32
154.96 ms 17.00 ms 9.29×
f = 128, f ′ = 128, h = w = 16, kh = kw = 7
16
15.13 ms
4.88 ms
3.10×
16
20.80 ms
4.71 ms
4.41×
16
18.17 ms
4.70 ms
3.86×
f = 384, f ′ = 384, h = w = 13, kh = kw = 3
14
39.82 ms
21.35 ms 1.86×
14
28.33 ms
20.22 ms 1.40×
14
47.84 ms
21.26 ms 2.25×
TRED/s
0.9
1.1
1.05
7.49
7.52
7.40
9.90
10.37
10.34
5.54
5.76
5.75
1.34
1.42
1.35
region (only once) and copy data from non-padded tensors to padded tensors. This memory consumption and spurious copies affect latency significantly. Instead, we devised an implementation for
batched 1-D FFT and 2-D FFT of sizes 2-256 and reaches up to 78% efficiency at 97.5% occupancy.
We also implemented an IFFT kernel based on our FFT kernel.
In our implementation we use clipping to conditionally load a value if reading within bounds or a
constant (0) otherwise. This is an approach used in automatic code generation tools such as Halide
(Ragan-Kelley et al. (2013)) and relies on aggressive if-conversion properties of the CUDA compiler.
It allows for more efficient control flow rather than using explicit loop prologues and epilogues. This
mechanism does not require any additional memory allocation and is zero-copy; this is particularly
desirable in the latency sensitive mode.
Additionally, since cuFFT and cuBLAS are closed source, it is impossible to take advantage of algorithmic simplifications that may be available. For instance, in the forward pass of our computation as
shown in Table 1, the result of the first cuFFT call is of the form S×f ×(h+ph)×(⌊(w+pw )/2⌋+1).
With fbfft we return it in the form S × f × (⌊(w + pw )/2⌋ + 1) × (h + ph ) where the two innermost data dimensions are transposed. This allows us to remove a full data transposition from each
of the FFT kernels. Another domain-specific optimization we have yet to explore is eliminating bit
reversal portions of the FFT and IFFT. This can be done by performing the FFT with decimation in
frequency (DIF) and the IFFT with decimation in time (DIT), discussed in the Supplement.
5.2
WARP - LEVEL 1-D FFT
AND
2-D FFT
FOR SIZE
n ≤ 32
For batched FFT of power of two sizes we view a single warp as a small distributed system with
lockstep collective communication capabilities and we program it in a bulk-synchronous fashion
(Valiant (1990)). We implement DIF and enforce the following invariants for the log2 n steps:
• each warp thread originally loads one real element of the input vector and locally computes
one complex twiddle factor (i.e. a root of unity);
• at each step, all warp threads exchange data with another thread in the warp in parallel and
produce a new value;
8
Published as a conference paper at ICLR 2015
• then, all warp threads exchange twiddle factors with another thread in the warp in parallel,
and produce a new value.
The two bulk-synchronous exchanges can be written each with one warp-wide instruction. After
the log2 n steps, the FFT is computed and stored in a distributed and bit reversed manner within 1
register across a warp. For sizes n ≤ 32, bit reversal can be implemented with a single warp shuffle.
We either load twiddle factors from device memory or compute them with the sincosf function
only once, and subsequently swap them within registers. This greatly reduces the reliance on either
memory bandwidth or on the special functional unit at the expense of a few additional registers.
The decision between explicitly loading twiddle factors from device memory or computing them is
a tradeoff between arithmetic intensity and memory bandwidth. For sizes 16 and 32 the arithmetic
pipeline is the bottleneck. Loading twiddle factors from memory for these two special sizes results
in a performance increase of 15% and 20% respectively.
The discussion above applies to 1-D FFT and to each independent FFT within a larger 2-D FFT.
A n-D Fourier transform is separable and can be implemented with sets of multiple 1-D FFT with
transpositions between each of these sets. In 2-D FFT R-to-C, the first set comprises n FFTs and the
second set comprises n/2 + 1 FFTs by Hermitian symmetry. Following standard techniques Lyons
(1996) we further pack 2 real FFTs into a single complex FFT . The extra 1 term in the quantity n/2+
1 makes the computation ill-balanced and can bring down performance by lowering occupancy. We
chose to dimension our kernels to have size n×(n/2) and introduce additional control flow to handle
the border case. This results in 30% additional performance. We implement the transposition in
SMEM across warps following Ruetsch & Micikevicius (2009). Data is already resident in registers
so our main concerns are limiting SMEM usage to keep occupancy high, and limiting load/stores by
using vector instructions to avoid saturating the load-store unit (LSU).
5.3
1-D FFT
AND
2-D FFT
FOR SIZE
32 < n ≤ 256
With size 32 as our building block, we extend our strategy to larger sizes. We use the same single
warp approach to compute a full 1-D FFT. The main difference is that the computation is now distributed across multiple registers across threads in a warp (⌈n/32⌉ Fourier coefficients and twiddle
factors in registers per thread). Because we perform a full FFT per warp, a performance cross-over
where cuFFT wins happens after register usage limits occupancy too much. We outperform 1-D
cuFFT for n ≤ 256, with a hard register limit at n = 512 (128 and 256 similarly for 2-D FFT). This
is still well within our application domain. The following modifications handle multiple registers
per thread:
• Hermitian symmetry allows us to perform half the computation. There is a tradeoff between adding control-flow divergence and performing less work. At n ≥ 64, benefits from
reduced computations dominate divergence losses;
• we take advantage of trigonometric symmetries and twiddle factor distribution to compute
only a fraction of the roots of unity needed for each FFT, distributed with register to register
copies;
• twiddle factor re-balancing across a warp and across registers requires a different implementation. We managed to implement it fully within registers;
• bit reversal occurs across registers and across warps. The high-order bits represent the register while the low-order bits represent the warp. Without a sophisticated implementation,
this results in indirect addressing of registers which is costly. We implement a simple bit
reversal in SMEM, which is an occupancy bottleneck at n ≥ 256 for 1-D FFT.
In the 2-D FFT case, the intermediate transpose becomes significantly more expensive. We experimented with various strategies to keep occupancy high, including partial transpositions within a
warp to use minimal amounts of SMEM.
5.4
D ISCUSSION
We report the relative performance of our implementation fbfft compared to cuFFT for various
batch and input sizes of interest. The number of batches to consider depends on the dimension of
9
Published as a conference paper at ICLR 2015
CNN layers as well as any multi-GPU parallelization strategy that may be involved. At typical sizes
of interest, fbfft is between 1.5× and 5× faster. We tried up to 4 million batches and at larger
sizes gains stabilize around 1.4× but efficiency goes down as more and more memory is used.
FBFFT-1D Speedup at various sizes and batches
4.5
3.5
FBFFT Speedup
3.0
3.5
2.5
2.0
1.5
3.0
2.5
2.0
1.5
1.0
1.0
0.5
0.5
0.0
4
32
128
1024
4096
Number of batches
16384
8
16
32
64
128
256
4.0
FBIFFT Speedup
4.0
FBIFFT-1D Speedup at various sizes and batches
4.5
8
16
32
64
128
256
0.0
65536
4
32
128
1024
4096
Number of batches
16384
65536
Figure 7: fbfft-1D FFT and IFFT (K40m, cuFFT 6.5 @ 1x)
FBFFT-2D Speedup at various sizes and batches
4.5
4.0
3.5
3.5
3.0
2.5
2.0
1.5
2.5
2.0
1.5
1.0
1.0
0.5
0.5
0.0
4
32
128
1024
4096
Number of batches
16384
8
16
32
64
128
4.0
FBIFFT Speedup
FBFFT Speedup
3.0
FBIFFT-2D Speedup at various sizes and batches
4.5
8
16
32
64
128
0.0
65536
4
32
128
1024
4096
Number of batches
16384
65536
Figure 8: fbfft-2D FFT and IFFT (K40m, cuFFT 6.5 @ 1x)
Figure 7 shows the performance in the 1-D case. These numbers do not exercise our implicit zerocopy padding, so we expect additional gains when we incorporate our FFT in the convolution. Our
implementation outperforms cuFFT for all cases of interest, more dramatically so for smaller batch
sizes. Small batch sizes also correspond to the latency sensitive regime in Figures 1-6 for which
the cuFFT based implementation performs quite worse than cuDNN. We achieve 78% efficiency at
97.5% occupancy for size 64 at batch size 16, 384, as reported by nvvp.
Figure 8 shows the performance in the 2-D case. Relative performance gains for sizes 64 are more
modest than in the 1-D case, even losing to cuFFT at size 128 and small batch sizes. The magnitude
of the relative gains at various batch sizes drops faster than in the 1-D case. Looking at the performance of the 32 × 32 FFT, we obtain 1.6× speedup over cuFFT at 1, 024 batches. The same ratio is
not obtained until 16, 384 batches in 1-D FFT.10 When coupled with the tiling strategy in Section 6,
we emphasize that the sizes of interest are actually 8-64, and depend on kh , kw but not input h, w.
Batch sizes can vary on the whole spectrum.
We interfaced fbfft into our convolution module and ran experiments with 3 × 3 kernels for the
3 different convolution passes over inputs of sizes x = h = w, x ∈ {13, 16, 27, 32, 57, 64}. For
problem size, we used p = S = f = f ′ , p ∈ {16, 32, 64, 128}. By swapping our FFT implementation we observed an overall mean speedup of 1.51× with standard deviation 0.21 and geometric
mean 1.49×. The minimum speedup was 1.21×, despite sometimes performing more computations
10
This is not unexpected because these two computations perform the same number of flops when accounting
for Hermitian symmetry, plus the fact that the efficiency of cuFFT increases while fbfft remains high but
almost constant.
10
Published as a conference paper at ICLR 2015
with fbfft which can only interpolate to a power of 2. These experiments exercise the zero-copy
padding and lower memory footprints of fbfft compared to cuFFT but do not yet reflect additional
optimizations such as tiling and bit twiddling elision.
6
C URRENT L IMITATIONS
AND
F UTURE WORK
In our current implementation, fbfft heavily relies on shuffle instructions. In spite of a good
efficiency, we only utilize 60% of the available memory bandwidth. This is due to the load and store
instructions in our kernel competing with the shuffle instructions for the Load-Store Unit (LSU). As
a consequence, our first bottleneck is the number of instructions issued on the LSU. For instance, on
Kepler (capability 3.5), the throughput for 32-bit floating point multiply-add operations is 192 per
cycle but the throughput for shuffles is only 32. In the future we will investigate and release faster
implementations as they become available.
Temporary memory overhead requirements are a common issue when performing convolutions in
the Fourier domain. In this first implementation, we introduced the following memory buffers to
support our implementation:
• for each of input, output and weight tensors we store 1 buffer for the frequency array and
1 buffer for its complex transpose. These buffers store the Fourier representation and are
generally limited by the weight tensor which is independent of the mini-batch size. Because
of the global memory pressure we introduce, we reuse buffers at each layer and pass on the
opportunity to (1) reuse 2 FFT results in each hidden layer, reducing the cost of forward
FFTs by 33%; and (2) asynchronously precompute FFTs of the weight tensors and their
gradients to better fill the gpu utilization pipeline,
• when using cuFFT we additionally pad the input, weight and output tensors explicitly to
the best performing common fft size
• when using cuFFT additional temporary memory is reserved by each cufftPlan
• with fbfft padding is implicit but and no temporary memory buffer is needed until we
reach size 64. On the other hand, fbfft only supports square convolutions whose size is
a power of 2. As a consequence, too much padding could occur and adversely affect both
performance and memory consumption. The tiling strategy we describe next is a good way
to circumvent the problem.
Additionally, we recently developed an in-place transposed batched CGEMM which permits the
removal of the complex transposed buffer. For this problem, a tool like MaxAS Lavin (2015) could
be valuable.
fbfft provides the most gains over cuFFT at sizes 8-64. A tiling strategy for the input can be used
to exploit this advantage. When the kernel is significantly smaller than the input, we can decompose
a large convolution into several smaller ones. For simplicity, we consider 1D convolution on a
single input plane, as it can trivially be extended. Let x be an input of size n, c a kernel of size w and
y = x ⋆ c. We write x[i,j] for the vector formed by contiguous elements of x: {xi , xi+1 , ..., xj−1 }.
Let d ≤ n. From the definition of the convolution, we have:
y[i,i+d] = x[i,i+d+w] ⋆ c
So the convolution of the input of size n can be computed with ⌊n/d⌋ convolutions with inputs of
size d + w. The cost of the convolution goes down from O(n log(n)) to O(⌊n/d⌋(d + w) log(d +
w)) = O((n + w/d) log(d + w)). From this formula, we see that the optimal d is of the order of
w, to get the complexity O(n log(w)). This strategy allows us to speed up forward and backward
propagation. Tiling can also be used to reduce memory cost for temporary storage by not running all
the tiles in parallel (just the tiles which do run in parallel need their scratch space), at the potential
expense of parallelism or efficiency.
For the gradient accumulation, we cannot reuse this strategy, since it involves a larger convolution
between an input x of size n and a kernel z = ∂L
∂y of size n − w + 1. However, we have a similar
formula:
11
Published as a conference paper at ICLR 2015
∂L
∂c
=
j
i=0
And so
∂L
∂c
n−1
X
⌊n/d⌋−1 d−1
xj+i · zi =
X X
k=0
i=0
xj+i+kd · zi+kd +
n−1
X
xj+i · zi
i=d⌊n/d⌋
⌊n/d⌋−1
=
X
x[dk,(d+1)k+w−1] ⋆ z[dk,(d+1)k] + x[d⌊n/d⌋,n] ⋆ z[d⌊n/d⌋,n−w+1]
k=0
We have a few other optimizations that are planned as well. Since much of the data we have is already available in registers or in shared memory, we are implementing our own in-place, in-register
transpose via recursive decomposition. The pointwise multiplications in the Fourier domain, especially with tiling, are rather small, so our own matrix multiplication routines integrated with the
rest of the convolution kernel code might win over cuBLAS, and prevent the need for multiple
CUDA kernel launches and their associated overhead. Finally, as mentioned earlier, bit reversal
portions can be eliminated with the FFT using DIF and the IFFT using DIT.
7
C ONCLUSION
To summarize, we achieve significant gains in CNNs using FFTs, with a cuFFT convolution implementation achieving 1.4 × −14.5× speedups over cuDNN for common sizes. In reaction to
cuFFT and cuBLAS limitations in the context of our specific application domain, we developed
our own FFT implementation, fbfft, which is more suited to deep learning problem sizes (large
batches, small feature planes). fbfft itself is ≥ 1.4× faster than cuFFT transforms for these problems of interest. For convolution, it is faster than the cuFFT as well, with a mean of 1.51× for sizes
that we wish to exploit.
Given our new efficient primitive for size 8-64 convolution, we are continuing work on bit twiddling,
transposition and pointwise multiplication optimizations, and continuing work on tiling to make the
computational advantage at that size apply to larger convolution problems. These will all allow for
reduced training time and use of ever larger and deeper CNNs.
ACKNOWLEDGMENTS
We would like to thank Julien Demouth from NVIDIA who suggested further improvements are still
possible by virtue of the current implementation being LSU throughput-bound rather than memorybound.
R EFERENCES
Bergstra, James, Breuleux, Olivier, Bastien, Frédéric, Lamblin, Pascal, Pascanu, Razvan, Desjardins, Guillaume, Turian, Joseph, Warde-Farley, David, and Bengio, Yoshua. Theano: a CPU
and GPU math expression compiler. In Proceedings of the Python for Scientific Computing Conference (SciPy), June 2010. Oral Presentation.
Bluestein, Leo I. A linear filtering approach to the computation of discrete Fourier transform. Audio
and Electroacoustics, IEEE Transactions on, 18(4):451–455, December 1970. ISSN 0018-9278.
Brosch, Tom and Tam, Roger C. Efficient training of convolutional deep belief networks in the
frequency domain for application to high-resolution 2d and 3d images. Neural Computation, 27
(1):211–227, 2015. doi: 10.1162/NECO a 00682. URL http://dx.doi.org/10.1162/
NECO_a_00682.
Burrus, C. Sidney. Fast fourier transforms, 2008. URL http://cnx.org/contents/
[email protected]:16/Fast_Fourier_
Transforms_(6x9_V.
Chellapilla, Kumar, Puri, Sidd, and Simard, Patrice. High Performance Convolutional Neural Networks for Document Processing. In Lorette, Guy (ed.), Tenth International Workshop on Frontiers
12
Published as a conference paper at ICLR 2015
in Handwriting Recognition, La Baule (France), October 2006. Université de Rennes 1, Suvisoft.
URL https://hal.inria.fr/inria-00112631. http://www.suvisoft.com.
Chetlur, Sharan, Woolley, Cliff, Vandermersch, Philippe, Cohen, Jonathan, Tran, John, Catanzaro,
Bryan, and Shelhamer, Evan. cudnn: Efficient primitives for deep learning. CoRR, abs/1410.0759,
2014. URL http://arxiv.org/abs/1410.0759.
Collobert, R., Kavukcuoglu, K., and Farabet, C. Torch7: A matlab-like environment for machine
learning. In BigLearn, NIPS Workshop, 2011a.
Collobert, Ronan, Weston, Jason, Bottou, Léon, Karlen, Michael, Kavukcuoglu, Koray, and Kuksa,
Pavel. Natural language processing (almost) from scratch. J. Mach. Learn. Res., 12:2493–2537,
November 2011b. ISSN 1532-4435. URL http://dl.acm.org/citation.cfm?id=
1953048.2078186.
Cooley, James W. and Tukey, John W. An algorithm for the machine calculation of complex fourier
series. Mathematics of computation, 19(90):297–301, 1965.
Garland, Michael, Le Grand, Scott, Nickolls, John, Anderson, Joshua, Hardwick, Jim, Morton,
Scott, Phillips, Everett, Zhang, Yao, and Volkov, Vasily. Parallel computing experiences with
cuda. IEEE Micro, 28(4):13–27, July 2008. ISSN 0272-1732. doi: 10.1109/MM.2008.57. URL
http://dx.doi.org/10.1109/MM.2008.57.
Giles, Mike. Course on cuda programming on nvidia gpus, lecture 3, 2014.
//people.maths.ox.ac.uk/gilesm/cuda/lecs/lec3.pdf.
URL http:
Gunnels, John A., Henry, Greg M., and van de Geijn, Robert A. A family of high-performance
matrix multiplication algorithms. In Proceedings of the International Conference on Computational Sciences-Part I, ICCS ’01, pp. 51–60, London, UK, UK, 2001. Springer-Verlag. ISBN
3-540-42232-3. URL http://dl.acm.org/citation.cfm?id=645455.653765.
Irigoin, F. and Triolet, R. Supernode partitioning. In Proceedings of the 15th ACM SIGPLANSIGACT Symposium on Principles of Programming Languages, POPL ’88, pp. 319–329, New
York, NY, USA, 1988. ACM. ISBN 0-89791-252-7. doi: 10.1145/73560.73588. URL http:
//doi.acm.org/10.1145/73560.73588.
Jia, Yangqing, Shelhamer, Evan, Donahue, Jeff, Karayev, Sergey, Long, Jonathan, Girshick, Ross,
Guadarrama, Sergio, and Darrell, Trevor. Caffe: Convolutional architecture for fast feature embedding. arXiv preprint arXiv:1408.5093, 2014.
Krizhevsky, Alex.
convnet2/.
cuda-convnet2, 2014.
URL https://code.google.com/p/cuda-
Krizhevsky, Alex, Sutskever, Ilya, and Hinton, Geoffrey E. Imagenet classification with deep
convolutional neural networks. In Pereira, F., Burges, C.J.C., Bottou, L., and Weinberger,
K.Q. (eds.), Advances in Neural Information Processing Systems 25, pp. 1097–1105. Curran Associates, Inc., 2012. URL http://papers.nips.cc/paper/4824-imagenetclassification-with-deep-convolutional-neural-networks.pdf.
Lavin, Andrew. maxdnn: An efficient convolution kernel for deep learning with maxwell gpus.
CoRR, abs/1501.06633, 2015. URL http://arxiv.org/abs/1501.06633.
Lyons, Richard G. Understanding Digital Signal Processing. Addison-Wesley Longman Publishing
Co., Inc., Boston, MA, USA, 1st edition, 1996. ISBN 0201634678.
Mathieu, Michaël, Henaff, Mikael, and LeCun, Yann. Fast training of convolutional networks
through ffts. CoRR, abs/1312.5851, 2013. URL http://arxiv.org/abs/1312.5851.
Ragan-Kelley, Jonathan, Barnes, Connelly, Adams, Andrew, Paris, Sylvain, Durand, Frédo, and
Amarasinghe, Saman P. Halide: a language and compiler for optimizing parallelism, locality,
and recomputation in image processing pipelines. In ACM SIGPLAN Conference on Programming Language Design and Implementation, PLDI ’13, Seattle, WA, USA, June 16-19, 2013, pp.
519–530, 2013. doi: 10.1145/2462156.2462176. URL http://doi.acm.org/10.1145/
2462156.2462176.
13
Published as a conference paper at ICLR 2015
Ruetsch, Greg and Micikevicius, Paulius. Optimizing matrix transpose in cuda. Technical report,
NVIDIA Corp., January 2009.
Sermanet, Pierre, Eigen, David, Zhang, Xiang, Mathieu, Michael, Fergus, Rob, and LeCun, Yann.
Overfeat: Integrated recognition, localization and detection using convolutional networks. In
International Conference on Learning Representations (ICLR 2014). CBLS, April 2014. URL
http://openreview.net/document/d332e77d-459a-4af8-b3ed-55ba.
Shin, Jaewook, Hall, Mary W., Chame, Jacqueline, Chen, Chun, Fischer, Paul F., and Hovland,
Paul D. Speeding up nek5000 with autotuning and specialization. In Proceedings of the 24th
ACM International Conference on Supercomputing, ICS ’10, pp. 253–262, New York, NY, USA,
2010. ACM. ISBN 978-1-4503-0018-6.
Valiant, Leslie G. A bridging model for parallel computation. Commun. ACM, 33(8):103–111,
August 1990. ISSN 0001-0782. doi: 10.1145/79173.79181. URL http://doi.acm.org/
10.1145/79173.79181.
Volkov, V. Better performance at lower occupancy. In GPU Technology Conference, 2010. URL
http://www.cs.berkeley.edu/˜volkov/volkov10-GTC.pdf.
14
Published as a conference paper at ICLR 2015
8
8.1
S UPPLEMENT
CU FFT
C ONVOLUTION P ERFORMANCE B REAKDOWN
We show a breakdown of cuFFT convolution performance for the steps indicated in Table 1. The
timings do not add up to 100% of the reported performance in the previous table because we do not
report additional copies needed for zero-padding here. We also enforce force extra synchronizations
to isolate the contribution of each operation. Abstracting from these details, the FFT and IFFT take
up a significant amount of compute resources, which we address in Section 5.
Table 5: cuFFT convolution performance breakdown (K40m, ms)
LAYER
L1
fprop
bprop
accGrad
L2
fprop
bprop
accGrad
L3
fprop
bprop
accGrad
L4
fprop
bprop
accGrad
L5
fprop
bprop
accGrad
FFT A
TRANS. A
FFT B
TRANS. B
CGEMM
TRANS. C
IFFT C
0.86
0.86
1.14
0.24
0.24
0.32
1.13
34.55
34.60
0.32
10.26
10.26
15.13
12.62
12.37
12.67
0.39
0.26
36.46
1.19
0.91
2.99
2.99
5.94
0.98
0.98
2.04
5.91
5.92
5.93
2.03
2.03
2.02
8.92
8.85
8.38
1.67
1.67
0.83
6.24
6.23
3.15
3.07
3.08
3.07
0.89
0.89
0.89
3.08
3.07
3.06
0.89
0.90
0.89
4.40
4.05
4.03
0.87
0.86
0.87
3.49
3.48
3.48
0.84
0.83
0.84
0.24
0.24
0.24
0.83
0.83
0.82
0.24
0.24
0.24
1.21
1.13
1.10
0.24
0.23
0.24
0.95
0.94
0.95
7.07
7.07
2.40
1.58
1.59
0.51
2.39
2.40
2.38
0.51
0.51
0.52
6.23
5.59
6.18
0.50
0.51
1.54
2.54
2.54
7.51
In the particular case of L1, the FFTs take more than 50% of the runtime. This is due to the wasteful
interpolation of the kernel tensor from a 11 × 11 up to 128 × 128, which is the minimal size to
compute the FFT of the input array without interpolation loss. In such cases, the tiling strategy we
are developing (see section 6) will result in large additional performance gains.
8.2
FFT : D ECIMATION
IN
T IME VS F REQUENCY
A Fourier transform projects R and C-valued functions onto a harmonic orthogonal basis. The
discrete Fourier transform of a vector {xk }, k ∈ [0, n − 1] is the vector:
{Xk } =
n−1
X
j=0
xj wnkj , k ∈ [0, n − 1]
where wnj = e−2πij/n is the j th n-root of unity. The traditional radix-2 Cooley-Tukey algorithm
recursively decomposes the computation between an odd and even part:
{Xk } =
(n−1)/2
(n−1)/2
X
X
j=0
xj wnk(2j) +
j=0
15
x2j+1 wnk(2j+1) , k ∈ [1, n]
Published as a conference paper at ICLR 2015
Figure 9: DIT output ordered (left); DIF input ordered (right) (Burrus (2008))
This decomposition is called decimation in time (DIT). An alternate decomposition performs decimation in frequency (DIF):
(n−1)/2
n
X
X
xj wnkj , k ∈ [1, n]
xj wnkj +
{Xk } =
j=0
j=(n−1)/2
When n is a power of 2, both decimations recursively decompose into a perfectly balanced tree and
take advantage of the symmetry properties of the roots of unity. The dataflow graph for the radix-2
FFT has a butterfly shape and is a good way of visualizing the computations. There is a symmetry
between DIT and DIF in both the order of operations applied and in whether the input or the output
order is shuffled (Figure 9).
8.3
GPU P ROGRAMMING
There are a variety of references available that describe CUDA and NVIDIA’s various GPU architectures (Garland et al. (2008)) which we won’t discuss in detail, but the implementation of fbfft very
much depends upon specifics of the Kepler GPU architecture.
NVIDIA GPUs execute code at the granularity of a warp which is defined as a set of 32 threads
in all existing architectures; each thread is assigned a lane within the warp. These threads execute
in a SIMT (single instruction, multiple thread) fashion, meaning that a warp is an atomic unit of
execution. It holds a single program counter (PC) and can thus only execute a single instruction
at a time across all of its threads. Collections of warps are brought together in blocks or CTAs,
which together share a region of fast shared memory resident on chip. Blocks themselves can only
exchange data via much slower global memory, resident on the GPU or in the host CPU’s address
space.
Individual threads within a warp are free to take divergent paths, but since a single PC is present,
each branch in the execution will be serialized. Threads that aren’t participating in the branch in
question are disabled. In other words, if all 32 threads were to take divergent code paths, we would
obtain only 1/32× of the computational efficiency.
Divergent code paths are hard to avoid, but the NVIDIA instruction set has means to reduce their
cost (Giles (2014)). One is with predicated instructions, which are used for small branches, in which
all warp threads execute both parts of the branch, with non-participating threads having no side
effects.
Block threads have access to a register file, with up to 255 registers per thread for Kepler. Registers
are allocated statically by the CUDA compiler. An important performance factor when writing
CUDA kernels is that data should be kept in registers as much as possible to avoid communications.
16
Published as a conference paper at ICLR 2015
Registers in CUDA are “addressable”: it is possible to declare a static array within registers and
operate on its elements. The limitation is that all addressing should be performed using statically
determined constants so the compiler can translate these accesses to a register number known at
compile time. Indirect addressing is also supported but results in copies to a local region within
global memory, which essentially constitutes register spilling. Even with the presence of caches,
using local memory usually comes with a performance hit.11 As a consequence, we design our
kernels using aggressive inlining, template parameters and unrolling directives to make all register
accesses statically addressable.
The Kepler architecture introduced specialized shuffle instructions to exchange data between registers within a warp synchronously, which avoids round-trips to shared or global memory. Interestingly, these shuffle instructions allow the dynamic indexing of an array held in registers, as long as
the array is distributed in a cyclic fashion across registers in each thread within a warp.
float arr[3];
...
// This simulates a linear array float realArr[96]:
// arr[0] holds elements 0-31 (lane i holds element i)
// arr[1] holds elements 32-63 (lane i holds element 32 + i)
// arr[2] holds elements 64-95 (lane i holds element 64 + i)
// Example: all warp threads read value held at realArr[34]
float val = __shfl(arr[1], 2); // ‘1‘ must be statically known
// ‘2‘ can be dynamic
Many warps run in parallel and can be switched by the GPU hardware at each cycle. When enough
parallelism is available (measured in occupancy of the GPU as a first approximation), long latency
operations are hidden thanks to fast context switching. Registers and shared memory come in finite
quantities on each GPU compute multiprocessor. These limited resources are partitioned by the
compiler and the hardware amongst computations at the level of a CUDA kernel. Increased usage
of registers or of shared memory can reduce GPU occupancy, which limits the ability to hide long
latency operations. Reduced occupancy does not necessarily result in performance loss (Volkov
(2010)). There are often non-obvious performance tradeoffs in increasing or decreasing threads per
block, shared memory per block or registers per thread that are hard to discover. This problem is
one of the many reasons why designing a one-size-fits-all implementation that aims to be efficient
for any problem is difficult.
11
There are bleeding edge cases where a little local memory consumption helps performance; for instance,
when restricting the number of registers per thread to increase occupancy.
17
| 9cs.NE
|
COMPOSITIONAL ABSTRACTION FOR NETWORKS OF CONTROL SYSTEMS: A
DISSIPATIVITY APPROACH
arXiv:1608.01590v2 [math.OC] 29 Dec 2016
MAJID ZAMANI1 AND MURAT ARCAK2
Abstract. In this paper we propose a compositional scheme for the construction of abstractions for networks
of control systems using the interconnection matrix and joint dissipativity-type properties of subsystems and
their abstractions. In the proposed framework, the abstraction, itself a control system (possibly with a lower
dimension), can be used as a substitution of the original system in the controller design process. Moreover, we
provide a procedure for constructing abstractions of a class of nonlinear control systems by using the bounds
on the slope of system nonlinearities. We illustrate the proposed results on a network of linear control systems
by constructing its abstraction in a compositional way without requiring any condition on the number or
gains of the subsystems. We use the abstraction as a substitute to synthesize a controller enforcing a certain
linear temporal logic specification. This example particularly elucidates the effectiveness of dissipativity-type
compositional reasoning for large-scale systems.
1. Introduction
Modern applications, e.g. power networks, biological networks, internet congestion control, and manufacturing
systems, are large-scale networked systems and inherently difficult to analyze and control. Rather than tackling
the network as a whole, an approach that severely restricts the capability of existing techniques to deal with
many numbers of subsystems, one can develop compositional schemes that provide network-level certifications
from main structural properties of the subsystems and their interconnections.
In the past few years, there have been several results on the compositional abstractions of control systems.
Early results include compositional abstractions of control systems [TPL04, Fre05, Kv10] which are useful for
verification rather than synthesis. Those results employ exact notions of abstractions based on simulation relations [Fre05, Kv10] and simulation maps [TPL04], for which constructive methodologies exist only for rather
restricted classes of control systems. In contrast to the exact notions, the compositional approximate abstractions were introduced recently which are useful for the controller synthesis. Examples include compositional
construction of finite abstractions of linear and nonlinear control systems [TI08, PPD16] and of infinite abstractions of nonlinear control systems [RZ15, RZ16a] and a class of stochastic hybrid systems [ZRMng]. In those
works, the abstraction (finite or infinite with possibly a lower dimension) can be used as a substitution of the
original system in the controller design process. The proposed results in [TI08, PPD16, RZ15, RZ16a, ZRMng]
use the small-gain type conditions to facilitate the compositional construction of abstractions. The resulting
small-gain type requirements intrinsically condition the spectral radius of the interconnection matrix which,
in general, depends on the size of the graph and can be violated or deteriorated as the number of subsystems
grows [DK04].
In this work we propose a novel compositional framework for the construction of infinite abstractions of
networks of control systems using dissipativity theory. First, we adapt the notion of storage function from
dissipativity theory [AMP16] to quantify the joint dissipativity-type properties of control subsystems and
their abstractions. Given a network of control subsystems and their storage functions, we propose conditions
based on the interconnection matrix and joint dissipativity-type properties of subsystems and their abstractions guaranteeing that the network of abstractions quantitatively approximate the behaviours of the network
of concrete subsystems. The proposed compositionality conditions can enjoy specific interconnection structures and provide scale-free compositional abstractions for large-scale control systems without requiring any
1
2
M. ZAMANI AND M. ARCAK
condition on the number or gains of the subsystems; we illustrate this point with an example in Section 6.
Furthermore, we provide a geometric approach on the construction of abstractions for a class of nonlinear
control systems and of their corresponding storage functions by using the bounds on the slope of system
nonlinearities.
Related Work. Compositional construction of infinite abstractions of networks of control systems is also
proposed in [RZ15, RZ16a]. While in [RZ15, RZ16a] small-gain type conditions are used to facilitate the
compositional construction of abstractions, here we use dissipativity-type conditions. The small-gain type
requirements inherently condition the spectral radius of the interconnection matrix which, in general, depends
on the size of the graph and can be dissatisfied as the number of subsystems grows [DK04]. On the other
hand, this is not necessarily the case with broader dissipativity-type conditions and in fact the compositionality
requirements may not condition the number or gains of the subsystems at all when the interconnection matrix
enjoys some properties (cf. Section 6). Although the results in [RZ15, RZ16a] provide constructive procedures
to determine abstractions of linear control systems, we propose techniques on the construction of abstractions
for a class of nonlinear control systems by using the bounds on the slope of systems nonlinearities. The
results in [RZ15, RZ16a] assume that the internal input and output space dimensions of each component in
a network are equal to the corresponding ones of its abstraction which is not the case in this paper. While
the interconnection matrix in [RZ15, RZ16a] is a permutation one, the one in this paper can be any general
interconnection matrix.
The recent results in [AMP16, MLAP15] establish only stability or stabilizability of networks of control systems
compositionally using dissipativity properties of components. On the other hand, the results here provide
construction of abstractions of networks of control systems compositionally using abstractions of components
and their joint dissipativity-type properties.
2. Control Systems
2.1. Notation. The sets of nonnegative integer and real numbers are denoted by N and R, respectively. Those
symbols are footnoted with subscripts to restrict them in the usual way, e.g. R>0 denotes the positive real
numbers. The symbol Rn×m denotes the vector space of real matrices with n rows and m columns. The
symbols 1n , 0n , In , and 0n×m denote the vector in Rn with all its elements to be one, the zero vector, identity
and zero matrices in Rn , Rn×n , and Rn×m , respectively. For a, b ∈ R with a ≤ b, the closed, open, and
half-open intervals in R are denoted by [a, b], ]a, b[, [a, b[, and ]a, b], respectively. For a, b ∈ N and a ≤ b,
the symbols [a; b], ]a; b[, [a; b[, and ]a; b] denote the corresponding intervals in N. Given N ∈ N≥1 , vectors
xi ∈ Rni , ni ∈ N≥1 and i ∈ [1; N ], we use x = [x1 ; . . . ; xN ] to denote the concatenated vector in Rn with
PN
n = i=1 ni . Given a vector x ∈ Rn , kxk denotes the Euclidean norm of x. Note that given any x ∈ RN ,
x ≥ 0 iff xi ≥ 0 for any i ∈ [1; N ]. Given a symmetric matrix A, λmax (A) and λmin (A) denote maximum and
minimum eigenvalues of A. We denote by diag(M1 , . . . , MN ) the block diagonal matrix with diagonal matrix
entries M1 , . . . , MN .
Given a function f : Rn → Rm and 0m ∈ Rm , we simply use f ≡ 0 to denote that f (x) = 0m for all x ∈ Rn .
Given a function f : R≥0 → Rn , the (essential) supremum of f is denoted by kf k∞ := (ess)sup{kf (t)k, t ≥ 0}.
A continuous function γ : R≥0 → R≥0 , is said to belong to class K if it is strictly increasing and γ(0) = 0; γ is
said to belong to class K∞ if γ ∈ K and γ(r) → ∞ as r → ∞. A continuous function β : R≥0 × R≥0 → R≥0
is said to belong to class KL if, for each fixed t, the map β(r, t) belongs to class K with respect to r and, for
each fixed nonzero r, the map β(r, t) is decreasing with respect to t and β(r, t) → 0 as t → ∞.
2.2. Control systems. The class of control systems studied in this paper is formalized in the following
definition.
Definition 2.1. A control system Σ is a tuple Σ = (Rn , Rm , Rp , U, W, f, Rq1 , Rq2 , h1 , h2 ), where Rn , Rm ,
Rp , Rq1 , and Rq2 are the state, external input, internal input, external output, and internal output spaces,
respectively, and
COMPOSITIONAL ABSTRACTION FOR NETWORKS OF CONTROL SYSTEMS: A DISSIPATIVITY APPROACH
3
• U and W are subsets of the sets of all measurable functions of time, from open intervals in R to Rm and
Rp , respectively;
• f : Rn × Rm × Rp → Rn is a continuous map satisfying the following Lipschitz assumption: for every
compact set D ⊂ Rn , there exists a constant Z ∈ R>0 such that kf (x, u, w) − f (y, u, w)k ≤ Zkx − yk for
all x, y ∈ D, all u ∈ Rm , and all w ∈ Rp ;
• h1 : Rn → Rq1 is the external output map;
• h2 : Rn → Rq2 is the internal output map.
A locally absolutely continuous curve ξ :]a, b[→ Rn is a state trajectory of Σ if there exist input trajectories
υ ∈ U and ω ∈ W satisfying:
˙
ξ(t) = f (ξ(t), υ(t), ω(t)),
Σ : ζ1 (t) = h1 (ξ(t)),
(2.1)
ζ2 (t) = h2 (ξ(t)),
for almost all t ∈ ]a, b[. We call the tuple (ξ, ζ1 , ζ2 , υ, ω) a trajectory of Σ, consisting of a state trajectory ξ,
output trajectories ζ1 and ζ2 , and input trajectories υ and ω, that satisfies (2.1). We also denote by ξxυω (t)
the state reached at time t under the inputs υ ∈ U, ω ∈ W from the initial condition x = ξxυω (0); the state
ξxυω (t) is uniquely determined due to the assumptions on f [Son98]. We also denote by ζ1xυω (t) and ζ2xυω (t)
the corresponding external and internal output value of ξxυω (t), respectively, i.e. ζ1xυω (t) = h1 (ξxυω (t)) and
ζ2xυω (t) = h2 (ξxυω (t)).
We call ζ1 an external output trajectory, ζ2 an internal output trajectory, υ an external input trajectory, and
ω an internal input trajectory mainly because ζ2 and ω are used only for the interconnection purposes and ζ1
and υ remain available after any interconnection; see Definition 4.1 later for more detailed information.
Remark 2.2. If the control system Σ does not have internal inputs and outputs, the definition of control
systems in Definition 2.1 reduces to tuple Σ = (Rn , Rm , U, f, Rq , h) and the map f becomes f : Rn × Rm → Rn .
Correspondingly, equation (2.1) describing the evolution of system trajectories reduces to:
(
˙ = f (ξ(t), υ(t)),
ξ(t)
Σ:
ζ(t) = h(ξ(t)).
3. Storage and Simulation Functions
First, we introduce a notion of so-called storage functions, adapted from the notion of storage functions from
dissipativity theory [Wil72, AMP16]. While the notion of storage functions in [Wil72, AMP16] characterizes
the correlation of inputs and outputs of a single control system, the proposed notion of storage functions here
characterizes the joint correlation of inputs and outputs of two different control systems. In the case that
two control systems are the same and have only internal inputs and outputs, our notion of storage functions
recovers the one of incremental storage functions introduced in [SS07].
Definition 3.1. Let Σ = (Rn , Rm , Rp , U, W, f, Rq1 , Rq2 , h1 , h2 ) and Σ̂ = (Rn̂ , Rm̂ , Rp̂ , Û, Ŵ, fˆ, Rq1 , Rq̂2 , ĥ1 , ĥ2 )
be two control systems with the same external output space dimension. A continuously differentiable function
V : Rn × Rn̂ → R≥0 is called a storage function from Σ̂ to Σ if there exist α, η ∈ K∞ , ρext ∈ K∞ ∪ {0}, some
matrices W, Ŵ , H of appropriate dimensions, and some symmetric matrix X of appropriate dimension with
conformal block partitions X ij , i, j ∈ [1; 2], where X 22 0, such that for any x ∈ Rn and x̂ ∈ Rn̂ one has
α(kh1 (x) − ĥ1 (x̂)k) ≤ V (x, x̂),
(3.1)
4
M. ZAMANI AND M. ARCAK
and ∀û ∈ Rm̂ ∃u ∈ Rm such that ∀ŵ ∈ Rp̂ ∀w ∈ Rp one obtains
X:=
T z 11 }| 12 {
W w − Ŵ ŵ
W w − Ŵ ŵ
X
X
T f (x, u, w)
∇V (x, x̂)
≤ −η(V (x, x̂)) + ρext (kûk) +
.
X 21 X 22 h2 (x) − H ĥ2 (x̂)
fˆ(x̂, û, ŵ)
h2 (x) − H ĥ2 (x̂)
(3.2)
We use notation Σ̂ Σ if there exists a storage function V from Σ̂ to Σ. Control system Σ̂ (possibly with
n̂ < n) is called an abstraction of Σ. There are several key differences between the notion of storage function
here and the corresponding one of simulation function in [RZ16a, Definition 2]. Definition 2 in [RZ16a] requires
internal signals w, ŵ and h2 (x), ĥ2 (x̂) to live in the same spaces, respectively, which is not necessarily the case
here. Moreover, the choice of input u here satisfying (3.2) only depends on x, x̂, and û, whereas in [RZ16a,
Definition 2] it also depends on internal input ŵ. Finally, we should point out that if in [RZ16a, Definition 2]
µ(s) := sT P s, for any s ∈ Rp≥0 and some positive definite matrix P , then the simulation function in [RZ16a,
Definition 2] is also a storage function as in Definition 3.1 with W = Ŵ = Ip , X 11 = P , and the rest of
conformal block partitions of X are zero.
Now, we recall the notion of simulation functions introduced in [GP09] with some modifications.
Definition 3.2. Let Σ = (Rn , Rm , U, f, Rq , h) and Σ̂ = (Rn̂ , Rm̂ , Û, fˆ, Rq , ĥ) be two control systems. A
continuously differentiable function V : Rn × Rn̂ → R≥0 is called a simulation function from Σ̂ to Σ if there
exist α, η ∈ K∞ and ρext ∈ K∞ ∪ {0} such that for any x ∈ Rn and x̂ ∈ Rn̂ one has
and ∀û ∈ R
m̂
∃u ∈ R
m
α(kh(x) − ĥ(x̂)k) ≤ V (x, x̂),
(3.3)
such that
T
∇V (x, x̂)
f (x, u)
≤ −η(V (x, x̂)) + ρext (kûk).
fˆ(x̂, û)
(3.4)
We use notation Σ̂ S Σ if there exists a simulation function V from Σ̂ to Σ.
Let us point out the differences between Definition 3.2 here and [GP09, Definition 1]. Here, for the sake
of brevity, we simply assume that for every x, x̂, û, there exists u so that (3.4) holds. Whereas in [GP09,
Definition 1] the authors use an interface function k : Rn × Rn̂ × Rm̂ → Rm to feed the input u = k(x, x̂, û)
enforcing (3.4). Function α in [GP09, Definition 1] is assumed to be the identity. Furthermore, we frame the
decay condition (3.4) in so-called “dissipative” form, while in [GP09, Definition 1] the decay condition is given
in so-called “implication” form.
Note that the notions of storage functions in Definition 3.1 and simulation functions in Definition 3.2 are not
comparable in general. The former is defined for control systems with internal inputs and outputs while the
latter is defined only for control systems without internal inputs and outputs. One can readily verify that
both notions coincide for control systems without internal inputs and outputs.
The next theorem shows the importance of the existence of a simulation function by quantifying the error
between the output behaviours of Σ and the ones of its abstraction Σ̂.
Theorem 3.3. Let Σ = (Rn , Rm , U, f, Rq , h) and Σ̂ = (Rn̂ , Rm̂ , Û, fˆ, Rq , ĥ). Suppose V is a simulation
function from Σ̂ to Σ. Then, there exist a KL function ϑ such that for any υ̂ ∈ Û, x ∈ Rn , and x̂ ∈ Rn̂ , there
exists υ ∈ U such that the following inequality holds for any t ∈ R≥0 :
kζxυ (t) − ζ̂x̂υ̂ (t)k ≤α−1 (2ϑ (V (x, x̂), t)) + α−1 (2η −1 (2ρext (kυ̂k∞ ))).
(3.5)
The proof of Theorem 3.3 is similar to the one of Theorem 3.5 in [ZRMng] and is omitted due to lack of space.
Let us illustrate the importance of the existence of a simulation function, correspondingly inequality (3.5),
on a simple example. Assume we are given a control system Σ = (Rn , Rm , U, f, Rq , h) and interested in
COMPOSITIONAL ABSTRACTION FOR NETWORKS OF CONTROL SYSTEMS: A DISSIPATIVITY APPROACH
5
computing a control input υ to keep the output ζxυ always inside a safe set D ⊂ Rq . Instead, one can
compute a control input υ̂ for the abstraction Σ̂ keeping the output ζ̂x̂υ̂ always inside D which is potentially
easier due to a lower dimension of Σ̂. The existence of a simulation function from Σ̂ to Σ and, hence,
the inequality (3.5) imply that there exists control input υ such that ζxυ is always inside Dε , where ε =
α−1 (2ϑ (V (x, x̂), 0)) + α−1 (2η −1 (2ρext (kυ̂k∞ ))) and Dε = {y ∈ Rp | inf y0 ∈D ky − y 0 k ≤ ε}. Note that one can
choose initial conditions x ∈ Rn and x̂ ∈ Rn̂ to minimize the first term in ε and, hence, to have a smaller error
in the satisfaction of the desired property.
Remark 3.4. Note that if α−1 and η −1 satisfy the triangle inequality (i.e., α−1 (a + b) ≤ α−1 (a) + α−1 (b)
and η −1 (a + b) ≤ η −1 (a) + η −1 (b) for all a, b ∈ R≥0 ), one can divide all the coefficients 2, appearing in the
right hand side of (3.5), by factor 2 to get a less conservative upper bound.
Remark 3.5. Note that if one is given an interface function k : Rn × Rn̂ × Rm̂ → Rm that maps every x, x̂,
û to an input u = k(x, x̂, û) such that (3.4) is satisfied (similar to [GP09, Definition 1]), then input υ realizing
ˆ υ̂). In Section 5 we show how the map k can be constructed for a class of
(3.5) is readily given by υ = k(ξ, ξ,
nonlinear control systems.
4. Compositionality Result
In this section, we analyze networks of control systems and show how to construct their abstractions together
with the corresponding simulation functions by using storage functions for the subsystems. The definition of
the network of control systems is based on the notion of interconnected systems described in [AMP16].
4.1. Interconnected control systems. Here, we define the interconnected control system as the following.
Definition 4.1. Consider N ∈ N≥1 control subsystems Σi = (Rni , Rmi , Rpi , Ui , Wi , fi , Rq1i , Rq2i , h1i , h2i ),
i ∈ [1; N ], and a static matrix M of an appropriate dimension defining the coupling of these subsystems. The
PN
interconnected control system Σ = (Rn , Rm , U, f, Rq , h), denoted by I(Σ1 , . . . , ΣN ), follows by n = i=1 ni ,
PN
PN
m = i=1 mi , q = i=1 q1i , and functions
f (x, u) := [f1 (x1 , u1 , w1 ); . . . ; fN (xN , uN , wN )],
h(x) := [h11 (x1 ); . . . ; h1N (xN )],
where u = [u1 ; . . . ; uN ], x = [x1 ; . . . ; xN ] and with the internal variables constrained by
[w1 ; . . . ; wN ] = M [h21 (x1 ); . . . ; h2N (xN )].
An interconnection of N control subsystems Σi is illustrated schematically in Figure 1.
⌃1
y
..
.
⌃N
M
[h21 (x1 ); . . . ; h2N (xN )]
[w1 ; . . . ; wN ]
u
Figure 1. An interconnection of N control subsystems Σ1 , . . . , ΣN .
6
M. ZAMANI AND M. ARCAK
4.2. Composing simulation functions from storage functions. We assume that we are given N control
subsystems Σi = (Rni , Rmi , Rpi , Ui , Wi , fi , Rq1i , Rq2i , h1i , h2i ) , together with their corresponding abstractions
Σ̂i = (Rn̂i , Rm̂i , Rp̂i , Ûi , Ŵi , fˆi , Rq1i , Rq̂2i , ĥ1i , ĥ2i ) and with storage functions Vi from Σ̂i to Σi . We use αi , ηi ,
ρiext , Hi , Wi , Ŵi , Xi , Xi11 , Xi12 , Xi21 , and Xi22 to denote the corresponding functions, matrices, and their
corresponding conformal block partitions appearing in Definition 3.1.
The next theorem, one of the main results of the paper, provides a compositional approach on the construction
of abstractions of networks of control systems and that of the corresponding simulation functions.
Theorem 4.2. Consider the interconnected control system Σ = I(Σ1 , . . . , ΣN ) induced by N ∈ N≥1 control
subsystems Σi and the coupling matrix M . Suppose each control subsystem Σi admits an abstraction Σ̂i with
the corresponding storage function Vi . If there exist µi ≥ 0, i ∈ [1; N ], and matrix M̂ of appropriate dimension
such that the matrix (in)equality
T
WM
WM
X(µ1 X1 , . . . , µN XN )
0,
(4.1)
Iq̃
Iq̃
W M H = Ŵ M̂ ,
are satisfied, where q̃ =
PN
i=1 q2i
(4.2)
and
W := diag(W1 , . . . , WN ), Ŵ := diag(Ŵ1 , . . . , ŴN ), H := diag(H1 , . . . , HN ),
µ1 X111
µ1 X112
..
..
.
.
11
12
µ
X
µ
X
N N
N N
X(µ1 X1 , . . . , µN XN ) :=
21
,
µ1 X122
µ1 X1
..
..
.
.
22
21
µN XN
µN XN
(4.3)
(4.4)
then
V (x, x̂) :=
N
X
µi Vi (xi , x̂i )
i=1
is a simulation function from the interconnected control system Σ̂ = I(Σ̂1 , . . . , Σ̂N ), with the coupling matrix
M̂ , to Σ.
Proof. First we show that inequality (3.3) holds for some K∞ function α. For any x = [x1 ; . . . ; xN ] ∈ Rn and
x̂ = [x̂1 ; . . . ; x̂N ] ∈ Rn̂ , one gets:
kh(x) − ĥ(x̂)k = k[h11 (x1 ); . . . ; h1N (xN )] − [ĥ11 (x̂1 ); . . . ; ĥ1N (x̂N )]k
≤
N
X
i=1
kh1i (x̂i ) − ĥ1i (xi )k ≤
where α is a K∞ function defined as
α(s) :=
max
→
s ≥0
s.t.
→
N
X
i=1
PN
i=1
αi−1 (Vi (xi , x̂i )) ≤ α(V (x, x̂)),
αi−1 (si )
→
µT s = s,
where s = [s1 ; . . . ; sN ] ∈ RN and µ = [µ1 ; . . . ; µN ]. By defining the K∞ function α(s) = α−1 (s), ∀s ∈ R≥0 ,
one obtains
α(kh(x) − ĥ(x̂)k) ≤ V (x, x̂),
satisfying inequality (3.3). Now we show that inequality (3.4) holds as well. Consider any x = [x1 ; . . . ; xN ] ∈
Rn , x̂ = [x̂1 ; . . . ; x̂N ] ∈ Rn̂ , and û = [û1 ; . . . ; ûN ] ∈ Rm̂ . For any i ∈ [1; N ], there exists ui ∈ Rmi , consequently,
COMPOSITIONAL ABSTRACTION FOR NETWORKS OF CONTROL SYSTEMS: A DISSIPATIVITY APPROACH
7
a vector u = [u1 ; . . . ; uN ] ∈ Rm , satisfying (3.2) for each pair of subsystems Σi and Σ̂i with the internal inputs
given by [w1 ; . . . ; wN ] = M [h21 (x1 ); . . . ; h2N (xN )] and [ŵ1 ; . . . ; ŵN ] = M̂ [ĥ21 (x̂1 ); . . . ; ĥ2N (x̂N )]. We derive
the following inequality
V̇ (x, x̂) =
N
X
µi V̇i (xi , x̂i )
(4.5)
i=1
≤
N
X
µi
i=1
T 11
Wi wi − Ŵi ŵi
Xi
Xi21
h2i (xi ) − Hi ĥ2i (x̂i )
− ηi (Vi (xi , x̂i )) + ρiext (kûi k) +
Xi12
Xi22
Wi wi − Ŵi ŵi
.
h2i (xi ) − Hi ĥ2i (x̂i )
Using conditions (4.1) and (4.2) and the definition of matrices W , Ŵ , H, and X in (4.3) and (4.4), the
inequality (4.5) can be rewritten as
V̇ (x, x̂) ≤
N
X
i=1
−µi ηi (Vi (xi , x̂i )) +
+
≤
N
X
i=1
N
X
i=1
µi ρiext (kûi k)
T
w1
ŵ1
w1
ŵ1
..
..
..
..
W . − Ŵ .
W . − Ŵ .
wN
ŵN
wN
ŵN
X(µ1 X1 , . . . , µN XN )
h21 (x1 ) − H1 ĥ21 (x̂1 )
h21 (x1 ) − H1 ĥ21 (x̂1 )
..
..
.
.
h2N (xN ) − HN ĥ2N (x̂N )
−µi ηi (Vi (xi , x̂i )) +
N
X
i=1
h2N (xN ) − HN ĥ2N (x̂N )
µi ρiext (kûi k)
T
h21 (x1 ) − H1 ĥ21 (x̂1 )
h21 (x1 ) − H1 ĥ21 (x̂1 )
T
WM
WM
..
..
+
X(µ1 X1 , . . . , µN XN )
.
.
Iq̃
Iq̃
h2N (xN ) − HN ĥ2N (x̂N )
h2N (xN ) − HN ĥ2N (x̂N )
≤
N
X
i=1
−µi ηi (Vi (xi , x̂i )) +
N
X
i=1
µi ρiext (kûi k).
Define the functions
η(s) :=
ρext (s) :=
min
→
PN
s.t.
max
→
µT s = s,
PN
i=1 µi ρiext (si )
s.t.
k s k = s,
s ≥0
s ≥0
i=1
µi ηi (si )
→
→
(4.6a)
(4.6b)
where η ∈ K∞ and ρext ∈ K∞ ∪ {0}. By construction, we readily have
V̇ (x, x̂) ≤ −η (V (x, x̂)) + ρext (kûk),
which satisfies inequality (3.4). Hence, we conclude that V is a simulation function from Σ̂ to Σ.
Figure 2 illustrates schematically the result of Theorem 4.2.
Remark 4.3. Let us assume, ∀i ∈ [1; N ], Wi = Ipi and each control subsystem Σi is single-internal-input
single-internal-output. Under these assumptions, analytical feasibility conditions for matrix inequality (4.1)
can be derived for special interconnection matrices M including negative and positive feedback interconnection,
M. ZAMANI AND M. ARCAK
ûi
ˆi
⌃
ŵi
ĥ2i (xi )
wi
ŷ
..
u
.
ˆN
⌃
M̂
h1i (xi )
⌃i
⌃1
S
[w1 ; . . . ; wN ]
ˆ1
⌃
ui
[ĥ21 (x1 ); . . . ; ĥ2N (xN )]
[ŵ1 ; . . . ; ŵN ]
û
ĥ1i (xi )
h2i (xi )
y
..
.
⌃N
M
[h21 (x1 ); . . . ; h2N (xN )]
8
Figure 2. Compositionality results provided that conditions (4.1) and (4.2) are satisfied.
skew symmetric interconnection, negative feedback cyclic interconnection, and finally extension to cactus graphs
as provided in details in [AMP16, Chapter 2].
5. Abstraction Synthesis for a Class of Nonlinear Control Systems
In this section, we concentrate on a specific class of nonlinear control systems Σ and quadratic storage functions
V . In the first part, we formally define the specific class of nonlinear control systems with which we deal in
this section. In the second part, we assume that an abstraction Σ̂ is given and we provide conditions under
which V is a storage function. In the third part it is shown geometrically how to construct the abstraction Σ̂
together with the storage function V . Finally, we discuss the feasibility of a key condition based on which the
results of this section hold.
5.1. A class of nonlinear control systems. The class of nonlinear control systems, considered in this
section, is given by
ξ˙ = Aξ + Eϕ(F ξ) + Bυ + Dω,
Σ : ζ1 = C1 ξ,
(5.1)
ζ2 = C2 ξ,
where ϕ : R → R satisfies
ϕ(v) − ϕ(w)
≤ b, ∀v, w ∈ R, v 6= w,
v−w
for some a ∈ R and b ∈ R>0 ∪ {∞}, a ≤ b, and
a≤
(5.2)
A ∈ Rn×n , E ∈ Rn×1 , F ∈ R1×n , B ∈ Rn×m , D ∈ Rn×p , C1 ∈ Rq1 ×n , C2 ∈ Rq2 ×n .
We use the tuple
Σ = (A, B, C1 , C2 , D, E, F, ϕ),
to refer to the class of control systems of the form (5.1).
Remark 5.1. If ϕ in (5.1) is linear including the zero function (i.e. ϕ ≡ 0) or E is a zero matrix, one can
remove or push the term Eϕ(F ξ) to Aξ and, hence, the tuple representing the class of control systems reduces
to the linear one Σ = (A, B, C1 , C2 , D). Therefore, every time we use the tuple Σ = (A, B, C1 , C2 , D, E, F, ϕ),
it implicitly implies that ϕ is nonlinear and E is nonzero.
COMPOSITIONAL ABSTRACTION FOR NETWORKS OF CONTROL SYSTEMS: A DISSIPATIVITY APPROACH
9
T c
c (A + BK) M
cZ M
c(BL1 + E)
c + C T X 22 C2 C T X 21 −F T
(A + BK) M
+M
−b
κM
2
2
c
(5.5)
X 12 C2
X 11
0
ZT M
0
0
2
T c
−F
0
(BL1 + E) M
0
0
b
——————————————————————————————————————————————–
Similar to what is shown in [AK01a], without loss of generality, we can assume a = 0 in (5.2) for the class of
nonlinear control systems in (5.1). If a 6= 0, one can define a new function ϕ(r)
e
:= ϕ(r) − ar which satisfies
e
(5.2) with e
a = 0 and b = b − a, and rewrite (5.1) as
e + E ϕ(F
e ξ) + Bυ + Dω,
ξ˙ = Aξ
Σ : ζ1 = C1 ξ,
ζ2 = C2 ξ,
e = A + aEF .
where A
Remark 5.2. For simplicity of derivations, we restrict ourselves to systems with a single nonlinearity as in
(5.1). However, it would be straightforward to obtain analogous results for systems with multiple nonlinearities
as
M
X
ξ˙ = Aξ +
Ei ϕi (Fi ξ) + Bυ + Dω,
Σ:
i=1
ζ = C1 ξ,
1
ζ2 = C 2 ξ
where ϕi : R → R satisfies (5.2) for some ai ∈ R and bi ∈ R>0 ∪ {∞}, Ei ∈ Rn×1 , and Fi ∈ R1×n , for
any i ∈ [1; M ]. Furthermore, the proposed results here can also be extended to systems with multivariable
nonlinearities satisfying a multivariable sector property along the same lines as in [FA03] in the context of
observer design.
Note that the class of nonlinear control systems in (5.1) and Remark 5.2 has been used widely to model many
physical systems including active magnetic bearing [AK01a], flexible joint robot [FA03], fuel cell [AGPV03],
the power generators [Sch04], underwater vehicles [AAFK01], and so on.
5.2. Quadratic storage functions. Here, we consider a quadratic storage function of the form
c(x − P x̂),
V (x, x̂) = (x − P x̂)T M
(5.3)
c 0 are some matrices of appropriate dimensions. In order to show that V in (5.3) is a storage
where P and M
function from an abstraction Σ̂ to a concrete system Σ, we require the following key assumption on Σ.
Assumption 1. Let Σ = (A, B, C1 , C2 , D, E, F, ϕ). Assume that for some constant κ
b ∈ R>0 there exist
c 0, K, L1 , Z, W , X 11 , X 12 , X 21 , and X 22 0 of appropriate dimensions such that the matrix
matrices M
equality
D = ZW,
(5.4)
and inequality (5.5) hold, where 0’s in (5.5) denote zero matrices of appropriate dimensions.
The next rather straightforward result provides a necessary and sufficient geometric condition for the existence
of matrix W appearing in condition (5.4).
Lemma 5.3. Given D and Z, condition (5.4) is satisfied for some matrix W if and only if
im D ⊆ im Z.
(5.6)
10
M. ZAMANI AND M. ARCAK
22
21
−b
κM + X
X
−M F T
Z BL1 + E
12
(5.7)
0
0
X
X 11
0
2
E T + LT1 B T
0
0
−F M
0
b
——————————————————————————————————————————————–
T
M AT + AM + K B T + BK
ZT
Note that the feasibility characterization of LMI (5.5) is more involved and will be discussed in details at the
end of this section.
c, K, L1 , Z, and linear in the variables X 11 ,
Note that matrix inequality (5.5) is bilinear in the variables M
12
21
22
X , X , and X when we fix the constant κ
b. However, by assuming C2 is a square and invertible matrix
−1
c
c−1 , X 22 = M
c−1 C T X 22 C2 M
c−1 , X 21 = M
c−1 C T X 21 ,
and introducing new variables K = K M , M = M
2
2
12
12
−1
c , and multiplying (5.5) from both sides by
X = X C2 M
c−1 0 0
M
0
In 0 ,
0
0 1
where 0’s denote zero matrices of appropriate dimension, one obtains the matrix inequality (5.7) which is an
22
21
12
b is a fixed
LMI (linear matrix inequality) in the variables M , K, L1 , Z, X , X , X , and X 11 when κ
constant.
Remark 5.4. Note that one can combine the compositionality condition (4.1) with a simultaneous search
for quadratic storage functions (5.3) for subsystems of the form (5.1). In particular, assume we are given
N control subsystems Σi = (Ai , Bi , C1i , C2i , Di , Ei , Fi , ϕi ), ∀i ∈ [1; N ]. For any i ∈ [1; N ], one can consider
matrices Xi in the LMI (4.1) as decision variables instead of being fixed and µi = 1 without loss of generality,
and solve the combined feasibility problems (5.5) and (4.1). Although the combined feasibility problem may be
huge for large networks and solving it directly may be intractable, one can use the alternating direction method
of multipliers (ADMM) to solve the feasibility problem in a distributed fashion along the same lines proposed
in [MLAP15].
Now, we provide one of the main results of this section showing under which conditions V in (5.3) is a storage
function.
Theorem 5.5. Let Σ = (A, B, C1 , C2 , D, E, F, ϕ) and Σ̂ = (Â, B̂, Ĉ1 , Ĉ2 , D̂, Ê, F̂ , ϕ) with q1 = q̂1 . Suppose
Assumption 1 holds and that there exist matrices P , Q, H, L2 , and Ŵ such that
AP = P Â − BQ
C1 P = Ĉ1
X 12 C2 P = X 12 H Ĉ2
22
22
X C2 P = X H Ĉ2
F P = F̂
E = P Ê − B(L1 − L2 )
P D̂ = Z Ŵ ,
(5.8a)
(5.8b)
(5.8c)
(5.8d)
(5.8e)
(5.8f)
(5.8g)
hold. Then, function V defined in (5.3) is a storage function from Σ̂ to Σ.
Before providing the proof, we point out that there always exist matrices {Â, B̂, Ĉ1 , Ĉ2 , D̂, Ê, F̂ } satisfying
(5.8) if P = In implying that n̂ = n. Naturally, it is better to have the simplest abstraction Σ̂ and, therefore,
one should seek a P with n̂ as small as possible. We elaborate on the construction of P satisfying (5.8) in
details in the next subsection.
COMPOSITIONAL ABSTRACTION FOR NETWORKS OF CONTROL SYSTEMS: A DISSIPATIVITY APPROACH
11
Proof. From (5.8b) and for all x ∈ Rn , x̂ ∈ Rn̂ , we have kC1 x − Ĉ1 x̂k2 = (x − P x̂)T C1T C1 (x − P x̂). It can be
c)
λmin (M
kC1 x − Ĉ1 x̂k2 ≤ V (x, x̂) holds for all x ∈ Rn , x̂ ∈ Rn̂ implying that inequality
λmax (C1T C1 )
c)
λmin (M
α(r) = λmax
r2 for any r ∈ R≥0 . We proceed with showing that the inequality (3.2)
(C1T C1 )
readily verified that
(3.1) holds with
holds. Note that
∂V (x, x̂)
c, ∂V (x, x̂) = −2(x − P x̂)T M
cP.
= 2(x − P x̂)T M
(5.9)
∂x
∂ x̂
Given any x ∈ Rn , x̂ ∈ Rn̂ , and û ∈ Rm̂ , we choose u ∈ Rm via the following linear interface function:
e + L1 ϕ(F x) − L2 ϕ(F P x̂),
u =k(x, x̂, û) := K(x − P x̂) + Qx̂ + Rû
(5.10)
e of appropriate dimension.
for some matrix R
By using the equations (5.8a), (5.8e), and (5.8f) and the definition of the interface function in (5.10), we get
Ax + Eϕ(F x) + Bk(x, x̂, û) + Dw − P (Âx̂ + Êϕ(F̂ x̂) + B̂ û + D̂ŵ) =
e − P B̂)û + (E + BL1 )(ϕ(F x) − ϕ(F P x̂)).
(A + BK)(x − P x̂) + (Dw − P D̂ŵ) + (B R
Using (5.9), (5.4), and (5.8g), we obtain the following expression for V̇ (x, x̂):
c (A + BK)(x − P x̂) + (ZW w − Z Ŵ ŵ)
V̇ (x, x̂) =2(x − P x̂)T M
e − P B̂)û + (E + BL1 )(ϕ(F x) − ϕ(F P x̂)) .
+ (B R
From the slope restriction (5.2), one obtains
ϕ(F x) − ϕ(F P x̂) = δ(F x − F P x̂) = δF (x − P x̂),
(5.11)
where δ is a constant and depending on x and x̂ takes values in the interval [0, b]. Using (5.11), the expression
for V̇ (x, x̂) reduces to:
c ((A + BK) + δ(E + BL1 )F )(x − P x̂) + Z(W w − Ŵ ŵ) + (B R
e − P B̂)û .
V̇ (x, x̂) = 2(x − P x̂)T M
Using Young’s inequality [You12] as
1
2
a + b2 ,
2
2
for any a, b ∈ R and any > 0, and with the help of Cauchy-Schwarz inequality, (5.5), (5.8c), and (5.8d), one
gets the following upper bound for V̇ (x, x̂):
c ((A + BK) + δ(E + BL1 )F )(x − P x̂) + Z(W w − Ŵ ŵ) + (B R
e − P B̂)û
V̇ (x, x̂) = 2(x − P x̂)T M
T
T c
c (A + BK) M
cZ M
c(BL1 + E) x − P x̂
(A + BK) M
+M
x − P x̂
c
= W w − Ŵ ŵ
ZT M
0
0
W w − Ŵ ŵ
c
δF (x − P x̂)
δF (x − P x̂)
(BL1 + E)T M
0
0
ab ≤
c(B R
e − P B̂)û
+ 2(x − P x̂)T M
T
c + C T X 22 C2
x − P x̂
−b
κM
2
≤ W w − Ŵ ŵ
X 12 C2
δF (x − P x̂)
−F
x − P x̂
−F T
c(B R
e − P B̂)û
0 W w − Ŵ ŵ + 2(x − P x̂)T M
2
δF (x − P x̂)
b
T 11
δ
W w − Ŵ ŵ
X
X 12
W w − Ŵ ŵ
=−κ
bV (x, x̂) − 2δ(1 − )(x − P x̂)T F T F (x − P x̂) +
X 21 X 22 C2 x − H Ĉ2 x̂
b
C2 x − H Ĉ2 x̂
C2T X 21
X 11
0
c(B R
e − P B̂)û
+ 2(x − P x̂)T M
p
T 11
c(B R
e − P B̂)k2
k M
W w − Ŵ ŵ
X
≤ − (b
κ − π)V (x, x̂) +
kûk2 +
X 21
π
C2 x − H Ĉ2 x̂
for any positive constant π < κ
b.
X 12
X 22
W w − Ŵ ŵ
,
C2 x − H Ĉ2 x̂
12
M. ZAMANI AND M. ARCAK
Using this computed upper bound, the inequality (3.2)
ρext ∈ K∞ ∪{0},
√ is satisfied with the functions η ∈ K∞ , 11
2
c
e
X
X 12
B̂)k 2
and the matrix X, as η(s) := (b
κ − π)s, ρext (s) := k M (B R−P
s , ∀s ∈ R≥0 , and X =
.
π
X 21 X 22
The next result shows that conditions (5.8a)-(5.8f) are actually necessary for (5.3) being a storage function
e
from Σ̂ to Σ provided that the structure of the interface function is as in (5.10) for some matrices K, Q, R,
L1 , and L2 of appropriate dimension.
Theorem 5.6. Let Σ = (A, B, C1 , C2 , D, E, F, ϕ) and Σ̂ = (Â, B̂, Ĉ1 , Ĉ2 , D̂, Ê, F̂ , ϕ) with q1 = q̂1 . Suppose that V defined in (5.3) is a storage function from Σ̂ to Σ with the interface k given in (5.10). Then,
equations (5.8a)-(5.8f) hold.
Proof. Since V is a storage function from Σ̂ to Σ, there exists a K∞ function α such that kC1 x − Ĉ1 x̂k ≤
α−1 (V (x, x̂)). From (5.3), it follows that kC1 P x̂ − Ĉ1 x̂k ≤ α−1 (V (P x̂, x̂)) = 0 holds for all x̂ ∈ Rn̂ which
implies (5.8b).
Let us consider the inputs υ̂ ≡ 0, ω ≡ 0, and ω̂ ≡ 0. Since X 22 0, inequality (3.2) reduces to
V̇ (x, x̂) ≤ −η(V (x, x̂)) + (h2 (x) − H ĥ2 (x̂))T X 22 (h2 (x) − H ĥ2 (x̂)) ≤ −η(V (x, x̂)),
(5.12)
for any x ∈ Rn and x̂ ∈ Rn̂ . Using the results in Lemma 4.4 in [LSW96] or Lemma 3.6 in [ZRMng], inequality
(5.12) implies the existence of a KL function ϑ such that
ˆ
ˆ
V (ξ(t), ξ(t))
≤ ϑ(V (ξ(0), ξ(0)),
t),
(5.13)
holds, where υ̂ ≡ 0, ω ≡ 0, ω̂ ≡ 0, and υ is given by the interface function k in (5.10). Then, for all
ˆ
ˆ
ξ(0) = P ξ(0),
t ≥ 0, and using (5.13), we obtain V (ξ(t), ξ(t))
= 0. Since M is positive definite, we have
ˆ
ξ(t) = P ξ(t)
and
˙ = P ξ(t),
ˆ˙
ξ(t)
from which we derive that
AP x̂ + BQx̂ + (E + B(L1 − L2 ))ϕ(F P x̂) = P Âx̂ + P Êϕ(F̂ x̂)
n̂
holds for all x̂ ∈ R and, hence, (5.8a), (5.8e), and (5.8f) follows. It remains to show that (5.8c) and (5.8d)
hold. First assume X 22 6= 0. Since V̇ (P x̂, x̂) = V (P x̂, x̂) = 0 and using the first inequality in (5.12), one gets
(C2 P x̂ − H Ĉ2 x̂)T X 22 (C2 P x̂ − H Ĉ2 x̂) ≥ 0,
for any x̂ ∈ Rn̂ . Since X 22 0 and by assumption X 22 6= 0, one obtains X 22 (C2 P − H Ĉ2 ) = 0 which implies
(5.8d). Now, let us consider the inputs υ̂ ≡ 0, ω 6≡ 0, and ω̂ 6≡ 0. Therefore, inequality (3.2) reduces to
V̇ (x, x̂) ≤ − η(V (x, x̂)) + (W w − Ŵ ŵ)T X 11 (W w − Ŵ ŵ) + 2(W w − Ŵ ŵ)T X 12 (C2 x − H Ĉ2 x̂),
(5.14)
for any x ∈ Rn , x̂ ∈ Rn̂ , w ∈ Rp , and ŵ ∈ Rp̂ . From (5.14) and by choosing x = 0n and x̂ = 0n̂ , one can
readily verify that X 11 0. Then, for all x = P x̂, we obtain
(W w − Ŵ ŵ)T X 11 (W w − Ŵ ŵ) + 2(W w − Ŵ ŵ)T X 12 (C2 P − H Ĉ2 )x̂ ≥ 0,
for any w, ŵ, and x̂, which implies X 12 (C2 P − H Ĉ2 ) = 0 and, hence, (5.8c) holds.
e is a free design parameter in the interface function (5.10). Using the results
Remark 5.7. Note that matrix R
e to minimize function ρext for V and, hence, reduce the upper bound in
in [GP09, Proposition 1], we choose R
e minimizing ρext is given by
(3.5) on the error between the output behaviors of Σ and Σ̂. The choice of R
e = (B T M
cB)−1 B T M
cP B̂.
R
(5.15)
COMPOSITIONAL ABSTRACTION FOR NETWORKS OF CONTROL SYSTEMS: A DISSIPATIVITY APPROACH
13
So far, we extracted various conditions on the original system matrices {A, B, C1 , C2 , D, E, F }, the abstraction
matrices {Â, B̂, Ĉ1 , Ĉ2 , D̂, Ê, F̂ }, and the ones appearing in (5.3) and (5.10). Those conditions ensure that
V in (5.3) is a storage function from Σ̂ to Σ with the corresponding interface function in (5.10) refining any
control signal designed for Σ̂ to the one for Σ. Apparently, those requirements do not enforce any condition
on matrix B̂. For example, one can select B̂ = In̂ making the abstract system Σ̂ fully actuated. On the other
hand, one can ask not only for the existence of a storage function from Σ̂ to Σ, but additionally require that
all the controllable behaviors (in the absence of internal inputs) of the concrete system Σ are preserved over
the abstraction Σ̂. We refer the interested readers to [GP09, Subsection 4.1] and [PLS00, Section V] for more
details on what we mean by preservation of controllable behaviors.
The next theorem requires a condition on B̂ in order to guarantee the preservation of controllable behaviors
of Σ over Σ̂.
Theorem 5.8. Let Σ = (A, B, C1 , C2 , D, E, F, ϕ) and Σ̂ = (Â, B̂, Ĉ1 , Ĉ2 , D̂, Ê, F̂ , ϕ) with q1 = q̂1 . Suppose
that there exist matrices P , Q, L1 , and L2 satisfying (5.8a) and (5.8f), and that matrix B̂ is given by
B̂ = [P̂ B P̂ AG],
(5.16)
where P̂ and G are assumed to satisfy
C1 = Ĉ1 P̂
(5.17a)
In = P P̂ + GT
(5.17b)
In̂ = P̂ P
(5.17c)
F = F̂ P̂ ,
(5.17d)
ˆ ζ̂1 , ζ̂2 , υ̂, 0) of Σ̂
for some matrix T . Then, for every trajectory (ξ, ζ1 , ζ2 , υ, 0) of Σ there exists a trajectory (ξ,
where ξˆ = P̂ ξ and ζ1 = ζ̂1 hold.
Proof. Let (ξ, ζ1 , ζ2 , υ, 0) be a trajectory of Σ. We are going to show that (P̂ ξ, ζ1 , ζ̂2 , υ̂, 0) with
υ − QP̂ ξ − (L1 − L2 )ϕ(F ξ)
υ̂ =
,
Tξ
is a trajectory of Σ̂. We use (5.17b) and derive
P̂ ξ˙ =P̂ Aξ + P̂ Eϕ(F ξ) + P̂ Bυ = P̂ AP P̂ ξ + P̂ A(In − P P̂ )ξ + P̂ Eϕ(F ξ) + P̂ Bυ
=P̂ AP P̂ ξ + P̂ AGT ξ + P̂ Eϕ(F ξ) + P̂ Bυ.
Now we use the equations (5.8a), (5.8f), (5.17c), and (5.17d) and the definition of B̂ and υ̂ to derive
P̂ ξ˙ =P̂ (P Â − BQ)P̂ ξ + P̂ AGT ξ + P̂ (P Ê − B(L1 − L2 ))ϕ(F ξ) + P̂ Bυ
=ÂP̂ ξ − P̂ BQP̂ ξ + P̂ AGT ξ + Êϕ(F̂ P̂ ξ) − P̂ B(L1 − L2 )ϕ(F ξ) + P̂ Bυ
=ÂP̂ ξ + Êϕ(F̂ P̂ ξ) + [P̂ B P̂ AG]υ̂ = ÂP̂ ξ + Êϕ(F̂ P̂ ξ) + B̂ υ̂,
showing that (P̂ ξ, ζ̂1 , ζ̂2 , υ̂, 0) is a trajectory of Σ̂. From C1 = Ĉ1 P̂ in (5.17a), it follows that ζ̂1 = ζ1 which
concludes the proof.
Remark 5.9. Note that the previous result establishes that Σ̂ (in the absence of internal inputs) is P̂ -related
to Σ as in [GP09, Definition 3]. We refer the interested readers to [PLS00] for more details about properties
(e.g. controllability) of Φ-related systems for some surjective smooth map Φ.
14
M. ZAMANI AND M. ARCAK
5.3. Construction of abstractions. Here, we provide several straightforward sufficient and necessary geometric conditions on matrices appearing in the definition of Σ̂, of storage function and its corresponding
interface function. The proposed geometric conditions facilitate the constructions of such matrices. First, we
recall [GP09, Lemma 2] providing necessary and sufficient conditions for the existence of matrices  and Q
appearing in condition (5.8a).
Lemma 5.10. Consider matrices A, B, and P . There exist matrices  and Q satisfying (5.8a) if and only if
im AP ⊆ im P + im B.
(5.18)
Now, we give necessary and sufficient conditions for the existence of matrices Ĉ2 , Ê, and L2 appearing in
conditions (5.8c), (5.8d), and (5.8f), respectively.
Lemma 5.11. Given P , C2 , and X 12 (resp. X 22 ), there exists matrix Ĉ2 satisfying (5.8c) (resp. (5.8d)) if
and only if
im X 12 C2 P ⊆ im X 12 H, (resp. im X 22 C2 P ⊆ im X 22 H)
(5.19)
for some matrix H of appropriate dimension.
Lemma 5.12. Given P , B, and L1 , there exist matrices Ê and L2 satisfying (5.8f) if and only if
im E ⊆ im P + im B.
(5.20)
Lemmas 5.10, 5.11, and 5.12 provide necessary and sufficient conditions on P and H resulting in the construction of matrices Â, Ĉ2 , and Ê together with the matrices Q and L2 appearing in the definition of the interface
function in (5.10). Matrices F̂ and Ĉ1 are computed as F̂ = F P and Ĉ1 = C1 P . The next lemma provides a
necessary and sufficient condition on the existence of matrix D̂ appearing in condition (5.8g).
Lemma 5.13. Given Z, there exists matrix D̂ satisfying (5.8g) if and only if
im Z Ŵ ⊆ im P,
(5.21)
for some matrix Ŵ of appropriate dimension.
Although condition (5.21) is readily satisfied by choosing Ŵ = 0, one should preferably aim at finding a
nonzero Ŵ to smooth later the satisfaction of compositionality condition (4.2).
As we already mentioned, the choice of matrix B̂ is free. One can also construct B̂ as in (5.16) ensuring
preservation of all controllable behaviors of Σ over Σ̂ under extra conditions given in (5.17). Lemma 3 in
[GP09], as recalled next, provides necessary and sufficient conditions on P and C1 for the existence of P̂ , G,
and T satisfying (5.17a), (5.17b), and (5.17c).
Lemma 5.14. Consider matrices C1 and P with P being injective and let Ĉ1 = C1 P . There exists matrix P̂
satisfying (5.17a), (5.17b), and (5.17c), for some matrices G and T of appropriate dimensions, if and only if
im P + ker C1 = Rn .
(5.22)
Similar to Lemma 5.14, we give necessary and sufficient conditions on P and F for the existence of P̂ satisfying
(5.17d).
Lemma 5.15. Consider matrices F and P with P being injective and let F̂ = F P . There exists matrix P̂
satisfying (5.17d) if and only if
im P + ker F = Rn .
(5.23)
Note that conditions (5.5), (5.6), and (5.18)-(5.21) (resp. (5.5), (5.6), and (5.18)-(5.23)) complete the characterization of mainly matrices P and Z which together with the matrices {A, B, C1 , C2 , D, E, F } result in the
COMPOSITIONAL ABSTRACTION FOR NETWORKS OF CONTROL SYSTEMS: A DISSIPATIVITY APPROACH
15
construction of matrices {Â, B̂, Ĉ1 , Ĉ2 , D̂, Ê, F̂ }, where B̂ can be chosen freely with appropriate dimensions
(resp. B̂ is computed as in (5.16)).
We summarize the construction of the abstraction Σ̂, storage function V in (5.3), and its corresponding
interface function in (5.10) in Table 1.
Table 1. Construction of Σ̂ = (Â, B̂, Ĉ1 , Ĉ2 , D̂, Ê, F̂ , ϕ), the corresponding storage function
V in (5.3), and interface function in (5.10) for a given Σ = (A, B, C1 , C2 , D, E, F, ϕ).
c, K, L1 , Z, X 11 , X 12 , and X 22 satisfying (5.5) and (5.6).
1. Compute matrices M
2. Pick an injective P with the lowest rank satisfying (5.18)-(5.21) (resp. (5.18)-(5.23));
3. Compute  and Q from (5.8a);
4. Compute Ê and L2 from (5.8f);
5. Compute F̂ = F P ;
6. Compute Ĉ1 = CP ;
6. Compute Ĉ2 satisfying H Ĉ2 = CP for some H;
7. Compute D̂ satisfying P D̂ = Z Ŵ for some (rather nonzero) Ŵ ;
8. Choose B̂ freely (resp. B̂ = [P̂ B P̂ AG]);
e appearing in (5.10), from (5.15);
9. Compute R,
5.4. Feasibility of LMI (5.5). In this subsection we discuss sufficient and necessary feasibility conditions
T c
for the LMI (5.5) in the restrictive case of X 12 = 0 and X 11 Z πM Z for any positive constant π < κ
b,
where 0 denotes a zero matrix of appropriate dimension. To do so, we convert the feasibility conditions for
the restricted version of LMI (5.5) into the ones for two dual control problems. When b = ∞ in (5.5), the
feasibility of restricted (5.5) is dual to the one of designing a controller rendering a linear system strictly
positive real (SPR) [AK01b]. When b < ∞, the duality is with a linear L2 -gain assignment control problem
[Isi99, Section 13.2].
When b = ∞, the restricted version of LMI (5.5) reduces to
T c
c (A + BK) ≺ 0,
+M
(A + BK) M
c(BL1 + E) + F T = 0.
M
(5.24)
(5.25)
By virtue of the Positive-Real Lemma [Yak62], conditions (5.24) and (5.25) mean that the linear control system
ξ˙ = Aξ + Bυ + Eω,
Σ:
(5.26)
ζ = −F ξ,
is enforced SPR from the disturbance ω to the output ζ by the control law
υ = Kξ + L1 ω.
(5.27)
Therefore, when b = ∞, the feasibility of the restricted version of LMI (5.5) is dual to the feasibility of the
control problem in which the system (5.26) is enforced SPR by the control law (5.27).
When b < ∞, using the Schur complement of −2/b, one can readily verify that the restricted version of LMI
(5.5) is equivalent to
T
b
c+M
c A + BK + b (BL1 + E)F
A + BK + (BL1 + E)F
M
2
2
bc
c + b F T F ≺ 0,
+ M
(BL1 + E)(BL1 + E)T M
2
2
16
M. ZAMANI AND M. ARCAK
which means that the L2 -gain of the dual system (5.26) from input ω
e := ω + (b/2)ζ to output ζ is enforced to
be strictly less than 2/b by the control law υ = Kξ + L1 ω [Isi99, Section 13.2].
We refer the interested readers to [AK01a, Theorem 3] deriving sufficient and necessary feasibility conditions
for the restricted version of LMI (5.5) by looking into the corresponding dual control problems, namely,
enforcing SPR and assigning a linear L2 -gain.
Note that in the context of observer design and observer-based control, the feasibility of those dual control
problems have been investigated for several physical problems in [AK01a, FA03, AGPV03, Sch04].
6. Example
Consider a linear control system Σ = (−L, In , C) satisfying
ξ˙ = −Lξ + υ,
Σ:
ζ = Cξ,
for some matrix C ∈ Rq×n and L ∈ Rn×n . Assume L is the Laplacian matrix [GR01] of an undirected graph,
e.g., for a complete graph:
n−1
−1
···
···
−1
−1
n−1
−1
···
−1
−1
−1
n
−
1
·
·
·
−1
(6.1)
L=
,
..
..
..
..
.
.
.
.
−1
···
···
−1 n − 1
and C has the following block diagonal structure
C = diag(C11 , . . . , C1N ),
q1i ×ni
where C1i ∈ R
. We partition ξ as ξ = [ξ1 ; . . . ; ξN ] and υ as υ = [υ1 ; . . . ; υN ] where ξi and υi are both
taking values in Rni , ∀i ∈ [1; N ]. Now, by introducing Σi = (0ni , Ini , C1i , Ini , Ini ) satisfying
ξ˙i = ωi + υi ,
Σi :
ζ1i = C1i ξi ,
ζ2i = ξi ,
one can readily verify that Σ = I(Σ1 , . . . , ΣN ) where the coupling matrix M is given by M = −L.
Our goal is to aggregate each ξi taking values in Rni into a scalar-valued ξˆi , governed by Σ̂i = (0, 1, C1i 1ni , 1, 1)
which satisfies:
˙
ξˆi = ω̂i + υ̂i ,
Σ̂i : ζ̂1i = C1i 1ni ξˆi ,
ζ̂2i = ξˆi .
ci = In , Ki =
One can readily verify that, for any i ∈ [1; N ], conditions (5.4) and (5.5) are satisfied with M
i
11
22
12
−λIni , for some λ > 0, κ
bi = 2λ, Zi = Ini , L1i = 0, Wi = Ini , X = 0, X = 0, and X = X 21 = Ini , where
0 denotes zero matrices of appropriate dimensions. Moreover, for any i ∈ [1; N ], Pi = 1ni satisfies conditions
(5.8) with Qi = L2i = 0ni , H = 1ni , and Ŵi = 1ni . Hence, function Vi (xi , x̂i ) = (xi − 1ni x̂i )T (xi − 1ni x̂i )
is a storage function from Σ̂i to Σi satisfying condition (3.1) with αi (r) = λmax (C1 T C1i ) r2 and condition (3.2)
1i
with η(r) = −2λr, ρext (r) = 0, ∀r ∈ R≥0 , Wi = Ini , Ŵi = Hi = 1ni , and
0 I ni
Xi =
,
Ini
0
(6.2)
where the input ui ∈ Rni is given via the interface function in (5.10) as ui = −λ(xi − 1ni x̂i ) + 1ni ûi . Note
ei = 1n was computed as in (5.15).
that R
i
COMPOSITIONAL ABSTRACTION FOR NETWORKS OF CONTROL SYSTEMS: A DISSIPATIVITY APPROACH
17
Now, we look at Σ̂ = I(Σ̂1 , . . . , Σ̂N ) with a coupling matrix M̂ satisfying condition (4.2) as follows:
− Ldiag(1n1 , . . . , 1nN ) = diag(1n1 , . . . , 1nN )M̂ .
(6.3)
Note that the existence of M̂ satisfying (6.3) for a graph Laplacian L means that the N subgraphs form an
equitable partition of the full graph [GR01]. Although this restricts the choice of a partition in general, for the
complete graph (6.1) any partition is equitable.
Choosing µ1 = · · · = µN = 1 and using Xi in (6.2), matrix X in (4.4) reduces to
0 In
X=
,
In 0
and condition (4.1) reduces to
T
−L
−L
X
= −L − LT 0
In
In
which always holds without any restrictions on the size of the graph. In order to show the above inequality,
we used L = LT 0 which is always true for Laplacian matrices of undirected graphs.
For the sake of simulation, we fix n = 9 and
1 0
C = 0 0
0 0
0
0
0
0
0
0
0
1
0
0
0
0
0
0
0
0
0
0
0
0 ,
1
where
C11 = 1
0
0 , C12 = 0
1
0 , C13 = 0
0
1 .
Let us now synthesize a controller for Σ via the abstraction Σ̂ to enforce the specification, defined by the LTL
formula [BK08]
5
^
(¬Oi ) ∧ 3T1 ∧ 3T2 ,
(6.4)
ψ = S ∧
i=1
which requires that any output trajectory ζ of the closed loop system evolves inside the set S, avoids sets
Oi , i ∈ [1; 5], indicated with blue boxes in Figure 6, and visits each Ti , i ∈ [1; 2], indicated with red boxed
in Figure 6, infinitely often. We use SCOTS [RZ16b] to synthesize a controller for Σ̂ to enforce (6.4). In the
synthesis process we restricted the abstract inputs to û1 , û2 , û3 ∈ [−14, 14]. Given that we can set the initial
states of Σ to xi = Pi x̂i , so that Vi (xi , x̂i ) = 0, and since ρext (r) = 0, ∀r ∈ R≥0 , we obtain kζ(t) − ζ̂(t)k = 0
for all t ≥ 0. A closed-loop output trajectory of Σ is illustrated in Figure 6. Note that it would not have
been possible to synthesize a controller using SCOTS for the original 9-dimensional system Σ, without the
3-dimensional intermediate approximation Σ̂.
Remark 6.1. This scale-free result highlights the advantage of dissipativity-type over small-gain type conditions proposed in [RZ15, RZ16a]: the storage function Vi from Σ̂i to Σi in this example also satisfies the
requirements of a simulation function defined in [RZ15, RZ16a]; however, the resulting small-gain type condin−1
< 1 which involves the spectral radius1 of L (ρ(L) = n). Hence,
tion, e.g., for L in (6.1) reduces to n−1+λ
using the results in [RZ15, RZ16a], one can readily verify that as the number of components increases, e.g.
n → ∞, the quality of approximation deteriorates unless the interface gain λ is increasing with n which is not
desirable because it results in high amplitude inputs ui .
1The spectral radius of a square matrix A ∈ Rn×n , denoted by ρ(A), is defined as ρ(A) := max{|λ |, · · · , |λ |} where λ , . . . , λ
n
n
1
1
are eigenvalues of A.
18
M. ZAMANI AND M. ARCAK
Figure 3. The specification with closed loop output trajectory of Σ. The sets S, Oi , i ∈ [1; 5],
and Ti , i ∈ [1; 2] are given by: S = [0, 10]3 , T1 = [1, 2]3 and T2 = [8, 9]3 , O1 = [4, 6]3 ,
O2 = [7, 9] × [1, 3] × [0, 10], O3 = [2, 3] × [7, 8] × [0, 10], O4 = [1, 2] × [1, 2] × [5, 10], and
O5 = [8, 9] × [8, 9] × [0, 5].
7. Conclusion
In this paper, we proposed for the first time a notion of so-called storage function relating a concrete control
system to its abstraction by quantifying their joint input-output correlation. This notion was adapted from
the one of storage function from dissipativity theory. Given a network of control subsystems together with
their corresponding abstractions and storage functions, we provide compositional conditions under which a
network of abstractions approximate the original network and the approximation error can be quantified compositionally using the storage functions of the subsystems. Finally, we provide a procedure for the construction
of abstractions together with their corresponding storage functions for a class of nonlinear control systems by
using the bounds on the slope of system nonlinearities. One of the main advantages of the proposed results
here based on a dissipativity-type condition in comparison with the existing ones based on a small-gain type
condition is that the former can enjoy specific interconnection matrix and provide scale-free compositional
conditions (cf. Section 6).
8. Acknowledgments
The authors would like to thank Mahmoud Khaled for the simulation in Section 6.
References
[AAFK01] O. M. Aamo, M. Arcak, T. I. Fossen, and P. V. Kokotovic. Global output tracking control of a class of Euler-Lagrange
systems with monotonic non-linearities in the velocities. International Journal of Control, 74(7):649–658, 2001.
[AGPV03] M. Arcak, H. Görgün, L. M. Pedersen, and S. Varigonda. An adaptive observer design for fuel cell hydrogen estimation.
In Proceedings of American Control Conference, pages 2037–2042, June 2003.
[AK01a]
M. Arcak and P. P. Kokotovic. Observer-based control of systems with slope-restricted nonlinearities. IEEE Transactions on Automatic Control, 46(7):1146–1150, July 2001.
[AK01b] M. Arcak and P. V. Kokotovic. Feasibility conditions for circle criterion designs. Systems & Control Letters, 42(5):405–
412, 2001.
[AMP16] M. Arcak, C. Meissen, and A. Packard. Networks of dissipative systems. SpringerBriefs in Electrical and Computer
Engineering. Springer International Publishing, 2016.
[BK08]
C. Baier and J. P. Katoen. Principles of model checking. The MIT Press, April 2008.
[DK04]
K. C. Das and P. Kumar. Some new bounds on the spectral radius of graphs. Discrete Mathematics, 281(1-3):149–161,
April 2004.
[FA03]
X. Fan and M. Arcak. Observer design for systems with multivariable monotone nonlinearities. Systems & Control
Letters, 50:319–330, 2003.
COMPOSITIONAL ABSTRACTION FOR NETWORKS OF CONTROL SYSTEMS: A DISSIPATIVITY APPROACH
19
[Fre05]
G. F. Frehse. Compositional verification of hybrid systems using simulation relations. PhD thesis, Radboud Universiteit Nijmegen, 2005.
[GP09]
A. Girard and G. J. Pappas. Hierarchical control system design using approximate simulation. Automatica, 45(2):566–
571, 2009.
[GR01]
C. Godsil and G. Royle. Algebraic Graph Theory. Graduate Texts in Mathematics. Springer New York, 2001.
[Isi99]
A. Isidori. Nonlinear Control Systems II. Communications and Control Engineering. Springer-Verlag London, 1 edition,
1999.
[Kv10]
F. Kerber and A. van der Schaft. Compositional analysis for linear control systems. In Proc. of the 13th ACM Int.
Conf. on Hybrid Systems: Computation and Control, pages 21–30, 2010.
[LSW96] Y. Lin, E. D. Sontag, and Y. Wang. A smooth converse Lyapunov theorem for robust stability. SIAM Journal on
Control and Optimization, 34:124–160, 1996.
[MLAP15] C. Meissen, L. Lessard, M. Arcak, and A. K. Packard. Compositional performance certification of interconnected
systems using ADMM. Automatica, 61:55–63, 2015.
[PLS00]
G. J. Pappas, G. Lafferriere, and S. Sastry. Hierarchically consistent control systems. IEEE Transactions on Automatic
Control, 45(6):1144–1160, June 2000.
[PPD16] G. Pola, P. Pepe, and M. D. Di Benedetto. Symbolic models for networks of control systems. IEEE Transactions on
Automatic Control, 2016.
[RZ15]
M. Rungger and M. Zamani. Compositional construction of approximate abstractions. In Proceedings of the 18th
International Conference on Hybrid Systems: Computation and Control, pages 68–77. ACM New York, NY, USA,
April 2015.
[RZ16a]
M. Rungger and M. Zamani. Compositional construction of approximate abstractions of interconnected control systems. IEEE Transactions on Control of Network Systems, 2016.
[RZ16b]
M. Rungger and M. Zamani. SCOTS: A tool for the synthesis of symbolic controllers. In Proceedings of the 19th
International Conference on Hybrid Systems: Computation and Control, pages 99–104. ACM New York, NY, USA,
April 2016.
[Sch04]
E. Scholtz. Observer-based monitors and distributed wave controllers for electromechanical disturbances in power
systems. PhD thesis, Massachusetts Institute of Technology, September 2004.
[Son98]
E. D. Sontag. Mathematical control theory, volume 6. Springer-Verlag, New York, 2nd edition, 1998.
[SS07]
G. B. Stan and R. Sepulchre. Analysis of interconnected oscillators by dissipativity theory. IEEE Transactions on
Automatic Control, 52(2):256–270, February 2007.
[TI08]
Y. Tazaki and J. Imura. Bisimilar finite abstractions of interconnected systems. in Proceedings of 11th International
Conference on Hybrid Systems: Computation and Control (HSCC), pages 514–527, 2008.
[TPL04]
P. Tabuada, G. J. Pappas, and P. Lima. Compositional abstractions of hybrid control systems. Discrete event dynamic
systems, 14(2):203–238, 2004.
[Wil72]
J. C. Willems. Dissipative dynamical systems part i: General theory. Archive for Rational Mechanics and Analysis,
45(5):321–351, 1972.
[Yak62]
V. A. Yakubovich. The solution of certain matrix inequalities in automatic control theory. Doklady Akademii Nauk
SSSR, 143(6):1304–1307, 1962.
[You12]
W. H. Young. On classes of summable functions and their fourier series. Proceedings of the Royal Society of London
A: Mathematical, Physical and Engineering Sciences, 87(594):225–229, 1912.
[ZRMng] M. Zamani, M. Rungger, and P. Mohajerin Esfahani. Approximations of stochastic hybrid systems: A compositional
approach. IEEE Transactions on Automatic Control, 2016 (forthcoming).
1 Department
of Electrical and Computer Engineering, Technical University of Munich, D-80290 Munich, Germany.
E-mail address: [email protected]
URL: http://www.hcs.ei.tum.de
2 Department
of Electrical Engineering and Computer Sciences, University of California, Berkeley, CA, USA.
E-mail address: [email protected]
URL: http://www.eecs.berkeley.edu/∼arcak
| 3cs.SY
|
TRUNCATED RANDOM MEASURES
arXiv:1603.00861v2 [math.ST] 1 Feb 2017
TREVOR CAMPBELL? , JONATHAN H. HUGGINS? , JONATHAN HOW,
AND TAMARA BRODERICK
Abstract. Completely random measures (CRMs) and their normalizations
are a rich source of Bayesian nonparametric priors. Examples include the beta,
gamma, and Dirichlet processes. In this paper we detail two major classes
of sequential CRM representations—series representations and superposition
representations—within which we organize both novel and existing sequential
representations that can be used for simulation and posterior inference. These
two classes and their constituent representations subsume existing ones that
have previously been developed in an ad hoc manner for specific processes.
Since a complete infinite-dimensional CRM cannot be used explicitly for computation, sequential representations are often truncated for tractability. We
provide truncation error analyses for each type of sequential representation,
as well as their normalized versions, thereby generalizing and improving upon
existing truncation error bounds in the literature. We analyze the computational complexity of the sequential representations, which in conjunction with
our error bounds allows us to directly compare representations and discuss
their relative efficiency. We include numerous applications of our theoretical
results to commonly-used (normalized) CRMs, demonstrating that our results
enable a straightforward representation and analysis of CRMs that has not
previously been available in a Bayesian nonparametric context.
1. Introduction
2. Background
2.1. CRMs and truncation
2.2. The gamma-Poisson process
3. Sequential representations
3.1. Series representations
3.2. Superposition representations
4. Truncation analysis
4.1. Series representations
4.2. Superposition representations
4.3. Stochastic mapping
4.4. Hyperpriors
5. Normalized truncation analysis
5.1. Series representations
5.2. Superposition representations
5.3. Hyperpriors
6. Simulation and computational complexity
7. Summary of results
8. Discussion
Acknowledgments
Date: February 2, 2017.
? First authorship is shared jointly by T. Campbell and J. H. Huggins.
1
2
5
5
6
6
7
9
12
12
16
18
19
19
21
21
23
24
24
27
28
2
T. CAMPBELL, J. H. HUGGINS, J. P. HOW, AND T. BRODERICK
Appendix A. Additional Examples
A.1. Beta process
A.2. Beta prime process
Appendix B. Technical lemmas
Appendix C. Proofs of sequential representation results
C.1. Correctness of B-Rep, DB-Rep, and PL-Rep
C.2. Proof of the expected number rejections of the R-Rep
C.3. Power-law behavior of the PL-Rep
Appendix D. Proofs of CRM truncation bounds
D.1. Protobound
D.2. Series representation truncation
D.3. Superposition representation truncation
D.4. Stochastic mapping truncation
D.5. Truncation with hyperpriors
Appendix E. Proofs of normalized truncation bounds
E.1. Normalized series representation truncation
E.2. Normalized superposition representation truncation
E.3. Truncation with hyperpriors
References
28
28
32
35
38
38
41
41
44
44
46
50
51
51
51
52
54
55
55
1. Introduction
In many data sets, we can view the data points as exhibiting a collection of
underlying traits. For instance, each document in the New York Times might
touch on a number of topics or themes, an individual’s genetic data might be a
product of the populations to which their ancestors belonged, or a user’s activity
on a social network might be dictated by their varied personal interests. When
the traits are not directly observed, a common approach is to model each trait as
having some frequency or rate in the broader population (Airoldi et al., 2014). The
inferential goal is to learn these rates as well as whether—and to what extent—
each data point exhibits each trait. Since the traits are unknown a priori, their
cardinality is also typically unknown.
As a data set grows larger, we can reasonably expect the number of traits to increase as well. In the cases above, for example, we expect to uncover more topics as
we read more documents, more ancestral populations as we examine more individuals’ genetic data, and more unique interests as we observe more individuals on a
social network. Bayesian nonparametric (BNP) priors provide a flexible, principled
approach to creating models in which the number of exhibited traits is random, can
grow without bound, and may be learned as part of the inferential procedure. By
generating a countable infinity of potential traits—where any individual data point
exhibits only finitely many—these models enable growth in the number of observed
traits with the size of the data set.
In practice, however, it is impossible to store a countable infinity of random
variables in memory or learn the distribution over a countable infinity of variables
in finite time. Conjugate priors and likelihoods have been developed (Orbanz, 2010)
that theoretically circumvent the infinite representation altogether and perform
exact Bayesian posterior inference (Broderick et al., 2017). However, these priors
and likelihoods are often just a single piece within a more complex generative model,
and ultimately an approximate posterior inference scheme such as Markov Chain
TRUNCATED RANDOM MEASURES
3
Monte Carlo (MCMC) or variational Bayes (VB) is required. These approximation
schemes often necessitate a full and explicit representation of the latent variables.
One option is to approximate the infinite-dimensional prior with a related finitedimensional prior: that is, to replace the infinite collection of random traits by
a finite subset of “likely” traits. To do so, first enumerate the countable infinity
of traits in the full model and write (ψk , θk ) for each paired trait ψk (e.g. a topic
in
a document) and its rate or frequency θk . Then the discrete measure Θ :=
P∞
k=1 θk δψk captures the traits/rates in a sequence indexed by k. The (ψk , θk )
pairs are random in the Bayesian model, so Θ is a random measure. In many
cases, the distribution of Θ can be defined by specifying a sequence of simple,
familiar distributions for the finite-dimensional ψk and θk , known as a sequential
representation. Given a sequential representation of Θ, a natural way to choose a
subset of traits is to keep the first K < ∞ traits and discard the rest, resulting in
an approximate measure ΘK . This approach is called truncation.1
Sequential representations have been shown to exist for completely random measures (CRMs) (Kingman, 1967; Ferguson and Klass, 1972), a large class of nonparametric priors that includes such popular models as the beta process (Hjort,
1990; Kim, 1999) and the gamma process (Ferguson and Klass, 1972; Kingman,
1975; Brix, 1999; Titsias, 2008; James, 2013). Numerous sequential representations
of CRMs have been developed in the literature (Ferguson and Klass, 1972; Bondesson, 1982; Rosiński, 1990, 2001; James, 2014; Broderick et al., 2017). CRM priors
are often paired with likelihood processes—such as the Bernoulli process (Thibaux
and Jordan, 2007), negative binomial process (Zhou et al., 2012; Broderick et al.,
2015), and Poisson likelihood process (Titsias, 2008). The likelihood process determines how much each trait is expressed by each data point. Sequential representations also exist for normalized completely random measures (NCRMs) (Perman
et al., 1992; Perman, 1993; James, 2002; Pitman, 2003; Regazzini et al., 2003; Lijoi
and Prünster, 2010),2 which provide random distributions over traits, such as the
Dirichlet process (Ferguson, 1973; Sethuraman, 1994). NCRMs are typically paired
with a likelihood that assigns each data point to a single trait using the NCRM as
a discrete distribution.
Since (N)CRMs have many possible sequential representations, a method is required for determining which to use for the application at hand and, once a representation is selected, for choosing a truncation level. Our main contributions
enable the principled selection of both representation and truncation level using
approximation error:
(1) We provide a comprehensive characterization of the different types of sequential representations for (N)CRMs, filling in many gaps in the literature
of sequential representations along the way. We classify these representations into two major groups: series representations, which are constructed
by transforming a homogeneous Poisson point process; and superposition
representations, which are the superposition of infinitely many Poisson
1It is also possible to truncate a sequential representation to a random level K by removing
atoms with weights less than a specified threshold (Argiento et al., 2016; Muliere and Tardella,
1998), though this approach is not straightforward for use in posterior inference algorithms.
2
NCRMs are sometimes referred to as normalized random measures with independent increments (NRMIs) (Lijoi and Prünster, 2010; Regazzini et al., 2003; James et al., 2009).
4
T. CAMPBELL, J. H. HUGGINS, J. P. HOW, AND T. BRODERICK
point processes with finite rate measures. We also introduce two novel
sequential representations for (N)CRMs.
(2) We provide theoretical guarantees on the approximation error induced when
truncating these sequential representations. We give the error as a function of the prior process, the likelihood process, and the level of truncation. While truncation error bounds for (N)CRMs have been studied
previously, past work has focused on specific combinations of (N)CRM
priors and likelihoods—in particular, the Dirichlet-multinomial (Sethuraman, 1994; Ishwaran and James, 2001; Ishwaran and Zarepour, 2002; Blei
and Jordan, 2006), beta-Bernoulli (Paisley et al., 2012; Doshi-Velez et al.,
2009), generalized beta-Bernoulli (Roy, 2014), and gamma-Poisson (Roychowdhury and Kulis, 2015) processes. In the current work, we give much
more general results for bounding the truncation error.
Our results fill in large gaps in the analysis of truncation error, which is often
measured in terms of the L1 (a.k.a. total variation) distance between the data distributions induced by the full and truncated priors. We provide the first analysis
of truncation error for some sequential representations of the beta process with
Bernoulli likelihood (Thibaux and Jordan, 2007), for the beta process with negative binomial likelihood (Zhou et al., 2012; Broderick et al., 2015), and for the
normalization of the generalized gamma process (Brix, 1999), the σ-stable process,
and the generalized inverse gamma (Lijoi et al., 2005; Lijoi and Prünster, 2010)
with discrete likelihood. Moreover, even when truncation results already exist in
the literature (Ishwaran and James, 2001; Doshi-Velez et al., 2009; Paisley et al.,
2012; Roychowdhury and Kulis, 2015), we improve on those error bounds by a factor of two. The reduction arises from our use of the point process machinery of
CRMs, circumventing the total variation bound used originally by Ishwaran and
James (2001, 2002) upon which most modern truncation analyses are built. We obtain our truncation error guarantees by bounding the probability that data drawn
from the full model will use a feature that is not available to the truncated model.
Thinking in terms of this probability provides a more intuitive interpretation of our
bounds that can be communicated to practitioners and used to guide them in their
choice of truncation level.
The remainder of this paper is organized as follows. In Section 2, we provide
background material on CRMs and establish notation. In our first main theoretical
section, Section 3, we describe seven different sequential CRM representations, including four series representations and three superposition representations, two of
which are novel. Next, we provide a general theoretical analysis of the truncation
error for series and superposition representations in Section 4. We provide analogous theory for the normalized versions of each representation in Section 5 via
an infinite extension of the “Gumbel-max trick” (Gumbel, 1954; Maddison et al.,
2014). We determine the complexity of simulating each representation in Section 6.
In Section 7, we summarize our results (Table 1) and provide advice on how to
select sequential representations in practice. Proofs for all results developed in this
paper are provided in the appendices.
TRUNCATED RANDOM MEASURES
5
2. Background
2.1. CRMs and truncation. Consider a Poisson point process on R+ := [0, ∞)
with rate measure ν(dθ) such that
Z
ν(R+ ) = ∞
and
min(1, θ)ν(dθ) < ∞.
(2.1)
∞
Such a process generates aPcountable infinity of values (θk )k=1 , θk ∈ R+ , having an
∞
almost surely finite sum k=1 θk < ∞. In a BNP trait model, we interpret each
θk as the rate or frequency of the k-th trait. Typically, each θk is paired with a
parameter ψk associated with the k-th trait (e.g., a topic in a document or a shared
interest on a social network). We assume throughout that ψk ∈ Ψ for some space
i.i.d.
Ψ and ψk ∼ G for some distribution G. Constructing a measure by placing mass
θk at atom location ψk results in a completely random measure (CRM) (Kingman,
1967). As shorthand, we will write CRM(ν) for the completely random measure
generated as just described:
X
Θ :=
θk δψk ∼ CRM(ν).
k
The trait distribution G is left implicit in the notation as it has no effect on our
results. Further, the possible fixed-location and deterministic components of a
CRM (Kingman, 1967) are not considered here for brevity; these components can
be added (assuming they are purely atomic) and the analysis modified without
undue effort. The CRM prior on Θ is typically combined with a likelihood that
generates trait counts for each data point. Let h(· | θ) be a proper probability mass
function on N ∪ {0} for all θ in the support of ν.3 Then a collection of conditionally
N
independent observations X1:N := {Xn }n=1 given Θ are distributed according to
the likelihood process LP(h, Θ), i.e.
X
i.i.d.
Xn :=
xnk δψk ∼ LP(h, Θ),
k
if xnk ∼ h(x | θk ) independently across k and i.i.d. across n. The desideratum that
each Xn expresses a finite number of traits is encoded by the assumption that
Z
(1 − h(0 | θ))ν(dθ) < ∞.
(2.2)
Since the trait counts are typically latent in a full generative model specification,
indep
define the observed data Yn | Xn ∼ f (· | Xn ) for a conditional density f with
∞
respect to a measure µ on some space. For instance, if the sequence (θk )k=1 represents the topic rates in a document corpus, Xn might capture how many words in
document n are generated from each topic and Yn might be the observed collection
of words for that document.
Since the sequence (θk )∞
k=1 is countably infinite, it may be difficult to simulate or
perform posterior inference in this model. One approximation scheme is to define
PK
the truncation ΘK := k=1 θk δψk . Since it is finite, the truncation ΘK can be
used for exact simulation or in posterior inference—but some error arises from not
using the full CRM Θ. To quantify this error, consider its propagation through the
3Likelihoods with support in R may also be considered. As long as Eq. (2.2) holds, all results
except those for the size-biased representation still hold, since they only rely upon the behavior
of h(0 | θ).
6
T. CAMPBELL, J. H. HUGGINS, J. P. HOW, AND T. BRODERICK
above Bayesian model. Define Z1:N and W1:N for ΘK analogous to the definitions
of X1:N and Y1:N for Θ:
i.i.d.
Zn | ΘK ∼ LP(h, ΘK ),
indep
Wn | Zn ∼ f (· | Zn ),
n = 1, . . . , N.
A standard approach to measuring the distance between Θ and ΘK is to use the
L1 metric between the marginal densities pN,∞ and pN,K (with respect to some
measure µ) of the final observations Y1:N and W1:N (Ishwaran and James, 2001;
Doshi-Velez et al., 2009; Paisley et al., 2012):
Z
1
1
kpN,∞ − pN,K k1 :=
|pN,∞ (y1:N ) − pN,K (y1:N )| µ(dy1:N ).
2
2 y1:N
All of our bounds on 21 kpN,∞ −pN,K k1 are also bounds on the probability that X1:N
contains a feature that is not in the truncation ΘK (cf. Sections 4 and 5). This
interpretation may be easier to digest since it does not depend on the observation
model f and is instead framed in terms of the underlying traits the practitioner is
trying to estimate.
2.2. The gamma-Poisson process. To illustrate the practical application of the
theoretical developments in this work, we provide a number of examples throughout involving the gamma process (Brix, 1999), denoted ΓP(γ, λ, d), with discount
parameter d ∈ [0, 1), scale parameter λ > 0, mass parameter γ > 0, and rate
measure
ν(dθ) = γ
λ1−d −d−1 −λθ
θ
e
dθ.
Γ(1 − d)
Setting d = 0 yields the undiscounted gamma process (Ferguson and Klass, 1972;
Kingman, 1975; Titsias, 2008). The gamma process is often paired with a Poisson
likelihood,
θx −θ
e .
x!
Throughout the present work, we use the rate parametrization of the gamma distribution (to match the gamma process parametrization), for which the density is
given by
h(x | θ) =
Gam(x; a, b) =
ba a−1 −bx
x
e .
Γ(a)
Appendix A provides additional example applications of our theoretical results for
two other CRMs: the beta process BP(γ, α, d) (Teh and Görür, 2009; Broderick
et al., 2012) with Bernoulli or negative binomial likelihood, and the beta prime
process BPP(γ, α, d) (Broderick et al., 2017) with odds-Bernoulli likelihood.
3. Sequential representations
Sequential representations are at the heart of the study of truncated CRMs.
They provide an iterative method that can be terminated at any point to yield a
finite approximation to the infinite process, where the choice of termination point
determines the accuracy of the approximation. Thus, the natural first step in providing a coherent treatment of truncation analysis is to do the same for sequential
representations. In past work, two major classes of sequential representation have
TRUNCATED RANDOM MEASURES
7
P∞
been used: series representations of the form k=1 θk δψk , and superposition repP∞ PCk
resentations of the form k=1 i=1
θki δψki , where each inner sum of Ck atoms
is itself a CRM. This section examines four series representations (Ferguson and
Klass, 1972; Bondesson, 1982; Rosiński, 1990, 2001) and three superposition representations (two of which are novel) (Broderick et al., 2012, 2017; James, 2014).
We show how previously-developed sequential representations for specific CRMs fit
into these seven general representations. Finally, we discuss a stochastic mapping
procedure that is useful in obtaining new representations from the transformation
of others. Proofs for the results in this section may be found in Appendix C.
3.1. Series representations. Series representations arise from the transformation
of a homogeneous Poisson point process (Rosiński, 2001). They tend to be somewhat difficult to analyze due to the dependence between the atoms but also tend
to produce very simple representations with small truncation error (cf. Sections 4
Pk
i.i.d.
and 7). Throughout the paper we let Γk = `=1 E` , E` ∼ Exp(1), be the ordered
jumps of a unit-rate homogeneous Poisson process on R+ , let ν be a measure on
i.i.d.
R+ satisfying the basic conditions in Eq. (2.1), and let ψk ∼ G.
:=
Inverse-Lévy (Ferguson and Klass, 1972). Define ν ← (u)
inf {x : ν ([x, ∞)) ≤ u}, the inverse of the tail measuer ν([x, ∞)). We say Θ
has an inverse-Lévy representation and write Θ ← IL-Rep(ν) if
∞
X
Θ=
θk δψk ,
with
θk = ν ← (Γk ).
k=1
Ferguson and Klass (1972) showed that Θ ← IL-Rep(ν) implies Θ ∼ CRM(ν). The
inverse-Lévy representation is analogous to the inverse CDF method for generating
an arbitrary random variable from a uniform random variable, with the homogenous
Poisson process playing the role of the uniform random variable. It is also the
optimal sequential representation in the sense that the sequence θk that it generates
is non-increasing. While an elegant and general approach, simulating the inverseLévy representation is difficult, as inverting the function ν ([x, ∞)) is analytically
intractable except in a few cases.
Example 3.1 (Gamma
process, ΓP(γ, λ, 0)). We have ν([x, ∞)) = γλE1 (λx),
R∞
where E1 (x) := x u−1 e−u du is the exponential integral function (Abramowitz
and Stegun, 1964). The inverse-Lévy representation for ΓP(γ, λ, 0) is thus
∞
X
Θ=
λ−1 E1−1 (γ −1 λ−1 Γk )δψk .
k=1
Neither E1 nor its inverse can be computed in closed form, so one must resort to
numerical approximations.
Bondesson (Bondesson, 1982). We say Θ has a Bondesson representation and
write Θ ← B-Rep(c, g) if for c > 0 and g a density on R+ ,
∞
X
i.i.d.
Θ=
θk δψk ,
with
θk = Vk e−Γk /c ,
Vk ∼ g.
k=1
Theorem 3.1 shows that Bondesson representations can be constructed for a large,
albeit restricted, class of CRM rate measures. We offer a novel proof of Theorem 3.1
in Appendix C using the induction strategy introduced by Banjevic et al. (2002).
8
T. CAMPBELL, J. H. HUGGINS, J. P. HOW, AND T. BRODERICK
Similar proof ideas are also used to prove truncation error bounds for sequential
representations in Section 4. We use a slight abuse of notation for brevity: if ν(dθ)
is a measure on R+ that is absolutely continuous with respect to Lebesgue measure,
then ν(θ) is the density of ν(dθ) with respect to the Lebesgue measure.
Theorem 3.1 (Bondesson representation (Bondesson, 1982)). Let ν(dθ) = ν(θ)dθ
be a rate measure satisfying Eq. (2.1). If θν(θ) is nonincreasing, limθ→∞ θν(θ) = 0,
d
and cν := limθ→0 θν(θ) < ∞, then gν (v) := −c−1
ν dv [vν(v)] is a density on R+ and
Θ ← B-Rep(cν , gν )
implies
Θ ∼ CRM(ν).
Example 3.2 (Bondesson representation for ΓP(γ, λ, 0)). The following representation for the gamma process with d = 0 was described by Bondesson (1982)
and Banjevic et al. (2002). Since θν(θ) = γλe−λθ is non-increasing and cν =
limθ→0 θν(θ) = γλ, we obtain gν (v) = λe−λv = Exp(v; λ). Thus, it follows from
Theorem 3.1 that if Θ ← B-Rep(γλ, Exp(λ)), then Θ ∼ ΓP(γ, λ, 0). The condition
that θν(θ) is non-increasing fails to hold if d > 0, so we cannot apply Theorem 3.1
to ΓP(γ, λ, d) when d > 0.
Thinning (Rosiński, 1990). Using the nomenclature of Rosiński (2001), we say
Θ has a thinning representation and write Θ ← T-Rep(ν, g) if g is a probability
measure on R+ such that ν is absolutely continuous with respect to g, i.e. ν g,
and
∞
X
dν
i.i.d.
(Vk ) ≥ Γk ,
Vk ∼ g.
θk δψk ,
Θ=
with
θk = Vk 1
dg
k=1
Rosiński (1990) showed that Θ ← T-Rep(ν, g) implies Θ ∼ CRM(ν). Note that
a.s.
Γk → ∞ as k → ∞, so the probability that dν
dg (Vk ) ≥ Γk is decreasing in k. Thus,
this representation generates atoms with θk = 0 (which have no effect and can be
removed) increasingly frequently and becomes inefficient as k → ∞.
Example 3.3 (Thinning representation for ΓP(γ, λ, d)). If we let g = Gam(1−d, λ),
then the thinning representation for ΓP(γ, λ, d) is
Θ=
∞
X
Vk 1(Vk Γk ≤ γ)δψk ,
with
i.i.d.
Vk ∼ Gam(1 − d, λ).
k=1
Rejection (Rosiński, 2001). Using the nomenclature of Rosiński (2001), we
say Θ has a rejection representation and write Θ ← R-Rep(ν, µ) if µ is a measure
dν
≤ 1, and
on R+ satisfying Eq. (2.1) and dµ
∞
X
dν
Θ=
θk δψk , with θk = Vk 1
(Vk ) ≥ Uk , (Vk )k∈N ∼ PoissP(µ),
dµ
k=1
i.i.d.
Uk ∼ Unif[0, 1].
Rosiński (2001) showed that Θ ← R-Rep(ν, µ) implies Θ ∼ CRM(ν). This representation is very similar to the thinning representation, except that the sequence
(Vk )k∈N is generated from a Poisson process on R+ rather than i.i.d. This allows
a.s.
Vk → 0 as k → ∞, causing the frequency of generating ineffective atoms θk = 0
dν
to decay as k → ∞, assuming µ is appropriately chosen such that dµ
(θ) → 1 as
θ → 0. This representation can thus be constructed to be more efficient than the
TRUNCATED RANDOM MEASURES
9
thinning representation. We can calculate the efficiency in terms of the expected
number of rejections (that is, the number of θk that are identically zero):
Proposition 3.2. For R-Rep(ν, µ), the expected number of rejections is
"∞
# Z
X
dν
(x) µ(dx).
E
1(θk = 0) =
1−
dµ
k=1
Remark. If µ and ν can be written as densities with respect
to Lebesgue measure,
R
then the integral in Proposition 3.2 can be rewritten as (µ(x) − ν(x))dx.
Example 3.4 (Rejection representation for ΓP(γ, λ, 0)). Following Rosiński (2001),
consider µ(dθ) = γλθ−1 (1 + λθ)−1 dθ. We call CRM(µ) the Lomax process,
LomP(γ, λ−1 ), after the related Lomax distribution. We can use the inverse-Lévy
method analytically with µ since µ← (u) = λ(e(γλ)1−1 u −1) . Thus, the rejection representation of ΓP(γ, λ, 0) is
∞
X
Vk 1(Uk ≤ (1 + λVk )e−λVk )δψk , with
Θ=
k=1
Vk =
1
λ(e(γλ)−1 Γk
− 1)
,
i.i.d.
Uk ∼ Unif[0, 1].
Unlike in the thinning construction given in Example 3.3, only a finite number
of rates will be set to zero almost surely. In particular, the expected number of
rejections is γλcγ , where cγ is the Euler-Mascheroni constant.
Example 3.5 (Rejection representation for ΓP(γ, λ, d), d > 0). For the case of
λ1−d
θ−1−d dθ. We can again use the inverse-Lévy
d > 0, we instead use µ(dθ) = γ Γ(1−d)
1−d
λ
. The
method analytically with µ since µ← (u) = (γ 0 u−1 )1/d , where γ 0 := γ dΓ(1−d)
rejection representation is then
∞
X
i.i.d.
1/d
Vk 1(Uk ≤ e−λVk )δψk , with Vk = (γ 0 Γ−1
Θ=
, Uk ∼ Unif[0, 1].
k )
k=1
1−d
The expected number of rejections is γ λ d , so the representation is efficient for
large d, but extremely inefficient when d is small.
3.2. Superposition representations. Superposition representations arise as an
infinite sum of CRMs with finite rate measure. These tend to be easier to analyze
than series representations as they decouple atoms between the summed CRMs,
but can produce representations with larger truncation error (cf. Sections 4 and 7).
Throughout, let ν be a measure on R+ satisfying the basic conditions in Eq. (2.1),
i.i.d.
and let ψk ∼ G.
Decoupled Bondesson. We say Θ has a decoupled Bondesson representation
and write Θ ← DB-Rep(c, g, ξ) if for c > 0, ξ > 0, and g a density on R+ ,
Θ=
Ck
∞ X
X
θki δψki ,
with
i.i.d.
θki = Vki e−Tki ,(3.1)
indep
Vki ∼ g.
Ck ∼ Poiss(c/ξ),
k=1 i=1
Tki ∼ Gam(k, ξ),
i.i.d.
This is a novel superposition representation, though special cases are already
known (Paisley et al., 2010; Roychowdhury and Kulis, 2015). Theorem 3.3 shows
that the decoupled Bondesson representation applies to the same class of CRMs as
the Bondesson representation from Section 3.1.
10
T. CAMPBELL, J. H. HUGGINS, J. P. HOW, AND T. BRODERICK
Theorem 3.3 (Decoupled Bondesson representation). Let ν(dθ) = ν(θ)dθ, cν , and
gν be as specified in Theorem 3.1. Then for any fixed ξ > 0,
Θ ← DB-Rep(cν , gν , ξ)
implies
Θ ∼ CRM(ν).
The proof of Theorem 3.3 in Appendix C generalizes the arguments from Paisley
et al. (2010) and Roychowdhury and Kulis (2015). The free parameter ξ controls
the number of atoms generated for each outer sum index k; its principled selection
can be made by trading off computational complexity (cf. Section 6) and truncation
error (cf. Section 4).
Example 3.6 (Decoupled Bondesson representation for ΓP(γ, λ, 0)). Arguments
paralleling those made in Example 3.2 show that the ΓP(γ, λ, 0) representation
from Roychowdhury and Kulis (2015) follows directly from an application of Theorem 3.3: if Θ ← DB-Rep(γλ, Exp(λ), ξ), then Θ ∼ ΓP(γ, λ, 0). As in the Bondesson
representation setting, Theorem 3.3 does not apply to ΓP(γ, λ, d) when d > 0 because the condition that θν(θ) is non-increasing fails to hold.
Size-biased (Broderick et al., 2017; James, 2014). Let π(θ) := h(0 | θ). We
say Θ has a size-biased representation and write Θ ← SB-Rep(ν, h) if
Θ=
Ck
∞ X
X
θki δψki ,
with
indep
Ck ∼ Poiss (ηk ) ,
k=1 i=1
indep
θki ∼
Z
ηk :=
1
π(θ)k−1 (1 − π(θ)) ν(dθ),
ηk
(3.2)
π(θ)k−1 (1 − π(θ)) ν(dθ).
Broderick et al. (2017) and James (2014) showed that Θ ← SB-Rep(ν, h) implies
Θ ∼ CRM(ν). If the rate measure ν and the P
likelihood h are selected to be a
∞
conjugate exponential family then, noting that x=1 h(x | θ) = 1 − π(θ), the rate
θki can be sampled from a mixture of exponential family distributions:
indep
θki | zki ∼
Z
ηkx :=
1
h(zki | θ)π(θ)k−1 ν(dθ),
ηkzki
indep
∞
zki ∼ Categorical ((ηkx /ηk )x=1 ) ,
h(x | θ)π(θ)k−1 ν(dθ).
Example 3.7 (Size-biased representation for ΓP(γ, λ, d)). For the Gamma process,
values for ηkx and ηk can be found using integration by parts and the standard
gamma distribution integral, while θki | zki is sampled from a gamma distribution
by inspection:
γλ1−d
γλ1−d Γ(x − d)
(λ + k)d − (λ + k − 1)d
d>0 ,
d
,
η
=
ηkx =
k
x!Γ(1 − d)(λ + k)x−d
γλ (log(λ + k) − log(λ + k − 1)) d = 0
indep
θki | zki ∼ Gam(x − d, λ + k).
Power-law. We say Θ has a power-law representation and write Θ ←
PL-Rep(γ, α, d, g) if for γ > 0, 0 ≤ d < 1, α > −d, and g a density on R+ ,
TRUNCATED RANDOM MEASURES
Θ=
Ck
∞ X
X
θki δψki , with
i.i.d.
Ck ∼ Poiss(γ),
θki = Vki Ukik
k=1 i=1
11
k−1
Y
(1 − Ukij ), (3.3)
j=1
i.i.d.
Vki ∼ g,
indep
Ukij ∼ Beta(1 − d, α + jd).
This is a novel superposition representation, although it was previously developed
in the special case of the beta process (where g(v) = δ1 ) (Broderick et al., 2012).
The name of this representation arises from the fact that it exhibits Types I and
II power-law behavior (Broderick et al., 2012) under mild conditions, as we show
in Theorem C.1 in the appendix. Theorem 3.5 below shows the conditions under
which Θ ← PL-Rep(γ, α, d, g) implies Θ ∼ CRM(ν). Its proof in Appendix C
relies on the notion of stochastic mapping (Lemma 3.4), a powerful technique for
transforming one CRM into another. Note that in Lemma 3.4, the case where u is
a deterministic function of θ via the mapping u = τ (θ) may be recovered by setting
κ(θ, du) = δτ (θ) .
P∞
Lemma 3.4 (CRM stochastic mapping). Let Θ = k=1 θk δψk ∼ CRM(ν). Then
for any probability kernel κ(θ, du), we have κ(Θ) ∼ CRM(νκ ), where
Z
∞
X
κ(Θ) :=
uk δψk , uk | θk ∼ κ(θk , ·), and νκ (du) := κ(θ, du)ν(dθ).
k=1
Theorem 3.5 (Power-law representation). Let ν(dθ) = ν(θ)dθ be a rate measure
satisfying Eq. (2.1), and let gν be a density on R+ such that
Z
ν(u) = θ−1 gν uθ−1 νBP (dθ) ,
where
νBP (dθ) = γ
Γ(α + 1)
1 [θ ≤ 1] θ−1−d (1 − θ)α+d−1 dθ
Γ(1 − d)Γ(α + d)
is the rate measure for the beta process BP(γ, α, d) from Eq. (A.1). Then
Θ ← PL-Rep(γ, α, d, gν )
implies
Θ ∼ CRM(ν).
Example 3.8 (Power-law representation for ΓP(γ, λ, d)). If we choose gν =
Gam(λ, λ), then using the change of variable w = u(θ−1 − 1),
Z
θ−1 gν uθ−1 νBP (dθ)
Z
−1
λλ
= γλ
uλ−1 θ−λ−d−1 e−λuθ (1 − θ)λ+d−1 dθ du
Γ(1 − d)Γ(λ + d)
Z
λλ
−1−d −λu
= γλ
u
e
wλ+d−1 e−λw dw du
Γ(1 − d)Γ(λ + d)
λ1−d
=γ
u−1−d e−λu du.
Γ(1 − d)
It follows immediately from Theorem 3.5 that if Θ ← PL-Rep(γ, λ, d, Gam(λ, λ)),
then Θ ∼ ΓP(γ, λ, d). To the best knowledge of the authors, this power-law representation for the gamma process is novel.
12
T. CAMPBELL, J. H. HUGGINS, J. P. HOW, AND T. BRODERICK
4. Truncation analysis
Each of the sequential representations developed in Section 3 shares a common
structural element—an outer infinite sum—which is responsible for generating a
countably infinite number of atoms in the CRM. In this section, we terminate
these outer sums at a finite truncation level K ∈ N, resulting in a truncated CRM
ΘK possessing a finite number of atoms. We develop upper bounds on the error
induced by this truncation procedure. All of the truncated CRM error bounds in
this section rely on Lemma 4.1, which is a tightening (by a factor of two) of the
bound in Ishwaran and James (2001, 2002).4 Proposition 4.2 shows that the bound
in Lemma 4.1 is tight without further assumptions on the data likelihood f .
Lemma 4.1 (CRM protobound). Let Θ ∼ CRM(ν). For any truncation ΘK , if
i.i.d.
Zn | ΘK ∼ LP(h, ΘK ),
i.i.d.
indep
Wn | Zn ∼ f (· | Zn ),
Xn | Θ ∼ LP(h, Θ),
indep
Yn | Xn ∼ f (· | Xn ),
then, with pN,∞ and pN,K denoting the marginal densities of Y1:N and W1:N , respectively,
1
kpN,∞ − pN,K k1 ≤ 1 − P (supp(X1:N ) ⊆ supp(ΘK )) ,
2
Proposition 4.2 (Protobound tightness). If G is non-atomic and Ψ is Borel, then
for any δ > 0, there exists a likelihood f such that Lemma 4.1 is tight up to a factor
of 1 − δ.
The proof of all results in this section (including Lemma 4.1 and Proposition 4.2)
can be found in Appendix D. All of the provided truncation results use the generative model in Lemma 4.1, and are summarized in Table 1 in Section 7. Throughout
this section, for a given likelihood model h(x | θ) we define π(θ) := h(0 | θ) for notational brevity. The asymptotic behavior of truncation error bounds is specified
with tilde notation:
a(K)
a(K) ∼ b(K), K → ∞
⇐⇒
lim
= 1.
K→∞ b(K)
4.1. Series representations. Each of the series representations can be viewed
a functional of a standard Poisson point process and a sequence of i.i.d. random
variables with some distribution g on R+ .5 In particular, we may write each in the
form
∞
X
i.i.d.
Θ=
θk δψk ,
with
θk = τ (Vk , Γk ),
Vk ∼ g,
(4.1)
k=1
where Γk are the jumps of a unit-rate homogeneous Poisson point process on
R+ , and τ : R+ × R+ → R+ is a non-negative measurable function such that
limu→∞ τ (v, u) = 0 for g-almost every v. The truncated CRM then takes the form
ΘK :=
K
X
θk δψk .
k=1
4See Lemma D.1 in Appendix D for its generalization to arbitrary discrete random measures.
5Our theory applies equally well if g is on some measurable space V, but we only use V = R .
+
TRUNCATED RANDOM MEASURES
13
Theorem 4.3 provides a general truncation error bound for series representations of
the form Eq. (4.1), specifies its range, and guarantees that the bound decays to 0
as K → ∞.
Theorem 4.3 (Series representation truncation error). The error in approximating
a series representation of Θ with its truncation ΘK satisfies
0≤
1
kpN,∞ − pN,K k1 ≤ 1 − e−BN,K ≤ 1,
2
where
Z
BN,K :=
∞
h
i
N
1 − E π (τ (V, u + GK ))
du,
(4.2)
0
indep
G0 := 0, GK ∼ Gam(K, 1) for K ≥ 1, and V
N, limK→∞ BN,K = 0.
indep
∼ g. Furthermore, ∀N ∈
Remark. An alternate form of BN,K that is sometimes easier to use in practice
can be found by applying the standard geometric series formula to Eq. (4.2), which
yields
N Z ∞ h
i
X
n−1
BN,K =
E π (τ (V, u + GK ))
(1 − π (τ (V, u + GK ))) du.
n=1
0
A simplified upper bound on BN,K can be derived by noting that π(θ) ≤ 1, so
Z ∞
BN,K ≤ N
(1 − E [π (τ (V, u + GK ))]) du.
(4.3)
0
This bound usually gives the same asymptotics in K as Eq. (4.2).
The main task in using Theorem 4.3 to develop a truncation error bound for a
series representation is evaluating the integrand in the definition of BN,K . Thus,
we next evaluate the integrand and provide expressions of the truncation error
bound for the four series representations outlined in Section 3.1. Throughout the
remainder of this section, GK is defined as in Theorem 4.3, F0 ≡ 1, and FK is the
CDF of GK .
Inverse-Lévy representation. For this representation we have
τ (v, u) = ν ← (u) := inf {y : ν ([y, ∞)) ≤ u} .
To evaluate the bound in Eq. (4.3), we use the transformation of variables x =
ν ← (u + GK ) and the fact that for a, b ≥ 0, ν ← (a) ≥ b ⇐⇒ a ≤ ν ([b, ∞)) to
conclude that
Z ∞
BN,K ≤ N
FK (ν[x, ∞))(1 − π(x)) ν(dx).
(4.4)
0
Recent work on the inverse-Lévy representation has developed Monte Carlo estimates of the error of the truncated random measure moments for those ν ([x, ∞))
with known inverse ν ← (Arbel and Prünster, 2017). In contrast, the result above
provides an explicit bound on the L1 truncation error. Our bound does not require
knowing ν ← , which is often the most challenging aspect of applying the inverse-Lévy
representation.
14
T. CAMPBELL, J. H. HUGGINS, J. P. HOW, AND T. BRODERICK
Example 4.1 (IL-Rep truncation for LomP(γ, λ−1 ) with Poisson likelihood). Recall from Example 3.4 that the Lomax process LomP(γ, λ−1 ) is the CRM with
rate measure ν(dθ) = γλθ−1 (1 + λθ)−1 dθ, so ν[x, ∞) = γλ log{1 + (λx)−1 }. Using
Eq. (4.4), we have
Z ∞
BN,K ≤ N γλ
FK (γλ log{1 + (λx)−1 })(1 − e−x )x−1 (1 + λx)−1 dx.
0
Since FK (t) ≤ tK /K! ≤ (3t/K)K , for any a > 0 the integral is upper bounded by
Z ∞
Z a
1
−1
x−1 (1 + λx)−1 dx
(1 + λx) dx + FK γλ log 1 +
λa
a
0
1
1
≤ a + FK γλ log 1 +
log 1 +
λa
λa
≤ λ−1 (eb − 1)−1 + b(3γλb/K)K
where
b := log{1 + (λa)−1 }.
(4.5)
Replacing (eb − 1)−1 with the approximation
e−b and then setting
the two terms in
n
o−1
K+2
1
Eq. (4.5) equal, we obtain b = kW0
3γλ K+1 (K + 1) K+1
, where W0 is the
product logarithm function, i.e.
W0 (y) = x ⇐⇒ xex = y.
(4.6)
Thus, using the fact that e−t ≤ (et − 1)−1 , we conclude that
BN,K ≤
2N γ[1 + (3γλ)−1 ]
n
o−1
K+2
1
K+1
K+1
exp KW0
(K + 1)
−1
3γλ
∼ 2N γ[1 + (3γλ)−1 ]e−KW0 ({3γλ}
Bondesson representation.
−1
)
K → ∞.
For this representation we have
τ (v, u) = ve−u/c ,
g(dv) = −c−1
d
(vν(v)) dv.
dv
Writing the expectation over V explicitly as an integral with measure g(v)dv, using
the transformation of variables u = −c log x/v (so x = ve−u/c ), and given the
d
(vν(v)) for the Bondesson representation, we have
definition of g(v) = −c−1 dv
Z ∞
h
i
BN,K ≤ N
1 − E π ve−GK /c
ν(dv).
0
Example 4.2 (Truncation of the Bondesson representation for ΓP(γ, λ, 0)). Let
D
G̃K = GK /(γλ). Since π(θ) = e−θ and c = γλ, we have
Z ∞
Z ∞
h
i
−GK /c
−ve−G̃K
−1 −λv
)v e
dv
1 − E π(ve
) ν(dv) = γλ E
(1 − e
0
h 0
i
= γλ E log(1 + e−G̃K /λ)
K
i
h
γλ
−G̃K
≤γE e
=γ
.
1 + γλ
TRUNCATED RANDOM MEASURES
15
The second equality follows by using the power series for the exponential integral (Abramowitz and Stegun, 1964, Chapter 5). Thus,
K
γλ
BN,K ≤ N γ
.
1 + γλ
Thinning representation. For this representation we have
dν
τ (v, u) = v1
(v) ≥ u ,
g any distribution on R+ s.t. ν g.
dg
Since π(0) = 1 by Lemma B.3, we have that 1 − π (v1(A)) = (1 − π (v)) 1(A) for
any event A. Using this fact, we have
Z ∞
Z dν
dg (v)
FK (u) du g(dv).
(4.7)
BN,K ≤ N
(1 − π(v))
0
0
Analytic bounds for the thinning representation of specific processes tend to be
opaque and notationally cumbersome, so we simply compare its truncation error in
Section 7 to the other representations by numerical approximation of Eq. (4.7).
Rejection representation. Assume that we can use the inverse-Lévy representation to simulate PoissP(µ). Then for the rejection representation we have
dν ←
τ (v, u) = µ← (u)1
(µ (u)) ≥ v ,
g(dv) = 1[0 ≤ v ≤ 1]dv,
dµ
dν
where µ satisfies ν µ, dµ
≤ 1, and µ← (u) := inf {x : µ ([x, ∞)) ≤ u}. Using the
same techniques as for the thinning and inverse-Lévy representations, we have that
Z ∞
BN,K ≤ N
FK (µ[x, ∞))(1 − π(x)) ν(dx).
(4.8)
0
Example 4.3 (R-Rep truncation for ΓP(γ, λ, 0) with Poisson likelihood). Using
Eq. (4.8) and the fact that 1 − e−x ≤ x, we have
Z ∞
BN,K ≤ N γλ
FK (γλ log{1 + (λx)−1 })e−λx dx.
(4.9)
0
Arguing as in Example 4.1, we see that the integral in Eq. (4.9) is upper bounded
by
Z ∞
Z a
1
e−λx dx + FK γλ log 1 +
e−λx dx
λa
0
a
1
≤ a + λ−1 FK γλ log 1 +
λa
−1
b
−1
=λ
(e − 1) + (3γλb/K)K ,
(4.10)
where b := log{1 + (λa)−1 }. Replacing (eb − 1)−1 with the approximation e−b
and then setting the two terms in Eq. (4.10) equal to each other, we obtain b =
KW0 ({3γλ}−1 ) (where W0 is defined in Eq. (4.6)) and conclude that
BN,K ≤
2N γ
eKW0 ({3γλ}−1 )
−1
∼ 2N γe−KW0 ({3γλ}
−1
)
K → ∞.
16
T. CAMPBELL, J. H. HUGGINS, J. P. HOW, AND T. BRODERICK
Example 4.4 (R-Rep truncation for ΓP(γ, λ, d) with Poisson likelihood, d > 0).
We have
Z ∞
γλ1−d
BN,K ≤ N
FK (γ 0 x−d )(1 − e−x )x−1−d e−λx dx.
Γ(1 − d) 0
The integral can be upper bounded as
Z
Z a
−d
0 −d
x dx + FK (γ a )
∞
(1 + e−x )x−1−d e−λx dx
a
0
≤ (1 − d)−1 a1−d + Γ(−d)(λd − (1 + λ)d )(3γ 0 K −1 a−d )K .
Setting the two terms equal and solving for a, we obtain
Kd(1−d)
d(1−d)+K
1−d
d(1−d)
γλ1−d
3γλ
BN,K ≤ 2N
[(1 − d)Γ(−d)] d(1−d)+K
Γ(2 − d)
dΓ(1 − d)K
d(1−d)
3γλ1−d
γλ1−d
K −d(1−d)
K → ∞.
∼ 2N
Γ(2 − d) dΓ(1 − d)
4.2. Superposition representations. For superposition representations, the
truncated CRM takes the form
ΘK :=
Ck
K X
X
θki δψki .
k=1 i=1
Let Θ+
K := Θ − ΘK denote the tail measure. By the superposition property of
Poisson point processes (Kingman, 1993), the tail measure is itself a CRM with
+
some rate measure νK
and is independent of ΘK :
Θ+
K =
Ck
∞
X
X
+
θki δψki ∼ CRM νK
,
Θ+
⊥ ΘK ,
K ⊥
Θ = Θ K + Θ+
K.
(4.11)
k=K+1 i=1
The following result provides a general truncation error bound for superposition
representations, specifies its range, and guarantees that the bound decays to 0 as
K → ∞.
Theorem 4.4 (Superposition representation truncation error). The error in approximating a superposition representation of Θ ∼ CRM(ν) with its truncation ΘK
satisfies
1
0 ≤ kpN,∞ − pN,K k1 ≤ 1 − e−BN,K ≤ 1,
2
where
Z
+
BN,K :=
1 − π(θ)N νK
(dθ).
(4.12)
Furthermore, ∀N ∈ N, limK→∞ BN,K = 0.
Remark. As for series representations, an alternate form of BN,K that is sometimes
easier to use can be found by applying the standard geometric series formula to
Eq. (4.12):
N Z
X
+
BN,K =
π(θ)n−1 (1 − π(θ)) νK
(dθ).
n=1
TRUNCATED RANDOM MEASURES
17
A simplified upper bound on BN,K can be derived by noting that π(θ) ≤ 1, so
Z ∞
+
BN,K ≤ N
(1 − π (θ)) νK
(dθ).
0
This bound usually gives the same asymptotics in K as Eq. (4.12).
The main task in using Theorem 4.4 to develop a truncation error bound for a
+
superposition representation is determining its tail measure νK
. In the following,
we provide the tail measure for the three superposition representations outlined in
Section 3.2.
Decoupled Bondesson representation. For each point process in the superposition, an average of c/ξ atoms are generated with independent weights of the
indep
indep
form V e−Tk where V ∼ g and Tk ∼ Gam(k, ξ). Therefore, the tail measure is
∞
c X
+
g̃k,ξ (θ)dθ,
νK
(dθ) =
ξ
k=K+1
−Tk
where g̃k,ξ is the density of V e
. The bound for the decoupled Bondesson representation can therefore be expressed as
∞
c X
BN,K ≤ N
E 1 − π V e−Tk .
ξ
k=K+1
Example 4.5 (Decoupled Bondesson representation truncation for ΓP(γ, λ, 0)).
Using the fact that 1 − e−θ ≤ θ, we have
∞
∞
N γλ X
N γλ X
BN,K =
E[1 − π(Vk1 e−Tk1 )] ≤
E[Vk1 e−Tk1 ]
ξ
ξ
k=K+1
k=K+1
k
K
∞
X
N γλ
1
ξ
ξ
=
= Nγ
,
ξ
λ 1+ξ
1+ξ
k=K+1
which is equivalent (up to a factor of 2) to the bound in Roychowdhury and Kulis
(2015).
Size-biased representation. The constructive derivation of the size-biased representation (Broderick et al., 2017, proof of Theorem 5.1) immediately yields
+
νK
(dθ) = π(θ)K ν(dθ).
Therefore, the size-biased representation truncation error bound can be expressed
using the formula for ηk from Eq. (3.2) as
N Z
N
X
X
BN,K =
π(θ)K+n−1 (1 − π(θ))ν(dθ) =
ηK+n .
(4.13)
n=1
n=1
Example 4.6 (Size-biased representation truncation for ΓP(γ, λ, d)). For d > 0,
the standard gamma integral yields
Z
γλ1−d
ηk = π(θ)k−1 (1 − π(θ))ν(dθ) =
(λ + k)d − (λ + k − 1)d .
d
The sum from Eq. (4.13) is telescoping, so canceling terms,
γλ1−d
BN,K ≤
(λ + K)d − (λ + K + N )d ∼ γN λ1−d K d−1
K → ∞,
d
18
T. CAMPBELL, J. H. HUGGINS, J. P. HOW, AND T. BRODERICK
where the asymptotic result follows from Lemma B.4. To analyze the d = 0 case,
we use L’Hospital’s rule to take the limit of the integral:
Z
lim π(θ)k−1 (1 − π(θ))ν(dθ) = γλ (log(λ + k) − log(λ + k − 1)) .
d→0
Canceling terms in the telescopic sum yields
BN,K ≤ γλ (log(λ + K) − log(λ + K + N )) ∼ γλN K −1
K → ∞,
where the asymptotic result follows from an application of Lemma B.4.
Power-law representation. For each point process in the superposition, an avQk−1
erage of γ atoms are generated with independent weights of the form V Uk `=1 (1−
indep
indep
U` ), where V ∼ g and U` ∼ Beta(1 − d, α + `d). Therefore, the tail measure is
+
νK
(dθ) = γ
∞
X
g̃k (θ)dθ,
k=K+1
Qk−1
where g̃k is the density of the random variable V Uk `=1 (1 − U` ). The truncation
error bound may be expressed as
"
!#
∞
k−1
X
Y
BN,K ≤ N γ
E 1 − π V Uk
(1 − U` )
.
k=K+1
`=1
Example 4.7 (Power-law representation truncation for ΓP(γ, λ, d)). Let βk be a
random variable with density g̃k (with λ in the place of α). Using 1 − e−θ ≤ θ, we
have
" ∞
#
∞
∞
K
X
X
X
Y
λ + kd
,
E[1 − π(βk )] ≤
E[βk ] = E
βk =
λ + kd − d + 1
k=K+1
k=K+1
k=K+1
k=1
where the final equality follows from Ishwaran and James (2001, Theorem 1). Thus,
K
λ
K
Y
d=0
λ + kd
λ+1
BN,K ≤ γN
∼ γN
K → ∞,
λ+1
−1
Γ
(
)
d
1−d
λ + kd − d + 1
0
<
d
<
1
λ+d K
k=1
Γ( d )
(4.14)
where the 0 < d < 1 case in Eq. (4.14) follows by Lemma B.5 applied to
K
Y
k=1
λ + kd
Γ((λ + 1)/d) Γ(λ/d + K + 1)
=
.
λ + kd − d + 1
Γ((λ + d)/d) Γ(λ/d + K + d−1 )
4.3. Stochastic mapping. We now show how truncation bounds developed elsewhere in this paper can be applied to CRM representations that have been transformed using Lemma 3.4. For Θ ∼ CRM(ν), we denote its transformation by
Θ̃ = κ(Θ). For any object defined with respect to Θ, the corresponding object
for Θ̃ is denoted with a tilde. For example, in place of N and X1:N (for Θ), we
use Ñ and X̃1:Ñ (for Θ̃). We make BN,K a function of π(θ) in the notation of
Proposition 4.5; when one applies stochastic mapping to a CRM, one usually also
wants to change the likelihood h(x | θ), and thus also changes π(θ) = h(0 | θ). The
proof of Proposition 4.5 may be found in Appendix D.
TRUNCATED RANDOM MEASURES
19
Proposition 4.5 (Truncation error under a stochastic mapping). Consider a representation for Θ ∼ CRM(ν) with truncation error bound BN,K (π). Then for
any likelihood h̃(x | u), if Θ̃ is a stochastic mapping of Θ under the probability kernel κ(θ, du), its truncation error bound is B1,K (πκ,Ñ ), where πκ,Ñ (θ) :=
R
h̃(0 | u)Ñ κ(θ, du).
4.4. Hyperpriors. In practice, prior distributions are often placed on the hyperparameters of the CRM rate measure (i.e. γ, α, λ, d, etc.). We conclude our
investigation of CRM truncation error by showing how bounds developed in this
section can be modified to account for the use of hyperpriors. Note that we make
the dependence of BN,K on the hyperparameters Φ explicit in the notation of
Proposition 4.6.
Proposition 4.6 (CRM truncation error with a hyperprior). Given hyperparameters Φ, consider a representation for Θ | Φ ∼ CRM(ν), and let BN,K (Φ) be given
by Eq. (4.2) (for a series representation) or Eq. (4.12) (for a superposition representation). The error of approximating Θ with its truncation ΘK satisfies
0≤
1
kpN,∞ − pN,K k1 ≤ 1 − e−E[BN,K (Φ)] ≤ 1.
2
Example 4.8 (Decoupled Bondesson representation truncation for ΓP(γ, λ, 0)).
A standard choice of hyperprior for the mass γ is a gamma distribution, i.e. γ ∼
Gam(a, b). Combining Proposition 4.6 and Example 4.5, we have that
E [BN,K (Φ)] ≤ N
a
b
ξ
ξ+1
K
.
5. Normalized truncation analysis
In this section, we provide truncation error bounds for normalized CRMs
(NCRMs). Examples include the Dirichlet process (Ferguson, 1973), the normalized gamma process (Brix, 1999; James, 2002; Pitman, 2003; Lijoi and Prünster,
2010), and the normalized σ-stable process (Kingman, 1975; Lijoi and Prünster,
2010). Given a CRM Θ on Ψ, we define the corresponding NCRM Ξ via Ξ(S) :=
Θ(S)/Θ(Ψ) for each measurable subset S ⊆ Ψ. Likewise, given a truncated CRM
ΘK , we define its normalization ΞK via ΞK (S) := ΘK (S)/ΘK (Ψ). Note that any
simulation algorithm for ΘK can be used for ΞK by simply normalizing the result.
This does not depend on the particular representation of the CRM, and thus applies
equally to all the representations in Section 3.
The first step in the analysis of NCRM truncations is to define their approximation error in a manner similar to that of CRM truncations. Since Ξ and ΞK are
both normalized, they are distributions on Ψ; thus, observations X1:N are generated
i.i.d. from Ξ, and Z1:N are generated i.i.d. from ΞK . Y1:N and W1:N have the same
definition as for CRMs. As in the developments of Section 4, the theoretical results
of this section rely on a general upper bound, provided by Lemma 5.1. Proposition 5.2 shows that the bound in Lemma 5.1 is tight without further assumptions
on the data likelihood f .
20
T. CAMPBELL, J. H. HUGGINS, J. P. HOW, AND T. BRODERICK
Lemma 5.1 (NCRM protobound). Let Θ ∼ CRM(ν), and let its truncation be
ΘK . Let their normalizations be Ξ and ΞK respectively. If
i.i.d.
Zn | ΞK ∼ ΞK ,
i.i.d.
indep
Wn | Zn ∼ f (· | Zn ),
Xn | Ξ ∼ Ξ,
indep
Yn | Xn ∼ f (· | Xn ),
then
1
kpN,∞ − pN,K k1 ≤ 1 − P (X1:N ⊆ supp(ΞK )) ,
2
where pN,∞ , pN,K are the marginal densities of Y1:N and W1:N , respectively.
Proposition 5.2 (Protobound tightness). If G is non-atomic and Ψ is Borel, then
for any δ > 0, there exists a likelihood f such that Lemma 5.1 is tight up to a factor
of 1 − δ.
The analysis of CRMs in Section 4 relied heavily on the Poisson process stucture
of the rates in Θ and X1:N ; unfortunately, the rates in Ξ do not possess the same
structure and thus lack many useful independence properties (the rates must sum
to one). Likewise, sampling Xn for each n does not depend on the atoms of Ξ
independently (Xn randomly selects a single atom based on their rates). Rather
than using the basic definitions of the above random quantities to derive an error
bound, we decouple the atoms of Ξ and X1:N using a technique from extreme
value theory. A Gumbel random variable T with location µ ∈ R and scale σ > 0,
denoted T ∼ Gumbel(µ, σ), is defined by the cumulative distribution function and
corresponding density
t−µ
−(
1 −( t−µ
σ )
σ )−e
e
.
σ
An interesting property of the Gumbel distribution is that if one perturbs the logprobabilities of a finite discrete distribution by i.i.d. Gumbel(0, 1) random variables,
the arg max of the resulting set is a sample from the discrete distribution (Gumbel,
1954; Maddison et al., 2014). This technique is invariant to normalization, as the
arg max is invariant to the corresponding constant shift in the log-transformed
space. For present purposes, we develop the infinite extension of this result:
P(T ≤ t) = e−e
−
t−µ
σ
and
∞
Lemma 5.3 (Infinite
P Gumbel-max sampling).pjLet (pi )i=1∞be a collection of positive
numbers such that i pi < ∞ and let p̄j := P pi . If (Ti )i=1 are i.i.d. Gumbel(0, 1)
i
random variables, then arg maxi∈N Ti + log pi exists, is unique a.s., and has distribution
∞
arg max Ti + log pi ∼ Categorical (p̄j )j=1 .
i∈N
The proof of this result, along with the others in this section, may be found
in Appendix E. The utility of Lemma 5.3 is that it allows the construction of Ξ
and X1:N without the problematic coupling of the underlying CRM atoms due to
normalization; rather than dealing directly with Ξ, we log-transform the rates of Θ,
perturb them by i.i.d. Gumbel(0, 1) random variables, and characterize the distribution of the maximum rate in this process. The combination of this distribution with
Lemma 5.3 yields the key proof technique used to develop the truncation bounds
in Theorems 5.4 and 5.5. The results presented in this section are summarized in
Table 1 in Section 7.
TRUNCATED RANDOM MEASURES
21
5.1. Series representations. The following result provides a general truncation
error bound for normalized series representations, specifies its range, and guarantees
that it decays to 0 as K → ∞. We again use the general series representation
notation from Eq. (4.1), where g is a distribution on R+ , and τ : R+ × R+ → R+
is a measurable function such that limu→∞ τ (v, u) = 0 for g-almost every v.
Theorem 5.4 (Normalized series representation truncation error bound). The error of approximating a series representation of Ξ ∼ NCRM(ν) with its truncation
ΞK satisfies
1
N
0 ≤ kpN,∞ − pN,K k1 ≤ 1 − (1 − BK ) ≤ 1,
2
where
Z ∞
Z 1
K−1
d R∞
BK := E
J (ΓK , t)
J (ΓK u, t) du
− e 0 (J(u+ΓK ,t)−1)du dt ,
dt
0
0
(5.1)
h
i
J(u, t) = E e−t·τ (V,u) ,
V ∼ g,
and
ΓK ∼ Gam(K, 1).
Furthermore, limK→∞ BK = 0.
Example 5.1 (Dirichlet process, DP(γ), B-Rep). The Dirichlet process with concentration γ > 0 is a normalized gamma process NΓP(γ, 1, 0). From Example 3.2
we have cν = γ and gν = Exp(1), and from Section 4.1 we have τ (v, u) = ve−u/cν .
Therefore J and its antiderivative are
Z
h
i
−1
−u/γ
J(u, t) = E e−tV e
= 1 + te−u/γ
and
J(u, t)du = γ log eu/γ + t .
Using the antiderivative to evaluate the integrals in the formula for BK , writing
the expectation over ΓK ∼ Gam(K, 1) explicitly, and making a change of variables
we have
K−1
K
Z Z
s+t
γ
γ K+1 ∞ ∞
−(γ+2)
log
(s + t)
ds dt =
,
BK =
Γ(K) 0
1+t
1+γ
1
where the last equality is found by multiplying and dividing the integrand by (1 +
s+t
t)−(γ+2) , and making the change of variables from s to x = log 1+t
. Therefore, the
truncation error can be bounded by
K !N
K
1
γ
γ
kpN,∞ − pN,K k1 ≤ 1 − 1 −
∼N
K → ∞.
2
γ+1
γ+1
The bound in Example 5.2 has exponential decay, and reproduces earlier DP
truncation error bound rates due to Ishwaran and James (2001) and Ishwaran and
Zarepour (2002). However, the techniques used in past work do not generalize
beyond the Dirichlet process, while those developed here apply to any NCRM.
5.2. Superposition representations. The following result provides a general
truncation error bound for normalized superposition representations, specifies its
range, and guarantees that it decays to 0 as K → ∞. We once again rely on the
property that the truncation ΘK and tail Θ+
K are mutually independent CRMs, as
+
expressed in Eq. (4.11), with the tail measure denoted νK
.
22
T. CAMPBELL, J. H. HUGGINS, J. P. HOW, AND T. BRODERICK
Theorem 5.5 (Truncation error bound for normalized superposition representations). The error of approximating a superposition representation of Ξ ∼ NCRM(ν)
with its truncation ΞK satisfies
1
N
0 ≤ kpN,∞ − pN,K k1 ≤ 1 − (1 − BK ) ≤ 1,
2
where
R
Z ∞ Z
−θt
−θt +
:=
θe νK (dθ) e (e −1)ν(dθ) dt.
(5.2)
BK
0
Furthermore, limK→∞ BK = 0.
This bound can be applied by using the tail measures derived earlier in Section 4.2.
Example 5.2 (Dirichlet process, DP(γ), DB-Rep). As in Example 5.1, we view
the Dirichlet process with concentration γ > 0 as a normalized gamma process
NΓP(γ, 1, 0). First, by Lemma B.8, the integral in the exponential is
Z
Z ∞
−tθ
−tθ
−1 −θ
exp
(e
− 1)ν(dθ) = exp γ
(e
− 1)θ e dθ = (t + 1)−γ .
0
Example 3.2 shows cν = γ and gν (v) = e−v , and Eq. (C.1) provides the tail measure
+
νK
for the decoupled Bondesson representation,
Z 1
∞
−1
γ X ξk
+
νK
(dθ) =
(− log x)k−1 xξ−2 e−θx dx dθ.
ξ
Γ(k)
0
k=K+1
Substituting this result, using Fubini’s theorem to swap the order of integration and
summation, evaluating the integral over θ, and making the substitution x = e−s
yields
ZZ
∞
γ X ξk
sk−1 e−(ξ−1)s (t + 1)−γ
BK =
ds dt.
ξ
Γ(k)
(es + t)2
s,t≥0
k=K+1
Noting that ∀s ≥ 0, es ≥ 1, we have for any a ∈ (0, 1] ∩ (0, γ),
Z ∞
Z ∞
∞
γ X ξk
k−1 −(ξ+a)s
BK ≤
s
e
ds
(t + 1)−(γ+1−a) dt
ξ
Γ(k) 0
0
k=K+1
k
K
∞
X
ξ
γ
ξ
γ
=
.
=
(γ − a)ξ
ξ+a
a(γ − a) ξ + a
k=K+1
Therefore, for any a ∈ (0, 1] ∩ (0, γ),
1
kpN,∞ − pN,K k1 ≤ 1 −
2
γ
1−
a(γ − a)
ξ
ξ+a
K !N
∼
Nγ
a(γ − a)
ξ
ξ+a
K
K → ∞.
To find the tightest bound, one can minimize with respect to a given γ, ξ, K.
Example 5.3 (Normalized gamma process, NΓP(γ, λ, d), SB-Rep). By
Lemma B.8, the integral in the exponential is
Z
( exp −γλ1−d d−1 ((t + λ)d − λd ) d > 0
γλ
exp
(e−θt − 1)ν(dθ) =
(5.3)
λ
d = 0,
t+λ
TRUNCATED RANDOM MEASURES
23
and the standard gamma integral yields
Z
Z
λ1−d
−θt +
θe νK (dθ) = γ
θ−d e−(K+t+λ)θ dθ = γλ1−d (K + t + λ)d−1 . (5.4)
Γ(1 − d)
When d > 0, multiplying the previous two displays and integrating over t ≥ 0
yields
Z ∞
1−d d
d−1
1−d γλ/d
(K + t)d−1 e−γλ t /d dt ≤ Cγ,λ,d (K + λ)
,
BK = γλ e
λ
d−1
d−1
where we have used (K + t)
≤ (K + λ)
for t ≥ λ and the change of variables
−1
u = γλ1−d d−1 td to find that Cγ,λ,d = eσ σ 1−d λ1−d Γ d−1 , σ , where σ = γλd−1
R∞
and Γ(a, x) := x θa−1 e−θ dθ is the upper incomplete gamma function. Therefore,
N
1
kpN,∞ − pN,K k1 ≤ 1 − 1 − Cγ,λ,d (K + λ)d−1
∼ N Cγ,λ,d K d−1 K → ∞.
2
Setting λ = 0 in the above expression immediately yields the bound for the normalized d-stable process (Kingman, 1975).
When d = 0, multiplying Eqs. (5.3) and (5.4) and integrating over t ≥ 0 yields
1
1−γλ
Z ∞
−λ1−γλ
γλ (K+λ)
−1
BK
(K
+
λ)
γλ 6= 1
1−γλ
=
(K + t)−1 t−γλ dt ≤
γλ1+γλ
λ
γλ = 1,
K −1 log K+λ
λ
where we obtain the bound for γλ 6= 1 by splitting the integral into the intervals
[λ, K + λ] and [K + λ, ∞) and bounding each section separately, and we obtain the
bound for γλ = 1 via the transformation u = t/(K + t). Therefore, asymptotically
1
Cγ,λ K − min(1,γλ) γλ 6= 1
K → ∞,
kpN,∞ − pN,K k1 . N
λK −1 log K
γλ = 1
2
γλ
γλ2
λ
where Cγ,λ := max 1−γλ
, γλ−1
.
Truncation of the NΓP(γ, λ, d) has been studied previously: Argiento et al.
(2016) threshold the weights of the unnormalized CRM to be beyond a fixed level
> 0 prior to normalization, and develop error bounds for that method of truncation. These results are not directly comparable to those of the present work due
to the different methods of truncation (i.e. sequential representation termination
versus weight thresholding).
5.3. Hyperpriors. As in the CRM case, we can place priors on the hyperparameters of the NCRM rate measure (i.e. γ, α, λ, d, etc.). We conclude our investigation
of NCRM truncation error by showing how bounds developed in this section can
be modified to account for hyperpriors. Note that we make the dependence of BK
on the hyperparameters Φ explicit in the notation of Proposition 5.6.
Proposition 5.6 (NCRM truncation error with a hyperprior). Given hyperparameters Φ, consider a representation for Θ | Φ ∼ CRM(ν), let Ξ | Φ be its normalization,
and let BK (Φ) be given by Eq. (5.1) (for a series representation) or Eq. (5.2) (for
a superposition representation). The error of approximating Ξ with its truncation
ΞK satisfies
1
N
0 ≤ kpN,∞ − pN,K k1 ≤ 1 − (1 − E [BK (Φ)]) ≤ 1.
2
24
T. CAMPBELL, J. H. HUGGINS, J. P. HOW, AND T. BRODERICK
Example 5.4 (Dirichlet process, DP(γ), B-Rep). If we place a Lomax prior on γ,
i.e. γ ∼ LomP(a, 1), then combining Proposition 5.6 and Example 5.1 yields
N
1
Γ(a + 1)Γ(K + 1)
kpN,∞ − pN,K k1 ≤ 1 − 1 −
∼ N Γ(a + 1)(K + 1)−a K → ∞.
2
Γ(a + K + 1)
6. Simulation and computational complexity
The sequential representations in Section 3 are each generated from a different
finite sequence of distributions, resulting in a different expected computational cost
for the same truncation level. Thus, the truncation level itself is not an appropriate
parameter with which to compare the error bounds for different representations and
we require a characterization of the computational cost. We investigate the mean
complexity E[R] of each representation, where R is the number of random variables
sampled, as a function of the truncation level for each of the representations in
Section 3.
We begin with the series representations. For each value of k = 1, . . . , K, each
series representation generates a single trait ψk ∼ G and a rate θk composed of
some transformation of random variables. Thus, all of the series representations in
this work satisfy E[R] = rK for some constant r: by inspection, the inverse-Lévy
representation has r = 2, and all the remaining series representations have r = 3.
The superposition representations, on the other hand, generate a Poisson random
variable to determine the number of atoms at each value of k = 1, . . . , K, and then
generate those atoms. Therefore, the mean simulation complexity takes the form
PK
E[R] = k=1 1 + rk E[Ck ] for some constants rk that might depend on the value
of k. For the decoupled Bondesson representation, rk = 3 since each atom
requires
generating three values (ψki , Vki , and Tki ), and E[Ck ] = c/ξ, so E[R] = 3c
ξ + 1 K.
For the size-biased representation, rk = 3 since each atom requires generating three
PK
values (ψki , zki , and θki ), and E[Ck ] = ηk , so E[R] = K + 3 k=1 ηk . Note that
here E[R] ∼ K, for K → ∞ since ηk is a decreasing sequence. For the power-law
representation, rk = k + 2, since each atom requires
ψki , Vki , and k beta
generating
γ
2
K
.
random variables, and therefore E[R] = 1 + 5γ
K
+
2
2
7. Summary of results
Table 1 summarizes our truncation and simulation cost results as applied to the
beta, (normalized) gamma, and beta prime processes. Results for the Bondesson
representation of BP(γ, 1, 0) as well as the decoupled Bondesson representations
of BP(γ, α, 0) and ΓP(γ, λ, 0) were previously known, and are reproduced by our
results. All other results in the table are novel to the best of the authors’ knowledge. It is interesting to note that the bounds and expected costs within each of the
representation classes often have the same form, aside from some constants. Across
classes, however, they vary significantly, indicating that the chosen sequential representation of a process has more of an influence on the truncation error than the
process itself.
Fig. 1 shows a comparison of how the truncation error bounds vary with the expected computational cost E[R] of simulation for the (normalized) gamma process
and Poisson likelihood with N = 5 observations. Results shown for the thinning,
rejection, and inverse-Lévy representations are computed by Monte-Carlo approximation of the formula for BN,K in Eq. (4.3), while all others use closed-form
TRUNCATED RANDOM MEASURES
25
Table 1. Asymptotic error bounds and simulation cost summary.
Error bounds are presented up to a constant that varies between
models. Be = Bernoulli, OBe = odds Bernoulli, Poi = Poisson.
Rep. Random Measure h
IL
B
T
R
DB
SB
PL
LomP(γ, α−1 )
Poi
BP(γ, α ≥ 1, 0)
ΓP(γ, α, 0)
BPP(γ, α, 0)
Be
Poi
OBe
Asymptotic Error Bound
−1
N e−KW0 ({3γα}
Nγ
γα
γα+1
γ
γ+1
)
—
BP(γ, α, 0)
ΓP(γ, α, 0)
BPP(γ, α, 0)
—
See Eq. (4.7)
Be
−KW0 ({3γα}−1 )
d=0
e
Poi
N
K −d(1−d)
d > 0 (ΓP)
OBe
K −1/d
d > 0 (BP, BPP)
BP(γ, α ≥ 1, 0)
ΓP(γ, α, 0)
BPP(γ, α > 1, 0)
Be
Poi
OBe
—
BP(γ, α, d)
ΓP(γ, α, d)
BPP(γ, α, d)
Be
Poi
OBe
NΓP(γ, α, d)
—
BP(γ, α, d)
ΓP(γ, α, d)
BPP(γ, α > 1, d)
Be
Poi
OBe
ξ
ξ+1
3K
K
—
N
2K
K
DP(γ)
DP(γ)
N
Complexity
3K
3K
K
Nγ
a(γ−a)
ξ
ξ+a
K
, a ∈ (0, 1] ∩ (0, γ)
N K d−1
d = 0, γα = 1
K −1 log K
N
K − min(1,γα) d = 0, γα 6= 1
K d−1
d>0
K
α
d = 0 (BP, ΓP)
α+1
N
−K
2
d
= 0 (BPP)
K 1−1/d
d>0
3c
ξ
+1 K
K
γ
K2
2
expressions from the examples in Sections 4 and 5. Note that the Bondesson and
decoupled Bondesson representations do not exist when d > 0. Further, only those
representations for which we provide closed-form bounds in the examples are shown
for the normalized gamma process; we leave the numerical approximation of the
results from Theorems 5.4 and 5.5 as an open problem. Similar figures for other
processes (in particular, the beta-Bernoulli and beta prime-odds Bernoulli) are provided in Appendix A. Note that all bounds presented are improved by a factor of
two versus comparable past results in the literature, due to the reliance on Lemmas 4.1 and 5.1 rather than the earlier bound found in Ishwaran and James (2001).
In Fig. 1, the top row shows results for the light-tailed process (γ = 1, λ = 1,
d = 0, and ξ = c = γλ). All representations except for thinning and size-biased
capture its exponential truncation error decay. This is due to the fact that the thinning representation generates increasingly many atoms of weight 0 as K → ∞, and
the expected number of atoms at each outer index for the size-biased representation
decays as K → ∞. The inverse-Lévy representation has the lowest truncation error
as expected, as it is the only representation that generates a nonincreasing sequence
of weights (and so must be the most efficient (Arbel and Prünster, 2017)). Based
on this figure and those in Appendix A for other processes, it appears that the Bondesson representation typically provides the best tradeoff between simplicity and
26
T. CAMPBELL, J. H. HUGGINS, J. P. HOW, AND T. BRODERICK
100
100
ILRep
BRep
TRep
RRep
SBRep
DBRep
PLRep
102
103
104
105
101
100
102
103
Mean Computational Cost
101
L1 Error
L1 Error
d = 0.0
101
104
L1 Error
L1 Error
d = 0.1
103
102
103
104
102
103
104
102
103
104
Mean Computational Cost
102
103
104
101
100
102
103
Mean Computational Cost
104
105
101
100
101
Mean Computational Cost
101
102
L1 Error
L1 Error
101
100
104
d = 0.5
105
101
102
103
104
105
103
104
101
105
102
102
103
104
101
102
103
Mean Computational Cost
ΓP
104
105
101
Mean Computational Cost
NΓP
Figure 1. Truncation error bounds for representations of the
(normalized) gamma-Poisson process, with γ = 1, λ = 2, and ξ =
γλ. The left column is for the unnormalized process, while the right
column is for the normalized process. Each row displays results for
a different value of the discount parameter d ∈ {0, 0.1, 0.5}.
efficiency, and should be used whenever its conditions in Theorem 3.1 are satisfied.
When the technical conditions are not satisfied, the rejection representation is a
good alternative. If ease of theoretical analysis is a concern, the decoupled Bondesson representation provides comparable efficiency with the analytical simplicity of
a superposition representation.
The bottom two rows of Fig. 1 show results for the heavy-tailed process (γ = 1,
λ = 2, and d ∈ {0.1, 0.5}). The representation options are more limited, as the
technical conditions of the Bondesson and decoupled Bondesson representations
are not satisfied. Here the rejection representation is often the best choice due to
its simplicity and competitive performance with the inverse-Lévy representation.
However, one must take care to check its efficiency beforehand using Proposition 3.2
given a particular choice of µ(dθ). For example, the choice of µ(dθ) ∝ θ−1−d dθ in
the present work makes the rejection representation very inefficient when d 1 for
both the gamma-Poisson (Fig. 1) and beta prime-odds Bernoulli (Fig. 3) processes,
but efficient for the beta-Bernoulli process (Fig. 2). If no µ(dθ) yields reasonable
results, the power-law representation is a good choice for d 1 as its truncation
TRUNCATED RANDOM MEASURES
27
bound approaches the exponential decay of the light-tailed process. For larger d > 0
the size-biased representation is a good alternative.
Based on the results in Fig. 1, it appears that there is no single dominant representation for all situations (provided the inverse-Lévy representation is intractable,
as it most often is). However, as a guideline, the rejection and Bondesson representations tend to be good choices for light-tailed processes, while the rejection,
size-biased, and power-law representations are good choices for heavy-tailed processes.
8. Discussion
We have investigated sequential representations, truncation error bounds, and
simulation algorithms for (normalized) completely random measures. In past work,
the development and analysis of these tools has occurred only on an ad hoc basis.
The results in the present paper, in contrast, provide a comprehensive characterization and analysis of the different types of sequential (N)CRM representations
available to the practitioner. However, there are a number of remaining open questions and limitations.
First, this work does not consider the influence of observed data: all analyses
assume an a priori perspective, as truncation is typically performed before data
are incorporated via posterior inference (e.g. in variational inference for the DP
mixture (Blei and Jordan, 2006) and BP latent feature model (Doshi-Velez et al.,
2009)). However, analysis of a posteriori truncation has been studied in past work
as well (Ishwaran and James, 2001; Gelfand and Kottas, 2002; Ishwaran and Zarepour, 2002). In the language of CRMs, observations introduce a fixed-location
component in the posterior process, while the unobserved traits are drawn from
the (possibly normalized) ordinary component of a CRM (Ishwaran and Zarepour,
2002; Broderick et al., 2017). We anticipate that this property makes observations
reasonably simple to include: the truncation tools provided in the present paper can
be used directly on the unobserved ordinary component, while the fixed-location
component may be treated exactly.
In addition, there are important open questions regarding the sequential representations developed in this work. It is unknown whether generalized versions
of the Bondesson and decoupled Bondesson representations can be developed for
larger classes of rate measures. The power-law representation does provide a partial answer in the decoupled Bondesson case. Regarding size-biased representations,
one might expect that the use of conjugate exponential family CRMs (Broderick
et al., 2017) would yield a closed-form expression for the truncation bound. In all
of the cases provided in this paper, this was indeed the case; the integrals were
evaluated exactly and a closed-form expression was found. However, we were unable to identify a general expression applicable to all conjugate exponential family
CRMs. Based on the examples provided, we conjecture that such an expression exists. Finally, fundamental connections between some of the representations were left
largely unexplored in this work. This is an open area of research, although progress
has been made by connecting decoupled Bondesson and size-biased representations
for (hierarchies of) generalized beta processes (Roy, 2014, Sec. 6.4).
A final remark is that one of the primary uses of sequential representations in
past work has been in the development of posterior inference procedures (Paisley
et al., 2010; Blei and Jordan, 2006; Doshi-Velez et al., 2009). The present work
100
101
102
103
104
105
ILRep
BRep
TRep
RRep
SBRep
DBRep
PLRep
101
102
103
Mean Computational Cost
d = 0.0
104
100
101
102
103
104
105
L1 Error
T. CAMPBELL, J. H. HUGGINS, J. P. HOW, AND T. BRODERICK
L1 Error
L1 Error
28
101
102
103
Mean Computational Cost
104
100
101
102
103
104
105
101
102
103
Mean Computational Cost
d = 0.1
104
d = 0.5
Figure 2. Truncation error bounds for the beta-Bernoulli process.
provides no guidance on which truncated representations are best paired with which
inference methods. We leave this as an open direction for future research, which
will require both theoretical and empirical investigation.
Acknowledgments
The authors thank the anonymous reviewers for their thoughtful comments,
which led to substantial improvements in both improvements in both the presentation and content of the paper. All authors are supported by the Office of Naval
Research under MURI grant N000141110688. J. Huggins is supported by the U.S.
Government under FA9550-11-C-0028 and awarded by the DoD, Air Force Office of
Scientific Research, National Defense Science and Engineering Graduate (NDSEG)
Fellowship, 32 CFR 168a. T. Campbell and T. Broderick are supported by DARPA
award FA8750-17-2-0019.
Appendix A. Additional Examples
In this section, we provide example applications of our theory to the beta process
and to the beta prime process.
A.1. Beta process. The beta process BP(γ, α, d) (Teh and Görür, 2009; Broderick
et al., 2012) with discount parameter d ∈ [0, 1), concentration parameter α > −d,
and mass parameter γ > 0, is a CRM with rate measure
ν(dθ) = γ
Γ(α + 1)
1 [θ ≤ 1] θ−1−d (1 − θ)α+d−1 dθ.
Γ(1 − d)Γ(α + d)
(A.1)
Setting d = 0 yields the standard beta process (Hjort, 1990; Thibaux and Jordan,
2007). The beta process is often paired with a Bernoulli likelihood or negative
binomial likelihood with s ∈ N failures:
Bern:
NegBinom:
h(x | θ) = 1 [x ≤ 1] θx (1 − θ)1−x ,
x+s−1
h(x | θ) =
(1 − θ)s θx .
x
Note that for the Bernoulli likelihood π(θ) = 1 − θ and for the negative binomial
likelihood π(θ) = (1 − θ)s .
TRUNCATED RANDOM MEASURES
29
Bondesson representation. If α > 1 and d = 0, then θν(θ) = γα(1 −
θ)α−1 1[θ ≤ 1] is non-increasing, cν = limθ→0 θν(θ) = γα, and gν (v) = (α −
1)(1 − v)α−2 = Beta(v; 1, α − 1). Thus, it follows from Theorem 3.1 that if
Θ ← B-Rep(γα, Beta(1, α − 1)), then Θ ∼ BP(γ, α, 0). In the case of α = 1,
gν (v) = δ1 , so Vk ≡ 1 and the Bondesson representation is equivalent to the inverseLévy representation. Since exp(−Ek /c) ∼ Beta(1, c), the representation used in Teh
et al. (2007) is equivalent to the Bondesson representation for BP(γ, 1, 0).
To obtain a truncation bound in the Bernoulli likelihood case, we can argue as
in Example 4.2:
Z ∞
h
i
1 − E π(ve−GK /(γα) ) ν(dv)
BN,K ≤ N
0
= N γα E[e−GK /(γα) ]
1
Z
(1 − v)α−1 dv
(A.2)
0
= Nγ
γα
1 + γα
K
.
This result generalizes that in Doshi-Velez et al. (2009), which applies only when
α = 1.
Theorem 3.1 does not apply directly when α < 1, since limθ→0 θν(θ) = ∞.
However, a representation can be obtained by using a trick from Paisley et al.
(2012). For α > 0, let
Θ0 =
C
X
θk0 δψk0 ,
k=1
i.i.d.
i.i.d.
where C ∼ Poiss(γ), θk0 ∼ Beta(1, α), and ψk0 ∼ G. Thus, Θ0 is a CRM with rate
measure γα(1 − θ)α−1 1[θ ≤ 1]dθ. If Θ ∼ BP(γα/(α + 1), α + 1, 0), which can be
generated according to Theorem 3.1, then Θ00 = Θ + Θ0 is a CRM with rate density
on [0, 1] given by
γαθ−1 (1 − θ)α + γα(1 − θ)α−1 = γαθ−1 (1 − θ)α−1 ,
hence Θ00 ∼ BP(γ, α, 0).
Thinning representation. If we let g = Beta(1 − d, α + d), then the thinning
representation for BP(γ, λ, d) is
Θ=
∞
X
Vk 1(Vk Γk ≤ γ)δψk ,
with
i.i.d.
Vk ∼ Beta(1 − d, α + d).
k=1
Rejection representation. To obtain a rejection representation for any d when
Γ(α+1)
α ≥ 1 − d, let µ be the rate measure for BP(γ Γ(2−d)Γ(α+d)
, 1 − d, d). We then have
that
(
(
γ 0 d−1 (x−d − 1) d > 0
(1 + du/γ 0 )−1/d d > 0
←
and
µ
(u)
=
µ[x, ∞) =
0
−γ 0 log x
d=0
e−u/γ
d = 0,
Γ(α+1)
where γ 0 := γ Γ(1−d)Γ(α+d)
. Thus, we can apply the inverse-Lévy method analytically for µ. Since we have constructed µ such that dν/dµ ≤ 1, we can use µ to
30
T. CAMPBELL, J. H. HUGGINS, J. P. HOW, AND T. BRODERICK
construct the rejection representation
∞
X
Θ=
Vk 1(Uk ≤ (1 − Vk )α+d−1 ) ∼ BP(γ, α, d),
α ≥ 1 − d,
k=1
(
with
Vk =
(1 + dΓk /γ 0 )−1/d
0
e−Γk /γ
d>0
,
d=0
i.i.d.
Uk ∼ Unif[0, 1].
The expected number of rejections is
h
i
(0,0,1,0)
(0,1,0,0)
−d−1 1 + 2d 2 F1
(−α − d, −d, −d; 1) + d 2 F1
(−α − d, −d, −d; 1) ,
where 2 F 1 is the ordinary hypergeometric function and the parenthetical superscripts indicate partial derivatives. This quantity monotonically diverges to ∞ as
d → 1.
To obtain a truncation bound in the Bernoulli likelihood case, we consider the
d > 0 and d = 0 settings separately. If d > 0, we have
Z 1
BN,K
≤
FK (γ 0 d−1 (x−d − 1))x−d (1 − x)α+d−1 dx
N γ0
0
Z a
Z 1
≤
x−d (1 − x)a+d−1 dx + FK (γ 0 d−1 (a−d − 1))
x−d (1 − x)α+d−1 dx
0
Z
≤
a
a
x−d dx + a−d FK (γ 0 d−1 (a−d − 1))
0
Z
1
(1 − x)α+d−1 dx
a
≤ (1 − d)−1 a1−d + a−d 3γ 0 d−1 (a−d − 1)/K
K
≤ (1 − d)−1 a1−d + 3γ 0 d−1 /K a−(K+1)d .
K
Setting the two terms equal and solving for a we obtain a1+d(K−1) =
K
3γ 0 (d−1 − 1)/K
and conclude that
K
0 −1
1/d
0 −1
3γ (d − 1)
3γ (d − 1) 1+d(K−1)
0
0
∼ 2N γ
K → ∞.
BN,K ≤ 2N γ
K
K
If d = 0, we have
BN,K
≤
N γ0
Z
1
FK (−γ 0 log x)(1 − x)α−1 dx
0
Z
≤
a
α−1
(1 − x)
0
1
Z
(1 − x)α−1 dx
dx + FK (−γ log x)
0
a
K
≤ a + (−3γ 0 log a/K) .
Setting the two terms equal and solving for a we conclude that
BN,K ≤ 2N γαe−KW0 ({3γα}
−1
)
,
where W0 is as defined in Eq. (4.6).
Decoupled Bondesson and power-law representations. The decoupled
Bondesson representation for BP(γ, α, 0) from Paisley et al. (2010) was extended
by Broderick et al. (2012) to the BP(γ, α, d) setting. The Broderick et al. (2012)
construction for the BP(γ, α, d) is in fact the “trivial” power-law representation
TRUNCATED RANDOM MEASURES
31
PL-Rep(γ, α, d, δ1 ) (the decoupled Bondesson representation is the special case
when d = 0).
In the Bernoulli likelihood case, the truncation bound for the decoupled Bondesson representation is
K
K
∞
∞
ξ
N γα X
ξ
N γα X 1
= Nγ
.
BN,K ≤
E[V e−Tk ] =
ξ
ξ
α 1+ξ
1+ξ
k=K+1
k=K+1
For the power law representation, by the same arguments as Example 4.7,
BN,K ≤ N γ
K
Y
k=1
α + kd
.
α + kd − d + 1
This result generalizes that in Paisley et al. (2012), which applies only when d = 0.
Size-biased representation. For the size-biased representation of the beta process is well-established and we refer the reader to Broderick et al. (2012, 2017) for
details. We note that the standard beta integral yields
Z
Γ(α + 1) Γ(α + d + k − 1)
,
ηk = ηk1 = π(θ)k−1 (1 − π(θ))ν(dθ) = γ
Γ(α + d)
Γ(α + k)
and for i > 1, ηki = 0. Hence zki = 1 almost surely and θki ∼ Beta(1 − d, α + d +
k − 1), demonstrating that the construction due to Thibaux and Jordan (2007) is
a special case of the size-biased representation for BP(γ, α, d).
To obtain a truncation bound for the Bernoulli likelihood case, first consider the
d > 0 setting. Using Lemma B.6 to simplify the sum in Eq. (4.13), we have
γ Γ(α + 1) Γ(α + d + K) Γ(α + d + K + N )
−
BN,K ≤
d Γ(α + d)
Γ(α + K)
Γ(α + K + N )
Γ(α + 1) d−1
∼ γN
K
K → ∞,
Γ(α + d)
where the asymptotic result follows from Lemmas B.4, B.7 and B.9. When d = 0,
we can again use Lemma B.6 to arrive at
BN,K ≤ γα (ψ(α + K) − ψ(α + N + K)) ∼ γαN K −1
K → ∞,
where ψ(·) is the digamma function, and the asymptotic result follows from
Lemma B.4.
We can also bound the truncation error in the case of the negative binomial
likelihood. For a fixed number of failures s > 0, and assuming α + d + (k − 1)s > 1,
integration by parts yields
Z
π(θ)k−1 (1 − π(θ))ν(dθ)
γ Γ(α + 1) Γ(α + d + ks) Γ(α + d + (k − 1)s)
=
−
,
(A.3)
d Γ(α + d)
Γ(α + ks)
Γ(α + (k − 1)s)
When d > 0, the sum from Eq. (4.13) is telescoping, so canceling terms,
γ Γ(α + 1) Γ(α + d + Ks) Γ(α + d + (K + N )s)
BN,K ≤
−
d Γ(α + d)
Γ(α + Ks)
Γ(α + (K + N )s)
Γ(α + 1) d−1
∼ γN sd
K
K → ∞,
Γ(α + d)
32
T. CAMPBELL, J. H. HUGGINS, J. P. HOW, AND T. BRODERICK
where the asymptotic result follows from Lemmas B.4, B.7 and B.9. To analyze
the case where d = 0, we can use L’Hospital’s rule to take the limit of Eq. (A.3) as
d → 0, yielding
Z
lim π(θ)k−1 (1 − π(θ))ν(dθ) = γα(ψ(α + ks) − ψ(α + (k − 1)s)).
d→0
Again computing the error bound by canceling terms in the telescoping sum,
BN,K ≤ γα (ψ(α + Ks) − ψ(α + (K + N )s)) ∼ γαN K −1
K → ∞,
where the asymptotic result follows from an application of Lemma B.4.
Stochastic mapping. We can transform the gamma process ΓP(γ, λ, 0) into the
beta process BP(γ, α, 0) by applying the stochastic mapping
θ 7→ θ/(θ + G),
G ∼ Gam(α, α),
κ(θ, du) = Gam(θ/(θ + u); α, α)
θ
du.
u2
Using Lemma 3.4 yields κ(Θ) ∼ BP(γ, α, 0). Applying this result to the Bondesson
representation for ΓP(γ, α, 0) yields
∞
X
θk δψk ∼ BP(γ, α, 0),
with
θk := (1 + Gk Vk−1 eΓk,αγ )−1 ,
i.i.d.
ψk ∼ G,
k=1
i.i.d.
Gk ∼ Gam(α, α),
i.i.d.
Vk ∼ Exp(α).
which, unlike the Bondesson representation, applies for all α > 0.
Hyperpriors. Consider truncating the Bondesson representation of the beta process, but with a hyperprior on the mass parameter γ. A standard choice of hyperprior for γ is a gamma distribution, i.e. γ ∼ Gam(a, b). Combining Proposition 4.6
and the beta-Bernoulli truncation bound in Eq. (A.2), we have that
BN,K
a
≤N
b
ξ
ξ+1
K
.
A.2. Beta prime process. The beta prime process BPP(γ, α, d) (Broderick et al.,
2017) with discount parameter d ∈ [0, 1), concentration parameter α > −d, and
mass parameter γ > 0, is a CRM with rate measure
ν(dθ) = γ
Γ(α + 1)
θ−1−d (1 + θ)−α dθ.
Γ(1 − d)Γ(α + d)
The beta prime process is often paired with an odds Bernoulli likelihood,
h(x | θ) = 1 [x ≤ 1] θx (1 + θ)−1 ,
in which case π(θ) = (1 + θ)−1 . All truncation results are for the odds Bernoulli
likelihood.
ILRep
BRep
TRep
RRep
SBRep
DBRep
PLRep
102
101
103
Mean Computational Cost
104
100
101
102
103
104
105
L1 Error
100
101
102
103
104
105
L1 Error
L1 Error
TRUNCATED RANDOM MEASURES
102
101
103
Mean Computational Cost
d = 0.0
104
100
101
102
103
104
105
33
101
102
103
Mean Computational Cost
d = 0.1
104
d = 0.5
Figure 3. Truncation error bounds for the beta prime-odds
Bernoulli process.
Bondesson representation. If d = 0, then θν(θ) = γα(1 + θ)−α is nonincreasing and cν = limθ→0 θν(θ) = γα, so gν (v) = α(1 + v)−α−1 = Beta0 (v; 1, α).
Thus, it follows from Theorem 3.1 that if Θ ← B-Rep(γα, Beta0 (1, α)), then
Θ ∼ BPP(γ, α, 0). For the truncation bound we have
Z ∞
h
i
BN,K = N
1 − E π(ve−GK /(γα) ) ν(dv)
0
Z ∞
= N γα E e−GK /(γα)
(1 + v)−α (1 + ve−GK /(γα) )−1 dv
Z 0∞
−GK /(γα)
≤ N γα E[e
]
(1 + v)−α (1 + vE[e−GK /(γα) ])−1 dv
0
K Z ∞
γα
(1 + v)−α−1 dv
≤ N γα
1 + γα
0
K
γα
= Nγ
,
1 + γα
where the first upper bound follows from Jensen’s inequality. Thus, the error bound
is the same as for the beta-Bernoulli process.
Thinning representation. If we let g = Beta0 (1 − d, α + d), then the thinning
representation for BPP(γ, α, d) is
Θ=
∞
X
Vk 1(Vk (Γk − γ) ≤ γ)δψk ,
with
i.i.d.
Vk ∼ Beta0 (1 − d, α + d).
k=1
Rejection representation. For d = 0 and α ≥ 1, we take µ to be the rate
measure for LomP(γ, 1), so the rejection representation is
Θ=
∞
X
Vk 1(Uk ≤ (1 + Vk )α−1 ) ∼ BPP(γ, α, 0),
α ≥ 1,
k=1
with
Vk = (eγ
−1
α−1 Γk
− 1)−1 ,
i.i.d.
Uk ∼ Unif[0, 1].
The expected number of rejections is cγ + ψ(α), where cγ is the Euler-Mascheroni
constant and ψ is the digamma function. Since ψ(α) ∼ log(α) for α → ∞, the representation remains efficient even for fairly large values of α. To obtain a truncation
34
T. CAMPBELL, J. H. HUGGINS, J. P. HOW, AND T. BRODERICK
bound, we use the same approach as in Example 4.3:
Z ∞
BN,K
=
FK (γα log(1 + x−1 ))(1 + x)−α−1 dx
N γα
0
Z a
Z
≤
(1 + x)−α−1 dx + FK (γα log(1 + a−1 ))
0
∞
(1 + x)−α−1 dx
a
≤ a + α−1 (3γα log(1 + a−1 )/K)K (1 + a)−α
≤ e−b + α−1 (3γα/K)K bK+α ,
where b := log(1 + a−1 ). Setting the two terms equal and solving for b, we conclude
that
BN,K ≤ 2N γαe−(K+α)W0 ({K+α}
∼ 2N γαe−(K+α)W0 ({3αγ}
−1
−1
{α/γ}1/(K+α) {3αγ/K}−K/(K+α) )
)
K → ∞,
where W0 is the product log function, as defined in Eq. (4.6).
Similarly to Example 3.5, for the case of d > 0 and α ≥ 0, we instead use µ(dθ) =
Γ(α+1)
Γ(α+1)
θ−1−d dθ. Since µ← (u) = (γ 0 u−1 )1/d , where γ 0 := γ dΓ(1−d)Γ(α+d)
, the
γ Γ(1−d)Γ(α+d)
rejection representation is
Θ=
∞
X
Vk 1(Uk ≤ (1 + Vk )−α )δψk ,
with
1/d
Vk = (γ 0 Γ−1
,
k )
k=1
i.i.d.
Uk ∼ Unif[0, 1].
The expected number of rejections is γα
d , so the representation is efficient for large
d, but extremely inefficient when d is small. We have
Z ∞
Γ(α + 1)
BN,K = N γ
FK (γ 0 x−d )x−d (1 + x)−1−α dx.
Γ(1 − d)Γ(α + d) 0
Following the approach of Example 4.4, the integral can be upper bounded as
Z a
Z ∞
x−d dx + FK (γ 0 a−d )
x−d (1 + x)−1−α dx
0
−1 1−d
≤ (1 − d)
a
+α
a
−1 −d
a
(3γ 0 K −1 a−d )K .
Setting the two terms equal and solving for a, we obtain
K
(d+1)K+1
− dK+1
dK+1
1
dK
γΓ(α + 1)
dK − dK+1
(α(1 − d) )
BN,K ≤ 2N
Γ(1 − d)Γ(α + d)
3
1+1/d
−1/d
γΓ(α + 1)
dK
∼ 2N
K → ∞.
Γ(2 − d)Γ(α + d)
3(1 − d)
Decoupled Bondesson representation. It follows from Theorem 3.3 and
the same arguments as those in the Bondesson case that if Θ ←
DB-Rep(γα, Beta0 (1, α), ξ), then Θ ∼ BPP(γ, α, 0). Using the trivial bound
θ/(1 + θ) ≤ θ and calculations analogous to those in the beta-Bernoulli case, for
α > 1 we obtain the upper bound
(
K )
K
ξ
γα
ξ
γα
∼N
K → ∞.
BN,K ≤ 1 − exp −N
α−1 1+ξ
α−1 1+ξ
TRUNCATED RANDOM MEASURES
35
Power-law representation. We can transform the gamma process ΓP(γ, 1, d)
into the beta prime process BPP(γ, α, d) by applying the stochastic mapping
θ
du.
u2
Using Lemma 3.4 yields κ(Θ) ∼ BPP(γ, α, d). Applying this result to the power-law
representation ΓP(γ, 1, d) from Example 3.8 yields the novel power-law representation
θ 7→ u = θ/G,
G ∼ Gam(α + d, 1),
κ(θ, du) = Gam(θ/u; α + d, 1)
Θ ← PL-Rep(γα, 1, d, Beta0 (1, α + d))
implies
Θ ∼ BPP(γ, α, d).
Using the trivial bound θ/(1 + θ) ≤ θ and calculations analogous to those in
the beta-Bernouli case, for α > 1 we obtain the upper bound and asymptotic
simplification
K
γα Y 1 + kd
α−1
2 + kd − d
k=1
( −K
2
γN α
2
Γ( d
∼
) 1−d−1
K
α−1
Γ( 1+d
d )
BN,K ≤ N
d=0
0<d<1
K → ∞.
Size-biased representation. We have
Z
Γ(α + 1)Γ(d + k + α − 1)
,
ηk1 = ηk = π(θ)k−1 (1 − π(θ))ν(dθ) = γ
Γ(α + d)Γ(k + α)
which is the same as for the beta-Bernoulli process. Thus, the error bound is also
the same as the beta-Bernoulli case.
Appendix B. Technical lemmas
Lemma B.1. If Yk is a uniformly bounded, non-negative sequence of random variP
ables such that limk→∞ E[Yk ] = 0, then Yk → 0.
Proof. Without loss of generality we assume that Yk ∈ [0, 1] a.s. Then for all
, δ > 0, by hypothesis there exists k 0 such that for all k ≥ k 0 , E[Yk ] ≤ δ. It then
follows from Markov’s inequality that for all k ≥ k 0 , P(Yk > ) ≤ δ.
Lemma B.2. If µ is a non-atomic measure on Rd , then for any x ∈ Rd and δ > 0,
there exists x,δ > 0 such that µ({y ∈ Rd | kx − yk2 ≤ x,δ }) ≤ δ.
Proof. Without loss of generality let x = 0. Suppose the implication does not
hold. Then there exists δ > 0 such that for all > 0, µ(B ) > δ, where
B := {y ∈ Rd | kyk2 ≤ }. Let n be a sequence such that n → 0 as n → ∞.
Then by continuity µ({0}) = limn→∞ µ(Bn ) > δ, hence µ is a atomic, which is a
contradiction.
Lemma B.3. If ν(dθ), an absolutely continuous σ-finite measure on R+ , and continuous φ : R+ → [0, 1] satisfy
Z
Z
ν(R+ ) = ∞,
min(1, θ)ν(dθ) < ∞,
and
φ(θ)ν(dθ) < ∞,
then
lim φ(θ) = φ(0) = 0.
θ→0
36
T. CAMPBELL, J. H. HUGGINS, J. P. HOW, AND T. BRODERICK
R
R1
R∞
Proof. min(1, θ)ν(dθ) < ∞ implies that 0 θν(dθ) < ∞ and 1 ν(dθ) < ∞. But
R1
ν(R+ ) = ∞, so 0 ν(dθ) = ∞. In fact, for all > 0, ν([0, ]) = ∞ since otherwise
R1
R1
ν(dθ) = ∞ =⇒ θν(dθ) = ∞, a contradiction. Since φ is continuous and has
bounded range, limθ→0 φ(θ) = c exists and is finite. Assume c > 0, so ∃ > 0 such
that ∀0 < , |φ(0 ) − c| < c/2, and in particular φ(0 ) > c/2. Thus, for any θ0 < ,
R θ0
R θ0
φ(θ)ν(dθ) ≥ c/2 0 ν(dθ) = ∞, a contradiction. Thus, c = 0.
0
Lemma B.4. Assume φ(x) is a twice continuously differentiable function with the
following properties:
(1) φ00 (x)/φ0 (x) → 0 as x → ∞
(2) for all δ > 0 there exists Bδ > 0 such that for any increasing sequence
(xn )∞
n=1 ,
lim sup
n→∞
φ00 (y)
= Bδ < ∞.
00
y∈[xn ,xn +δ] φ (xn )
sup
Then for any constant c > 0 and any increasing sequence (xn )∞
n=1 ,
φ(xn + c) − φ(xn ) ∼ cφ0 (xn )
for n → ∞.
Proof. A second-order Taylor expansion of φ(xn + c) about xn yields
φ(xn + c) − φ(xn ) = cφ0 (xn ) +
c2 00 ∗
φ (xn ),
2
where x∗n ∈ [xn , xn + c]. Our assumptions on φ ensure that,
φ00 (x∗n )
Bc φ00 (xn )
≤ lim
=0
0
n→∞ φ (xn )
n→∞ φ0 (xn )
lim
and hence
φ(xn + c) − φ(xn )
c φ00 (x∗n )
= lim 1 +
= 1.
0
n→∞
n→∞
cφ (xn )
2 φ0 (xn )
lim
Lemma B.5 (Gautschi (1959)).
(1 + x)d−1 ≤
Γ(x + d)
≤ xd−1
Γ(x + 1)
0 ≤ d ≤ 1, x ≥ 1,
and thus for 0 ≤ d ≤ 1,
Γ(x + d)
∼ xd−1
Γ(x + 1)
Lemma B.6. For α > 0 and x ≥ −1,
(
M
Γ(α+M +x+1)
1
X
Γ(α + m + x)
−
1+x
Γ(α+M )
=
Γ(α
+
m)
ψ(α
+
M
)
−
ψ(α)
m=1
where ψ(·) is the digamma function.
x → ∞.
Γ(α+x+1)
Γ(α)
x > −1
x = −1
TRUNCATED RANDOM MEASURES
37
Proof. When M = 1 and x > −1, analyzing the right hand side yields
Γ(α + M + x + 1) Γ(α + x + 1)
Γ(α + x + 2) Γ(α + x + 1)
−
=
−
Γ(α + M )
Γ(α)
Γ(α + 1)
Γ(α)
Γ(α + x + 1) α + x + 1
=
−1
Γ(α)
α
Γ(α + x + 1)
= (x + 1)
.
Γ(α + 1)
By induction, supposing that the result is true for M − 1 ≥ 1 and x > −1,
M
M
−1
X
X
Γ(α + m + x)
Γ(α + m + x) Γ(α + M + x)
=
+
Γ(α
+
m)
Γ(α + m)
Γ(α + M )
m=1
m=1
Γ(α + M + x) Γ(α + x + 1)
1
Γ(α + M + x)
−
=
+
1 + x Γ(α + M − 1)
Γ(α)
Γ(α + M )
Γ(α + M + x)
α+M +x
Γ(α + x + 1)
=
−
Γ(α + M − 1) (1 + x)(α + M − 1)
(1 + x)Γ(α)
Γ(α + M + x + 1) Γ(α + x + 1)
1
=
−
.
1+x
Γ(α + M )
Γ(α)
This demonstrates the desired result for x > −1. Next, when x = −1, we have that
M
M
X
X
Γ(α + m − 1)
1
=
.
Γ(α
+
m)
α
+
m
−1
m=1
m=1
We proceed by induction once again. For M = 1, using the recurrence relation
ψ(x + 1) = ψ(x) + x−1 (Abramowitz and Stegun, 1964, Chapter 6), the right hand
side evaluates to
ψ(α + 1) − ψ(α) = ψ(α) + α−1 − ψ(α) = α−1 .
Supposing that the result is true for M − 1 ≥ 1 and x = −1,
M
−1
X
1
1
1
=
+
α+m−1
α+m−1 α+M −1
m=1
m=1
M
X
= ψ(α + M − 1) − ψ(α) +
1
α+M −1
= ψ(α + M ) − ψ(α),
demonstrating the result for x = −1.
Lemma B.7. For a > 0, d ∈ R, and xn → ∞,
d Γ(a + xn + d)
∼ dxd−1
n .
dxn Γ(a + xn )
Proof. We have
d Γ(a + x + d)
Γ(a + x + d)
=
(ψ(a + x + d) − ψ(a + x)),
dx Γ(a + x)
Γ(a + x)
38
T. CAMPBELL, J. H. HUGGINS, J. P. HOW, AND T. BRODERICK
where ψ is the digamma function. Using Lemma B.4 and the asymptotic expansion
of ψ 0 (Abramowitz and Stegun, 1964, Chapter 6), we obtain
ψ(a + xn + d) − ψ(a + xn ) ∼ dψ 0 (a + xn ) ∼
d
∼ dx−1
n .
xn + a
Since
Γ(a + xn + d)
∼ (a + xn )d ∼ xdn ,
Γ(a + xn )
using Lemma B.9(2) with the previous two displays yields the result.
Lemma B.8. For 0 ≤ d < 1,
(
Z ∞
d
d
Γ(−d)
−1−d −λθ
+ t) − λ
(λ
−tθ
e
−1 θ
e
dθ =
λ
log t+λ
0
0<d<1
d=0
.
Proof. By integration by parts and the standard gamma integral,
Z ∞
Z ∞h
i θ−d
e−tθ − 1 θ−1−d e−λθ dθ =
(λ + t)e−(λ+t)θ − λe−λθ
dθ
−d
0
0
= Γ(−d) (λ + t)d − λd .
Taking the limit as d → 0 via L’Hospital’s rule yields
λ
lim Γ(−d) (λ + t)d − λd = log
.
d→0
λ+t
Lemma B.9 (Standard asymptotic equivalence properties).
(1) If an ∼ bn and bn ∼ cn , then an ∼ cn .
(2) If an ∼ bn and cn ∼ dn , then an cn ∼ bn dn .
(3) If an ∼ bn , cn ∼ dn and an cn > 0, then an + cn ∼ bn + dn .
Appendix C. Proofs of sequential representation results
C.1. Correctness of B-Rep, DB-Rep, and PL-Rep.
Proof of Theorem 3.1. First, we show that gν (v) is a density. Since vν(v) is nond
d
decreasing, dv
[vν(v)] exists almost everywhere, dv
[vν(v)] ≤ 0 and hence gν (v) ≥ 0.
Furthermore,
Z ∞
Z ∞
∞
d
−1
gν (v)dv = −cν
[vν(v)]dv = −c−1
= 1,
ν vν(v)
dv
v=0
0
0
where the final equality follows from the assumed behavior of vν(v) at 0 and ∞.
Since for a partition A1 , . . . , An , the random variables Θ(A1 ), . . . , Θ(An ) are independent, it suffices to show that for any measurable set A (with complement Ā),
the random variable Θ(A) has the correct characteristic function. Define the family
of random measures
∞
X
Θt =
Vk e−(Γk +t)/cν δψk ,
t ≥ 0,
k=1
so Θ0 = Θ. Conditioning on Γ11 ,
D
(Θt (A) | Γ1 = u) = V1 e−(u+t)/cν 1[ψ1 ∈ A] + Θt+u (A),
TRUNCATED RANDOM MEASURES
39
and note that the two terms on the left hand side are independent. We can thus
write the characteristic function of Θt (A) as
ϕ(ξ, t, A) := E[eiξΘt (A) ]
= E[E[eiξΘt (A) | Γ1 = u]]
= E[E[eiξV1 e
−(u+t)/cν
1[ψ1 ∈A] iξΘt+u (A)
e
| Γ1 = u]]
iξV1 e−(u+t)/cν
= E[E[(G(A)e
+ G(Ā))ϕ(ξ, t + u, A) | Γ1 = u]]
Z ∞Z ∞
−(u+t)/cν
eiξve
ϕ(ξ, t + u, A)gν (v)e−u du dv
= G(A)
0
0
Z ∞
ϕ(ξ, t + u, A)e−u du,
+ G(Ā)
0
where Ā is the complement of A. Multiplying both sides by e−t and making the
change of variable w = u + t yields
e−t ϕ(ξ, t, A) = G(A)
Z
∞
t
Z
∞
eiξve
−w/cν
ϕ(ξ, w, A)gν (v)e−w dv dw
0
∞
Z
+ G(Ā)
ϕ(ξ, w, A)e−w dw
t
Z ∞
= G(A)
ϕgν (ξe−w/cν )ϕ(ξ, w, A)e−w dw
t
Z ∞
+ G(Ā)
ϕ(ξ, w, A)e−w dw,
t
R∞
where ϕgν (a) := 0 eiav gν (v) dv is the characteristic function of a random variable
with density gν . Differentiating both sides with respect to t and rearranging yields
∂ϕ(ξ, t, A)
= ϕ(ξ, t, A) − G(A)ϕgν (ξe−t/cν )ϕ(ξ, t, A) − (1 − G(A))ϕ(ξ, t, A)
∂t
= ϕ(ξ, t, A)G(A)(1 − ϕgν (ξe−t/cν )),
so we conclude that
Z
ϕ(ξ, t, A) = exp −G(A)
∞
(1 − ϕgν (ξe
−u/cν
)) du .
t
Using integration by parts and the definition of gν , rewrite
Z
∞
d
[vν(v)]eiav dv
dv
0
Z ∞
∞
iav
−1
= −c−1
vν(v)e
+
c
iavν(v)eiav dv
ν
ν
v=0
0
Z ∞
iav
=1+
ν(v)eiav dv,
cν
0
ϕgν (a) = −c−1
ν
40
T. CAMPBELL, J. H. HUGGINS, J. P. HOW, AND T. BRODERICK
where the final equality follows from the assumed behavior of vν(v) at 0 and ∞.
Combining the previous two displays and setting t = 0 concludes the proof:
Z ∞Z ∞
iξv −u/cν iξve−u/cν
ϕ(ξ, 0, A) = exp −G(A)
e
ν(v) dv du
e
cν
0
0
Z ∞Z ∞
∂ h iξve−u/cν i
= exp −G(A)
−e
ν(v) du dv
∂u
0
0
Z ∞
= exp −G(A)
(eiξv − 1)ν(v) dv .
0
Proof of Theorem 3.3. It was already shown that gν (v) is a density. Let Θ0k =
PCk
P∞
cν 0
0
0
i=1 θki δψki , so Θ =
k=1 Θk . Each Θk is a CRM with rate measure ξ νk (dθ),
0
where νk (dθ) is the law of θki . Using the product distribution formula we have
νk0 (dθ)
1
Z
=
0
ξk
(− log w)k−1 wξ−2 gν (θ/w) dw dθ.
Γ(k)
(C.1)
Rv
Let Gν (v) = 0 gν (x) dx be the cdf derived from gν . From the preceding arguments,
conclude that the rate measure of Θ is
∞
cν X 0
νk (dθ)
ξ
k=1
Z 1X
∞
ξ k−1
(− log w)k−1 wξ−2 gν (θ/w) dw dθ
= cν
Γ(k)
0 k=1
Z 1
= cν
ξw−2 gν (θ/w) dw dθ
ν 0 (dθ) :=
0
Z
= cν
0
1
ξθ−1
∂
[−Gν (θ/w)] dw dθ
∂w
= −cν θ−1 Gν (θ/w)
1
dθ
w=0
= cν θ−1 (1 − Gν (θ)) dθ.
The cdf can be rewritten as
Z θ
d
−1
1 − Gν (θ) = 1 + cν
[xν(x)] dx = 1 + c−1
ν xν(x)
dx
0
θ
x=0
= c−1
ν θν(θ).
Combining the previous two displays, conclude that ν 0 (dθ) = ν(θ) dθ.
Proof of Theorem 3.5. Since the power-law representation in Eq. (3.3) for the case
when Vki = 1 almost surely was previously shown to be BP(γ, α, d) (Broderick
et al., 2012), we simply apply the stochastic mapping result in Lemma 3.4 with
κ(θ, du) = θ−1 gν (uθ−1 )du, where θ−1 gν (uθ−1 ) is the density of U = V θ | θ under
V ∼ gν .
TRUNCATED RANDOM MEASURES
41
C.2. Proof of the expected number rejections of the R-Rep.
Proof of Proposition 3.2. We have
"∞
#
"∞
#
X
X
dν
E
1(θk = 0) = E
1
(Vk ) ≥ Uk
dµ
k=1
k=1
"∞
#
X
dν
=E
1−
(Vk )
dµ
k=1
Z
dν
=
(x) µ(dx),
1−
dµ
where the equalities follow from the definition of θk , integrating out Uk , and applying Campbell’s theorem.
C.3. Power-law behavior of the PL-Rep. We now formalize the sense in which
i.i.d.
power-law representations do in fact produce power-law behavior. Let Zn | Θ ∼
PN
LP(Poiss, Θ) and yk :=
n=1 1[znk ≥ 1]. We analyze the number of non-zero
features after N observations,
KN :=
∞
X
1[yk ≥ 1],
k=1
and the number of features appearing j > 1 times after N observations,
KN,j :=
∞
X
1[yk = j].
k=1
In their power law analysis of the beta process, Broderick et al. (2012) use a
Bernoulli likelihood process. However, the Bernoulli process is only applicable if
θk ∈ [0, 1], whereas in general θk ∈ R+ . Replacing the Bernoulli process with a
Poisson likelihood process is a natural choice since 1[znk ≥ 1] ∼ Bern(1−e−θk ), and
asymptotically 1 − e−θk ∼ θk a.s. for k → ∞ since limk→∞ θk = 0 a.s. Thus, the
Bernoulli and Poisson likelihood processes behave the same asymptotically, which
is what is relevant to our asymptotic analysis. We are therefore able to show that
all CRMs with power-law representations, not just the beta process, have what
Broderick et al. (2012) call Types I and II power law behavior. Our only condition
is that the tails of g are not too heavy.
Theorem C.1. Assume that g is a continuous density such that for some > 0,
g(x) = O(x−1−d− ).
(C.2)
Then for Θ ← PL-Rep(γ, α, d, g) with d > 0, there exists a constant C depending
on γ, α, d, and g such that, almost surely,
KN ∼ Γ(1 − d)CN d ,
KN,j
d Γ(j − d)
CN d ,
∼
j!
N →∞
N →∞
(j > 1).
In order to prove Theorem C.1, we require a number of additional definitions
and lemmas. Our approach follows that in Broderick et al. (2012), which the
reader is encouraged to consult for more details and further discussion of power law
42
T. CAMPBELL, J. H. HUGGINS, J. P. HOW, AND T. BRODERICK
behavior of CRMs. Throughout this section, Θ ← PL-Rep(γ, α, d, g) with d > 0.
By Lemma 3.4, Θ ∼ CRM(ν), where
Z
ν(dθ) := g(θ/u)u−1 νBP (du) dθ
and νBP (dθ) is the rate measure for BP(γ, α, d). Let Πk be a homogeneous Poisson
point process on R+ with rate θk and define
K(t) :=
Kj (t) :=
∞
X
k=1
∞
X
1[|Πk ∩ [0, t]| > 0]
1[|Πk ∩ [0, t]| = j].
k=1
Furthermore, for N ∈ N, let
ΦN := E[KN ]
and
ΦN,j := E[KN,j ]
and
Φj (t) := E[Kj (t)]
(j > 1)
and for t > 0, let
Φ(t) := E[K(t)]
(j > 1).
If follows from Campbell’s Theorem (Kingman, 1993) that
"
# Z
X
−tθk
Φ(t) = E
(1 − e
) = (1 − e−tθ )ν(dθ)
k
"
ΦN
#
X
=E
(1 − e−N θk ) = Φ(N )
k
#
"X j
Z
tj
N
t (1 − e−θk )j −tθk
e
=
(1 − e−θ )j e−tθ ν(dθ)
Φj (t) =
E
j!
j!
j
k
# Z
"X
N
N
−θk j −(N −j)θk
ΦN,j =
E
(1 − e ) e
=
(1 − e−θ )j e−(N −j)θ ν(dθ).
j
j
k
The first lemma characterizes the power law behavior of Φ(t) and Φj (t). A slowly
varying function ` satisfies `(ax)/`(x) → 1 as x → ∞ for all a > 0.
Lemma C.2 (Broderick et al. (2012), Proposition 6.1). If for some d ∈ (0, 1),
C > 0, and slowly varying function `,
Z x
d
C`(1/x)x1−d ,
x → 0,
(C.3)
ν̄[0, x] :=
θν(dθ) ∼
1−d
0
then
Φ(t) ∼ Γ(1 − d)Ctd ,
Φj (t) ∼
d Γ(j − d) d
Ct ,
j!
t→∞
t→∞
(j > 1).
Transferring the power law behavior from Φ(t) to ΦN is trivial since Φ(N ) = ΦN .
The next lemma justifies transferring the power law behavior from Φj (t) to ΦN,j .
TRUNCATED RANDOM MEASURES
43
Lemma C.3 (Broderick et al. (2012), Lemmas 6.2 and 6.3). If ν satisfies Eq. (2.1),
then
K(t) ↑ ∞ a.s.,
Φ(t) ↑ ∞,
Φ(t)/t ↓ 0.
Furthermore,
|ΦN,j − Φj (N )| <
Cj
max{Φj (N ), Φj+2 (N )} → 0.
N
The final lemma confirms that the asymptotic behaviors of KN and KN,j is
almost surely the same as the expectations of KN and KN,j .
Lemma C.4. Assume ν satisfies Eq. (2.1) and that for some d ∈ (0, 1), C > 0,
Cj > 0, and slowly varying functions `, `0 , Φ(t) ∼ C`(t)td and Φj (t) ∼ Cj `(t)td .
Then for N → ∞, almost surely
X
X
Kn ∼ ΦN
and
KN,i ∼
ΦN,i .
i<j
i<j
Proof of Theorem C.1. Combining the three lemmas, the result follows as soon as
we show that ν(dθ) satisfies Eq. (C.3). C will be a constant that may change from
line to line. We begin by rewriting ν(dθ) using the change of variable w = θ(u−1 −1):
Z 1
ν(dθ) = C
g(θ/u)u−2−d (1 − u)α+d−1 du dθ
0
Z ∞
wα+d−1
−1−d
dw dθ.
= Cθ
g(w + θ)
(w + θ)α−1
0
Since g(x) is integrable and continuous, for x ∈ [0, 1], it is upper-bounded by the
non-integrable function C0 x−1 for some C0 > 0. Combining this upper bound with
Eq. (C.2) yields g(x) ≤ φ(x) := C0 x−1 1[x ≤ 1] + C1 x−1−d− 1[x > 1] for some
C1 > 0, so
g(w + θ)
wα+d−1
≤ φ(w)wd .
(w + θ)α−1
Since φ(w)wd is integrable, by dominated convergence the limit
Z ∞
wα+d−1
L = lim
g(w + θ)
dw
θ→0 0
(w + θ)α−1
exists and is finite. Moreover, since g(x) is a continuous density, there exists M > 0
and 0 < a < b < ∞ such that g(x) ≥ M for all x ∈ [a, b]. Hence, for θ < a,
Z
∞
g(w + θ)
0
wα+d−1
dw ≥ M
(w + θ)α−1
Z
b−θ
a−θ
wα+d−1
dw > 0,
(w + θ)α−1
so L > 0. Thus,
Z
ψ(θ) := θ
0
1
g(θ/u)u−2−d (1 − u)α+d−1 du → Cθ−d ,
θ→0
44
T. CAMPBELL, J. H. HUGGINS, J. P. HOW, AND T. BRODERICK
and hence for δ > 0 and θ sufficiently small, |ψ(θ) − Cθ−d | < δ. Thus, for x
sufficiently small,
Z x
Z x
Z x
ψ(θ) dθ ≤
Cθ−d dθ +
|ψ(θ) − Cθ−d | dθ
0
0
0
C x1−d
≤
+ δx
1−d
C x1−d
,
x → 0,
∼
1−d
which shows that Eq. (C.3) holds.
Appendix D. Proofs of CRM truncation bounds
D.1. Protobound.
Lemma D.1 (Protobound). Let Θ and Θ0 be two discrete random measures. Let
X1:N be a collection of random measures generated i.i.d. from Θ with supp(Xn ) ⊆
supp(Θ), and let Y1:N be a collection of random variables where Yn is generated from
Xn via Yn | Xn ∼ f (· | Xn ). Define Z1:N and W1:N analogously for Θ0 . Finally,
D
define Q := 1 [supp(X1:N ) ⊆ supp(Θ0 )]. If (X1:N |Θ, Θ0 , Q = 1) = (Z1:N |Θ0 , Θ)
almost surely under the joint distribution of Θ, Θ0 , then
1
kpY − pW k1 ≤ 1 − P(Q = 1),
2
where pY , pW are the marginal densities of Y1:N and W1:N .
Proof of Lemma 4.1. This is the direct application of Lemma D.1 to CRMs, where
Θ ∼ CRM(ν), and Θ0 is a truncation Θ0 = ΘK . The technical condition is satisfied
because the weights in X1:N are sampled independently for each atom in Θ.
Proof of Lemma 5.1. This is the direct application of Lemma D.1 to NCRMs, where
Θ is the normalization of a CRM with distribution CRM(ν), and Θ0 is the normalization of its truncation. The technical condition is satisfied because the conditioning on X1:N ⊆ supp(Θ0 ) is equivalent to normalization of Θ0 .
Proof of Lemma D.1. We begin by expanding the 1-norm and conditioning on both
Θ and Θ0 (denoted by conditioning on Θ̃ := (Θ, Θ0 ) for brevity):
"N
#
"N
#
Z
Y
Y
kpY − pW k1 =
E
f (yn |Zn ) − E
f (yn |Xn ) dy
n=1
Z
=
" "
E E
n=1
N
Y
#
f (yn |Zn )|Θ̃ − E
n=1
"
N
Y
##
f (yn |Xn )|Θ̃
dy.
n=1
Then conditioning on Q,
"N
#
" "N
# #
Y
Y
E
f (yn |Xn )|Θ̃ = E E
f (yn |Xn )|Θ̃, Q |Θ̃
n=1
n=1
"
a.s.
= P(Q = 1|Θ̃)E
N
Y
n=1
#
f (yn |Zn )|Θ̃ + P(Q = 0|Θ̃)E
"
N
Y
n=1
#
f (yn |Xn )|Θ̃, Q = 0 ,
TRUNCATED RANDOM MEASURES
45
where the first term arises from the fact that for any function φ,
h
i
h
i
a.s.
E φ(Z1:N )|Θ̃ = E φ(X1:N )|Θ̃, Q = 1 ,
because X1:N |Θ̃, Q = 1 is equal in distribution to Z1:N |Θ̃ a.s. by assumption. Substituting this back in above,
kpY − pW k1
(D.1)
"
"N
#
"N
#!#
Z
Y
Y
dy
E P(Q = 0|Θ̃) E
f (yn |Zn )|Θ̃ − E
f (yn |Xn )|Θ̃, Q = 0
=
n=1
"
Z
≤
"
E P(Q = 0|Θ̃) E
n=1
#
N
Y
"
f (yn |Zn )|Θ̃ − E
n=1
"
Z
≤
"
E P(Q = 0|Θ̃) E
N
Y
##
f (yn |Xn )|Θ̃, Q = 0
dy
n=1
#
N
Y
"
f (yn |Zn )|Θ̃ + E
n=1
#!#
N
Y
f (yn |Xn )|Θ̃, Q = 0
dy,
n=1
and finally by Fubini’s Theorem,
kpY − pW k1
"
"Z
≤ E P(Q = 0|Θ̃) E
N
Y
#
"Z
f (yn |Zn )dy|Θ̃ + E
n=1
N
Y
#!#
f (yn |Xn )dy|Θ̃, Q = 0
n=1
h
h
i
h
ii
= E P(Q = 0|Θ̃) E 1|Θ̃ + E 1|Θ̃, Q = 0
= 2 P(Q = 0) = 2(1 − P(Q = 1)).
Proof of Propositions 4.2 and 5.2. The same proof applies to both results. Since
G is non-atomic and Ψ is Borel, there exists a measurable mapping T : Ψ → R
such that the random variable T (ψ), ψ ∼ G, is non-atomic. For an atomic measure
PK%
% = k=1
wk δψk on Ψ, define
S(%, T ) :=
K%
X
T (ψk ).
k=1
i.i.d.
Let ψi0 ∼ G, TX := #{k | xk > 0∧zk = 0} < ∞, and TZ := #{k | zk > 0∧xk = 0}.
Conditional on Q = 0, there exists n ∈ [N ] such that TXn ≥ 1 since Xn has at least
one atom that Zn lacks. Using these definitions and facts, observe that
E[S(Xn , T ) − S(Zn , T ) | Q = 0]
TXn
TZ n
X
X
= E
T (ψi0 ) −
T (ψj0 ) | Q = 0
i=1
"
=E
j=1
∞
X
P(TXn = tX , TZn = tZ | Q = 0)
tX =1,tZ =0
"
×E
tX
X
i=1
T
(ψi0 )
−
tZ
X
j=1
##
T
(ψj0 ) | TXn
= tX , TZn = tz , Q = 0
.
46
T. CAMPBELL, J. H. HUGGINS, J. P. HOW, AND T. BRODERICK
PtZ
PtX
0
0
Since TXn ≥ 1, V (tX , tZ ) :=
j=1 T (ψj ) is a finite sum of one
i=1 T (ψi ) −
or more non-atomic random variables, so V (tX , tZ ) is itself non-atomic. It then
follows that ∆ := S(X, T ) − S(Z, T ) is non-atomic. Thus, by Lemma B.2, for any
δ, there exists δ > 0 such that P(|∆| < δ ) < δ. Define the family of likelihoods
fδ (· | %) = Unif[S(%, T ) − δ /2, S(%, T ) + δ /2]. Let f = fδ . Then, conditioned on
Q = 0, with probability at least 1 − δ, f (y | Xn ) > 0 =⇒ f (y | Zn ) = 0 and
f (y | Zn ) > 0 =⇒ f (y | Xn ) = 0, which implies that both inequalities in Eq. (D.1)
are equalities. Hence, we conclude that kpY − pW k1 ≥ 2(1 − δ)P(Q = 0).
D.2. Series representation truncation. Recall from Section 4.1 that a series
representation generally has the form
Θ=
∞
X
θk δψk
θk = τ (Vk , Γk )
i.i.d.
Vk ∼ g,
k=1
Pk
i.i.d.
where Γk = `=1 E` , E` ∼ Exp(1), are the jumps in a unit-rate homogeneous
Poisson point process, τ : R+ × R+ → R+ is a measurable function, g is a distribution on R+ , and limu→∞ τ (v, u) = 0 for g-almost every v. Note that by Lemma B.3
π̄(0) = 0, where π̄(x) := 1 − π(x). This fact will repeatedly prove useful for the
proofs in this section.
The proof of Theorem 4.3 is based on the following lemma.
Lemma D.2. Under the same hypotheses as Theorem 4.3,
P(supp(X1:N ) ⊆ supp(ΘK ))
Z ∞
Z ∞
N
= E exp −
1−
π(τ (v, u + GK )) g(dv) du .
0
(D.2)
0
Proof. Let
"
p(t, K) := E
∞
Y
#
π (τ (Vk , Γk + t))
N
,
k=K+1
so p(0, K) = P(supp(X1:N ) ⊆ supp(ΘK )). We use the proof strategy from Banjevic
et al. (2002) and induction in K. For K = 0,
!!N
∞
Y
X
N
p(t, 0) = E E π (τ (V1 , u + t))
π τ Vk ,
Ej + u + t
Γ1 = u
k=2
Z
=
∞
Z
1<j≤k
N
π (τ (v, u + t)) p(u + t, 0)e−u g(dv) du
0
since the Vk are i.i.d. Multiplying both sides by e−t and making the change of
variable w = u + t yields
Z ∞Z
N
−t
e p(t, 0) =
π (τ (v, w)) p(w, 0)e−w g(dv) dw.
t
Differentiating both sides with respect to t and rearranging yields
Z
∂p(t, 0)
N
= p(t, 0) 1 − π (τ (v, t)) g(dv) .
∂t
(D.3)
TRUNCATED RANDOM MEASURES
47
Since limu→∞ τ (v, u) = 0 and π(0) = 1 by Lemma B.3, we can solve Eq. (D.3) and
conclude that
Z ∞
Z
N
p(t, 0) = exp −
1 − π (τ (v, u)) g(dv) du
Zt ∞
Z
N
1 − π (τ (v, u + t)) g(dv) du .
= exp −
0
We use the inductive hypothesis that
GK ∼ Gam(K, 1),
p(t, K) = E[p(t + GK , 0)],
G0 = 0,
which trivially holds for K = 0. If the inductive hypothesis holds for some K ≥ 0,
then using the tower property,
!!N
∞
Y
X
p(t, K + 1) = E E
π τ (Vk ,
E1j + u + t
Γ1 = u
k=K+2
1<j≤k
= E[p(t + E1 , K)],
E1 ∼ Exp(1)
= E[p(t + GK + E1 , 0)]
= E[p(t + GK+1 , 0)].
Eq. (D.2) follows by setting t = 0.
Proof of Theorem 4.3. The main result follows by combining Lemmas 4.1 and D.2,
applying Jensen’s inequality, then using monotone convergence. The upper bound
1 − e−BN,K ≤ 1 follows immediately from the fact that the integral
Z ∞
Z ∞
N
1−E
π(τ (v, u + GK )) g(dv)
du
0
0
is non-negative.
Fix N . It follows from Eq. (2.2) that
lim 1 − P(supp(X1:N ) ⊆ supp(ΘK )) → 0.
K→∞
P
It
from Lemma B.1 that 1 − e−ωN,K → 0, where ωN,K :=
R ∞then Rfollows
∞
1 − 0 π(τ (v, u + GK ))N g(dv) du. By the continuous mapping theorem,
0
P
conclude that ωN,K → 0 as K → ∞ and hence BN,K = E[ωN,K ] → 0 as
K → ∞.
Theorem D.3 (Inverse-Lévy representation truncation error). For Θ
IL-Rep(ν), the conclusions of Theorem 4.3 hold with
Z ∞
BN,K = N
FK (ν[x, ∞))(1 − π(x)) ν(dx).
0
Proof. We have from Theorem 4.3 that
Z ∞
BN,K = N E
(1 − π(ν ← (u + GK )) du .
0
We first make the change of variables x = ν ← (u) to obtain
Z ∞
Z ∞
Z ν ← (GK )
π̄(ν ← (u + GK )) du =
π̄(ν ← (u)) du =
π̄(x)ν(dx)
0
GK
0
←
48
T. CAMPBELL, J. H. HUGGINS, J. P. HOW, AND T. BRODERICK
Finally, use the fact that for all a, b ≥ 0, ν ← (a) ≥ b ⇐⇒ a ≤ ν ([b, ∞)) and
monotone convergence:
"Z ←
#
ν (GK )
π̄(x) ν(dx)
BN,K = N E
0
∞
1[x ≤ ν (GK )]π̄(x) ν(dx)
= NE
Z0 ∞
1[GK ≤ ν[x, ∞)]π̄(x) ν(dx)
= NE
Z ∞0
=N
FK (ν[x, ∞))π̄(x) ν(dx).
Z
←
0
Theorem D.4 (Thinning representation truncation error). For Θ ← T-Rep(ν),
the conclusions of Theorem 4.3 hold with
Z ∞
Z dν
dg (v)
dν
BN,K = N
(1 − π(v))
FK
(v) − u du g(v)dv.
dg
0
0
Proof. We have from Theorem 4.3 that
Z ∞ Z ∞
dν
BN,K = N E
(v) ≥ u + GK
du g(v)dv
π̄ v1
dg
0
0
Since π̄(0) = 0, using monotone convergence we have
Z ∞
Z ∞
dν
(v) ≥ u + GK du g(v)dv
BN,K = N
π̄(v)
E 1
dg
0
0
Z ∞
Z ∞
dν
(v) − u du g(v)dv
=N
π̄(v)
FK
dg
0
0
Z ∞
Z dν
(v)
dg
dν
=N
π̄(v)
(v) − u du g(v)dv
FK
dg
0
0
Z ∞
Z dν
dg (v)
=N
π̄(v)
FK (u) du g(v)dv.
0
0
Theorem D.5 (Rejection representation truncation error). For Θ ← R-Rep(ν),
the conclusions of Theorem 4.3 hold with
Z ∞
BN,K = N
FK (µ[x, ∞))(1 − π(x)) ν(dx).
0
Proof. We have from Theorem 4.3 that
Z ∞ Z 1
dν ←
←
(µ (u + GK )) ≥ v dv du
BN,K = N E
π̄ µ (u + GK )1
dµ
0
0
Since π̄(0) = 0, we can eliminate the innermost integral:
Z 1
dν ←
←
π̄ µ (u + GK )1
(µ (u + GK )) ≥ v dv
dµ
0
dν ←
=
(µ (u + GK ))π̄(µ← (u + GK )).
dµ
TRUNCATED RANDOM MEASURES
49
Making the change of variable x = µ← (u + GK ) and reasoning analogously to the
proof of Theorem D.3, we obtain
"Z ←
#
Z ∞
µ (GK )
dν
BN,K = N E
(x)π̄(x)µ(dx) = N
FK (µ[x, ∞))π̄(x)ν(dx).
dµ
0
0
Theorem D.6 (Bondesson representation truncation error). For Θ ← R-Rep(ν),
the hypotheses of Theorem 4.3 are satisfied and its conclusions hold with
Z ∞
1 − E π(ve−GK ) ν(dv).
BN,K = N
0
Proof. While Theorem D.6 can be proved using Theorem 4.3, we take an alternative
approach using more direct Poisson process arguments.
Lemma D.7. For K ≥ 0, GK ∼ Gam(K, c), and G0 = 0,
Z ∞
−GK N
P(supp(X1:N ) ⊆ supp(ΘK )) = E exp −
1 − π(ve
) ν(dv) .
0
Proof of Lemma D.7. For t ≥ 0, the measure tΘ has distribution tΘ ∼ CRM(νt )
i.i.d.
where νt (dθ) := ν(dθ/t). Further, define X̃n | Θ ∼ LP(h, tΘ), and
" ∞
#
N
Y
−Γk /c
p(t, K) := P(supp(X̃1:N ) ⊆ supp(tΘK )) = E
π tVk e
.
k=K+1
We will prove that
Z
p(t, K) = E exp −
∞
−GK N
1 − π(tve
)
ν(dv)
,
0
and then set t = 1 to obtain the desired result. The proof proceeds by induction.
For K = 0, the event supp(X̃1:N ) ⊆ supp(tΘK ) is equal to supp(X̃1:N ) ⊆ ∅ and
thus supp(X̃1:N ) = ∅. This is in turn equivalent to the probability that after
thinning a CRM(νt ) by π(θ) N times,
the remaining process has no atoms, i.e. the
probability that CRM 1 − π(θ)N νt (dθ) has no atoms.
Since a Poisson process
R
with measure µ(dθ) has no atoms with probability e− µ(dθ) ,
Z ∞
p(t, 0) = exp −
1 − π(θ)N νt (dθ)
0 Z ∞
= E exp −
1 − π(tve−G0 )N ν(dv) .
0
The second equality follows by the change of variables v = θ/t and because G0 = 0
with probability 1. The inductive hypothesis is that for K ≥ 0,
p(t, K) = E p te−GK , 0 ,
GK ∼ Gam(K, c).
50
T. CAMPBELL, J. H. HUGGINS, J. P. HOW, AND T. BRODERICK
Using the tower property to condition on Γ1 /c and the fact that the Vk are i.i.d.,
#
" ∞
N
Y
−Γk /c
p(t, K + 1) = E
π tVk e
k=K+2
" "
=E E
∞
Y
π te
−Γ1 /c
Vk e
−Γk /c
N
##
| Γ1 /c
k=K+1
h
i
h
i
= E p te−Γ1 /c , K = E p te−(Γ1 /c+GK ) , 0 = E p te−GK+1 , 0 ,
since Γ1 /c ∼ Exp(c). The desired result follows by setting t = 1.
First combine Lemmas 4.1 and D.7, then apply Jensen’s inequality. The bounds
on BN,K and fact that limK→∞ BN,K = 0 follows by the same arguments as in the
proof of Theorem 4.3.
D.3. Superposition representation truncation.
Proof of Theorem 4.4. We begin with Lemma 4.1, and note that
P (supp(X1:N ) ⊆ supp(ΘK )) = P supp(X1:N ) ∩ supp(Θ+
K) = ∅ .
After generating X1:N from Θ, we can view the point process representing the atoms
+
N
in Θ+
(i.e. the Bernoulli trial
K not contained in any Xn as ΘK thinned by π(θ)
with success probability 1 − π(θ) to generate an atom failed N times), and thus
N
the remaining process is Θ+
K thinned by 1 − π(θ) . Therefore, the above event is
+
equivalent to the event that ΘK thinned by 1 − π(θ)N has no atoms. Using the
fact
R
that a Poisson process with measure µ(dθ) has no atoms with probability e− µ(dθ) ,
we have the formula for BN,K ,
+
N
P (supp(X1:N ) ⊆ supp(ΘK )) = e− (1−π(θ) )νK (dθ) .
+
R
Since BN,K :=
1 − π(θ)N νK
(dθ) is nonnegative, the error bound lies in the
interval [0, 1]. To show that limK→∞ BN,K = 0, first note that
R
Z
Z
∞ Z
X
1 − π(θ)N ν(dθ) ≤ N (1 − π(θ)) ν(dθ) = N
h(x | θ)ν(dθ) < ∞,
(D.4)
x=1
by Eq. (2.2). Further, splitting ν into its individual summed components, we have
that
Z
Z
K Z
X
+
(1 − π(θ))ν(dθ) =
(1 − π(θ))νk (dθ) + (1 − π(θ))νK
(dθ).
(D.5)
k=1
Combining the results from Eqs. (D.4) and (D.5) yields
Z
+
lim
(1 − π(θ))νK
(dθ) = 0.
K→∞
TRUNCATED RANDOM MEASURES
51
D.4. Stochastic mapping truncation.
Proof of Proposition 4.5. Let π̃(u) = h̃(0 | u). For notational brevity, define Q to
be the event where supp(X1:N ) ⊆ supp(ΘK ), and Q̃ to be the corresponding transformed event where supp(X̃1:N ) ⊆ supp(Θ̃K ). Then we have
hQ
i
∞
Ñ
P(Q̃) = E
.
k=K+1 π̃(uk )
If h(x | θ) = Bern(x; 1 − πκ,Ñ (θ)) and N = 1, then
hQ
i
R
∞
Ñ
P(Q̃) = E
π̃(u)
κ(θ
,
du)
= P(Q).
k
k=K+1
D.5. Truncation with hyperpriors.
Proof of Proposition 4.6. By repeating the proof of Lemma 4.1 in Appendix D.1,
except with an additional use of the tower property to condition on the hyperparameters Φ, an additional use of Fubini’s theorem to swap integration and expectation,
and Jensen’s inequality, we have
1
kpY − pW k1 ≤ E [1 − P (supp(X1:N ) ⊆ supp(ΘK ) | Φ)]
2
h
i
≤ E 1 − e−BN,K (Φ)
≤ 1 − e−E[BN,K (Φ)] .
Appendix E. Proofs of normalized truncation bounds
Proof of Lemma 5.3. First, we demonstrate that the arg max is well-defined. Note
that
arg max Ti + log pi = arg max exp (Ti + log pi )
i∈N
i∈N
if it exists, due to the monotonicity of exp. Similarly, existence of either proves the
existence of the other. Since Ti are i.i.d. Gumbel(0, 1),
P (exp (Ti + log pi ) > ) = 1 − exp −e−(log −log pi )
= 1 − exp (−pi /) ≤ 1 − (1 − pi /).
Therefore,
∞
X
P (exp (Ti + log pi ) > ) ≤
i=1
∞
X
1 − (1 − pi /) = −1
i=1
∞
X
pi < ∞.
i=1
This is sufficient to demonstrate that
a.s.
exp (Ti + log pi ) → 0
as i → ∞.
Finally, since any positive sequence converging to 0 can have only a finite number
of elements greater than any > 0, set = exp(T1 + log p1 ), and thus
arg max exp (Ti + log pi ) = arg max exp (Ti + log pi )
i∈N
i:Ti +log pi ≥
52
T. CAMPBELL, J. H. HUGGINS, J. P. HOW, AND T. BRODERICK
where the right hand side exists because it computes the maximum of a finite,
nonempty set of numbers. Note that the arg max is guaranteed to be a single
element, since Ti + log pi has a purely diffuse distribution on R.
Now that the a.s. existence and uniqueness of the arg max has been demonstrated, we can compute its distribution. First, note that
∞
Y
P (Ti + log pi ≤ x ∀i ∈ N, i 6= j) =
exp −e−(x−log pi ) = exp −e−x (s − pj ) ,
i=1,i6=j
P
where s := i pi . So then
P j = arg max Ti + log pi = P (Ti + log pi ≤ Tj + log pj ∀i ∈ N)
i∈N
Z
−(x−log pj )
−x
) dx
= e−e (s−pj ) e−(x−log pj +e
Z
−x
= pj e−se e−x dx
Z
−(x−log s)
pj
e−e
e−(x−log s) dx
=P
p
i
i
pj
P
=
,
i pi
where the last integral is 1 since its integrand is the Gumbel(s, 1) density.
E.1. Normalized series representation truncation.
Proof of Theorem 5.4. First, we apply Lemma 5.1,
1
kpY − pW k1 ≤ 1 − P (X1:N ⊆ supp(ΞK )) .
2
Next, by Jensen’s inequality,
"
N #
N
ΘK (Ψ)
ΘK (Ψ)
P (X1:N ⊆ supp(ΞK )) = E
≥E
Θ (Ψ)
Θ (Ψ)
= P(X1 ∈ supp(ΞK ))N .
The remaining part of this proof quantifies the probability that sampling X1 from
Ξ generates an atom in the support of ΞK (equal to the support of ΘK , since Ξ is
just the normalization of Θ). To do this, we use the trick based on Lemma 5.3: we
log-transform the rates in Θ, perturb them all by i.i.d. Gumbel random variables,
and quantify the probability that the max occurs within the atoms of ΘK .
First, we split the sequential representation of Θ into the truncation ΘK and its
tail Θ+
K , using the form from Eq. (4.1),
Θ=
∞
X
k=1
τ (Vk , Γk )δψk =
K
X
τ (Vk , Γk )δψk +
k=1
∞
X
τ (Vk , Γk )δψk := ΘK + Θ+
K.
k=K+1
Next, we define the maximum of the log-transformed, Gumbel perturbed rates in
ΘK as
MK := max log τ (Vk , Γk ) + Wk ,
1≤k≤K
TRUNCATED RANDOM MEASURES
53
i.i.d.
where Wk ∼ Gumbel(0, 1). Since Γk are from a unit-rate homogeneous Poisson
process, if we condition on the value of ΓK , this is equivalent to conditioning on
the event that the Poisson process has exactly K − 1 atoms on [0, ΓK ], and the K th
atom is at ΓK . And since MK does not depend on the ordering of (Γk )K
k=1 ,
D
MK | ΓK = max log τ (VK , ΓK ) + WK , max log τ (Vk , Uk ) + Wk ,
1≤k≤K−1
i.i.d.
where Uk ∼ Unif(0, ΓK ). Therefore, MK | ΓK is the maximum of a collection of
independent random variables, so we can compute its CDF by simply taking the
product of the CDFs of each of those random variables. Using standard techniques
for transformation of independent random variables, we have that
Z ∞Z 1
−x
g(v)e−τ (v,ΓK u)e dudv, k < K
P (log τ (Vk , Uk ) + Wk ≤ x | ΓK ) =
Z0 ∞ 0
−x
P (log τ (VK , ΓK ) + WK ≤ x | ΓK ) =
g(v)e−τ (v,ΓK )e dv,
0
so
∞
Z
Z
1
P (MK ≤ x | ΓK ) =
−τ (v,ΓK u)e−x
g(v)e
0
K−1 Z
dudv
∞
g(v)e
0
−τ (v,ΓK )e−x
dv .
0
Defining the function
h
i
J(u, t) := E e−t·τ (V,u) ,
V ∼ g,
we have
Z
K−1
J ΓK u, e−x du
J ΓK , e−x .
1
P (MK ≤ x | ΓK ) =
0
Next, we define an analogous maximum for the tail process rates in Θ+
K,
+
:= sup log τ (Vk , Γk ) + Wk .
MK
k>K
Conditioning on ΓK , we have
D
+
MK
| ΓK = sup log τ (Vk , Γ0k + ΓK ) + Wk ,
k≥1
where Γ0k is a unit-rate homogeneous Poisson process on R+ . Now note that since
Γ0k is a Poisson point process on R+ , so is log τ (Vk , Γ0k + ΓK ) + Wk (using Poisson
process stochastic mapping), with rate measure
Z ∞ Z ∞
−(t−log τ (v,u+ΓK ))
e−(t−log τ (v,u+ΓK ))−e
g(v)dudv dt.
0
0
+
MK
Therefore, P
≤ x | ΓK is equal to the probability that the above Poisson point
process has no atoms with position greater than x. Since aR Poisson process on R+
∞
with measure µ has no atoms above x with probability e− x µ(dt) ,
P
+
MK
≤ x | ΓK = e
−
R∞ R∞R∞
x
0
0
e−(t−log τ (v,u+ΓK ))−e
−(t−log τ (v,u+ΓK ))
g(v)dudv dt
.
54
T. CAMPBELL, J. H. HUGGINS, J. P. HOW, AND T. BRODERICK
Noticing that the integrand in t is a Gumbel density, we can use Fubini’s theorem
to swap integrals and evaluate:
R∞R∞
+
P MK
≤ x | ΓK = e 0 0
=e
R∞
0
e−τ (v,u+ΓK )e
−x
−1 g(v)dudv
(J (u+ΓK ,e−x )−1)du .
+
Taking the derivative yields the density of MK
| ΓK with respect to the Lebesgue
measure. Therefore, using the Gumbel-max trick from Lemma 5.3, we substitute
+
the results for MK | ΓK and MK
| ΓK into the original bound yielding
1
N
kpY − pW k1 ≤ 1 − P (X1 ⊆ supp(ΞK ))
2
N
+
= 1 − 1 − E P MK < MK
| ΓK
,
where (using the substitution t = e−x )
Z ∞
d
+
+
P MK < MK
| ΓK =
P (MK ≤ x | ΓK )
P MK
≤ x | ΓK dx
dx
−∞
Z 1
K−1
Z ∞
d R∞
=
J (ΓK , t)
J (ΓK u, t) du
− e 0 (J(u+ΓK ,t)−1)du dt.
dt
0
0
The fact that the bound is between 0 and 1 is a simple consequence of the fact that
P (X1 ∈ supp(ΞK )) ∈ [0, 1], and the asymptotic result follows from the fact that
P (X1 ∈ supp (ΞK )) → 1 as K → ∞.
E.2. Normalized superposition representation truncation.
Proof of Theorem 5.5. The same initial technique as in the proof of Theorem 5.5
yields
1
N
kpY − pW k1 ≤ 1 − P (X1 ∈ supp (ΞK )) .
2
The remaining part of this proof quantifies the probability that sampling X1 from Ξ
generates an atom in the support of ΞK . Since most of the following developments
+
are similar for ΘK , νK and Θ+
K , νK , we will focus the discussion on ΘK , νK and
reintroduce the tail quantities when necessary. First, we transform the rates of ΘK
i.i.d.
under the stochastic mapping w = log θ + W , where W ∼ Gumbel(0, 1), resulting
in a new Poisson point process with rate measure
Z
−(t−w)−e−(t−w) w
w
e
e νK (e )dw dt.
The probability that all points in this Poisson point process are less than a value
x is equal to the probability that there are no atoms above x. Defining MK to be
the supremum of the points in this process, combined with the basic properties of
Poisson point processes, we have
Z ∞Z
−(t−w)−e−(t−w) w
w
P(MK ≤ x) = exp −
e
e νK (e )dwdt
x
TRUNCATED RANDOM MEASURES
55
Using Fubini’s theorem to swap the integrals, we can evaluate the inner integral
analytically by noting the integrand is a Gumbel(w, 1) density,
Z
−e−(x−w) w
w
P(MK ≤ x) = exp − (1 − e
)e νK (e )dw
Z
−θe−x
e
− 1 νK (dθ) .
= exp
We can take the derivative with respect to x to obtain its density with respect to
the Lebesgue measure. The above derivation holds true for Θ+
K , replacing νK with
+
+
νK
and MK with MK
. Therefore, using the Gumbel-max trick from Lemma 5.3,
+
we substitute the results for MK and MK
into the original bound, yielding
1
N
kpY − pW k1 ≤ 1 − P (X1 ∈ supp (ΘK ))
2
+ N
= 1 − 1 − P MK < MK
,
where (using the substitution t = e−x , and the fact that θe−θt is dominated by θ
to swap integration and differentiation)
Z ∞
d
+
+
≤ x dx
P MK < MK
=
P(MK ≤ x) P MK
dx
−∞
Z
Z ∞ R −θe−x
e
−1 ν(dθ) d
+
−θe−x
=
e
e
− 1 νK (dθ) dx
dx
−∞
Z
Z ∞ R
+
e−θt −1)ν(dθ) d
−θt
(
e
− 1 νK (dθ) dt
=−
e
dt
0
Z
Z ∞ R
e−θt −1)ν(dθ)
−θt +
(
=
e
θe νK (dθ) dt.
0
The fact that the bound lies between 0 and 1 is a simple consequence of the fact
that P (X1 ∈ supp(ΞK )) ∈ [0, 1]. The fact that the error bound asymptotically
approaches 0 is a consequence of the monotone convergence theorem applied to the
+
decreasing sequence of functions θνK
(dθ).
E.3. Truncation with hyperpriors.
Proof of Proposition 5.6. By repeating the proof of Lemma 5.1 in Appendix D.1,
except with an additional use of the tower property to condition on the hyperparameters Φ, an additional use of Fubini’s theorem to swap integration and expectation,
and Jensen’s inequality, we have
1
kpY − pW k1 ≤ E [1 − P (X1:N ⊆ supp(ΘK ) | Φ)]
2
≤ 1 − E (1 − BK (Φ))N
N
≤ 1 − (1 − E [BK (Φ)]) .
References
Abramowitz, M. and Stegun, I. (eds.) (1964). Handbook of Mathematical Functions.
Dover Publications.
56
T. CAMPBELL, J. H. HUGGINS, J. P. HOW, AND T. BRODERICK
Airoldi, E. M., Blei, D., Erosheva, E. A., and Fienberg, S. E. (2014). Handbook of
Mixed Membership Models and Their Applications. CRC Press.
Arbel, J. and Prünster, I. (2017). “A moment-matching Ferguson & Klass algorithm.” Statistics and Computing, 27(1): 3–17.
Argiento, R., Bianchini, I., and Guglielmi, A. (2016). “A blocked Gibbs sampler for
NGG-mixture models via a priori truncation.” Statistics and Computing, 26(3):
641–661.
Banjevic, D., Ishwaran, H., and Zarepour, M. (2002). “A recursive method for
functionals of Poisson processes.” Bernoulli , 8(3): 295–311.
Blei, D. M. and Jordan, M. I. (2006). “Variational inference for Dirichlet process
mixtures.” Bayesian Analysis, 1(1): 121–144.
Bondesson, L. (1982). “On simulation from infinitely divisible distributions.” Advances in Applied Probability, 14: 855–869.
Brix, A. (1999). “Generalized gamma measures and shot-noise Cox processes.”
Advances in Applied Probability, 31: 929–953.
Broderick, T., Jordan, M. I., and Pitman, J. (2012). “Beta processes, stick-breaking
and power laws.” Bayesian Analysis, 7(2): 439–476.
Broderick, T., Mackey, L., Paisley, J., and Jordan, M. I. (2015). “Combinatorial clustering and the beta negative binomial process.” IEEE Transactions on
Pattern Analysis and Machine Intelligence, 37(2): 290–306.
Broderick, T., Wilson, A. C., and Jordan, M. I. (2017). “Posteriors, conjugacy, and
exponential families for completely random measures.” Bernoulli .
Doshi-Velez, F., Miller, K. T., Van Gael, J., and Teh, Y. W. (2009). “Variational
inference for the Indian buffet process.” In International Conference on Artificial
Intelligence and Statistics.
Ferguson, T. S. (1973). “A Bayesian analysis of some nonparametric problems.”
The Annals of Statistics, 1(2): 209–230.
Ferguson, T. S. and Klass, M. J. (1972). “A representation of independent increment
processes without Gaussian components.” The Annals of Mathematical Statistics,
43(5): 1634–1643.
Gautschi, W. (1959). “Some elementary inequalities relating to the gamma and
incomplete gamma function.” Journal of Mathematics and Physics, 38(1): 77–
81.
Gelfand, A. and Kottas, A. (2002). “A computational approach for nonparametric
Bayesian inference under Dirichlet process mixture models.” Journal of Computational and Graphical Statistics, 11(2): 289–305.
Gumbel, E. J. (1954). Statistical theory of extreme values and some practical applications: A series of lectures. National Bureau of Standards Applied Mathematics
Series 33. Washington, D.C.: U.S. Government Printing Office.
Hjort, N. L. (1990). “Nonparametric Bayes estimators based on beta processes in
models for life history data.” The Annals of Statistics, 18(3): 1259–1294.
Ishwaran, H. and James, L. F. (2001). “Gibbs sampling methods for stick-breaking
priors.” Journal of the American Statistical Association, 96(453): 161–173.
— (2002). “Approximate Dirichlet Process Computing in Finite Normal Mixtures:
Smoothing and Prior Information.” Journal of Computational and Graphical
Statistics, 11(3): 508–532.
Ishwaran, H. and Zarepour, M. (2002). “Exact and approximate sum representations for the Dirichlet process.” Canadian Journal of Statistics, 30(2): 269–283.
TRUNCATED RANDOM MEASURES
57
James, L. F. (2002). “Poisson Process Partition Calculus with applications
to Exchangeable models and Bayesian Nonparametrics.”
arXiv preprint
arXiv:0205093 .
— (2013). “Stick-breaking PG(α,ζ)-Generalized Gamma Processes.” arXiv preprint
arXiv:1308.6570 .
— (2014). “Poisson Latent Feature Calculus for Generalized Indian Buffet Processes.” arXiv preprint arXiv:1411.2936v3 .
James, L. F., Lijoi, A., and Prünster, I. (2009). “Posterior Analysis for Normalized Random Measures with Independent Increments.” Scandinavian Journal of
Statistics, 36(1): 76–97.
Kim, Y. (1999). “Nonparametric Bayesian estimators for counting processes.” The
Annals of Statistics, 27(2): 562–588.
Kingman, J. F. C. (1967). “Completely random measures.” Pacific Journal of
Mathematics, 21(1): 59–78.
— (1975). “Random discrete distributions.” Journal of the Royal Statistical Society
B , 37(1): 1–22.
Kingman, J. F. C. (1993). Poisson Processes. Oxford Studies in Probability. Oxford
University Press.
Lijoi, A., Mena, R., and Prünster, I. (2005). “Bayesian nonparametric analysis for a
generalized Dirichlet process prior.” Statistical Inference for Stochastic Processes,
8: 283–309.
Lijoi, A. and Prünster, I. (2010). “Models beyond the Dirichlet process.” In Hjort,
N. L., Holmes, C., Müller, P., and Walker, S. (eds.), Bayesian Nonparametrics,
80–136. Cambridge University Press.
Maddison, C., Tarlow, D., and Minka, T. P. (2014). “A* Sampling.” In Advances
in Neural Information Processing Systems.
Muliere, P. and Tardella, L. (1998). “Approximating distributions of random functionals of Ferguson–Dirichlet priors.” Canadian Journal of Statistics, 26(2): 283–
297.
Orbanz, P. (2010). “Conjugate projective limits.” arXiv preprint arXiv:1012.0363 .
Paisley, J. W., Blei, D. M., and Jordan, M. I. (2012). “Stick-breaking beta processes
and the Poisson process.” International Conference on Artificial Intelligence and
Statistics.
Paisley, J. W., Zaas, A. K., Woods, C. W., Ginsburg, G. S., and Carin, L. (2010).
“A stick-breaking construction of the beta process.” International Conference on
Machine Learning.
Perman, M. (1993). “Order statistics for jumps of normalised subordinators.” Stochastic Processes and their Applications, 46(2): 267–281.
Perman, M., Pitman, J., and Yor, M. (1992). “Size-biased sampling of Poisson
point processes and excursions.” Probability Theory and Related Fields, 92(1):
21–39.
Pitman, J. (2003). “Poisson-Kingman partitions.” Lecture Notes-Monograph Series.
Regazzini, E., Lijoi, A., and Prünster, I. (2003). “Distributional results for means
of normalized random measures with independent increments.” The Annals of
Statistics, 31(2): 560–585.
Rosiński, J. (1990). “On series representations of infinitely divisible random vectors.” Annals of Probability, 18: 405–430.
58
T. CAMPBELL, J. H. HUGGINS, J. P. HOW, AND T. BRODERICK
— (2001). “Series representations of Lévy processes from the perspective of point
processes.” In Barndorff-Nielson, O., Resnick, S., and Mikosch, T. (eds.), Lévy
processes: theory and applications, chapter VI, 401–415. Bikhäuser Boston.
Roy, D. (2014). “The continuum-of-urns scheme, generalized beta and Indian buffer
processes, and hierarchies thereof.” arXiv preprint arXiv:1501.00208 .
Roychowdhury, A. and Kulis, B. (2015). “Gamma Processes, Stick-Breaking, and
Variational Inference.” In International Conference on Artificial Intelligence and
Statistics.
Sethuraman, J. (1994). “A constructive definition of Dirichlet priors.” Statistica
Sinica, 4: 639–650.
Teh, Y. W. and Görür, D. (2009). “Indian buffet processes with power-law behavior.” In Advances in Neural Information Processing Systems.
Teh, Y. W., Görür, D., and Ghahramani, Z. (2007). “Stick-breaking construction for
the Indian buffet process.” In International Conference on Artificial Intelligence
and Statistics.
Thibaux, R. and Jordan, M. I. (2007). “Hierarchical beta processes and the Indian buffet process.” In International Conference on Artificial Intelligence and
Statistics.
Titsias, M. (2008). “The infinite gamma-Poisson feature model.” In Advances in
Neural Information Processing Systems.
Zhou, M., Hannah, L., Dunson, D., and Carin, L. (2012). “Beta-negative binomial
process and Poisson factor analysis.” In Artificial Intelligence and Statistics.
Computer Science and Artificial Intelligence Laboratory (CSAIL), Massachusetts
Institute of Technology
URL: http://www.trevorcampbell.me/
E-mail address: [email protected]
Computer Science and Artificial Intelligence Laboratory (CSAIL), Massachusetts
Institute of Technology
URL: http://www.jhhuggins.org/
E-mail address: [email protected]
Laboratory for Information and Decision Systems (LIDS), Massachusetts Institute
of Technology
URL: http://www.mit.edu/~jhow/
E-mail address: [email protected]
Computer Science and Artificial Intelligence Laboratory (CSAIL), Massachusetts
Institute of Technology
URL: http://www.tamarabroderick.com
E-mail address: [email protected]
| 10math.ST
|
PURE O-SEQUENCES: KNOWN RESULTS,
APPLICATIONS AND OPEN PROBLEMS
arXiv:1204.5247v2 [math.AC] 18 Feb 2013
JUAN MIGLIORE∗ , UWE NAGEL+ , AND FABRIZIO ZANELLO
Abstract. This note presents a discussion of the algebraic and combinatorial aspects
of the theory of pure O-sequences. Various instances where pure O-sequences appear are
described. Several open problems that deserve further investigation are also presented.
Dedicated to David Eisenbud on the occasion of his 65th birthday
1. Introduction
Pure O-sequences are fascinating objects that arise in several mathematical areas. They
have been the subject of extensive research, yet our knowledge about pure O-sequences
is limited. The goal of this note is to survey some of the known results and to motivate
further investigations by pointing out connections to various interesting problems. Much
of the material for this paper has been drawn from the recent monograph [2] by the
authors with Mats Boij and Rosa Miró-Roig, cited in this article as BMMNZ. We often
quote results giving the name of the author(s) and the year the result was published, so as
to also hint at the history of ideas in the development of the theory of pure O-sequences.
The multi-faceted interest in pure O-sequences is already indicated in their definition.
On the one hand, a pure O-sequence can be defined as the vector whose entries record
the number of monomials of a fixed degree in an order ideal generated by monomials of
the same degree. On the other, a pure O-sequence is the Hilbert function of a finitedimensional graded algebra that is level, i.e., its socle is concentrated in one degree,
and has monomial relations. There is an extensive literature on monomial ideals and an
extensive literature on level algebras. Pure O-sequences form a bridge between the two
theories, and we will outline the work on pure O-sequences from this point of view. But
more than that, these sequences have a broad array of applications, and occur in many
settings, largely combinatorial ones.
In Section 2, we review some of the basic results in the theory of pure O-sequences
and focus on qualitative aspects of their shape. For instance, in some cases pure Osequences are known to be unimodal; that is, they are first weakly increasing, and once
the peak is reached they are weakly decreasing. However, pure O-sequences may fail to be
unimodal, even with arbitrarily many “valleys.” We include a discussion of recent results
on conditions that force unimodality.
Connections to various combinatorial problems are the subject of Section 3. Face
vectors of pure simplicial complexes are examples of pure O-sequences. In particular, the
existence of certain block designs, such as Steiner systems, is related to that of some pure
O-sequences. As a special case, the existence of finite projective planes is equivalent to
the existence of particular pure O-sequences.
∗
The work for this paper was done while the first author was sponsored by the National Security
Agency under Grant Number H98230-12-1-0204, and by the Simons Foundation under grant #208579.
+
The work for this paper was done while the second author was sponsored by the National Security
Agency under Grant Number H98230-12-1-0247, and by the Simons Foundation under grant #208869.
1
2
JUAN MIGLIORE, UWE NAGEL, AND FABRIZIO ZANELLO
Another challenging problem is Stanley’s conjecture that the h-vector of any matroid
complex is a pure O-sequence. We discuss some recent progress. However, the conjecture
remains open in general.
In Section 4, we describe results on the enumeration of pure O-sequences, with a focus
on asymptotic properties. In particular, it follows that, when the number of variables is
large, “almost all” pure O-sequences are unimodal.
We conclude this note with a collection of open problems, most of which are mentioned
in the earlier sections.
2. Monomial level algebras
A finite, nonempty set X of (monic) monomials in the indeterminates y1 , . . . , yr is
called a monomial order ideal if, whenever M ∈ X and N is a monomial dividing M , then
N ∈ X. The h-vector of X is defined to be the vector h = (h0 = 1, h1 , . . . , he ) counting
the number of monomials of X in each degree. A monomial order ideal, X, is called pure
if all maximal monomials of X (in the partial ordering given by divisibility) have the same
degree. A pure O-sequence is the h-vector of a pure monomial order ideal. For reasons
that will be clear shortly, we call e the socle degree of h. The type of a pure O-sequence
is the number of maximal monomials.
Notice that if we think of y1 , . . . , yr as the indeterminates of a polynomial ring R =
K[y1 , . . . , yr ] over a field, the question of whether a given sequence is or is not a pure
O-sequence does not depend on the choice of K. Thus, for many of our results it does
not matter what we choose for K. However, in some situations choosing a “nice” field K
allows us to use special algebraic tools to say something about pure O-sequences, so in
this case we make whatever additional assumptions we need for K.
We now let R = K[x1 , . . . , xr ], where K is an infinite field. We will consider standard
graded artinian K-algebras A = R/I, where I will usually be a monomial ideal. Without
loss of generality we will assume that I does not contain nonzero linear forms, so we will
define r to be the codimension of A.
Let R = K[y1 , . . . , yr ], and consider the action of R on monomials of R by contraction.
By this we mean the action generated by
a1 a2
y1 y2 · · · yia1 −1 · · · yrar , if ai > 0,
a1 a2
ar
xi ◦ y1 y2 · · · yr =
0,
if ai = 0.
For a monomial ideal I ⊂ R, we define the inverse system to be the R-module I ⊥ =
annR (I) ⊂ R. One can check that I ⊥ consists of the monomials not in I (identifying xi
with yi ), and as such it can be viewed as a monomial order ideal. Recalling that for a
standard graded algebra R/I the Hilbert function is defined to be hR/I (t) = dim[R/I]t ,
we observe that the h-vector (as defined above) of the order ideal I ⊥ coincides with the
Hilbert function of R/I.
Furthermore, I ⊥ is a pure monomial order ideal if and only if R/I is a level algebra;
that is, the socle of R/I (i.e., the annihilator of the homogeneous maximal ideal of R/I)
is concentrated in one degree, called the socle degree of R/I; it is necessarily the degree
of the maximal monomials of I ⊥ . The dimension of the socle as a K-vector space is equal
to the type of the pure O-sequence. See [25, 39] for more details on inverse systems.
Thus the study of pure O-sequences boils down to a study of the possible Hilbert
functions of artinian monomial level algebras. In some cases we can rule out candidates
for pure O-sequences by showing that there is not even a level algebra with that Hilbert
function, but more often we need to use the structure of monomial algebras themselves.
PURE O-SEQUENCES: KNOWN RESULTS, APPLICATIONS AND OPEN PROBLEMS
3
The most basic tool is Macaulay’s theorem to determine if the sequence is even an
O-sequence; that is, to determine if it is the Hilbert function of some artinian algebra.
We refer to [8, 47] for details of Macaulay’s theorem, but we recall the statement. Let n
and d be positive integers. There exist uniquely determined integers kd > kd−1 > · · · >
kδ ≥ δ ≥ 1 such that
kd
kd−1
kδ
n = n(d) =
+
+ ··· +
.
d
d−1
δ
This is called the d-binomial expansion of n. We set
kd + 1
kd−1 + 1
kδ + 1
1
(n(d) )1 =
+
+ ··· +
d+1
d
δ+1
and (0(d) )11 = 0, for each d. Then Macaulay’s theorem is the following.
Theorem 2.1 (Macaulay 1927). Let A be a standard graded algebra with Hilbert function
hA (t) := ht . Then for all t ≥ 1, ht+1 ≤ ((ht )(t) )11 .
An O-sequence is a (possibly infinite) sequence of integers (1, h1 , h2 , . . . ) that satisfies
the growth condition of Theorem 2.1 for every value of t. Thus the O-sequences are the
sequences that occur as the Hilbert function of some standard graded algebra.
Example 2.2. The sequence (1, 3, 6, 8, 8, 10) is not a pure O-sequence because it is
not even an O-sequence (the growth from degree 4 to degree 5 is too big). Similarly,
(1, 3, 5, 5, 4, 4) is an O-sequence but it is not a pure O-sequence because it is not the
Hilbert function of a level algebra (see [26]). Finally, h = (1, 3, 6, 10, 15, 21, 28, 27, 27, 28)
is the Hilbert function of a level algebra, but it is not a pure O-sequence [4] because
there is no monomial level algebra with this Hilbert function. In fact, h has been the
first nonunimodal level Hilbert function discovered in codimension 3 (see the third author [83]), and Boyle [4] has shown that this is in fact the smallest possible such Hilbert
function.
So the challenge is to determine what additional conditions on an O-sequence are imposed by requiring that it be the Hilbert function of an artinian level monomial algebra.
The first result, due originally to Stanley [67] with subsequent proofs given by J. Watanabe
[78], Ikeda [65], Reid, Roberts and Roitman [63], Herzog and Popescu [33], and Lindsey
[45], concerns monomial complete intersections. (Note though that an equivalent property was proven earlier by de Bruijn, van Ebbenhorst Tengbergen and Kruyswijk [21].)
This result requires us to introduce here the notion of the Weak and Strong Lefschetz
Properties. The consequence of this result for pure O-sequences of type 1 is perhaps not
so critical, as alternative proofs could be given. But its influence in the study of the SLP
and the WLP in general, and on related topics in commutative algebra, can hardly be
overstated. In the result below, and throughout the paper, we will call a sequence unimodal if it is nondecreasing up to some degree and then nonincreasing past that degree.
We will call it strictly unimodal if it is strictly increasing up to some degree, then possibly
constant for some range, then strictly decreasing until it reaches zero, and then zero past
that point. Unimodality is a central concept in combinatorics and combinatorial commutative algebra, but also in other branches of mathematics. See for instance the classical
surveys of Stanley ([68], 1989) and Brenti ([6], 1994).
Theorem 2.3 (see above for sources). Let R = K[x1 , . . . , xr ], where K has characteristic
zero, and let I be an artinian monomial complete intersection, i.e.,
I = hxa11 , . . . , xar r i.
4
JUAN MIGLIORE, UWE NAGEL, AND FABRIZIO ZANELLO
Let L be a general linear form. Then for any positive integers d and i, the homomorphism
induced by multiplication by Ld ,
×Ld : [R/I]i → [R/I]i+d ,
has maximal rank. (In particular, this is true when d = 1.) As a consequence, a pure
O-sequence of type 1 is strictly unimodal.
When I is an arbitrary artinian homogeneous ideal, the above maximal rank property
for all d and i is called the Strong Lefschetz Property (SLP), and the case d = 1 is called
the Weak Lefschetz Property (WLP). A consequence of the WLP is that in the range
where (×L) is injective, we have a short exact sequence
0 → [R/I]i → [R/I]i+1 → [R/(I, L)]i+1 → 0.
Thus, the first difference ∆hR/I (t) = hR/I (t) − hR/I (t − 1) is the Hilbert function of a
standard graded algebra in this range, that is, it is again an O-sequence. We then say
that hR/I is a differentiable O-sequence in this range. It also follows from this sequence
that if the WLP holds, then once the peak of the Hilbert function is reached, the Hilbert
function will be nonincreasing, and therefore the whole Hilbert function is unimodal.
Remark 2.4. Simple examples show that Theorem 2.3 may fail in positive characteristic.
It turns out that the question in which positive characteristics a monomial complete
intersection has the weak or strong Lefschetz property leads to unexpected connections
to the problem of determining the number of certain plane partitions, lozenge tilings, or
families of lattice paths (see [11, 16, 18, 19, 43]).
One of the early important results on pure O-sequences is due to Hibi [34].
Theorem 2.5 (Hibi 1989). Let h be a pure O-sequence of socle degree e. Then
hi ≤ hj
whenever 0 ≤ i ≤ j ≤ e − i. This has the following two important consequences:
(a) h is flawless, i.e., hi ≤ he−i for all 0 ≤ i ≤ b 2e c.
(b) The “first half ” of h is nondecreasing:
1 = h0 ≤ h1 ≤ h2 ≤ · · · ≤ hb 2e c .
This latter result was later improved by the following algebraic g-theorem of Hausel
[31]:
Theorem 2.6 (Hausel 2005). Let A be a monomial Artinian level algebra of socle degree
e. If the field K has characteristic zero, then for a general linear form L, the induced
multiplication
×L : Aj → Aj+1
is an injection, for all j = 0, 1, . . . , b e−1
c. In particular, over any field, the sequence
2
1, h1 − 1, h2 − h1 , . . . , hb e−1 c+1 − hb e−1 c
2
2
is an O-sequence, i.e., the “first half ” of h is a differentiable O-sequence.
We have the following additional results on differentiability from [2]:
Theorem 2.7 (BMMNZ 2012).
(a) Every finite differentiable O-sequence h is the “first
half” of some pure O-sequence. (This is the converse of Hausel’s theorem.)
(b) In particular, any finite differentiable O-sequence is pure (by truncation).
(c) Any nondecreasing pure O-sequence of socle degree ≤ 3 is differentiable.
PURE O-SEQUENCES: KNOWN RESULTS, APPLICATIONS AND OPEN PROBLEMS
5
It turns out that (c) is the best possible result in this direction:
Proposition 2.8 (BMMNZ 2012). There exist nondecreasing pure O-sequences of any
socle degree e ≥ 4 that are not differentiable.
Example 2.9. We illustrate the preceding result with an example from [2]. Observe first
that the h-vector h0 = (1, 4, 10, 20, 35) is a pure O-sequence since it is the h-vector of the
truncation of a polynomial ring in four variables, w, x, y, z; the pure order ideal arises
using all 35 monomials of degree 4 in w, x, y, z. The h-vector h00 = (1, 4, 6, 4, 1) is also
a pure O-sequence, since it is the order ideal generated by a monomial abcd in four new
variables. Now we work in a polynomial ring in the eight variables w, x, y, z, a, b, c, d, and
we consider the pure order ideal generated by the above 36 monomials of degree 4. The
resulting h-vector h is
1 4 10 20 35
+
4 6 4 1
1 8 16 24 36
Since the first difference of h = (1, 8, 16, 24, 36) is (1, 7, 8, 8, 12), which is not an O-sequence
(because 12 > (8(3) )11 = 10), we have constructed the desired example.
Putting aside the class of nondecreasing pure O-sequences, we now turn to the question
of unimodality. There are three factors that go into whether a nonunimodal example will
exist: the codimension (i.e., the number of variables), the socle degree, and the type.
An easy application of Macaulay’s theorem gives that any standard graded algebra of
codimension two has unimodal Hilbert function, and in fact it is not hard to show that
if K[x, y]/I is level (monomial or not) then the Hilbert function is strictly unimodal (see
e.g. Iarrobino [38], 1984). Hence the interesting questions arise for codimension r ≥ 3.
We will begin with some results involving the socle degree (some of which will also
bring in the codimension). It follows from Hibi’s result on flawlessness that any pure
O-sequence of socle degree ≤ 3 is unimodal. The next case, socle degree 4, already is not
necessarily unimodal, again thanks to an example from [2]:
Example 2.10. Observe that h0 = (1, 5, 15, 35, 70) is a pure O-sequence, since it is
the h-vector of the truncation of a polynomial ring in five variables, and as before h00 =
(1, 4, 6, 4, 1) is also pure, since it corresponds to the maximal monomial abcd ∈ K[a, b, c, d].
Hence, reasoning as above, we now consider one copy of h0 and eleven copies of h00 as hvectors of pure O-sequences in twelve different rings, and we work in the tensor product
of those rings. It follows that
h = (1, 5, 15, 35, 70) + 11 · (0, 4, 6, 4, 1) = (1, 49, 81, 79, 81)
is a nonunimodal pure O-sequence of socle degree 4.
A natural question, then, is what is the smallest codimension for which nonunimodal
pure O-sequences exist with socle degree 4. This remains open. We do have the following
results from [4], however.
Theorem 2.11 (Boyle 2012).
(a) All pure O-sequences of socle degree ≤ 9 in three
variables are unimodal.
(b) All pure O-sequences of socle degree ≤ 4 in four variables are unimodal.
(c) In four or more variables, there exist nonunimodal pure O-sequences in all socle
degrees ≥ 7.
In [3], Boij and the third author gave a nonunimodal pure O-sequence of codimension 3
and socle degree 12. This is the smallest known example in codimension 3. It follows
6
JUAN MIGLIORE, UWE NAGEL, AND FABRIZIO ZANELLO
from this and Boyle’s result that in codimension 3, the only open cases are socle degrees
10 and 11. Notice that, in codimension 4, the previous theorem leaves only open the socle
degrees 5 and 6.
Now we turn to questions involving the type. Of course it follows immediately from
Theorem 2.3 that pure O-sequences of type 1 are unimodal. What else can be deduced
about pure O-sequences using the WLP? A collection of results was obtained in [2] which
showed, in some sense, the limits of the WLP in the study of pure O-sequences.
Theorem 2.12 (BMMNZ 2012). Over a field of characteristic zero the following hold:
(a) Any monomial artinian level algebra of type 2 in three variables has the WLP. Thus
a pure O-sequence of type 2 and codimension 3 is differentiable until it reaches its
peak, is possibly constant, and then is nonincreasing until it reaches zero.
(b) Fix two positive integers r and d. Then all monomial artinian level algebras of
codimension r and type d possess the WLP if and only if at least one the following
is true:
(i) r = 1 or 2;
(ii) d = 1;
(iii) r = 3 and d = 2.
The proof of (a) was surprisingly long and intricate. The main point of (b) is that in all
other cases, we were able to show that artinian monomial level algebras exist that do not
have the WLP.
Notice that Theorems 2.3, 2.6 and 2.12(a) require that K have characteristic zero. The
statements about injectivity and surjectivity require this property of the characteristic,
and indeed a great deal of research has been carried out to see what happens when
K has positive characteristic; we refer to [53] for an overview of these results. The
consequences on the shape of the pure O-sequences are indeed characteristic free, as has
been noted above, since the Hilbert function of a monomial ideal does not depend on the
characteristic.
If one is studying all artinian level monomial algebras of fixed type, Theorem 2.12 is
a serious limitation on the usefulness of the WLP. However, in the study of pure Osequences (e.g. to determine combinations of codimension, type and socle degree that
force unimodality) it is still conceivable that the WLP will play a useful role. As a trivial
example, we know that monomial complete intersections in any codimension possess the
WLP, thanks to Theorem 2.3. It is not known (except in codimension ≤ 3, as noted
below) whether all complete intersections have this property, but the knowledge in the
special case of monomial ideals is enough to say that all complete intersection Hilbert
functions are unimodal. Perhaps a similar phenomenon will allow the WLP to continue
to play a role in the study of pure O-sequences. A first approach using this philosophy
was obtained by Cook and the second author [17], where they lifted a monomial ideal to
an ideal of a reduced set of points in one more variable, showed that the general artinian
reduction has the WLP, and concluded that the Hilbert function of the original monomial
algebra is unimodal, regardless of whether it has the WLP or not.
For this reason we mention a useful tool in studying the WLP, that was introduced by
the first and second authors with T. Harima and J. Watanabe in [30] and whose study
was continued by H. Brenner and A. Kaid in [5]. This is the study of the syzygy bundle,
and the use of the Grauert-Mülich theorem (see [58]). It was used in [30] to show that
any complete intersection I in K[x, y, z] over a field of characteristic zero has the WLP.
The idea is to restrict to a general line, say one defined by a general linear form L. The
key is that the restricted ideal (I, L)/(L) (which now has codimension 2, hence has a
PURE O-SEQUENCES: KNOWN RESULTS, APPLICATIONS AND OPEN PROBLEMS
7
Hilbert-Burch matrix) should have minimal syzygies in two consecutive degrees, at most.
The idea of [30] was that for height 3 complete intersections, this information can be
obtained by considering the module of syzygies of I, sheafifying it to obtain the syzygy
bundle, and applying the Grauert-Mülich theorem to the general line defined by L. Of
course this introduces questions about the semistability of the syzygy bundle, which we
mostly omit here. In a more general setting, the following result from [5] summarizes the
idea nicely, at least for codimension three.
Theorem 2.13 (Brenner-Kaid 2007). Let I = hf1 , . . . , fk i ⊂ K[x, y, z] = R be an artinian
homogeneous ideal whose syzygy bundle S is semistable on P2 . Then
(a) If the restriction of S splits on a general line L as
SL ∼
=
s
M
i=1
OL (a + 1) ⊕
k−1
M
OL (a),
i=s+1
then A = R/I has the WLP.
(b) If the restriction of S splits on a general line L as
SL ∼
= OL (a1 ) ⊕ · · · ⊕ OL (an−1 )
with a1 ≥ a2 ≥ · · · ≥ an−1 and a1 − an−1 ≥ 2, then A = R/I does not have the
WLP.
Other applications of this approach, for higher codimension, can be found for instance
in [51].
Returning to unimodality questions, tools other than the WLP will also be needed. We
have the following theorems of Boyle [4], which relied on decomposition results and an
analysis of complete intersection Hilbert functions but did not use the WLP.
Theorem 2.14 (Boyle 2012).
(a) In codimension 3, all pure O-sequences of type 3
are strictly unimodal.
(b) In codimension 4, all pure O-sequences of type 2 are strictly unimodal.
A natural question is whether all pure O-sequences of type 2 and arbitrary codimension
are unimodal. Also, for any fixed codimension, it is an interesting problem to determine
which types force unimodality. The “record” for the smallest known nonunimodal example
in codimension 3 is type 14, given in [2].
Finally, one can ask “how nonunimodal” a pure O-sequence can be. The answer is “as
nonunimodal as you want.” We have:
Theorem 2.15 (BMMNZ 2012). For any integers M ≥ 2 and r ≥ 3, there exists a pure
O-sequence in r variables which is nonunimodal and has exactly M maxima.
Of course the “price” in Theorem 2.15 is paid in having a large socle degree and type.
In fact, even Cohen-Macaulay f -vectors (i.e., the face vectors of Cohen-Macaulay simplicial complexes, which are a much smaller subset of pure O-sequences) can be nonunimodal with arbitrarily many peaks (see [61]). This result considerably extends Theorem
2.15, even though, unlike for arbitrary pure O-sequences, here the number of variables
becomes necessarily very large as the number of peaks increases.
A good topic to build a bridge between the algebraic and the combinatorial sides of the
theory of pure O-sequences is the Interval Conjecture.
The Interval Property (IP) was introduced in 2009 by the third author [84], where he
conjectured its existence for the set of Hilbert functions of level — and, in a suitably
8
JUAN MIGLIORE, UWE NAGEL, AND FABRIZIO ZANELLO
symmetric way, Gorenstein — algebras. Namely, the IP says that if two (not necessarily
finite) sequences, h and h0 , of a class S of integer sequences coincide in all entries but one,
say
h = (h0 , . . . , hi−1 , hi , hi+1 , . . . ) and h0 = (h0 , . . . , hi−1 , hi + α, hi+1 , . . . ),
for some index i and some positive integer α, then the sequences
(h0 , . . . , hi−1 , hi + β, hi+1 , . . . )
are also in S, for all β = 1, 2, . . . , α − 1.
Given that level and Gorenstein Hilbert functions are nearly impossible to characterize,
the IP appears to be both a very natural property and one of the strongest structural
results that we might hope to achieve for the set of such sequences.
For example, it is proved in [84] that the IP holds for all Gorenstein Hilbert functions
of socle degree 4. Since Gorenstein Hilbert functions are symmetric, this means that,
for any fixed r, (1, r, a, r, 1) is Gorenstein
if and only if a ranges between some minimum
possible value, say f (r), and r+1
.
This
latter
is the maximum allowed by a polynomial
2
ring in r variables, and is achieved by the so-called compressed Gorenstein algebras (see,
e.g., the 1984 papers of Fröberg-Laksov and Iarrobino [24, 38] or the third author’s works
[81, 82]).
Notice that the existence of the IP for these Hilbert functions is especially helpful in
view of the fact that, for most codimensions r, the value of f (r) is not known. In fact,
such Gorenstein Hilbert functions are “highly” nonunimodal. Asymptotically, we have
f (r)
lim 2/3 = 62/3 ,
r→∞ r
as proved by these three authors in [54] (2008), solving a longstanding conjecture of
Stanley [69] (see also [55], 2009, for some broad generalizations).
The Interval Property is still wide open today for both level and Gorenstein algebras.
In a more combinatorial direction, in BMMNZ 2012 the IP has then also been conjectured for: 1) pure O-sequences (under the name “ICP”); and 2) the f -vectors of pure
simplicial complexes, a topic of discussion of the next section.
As for pure O-sequences, the ICP has been proved in a number of special cases. Most
importantly, it is known when the socle degree is at most 3, in any number of variables.
Theorem 2.16 (BMMNZ 2012). The ICP holds for the set of all pure O-sequences of
socle degree e ≤ 3.
Thanks to this result, H.T. Hà, E. Stokes and the third author [27] have recently
developed a new approach leading to a proof of Stanley’s matroid h-vector conjecture in
Krull-dimension 3, as we will see in the next section.
While the ICP remains open in most instances — e.g., in three variables — it must be
pointed out that, just recently, it has been disproved in the four variable case by Constantinescu and Varbaro (see [15, Remark 1.10]), who found the following counterexample.
Example 2.17 (Constantinescu-Varbaro 2012). Consider the pure order ideals generated by {x3 y 2 z, x3 yt2 , x3 z 2 t} and {x4 y 2 , x3 yzt, x2 z 2 t2 }. Their h-vectors are the pure Osequences (1, 4, 10, 13, 12, 9, 3) and (1, 4, 10, 13, 14, 9, 3). However, an exhaustive computer
search over all sets of three monomials of degree 6 in four variables reveals that the sequence (1, 4, 10, 13, 13, 9, 3) is not pure, contrary to the ICP.
It is worth remarking that h = (1, 4, 10, 13, 13, 9, 3) is, however, a level h-vector, and so
this does not provide a counterexample to the IP for arbitrary level algebras. Indeed, h is
the h-vector of a level algebra in four variables whose inverse system is generated by two
PURE O-SEQUENCES: KNOWN RESULTS, APPLICATIONS AND OPEN PROBLEMS
9
sums of sixth powers of six general linear forms each, and the sixth power of one general
linear form.
As for pure f -vectors, the IP is still wide open, and little progress has been made so
far.
In general, at this time it is still unclear what the exact scope of the Interval Property
is, and if it can also be of use in other areas of combinatorial algebra or even enumerative
combinatorics. It is well known to hold, e.g., for the set of Hilbert functions of graded
algebras of any Krull-dimension (see Macaulay’s theorem), the f -vectors of arbitrary
simplicial complexes (the Kruskal-Katona theorem), and the f -vectors of Cohen-Macaulay
complexes (BMMNZ 2012). Instead, the IP fails quite dramatically, for example, for
matroid h-vectors, which are conjecturally another subset of pure O-sequences, as we will
see in the next section (we refer to BMMNZ 2012 and [69] for details). Stanley and the
third author [72] recently looked at the IP in the context of r-differential posets, a class of
ranked posets generalizing the Young lattice of integer partitions and the Young-Fibonacci
lattice. Here, even though the IP fails in general, it might be a reasonable property to
conjecture, for instance, for r = 1, which is the most natural class of differential posets.
3. Pure O-sequences and combinatorics
Much of the motivation for the study of pure O-sequences comes from combinatorics,
and in this section we give an overview of this side of the theory. In order to put in context
the definition of a pure O-sequence given in the introduction, we quickly recall the notion
of posets and order ideals. For an introduction to this theory, we refer to Chapter 3 of
Stanley’s new edition of “EC1” ([71], 2012). A poset (short for partially ordered set) is a
set S equipped with a binary relation, “≤,” that is: 1) reflexive (i.e., a ≤ a for all a ∈ S);
2) antisymmetric (a ≤ b ≤ a implies a = b); and 3) transitive (a ≤ b ≤ c implies a ≤ c).
An order ideal in a poset S is a subset I of S that is closed with respect to “≤ .”
That is, if t ∈ I and s ≤ t, then s ∈ I. Thus, our monomial order ideals are the (finite)
order ideals of the poset P of all monomials in the polynomial ring R = K[y1 , . . . , yr ],
where the binary relation of P is divisibility. Notice that P is a ranked poset, where the
rank of a monomial is its degree in R. Therefore, Macaulay’s O-sequences are exactly the
possible rank functions of the order ideals of P , since every Hilbert function satisfying
Macaulay’s theorem can be achieved by a monomial algebra. Similarly, as we have seen,
a pure O-sequence is the rank function of some monomial order ideal whose generators
(i.e., the antichain of maximal monomials) are all of the same degree.
Another fundamental class of order ideals are those contained in the Boolean algebra Br ,
the poset of all subsets of {1, 2, . . . , r}, ordered by inclusion. By identifying the integer
i with a vertex vi , the order ideals of Br are usually called simplicial complexes (on r
vertices).
The elements of a simplicial complex ∆ are dubbed faces, and the maximal faces are
the facets of ∆. The dimension of a face is its cardinality minus 1, and the dimension of
∆ is the largest of the dimensions of its faces.
Notice that if we identify vi with a variable yi ∈ R = K[y1 , . . . , yr ], then simplicial
complexes also coincide with the order ideals of P generated by squarefree monomials. In
particular, if we define as pure those simplicial complexes whose facets have all the same
dimension, then clearly their rank vectors, called pure f -vectors, are the special subset of
pure O-sequences that can be generated by squarefree monomials.
10
JUAN MIGLIORE, UWE NAGEL, AND FABRIZIO ZANELLO
Example 3.1. The simplicial complex
∆ = {{1, 2, 3}, {2, 3, 4}, {1, 2}, {1, 3}, {2, 3}, {2, 4}, {3, 4}, {1}, {2}, {3}, {4}, ∅}
is the order ideal of B4 generated by {1, 2, 3} and {2, 3, 4}.
Thus, ∆ is a pure complex of dimension 2, whose pure f -vector is f∆ = (1, 4, 5, 2).
Equivalently, f∆ is the pure O-sequence generated by the two squarefree monomials
y1 y2 y3 and y2 y3 y4 .
Similarly to Macaulay’s theorem for arbitrary O-sequences, we know a characterization
of the class of pure f -vectors thanks to the classical Kruskal-Katona theorem (see e.g.
[69]). However, analogously, things become dramatically more complicated (hopeless, we
should say) when it comes to attempting a characterization of pure f -vectors.
In the last section of BMMNZ 2012, we have begun a study of pure f -vectors, but still
very little is known today beyond what is known for arbitrary pure O-sequences.
Besides their obvious intrinsic importance — simplicial complexes are a central object in
algebraic combinatorics, combinatorial algebra and topology, just to name a few subjects
— pure f -vectors also carry fascinating applications. It is on their connections to finite
geometries and design theory that we want to focus in the next portion of this section.
It will follow from our discussion, as probably first observed by Björner ([1], 1994),
that a characterization of pure simplicial complexes and their f -vectors would imply, for
instance, that of all Steiner systems, and as a further special case, a classification of all
finite projective planes, one of the major open problems in geometry.
A Steiner system S(l, m, r) is an r-element set V , together with a collection of m-subsets
of V , called blocks, such that every l-subset of V is contained in exactly one block. Steiner
systems are a special family of the so-called block designs. We refer our reader to the two
texts [14, 44], where she can find a truly vast amount of information on combinatorial
designs. For instance, a Steiner triple system (STS) is a Steiner system S(2, 3, r), while
S(3, 4, r) is dubbed a Steiner quadruple system, where r is the order of the system.
Since we are dealing with maximal sets of the same cardinality (m, in this case), it
is clear that if we identify each element of V with a variable yi , then the existence of
Steiner systems (and similarly for other block designs) will be equivalent to the existence
of certain pure f -vectors.
Example 3.2. Let us consider STS’s of order 7, i.e., S(2, 3, 7). Constructing such a
design is tantamount to determining a family of squarefree degree 3 monomials of R =
K[y1 , y2 , . . . , y7 ], say M1 , M2 , . . . , Mt , such that each squarefree degree 2 monomial of R
divides exactly one of the Mi .
Clearly, since
there are 72 = 21 squarefree degree 2 monomials in R, if the Mi exist,
then t = 21/ 32 = 7. In other words, S(2, 3, 7) exists if and only if
f = (1, 7, 21, 7)
is a pure f -vector.
Notice also that f exists as a pure f -vector if and only if it exists as a pure O-sequence,
since for seven degree 3 monomials to have a total of 21 degree 2 divisors, each needs to
have exactly three linear divisors, i.e., it must be squarefree.
It is easy to see that an STS of order 7, and so the pure f -vector f = (1, 7, 21, 7), do
indeed exist, using the monomial order ideal generated by:
y1 y2 y3 , y3 y4 y5 , y3 y6 y7 , y1 y4 y7 , y2 y4 y6 , y2 y5 y7 , y1 y5 y6 .
PURE O-SEQUENCES: KNOWN RESULTS, APPLICATIONS AND OPEN PROBLEMS
11
Some simple numerical observations show that a necessary condition for an STS of
order r to exist is that r be congruent to 1 or 3 modulo 6, and it is a classical result of
Kirkman ([40], 1847) that this is also sufficient. In other words,
r
r
f = 1, r,
,
/3
2
2
is a pure f -vector, if and only if it is a pure O-sequence, if and only if r is congruent to 1
or 3 modulo 6.
A different, and more challenging, problem is the classification of all Steiner systems,
up to isomorphism. Even the existence of particular systems sometimes brings into the
story a nontrivial amount of interesting algebra. As an illustration, we mention here the
case of the Steiner systems S(4, 5, 11), S(5, 6, 12), S(3, 6, 22), S(4, 7, 23), and S(5, 8, 24),
which are intimately connected to the first sporadic finite simple groups ever discovered,
called the Mathieu groups (see E.L. Mathieu [48, 49], 1861 and 1873). These five groups —
denoted respectively by M11 , M12 , M22 , M23 , and M24 — in fact arise as the automorphism
groups of the above Steiner systems (i.e., the transformations of the systems that preserve
the blocks).
There exists only one STS of order 7, which is called the Fano plane (see the figure
below) for reasons that will be clear in a minute. In other words, the seven monomials
of the previous example are, up to isomorphism, the only possible set of generators for a
pure order ideal in K[y1 , y2 , . . . , y7 ] with (1, 7, 21, 7) as its pure O-sequence.
y1
y2
y5
y4
y3
y6
y7
Also for r = 9, there exists a unique STS. However, it is reasonable to believe that the
number of nonisomorphic STS increases extremely quickly for r large. For instance, there
are 80 nonisomorphic STS of order 15, and there are 11,084,874,829 of order 19.
Similarly, the possible orders of Steiner quadruple systems are known and nicely characterized, since the obvious necessary conditions again turn out to be also sufficient:
S(3, 4, r) exists if and only if its order r is congruent to 2 or 4 modulo 6, as proved by
Hanani ([28], 1960). In other words, reasoning as above, since there are four possible
3-subsets of any given 4-set, we have that
r
r
r
f = 1, r,
,
,
/4
2
3
3
is a pure f -vector if and only if it is a pure O-sequence, if and only if r is congruent to 2
or 4 modulo 6.
However, as we increase the cardinality m of the blocks, things become more and more
obscure. This is due to the high complexity of computing combinatorial designs over a
large vertex set, as well as to the lack of a general theory.
12
JUAN MIGLIORE, UWE NAGEL, AND FABRIZIO ZANELLO
Already for Steiner quintuple systems, the trivial necessary conditions (r congruent to
3 or 5, but not to 4, modulo 6) are no longer sufficient. For example, no Steiner quintuple
system S(4, 5, 17) exists (see [59], 2008). The smallest value of r for which the existence
of S(4, 5, r) is currently open is 21. In other words, it is unknown whether
21
21
21
21
1, 21,
,
,
,
/5 = (1, 21, 210, 1330, 5985, 1197)
2
3
4
4
is a pure O-sequence.
Perhaps the best-known family of examples of Steiner systems is that of finite projective
planes, so they deserve a special mention here. Recall that a projective plane is a collection
of points and lines such that any two lines “intersect at” exactly one point, and any two
points “lie on” exactly one line (one also assumes that there exist four points no three of
which are collinear, in order to avoid uninteresting pathological situations).
If the projective plane is finite, it can easily be seen that the number of points is equal to
the number of lines, and that this number must be of the form q 2 +q +1, where the integer
q is the order of the plane. Further, in a projective plane of order q, any line contains
exactly q + 1 points, and by duality, any point is at the intersection of exactly q + 1 lines.
In other words, finite projective planes are the Steiner systems S(2, q + 1, q 2 + q + 1). The
reader may want to consult, e.g., [23] for an introduction to this area.
Thus, the above example of a Steiner system S(2, 3, 7), the Fano plane, is the unique
smallest possible projective plane.
Similarly to how we argued earlier in terms of pure f -vectors, one can show that a
projective plane of order q exists if and only if
q+1
q+1
q+1
2
2
2
2
h = 1, q + q + 1, (q + q + 1)
, (q + q + 1)
, . . . , (q + q + 1)
2
3
q+1
is a pure f -vector, if and only if it is a pure O-sequence.
A major open problem in geometry asks for a classification of all finite projective planes,
or even just of the possible values that q may assume. Conjecturally, q is always the power
of a prime, and it is a standard algebraic exercise, using finite field theory, to construct a
projective plane of any order q = pn .
The best general necessary condition known today on q is still the following theorem
from [7, 13]:
Theorem 3.3 (Bruck-Ryser-Chowla [7, 13]). If q is the order of a projective plane and q
is congruent to 1 or 2 modulo 4, then q is the sum of two squares.
Thus, for instance, as a consequence of the Bruck-Ryser-Chowla Theorem, no projective
plane of order 6 exists. In other words,
(1, 43, 903, 1505, 1505, 903, 301, 43)
is not a pure O-sequence. However, already ruling out the existence of projective planes
of order 10 has required a major computational effort (see Lam [41], 1991). The case
q = 12 is still open.
Notice that, at least for certain values of q, the number of nonisomorphic projective
planes of order q can be very large, and a general classification seems entirely out of reach.
The smallest q for which there exists more than one nonisomorphic projective plane is 9,
where the four possible cases were already known to O. Veblen ([77], 1907).
The second important application of pure O-sequences that we want to discuss brings
our attention to a very special class of simplicial complexes, called matroid complexes.
PURE O-SEQUENCES: KNOWN RESULTS, APPLICATIONS AND OPEN PROBLEMS
13
Matroids are ubiquitous in mathematics, where they often show up in surprising ways
(see [56, 60, 79, 80]).
The algebraic theory of matroids began in the same 1977 seminal paper of Stanley [66]
that introduced pure O-sequences. A finite matroid can be naturally identified with a
pure simplicial complex over V = {1, 2, . . . , r}, such that its restriction to any subset of
V is also a pure complex.
One associates, to any given simplicial complex ∆ over V = {1, 2, . . . , r}, the following
squarefree monomial ideal in S = K[x1 , . . . , xr ], where K is a field:
*
+
Y
I∆ = xF =
xi | F 6∈ ∆ .
i∈F
I∆ is called the Stanley-Reisner ideal of the complex ∆, and the quotient algebra S/I∆ is
its Stanley-Reisner ring.
It is a standard fact of combinatorial commutative algebra (see e.g. [69]) that the
Stanley-Reisner ring S/I∆ of a matroid complex is Cohen-Macaulay and level, although
of course of positive Krull-dimension (except in degenerate cases). Thus, the h-vector of
S/I∆ is level, since it is the h-vector of an artinian reduction of S/I∆ . However, even
though S/I∆ is presented by monomials, notice that its artinian reductions will in general
be far from monomial, for they require taking quotients by “general enough” linear forms.
The following spectacularly simple conjecture of Stanley ([66], 1977) predicts that, for
any matroid complex ∆, we can nonetheless find some artinian monomial level algebra
having the h-vector of S/I∆ as its h-vector:
Conjecture 3.4. Any matroid h-vector is a pure O-sequence.
The problem of characterizing matroid h-vectors appears to be once again hopeless, and
Conjecture 3.4 has motivated much of the algebraic work done on matroids over the past
35 years (see, as a highly nonexhaustive list, [9, 10, 15, 22, 32, 50, 57, 62, 64, 73, 74, 75]).
The main approach to Conjecture 3.4 has been, given the h-vectors of a certain class
of matroids, to explicitly produce some pure monomial order ideals having those matroid
h-vectors as their pure O-sequences. Recently, H.T. Hà, E. Stokes and the third author
[27] introduced a “more abstract” approach to Conjecture 3.4. Their main idea, inspired
by the latest progress on pure O-sequences made in BMMNZ 2012, and in particular
the proof of the Interval Conjecture (ICP) in socle degree 3, has been to try to reduce
Stanley’s conjecture, as much as possible, to one on the properties of pure O-sequences,
thus avoiding explicit construction of a monomial ideal for each matroid h-vector.
The approach of [27] has already led to a proof of Conjecture 3.4 for all matroid complexes of Krull-dimension at most 2 (the dimension 1 case, that had been the focus of a
large portion of the thesis [73], simply followed in a few lines).
Theorem 3.5 (Hà-Stokes-Zanello [27]). All matroid h-vectors (1, h1 , h2 , h3 ) are pure Osequences.
More generally, the following is a first concrete, if still tentative, general approach to
Conjecture 3.4 (see [27]).
Assuming Conjecture 3.4 holds for all matroid complexes whose deletions with respect
to any vertex are cones (which may not be too difficult to show with the techniques of
paper [27]), Conjecture 3.4 is true in general under the following two natural (but still
too bold?) assumptions:
14
JUAN MIGLIORE, UWE NAGEL, AND FABRIZIO ZANELLO
(A) Any matroid h-vector is differentiable for as long as it is nondecreasing. (In fact,
incidentally, would a g-element, that Hausel [31] and Swartz [74] proved to exist
in the “first half” of a matroid, carry on all the way?)
(B) Suppose that the shifted sum, h00 = (1, h1 + 1, h2 + h01 , . . . , he + h0e−1 ), of two pure
O-sequences h and h0 is differentiable for as long as it is nondecreasing. Then h00
is also a pure O-sequence.
4. Enumerations of pure O-sequences
As we have seen above, pure O-sequences arise in several areas, yet their properties are
not well understood, and there are other important questions that should be addressed
even if a classification is not available. For example, one would like to estimate the number
of pure O-sequences of given codimension and socle degree. What happens asymptotically? Moreover, we have seen that pure O-sequences can be as far from being unimodal
as we want. Nevertheless, one may ask: What are the odds for a pure O-sequence to be
unimodal?
In order to discuss such questions, let us denote by O(r, e), P (r, e), and D(r, e) the sets
of O-sequences, pure O-sequences, and differentiable O-sequences, respectively, that have
codimension r and socle degree e. Recall that given two functions f, g : R → R, one says
that f is asymptotic to g, and writes f (r) ∼r g(r), if limr f (r)/g(r) = 1. All limits are
taken for r approaching infinity.
Consider now an O-sequence (1, r − 1, h2 , . . . , he ) in O(r − 1, e). Integrating it, that is,
passing to (1, r, r − 1 + h2 , . . . , he−1 + he ), provides a differentiable O-sequence in D(r, e).
Thus, since finite differentiable O-sequences are pure by Theorem 2.7(b), we have the
following inclusions:
O(r − 1, e) ,→ D(r, e) ⊂ P (r, e) ⊂ O(r, e).
Results by Linusson (see [46]) imply that, for r large, the cardinalities of O(r − 1, e)
and O(r, e) are asymptotically equal. It follows that in large codimensions almost all
O-sequences are pure. More precisely, one has:
Theorem 4.1 (BMMNZ 2012). Fix a positive integer e. Then, for r large, almost all
O-sequences of socle degree e are differentiable. Namely,
e+1
#O(r, e) ∼ #P (r, e) ∼ #D(r, e) ∼ c · r( 2 )−1 ,
r
where
r
e−2
Y
ce =
i=0
r
e+1
2
−
i+1
2
e
−1
i
e+1
−1 !
2
.
Since pure O-sequences are Hilbert functions of level algebras, we immediately get the
following consequence.
Corollary 4.2 (BMMNZ 2012). Fix a positive integer e. Let L(r, e) be the set of level
Hilbert functions of codimension r and socle degree e. Then, for r large, almost all level
sequences are pure and unimodal, and
e+1
#L(r, e) ∼ c r( 2 )−1 .
r
e
The outcome changes drastically if we fix as additional parameter the socle type t.
Since each monomial of degree t is divisible by at most t distinct variables, we observe:
PURE O-SEQUENCES: KNOWN RESULTS, APPLICATIONS AND OPEN PROBLEMS
15
Proposition 4.3 (BMMNZ 2012). Let P (r, e, t) be the set of pure O-sequences of codimension r, socle degree e, and type t. Then #P (r, e, t) = 0 for r > te, this bound being
sharp.
Note that, in contrast to this result, there is no analogous restriction on level Hilbert
functions. For example, Gorenstein algebras have type 1 and admit any positive socle
degree and codimension. In fact, asymptotically, the number of their Hilbert functions is
known.
Recall that an SI-sequence of socle degree e is an O-sequence that is symmetric about 2e
and differentiable up to degree b 2e c. Initially, Stanley and Iarrobino (see [69]) had hoped
that all Hilbert functions of Gorenstein algebras were SI-sequences. Although this is not
true (see [67] for the first counterexample), it is almost true! In fact, any differentiable
O-sequence in D(r, b 2e c) can be extended to a symmetric sequence, so that the result is
an SI-sequence. Moreover, every SI-sequence is the Hilbert function of some Gorenstein
algebra (see [12, 29, 52]). Obviously, the first half of the Hilbert function of a Gorenstein
algebra is an O-sequence. Taken together, it follows that the number of Gorenstein Hilbert
functions that are not SI-sequences is negligible:
Theorem 4.4 (BMMNZ 2012). Fix a positive integer e. Let G(r, e) be the set of Gorenstein Hilbert functions of codimension r and socle degree e, and let SI(r, e) be the set of
SI-sequences of codimension r and socle degree e. Then, for r large, almost all Gorenstein
Hilbert functions are SI-sequences. More precisely,
be/2c+1
#G(r, e) ∼ #SI(r, e) ∼ c
r( 2 )−1 .
r
r
be/2c
Returning to pure O-sequences, it would be very interesting to determine, or at least to
find a good estimate of, the number of pure O-sequences of codimension r, socle degree
e, and type t. This seems a difficult problem. However, in the simplest case, where t = 1,
there is an easy combinatorial answer. In fact, any pure O-sequence of type 1 is the
Hilbert function of a (complete intersection) algebra whose inverse system is a monomial
of the form y1a1 · · · yrar , where a1 + · · · + ar = e and ai ≥ 1 for all i. We may assume that
a1 ≥ a2 ≥ · · · ≥ ar , so that (a1 , . . . , ar ) is a partition of e. Since it is easy to see that
distinct partitions lead to different Hilbert functions, we arrive at the following result.
Proposition 4.5 (BMMNZ 2012). #P (r, e, 1) = pr (e), the number of partitions of the
integer e having exactly r parts.
We now consider a slightly different asymptotic enumeration question. Fix positive
integers e and t. Since the number of monomials dividing t monomials of degree e is
finite, there are only finitely many pure O-sequences of socle degree e and type t. Denote
their set by P (∗, e, t). Determining #P (∗, e, t) exactly seems out of reach. However, one
may hope to at least find its order for t large. The first interesting case, namely e = 3,
has been settled.
Theorem 4.6 (BMMNZ 2012). Let #P (∗, 3, t) denote the number of pure O-sequences
of socle degree 3 and type t. Then
#P (∗, 3, t)
9
lim
=
.
t→∞
t2
2
Its proof gives some further information. Consider a pure O-sequence (1, r, a, t). Then
r ≤ a ≤ 3t,
by Hibi’s theorem and the fact that t monomials of degree 3 are divisible by at most
3t distinct quadratic monomials. It follows that #P (∗, 3, t) ≤ 29 t2 . To see that the two
16
JUAN MIGLIORE, UWE NAGEL, AND FABRIZIO ZANELLO
functions are in fact asymptotically equal, notice that, for fixed t, the possible values of
r and a fall into one of the following three regions, illustrated in the figure below:
Region I: t ≤ r ≤ a ≤ 3t;
Region II: 0 < r < t ≤ a ≤ 3t;
Region III: 0 < r ≤ a < t.
a
6
3t
II
I
t
III
-
t
3t
r
Using superscripts to denote the sets of pure O-sequences in each region, it is shown in
BMMNZ that
#P (I) (∗, 3, t)
#P (II) (∗, 3, t)
#P (III) (∗, 3, t)
1
=
2,
lim
=
2,
and
lim
= .
2
2
2
t→∞
t→∞
t→∞
t
t
t
2
The arguments use in a crucial way the interval property for pure O-sequences of socle
degree 3. Unfortunately, this property fails in general. Nevertheless it would be very
interesting to extend the above results to socle degree e ≥ 4.
lim
5. Open problems
In this section we collect a few interesting problems that remain open in the area of
pure O-sequences. Most of them have been discussed in the previous sections, but some
are related problems that were not addressed above.
1. What is the largest type t for which all pure O-sequences are unimodal (independently of the codimension or socle degree)? Even proving that t ≥ 2, i.e., that pure
O-sequences of type 2 are unimodal in any codimension, would be very interesting.
2. For a fixed codimension r, what is the largest type t for which all pure O-sequences
are unimodal?
3. For a fixed codimension r, what is the largest socle degree e for which all pure
O-sequences are unimodal?
4. What is the smallest codimension r for which there exists a nonunimodal pure
O-sequence of socle degree 4?
5. Determine asymptotically the number #P (∗, e, t) of pure O-sequences of socle
degree e and type t, for t large. What is the order of magnitude of #P (∗, e, t)?
6. The first example of a nonunimodal pure O-sequence (due to Stanley [66], 1977)
was (1, n = 505, 2065, 3395, 3325, 3493), which is in fact the f -vector of a CohenMacaulay simplicial complex, hence in particular a pure f -vector. What is the
PURE O-SEQUENCES: KNOWN RESULTS, APPLICATIONS AND OPEN PROBLEMS
17
smallest number of variables n (i.e., the number of vertices of the complex) allowing
the existence of a nonunimodal pure f -vector?
A. Tahat [76] has recently determined the sharp lower bound n = 328 for nonunimodal Cohen-Macaulay f -vectors of socle degree 5 (the socle degree of Stanley’s
original example), and has produced examples in larger socle degree with n as low
as 39.
7. Stanley’s Twenty-Fifth Problem for the year 2000 IMU Volume “Mathematics:
frontiers and perspectives” [70] asks, among a few other things: are all matroid
f -vectors unimodal, or even log-concave? What about matroid h-vectors?
Notice that matroid f -vectors are a much smaller subset of Cohen-Macaulay f vectors. For a major recent breakthrough on this problem, see Huh [35], and Huh
and Katz [37]; as a consequence of their work, log-concavity (hence unimodality)
is now known for all f -vectors (see Lenz [42], 2012) and h-vectors (Huh [36]) of
representable matroids. Notice also that, in general, the log-concavity of matroid
h-vectors would imply the log-concavity of matroid f -vectors, as proved by Dawson
[20] (1984) and Lenz [42].
Acknowledgements
We wish to thank David Cook II and Richard Stanley for helpful comments. The third
author also thanks Jürgen Bierbrauer for an interesting discussion on the connections
between group theory and Steiner systems. We thank the referee for a careful reading of
the paper.
References
[1] A. Björner: Nonpure shellability, f -vectors, subspace arrangements and complexity, in “Formal Power
Series and Algebraic Combinatorics,” Series Formelles et Combinatoire Algébrique (1994), DIMACS
Series in Discrete Mathematics and Theoretical Computer Science, volume 24, 1994, 25–54.
[2] M. Boij, J. Migliore, R. Miró-Roig, U. Nagel and F. Zanello: “On the shape of a pure O-sequence,”
Mem. Amer. Math. Soc. 218 (2012), no. 2024, vii + 78 pp..
[3] M. Boij and F. Zanello: Level algebras with bad properties, Proc. Amer. Math. Soc. 135 (2007), no.
9, 2713–2722.
[4] B. Boyle: “On the unimodality of pure O-sequences,” Ph.D. Thesis, University of Notre Dame (2012).
[5] H. Brenner and A. Kaid: Syzygy bundles on P2 and the Weak Lefschetz property, Illinois J. Math. 51
(2007), no. 4, 1299–1308.
[6] F. Brenti: Log-concave and unimodal sequences in algebra, combinatorics, and geometry: an update,
in “Jerusalem Combinatorics ’93,” Contemporary Mathematics 178 (1994), 71–89.
[7] R.H. Bruck and H.J. Ryser: The nonexistence of certain finite projective planes, Canadian J. Math.
1 (1949), 88–93.
[8] W. Bruns and J. Herzog: “Cohen-Macaulay rings. Rev. ed.”, Cambridge Studies in Advanced Mathematics 39, Cambridge University Press, Cambridge (1998).
[9] M.K. Chari: Matroid inequalities, Discrete Math. 147 (1995), no. 1–3, 283–286.
[10] M.K. Chari: Two decompositions in topological combinatorics with applications to matroid complexes,
Trans. Amer. Math. Soc. 349 (1997), no. 10, 3925–3943.
[11] C. Chen, A. Guo, X. Jin and G. Liu, Trivariate monomial complete intersections and plane partitions,
J. Commut. Algebra 3 (2011), no. 4, 459–490.
[12] Y.H. Cho and A. Iarrobino: Hilbert Functions and Level Algebras, J. Algebra 241 (2001), no. 2,
745–758.
[13] S. Chowla and H.J Ryser: Combinatorial problems, Canadian J. Math. 2 (1950) 93–99.
[14] C.J. Colbourn and J.H. Dinitz, Eds.: “Handbook of Combinatorial Designs,” CRC Press, Boca
Raton, FL (1996).
[15] A. Constantinescu and M. Varbaro: h-vectors of matroid complexes, preprint (2012).
18
JUAN MIGLIORE, UWE NAGEL, AND FABRIZIO ZANELLO
[16] D. Cook II: The Lefschetz properties of monomial complete intersections in positive characteristic,
J. Algebra 369 (2012) 42–58.
[17] D. Cook II and U. Nagel: Hyperplane sections and the subtlety of the Lefschetz properties, J. Pure
Appl. Algebra 216 (2012), no. 1, 108–114.
[18] D. Cook II and U. Nagel: The Weak Lefschetz Property, monomial ideals, and lozenges, Illinois J.
Math. (to appear); (arXiv:0909.3509).
[19] D. Cook II and U. Nagel: Enumerations deciding the Weak Lefschetz Property, preprint
(arXiv:1105.6062).
[20] J.E. Dawson: A collection of sets related to the Tutte polynomial of a matroid, in “Graph theory,”
Singapore (1983), Lecture Notes in Math., vol. 1073, Springer, Berlin (1984), 193–204.
[21] N. G. de Bruijn, Ca. van Ebbenhorst Tengbergen and D. Kruyswijk: On the set of divisors of a
number, Nieuw Arch. Wiskunde (2) 23 (1951), 191–193.
[22] J.A. De Loera, Y. Kemper and S. Klee: h-vectors of small matroid complexes, Electron. J. Combin.
19 (2012), P14.
[23] P. Dembowski: “Finite geometries,” Reprint of the 1968 original, Classics in Mathematics, SpringerVerlag (1997), xii + 375 pp..
[24] R. Fröberg and D. Laksov: Compressed Algebras, Conference on Complete Intersections in Acireale,
Lecture Notes in Mathematics, vol. 1092, Springer, Berlin (1984), 121–151.
[25] A.V. Geramita: Inverse systems of fat points: Waring’s problem, secant varieties and Veronese varieties and parametric spaces of Gorenstein ideals, Queen’s Papers in Pure and Applied Mathematics,
no. 102, The Curves Seminar at Queen’s (1996), vol. X, 3–114.
[26] A.V. Geramita, T. Harima, J. Migliore and Y. Shin: “The Hilbert function of a level algebra”, Mem.
Amer. Math. Soc. 186 (2007), no. 872, vi + 139 pp.
[27] H.T. Hà, E. Stokes and F. Zanello: Pure O-sequences and matroid h-vectors, Ann. Comb. (to appear);
(arXiv:1006.0325).
[28] M. Hanani: On Quadruple Systems, Canad. J. Math. 12 (1960), 145–157.
[29] T. Harima: Characterization of Hilbert functions of Gorenstein Artin algebras with the weak Stanley
property, Proc. Amer. Math. Soc. 123 (1995), no. 29, 3631–3638.
[30] T. Harima, J. Migliore, U. Nagel and J. Watanabe: The weak and strong Lefschetz properties for
Artinian K-algebras, J. Algebra 262 (2003), no. 1, 99–126.
[31] T. Hausel: Quaternionic geometry of matroids, Cent. Eur. J. Math. 3 (2005), no. 1, 26–38.
[32] T. Hausel and B. Sturmfels: Toric hyperkähler varieties, Doc. Math. 7 (2002), 495–534.
[33] J. Herzog and D. Popescu, The strong Lefschetz property and simple extensions, preprint
(arXiv:math/05065537).
[34] T. Hibi: What can be said about pure O-sequences?, J. Combin. Theory Ser. A 50 (1989), no. 2,
319–322.
[35] J. Huh: Milnor numbers of projective hypersurfaces and the chromatic polynomial of graphs, J. Amer.
Math. Soc. 25 (2012), 907–927.
[36] J. Huh: h-vectors of matroids and logarithmic concavity, preprint (arXiv:1201.2915).
[37] J. Huh and E. Katz: Log-concavity of characteristic polynomials and the Bergman fan of matroids,
Math. Ann. 354 (2012), 1103–1116.
[38] A. Iarrobino: Compressed Algebras: Artin algebras having given socle degrees and maximal length,
Trans. Amer. Math. Soc. 285 (1984), no. 1, 337–378.
[39] A. Iarrobino and V. Kanev: “Power sums, Gorenstein algebras, and determinantal loci”, Springer
Lecture Notes in Mathematics 1721, Springer, Heidelberg (1999).
[40] T.P. Kirkman: On a Problem in Combinatorics, Cambridge Dublin Math. J. 2 (1847), 191–204.
[41] C.W.H. Lam: The Search for a Finite Projective Plane of Order 10, Amer. Math. Monthly 98 (1991),
no. 4, 305–318.
[42] M. Lenz: The f -vector of a realizable matroid complex is strictly log-concave, Combinatorics Probability, and Computing (to appear); (arXiv:1106.2944).
[43] J. Li and F. Zanello: Monomial complete intersections, the Weak Lefschetz Property and plane
partitions, Discrete Math. 310 (2010), no. 24, 3558–3570.
[44] C.C. Lindner and C.A. Rodger: “Design Theory,” CRC Press, Boca Raton, FL (1997).
[45] M. Lindsey: A class of Hilbert series and the strong Lefschetz property, Proc. Amer. Math. Soc. 139
(2011), no. 1, 79–92.
[46] S. Linusson: The number of M -sequences and f -vectors, Combinatorica 19 (1999), no. 2, 255–266.
PURE O-SEQUENCES: KNOWN RESULTS, APPLICATIONS AND OPEN PROBLEMS
19
[47] F.H.S. Macaulay: Some properties of enumeration in the theory of modular systems, Proc. London
Math. Soc. 26 (1927), no. 1, 531–555.
[48] E. Mathieu: Mémoire sur l’étude des fonctions de plusieurs quantités, sur la manière de les former
et sur les substitutions qui les laissent invariables, J. Math. Pures Appl. (Liouville) (2), VI (1861),
241–323.
[49] E. Mathieu: Sur la fonction cinq fois transitive de 24 quantités, Liouville Journ. (2), XVIII (1873),
25–47.
[50] C. Merino: The chip firing game and matroid complexes, Discrete models: combinatorics, computation, and geometry (2001), Discrete Math. Theor. Comput. Sci. Proc., AA, Maison Inform. Math.
Discrete. (MIMD), Paris (2001), 245–255.
[51] J. Migliore, R. Miró-Roig and U. Nagel: Monomial ideals, almost complete intersections, and the
Weak Lefschetz property, Trans. Amer. Math. Soc. 363 (2011), no. 1, 229–257.
[52] J. Migliore and U. Nagel: Reduced arithmetically Gorenstein schemes and simplicial polytopes with
maximal Betti numbers, Adv. Math. 180 (2003), no. 1, 1–63.
[53] J. Migliore and U. Nagel: A tour of the Weak and Strong Lefschetz Properties, preprint
(arXiv:1109.5718).
[54] J. Migliore, U. Nagel and F. Zanello: A lower bound on the second entry of a Gorenstein h-vector
and a conjecture of Stanley, Proc. Amer. Math. Soc. 136 (2008), no. 8, 2755–2762.
[55] J. Migliore, U. Nagel and F. Zanello: Bounds and asymptotic minimal growth for Gorenstein Hilbert
functions, J. Algebra 210 (2009), no. 5, 1510–1521.
[56] D.L. Neel and N.A. Neudauer: Matroids you have known, Math. Mag. 82 (2009), no. 1, 26–41.
[57] S. Oh: Generalized permutohedra, h-vectors of cotransversal matroids and pure O-sequences, preprint
(arXiv:1005.5586).
[58] C. Okonek, M. Schneider and H. Spindler: “Vector bundles on complex projective spaces.” Corrected reprint of the 1988 edition. With an appendix by S. I. Gelfand. Modern Birkhäuser Classics.
Birkhäuser/Springer Basel AG, Basel (2011).
[59] P.R.J. Östergard and O. Pottonen: There exists no Steiner system S(4, 5, 17), J. Combin. Theory
Ser. A 115 (2008), no. 8, 1570–1573.
[60] J.G. Oxley: “Matroid theory,” Oxford University Press, Oxford (2006).
[61] A. Pastine and F. Zanello: Two unfortunate properties of pure f -vectors, in preparation.
[62] N. Proudfoot: On the h-vector of a matroid complex, unpublished note.
[63] L. Reid, L. Roberts and M. Roitman: On complete intersections and their Hilbert functions, Canad.
Math. Bull. 34 (1991), no. 4, 525–535.
[64] J. Schweig: On the h-Vector of a Lattice Path Matroid, Electron. J. Combin. 17 (2010), no. 1, N3.
[65] H. Sekiguchi: The upper bound of the Dilworth number and the Rees number of Noetherian local
rings with a Hilbert function, Adv. Math. 124 (1996), no. 2, 197–206.
[66] R. Stanley: Cohen-Macaulay Complexes, in “Higher Combinatorics” (M. Aigner, Ed.), Reidel, Dordrecht and Boston (1977), 51–62.
[67] R. Stanley: Hilbert functions of graded algebras, Adv. Math. 28 (1978), no. 1, 57–83.
[68] R. Stanley: Log-concave and unimodal sequences in algebra, combinatorics, and geometry, Ann. New
York Acad. Sci. 576 (1989), 500–535.
[69] R. Stanley: “Combinatorics and commutative algebra,” Second Ed., Progress in Mathematics 41,
Birkhäuser Boston, Inc., Boston, MA (1996).
[70] R. Stanley: Positivity problems and conjectures in algebraic combinatorics, in “Mathematics: frontiers and perspectives”, Amer. Math. Soc., Providence, RI (2000), 295–319.
[71] R. Stanley: “Enumerative Combinatorics,” Vol. I, Second Ed., Cambridge University Press, Cambridge (2012).
[72] R. Stanley and F. Zanello: On the rank function of a differential poset, Electron. J. Combin. 19
(2012), no. 2, P13, 17 pp..
[73] E. Stokes: “The h-vectors of matroids and the arithmetic degree of squarefree strongly stable ideals,”
Ph.D. Thesis, University of Kentucky (2007).
[74] E. Swartz: g-elements of matroid complexes, J. Combin. Theory Ser. B 88 (2003), no. 2, 369–375.
[75] E. Swartz: g-elements, finite buildings, and higher Cohen-Macaulay connectivity, J. Combin. Theory
Ser. A 113 (2006), no. 7, 1305–1320.
[76] A. Tahat: “On the nonunimodality of Cohen-Macaulay f -vectors,” M.S. Thesis, Michigan Technological University, in preparation.
20
JUAN MIGLIORE, UWE NAGEL, AND FABRIZIO ZANELLO
[77] O. Veblen and J.H.M. Wedderburn: Non-Desarguesian and non-Pascalian geometries, Trans. Amer.
Math. Soc. 8 (1907), no. 3, 379–388.
[78] J. Watanabe: The Dilworth number of Artinian rings and finite posets with rank function, in: Commutative Algebra and Combinatorics, Advanced Studies in Pure Math. Vol. 11, Kinokuniya Co. North
Holland, Amsterdam (1987), 303–312.
[79] N. White, Ed.: “Theory of Matroids,” Encyclopedia of Mathematics and Its Applications 26, Cambridge Univ. Press, Cambridge (1986).
[80] N. White, Ed.: “Matroids Applications,” Encyclopedia of Mathematics and Its Applications 40,
Cambridge Univ. Press, Cambridge (1992).
[81] F. Zanello: Extending the idea of compressed algebra to arbitrary socle-vectors, J. Algebra 270 (2003),
no. 1, 181–198.
[82] F. Zanello: Extending the idea of compressed algebra to arbitrary socle-vectors, II: cases of nonexistence, J. Algebra 275 (2004), no. 2, 730–748.
[83] F. Zanello: A non-unimodal codimension 3 level h-vector, J. Algebra 305 (2006), no. 2, 949–956.
[84] F. Zanello: Interval Conjectures for level Hilbert functions, J. Algebra 321 (2009), no. 10, 2705–2715.
Department of Mathematics, University of Notre Dame, Notre Dame, IN 46556, USA
E-mail address: [email protected]
Department of Mathematics, University of Kentucky, 715 Patterson Office Tower,
Lexington, KY 40506-0027, USA
E-mail address: [email protected]
Department of Mathematical Sciences, Michigan Technological University, Houghton,
MI 49931, USA
E-mail address: [email protected]
| 0math.AC
|
1
Optimal Power Allocation by Imperfect
Hardware Analysis in Untrusted Relaying
arXiv:1709.01334v1 [cs.CR] 5 Sep 2017
Networks
Ali Kuhestani, Student Member, IEEE, Abbas Mohammadi, Senior Member, IEEE,
Kai-Kit Wong, Fellow, IEEE, Phee Lep Yeoh, Member, IEEE,
Majid Moradikia, and Muhammad R. A. Khandaker, Member, IEEE
Abstract
By taking a variety of realistic hardware imperfections into consideration, we propose an optimal
power allocation (OPA) strategy to maximize the instantaneous secrecy rate of a cooperative wireless
network comprised of a source, a destination and an untrusted amplify-and-forward (AF) relay. We
assume that either the source or the destination is equipped with a large-scale multiple antennas
(LSMA) system, while the rest are equipped with a single antenna. To prevent the untrusted relay from
intercepting the source message, the destination sends an intended jamming noise to the relay, which is
referred to as destination-based cooperative jamming (DBCJ). Given this system model, novel closedform expressions are presented in the high signal-to-noise ratio (SNR) regime for the ergodic secrecy
rate (ESR) and the secrecy outage probability (SOP). We further improve the secrecy performance
of the system by optimizing the associated hardware design. The results reveal that by beneficially
distributing the tolerable hardware imperfections across the transmission and reception radio-frequency
(RF) front ends of each node, the system’s secrecy rate may be improved. The engineering insight is
that equally sharing the total imperfections at the relay between the transmitter and the receiver provides
the best secrecy performance. Numerical results illustrate that the proposed OPA together with the most
appropriate hardware design significantly increases the secrecy rate.
A. Kuhestani and A. Mohammadi are with the Electrical Engineering Department, Amirkabir University of Technology,
Tehran, Iran.
K.-K Wong and M. R. A. Khandaker are with the Department of Electronic and Electrical Engineering, University College
London, London WC1E 6BT, U.K.
P. L. Yeoh is with the School of Electrical and Information Engineering, The University of Sydney, NSW, Australia.
M. Moradikia is with the Electrical Engineering Department, Shiraz University of Technology, Shiraz, Iran.
2
Index Terms
Physical layer security, Untrusted relay, Hardware imperfections, Optimal power allocation, Hardware design
I. I NTRODUCTION
Security in wireless communication networks is conventionally implemented above the physical layer using key based cryptography [1]. To complement these highly complex schemes,
wireless transmitters can also be validated at the physical layer by exploiting the dynamic
characteristics of the associated communication links [2], [3]. Physical layer security (PLS) is
a promising paradigm for safeguarding fifth-generation (5G) wireless communication networks
without incurring additional security overhead [3].
Massive multiple-input multiple-output (MIMO) systems as a key enabling technology of
5G wireless communication networks provide significant performance gains in terms of spectral
efficiency and energy efficiency [4], [5]. This new technology employs coherent processing across
arrays of hundreds or even thousands of base station (BS) antennas and supports tens or hundreds
of mobile terminals [4]–[6]. As an additional advantage, massive MIMO is inherently more secure
than traditional MIMO systems, as the large-scale antenna array exploited at the transmitter can
precisely aim a narrow and directional information beam towards the intended receiver, such that
the received signal-to-noise ratio (SNR) is several orders of magnitude higher than that at any
incoherent passive eavesdropper [7]. However, these security benefits are severely hampered in
cooperative networks where the intended receivers may also be potential eavesdroppers [8]–[10].
In the context of PLS, cooperative jamming which involves the transmission of additional
jamming signals to degrade the received SNR at the potential eavesdropper can be applied by
source [9], [10], the intended receiver node [8] or a set of nodes, i.e., source and destination
or source and relay to beamform the jamming noise orthogonal to the spatial dimension of the
desired signal [11], [12]. Recently, several works have considered the more interesting scenario
of untrusted relaying [13]–[21] where the cooperative jamming is performed by the intended
receiver, which is referred to as destination-based cooperative jamming (DBCJ).
In real life, an untrusted, i.e., honest-but-curious, relay may collaborate to provide a reliable
communication. Several practical scenarios may include untrusted relay nodes, e.g., in ultra-dense
heterogenous wireless networks where low-cost intermediate nodes may be used to assist the
source-destination transmission. In these networks, it is important to protect the confidentiality
3
of information from the untrustworthy relay, while concurrently relying on it to increase the
reliability of communication. Thanks to the DBCJ strategy [8], positive secrecy rate can still
be attained in untrusted relay networks. In recent years, several works have focused on the
performance analysis [13]–[16], power allocation [17]–[20] and security enhancement [18], [21]
of untrusted relaying networks. To be specific, the authors in [17] proposed an optimal power
allocation (OPA) strategy to maximize the instantaneous secrecy rate of one-way relaying network
while two-way relaying scenario was considered in [18]. By exploiting the direct link, a sourcebased artificial noise injection scheme was proposed in [19] to hinder the untrusted relay from
intercepting the confidential message. A power allocation strategy was also proposed in [19] to
optimally determine the information and jamming signal powers transmitted by the source. In
[20], the OPA problem with imperfect channel state information was investigated. Notably, all
the aforementioned works considered perfect hardware in the communication network.
In practice, hardware equipments experience detrimental impacts of phase noise, I/Q imbalance, amplifier non-linearities, quantization errors, converters, mixers, filters and oscillators [22],
[23]. Each of the imperfections distorts the signals in its own way. While hardware imperfections
are unavoidable, the severity of the imperfections depends on the quality of the hardware used in
the radio-frequency (RF) transceivers. The non-ideal behavior of each component can be modeled
in detail for the purpose of designing compensation algorithms, but even after compensation there
remain residual transceiver imperfections [23]. This problem is more challenging especially in
high rate systems such as LTE-Advanced and 5G networks exploiting inexpensive equipments
[22]. Although most contributions in security based wireless networks have assumed perfect
transceiver hardware [8]–[21], or only investigated the impact of particular imperfections such
as I/Q imbalance [25] or phase noise [26] in the presence of an external eavesdropper, this paper
goes beyond these investigations by considering residual hardware imperfections in physical
layer security design [22]–[24].
In this paper, we take into account the OPA in a two-hop amplify-and-forward (AF) untrusted
relay network where all the nodes suffer from hardware imperfections and either the source or
the destination is equipped with large-scale multiple antennas (LSMA) [27], [28] while the other
nodes are equipped with a single antenna. The DBCJ protocol is operated in the first phase and
then the destination perfectly removes the jamming signal via self interference cancelation in
the second phase. For this system model, the main contributions of the paper are summarized
as follows:
4
•
Inspired by [22]–[24], we first present the generalized system model for transceiver hardware
imperfections in our secure transmission network. Based on this, we calculate the received
instantaneous signal-to-noise-plus-distortion-ratio (SNDR) at the relay and destination.
•
We formulate the OPA between the source and destination that maximizes the instantaneous
secrecy rate of untrusted relaying. Accordingly, novel closed-form solutions are derived for
the exact OPA. In addition, new simple solutions are derived for the OPA in the high SNR
regime.
•
According to our OPA solutions, novel compact expressions are derived for the ergodic
secrecy rate (ESR) and secrecy outage probability (SOP) in the high signal-to-noise-ratio
(SNR) regime that can be applied to arbitrary channel fading distributions. To gain further
insights, new closed-form expressions are presented over Rayleigh fading channels. The
asymptotic results highlight the presence of a secrecy rate ceiling which is basically different
from the prefect hardware case. We highlight that this ceiling phenomenon is independent
of the fading characteristic of the two hops.
•
We provide new insights for hardware design in DBCJ-based secure communications. To
this end, under the cost constraint of transceiver hardware at each node, we formulate the
hardware design problem for the aforementioned network to maximize the secrecy rate. The
results reveal that the secrecy rate can be improved by optimally distributing the level of
hardware imperfections between the transmit and receive RF front ends of each node.
Notation: We use bold lower case letters to denote vectors. IN and 0N ×1 denote the Identity
matrix and the zeros matrix, respectively. k.k, (.)H and (.)T denote the Euclidean norm, conjugate
transpose and transpose operators, respectively; Ex {·} stands for the expectation over the random
variable (r.v.) x; Pr(·) denotes the probability; fX (·) and FX (·) denote the probability density
function (pdf) and cumulative distribution function (cdf) of the r.v. X, respectively; the CN (µ, σ 2 )
denotes a circularly symmetric complex Gaussian RV with mean µ and variance σ 2 ; diag(A)
stands for the main diagonal elements of matrix A; Ei(x) is the exponential integral [34, Eq.
(8.211)]. [·]+ = max{0, x} and max stands for the maximum value.
II. S IGNAL
AND
S YSTEM R EPRESENTATION
A. System Model
As shown in Fig. 1, the system model under consideration is a wireless network with one
source (S), one destination (D) and one untrusted AF relay (R). While R is equipped with one
5
Fig. 1. Secure transmission under the presence of transceiver imperfections. The first and the second figures show the downlink
and the uplink transmissions, respectively. The relay acts as both helper and eavesdropper. The solid lines represent the first
phase of transmission while the dashed line represents the second phase of transmission.
antenna, S or D is equipped with LSMA denoted by Ns or Nd , respectively [29], [30]. This
corresponds to the downlink (DL) and uplink (UL) scenarios in a cellular system where the
base station is equipped with a LSMA and the mobile user and relay are equipped with a single
antenna [29]. In the DL scenario, S has many antennas, whereas D has a single antenna. The
opposite applies to the UL scenario.
All the nodes operate in a half-duplex mode. Accordingly, D cannot receive the transmitted
signal from S while transmitting the jamming signal and hence, the direct link between S and
D is unavailable. We also assume that the channels satisfy the reciprocity theorem [8]. In DL
transmission, the complex Gaussian channel from S to R and R to D are denoted by hsr ∼
CN (0Ns ×1 , µsr INs ) and hrd ∼ CN (0, µrd), respectively, and for UL transmission, they are denoted
by hsr ∼ CN (0, µsr ) and hrd ∼ CN (0Nd ×1 , µrdINd ), respectively. We consider slow fading such
that the channel coefficients vary independently from one frame to another frame and, they do not
change within one frame. The additive white noise ni (i ∈ {R, D}) at each receiver is represented
by a zero-mean complex Gaussian variable with variance N0 . We define the SNRs per link as
6
∆
∆
γsr =ρkhsr k2 and γrd=ρkhrd k2 and hence, the average SNRs per branch is given by γ sr = ρµsr
and γ rd = ρµrd , where ρ =
P
N0
represents the transmit SNR of the network. The maximum ratio
transmission (MRT) beamforming and maximal ratio combining (MRC) processing are applied
at the multi-antenna node to improve the overall system performance [29]. Let ν =
γsr
γrd
represent
the ratio between the source-to-relay and relay-to-destination SNRs. With LSMA, we consider
ν ≫ 1 in the DL scenario while ν ≪ 1 in the UL scenario. These two special cases are taken
into account in this paper. As observed in the numerical results, the analysis are satisfactory
even for moderate values of ν.
The DBCJ technique is applied to degrade the received signal at the untrusted relay such
that it cannot decipher the desired information. The whole transmission is performed based on
a time-division multiple-access (TDMA) based protocol such that the message transmission is
divided into two phases, i.e. the broadcast phase and the relaying phase. We consider a total
transmit power budget for S and D of P with power allocation factor λ ∈ (0, 1) such that the
transmit powers at S and D are λP and (1 − λ)P , respectively [14], [18]. As such, during the
first phase, while S transmits the intended signal with power λP , concurrently D jams with a
Gaussian noise to confuse the untrusted relay with power (1 − λ)P . For simplicity, the transmit
power at R is set to P and accordingly, in the second phase of transmission, R simply broadcasts
the amplified version of the received signal with power of P .
In order to consider the residual transceiver imperfections (after conventional compensation
algorithms have been applied) at node i, i ∈ {S, R, D}, the generalized system model from
[22] is taken into account. The imperfection at transmission and reception segments denoted
by ηit and ηir respectively, are introduced as distortion noises. The experimental results in [23]
and many theoretical investigations in [23], [31] have verified that these distortion noises are
well-modeled as Gaussian distributions due to the central limit theorem. A key property is that
the variance of distortion noise at an antenna is proportional to the signal power at that antenna
[23]. Accordingly, for the DL scenario, we have [23]
λP k t 2
S
2
2
|
)
,
diag(|h
|
...
|h
η tS ∼ CN 0,
sr1
srNs
khsr k2
t
t 2
ηD ∼ CN 0, (1 − λ)P kD ,
r
r2
ηD
∼ CN 0, P kD
|hrd |2 .
(1)
7
Furthermore, we have
ηSt ∼ CN
2
0, λP kst
,
t
(1 − λ)P kD
2
2
|
)
,
diag(|h
|
...
|h
0,
rd
rd
1
Nd
khrd k2
r2
η rD ∼ CN 0, P kD
diag(|hrd1 |2 ... |hrdNd |2 ) ,
η tD ∼ CN
2
(2)
for the UL scenario. The imperfections at R of both cases are also given by
2
ηRt ∼ CN 0, P kRt ,
i
h
2
2
r
r2
ηR ∼ CN 0, P kR λkhsr k + (1 − λ)khrd k ,
(3)
where the design parameters kit , kir > 0 for i ∈ {S, R, D} characterize the level of imperfections
in the transmitter and receiver hardware, respectively. These parameters can be interpreted as the
error vector magnitudes (EVMs). EVM determines the quality of RF transceivers and is defined
as the ratio of the average distortion magnitude to the average signal magnitude. Since the EVM
measures the joint effect of different hardware imperfections and compensation algorithms, it
can be measured directly in practice [22]. 3GPP LTE has EVM requirements in the range of
kit , kir ∈ [0.08, 0.175], where smaller values are needed to achieve higher spectral efficiencies
[24].
Remark 1 (Co-channel Interference): The analysis in this paper supports the scenario of
large enough number of interfering signals, which is typical in wireless environments where
the Gaussian assumption for the interference is valid by applying the central limit theorem [32].
B. Signal Representation
Let us denote xS and xD as the unit power information signal and the jamming signal, respectively. According to the combined impact of hardware imperfections which is well-addressed
by a generalized channel model [22], the received signal at R for DL and UL scenarios can be
expressed, respectively as
=
√
yRUL =
√
yRDL
λP wST xS
+
T
η tS
p
t T
hsr +
(1 − λ)P xD + ηD hrd + ηRr + nR ,
(4)
and
λP xS + ηSt
T
hsr +
p
T
(1 − λ)P wD
xD + η tD
T
hrd + ηRr + nR ,
(5)
8
where wS =
hsr H
khsr k
and wD =
hrd H
khrd k
represent the MRT transmit weight vectors at S and D,
respectively. Observe from (4) and (5) that the propagated distortion noises by S and D, and the
self-distortion noise at R are treated as interference at the untusted relay which is a potential
eavesdropper. As a result, the engineering insight is to beneficially forward these hardware
imperfections to make the system secure instead of injecting more artificial noise by S [9], [10],
[19], D [13]–[17], [20] or a friendly jammer [18].
Then the relay amplifies its received signal in the first phase by an amplification factor of 1
s
r
P
ρ
G=
,
(6)
=
2
E|yR |
AG λ + BG
2
2
2
t
t
where AG = (γsr − γrd)(1 + kRr 2 ) + kSt γu − kD
γv and BG = γrd(1 + kRr 2 ) + kD
γv + 1. Note
P
that in the DL scenario γu = ρ |hsri |4 /khsr k2 and γv = γrd , and in the UL scenario γu = γsr
P
and γv = ρ |hrdi |4 /khrd k2 . Then, the received signal at D for DL and UL scenarios after
self-interference (or jamming signal) cancelation are respectively, given by
√
T
t T
r
DL
,
hrd hdr + ηRt hrd + ηD
yD
=G λP wSH hsr hrd xS + GhrdnR + nD + Gη tS hsr hrd + GηRr hrd + GηD
{z
} |
|
{z
} |
{z
}
Noise
Information signal
Distortion noise
(7)
and
√
T
T
UL
yD
= G λP hsr hrd xS + Ghrd nR + nD + GηSt hsr hrd + GηRr hrd + Gη tD hrd hdr + ηRt hrd + η rD .
|
{z
} |
{z
} |
{z
}
Noise
Information signal
Distortion noise
(8)
According to (4) and (5) and after some algebraic manipulations, the SNDR at R is given by
γR =
where AR = kRr 2 ν + kSt
2 γu
γrd
t
− kD
2 γv
γrd
λν
,
AR λ + BR
(9)
t
− kRr 2 − 1 and BR = 1 + kRr 2 + kD
2 γv
γrd
+
1
.
γrd
Based on (7)
and (8) and after some manipulations, the SNDR at D can be calculated as
γD =
2
λγsr
,
AD λ + BD
(10)
2
2
2
2
2
r2 r2
r2
r2 t
r2 t
where AD = (γsr −γrd )(kD
kR + kRr 2 kRt + kR2 + kD
) + γu(kD
kS + kRt kSt + kSt ) + γv (kD
kD −
1
In our analysis, we assume that the EVMs are perfectly known and will consider estimation errors in our future work.
9
2
2
2
γu t 2
t 2
k − γγrdv kD
γrd S
t 2
+ γ1rd +
+ γγrdv kD
t
t
kRt kD
− kD
) + (ν − 1)(1 + kRr 2 ) +
2
2
2
2
r2
r2 t
t
t
kD
) + γv (kD
kD + kRt kD
+ kD
)
∆
2
r2
and BD = γrd(kRr 2 kD
+ kRr 2 kRt + kR2 +
∆
2
r2
kR2 + kD
+ 2. We define kR2 =kRt + kRr 2
2
2
t
r2
and kD
=kD
+ kD
as the total imperfection level at R and D, respectively.
Remark 2 (Perfect Hardware): The received SNRs at R and D with perfect hardware were
derived in [13], [14], [17]. When setting the level of imperfections at the nodes to zero, the
derived SNDRs in this section reduce to the special case as follows [14]
γRperfect =
λγsr
(1 − λ)γrd + 1
γDperfect =
and
λγsr γrd
.
λγsr + (2 − λ)γrd + 1
(11)
As can be seen, the mathematical structure of the derived SNDRs in (9), (10) are more complicated compared to the perfect hardware case in (11), since the terms
γu
γrd
and
γu
γrd
manifest in
the denominator. As such, it is non-trivial to propose an OPA solution for the general scenario
of imperfect hardware. This generalization is done in Section III and is a main contribution of
this work.
For DL scenario, (9) and (10) are simplified to
γR =
aL λ
λ + bL
and
γD =
cL λ
,
λ + dL
(12)
where
aL =
1
,
ξ1 − 1
bL =
τ1
,
(ξ1 − 1)ν
cL =
2
γrd
τ2 γrd + ξ1
and
2
dL =
τ3 γrd + τ4
,
ν(τ2 γrd + ξ1 )
2
2
(13)
2
2
t
r2 r2
r2
t
r2
t
t
and, τ1 = 1 + kRr 2 + kD
, τ2 = kD
kR + kRr 2 kRt + kR2 + kD
, τ3 = τ2 + kD
kD
+ kRt kD
+ kD
,
2
τ4 = 2 + kR2 + kD
and ξ1 = 1 + kRr 2 . Moreover, for UL scenario, we obtain
γR =
aS λ
1−λ
and
γD =
bS λ
,
λ + cS
(14)
where
aS =
ν
,
ξ1
bS =
γsr
(γsr − γrd )τ2 + (ν − 1)ξ1
and
cS =
τ2 γrd + ξ2
,
(γsr − γrd)τ2 + (ν − 1)ξ1
(15)
r2
and ξ2 = 2 + kR2 + kD
. Based on (12) and (14), we can conclude that although the intercept
probability is reduced by increasing the imperfection at R, the secrecy rate is also degraded.
It is, therefore, of great interest to intelligently distribute the tolerable hardware imperfections
across the transmission and reception radio frequency (RF) front ends of R (and other nodes) to
improve the secrecy rate of the network. This hardware design approach is analyzed in Section
10
VI and is a main contribution of this paper.
III. O PTIMAL P OWER A LLOCATION
This section proceeds to analyze the optimal power allocation problem with the aim of
maximizing the instantaneous secrecy rate. Extending the results in [17], [19], [20] where the
OPA was solved for perfect hardware, we investigate the power allocation factor λ under the
presence of hardware imperfections. To do so, the instantaneous secrecy rate is evaluated by [8]
i+
1 h
ln(1 + γD ) − ln(1 + γR ) .
Rs =
2 ln 2
(16)
By substituting λ = 0 into (9), (10) and then (16), we find Rs = 0. Since our goal is to distribute
the power optimally between S and D, a non-negative secrecy rate is achievable. As such, the
instantaneous secrecy rate can be reformulated as
i
1 h
ln(1 + γD ) − ln(1 + γR ) .
Rs =
2 ln 2
(17)
Given that log(·) is monotonically increasing, the maximization of Rs is equivalent to the
maximization of
∆
φ(λ) =
1 + γD
.
1 + γR
(18)
Therefore, the OPA factor λ⋆ can be obtained by solving the following constrained optimization
problem
λ⋆ = arg max
s.t.
n
o
φ(λ)
0<λ≤1
(19)
Lemma 1: f (x) is a quasi-concave function in R, if and only if [33, Section 3.4.3]
∂f (x)
∂ 2 f (x)
=0 ⇒
≤ 0.
∂x
∂x2
(20)
Based on lemma 1, we have the following corollary.
Corollary 1: f (x) is a quasi-concave function in x ∈ [x1 , x2 ], if
∂f (x)
|
∂x x=x1
and there is only one maximum over [x1 , x2 ] (despite constant functions).
> 0,
∂f (x)
|
∂x x=x2
<0
11
Proposition 1: φ(λ) is a quasiconcave function of λ in the feasible set 0 < λ ≤ 1 and the
optimal point is given by
√
bL dL (cL −aL )+ −aL bL cL dL (bL −dL )(aL dL −bL cL −bL +dL )
aL bL (cL +1)−cL dL (aL +1)
λ⋆E =
q
1 − aS (cS +1)(bS +cS +1)
bS cS
;ν ≫ 1
(21)
;ν ≪ 1
Proof: The first-order derivative of φ(λ) on λ is given by
AL λ2 +BL λ+CL
∂φ(λ) [(aL +1)λ+bL ]2 [λ+dL ]2 ; ν ≫ 1
=
,
∂λ
AS λ2 +BS2λ+CS 2
;ν ≪ 1
(22)
[1+(aL −1)λ] [λ+cL ]
where AL = −aL bL (cL +1)+cL dL (aL +1), BL = −2bL dL (aL −cL ), CL = −aL bL dL 2 +bL 2 cL dL ,
AS = (−bL (aL − 1) cL − aL (bL + 1)), BS = −2 cL (aL + bL ) and CS = −aL cL 2 + bL cL . As can
be seen from (22),
∂φ(λ)
∂λ
= 0 leads to two solutions on λ. It is easy to examine that the feasible
solution for practical values of kit and kir [22] are derived as (21). According to Corollary 1, we
find that φ(λ) is a quasiconcave function in the feasible set.
To make the further analysis tractable, we provide new compact expressions for the OPA in
the high SNR regime. In the case of DL, the expressions in (13) are simplified to
aL =
τ1
1
τ3
1
, bL =
, cL = , d L =
,
ξ1 − 1
(ξ1 − 1)ν
τ2
τ2 ν
(23)
and in the case of UL scenario, the expressions in (15) are changed to
aS =
ν
1
ν
, bS =
, cS =
.
ξ1
(ν − 1)τ2
ν −1
(24)
By substituting (23) and (24) into (21), the OPA solutions in the high SNR regime can be
expressed in the following tractable forms
θL
ν
⋆
λHigh =
1 − θS ν
where θL =
q
τ3
(τ
τ2 1
− τ3 ) +
τ3
(ξ
τ2 1
; DL
; UL
− 1) − τ3 and θS =
q
1+τ2
.
ξ1
(25)
12
IV. E RGODIC S ECRECY R ATE
In this section, we derive the ESR of the proposed secure transmission scheme in each case
of DL and UL scenarios. Since it is not straightforward to obtain a closed-form expression for
the exact ESR of DL and UL scenarios (the exact ESR includes double integral expressions
due to the complicated structures of (21)), we therefore proceed by first deriving new analytical
expressions for the ESR in the high SNR regime that can be applied to arbitrary channel fading
distributions. Based on these, new closed-form expressions are derived for the ESR in Rayleigh
fading channels. Despite prior works in the literature [14], [16]–[21] that investigated the ESR
based on perfect hardware assumption in various untrusted relaying networks, we take into
account hardware imperfections. The new results in this section generalize the recent results in
[14], [17].
The ESR as a useful secrecy metric representing the rate below which any average secure
transmission rate is achievable [2]. As such, the ESR is given by
n o
Rs = E Rs =
o
n
oi
1 h n
E ln(1 + γD ) − E ln(1 + γR ) .
2 ln 2 |
{z
} |
{z
}
T1
(26)
T2
In the following, we proceed to evaluate the parts T1 and T2 and then Rs for each case of DL
and UL scenarios. Towards this goal, we present the following useful lemma.
Lemma 2: For positive constants α1 , α2 and α3 , and non-negative r.v. Γ, the cdf of the new
b = α1 Γ is derived as
r.v. Γ
α2 Γ+α3
FΓb (x) =
FΓ α3 x
α1 −α2 x
1
;0 ≤ x <
;x ≥
α1
α2
(27)
α1
α2
Proof: We start from the definition of the cdf as follows
FΓb (x) = Pr
n
o
n
o
α1 Γ
≤ x = Pr Γ(α1 − α2 x) ≤ α3 x ,
α2 Γ + α3
where the last probability equals to one for α1 − α2 x < 0. Otherwise, it equals to FΓ
(28)
α3 x
α1 −α2 x
.
13
1) Downlink Scenario: By plugging (25) into (12), we obtain
θL
,
θL (ξ1 − 1) + τ1
θL γrd
γD =
.
(τ2 θL + τ3 )γrd + ξ1 θL + τ4
γR =
(29)
(30)
We find that all the terms in (29) and (30) are deterministic constants which leads to the secrecy
rate ceiling in the high SNR regime.
Based on lemma 2 and (30), the part T1 in (26) is given by
n
T1 = E ln 1 +
o
θL γrd
=
(τ2 θL + τ3 )γrd + ξ1 θL + τ4
Z
θL
τ2 θL +τ3
0
1 − Fγrd
(ξ1 θL +τ4 )x
θL −(τ2 θL +τ3 )x
1+x
dx, (31)
where the last equation follows from the integration by parts. The expression in (31) is straightforwardly evaluated for any channel fading distribution, either directly or by a simple numerical
integration.
Furthermore, based on (30) the part T2 is a constant value as
T2 = ln 1 +
θL
.
θL (ξ1 − 1) + τ1
(32)
We conclude from (32) that the amount of information leakage is independent of the transmit
SNR and the position of the relay, and only depends on the EVMs at network nodes. By replacing
(31) and (32) into (26), the compact ESR expression is achieved for any channel distribution.
For the case of Rayleigh fading, due to the fact that γrd is an exponential r.v. and applying
[34, Eq. (4.337.2)], the part T1 can be expressed in a closed-form solution. By substituting this
and (32) into (26), the closed-form ESR expression becomes
Rs
DL
=
where r1 =
i
1
1
1
θL
1 h r γ1
e 2 rd Ei(−
,
) − e r1 γ rd Ei(−
) − ln 1 +
2 ln 2
r2 γ rd
r1 γ rd
θL (ξ1 − 1) + τ1
(1+τ2 )θL +τ3
ξ1 θL +τ4
and r2 =
τ2 θL +τ3
.
ξ1 θL +τ4
(33)
We conclude from (33) that the ESR is exclusively
characterized by the level of imperfections over nodes and γ rd which is a function of the transmit
SNR and the distance-dependent channel gain µrd.
14
2) Uplink Scenario: Substituting (25) into (14) yields
1
λ⋆H ν
≈p
,
⋆
(1 − λH )ξ1
ξ1 (1 + τ2 )
γsr
q
.
γD ≈
2
)γ
+
ξ
−
ξ
τ2 (1 + 1+τ
sr
2
1
ξ1
γR =
(34)
(35)
Observe from (34) and (35) that only the first hop SNR, γsr contributes to the secrecy rate
performance. Following the similar procedure as the DL scenario, the ESR performance of the
UL case for arbitrary fading distribution can be expressed as
Rs
UL
1
=
2 ln 2
Z
τ2 (1+
1
r
1+τ2
)
ξ1
1 − Fγsr (
(ξ2 −ξ1 )x
q
1−τ2 (1+
1+x
0
1+τ2
)x
ξ1
)
dx − ln(1 + p
1
) .
ξ1 (1 + τ2 )
(36)
For the case of Rayleigh fading, the closed-form ESR expression is given by
Rs
where t1 =
UL
=
i
1
1
1
1
1 h t 1γ
,
e 2 sr Ei(−
) − e t1 γsr Ei(−
) − ln 1 + p
2 ln 2
t2 γ sr
t1 γ sr
ξ1 (1 + τ2 )
1+τ2 (1+
q
1+τ2
)
ξ1
ξ2 −ξ1
and t2 =
τ2 (1+
q
1+τ2
)
ξ1
ξ2 −ξ1
(37)
. It is observed from (37) that the ESR is entirely
determined by the average channel gain of the first hop, the transmit SNR and the level of
imperfections of all the network nodes. We also find that increasing the number of antennas at
D has no impact on the ESR when Nd is large.
V. S ECRECY O UTAGE P ROBABILITY
In this section, similar to our ESR results, general expressions are first presented for the SOP
that can be applied to any channel distribution, under the presence of transceiver imperfections
and in the high SNR regime. Based on these, we derive novel closed-form expressions for the
SOP in Rayleigh fading channels.
The SOP denoted by Pso is a criterion that determines the fraction of fading realizations
where a secrecy rate Rt cannot be supported [13]. Accordingly, the overall SOP is defined as
the probability that a system with the instantaneous secrecy rate Rs is not able to support the
n
o
target transmission rate Rt ; Pso = Pr Rs < Rt .
In the following, we focus on each case of DL and UL, respectively.
15
1) Downlink Scenario: Substituting (30) into (17) and then based on the SOP definition we
obtain
θL γrd
ft
≤R
PsoDL = Pr
(τ2 θL + τ3 )γrd + ξ1 θL + τ4
1+ θL
ft
τ2 θL +τ3
(ξ1 θL +τ4 )R
1
Fγ
;
R
<
log
t
2
rd
ft
2
1+γR
θL −(τ2 θL +τ3 )R
=
1+ θL
τ2 θL +τ3
1
1
; Rt ≥ 2 log2
1+γR
(38)
ft = 22Rt (1+γR)−1 and γR is in (29), and the last equation follows from using lemma 2.
where R
It is worth pointing out that the SOP expressions in (38) allows the straightforward evaluation of
the SOP for any channel fading distribution by a simple numerical integration. We can conclude
from (38) that the SOP is always 1 for target transmission rates more than a threshold (which
only depends on the EVMs of the nodes). Interestingly, this event holds for any channel fading
distribution, any network topology and any transmit SNR. Therefore as explained in Section IV,
some secrecy rates can never be achieved due to secrecy rate ceiling. Furthermore, we conclude
that for target transmission rates smaller than the threshold, Pso approaches zero with increasing
SNR (similar to perfect hardware) whereas the SOP always equals one for target transmission
rates larger than the threshold. This result is fundamentally different to the perfect hardware
case where the SOP goes to zero with increasing SNR and for any target transmission rate [13],
[15], [17].
For Rayleigh fading channels, γrd is an exponential r.v. and therefore, our new and simple
closed-form SOP expression in the presence of transceiver hardware imperfection is given by
PsoDL =
1 − exp −
1
ft
(ξ1 θL +τ4 )R
ft )γ
(θL −(τ2 θL +τ3 )R
rd
; Rt < 21 log2
; Rt ≥ 21 log2
1+
1+
θL
τ2 θL +τ3
1+γR
θL
τ2 θL +τ3
1+γR
(39)
We note that the results of this section generalize the results of [17] which were derived for the
case of untrusted relaying with perfect hardware.
2) Uplink Scenario: Similar to the DL case, the SOP can be obtained by substituting γD in
16
(35) into (17) and using the SOP definition. This yields
1
1+ τ2 (1+r 1+τ2 )
f
ξ1
(ξ2 −ξ1 )Rt
q
Fγsr
; Rt < 21 log2
1+τ2 f
1+γR
1−τ2 (1+
)
R
UL
t
ξ1
Pso =
1
1+ τ2 (1+r 1+τ2 )
ξ1
1
; Rt ≥ 21 log2
1+γR
(40)
the closed-form SOP is derived as
1
1+ τ2 (1+r 1+τ2 )
f
ξ1
−(ξ
−ξ
)
R
t
1
2
1
q
1 − exp
; Rt < 2 log2
1+τ2 f
1+γR
(1−τ
(1+
)
R
)γ
t sr
2
ξ1
PsoUL =
1
1+ τ2 (1+r 1+τ2 )
ξ1
1
; Rt ≥ 12 log2
1+γR
(41)
ft = 22Rt (1 + γR ) − 1 and γR is in (34). For the special case of Rayleigh fading channels,
where R
As observed in the numerical results, the closed-form expressions (39) and (41) are sufficiently
tight at medium and high transmit SNRs.
VI. H ARDWARE D ESIGN
In this section, we aim to gain some insights into the design of RF hardware to maximize the
secrecy rate in untrusted relaying networks. Depending on the fixed cost of each network node,
we show how the RF segments at the transmission and reception front ends can be designed.
Specially, given the maximum tolerable hardware imperfection of each node, we derive new
analytical results characterizing how the hardware imperfections should be distributed between
the transmission RF segment and the reception RF segment of each node to maximize the secrecy
rate. Therefore, we should find kit and kir , i ∈ {S, R, D} to maximize the secrecy rate such that
kit + kir = kitot . Mathematically speaking, our goal is to solve the following optimization problem
t
r
(kRt , kRr , kD
, kD
) = arg max φ(λ⋆ )
(42)
s.t. kRt + kRr = kRtot
t
r
tot
kD
+ kD
= kD
Based on (30), (35) and (17), the instantaneous secrecy rate is an increasing function of the
transmit SNR. Since it is our aim to achieve high transmission rates, we consider the asymptotic
SNR regime ρ → ∞ [22] to solve the hardware design problem (42). As observed in numerical
studies, the results of the high SNR analysis can be applied successfully at finite SNRs.
17
In the asymptotic SNR regime and for any random distributions on γsr and γrd, the asymptotic
received SNDRs at R and D are respectively, given by
θL
; DL
θL (ξ1 −1)+τ1
∞
,
γR =
√ 1
; UL
(43)
ξ1 (1+τ2 )
and
γD∞
=
τ
θL
2 θL +τ3
; DL
1
q
; UL
τ2 (1+
.
1+τ2
)
ξ1
By substituting (43) and (44) into (18), the secrecy rate ceiling is given by
((1+τ2 )θL +τ3 )((ξ1 −1)θL +τ1 )
; DL
(ξ1 θL +τ1 )(τ2 θL +τ3 )
∞
√
√ √
φ =
√ξ1 (τ2√+1)(τ2 +√ ξ1√ τ2 +1)
; UL
τ2 ( ξ1 + τ2 +1)( ξ1 τ2 +1+1)
(44)
(45)
Some conclusions and insights can be concluded from (45). First, the secrecy rate ceiling
event appears in the asymptotic SNR regime, which significantly limits the performance of the
system. This event is different from the perfect hardware case, in which the ESR increases with
increasing SNR. Note that this ceiling effect is independent of the fading distribution.
In the following, we focus on the of DL and UL scenarios separately and then conclude about
the hardware design of the overall network.
A. Downlink Scenario:
In the following, we proceed to solve the optimization problem (42) by independently discussing on the hardware design at R and D as follows.
Proposition 2: Suppose kRt + kRr = kRtot , hence the secrecy rate ceiling is maximized if kRt =
kRr =
tot
kR
.
2
Proof: Please see Appendix A.
t
r
tot
Proposition 3: Suppose kD
+ kD
= kD
, thus the secrecy rate ceiling is maximized if
q
2
tot 2
tot 2
tot 4
tot 2
+4kD
+12kR2 −4kD
+9
2kR+2kD +3− 4kR4 +8kR2 kD
t
kD =
.
tot
4kD
(46)
18
Proof: Please see Appendix B.
B. Uplink Scenario:
Similar to DL scenario, two propositions are provided as follows.
Proposition 4: Suppose kRt + kRr = kRtot , thus the secrecy rate ceiling is maximized if kRt =
kRr =
tot
kR
.
2
Proof: Please see Appendix C.
t
r
tot
Proposition 5: Suppose kD
+ kD
= kD
, hence the secrecy rate ceiling is a monotonically
r
decreasing function of kD
.
Proof: In this case, we have
∂φ∞ ∂τ2
∂φ∞
=
,
r
r
∂kD
∂τ2 ∂kD
where
∂τ2
r
∂kD
r
r
+2kD
> 0 and
= 2kRr 2 kD
∂φ∞
∂τ2
in (67) is negative in the feasible set. As such,
(47)
∂φ∞
r
∂kD
< 0.
Based on Propositions 2–5, we provide the following corollary as a conclusion of the analysis
which provides new insights into the system design.
Corollary 2: Consider a cooperative network in which one multiple antennas node communicates with a single antenna node via a single antenna untrusted relay. Let us assume a predefined
cost can be assigned to each node. To maximize the secrecy rate of this network the following
considerations should be taken into account:
•
According to Propositions 2 and 4, the total cost for the relay node should be divided by
half between the transmission and reception RF front ends, i,e., it is better to apply the same
level of imperfections at every transceiver chain, instead of utilizing a mix of high-quality
and low-quality transceiver chains.
•
According to Proposition 5, to design the multiple antennas node, the designers are persuaded to use higher-quality hardware in reception RF front end and lower-quality hardware
19
4
LSMA−S, Exact OPA Simulation
LSMA−D, Exact OPA Simulation
LSMA−S, Approximate OPA using eq. (33)
LSMA−D, Approximate OPA using eq. (37)
LSMA−S, EPA
LSMA−D, EPA
Secrecy rate ceiling using eq. (45)
Ergodic Secrecy Rate (bits/s/Hz)
3.5
3
2.5
Perfect hardware
2
1.5
1
Imperfect hardware
0.5
0
0
5
10
15
20
25
30
35
40
45
Transmit SNR [dB]
Fig. 2. Ergodic secrecy rate versus transmit SNR for exact and the derived closed-form expressions under perfect and imperfect
transceiver hardwares. Number of antennas at source or destination is set to 16. For imperfect case with k = 0.1, the secrecy
rate ceiling is observed.
in the transmission RF front end, i.e, the hardware imperfections at the reception end of
the multiple antennas node should be close to zero.
•
According to Proposition 3, to design the single antenna node, the quality of RF requirements
at the transmission end should obeys from (46). As observed in the numerical examples,
t
we obtain kD
>
tot
kD
2
for typical values of EVMs.
VII. N UMERICAL R ESULTS
AND
D ISCUSSIONS
In this section, numerical results are provided to verify the accuracy of the derived closedform expressions in Section IV and V for LSMA at S (LSMA-S) and LSMA at D (LSMA-D),
respectively, and also the cases of multiple antennas at S (MA-S) and multiple antennas at D
(MA-D). We compare our LSMA-based ESR performance with the exact ESR with Monte-Carlo
simulations where the OPA is numerically evaluated for finite numbers of antennas using the
bisection method. In addition, the equal power allocation (EPA) between S and D (i.e., λ = 0.5)
is plotted as a benchmark. Furthermore, the concepts of secrecy rate ceiling and the practical
hardware insights from Section VI are numerically presented. In our numerical evaluations, the
20
2.4
2.2
Ergodic Secrecy Rate (bits/s/Hz)
Case of k=0.05
2
1.8
1.6
1.4
1.2
Case of k=0.1
1
0.8
MA−S, Exact OPA Simulation
MA−D, Exact OPA Simulation
MA−S, Approximate OPA using eq. (33)
MA−D, Approximate OPA using eq. (37)
0.6
0.4
0
5
10
15
20
25
30
35
40
45
Transmit SNR [dB]
Fig. 3. Ergodic secrecy rate versus transmit SNR for exact and the derived closed-form expressions under different levels of
hardware imperfections; k ∈ {0.05, 0.1}. Number of antennas at source or destination is set to 4.
transmission links between nodes are modeled by the Rayleigh fading channel and the average
channel gains are specified as µsr = µrd = 10. Moreover, for LSMA the number of antennas is
set to 16, and for MA the number of antennas is set to 4.
Fig. 2 depicts the ESR versus transmit SNR ρ in dB for both cases of DL and UL and for
t
r
perfect (k = kRt = kRr = kSt = kD
= kD
= 0) and imperfect (k = 0.1) cases. The number of
antennas at S and D are set to Ns = Nd = 16. It is observed from the figure that the Monte-Carlo
simulation of the exact OPA evaluated using the bisection method is in good agreement with
the derived high SNR closed-form solutions in (33) and (37) for both perfect and imperfect
hardwares. In contrast to perfect hardware, the figure shows that the ESR ceiling phenomenon
occurs for imperfect hardware which reveals the performance limits of hardware-constrained
realistic networks in the high SNR regime. This figure also reveals that hardware imperfections
have low impact at low SNRs, but are significant in the high SNR regime. Furthermore, it is
observed that the proposed OPA increases the secrecy rate floor by approximately 1 bits/s/Hz
and 0.9 bits/s/Hz for DL and UL scenarios, respectively compared to the EPA (λ = 0.5).
In Fig. 3, we examine the accuracy of the derived closed-form solutions for MA-S and MA-D
by considering Ns = Nd = 4. As can be seen, the numerical and the theoretical curves are in
21
0
10
OPA, R =2 bits/s/Hz
t
EPA, R =0.25 bits/s/Hz
−1
Secrecy Outage Probability
10
t
OPA, R =1 bits/s/Hz
t
−2
10
OPA, Rt=0.25 bits/s/Hz
−3
10
−4
10
0
Imperfect, Exact OPA Simulation
Perfect, Exact OPA Simulation
Imperfect, Approximate OPA using eq. (39)
Perfect, Approximate OPA using eq. (39)
Imperfect, EPA
Perfect, EPA
5
10
15
20
25
30
35
40
Transmit SNR [dB]
Fig. 4. Secrecy outage probability versus transmit SNR for DL transmission, with different target transmission rates and under
perfect (k = 0) and imperfect hardware (k = 0.1).
good agreement across all SNR regimes. Moreover, it is observed that by increasing the level
of hardware imperfections from k = 0.05 to k = 0.1, the achievable secrecy rate is degraded
approximately 1 bits/s/Hz in the high SNR regime.
Figs. 4 and 5 show the SOP as a function of the transmit SNR for LSMA based DL and UL
scenarios, respectively, and for different target secrecy rates. The theoretical curves were plotted
by the derived analytical expressions in (39) and (41) which are well-tight with the marker
symbols generated by the Monte-Carlo simulations. As observed from these figures, there is
only a negligible performance loss caused by transceiver hardware imperfections in the low
target secrecy rate of Rt = 0.25 bits/s/Hz, but by increasing the target secrecy rate to Rt = 1
bits/s/Hz or Rt = 2 bits/s/Hz, substantial performance loss is revealed. Interestingly, for Rt = 2
bits/s/Hz, the network with imperfect hardware is always in outage and secure communications is
unattainable-irrespective of the transmit SNR. This is exactly predicted by our analytical results
in section V. The reason is that this target secrecy rate is more than the derived thresholds
in (39) and (41), and as mentioned, the SOP of the system always equals one for Rt more
the thresholds. It can also be seen from the figures that despite the OPA technique that the
SOP curves with imperfect hardware and with perfect hardware have the same slope (and thus,
22
0
10
OPA, R =2 bits/s/Hz
t
−1
10
Secrecy Outage Probability
EPA, R =0.25 bits/s/Hz
t
OPA, R =1 bits/s/Hz
−2
t
10
−3
10
OPA, Rt=0.25 bits/s/Hz
−4
10
−5
10
0
Imperfect, Exact OPA Simulation
Perfect, Exact OPA Simulation
Perfect, Approximate OPA using eq. (41)
Perfect, Approximate OPA using eq. (41)
Imperfect, EPA
Perfect, EPA
5
10
15
20
25
30
35
40
Transmit SNR [dB]
Fig. 5. Secrecy outage probability versus transmit SNR for UL transmission, with different target transmission rates and under
perfect (k = 0) and imperfect hardware (k = 0.1).
hardware imperfections lead to only an SNR offset which is unveiled as a curve shifting to the
right), the SOP performance of the EPA technique approaches a non-zero saturation value in
the high SNR regime for imperfect hardware. This observation reveals the secrecy performance
advantage of the proposed OPA scheme compared with EPA.
Finally, we provide Figs. 6 and 7 to illustrate the insights for designing practical systems that
were presented in Section VI. In the simulation, we assume that the total hardware imperfection
tot
over each node equals to 0.2, i.e., kRtot = kD
= 0.2. Based on Propositions 2 and 4, to maximize
the secrecy rate, we should design the transmission and reception RF front ends at R such that
t
r
kRt = kRr = 0.1. For LSMA at S, based on Proposition 3, we obtain kD
= 0.13 and kD
= 0.07
while for LSMA at D and based on Proposition 5, we should design the hardwares such that
t
r
t
r
kD
= 0.2 and kD
= 0. By defining the hardware imperfection vector as IV = [kRt , kRr , kD
, kD
],
we consider the following four different hardware design schemes:
•
Design 1: R and D are designed randomly, for example IV= [0.15, 0.05, 0.1, 0.1],
•
Design 2: R is designed optimally based on Propositions 2 and 4 while D is designed
randomly; IV= [0.1, 0.1, 0.1, 0.1],
•
Design 3: R is designed randomly while D is designed optimally; For LSMA at S, IV=
23
1.6
Ergodic Secrecy Rate (bits/s/Hz)
1.4
1.2
OPA, Design 1
OPA, Design 2
OPA, Design 3
OPA, Design 4
EPA, Design 4
1
0.8
0.6
0.4
0.2
0
0
5
10
15
20
25
30
35
40
45
Transmit SNR [dB]
Fig. 6. Ergodic secrecy rate versus transmit SNR for LSMA at S. Various imperfection distributions over RF transmission and
tot
tot
reception ends are considered for kR
= kD
= 0.2.
[0.15, 0.05, 0.13, 0.07] and for LSMA at D, IV= [0.15, 0.05, 0.2, 0], and
•
Design 4: R and D are designed optimally; For LSMA at S, IV= [0.1, 0.1, 0.13, 0.07] and
for LSMA at D, IV= [0.1, 0.1, 0.2, 0].
The results depict that the hardware design 4 which is based on Propositions 2-5 provides
higher ESR performance compared to the case of random hardware design (Design 1) and
the cases of optimizing only one node (Designs 2 and 3). Furthermore, they show that the
analysis presented in Section VI (which was based on high SNR analysis), can be utilized
auspiciously at medium SNRs. In addition, as can be seen from these figures and mentioned
before, different hardware designs have the ESR performance close together at low SNR regime,
while the difference between the ESR performance of the designs is large at high SNR regime.
Finally, we can understand from the figure that the proposed OPA together with Design 4
significantly outperforms the scenario of EPA with Design 4.
VIII. C ONCLUSION
Physical radio-frequency (RF) transceivers are inseparable segments in both traditional and
new emerging wireless networks. In the literature, very few works have considered the impact
24
2
Ergodic Secrecy Rate (bits/s/Hz)
1.8
1.6
1.4
OPA, Design 1
OPA, Design 2
OPA, Design 3
OPA, Design 4
EPA, Design 4
1.2
1
0.8
0.6
0.4
0
5
10
15
20
25
30
35
40
45
Transmit SNR [dB]
Fig. 7. Ergodic secrecy rate versus transmit SNR for LSMA at D. Various imperfection distributions over RF transmission and
tot
tot
reception ends are considered for kR
= kD
= 0.2.
of hardware imperfections on security based transmissions and little is understood regarding
this impact on untrusted relaying networks. In this paper, by taking hardware imperfections
into consideration, we proposed an optimal power allocation (OPA) strategy to maximize the
instantaneous secrecy rate of a cooperative wireless network comprised of a source, a destination
and an untrusted amplify-and-forward (AF) relay. Based on our OPA solutions, new closed-form
expressions were derived for the ergodic secrecy rate (ESR) and secrecy outage probability (SOP)
with Rayleigh fading channels. The expressions effectively characterize the impact of hardware
imperfections and manifest the existence of a secrecy rate ceiling that cannot be enhanced by
increasing SNR or improving fading conditions. They also illustrate that hardware imperfections
have low impact at low SNRs, but are significant in the high SNR regime. This issue reveals that
hardware imperfections should be taken into account when developing high rate systems such as
LTE-Advanced and 5G networks. To improve the secrecy performance of the network, we finally
presented the hardware design approach. Numerical results depict that optimally distributing the
hardware imperfections between the transmission and reception RF segments can further improve
the secrecy performance.
25
A PPENDIX A
Let take the first-order derivative of φ∞ on kRt using the chain rule in partial derivations as
follows
∂φ∞ ∂θL ∂τ1
∂θL ∂τ2
∂θL ∂τ3
∂θL ∂ξ1
∂φ∞
=
+
+
+
∂kRt
∂θL ∂τ1 ∂kRt
∂τ2 ∂kRt
∂τ3 ∂kRt
∂ξ1 ∂kRt
∂φ∞ ∂τ1
∂φ∞ ∂τ2
∂φ∞ ∂τ3
∂φ∞ ∂ξ1
+
+
+
+
,
∂τ1 ∂kRt
∂τ2 ∂kRt
∂τ3 ∂kRt
∂ξ1 ∂kRt
(48)
where using (45), we obtain
κ1 θL 2 + κ2 θL + κ3
∂φ∞
=
,
∂θL
(θL ξ1 + τ1 )2 (τ2 θL + τ3 )2
τ3
∂θL
= p
,
∂τ1
2 τ2 τ3 (τ1 − τ3 )
p
τ3 (τ1 − τ3 ) τ3 (ξ1 − 1)
∂θL
,
=−
−
√
∂τ2
2τ2 τ2
τ22
ξ1 − 1
τ1 − 2τ3
∂θL
+
= p
− 1,
∂τ3
τ2
2 τ2 τ3 (τ1 − τ3 )
τ3
∂θL
= ,
∂ξ1
τ2
∞
θL (τ2 θL + τ3 + θL )
∂φ
=
,
∂τ1
(θL ξ1 + τ1 )2 (τ2 θL + τ3 )
∂φ∞
θ2 (ξ1 θL + τ1 − θL )
=− L
,
∂τ2
(τ2 θL + τ3 )2 (ξ1 θL + τ1 )
θL (ξ1 θL + τ1 − θL )
∂φ∞
=−
,
∂τ3
(τ2 θL + τ3 )2 (ξ1 θL + τ1 )
∂φ∞
θL2 (τ2 θL + τ3 + θL )
=
,
∂ξ1
(ξ1 θL + τ1 )2 (τ2 θL + τ3 )
(49)
(50)
(51)
(52)
(53)
(54)
(55)
(56)
(57)
26
where κ1 = −τ1 τ2 (τ2 + 1) + τ3 ξ1 (ξ1 − 1), κ2 = −2τ1 τ3 (τ2 − ξ1 + 1) and κ3 = τ1 τ3 (τ1 − τ3 ). By
substituting kRr = kRtot − kRt into (45), we obtain
∂τ1
∂kRt
∂τ2
∂kRt
∂τ3
∂kRt
∂ξ1
∂kRt
= −2kRtot + 2kRt ,
(58)
2
r 2 tot
= −2kD
(kR − kRt ) − 2(kRtot − kRt )kRt + 2(kRtot − kRt )2 kRt + 4kRt − 2kRtot ,
2
(59)
2
r 2 tot
t
= −2kD
(kR − kRt ) − 2(kRtot − kRt )kRt + 2(kRtot − kRt )2 kRt + 4kRt − 2kRtot + 2kRt kD
, (60)
= −2kRtot + 2kRt .
(61)
Substituting (49)–(61) into (48) and after tedious manipulations yields
r2
4(1 − kD
)(kRtot − 2kRt )
∂φ∞
=
2 .
∂kRt
t 2
t tot
tot 2
r2
t 2
4kR − 4kR kR + 2kR + 2kD + kD
(62)
Expression (62) shows that φ∞ is a concave function of kRt in the feasible set and kRt =
the single solution to
∂φ∞
t
∂kR
tot
kR
2
is
= 0.
A PPENDIX B
Following the similar approach in Proposition 2, we should evaluate
∂φ∞
t .
∂kD
r
Let substitute kD
=
tot
t
kD
− kD
into τ1 , τ2 , τ3 and then compute the following derivations
∂τ1
∂τ2
tot
t
tot
t
= 1,
= −2kRr 2 (kD
− kD
) − 2kD
+ 2kD
,
(63)
t
t
∂kD
∂kD
∂τ3
2
tot
t
tot
t
t
t
tot
t 2
t 2 tot
t
= −2kRr 2 (kD
− kD
) − 2kD
+ 2kD
+ 2kD
(kRt + 1) + 2kD
(kD
− kD
) − 2kD
(kD − kD
).
t
∂kD
(64)
The expression
∂φ∞
t
∂kD
t
can be obtained similar to (48) by changing kRt to kD
. Then by substituting
(49)–(57) and (63), (64) into
∞
∂φ
t
∂kD
∂φ∞
t ,
∂kD
and after manipulations, we obtain
t
t 2 tot
t tot 2
t
tot
2 2kR2 kD
− 2kD
kD + 2kD
kD + 3kD
− 2kD
=−
.
2
2
t 2
t tot
tot 2
2kR + 3kD − 4kD kD + 2kD
(65)
t
It is straightforward to see that (65) is a concave function of kD
in the feasible set and the single
solution to
∂φ∞
t
∂kD
= 0 is simply calculated.
27
A PPENDIX C
We can write
∂φ∞ ∂τ2
∂φ∞ ∂ξ1
∂φ∞
=
+
.
∂kRr
∂τ2 ∂kRr
∂ξ1 ∂kRr
(66)
Using (45) yields
∞
∂φ
∂τ2
∂φ∞
∂ξ1
i
h
p
ξ1 (2 (τ2 +2) ξ1−τ2 2 ) ξ1 (1+τ2 )+2ξ1(τ2 (ξ1+1−τ2 )+ξ1+1)
,
=−
2
2
√ √
√ √
2τ2 2
ξ1 1+τ2+ 1 ξ1+ ξ1 1+τ2
h
i
p
(1 + τ2 ) (τ2 + 2ξ1) ξ1 (1 + τ2 ) + 2 (1 + τ2 ) ξ1
=
2
2 .
√ √
√ √
2τ2
ξ1 1 + τ2 + 1 ξ1 + ξ1 1 + τ2
(67)
(68)
Considering kRt = kRtot − kRr , one can obtain
∂τ2
2
r2
= 4kRr 3 − 6kRr 2 kRtot + (2kD
+ 2kRtot + 4)kRr − 2kRtot ,
∂kRr
∂ξ1
= 2kRr .
r
∂kR
By substituting (67)–(70) into (66) and solving
∂φ∞
r
∂kR
= 0 yields kRr =
(69)
(70)
tot
kR
.
2
Acknowledgements
The authors would like to thank Prof. Lajos Hanzo for helpful comments to improve the paper.
R EFERENCES
[1] A. Mukherjee, S. A. A. Fakoorian, J. Huang, and A. L. Swindlehurst, “Principles of physical-layer security in multiuser
wireless networks: A survey,” IEEE Commun. Surveys and Tutorials, vol. 16, no. 3, pp. 3062-3080, Feb. 2014.
[2] X. Chen, D. W. K. Ng, W. Gerstacker, and H-H. Chen, “A Survey on Multiple-Antenna Techniques for Physical Layer
Security,” IEEE Commun. Surveys Tuts., doi: 10.1109/COMST.2016.2633387.
[3] N. Yang, L. Wang, G. Geraci, M. Elkashlan, J. Yuan, and M. D. Renzo, “Safeguarding 5G wireless communication networks
using physical layer security,” IEEE Commun. Mag., vol. 53, no. 4, pp. 20-27, Apr. 2015.
[4] F. Rusek, D. Persson, B. K. Lau, E. G. Larsson, T. L. Marzetta, O. Edfors, and F. Tufvesson, “Scaling up MIMO:
Opportunities and challenges with very large arrays,” IEEE Sig. Proc. Mag., vol. 30, no. 1, pp. 40-46, Jan. 2013.
[5] E. G. Larsson, F. Tufvesson, O. Edfors, and T. L. Marzetta, “Massive MIMO for next generation wireless systems,” IEEE
Commun. Mag., vol. 52, no. 2, pp. 186-195 , Feb. 2014.
[6] T. L. Marzetta, “Noncooperative cellular wireless with unlimited numbers of BS antennas,” IEEE Trans. Wireless Commun.,
vol. 9, no. 11, pp. 3590-3600, Nov. 2010.
[7] D. Kapetanovic, G. Zheng, and F. Rusek, “Physical layer security for massive MIMO: An overview on passive
eavesdropping and active attacks,” IEEE Commun. Mag., vol. 53, no. 6, pp. 21-27, Jun. 2015.
28
[8] X. He and A. Yener, “Two-hop secure communication using an untrusted relay: A case for cooperative jamming,” in Proc.
IEEE Globecom, New Orleans, LA, Dec. 2008, pp. 15.
[9] M. R. A. Khandaker and K-K Wong, “Masked Beamforming in the Presence of Energy-Harvesting Eavesdroppers,” IEEE
Trans. Inf. Forens. Sec., vol. 10, no. 1, pp. 40-54, Jan. 2015.
[10] M. R. A. Khandaker and K.-K. Wong, “Robust secrecy beamforming in the presence of energy-harvesting eavesdroppers,”
IEEE Wireless Commun. Lett., vol. 4, no. 1, pp. 10-13, Feb. 2015.
[11] Y. Liu, A. P. Petropulu, and H. V. Poor, “Joint decode-and-forward and jamming for wireless physical layer security with
destination assiatance,” in Proc. Asilomar Conference on Signals, Systems and Computer (ASILOMAR11), Pacific Grove,
USA, Nov. 2011.
[12] J. Huang and A. L. Swindlehurst, “Cooperative jamming for secure communications in MIMO relay networks,” IEEE J.
Sel. Areas Commun., vol. 59, no. 10, pp. 4871-4884, Oct. 2011.
[13] J. Huang, A. Mukherjee, and A. L. Swindlehurst, “Secure communication via an untrusted non-regenerative relay in fading
channels,” IEEE Trans. Signal Process., vol. 61, no. 10, pp. 2536-2550, May 2013.
[14] L. Sun, T. Zhang, Y. Li, and H. Niu, “Performance study of two-hop amplify-and-forward systems with untrustworthy
relay nodes,” IEEE Trans. Veh. Technol., vol. 61, no. 8, pp. 3801-3807, Oct. 2012.
[15] W. Wang, K. C. Teh, and K. H. Li, “Secure Cooperative AF Relaying Networks with Untrustworthy Relay Nodes,” in
IEEE Globecom, Washington, DC. 2016, pp. 1-6.
[16] A. Mabrouk, K. Tourki, and N. Hamdi, “Secure cooperative untrusted-relay network with outdated CSI,” in IEEE
International Wireless Communications and Mobile Computing Conference (IWCMC), Paphos, 2016, pp. 90-95.
[17] A. Kuhestani, A. Mohammadi, and M. Noori, “Optimal Power Allocation to Improve Secrecy Performance of NonRegenerative Cooperative Systems Using an Untrusted Relay,” IET Commun., vol. 10, no. 8, pp. 962-968, May 2016.
[18] A. Kuhestani, A. Mohammadi, and P. L. Yeoh, “Optimal Power Allocation and Secrecy Sum Rate in Two-Way Untrusted
Relaying,” Submitted to IEEE Trans. Veh. Technol., Apr. 2017.
[19] L. Lv, J. Chen, L. Yang, and Y. Kuo, “Improving physical layer security in untrusted relay networks: cooperative jamming
and power allocation,” IET Commun., vol. 11, no. 3, pp. 393-399, Jul. 2017.
[20] T. Mekkawy, R. Yao, F. Xu, and L. Wang, “Optimal power allocation for achievable secrecy rate in an untrusted relay
network with bounded channel estimation error,” in 26th Wireless and Optical Communication Conference (WOCC),
Newark, NJ, USA, 2017, pp. 1-5.
[21] L. Sun, P. Ren, Q. Du, Y. Wang, and Z. Gao, “Security-Aware Relaying Scheme for Cooperative Networks with Untrusted
Relay Nodes,” IEEE Commun. Lett., vol. 19, no. 3, pp. 463-466, Sep. 2014.
[22] T. Schenk, RF Imperfections in High-Rate Wireless Systems: Impact and Digital Compensation, Springer, 2008.
[23] C. Studer, M. Wenk, and A. Burg, “MIMO transmission with residual transmit-RF impairments,” in Proc. ITG/IEEE
Workshop on Smart Antennas, Feb. 2010.
[24] E. Bjornson, A. Papadogiannis, M. Matthaiou, and M. Debbah, “On the impact of transceiver impairments on AF relaying,”
in Proc. 2013 IEEE Int. Conf. Acoustics, Speech, Signal Process.
[25] A. A. A. Boulogeorgos, D. S. Karas, and G. K. Karagiannidis, “How Much Does I/Q Imbalance Affect Secrecy Capacity?,”
IEEE Commun. Lett., vol. 20, no. 7, pp. 1305-1308, July 2016.
[26] J. Zhu, R. Schober, and V. K. Bhargava, ”Physical Layer Security for Massive MIMO Systems Impaired by Phase Noise,”
IEEE 17th International Workshop on Signal Processing Advances in Wireless Communications, pp. 1-5, Jul. 2016.
[27] J. Zhu, R. Schober and V. K. Bhargava, “Linear Precoding of Data and Artificial Noise in Secure Massive MIMO Systems,”
IEEE Trans. Wireless Commun., vol. 15, no. 3, pp. 2245-2261, Mar. 2016.
29
[28] J. Chen, X. Chen, W. H. Gerstacker and D. W. K. Ng, “Resource Allocation for a Massive MIMO Relay Aided Secure
Communication,” IEEE Trans. Inf. Forens. Sec., vol. 11, no. 8, pp. 1700-1711, Aug. 2016
[29] J. Chen, H. Chen, H. Zhang and F. Zhao, “Spectral-Energy Efficiency Tradeoff in Relay-Aided Massive MIMO Cellular
Networks With Pilot Contamination.,” IEEE Access, vol. 4, no. , pp. 5234-5242, Sept. 2016.
[30] L. Wang, Y. Cai, Y. Zou, W. Yang and L. Hanzo, “Joint Relay and Jammer Selection Improves the Physical Layer Security
in the Face of CSI Feedback Delays,” IEEE Trans. Veh. Technol., vol. 65, no. 8, pp. 6259-6274, Aug. 2016.
[31] B. E. Priyanto, T. B. Sorensen, O. K. Jensen, T. Larsem, T. Kolding, and P. Mogensen, “Assessing and modelling the effect
of RF impairments on UTRA LTE uplink performance,” in Proc. 2007 IEEE Vehic. Techn. Conf. Fall, pp. 12131217.
[32] K. S. Ahn and R. W. Heath, “Performance analysis of maximum ratio combining with imperfect channel estimation in the
presence of cochannel interferences,” IEEE Trans. Wireless Commun., vol. 8, pp. 1080-1085, Mar. 2009.
[33] S. Boyd and L. Vandenberghe, Convex Optimization, Cambridge University Press, 2004.
[34] I. S. Gradshteyn and I. M. Ryzhik, Table of Integrals, Series, and Products, 7th ed. New York: Academic, 2007.
| 7cs.IT
|
Constrained Deep Transfer Feature Learning and its Applications
Yue Wu
Qiang Ji
ECSE Department, Rensselaer Polytechnic Institute
110 8th street, Troy, NY, USA
arXiv:1709.08128v1 [cs.CV] 23 Sep 2017
{wuy9, jiq}@rpi.edu
Abstract
Feature learning with deep models has achieved impressive results for both data representation and classification
for various vision tasks. Deep feature learning, however,
typically requires a large amount of training data, which
may not be feasible for some application domains. Transfer
learning can be one of the approaches to alleviate this problem by transferring data from data-rich source domain to
data-scarce target domain. Existing transfer learning methods typically perform one-shot transfer learning and often
ignore the specific properties that the transferred data must
satisfy. To address these issues, we introduce a constrained
deep transfer feature learning method to perform simultaneous transfer learning and feature learning by performing
transfer learning in a progressively improving feature space
iteratively in order to better narrow the gap between the target domain and the source domain for effective transfer of
the data from source domain to target domain. Furthermore, we propose to exploit the target domain knowledge
and incorporate such prior knowledge as constraint during
transfer learning to ensure that the transferred data satisfies
certain properties of the target domain.
To demonstrate the effectiveness of the proposed constrained deep transfer feature learning method, we apply
it to thermal feature learning for eye detection by transferring from the visible domain. We also applied the proposed
method for cross-view facial expression recognition as a
second application. The experimental results demonstrate
the effectiveness of the proposed method for both applications.
1. Introduction
Feature learning with deep models is an active research
area. Recent research has demonstrated that with feature
learning methods, effective features can be learnt for both
representation and classification of the input data for many
computer vision tasks including face recognition [20], object detection [5], and scene classification [24]. Feature
Second feature space
Step 4: Feature learning with
both transferred and target data
Step 3: Constrained
transfer learning
First feature space
Step 2: Feature learning with
both transferred and target data
Step 1: Constrained
transfer learning
Source
data
Source domain
Transferred
data
Target
data
Raw data space
Intermediate domains Target domain
Figure 1. The framework of the proposed constrained deep transfer
feature learning method. The algorithm iteratively performs transfer learning and feature learning in more and more deeper feature
spaces.
learning with deep model, however, typically requires a
large amount of training data. Hence, feature learning for
domains with scarce data is not feasible.
Transfer learning can be one of the approaches to address
this problem and help feature learning in the data-scarce target domain by transferring data or knowledge from datarich source domain. Transfer learning not only can compensate for the lack of data in the target domain but can
also benefit the tasks in the target domain from the experience gained from the source domain. The transfer learning
techniques usually involve instance transfer, feature transfer, model parameter transfer, or relational knowledge transfer [12]. Those transfer learning techniques have been
used in natural language processing [1], document classification [23], etc. However, typical transfer learning techniques usually perform one-shot transfer in a fixed or shallow feature space, while a fixed feature space may not effectively fill the semantic gap between the target and source
domains. Another issue with transfer learning is that it is
purely data driven, without adequately considering certain
inherent properties of the target data.
To tackle these issues, we propose a constrained deep
transfer feature learning method to perform simultaneous
transfer learning and feature learning by exploiting the
knowledge in both target and source domains. The general
framework is shown in Figure 1. Specifically, we propose
15101
to iteratively perform constrained transfer learning and feature learning in several increasingly deeper feature spaces to
gradually transfer the knowledge from source domain to target domain and learn the features in the target domain. Furthermore, we propose to exploit the target domain knowledge and incorporate such prior knowledge as constraint
during transfer learning to ensure that the transferred data
satisfies certain properties of the target domain principles.
The proposed framework has the following merits. First,
the progressive transfer allows the creation of several intermediate pseudo domains to bridge the gap between the
source and target domains. As the knowledge transfer continues, the intermediate domains gradually approach the target domain. Second, feature learning is performed at each
level as knowledge transfer happens, and it is also performed progressively at a higher level as knowledge transfer
continues. Hence, knowledge transfer and feature learning
intertwine at each step, improving both feature learning and
knowledge transfer. Finally, by imposing constraints on the
transferred data, we can ensure the transferred data not only
possess certain desired characteristics but also to be semantically meaningful.
The remaining part of this paper is organized as follows.
Section 2 reviews the related work. Section 3 discusses the
Restricted Boltzmann Machine and Deep Boltzmann Machine models. Section 4 introduces the proposed method.
Section 5 shows the experimental results. Section 6 concludes the paper.
2. Related Work
Recently, an increasing number of works concentrate
on learning good representations for data with deep models. In [18], Salakhutdinov et al. build a hierarchical deep
model to learn features for object recognition and handwritten character recognition. Similar to the proposed method,
some works perform feature learning based on multimodal
data in different domains. For example, in [11], Ngiam et
al. use deep autoencoder to learn common features from
audio and video data. In [19], a deep multimodal DBM is
constructed to learn shared features for images and texts.
However, in contrast to the existing works that learn shared
representations across domains [11][19], our work focuses
on transferring knowledge from one domain to help feature
learning for another domain. Multimodal feature learning
methods usually require a lot of paired training data, while
we specifically handle the case where there is limited data
in the target domain and rich data in the source domain.
Transfer learning refers to the learning methods that
leverage the knowledge from the source domain to help
learning in the target domain. The transferred knowledge
includes training instances, features, model parameters and
relational knowledge [12]. For example, in [3], Triadaboost
is proposed to transfer the instances within the framework
ࢎ
ࣈ
ࢎ
࢚
ࣈ
ࢎ
(a)
ࣈ
ࢎ
࢚
ࣂ
࢙
(c)
࢚
(b)
Figure 2. (a) RBM for data in one domain. (b) RBM for transfer
learning. (c) DBM model.
of adaboost. In [15], feature spaces are found that minimize
the inter-domain differences.
There is also work that combines transfer learning with
feature learning. In [23], deep features learned with the
DBM model using data in the source domain are selected
for document classification in the target domain. Different
from the work [23] that first learns deep features and then
transfers the learned features among multimodal data, our
work preforms transfer feature learning at multiple levels of
the deep architecture. More importantly, we add target domain knowledge to constrain the transfer learning and feature learning procedure.
3. Background
Before we introduce the proposed method, we first review the Restricted Boltzmann Machine (RBM) and the
Deep Boltzmann Machine (DBM) models. RBM is a undirected probabilistic graphical model (Figure 2 (a)) that captures the joint probability of the binary input data t (e.g. data
in the target domain) with multiple binary hidden nodes h.
p(t; ξ) =
P
h exp(−E(t, h; ξ))
Z(ξ)
,
(1)
− E(t, h; ξ) = tT W h + cT t + bT h,
(2)
X
p(ti = 1|h; ξ) = σ(
Wi,j hj + ci ),
(3)
where
E(.) is the energy function, Z(ξ)
=
P
t,h exp(−E(t, h; ξ)) is the partition function, and
ξ = {W, c, b} are the parameters. The conditional
probabilities are as follows:
j
X
p(hj = 1|t; ξ) = σ(
ti Wi,j + bj ),
(4)
− E(t, h1 , h2 ; ξ 1 , ξ 2 ) = tT W 1 h1 + (h1 )T W 2 h2 ,
(5)
i
where σ(.) denotes the sigmoid function. Given the training
data, the parameters are learned by maximizing the log likelihood with stochastic gradient ascend algorithm approximated by the Contrastive Divergence (CD) algorithm [7].
DBM model [16] shown in Figure 2 (c) (assume two
layer model thereafter) consists of one layer of visible nodes
and multiple layers of binary hidden nodes. It represents the
probability of visible nodes (similar as Equation 1) with the
new energy function (ignoring the bias terms):
5102
Table 1. Notations. t, s and et represent target data, source data, and
the pseudo target data transferred from the source domain. If necessary, we use the superscript to indicate the pairwise (P)/unpaired
(U) data (corresponding data in both domains, and data in one domain).
Data
Notations
P
P
Pairwise target and source data
Data = {tk , sk }N
k=1
Unpaired target data
Unpaired source data
Algorithm 1: Constrained deep transfer feature learning
Data: Pairwise source and target data (DataP ); Unpaired
U
target data (DataU
t ); Unpaired source data (Datas ).
1
2
Result: Learned features: ξ , ξ ,...
/* Pre-training stage
*/
for Layer l=1 to L do
/* Constrained transfer learning
*/
• Learn the joint probability of target and source data
p(tl , sl ; θl ) with constraint C(.; θl ).
• Transfer the unpaired source data sli by sampling the pseudo
l
target data et through p(tl |sli ; θl ).
NU
t
DataU
t = {ti }i=1
NU
v
DataU
s = {sj }j=1
U
Unpaired pseudo target data
Ne
t
DataU
= {etn }n=1
te
where ξ = {ξ 1 , ξ 2 } are the parameters for different layers.
DBM is learned in a layer-wise manner and then jointly fine
tuned [16].
For feature learning using RBM or DBM, the hidden
nodes in the top layer are inferred for each input data using Equation 4 or the mean-field techniques [16], and they
are considered as the new feature representations. For transfer learning, the input data is the concatenated data from the
source s and target domains t, and the RBM (Figure 2 (b))
will learn their joint probability distribution. Transfer learning can then be performed using the joint probability.
4. Constrained deep transfer feature learning
4.1. The general framework
The proposed constrained deep transfer feature learning method is motivated by the following intuitions. First,
it’s straightforward to think of directly applying the DBM
model illustrated in section 3 to learn the deep features for
the target domain. However, DBM learning usually requires
a lot of training data with variations, while there is limited
data in the target domain. But, if we can capture the joint
distribution of target and source data as a bridge between
the source and target domains, we could transfer additional
large amount of source data to the target domain as pseudo
data for further target domain feature learning. Second, because of the significant domain differences, one time transfer may not be effective as stated above. Therefore, we propose to perform the transfer feature learning progressively
at different levels of feature spaces to generate intermediate
pseudo target spaces to gradually approach the final target
space. Third, because of the underlying mechanism that
produces the data, target data must follow certain properties
that we want to preserve in the transferred data. We hence
constrain the transfer learning to ensure the satisfaction of
these properties. Therefore, we propose the constrained
deep transfer feature learning method. Table 1 shows the
notations.
The general framework is shown in Figure 1 and Algorithm 1. First, in the pre-training stage, we perform the
transfer learning and feature learning iteratively. In the
transfer learning step, we capture the joint distribution of
/* Feature learning
*/
• Learn the features ξ l with target and pseudo target data.
• Project the data to the learned feature space for further
constrained transfer feature learning.
end
/* Joint fine-tuning stage
*/
• Jointly fine-tune the features ξ 1 , ξ 2 , ... based on the deeply
transferred pseudo target data and the real target data.
target and source data p(t, s; θ) using RBM (Figure 2 (b))
as a bridge and transfer additional large amount of source
data as pseudo target data et by sampling through the conditional distribution p(t|s; θ). In addition, we impose constraint C(.; θ) in the learning. Third, in the feature learning
step, the transferred pseudo target data are combined with
the original target data (shaded area in Figure 1) for feature
learning using RBM (Figure 2 (a)). Then, the pseudo target
data and real target data are projected to the learned feature
space. In the new feature space of the next level, the transferred pseudo target data are considered as “source data” for
further constrained transfer feature learning in the next iteration. Finally, similar as the DBM model (Figure 2 (c)), the
layer-wisely learned features with parameters ξ 1 , ξ 2 ,... are
fine-tuned jointly based on the deeply transferred pseudo
target data and the real target data. The iterative transfer learning and feature learning should converge, because
deep feature learning has been shown converging [8] and
the transfer learning also converges due to the gradually
reduced gaps between target domain and source domain
through the iterations. In the following subsections, we discuss each step in details.
4.2. Constrained semi-supervised transfer learning
in one layer
In this section, we first discuss how to learn the joint
probability of target and source data p(t, s; θ) in a semisupervised manner with constraints. Then, we discuss how
to generate the pseudo target data by sampling through
p(t|s; θ).
5103
4.2.1
Semi-supervised learning of the joint probability
Knowledge transfer from visible to thermal
In this work, we use RBM model illustrated in section 3 to
learn the joint probability of target and source data p(t, s; θ).
As shown in Figure 2 (b), the model defines the joint probability:
P
p(t, s; θ) =
h exp(−E(t, s, h; θ))
Z(θ)
,
(6)
where the energy function E(.) has similar format as that in
Equation 2.
Recall that RBM is usually learned with the Contrastive
Divergence (CD) algorithm [7]. However, standard CD
learning for the RBM model requires a large amount of
pairwise source and target data, which are limited in our
application. To tackle this problem, we propose the semisupervised learning method. Specifically, we learn the
model parameters by maximizing the log likelihood w.r.t the
pairwise source and target data DataP , the unpaired target
U
data DataU
t , and the unpaired source data Datas .
θ∗ = arg max L(θ; Data)
θ
k=1
U
Nt
1 X
log(p(ti ; θ))
= U
Nt i=1
(10)
U
L(θ; DataU
s )
Ns
1 X
= U
log(p(sj ; θ))
Ns j=1
(11)
Here, α and β are two parameters that balance different
terms. p(t; θ) and p(s; θ) in Equation 10 and Equation 11
are the marginal distributions of data in one domain by summing out the missing data in the other domain.
Following the CD algorithm, we solve the optimization
problem in Equation 7 using gradient ascent algorithm. The
gradient of model parameters for each log likelihood term
is calculated as:
∂L(θ; Data)
∂E
∂E
+h
,
= −h
iP
iP
∂θ
∂θ data
∂θ model
(b) Thermal
(c) Thermal mean eye
Figure 3. Constrained deep transfer feature learning for thermal
eye detection. (a)(b) Pairwise visible and thermal facial images [13]. (c) Thermal mean eye.
generate samples h, t from p(h, t|sj ; θ). Following the CD
algorithm, we estimate the model dependent expectations
with k-step (k=5) Gibbs update through the model, starting
from the samples calculated using the data dependent probabilities.
(7)
U
L(θ; Data) = L(θ; DataP ) + αL(θ; DataU
t ) + βL(θ; Datas )
(8)
NP
1 X
P
L(θ; Data ) = P
log(p(tk , sk ; θ))
(9)
N
L(θ; DataU
t )
(a) Visible
(12)
where h.iPdata and h.iPmodel represent the data dependent
and model dependent expectations.
In the CD algorithm [7], to approximately calculate the
data dependent expectation h.iPdata , we need to sample
the unknown variables from the data dependent probabilities Pdata . Note that, in the semi-supervised learning setting, the data dependent probabilities Pdata differ for each
data type. Specifically, PdataP = p(h|tk , sk ; θ), PdataUt =
p(h, s|ti ; θ) and PdataUs = p(h, t|sj ; θ). For the pairwise
data, we could directly sample h from p(h|tk , sk ; θ) using
Equation 4. However, for the unpaired data, p(h, s|ti ; θ)
and p(h, t|sj ; θ) are intractable. Thus, we use Gibbs sampling to generate h, s from p(h, s|ti ; θ). Similarly, we can
4.2.2
Transfer learning with constraints
In every domain, its data only represent the observations
of the underlying latent objects. It is often through either
a physical or biological process that the latent objects produce the observed data in the target domain. For example,
as shown in Figure 3, in the case of eye detection on thermal
images by transferring the knowledge from visible domain,
the intensities of the eyes measure the temperature near the
eye skin surface and the eye temperature is determined by
the the blood flow to the eye skin surface. Hence, the intensity distribution near the eye on an thermal image is mainly
determined by the underlying latent vascular structure distribution. For example, the tear dust area usually has the
high temperature due to the blood flow in the artery beneath
(Figure 3 (c)). On the other hand, in the pupil area and the
hair-insulated eye lash region, the temperature is expected
to be low because of lack of blood flow. These unique facial
anatomy structures in the eye regions lead to the unique and
distinct target data pattern that is universal across subjects.
As a result, to ensure a physically and semantically meaningful transfer, we propose to impose certain constraints
during the transfer learning in order to preserve such target data properties, such as the certain unique shape and
appearance characteristics.
Based on the intuitions illustrated above, we modify the
parameter learning problem in Equation 7 by adding the
constraint C(.; θ):
θ∗ = arg max L(θ; Data) − λC(.; θ)
θ
(13)
Here, the first term is the log likelihood function defined in
Equation 8. The second term C(.; θ) is a cost function to
ensure that the transferred data satisfies certain properties.
5104
To solve this optimization problem in Equation 13, depending on the type of C(.; θ), we could still use the gradient ascent algorithm and we have derived the gradient of
the first term w.r.t the parameters in section 4.2.1. Then,
we only need to calculate the gradient of the second term
C(.; θ) w.r.t the parameters, i.e., dC(.;θ)
dθ . It’s detailed calculated will be discussed in section 5.2.
4.2.3
Pseudo target data generation by sampling
Given the learned joint probability p(t, s; θ), we could transfer the large amount of additional source data sj ∈ DataU
s
to the target domain to generate the pseudo target data
et. This can be done by sampling through p(t|sj ; θ), sj ∈
DataU
s using Gibbs Sampling method that iteratively calls
Equation 3 and 4.
4.3. Feature learning in one layer
In each iteration, combining the transferred pseudo target data and the real target data, the goal of feature learning is to learn the RBM model (Figure 2 (a)) that captures
the variations of target and pseudo target data and uses the
hidden nodes as new features. More formally, the RBM is
defined in Equation 1 and parameter learning is formulated
as:
ξ ∗ = arg max L(ξ; Datat ) + γL(ξ; DataU
te )
ξ
(14)
The first term and second term represent S
the log likelihood
w.r.t the real target data Datat = DataP
DataU
t
t , and the
U
pseudo target data Dataet . γ is the parameter that balances
the two terms. To learn the features, we apply the standard Contrastive Divergence (CD) algorithm [7]. The only
difference is that the gradient is calculated based on both
terms in Equation 14. Given the learned model, for each
target and pseudo target data, its new representation can be
calculated using Equation 4. Then, for the next iteration in
the learned feature space, constrained transfer learning and
feature learning continue and interact until convergence.
4.4. Joint feature fine-tuning
After multi-layer constrained transfer feature learning in
the pre-training stage, we have the initially learned features ξ 1 , ξ 2 , .... In addition, we have the deeply transferred
L
pseudo target data in the top feature layer denoted as et .
Then, for joint feature fine-tuning, we need to project the
pseudo data back to the original space by repeatedly calling Equation 3 with parameters ξ L ,..., ξ 1 and get the deeply
1
transferred pseudo data in the original space bt . Then, based
1
on the pseudo target data bt and the real target data t1 ,
we apply standard fine-tuning algorithm [16] for the Deep
Boltzmann Machine model, where the mean-field fix point
equation is used to estimate the data-dependent expectation and the Persistent Markov Chain is used to estimate
the model dependent expectation.
5. Experimental Results
5.1. Eye detection on thermal images by transferring from visible domain
To demonstrate the proposed framework for constrained
deep transfer feature learning, we applied it to eye detection
on thermal images. Comparing to facial analysis in the visible domain, thermal facial analysis is more sensitive to eye
localizations [2]. However, there are limited works about
eye detection on thermal images. Furthermore, the existing thermal eye detection techniques typically use the visible image features, which are suboptimal, since thermal and
visible images are formed based on different principles [4].
For example, thermal images contain limited texture and
gradient information due to the heat diffusion phenomenon,
while the visible image features usually focus on encoding
these detailed information.
While feature learning on thermal images can alleviate
this problem, limited thermal training images make it very
difficult to leverage on the existing deep learning models.
To address this challenge, we propose to apply the proposed constrained transfer feature learning method to perform thermal feature learning by transferring the eye data
in the visible domain to the thermal domain to compensate
for the lack of thermal data. The target domain refers to the
thermal patches, including the eye patches and the background patches. The source domain refers to the visible eye
patches. Note that, we only need to transfer the eye patches
as positive data, since we can generate many negative data
by sampling from the background. For thermal eye detection, with the learned thermal features using the proposed
method, we train SVM classifier to search the eye with a
scanning window manner.
5.2. Implementation details
Databases: We use four databases including the visible and thermal facial behavior database(VTFB), the MAHNOB laughter database [13], the Natural visible and thermal facial expression database (NVIE) [21], and the Facial
Recognition Technology (FERET) database [14]. VTFB
has synchronized visible and thermal videos (FLIR SC6800
thermal camera) for 7 subjects and additional thermal
videos for 13 subjects with spontaneous facial expressions and arbitrary head poses. The MAHNOB Laughter database provides thermal facial sequences of 22 subjects with moderate head poses, neutral and happy facial
expressions. The NVIE database contains thermal facial sequences of 215 subjects with spontaneous and posed expressions. FERET database provides visible facial images
with different head poses and moderate expressions.
During training, we use the synchronized thermal and
visible images of the first 7 subjects from VTFB (1295 images) as pairwise data. We use the thermal images of ad5105
ditional 3 subjects from the VTFB (573 images), and visible images from FERET (3830 images) as unpaired data.
We test the method on the thermal images of remaining 10
subjects from VTFB (542 images), the MAHNOB database
(114 images) and the NVIE database (35550 images).
The method: We use two layer DBM with 800 and 600
hidden nodes to learn features. The RBM models used
to learn the joint distributions p(t, s; θl ) have 400 hidden
nodes. The hyper-parameters α, β, λ and γ are 0.4, 0.3,
0.002, and 0.3, respectively. In our experiments, the height
and width of the image patches are about 1/2 the inter-ocular
distance and the patches are normalized to 40×40. The
background patches are at least 1/4 inter-ocular distance
away from the eyes. We augment the training data by rotating and resizing the images. Following [16], we learn
the Gaussian-binary RBMs [9] with 1000 hidden nodes for
the raw visible and thermal image patches, respectively, and
then treat the values of the hidden layers as the preprocessed
data to speed up the learning procedure. With Matlab implementation, it takes about 30 hours to train the full model
with two layers and the constraint on a single core machine.
To impose the target domain constraint as discussed in
section 4.2.2, we propose to impose the constraint that the
transferred thermal eyes must satisfy the general appearance
pattern as shown in Fig. 3(c) for the thermal eye. To achieve
this goal, we propose to capture such a target data pattern
by using a thermal mean eye, which can be obtained by averaging the existing target data. The basic idea is that while
individual thermal eye may vary because of each person’s
unique eye structure, through averaging, the thermal mean
eye can capture their commonality while canceling out their
differences. The commonality is resulted from the shared
underlying eye structure. Figure 3(c) shows the mean eye
as an example. It apparently can capture the unique pattern of a thermal eye, i.e., brighter in the inner eye corner
and darker in the eye center that are true across subjects at
different conditions.
For transferring the visible eyes to thermal domain to
help the thermal feature learning, the transfer learning step
in Equation 13 becomes:
θ∗ = arg max L(θ; Data+ ) − λk
θ
1
NsU +
X
j
ht+ iQj − m+ k22 .
(15)
Here, “+” denotes the positive eye data. The second term
enforces
mean of the transferred pseudo thermal eye
P the
1
+
+
ht
i
Qj is close to the given mean eye m . SpecifU+
j
N
s
U+
ically, for each source data s+
j ∈ Datas , htiQj represents the expected transferred
thermal eye and
P pseudo
1
+
will be the mean
ht
i
Qj = p(t+ |s+
;
θ).
Then,
Q
U
+
j
j
j
Ns
of the transferred pseudo thermal eyes and it should be close
to the given mean eye m+ . Assume the gradient of the second constraint term of the objective function in Equation 15
w.r.t θ is denoted as δ. Then, it is calculated as follows:
δ =2 ∗ [
∗
1
NsU +
1
NsU +
X
j
X
ht+ iQj − m+ ]T
[−ht+
j
∂E
∂E
iR + h
iR ht+ iQj ],
∂θ j
∂θ j
(16)
+
U+
where Rj = p(t , h|s+
j ; θ), sj ∈ Datas .
Evaluation criterion: The eye detection error is de−Gl ||2 ,||Dr −Gr ||2 )
, where D and
fined as error = max(||Dl||G
l −Gr ||2
G represent the detected and ground truth eye locations,
and the subscript denotes left and right eyes. We regard
error < 0.15 as the successful detection. The evaluation
criteria follows the methods [10] [22] for fair comparison.
+
5.3. Evaluation of the proposed method
In this section, we evaluate the proposed constrained
deep transfer feature learning method on the VTFB
database. Note that, in the constrained transfer learning
step (discussed in section 4.2), the joint probability of thermal and visible eyes p(t+ , s+ ; θ) can be learned with different approaches. They are (M1) learning with pairwise
data using standard CD algorithm [7], (M2) proposed semisupervised learning with paired and unpaired data as discussed in section 4.2.1, (M3) learning with pairwise data
and constraints (second term in Equation 13), and (M4)
learning with semi-supervised data and constraints (the full
model). In the experiments, we compare the four variations
and study the other properties of the proposed method.
5.3.1
Evaluation of the constrained transfer learning
method in layer one
First, we evaluate the constrained transfer learning method
in layer one. We evaluate the learned joint probabilities
p(t+ , s+ ; θ) with different learning strategies (M1 to M4)
based on two criteria, including the reconstruction error and
the log likelihood on the hold-out pairwise data. The reconstruction error refers to the average pixel difference between the transferred pseudo thermal eyes sampled from
p(t+ |s+
j ; θ) and the ground truth thermal eyes. We calculate the log-likelihood on the hold-out testing set using the
method in [17].
As can be seen in Table 2, while both the semisupervised learning and the constraint (M2&M3) improve
the performances over standard learning method M1, combining them together (M4, the full model) achieves the best
performance.
Table 2. Evaluation of different constrained transfer learning methods in layer one.
Methods
M1: Pairwise data
M2: Semi-supervised
M3: Pairwise + constraint
M4: Semi + constraint
5106
Reconstruction error
0.0627
0.0601
0.0588
0.0580
Log likelihood
-514.1
-484.8
-492.0
-476.9
Evaluation of the thermal eye detector with the
constrained deep transfer features
In this subsection, we further evaluate the eye detection
performance based on the features learned using the proposed method with two-layer DBM (two iterations in Algorithm 1). For the constrained transfer learning step in
each layer of the deep model, we use four different learning
strategies (M1-M4).
(1) Different methods: We compare the proposed
method with different strategies (M1-M4) to the baseline
method (M0). The baseline method learns features using
the standard Deep Boltzmann Machine model based on the
thermal data without any knowledge transfer. As shown in
Table 3, even with only the pairwise data, the proposed
transfer feature learning method (M1) improves over the
baseline (M0). The semi-supervised learning (M2) and
constraint (M3) further boost the performances. By combining them together (M4), the constrained deep transferred
features learned from multi-modality data increase the detection rate by 6.09%, comparing to the baseline (M0).
We also implement two other baselines using DBM and
visible training data. (V0) refers to DBM feature learning
based on the visible data. (V1) fine-tune the visible features
learned with visible data on thermal data, so that the model
parameters for thermal feature learning are initialized as the
parameters of visible features. As can be seen in Table 3,
they are all inferior to the proposed method (M4).
Table 3. Eye detection rates using different methods
Methods
M0: DBM (baseline)
M1: Pairwise data
M2: Semi-supervised
M3: Pairwise + constraint
M4: Semi + constraint
V0: DBM visible (baseline)
V1: DBM visible finetune (baseline)
Eye detection rate
87.45%
89.48%
90.04%
91.88%
93.54%
73.25%
86.72%
(3) Learning with different amount of training data
We vary the training data to train the proposed method (features and classifier) and the baseline method, and the eye
detection rates are shown in Figure 5. Specifically, we keep
reducing the number of training subjects from “P7+UP3” (7
subjects with pairwise data + 3 subjects with unpaired data)
to “P2+UP2”. Comparing M4 to M0, the proposed method
always learns better features than the baseline. Comparing
M4 to M2, the constraints are important and they improve
the performances with different sets of data.
M0: DBM
M2: Semi−supervise
M4: Semi + constraint
0.95
0.9
Eye detection rate
5.3.2
0.85
0.8
0.75
0.7
P7+UP3
P7
P4+UP3
P4
Different training data sets.
P2+UP2
Figure 5. Learning with different amounts of training data.
(4) Comparison with other features: We compare the
learned features with the proposed method (M4) to standard
image features, such as the Scale Invariant Feature Transform (SIFT) feature, and the Histogram of Oriented Gradients (HOG) feature. For fair comparison, we train the
linear SVM with same training data (P7+UP3), and the experimental results only differ due to the features. As shown
in Figure 6, the learned features with the proposed method
significantly outperforms the other designed features.
1
percentage of images with successful detections
0.9
(2) Transfer in different layers: Figure 4 shows that
it’s important to perform the constrained transfer learning
in multiple layers, since it further boosts the performance
comparing to one layer transfer. In addition, transferring
in the deeper feature space (layer 2) is slightly better than
transferring in the shallow feature space (layer 1).
0.8
0.7
0.6
0.5
0.4
0.3
SIFT features
HOG features
Proposed method (M4: Semi+ constraint)
0.2
0.1
0.942
M4: semi + constraint
0
0.94
0.938
Eye detection rate
0.05
0.1
0.15
error
0.2
0.25
0.3
Figure 6. Cumulative error distribution curves using eye detectors
with different features.
0.936
0.934
0.932
0.93
0.928
0.926
0.924
0.922
transfer in layer 1
0
transfer in layer 2
transfer in both layers
Different constrained deep transfer feature learning methods
Figure 4. Transfer in different layers.
(5) Visualization: Figure 7 (a)(b) show the learned filters (parameters ξ 1 and ξ 2 ) with the proposed method (M4).
Filters in the lower level capture local dependencies. They
represent small dots, and some of them are similar to the
Gabor Filters. Filters in the higher level capture more global
variations.
5107
(a) Frontal image (source)
(b) Non-frontal image (target)
Figure 8. Cross-view neutral (left) and non-neutral (right) facial
expression recognition.
5.4. Comparison with existing thermal eye detectors
on other databases
To the best of our knowledge, there are only two
works [10][22] that report eye detection results on publicly
available thermal face databases. In [10], image patches
are represented by Haar wavelet coefficients. Two successive classifiers (coarse and fine) are constructed based on
features in different levels of details in a cascade manner.
In [22], information from the whole face region is used to
support eye detection on thermal facial images. It identifies
15 sub-regions on the face. Haar-like features and SVM are
used to learn the eye detector.
Table 4 illustrates that our detector outperforms the
method [10] on the MAHNOB Laughter database [13]. On
the NVIE database [21], our detector is significantly better
than the method in [22].
different features, including the learned features using the
proposed constrained deep transfer feature learning method,
the learned features using the conventional DBM, the handcrafted HOG and LBP features. In addition, we vary the
number of training subjects in the target domain (100, 50,
20). There are a few observations. First, the learned features are significantly better than the hand-crafted features.
Second, for all settings, by transferring the knowledge in
the source domain, the learned features using the proposed
method outperform the features learned with DBM.
0.9
facial expression recognition rate
(a) learned filters in layer 1 (b) learned filters in layer 2
Figure 7. (a)(b): Learned filters in different layers.
0.85
0.8
0.75
0.7
0.65
0.6
0.55
0.5
HOG
LBP
DBM
Proposed method
100
50
20
number of training subjects in the target domain
Table 4. Comparison with existing thermal eye detectors on the
MAHNOB laughter database and the NVIE database.
MAHNOB Laughter Database:
Methods
reported in [10] Proposed method (M4)
Detection rate
83.3%
87.72%
NVIE Database:
Methods
reported in [22] Proposed method (M4)
Detection rate
68%
81.60%
5.5. Second Application
The proposed constrained deep transfer feature learning
method is not limited to visible and thermal domains. It can
be extended to other domains for simultaneous knowledge
transfer and deep feature learning. In this section, we apply the proposed framework to cross-view facial expression
recognition as the second application. The goal of this application is to learn the effective feature representation to
classify neutral and non-neutral face in non-frontal view by
transferring the data in frontal view (source) to non-frontal
view (target). The unique facial structure in non-frontal
view is used to constrain the transferring learning. In the
experiments, we used the Multi-Pie database [6] (Figure 8),
which contains facial images with varying head poses, facial expressions, and illumination conditions (training: id
1-200, testing: id 201-337).
Figure 9 below shows the results. In the experiments, we
compared the classification accuracies (linear SVM) with
Figure 9. Facial expression classification accuracies with different
features and numbers of training subjects in the target domain.
6. Conclusion
In this paper, we propose a constrained deep transfer feature learning framework in order to perform feature learning
in a data-scarce target domain. Instead of performing transfer learning in a fixed feature space, we propose to simultaneously perform transfer and feature learning iteratively in
increasingly higher level of feature spaces in order to minimize the semantic gap between the source and target domains. Furthermore, to ensure the transferred data to be semantically meaningful and to be consistent with the underlying properties of the target domain, we incorporate target
domain knowledge as constraints into the transfer learning.
We applied the proposed framework for thermal feature learning for thermal eye detection by transferring the
knowledge from visible domain. We also applied the it for
cross-view facial expression recognition. The experiments
demonstrate the effectiveness of the proposed framework
for both applications. In the future, we would apply the
framework to other vision applications.
Acknowledgements: The work described in this paper was supported in part by the National Science Foundation under the grant number NSF CNS-1205664 and in
part by Army Research Office under the grant number ARO
W911NF-12-C-0017.
5108
References
[1] J. Blitzer, R. McDonald, and F. Pereira. Domain adaptation
with structural correspondence learning. In Proceedings of
the 2006 Conference on Empirical Methods in Natural Language Processing, EMNLP ’06, pages 120–128, Stroudsburg, PA, USA, 2006. Association for Computational Linguistics. 1
[2] X. Chen, P. J. Flynn, and K. W. Bowyer. Ir and visible light
face recognition. J. Comput. Vis. Image Underst, 99:332–
358, 2005. 5
[3] W. Dai, Q. Yang, G.-R. Xue, and Y. Yu. Boosting for transfer
learning. In Proceedings of the 24th international conference
on Machine learning, ICML ’07, pages 193–200, 2007. 2
[4] V. Espinosa-Duro, M. Faundez-Zanuy, and J. Mekyska. Beyond cognitive signals. Cognitive Computation, 3(2):374–
381, 2011. 5
[5] R. Girshick, J. Donahue, T. Darrell, and J. Malik. Rich feature hierarchies for accurate object detection and semantic
segmentation. In Proceedings of the IEEE Conference on
Computer Vision and Pattern Recognition (CVPR), 2014. 1
[6] R. Gross, I. Matthews, J. Cohn, T. Kanade, and S. Baker.
Multi-pie. Image Vision Comput., 28(5):807–813, 2010. 8
[7] G. Hinton. Training products of experts by minimizing contrastive divergence. Neural Computation, 14(8):1771–1800,
2002. 2, 4, 5, 6
[8] G. E. Hinton and S. Osindero. A fast learning algorithm for
deep belief nets. Neural Computation, 18:2006, 2006. 3
[9] G. E. Hinton and R. R. Salakhutdinov. Reducing the
dimensionality of data with neural networks. Science,
313(5786):504–507, July 2006. 6
[10] B. Martinez, X. Binefa, and M. Pantic. Facial component detection in thermal imagery. In Computer Vision and Pattern
Recognition Workshops (CVPRW), IEEE Computer Society
Conference on, pages 48–54, 2010. 6, 8
[11] J. Ngiam, A. Khosla, M. Kim, J. Nam, H. Lee, and A. Y.
Ng. Multimodal deep learning. In Internation conference on
Machine Learing (ICML), pages 689–696, 2011. 2
[12] S. J. Pan and Q. Yang. A survey on transfer learning.
IEEE Transactions on Knowledge and Data Engineering,
22(10):1345–1359, 2010. 1, 2
[13] S. Petridis, B. Martinez, and M. Pantic. The mahnoblaughter database. Image and Vision Computing Journal,
31(2):186–202, 2013. 4, 5, 8
[14] P. J. Phillips, H. Moon, S. A. Rizvi, and P. J. Rauss. The feret
evaluation methodology for face-recognition algorithms.
IEEE Trans. Pattern Anal. Mach. Intell., 22(10):1090–1104,
Oct. 2000. 5
[15] K. Saenko, B. Kulis, M. Fritz, and T. Darrell. Adapting visual category models to new domains. In Proceedings of
the 11th European conference on Computer vision: Part IV,
pages 213–226, 2010. 2
[16] R. Salakhutdinov and G. Hinton. Deep boltzmann machines.
In Proceedings of the International Conference on Artificial
Intelligence and Statistics, volume 5, pages 448–455, 2009.
2, 3, 5, 6
[17] R. Salakhutdinov and I. Murray. On the quantitative analysis of deep belief networks. In In Proceedings of the International Conference on Machine Learning, pages 872–879,
2008. 6
[18] R. Salakhutdinov, J. Tenenbaum, and A. Torralba. Learning
with hierarchical-deep models. Pattern Analysis and Machine Intelligence, IEEE Transactions on, 35(8):1958–1971,
2013. 2
[19] N. Srivastava and R. Salakhutdinov. Multimodal learning
with deep boltzmann machines. In P. Bartlett, F. Pereira,
C. Burges, L. Bottou, and K. Weinberger, editors, Advances
in Neural Information Processing Systems 25, pages 2231–
2239. 2012. 2
[20] Y. Taigman, M. Yang, M. Ranzato, and L. Wolf. Deepface:
Closing the gap to human-level performance in face verification. June 2014. 1
[21] S. Wang, Z. Liu, S. Lv, G. Wu, P. Peng, F. Chen, and
X. Wang. A natural visible and infrared facial expression
database for expression recognition and emition inference.
IEEE Trans. on Multimedia, 12(7):682–691, Nov. 2010. 5, 8
[22] S. Wang, Z. Liu, P. Shen, and Q. Ji. Eye localization from
thermal infrared images. Pattern Recognition, 46(10):2613
– 2621, 2013. 6, 8
[23] J. Zhang. Deep transfer learning via restricted boltzmann
machine for document classification. In Machine Learning and Applications and Workshops (ICMLA), 10th International Conference on, volume 1, pages 323–326, 2011. 1,
2
[24] B. Zhou, A. Lapedriza, J. Xiao, A. Torralba, and A. Oliva.
Learning deep features for scene recognition using places
database. In Z. Ghahramani, M. Welling, C. Cortes,
N. Lawrence, and K. Weinberger, editors, Advances in Neural Information Processing Systems 27, pages 487–495. Curran Associates, Inc., 2014. 1
5109
| 1cs.CV
|
PRIME IDEALS AND REGULAR SEQUENCES OF SYMMETRIC
POLYNOMIALS
arXiv:1309.1098v1 [math.AC] 4 Sep 2013
NEERAJ KUMAR
A BSTRACT. Let S = C[x1 , . . . , xn ] be a polynomial ring. Denote by pa the power
sum symmetric polynomial xa1 + · · · + xan . We consider the following two questions: Describe the subsets A ⊂ N such that the set of polynomials pa with a ∈ A
generate a prime ideal in S or the set of polynomials pa with a ∈ A is a regular
sequence in S. We produce a large families of prime ideals by exploiting Serre’s
criterion for normality [4, Theorem 18.15] with the help of arithmetic considerations, vanishing sums of roots of unity [9]. We also deduce several other results
concerning regular sequences of symmetric polynomials.
1. I NTRODUCTION
Let S = C[x1 , . . . , xn ] be a polynomial ring. A sequence of elements y1 , y2 , . . . , yd
in a ring S is called regular sequence on S if the ideal hy1 , y2 , . . . , yd i is proper and
for each i, the image of yi+1 is a nonzero divisor in S/hy1 , . . . , yi i.
Following the notation of Macdonald [10], let pa , ha and ea denote the power
sum symmetric polynomial, complete symmetric polynomial, and the elementary
symmetric polynomial of degree a in S respectively. Let N be the set of positive
integers. For a given set A ⊂ N, we denote by the set of power sum symmetric
polynomials as pA = {pa | a ∈ A}. In this paper, we discuss the following two
questions:
Question 1.1. Let S = C[x1 , . . . , xn ] be a polynomial ring. For which subsets A ⊂ N,
the ideal generated by the set of polynomials pA is a prime ideal in S.
Question 1.2. Let S = C[x1 , . . . , xn ] be a polynomial ring. For which subsets A ⊂ N,
the set of polynomials pA is a regular sequence in S.
Similarly, we ask these questions for the complete symmetric polynomials and
the elementary symmetric polynomials. For obvious reasons, whenever pA is a
regular sequence generating a prime ideal and pb ∈
/ hpA i, then pA , pb is a regular
sequence in S as well. More specifically, we will focus on Question 1.1.
These problems are fundamental in nature, and interesting, both in algebraic and
geometric point of view. In a more general setting of Question 1.1, it is mentioned
by Eisenbud that “ In general it is extremely difficult to prove that a given ideal
of polynomial is prime” [4, Chapter 10: pg.241]. The most powerful methods
known for showing primeness of ideal are Hochster’s method of “principal radical
system” [7] and Serre’s criterion for normality [4, Theorem 18.15]. We will use
the Serre’s criterion for normality in this paper.
The study of Question 1.2 began in the paper [1] in the dimention zero case by
Conca, Krattenthaler, and Watanabe. The Question 1.2 is highly non-trivial for
Key words: Regular sequence, Prime ideal, Symmetric polynomial.
2010 Mathematics Subject Classification: Primary 13A15, Secondary 14M10, 12E05.
1
2
NEERAJ KUMAR
n ≥ 3. For n = 3, a beautiful conjecture of Conca, Krattenthaler, and Watanabe
states that given a positive integers a < b < c with gcd(a, b, c) = 1, pa , pb , pc form
a regular sequence in C[x1 , x2 , x3 ] if and only if abc ≡ 0( mod 6), see [1, Conjecture 2.15]. The necessary condition follows from [1, Lemma 2.8]. For partial
evidence in support of sufficient condition, see [1, Theorem 2.11]. Similarly the
authors, also formulated a conjecture, when three complete symmetric polynomials form a regular sequence in C[x1 , x2 , x3 ], see [1, Conjecture 2.17]. In a joint
paper with Martino [8], we could provide evidence for these conjectures by proving it in special instances. Then we employed the technique of Serre’s criterion
to show the primeness of an ideal. For instance, we have shown that the ideal
I = hpa , pa+1 , . . . , pa+m−1 i is prime in S if m < n − 1, see [8, Theorem 3.3]. We
have also shown that the ideal I = hp1 , p2m i, where m ∈ N, is prime in C[x1 , . . . , x4 ],
see [8, Proposition 4.1]. In this way, we succeeded to give more families of regular sequences. With the help of Computer calculations, we proposed, Conjecture
4.5 and Conjecture 4.6 in [8]. One of the main result of this paper, Theorem 3.8,
partially answers the Conjecture 4.6 in [8].
In this paper, we have managed to produce families of prime ideals by exploiting
Serre’s criterion with the help of arithmetic considerations, vanishing sums of roots
of unity [9]. The main results of the paper are the following:
(i) Let S = C[x1 , x2 , . . . , xn ] be a polynomial ring with n ≥ 4. Let I = hpa , pb i,
where a, b ∈ N. Let b − a = n0 . Suppose q1 is the smallest prime factor in
the factorization of n0 . If q1 > max{n, a}, then I is a prime ideal in S.
(ii) Let S = C[x1 , x2 , . . . , xn ] be a polynomial ring with n ≥ 3. Let a ∈ N and
m < n − 1. Let I = hpa , p2a , . . . , pma i. Then I is a prime ideal in S.
Section 2 contains preliminary results. In Section 3, we discuss the problem of
whether two power sum symmetric polynomials generate a prime ideal in S for
n ≥ 4. We answer this to certain extent purely in terms of arithmetic conditions of
the degree of polynomials and the number of indeterminates, see Theorem 3.8. Let
I = hpa , p2a , . . . , pma i, where a ∈ N, and m < n − 1. We show that I is a prime ideal
in S for all n ≥ 3, see Theorem 3.13. For n ≥ 3, we show that ∂∂ hx1a , · · · , ∂∂ hxna form a
regular sequence in S for all a ≥ 2. We also show that any two complete symmetric
polynomial form a regular sequence in S for all n ≥ 3. Similar results also hold
for the power sum and elementary symmetric polynomials, see Lemma 3.1 and
Proposition 3.7. Computer calculations in CoCoA [2] suggest that I = hh1 , h2m i,
where m ∈ N, should be a prime ideal in C[x1 , . . . , x4 ]. It is obvious for m = 1. We
provide evidence for m = 2 in the Example 3.16.
2. G ENERALITIES
AND PRELIMINARY RESULTS
Let S = C[x1 , . . . , xn ] be a polynomial ring. Let pa , ha and ea be the power
sum symmetric polynomial, complete symmetric polynomial, and the elementary
symmetric polynomial of degree a in S respectively, that is,
n
pa (x1 , x2 , . . . , xn ) := ∑ xai ,
i=1
ha (x1 , x2 , . . . , xn ) :=
∑
xi1 xi2 · · · xia ,
∑
xi1 xi2 · · · xia .
1≤i1 ≤i2 ≤···≤ia ≤n
ea (x1 , x2 , . . . , xn ) :=
1≤i1 <i2 <···<ia ≤n
PRIME IDEALS AND REGULAR SEQUENCES
3
For instance, for n = 3 and a = 2, one has
p2 (x1 , x2 , x3 ) =x21 + x22 + x23 ,
h2 (x1 , x2 , x3 ) =x21 + x22 + x23 + x1 x2 + x1 x3 + x2 x3 ,
e2 (x1 , x2 , x3 ) =x1 x2 + x1 x3 + x2 x3 .
For these symmetric polynomials, we have the following Newton’s formula, see
[10, Equation’s 2.6′ , 2.11, 2.11′ respectively]:
n
∑ (−1)i ei hn−i = 0 for all n ≥ 1,
(1)
i=0
a
aha = ∑ pi ha−i for all a ≥ 1,
(2)
i=1
n
nen = ∑ (−1)i−1 en−i pi for all n ≥ 1.
(3)
i=1
We will use the following lemma to prove the smoothness of symmetric polynomials ha , ea , and pa in the Lemma 3.2. We will also use Lemma 2.1 in the Example
3.16.
Lemma 2.1. (Technical Lemma)
Let S = C[x1 , . . . , xn ] be a polynomial ring. Then for the symmetric polynomials
ha , ea , and pa , one has the following:
(i)
(ii)
(iii)
∂ ha
∂ xi
∂ ea
∂ xi
∂ pa
∂ xi
n ∂ ha
= ha−1 + xi ∂ ∂ha−1
xi and ∑i=1 ∂ xi = (n + a − 1)ha−1 .
n ∂ ea
= ea−1 − xi ∂ ∂ea−1
xi and ∑i=1 ∂ xi = (n − a + 1)ea−1 .
= axa−1
and ∑ni=1 ∂∂ pxai = apa−1 .
i
Proof. We can write ha = xi ha−1 + g for some polynomial g not involving xi . Taking partial derivative of ha w.r.t. xi , one obtains
∂ ha
∂ ha−1
= ha−1 + xi
.
∂ xi
∂ xi
(4)
By Euler’s formula, we have ∑ni=1 xi ∂ ∂ha−1
xi = (a − 1)ha−1 . Thus, we conclude that
∂ ha
n
∑ ∂ xi
i=1
= (n + a − 1)ha−1 .
Taking partial derivative of ea w.r.t. xi , one obtains
∂ ea
∂ ea−1
= ea−1 − xi
.
∂ xi
∂ xi
(5)
Proceeding as before, we conclude that
n
∂ ea
∑ ∂ xi
i=1
The claim of (iii) is obvious.
= (n − a + 1)ea−1 .
4
NEERAJ KUMAR
For a given natural number m, consider the m-th roots of unity in the field of
complex number C. We may ask ourself for which natural numbers n and k, do
there exist m-th roots of unity α1 , . . . , αn ∈ C such that α1k + α2k + · · · + αnk = 0? We
define such an equation to be Vanishing sum of k-th power of m-th roots of unity
of weight n. For k = 1, such an equation is said to be a vanishing sum of m-th
roots of unity of weight n. For instance, for m = 10 and k = 1, the set of n’s is
{0, 2, 4, 5, 6, 7, 8, 9, 10, . . . }. For a given m and k, let W (m, k) be the set of weights
n for which there exists a vanishing sum α1k + α2k + · · · + αnk = 0, where αi is a m-th
roots of unity. If m has a prime factorization of the form qa11 qa22 . . . qar r , then by the
theorem of Lam and Leung [9, Theorem 5.2], the weight set W (m, 1) is exactly
given by Nq1 + · · · + Nqr . Poonen and Rubinstein [13] have classified all minimal
vanishing sums α1 + · · · + αn = 0 of weight n ≤ 12. For similar treatment by Mann,
one may also see [11].
Remark 2.2. Note that by [9, Theorem 5.2], for a given m ∈ N, W (m, 1) depends
only on the prime divisors of m, and not on the multiplicity to which they occur
in the factorization of m. The theorem also shows that any (non-empty) vanishing
sum of m-th roots of unity must have weight ≥ q1 , where q1 is the smallest prime
divisor of m.
Lemma 2.3. Let m ∈ N and α1 , . . . , αn be a subset of m-th roots of unity in C.
Suppose that q1 is the smallest prime factor in the factorization of m. Let k ∈ N be
any natural number such that q1 > max{n, k}. Then one has α1k + α2k +· · ·+ αnk 6= 0.
Proof. The claim follows from Remark 2.2.
3. P RIME
IDEALS AND REGULAR SEQUENCES
We begin this section with a useful lemma, which states that all the partial
derivatives of a complete symmetric polynomial form a regular sequence in the
polynomial ring, see Lemma 3.1. Then we use Lemma 3.1 to show that the complete symmetric polynomials and their partial derivatives are smooth polynomials,
see Lemma 3.2 and Lemma 3.3. Then we recall the irreducibility of Schur polynomial in the polynomial ring, see [3, Theorem 3.1]. In a special case, we discuss
the smoothness of Schur polynomial, see Example 3.5. Then we discuss the main
results of this paper, Theorem 3.8 and Theorem 3.13 respectively. We also discuss
the case of two complete symmetric polynomials generating a prime ideal in the
polynomial ring.
Lemma 3.1. Let n ∈ N with n ≥ 3. Let S = C[x1 , x2 , . . . , xn ] be a polynomial ring.
Let a ≥ 2. Then the following holds:
(i)
(ii)
(iii)
∂ ha
∂ ha
∂ x1 , · · · , ∂ xn form a regular sequence in S.
∂ pa
∂ pa
∂ x1 , · · · , ∂ xn form a regular sequence in S.
∂ ea
∂ ea
∂ x1 , · · · , ∂ xn form a regular sequence in S for
all a < n.
Proof. Let the ideal generated by all the partial derivatives of ha be Ja , that is,
Ja = h
∂ ha
∂ ha
,··· ,
i.
∂ x1
∂ xn
(6)
PRIME IDEALS AND REGULAR SEQUENCES
5
We have aha = ∑ xi ∂∂hxai . Thus, clearly ha ∈ Ja . By Lemma 2.1 (i), we have
∂ ha
∈ Ja .
i=1 ∂ xi
n
(n + a − 1)ha−1 = ∑
That is, ha−1 ∈ Ja . By Lemma 2.1 (i), we have
∂ ha+1
∂ ha
= ha + xi
∈ Ja .
∂ xi
∂ xi
Thus, we observe that Ja+1 ⊂ Ja . Moreover, ha+1 ∈ Ja+1 . Proceeding similarly, we
have a chain of containment of ideals
Ja+n−2 ⊂ · · · ⊂ Ja+1 ⊂ Ja .
Therefore, we have ha−1 , ha , ha+1 , · · · ⊂ ha+n−2 ∈ Ja . Recall that by [1, Proposition
2.9], any n consecutive complete symmetric polynomials ha−1 , ha , . . . , ha+n−2 form
a regular sequence in S for all a ≥ 2. Thus, we conclude that ht(Ja ) = n, moreover
Ja is a complete intersection ideal in S. The claim (ii) is obvious. Let the ideal
generated by all the partial derivatives of ea be Ea . The proof of (iii) is similar
to above, except the fact that this time, we choose carefully a’s such that ea ∈
Ea . The reason for this is that ea = 0 for all a > n. We want to use the fact that
e1 , e2 , . . . , en ∈ Ea . As we know that e1 , · · · , en form a regular sequence in S.
In the following lemma, we discuss the smoothness of the symmetric polynomials ha , ea and pa .
Lemma 3.2. Let n ∈ N with n ≥ 3. Let S = C[x1 , x2 , . . . , xn ] be a polynomial ring.
Then the following holds:
(i) ha is smooth, hence an irreducible element in S for all a ≥ 1.
(ii) ea is smooth, hence an irreducible element in S for all 1 ≤ a ≤ n − 1.
(iii) pa is smooth, hence an irreducible element in S for all a ≥ 1.
Proof. If a = 1, the claims are obvious. We will give an elementary proof of (i).
The proof of (ii) and (iii) are similar.
Proof of (i): If ha = f · g with f and g non constant polynomial, then f and g
have to be homogeneous. By Bezout theorem, the hypersurfaces f = 0 and g = 0
intersects in the projective space Pn−1 , since n ≥ 3. This gives a singular point
on the hypersurface ha = 0. So, it suffices to prove that ha , ∂∂ hx1a , . . . , ∂∂ hxna have no
common zero in Cn − {0}. This claim follows from Lemma 3.1 (i).
In the following lemma, we discuss the smoothness of the partial derivatives of
the complete symmetric polynomials.
Lemma 3.3. Let n ∈ N with n ≥ 3. Let S = C[x1 , x2 , . . . , xn ] be a polynomial ring.
Let a ≥ 3. Then ∂∂hxai is smooth, and hence irreducible in S.
Proof. It is enough to show that
∂ ha
∂ x1
is smooth. Similar to the proof given in Lemma
3.2, it suffices to show that all its partial derivatives ∂∂x1 ∂hax1 , ∂∂x1 ∂hax2 , . . . , ∂∂x1 ∂haxn have
no common zero in Cn − {0}. We set few notations for convenience, for instance,
let ga,i = ∂∂hxai . Let the ideal generated by all the partial derivatives of ga,1 be Ia :
Ia = h
2
∂ ga,1 ∂ ga,1
∂ ga,1
,
··· ,
i.
∂ x1 ∂ x2
∂ xn
2
2
6
NEERAJ KUMAR
Also note that
∂ ga, j
∂ xi
=
∂ ga,i
∂xj .
With this observation, one obtains
Ia = h
∂ ga,1 ∂ ga,2
∂ ga,n
,
··· ,
i.
∂ x1 ∂ x1
∂ x1
Taking partial derivatives of ga,i w.r.t. x1 , for i = 1, . . . , n,
∂ ga,1 ∂ ha−1
∂ 2 ha
∂ ha−1
=
+ x1
+
,
∂ x1
∂ x1
∂ x1 ∂ x1
∂ xi
∂ ga,2 ∂ ha−1
∂ 2 ha
=
+ x2
∂ x1
∂ x1
∂ x2 ∂ x1
..
.
∂ ga,n ∂ ha−1
∂ 2 ha
=
+ xn
∂ x1
∂ x1
∂ xn ∂ x1
Then summing up, we get
n
∂ ga,i
∂ ha−1
∂ ∂ ha−1
=(n + 1)
+ ∑ xi
(
)
∂ x1
∂ x1
i=1 ∂ xi
i=1 ∂ x1
n
∑
=(♯)
∂ ha−1
,
∂ x1
for some integer number ♯, which is irrelevant. Thus, we conclude that
Also ga,1 ∈ Ia . We also know that ga,1 =
ha−1 ∈ Ia . Now consider the ideal
Ia+1 = h
∂ ha
∂ x1
=
ha−1 + x1 ∂∂hxa−1
.
1
∂ ha−1
∂ x1
∈ Ia .
Thus, we get
∂ ga+1,1 ∂ ga+1,1
∂ ga+1,1
,
··· ,
i.
∂ x1
∂ x2
∂ xn
Proceeding similarly, we obtain
n
∂ ga+1,i
∂ ha
∂ ∂ ha
=(n + 1)
+ ∑ xi
(
)
∂ x1 i=1 ∂ xi ∂ x1
i=1 ∂ x1
n
∑
=(♯1 )
∂ ha
,
∂ x1
for some integer number ♯1 , which is irrelevant. Thus, we conclude that
Also ga+1,1 ∈ Ia+1 . We also know that ga+1,1 =
∂ ga+1,1
∂ xi
∂ ha+1
∂ x1
=
ha + x1 ∂∂ hx1a .
∂ ha
∂ x1
∈ Ia+1 .
Thus, we get
ha ∈ Ia+1 . Also note that
∈ Ia for all i. That is, Ia+1 ⊂ Ia . Proceeding in a
similar way, we obtain the following relations
Ia+ j+1 ⊂ Ia+ j , and ha+ j−1 ∈ Ia+ j
for all j ≥ 0. Thus, using similar argument as in the Lemma 3.1, we conclude
that ht(Ia ) = n, moreover Ia is a complete intersection ideal in S. Thus the claim
follows.
Following Macdonald [10], the Schur polynomial is defined as
λ +n− j
sλ = sλ (x1 , . . . , xn ) =
det(xi j
)1≤i, j≤n
,
n− j
det(xi )1≤i< j≤n
PRIME IDEALS AND REGULAR SEQUENCES
7
where λ = (λ1 , λ2 , . . . , λn ) is the partition of non-negative integers with λi ≥ λi+1
for i = 1, . . . , n − 1. For Schur polynomial, we have the following relation, see [10,
Equation 3.4]:
where n ≥ l(λ ).
sλ = det(hλi −i+ j )1≤i, j≤n
(7)
Remark 3.4. Irreducibility of Schur polynomial is discussed by Dvornicich and
Zannier in [3]. Let n ≥ 3. For a given partition λ = (λ1 , λ2 , . . . , λn ), where λ1 >
λ2 > · · · > λn with λn = 0 and gcd(λ1 , λ2 , . . . , λn−1 ) = 1, then the Schur polynomial
sλ (x1 , . . . , xn ) is irreducible in S = C[x1 , x2 , . . . , xn ], see [3, Theorem 3.1]. Recall
that ha and ea are special forms of a Schur polynomial. Simply note that the irreducibility of ha and ea does not follow from [3, Theorem 3.1]. In the Lemma 3.2,
we not only show the irreducibility of these polynomials, but also the smoothness.
The obtained results partially extend the domain of partition for the irreducibility
of Schur polynomial.
Assuming that the Schur polynomial sλ (x1 , . . . , xn ) is irreducible in S. One may
ask, whether is it true that the Schur polynomial is also smooth? The answer is
positive in the case of ha and ea . Computational evidence shows that the answer
is negative in general. However, in a special case, when the partition λ is of the
form (λ1 , 1, 0) for λ1 ≥ 2, then the Schur polynomial sλ (x1 , x2 , x3 ) turns out to be
smooth in C[x1 , x2 , x3 ]. We discuss the proof of this in the following example:
Example 3.5. Let S = C[x1 , x2 , x3 ] be a polynomial ring. Let λ = (λ1 , 1, 0) be the
partition. Then the Schur polynomial sλ (x1 , x2 , x3 ) is smooth in S for all λ1 ≥ 2.
Proof. By (7), we have
sλ = s(λ1 ,1,0) = h1 hλ1 − hλ1 +1 .
Similar to the proof given in Lemma 3.2, it suffices to show that all its partial
derivatives ∂∂ sxλ1 , ∂∂ sxλ2 , ∂∂ sxλ3 have no common zero in C3 − {0}. Taking partial derivatives of sλ w.r.t. xi , we get
∂ hλ1 ∂ hλ1 +1
∂ sλ
=hλ1 + h1
−
∂ xi
∂ xi
∂ xi
∂ hλ1
=(h1 − xi )
for all i = 1, 2, 3.
∂ xi
∂ hλ1 ∂ hλ1 ∂ hλ1
∂ x1 , ∂ x2 , ∂ x3
By Lemma 3.1, we conclude that h ∂∂ sxλ1 , ∂∂ sxλ2 , ∂∂ sxλ3 i
We see that h1 6= xi , unless xi = 0 for all i. Thus the common zero of
is also a common zero of ∂∂ sxλ1 , ∂∂ sxλ2 , ∂∂ sxλ3 .
is a complete intersection ideal in S. Thus the claim follows.
We record the following convention, which we will follow from here onwards
throughout this paper:
Conventions 3.6. When we list the symmetric polynomials fi1 , fi2 , . . . , fik with respective degrees deg( fi j ) = i j , we always assume that i1 < i2 < · · · < ik , unless
otherwise specified.
In the following proposition, we will see that any two complete symmetric polynomials always form a regular sequence in S = C[x1 , . . . , xn ] for n ≥ 3. Similar
results also hold for the power sum and elementary symmetric polynomials.
8
NEERAJ KUMAR
Proposition 3.7. Let n ∈ N with n ≥ 3. Let S = C[x1 , x2 , . . . , xn ] be a polynomial
ring. Then the following holds:
(i) ha , hb form a regular sequence.
(ii) ea , eb form a regular sequence for all 1 ≤ a < b ≤ n − 1.
(iii) pa , pb form a regular sequence.
Proof. We will prove (i). By Lemma 3.2, ha is an irreducible polynomial in S.
Hence S/hha i is a domain. Now hb being an irreducible polynomial in S, can not
be factored into lower degree complete symmetric polynomials ha . So, hb is a
nonzero divisor in S/hha i for b > a. Hence ha , hb form a regular sequence in S.
Proof of (ii) and (iii) are similar.
We have seen that any two power sum polynomials pa , pb form a regular sequence in S = C[x1 , x2 , . . . , xn ] for n ≥ 3. We would like to know when two power
sum polynomials generate a prime ideal in S for n ≥ 4. In a special case, some
answers are known due to [8, Theorem 4.3 and Proposition 4.3]. In the following
theorem, we will answer this to certain extent, when two power sum polynomials
generate a prime ideal in S for n ≥ 4, purely in terms of arithmetic conditions of
the degree of polynomials and the number of indeterminates.
Theorem 3.8. Let S = C[x1 , x2 , . . . , xn ] be a polynomial ring with n ≥ 4. Let I =
hpa , pb i, where a, b ∈ N. Let b − a = n0 . Suppose q1 is the smallest prime factor in
the factorization of n0 . If q1 > max{n, a}, then I is a prime ideal in S. Moreover
pa , pb , pc form a regular sequence in S for all pc ∈
/ hpa , pb i.
Proof. If n0 = 1, then I is a prime ideal in S follows from [8, Theorem 4.3]. Thus,
we assume that n0 ≥ 2. Let R = S/I. We compute the Jacobian of I up to scalar
(We can ignore the coefficients, since we are in the field of characteristic zero.),
say Jacobian is J:
a−1 a−1
x
x
. . . xa−1
n
J = 1b−1 2b−1
.
x1
x2
. . . xb−1
n
Let J ′ = I2 (J), denotes the ideal generated by 2 × 2 minors of Jacobian. Also
ht(I) = 2, since I is generated by a regular sequence of length 2. The determinants
of 2 × 2 minors of the Jacobian can be written as
a−1 b−a
J ′ = h xa−1
(x j − xb−a
) i for 1 ≤ i < j ≤ n.
j xi
i
Thus, we have
a−1 b−a
I + J ′ = h pa , pb , xa−1
(x j − xb−a
) i for 1 ≤ i < j ≤ n.
j xi
i
√
Claim: I + J ′ = (x1 , x2 , . . . , xn ).
Suppose not, that is, there exists w = (w1 , w2 , . . . , wn ) ∈ Pn−1 with w ∈ Z(I + J ′ ).
Since w is in Pn−1 , we may assume w = (1, y1 , y2 , . . . , yn−1 ). As w ∈ Z(I + J ′ ), we
have that yb−a
= yni 0 = 1 for all i = 1, . . . n − 1, moreover w also satisfies pa , pb .
i
Therefore we have
1 + ya1 + ya2 + · · · + yan−1 = 0 and 1 + yb1 + yb2 + · · · + ybn−1 = 0.
Both the equations reduce to the existence of solution of 1+ya1 +ya2 +· · ·+yan−1 = 0.
We use the fact that all the yi ’s are n0 -th roots of unity, say 1, ζ1 , . . . , ζn0 −1 . Suppose
q1 is the smallest prime factor in the factorization of n0 . If q1 > max{n, a}, then
PRIME IDEALS AND REGULAR SEQUENCES
9
it follows from Lemma 2.3 that 1 + ya1 + ya2 + · · · + yan−1 6= 0. So, the only possible
solution has to be the trivial solution. Hence the claim is proved.
Thus ht(I + J ′ ) = n and dim S/(I + J ′ ) = 0. The co-dimension of J ′ in S is n − 2.
By [4, Theorem 18.15], R is a product of normal domain, since n ≥ 4. Thus, we
can write R = R1 × · · · × Rk . Since R is a standard graded C-algebra with R0 = C,
also R0 = (R1 )0 × · · · × (Rk )0 = Ck . Hence k = 1. Thus R is a normal domain and
I is a prime ideal in S.
Remark 3.9. Let p1 , p2 , p5 ∈ C[x1 , . . . , x4 ]. By Newton’s formula (3), we observe
that p5 ≡ 0 mod (p1 , p2 ). Hence p5 ∈ hp1 , p2 i. Replacing xi by xdi , we may also
conclude that p5d ∈ hpd , p2d i. Thus in the hypothesis of Theorem 3.8, we need the
condition pc ∈
/ hpa , pb i.
Remark 3.10. Computer calculations in CoCoA [2] suggest that whenever I =
hpa , pb i is a prime ideal in C[x1 , x2 , x3 , x4 ], then pa , pb , pc form a regular sequence,
except pa , p2a , p5a . It is clear from Remark 3.9 that pa , p2a , p5a do not form a
regular sequence. If this computational claim can be answered, then it will prove
[8, Conjecture 4.5] to certain extent.
Proposition 3.11. Let I and J be the prime ideal in K[x1 , . . . , xn ] and K[y1 , . . . , ym ]
respectively, where K is an algebraically closed field. Let (I, J) be the ideal generated by elements of I and J in K[x1 , . . . , xn , y1 , . . . , ym ]. Then (I, J) is a prime ideal
in K[x1 , . . . , xn , y1 , . . . , ym ].
Proof. There is a standard isomorphism
K[x1 , . . . , xn ]/I ⊗ K[y1 , . . . , ym ]/J ∼
= K[x1 , . . . , xn , y1 , . . . , ym ]/(I, J)
by sending f ⊗ g 7→ f g. We see that both K[x1 , . . . , xn ]/I and K[y1 , . . . , ym ]/J are
integral domains as well as K-algebras. By [12, Proposition 4.15(b)], the tensor
product of K[x1 , . . . , xn ]/I and K[y1 , . . . , ym ]/J is also an integral domain, since K
is algebraically closed field. Hence the claim follows.
Remark 3.12. Note that the goal of Proposition 3.11 is to generate more families
of prime ideals from given prime ideals.
By [1, Proposition 2.9], we know that p1 , p2 , . . . , pn form a regular sequence in
S = C[x1 , x2 , . . . , xn ]. We also know that a subset of a regular sequence is again a
regular sequence. Thus p1 , p2 , . . . , pm also form a regular sequence for all m < n.
Then by [1, Lemma 2.2], we conclude that pa , p2a , . . . , pma also form a regular
sequence. Let I = hpa , p2a , . . . , pma i, where a ∈ N, and m < n − 1. Let R = S/I.
Then R is a Cohen-Macaulay ring. In the following theorem, we will show that I
is a prime ideal in S. We show this by proving that R is a normal domain using
Serre’s criterion for normality.
Theorem 3.13. Let S = C[x1 , x2 , . . . , xn ] be a polynomial ring with n ≥ 3. Let a ∈ N.
Let I = hpa , p2a , . . . , pma i, where m < n − 1. Then I is a prime ideal in S.
Proof. For a = 1, it follows from [8, Proposition 4.3]. Assume a > 1. Let R = S/I.
We compute the Jacobian of I up to scaler, say Jacobian is J:
a−1
a−1
x1
xa−1
·
·
·
x
n
2
x2a−1 x2a−1 · · · x2a−1
n
1
2
J= .
.. .
..
..
.
.
···
x1ma−1 x2ma−1 · · ·
xnma−1
10
NEERAJ KUMAR
We can ignore the coefficients, since we are in the field of characteristic zero. We
have ht(I) = m, since I is generated by a regular sequence of length m. Let J ′ =
Im (J), denotes the ideal generated by m×m minors of Jacobian J. The determinants
of m × m minors of the Jacobian can be written as
J ′ = h xij11 xij22 · · · xijmm
∏
(xijaa − xijbb ) i for 1 ≤ i1 < i2 < · · · < im ≤ n,
1≤a<b≤m
where j1 , j2 , . . . , jm are some positive integers. Therefore
I + J ′ = hpa , p2a , . . . , pma , xij11 xij22 · · · xijmm
∏
(xijaa − xijbb )i.
1≤a<b≤m
√
Claim: I + J ′ = (x1 , x2 , . . . , xn ).
Suppose not, that is, there exists w ∈ Pn−1 with w ∈ Z(I + J ′ ). Then the vector
w can have at the most m − 1 distinct nonzero coordinates. If w has m or more
than m distinct nonzero coordinates, then w ∈
/ Z(J ′ ). Say w has v distinct nonzero
coordinates. We can write
w = (w1 , . . . w1 , w2 , . . . , w2 , . . . , wv , . . . , wv , 0, 0, . . . , 0),
where wi appears βi times and v ≤ m − 1. Also w should satisfy pia for i = 1, . . . , m.
Thus, we have
ia
ia
β1 wia
1 + β2 w2 + · · · + βv wv = 0 for i = 1, 2, . . . , m.
This is a system of equation, which can be represented in the matrix form with m
rows and v columns as
1
1
···
1
0
β1 wa1
a
a
a
wa1
w
·
·
·
w
w
β
v
2
2 2 0
..
.. .. = .. .
..
.
.
···
. . .
(m−1)a
w1
(m−1)a
w2
···
(m−1)a
wv
βv wav
0
We know that neither βi = 0 nor wi = 0 for i = 1, . . . , v. So βi wai 6= 0 for i = 1, . . . , v.
We can choose the matrix say M with first v rows out of m rows and look for
the solution. The matrix M is of full rank since wi 6= w j for i 6= j, so the only
possible solution has to be the trivial solution. Hence the claim is proved. By
similar argument as used in Theorem 3.8, we conclude that R is a normal domain
and I is a prime ideal in S.
For n ≥ 4, we know that the ideal hp1 , p2 , . . . , pm i is prime in S = C[x1 , x2 , . . . , xn ]
for all m < n − 1, see [8, Theorem 4.3]. We will see in the following theorem that
similar result holds for the complete symmetric polynomials and the elementary
symmetric polynomials.
Proposition 3.14. Let pa , ha , and ea in the polynomial ring S = C[x1 , x2 , . . . , xn ],
with n ≥ 4. Let m < n − 1. Then one has
hp1 , p2 , . . . , pm i = hh1 , h2 , . . . , hm i = he1 , e2 , . . . , em i.
Therefore the ideals hh1 , h2 , . . . , hm i and he1 , e2 , . . . , em i are also prime in S.
Proof. It follows from simple observation in the ring of symmetric polynomials
that the algebra generated by the power sum polynomials p1 , p2 , . . . , pm is same as
the algebra generated by h1 , h2 , . . . , hm , and also by e1 , e2 , . . . , em . Thus one has
hp1 , p2 , . . . , pm i = hh1 , h2 , . . . , hm i = he1 , e2 , . . . , em i.
(8)
PRIME IDEALS AND REGULAR SEQUENCES
11
One may also conclude (8) from Newton’s formula (1) and (3). By [8, Theorem 4.3], hp1 , p2 , . . . , pm i is a prime ideal in S. Thus, we can conclude that
hh1 , h2 , . . . , hm i and he1 , e2 , . . . , em i are also prime ideals in S.
Remark 3.15. Let n = 4 in Proposition 3.14. Then h1 , h2 , ha form a regular sequence in S provided ha ∈
/ hh1 , h2 i. By Newton’s formula (1), we observe that
h5 ≡ 0 mod (h1 , h2 ). Hence h5 ∈ hh1 , h2 i. We see that h1 , h2 generate a prime
ideal in S, but h1 , h2 , h5 do not form a regular sequence in S. Thus, we need the
condition ha ∈
/ hh1 , h2 i.
Computer calculations in CoCoA [2] suggest that I = hh1 , h2m i, where m ∈ N,
should be a prime ideal in S = C[x1 , x2 , x3 , x4 ]. For m = 1, it follows from Proposition 3.14. We prove for m = 2 in the following example.
Example 3.16. Let S = C[x1 , x2 , x3 , x4 ] be a polynomial ring. Let I = hh1 , h4 i.
Then I is a prime ideal in S.
Proof. Let R = S/I. We compute the Jacobian of I, say Jacobian is J. Let J ′ =
I2 (J), denotes the ideal generated by 2 × 2 minors of Jacobian. Also ht(I) = 2,
since I is generated by a regular sequence of length 2. The determinants of 2 × 2
minors of the Jacobian can be written as
∂ h4 ∂ h4
−
i for 1 ≤ i < j ≤ 4.
J′ = h
∂ xi ∂ x j
By Lemma 2.1 (i), we may write J ′ as
J ′ = h (x j − xi )h2 + (x2j − x2i )h1 + (x3j − x3i )i for 1 ≤ i < j ≤ 4.
Thus, we have
I + J ′ = hh1 , h4 , (x j − xi )h2 + (x2j − x2i )h1 + (x3j − x3i )i for 1 ≤ i < j ≤ 4.
√
Claim: I + J ′ = (x1 , x2 , x3 , x4 ).
Suppose not, that is, there exists w = (w1 , w2 , w3 , w4 ) ∈ P3 with w ∈ Z(I + J ′ ). We
assume that none of wi is zero. Also assume that wi 6= w j for i 6= j. Since w is in
P3 , we can make w1 = 1 as w1 6= 0. So, let w = (1, x, y, z). As w ∈ Z(I + J ′ ), we
have h1 (w) = 0 = h4 (w), moreover
(x j − xi )h2 (w) + (x3j − x3i ) = 0 for 1 ≤ i < j ≤ 4.
(9)
By (9), either xi = x j or h2 (w) = −(x2j + xi x j + x2i ) for 1 ≤ i < j ≤ 4. By assumption
xi 6= x j . Thus we have h2 (w) = −(x2j + xi x j + x2i ) for 1 ≤ i < j ≤ 4. An easy
simplification shows that it is not possible. We may argue similarly when wi = w j
for i 6= j. Hence there is no nontrivial solution. By similar argument as used in
Theorem 3.8, we conclude that R is a normal domain and I is a prime ideal.
Simply note that by [1, Lemma 2.2 and Proposition 2.9], the sequence of polynomials pa , p2a , . . . , pna and ha , h2a , . . . , hna form a regular sequence in S = C[x1 , x2 , . . . , xn ]
respectively. We partially extend the above conclusion in the following proposition:
Proposition 3.17. Let S = C[x1 , x2 , . . . , xn ] be a polynomial ring. Let n, b, k ∈ N.
Then the following holds:
(i) pa , p2a , . . . , p(n−1)a , pb form a regular sequence in S if and only if b = nak.
(ii) ha , h2a , . . . , h(n−1)a , hb form a regular sequences in S if and only if b = nak.
12
NEERAJ KUMAR
Proof. By Newton’s formula (3), one has
(
(−1)k nekn mod (p1 , p2 , . . . , pn−1 ),
pc =
0 mod (p1 , p2 , . . . , pn−1 ),
if c = nk;
otherwise.
Clearly p1 , p2 , . . . , pn−1 , pc form a regular sequence in S if and only if c = nk. Thus
the claim (i) follows from [1, Lemma 2.2 ]. By Newton’s formula (1), one has
(
(−1)k ekn mod (h1 , h2 , . . . , hn−1 ), if c = nk;
hc =
0 mod (h1 , h2 , . . . , hn−1 ),
otherwise.
Clearly h1 , h2 , . . . , hn−1 , hc form a regular sequence in S if and only if c = nk. Thus
the claim (ii) also follows from [1, Lemma 2.2 ].
4. F INAL R EMARKS
Following question is a special case of Question 1.1:
Question 4.1. Let S = C[x1 , x2 , . . . , xn ] be a polynomial ring with n ≥ 4. Let I =
hpa , pb i, where a, b ∈ N. For which pairs of integers a, b, I is a prime ideal in S.
We answer the previous question to some extent in the Theorem 3.8. We observe
that the Theorem 3.8 is still in weaker form, due to arithmetic condition q1 >
max{n, a}. In fact one can answer all the prime ideals which arises by using Serre
criterion for normality and vanishing sums of roots of unity, by answering the
following problem:
Problem 4.2. Let m ∈ N. Consider m-th roots of unity in the field of complex
number C. For which natural numbers n and k ≥ 2, do there exist m-th roots of
unity α1 , . . . , αn ∈ C such that α1k + α2k + · · · + αnk 6= 0.
Similar to Question 4.1, one may also ask the following question:
Question 4.3. Let S = C[x1 , x2 , . . . , xn ] be a polynomial ring with n ≥ 4. Let I =
hha , hb i, where a, b ∈ N. For which pairs of integers a, b, I is a prime ideal in S.
For n = 4 in the previous question, the computational calculations in CoCoA [2]
suggest that I = hh1 , h2m i, where m ∈ N, should be a prime ideal in S = C[x1 , x2 , x3 , x4 ].
Recall the following definition:
Definition 4.4. Let R = ci=0 Ri , Rc 6= 0 be a graded Artinian algebra. We say that
R has the strong Lefschetz property (SLP) if there exists an element L ∈ R1 such
that the multiplication map
×Ld : Ri −→ Ri+1
has full rank for all 0 ≤ i ≤ c − 1 and 1 ≤ d ≤ c − i. We call an L ∈ R1 with this
property a strong Lefschetz element.
L
Let J = hpa , pa+1 , . . . , pa+n−1 i in the polynomial ring S = K[x1 , x2 , . . . , xn ] over
a field K. Then the Artinian ring R = S/J has the SLP, see [6, Proposition 7.1].
In the following example, we will show that for a complete intersection ideal I =
hha , ha+1 , . . . , ha+n−1 i, the Artinian ring R = S/I have the SLP.
Example 4.5. Let K be a field of characteristic zero. Let S = K[x1 , x2 , . . . , xn ]
be a polynomial ring over K with standard grading, i.e. deg xi = 1 for all i. Let
I = hha , ha+1 , . . . , ha+n−1 i. Then R = S/I has the SLP.
PRIME IDEALS AND REGULAR SEQUENCES
13
Proof. Consider the initial ideal of I with respect to lexicographic term order, one
has
a+n−1
in(I) = hxa1 , xa+1
i.
2 , . . . , xn
Stanley [14] proved that every monomial complete intersection
a+n−1
K[x1 , x2 , . . . , xn ]/hxa1 , xa+1
i
2 , . . . , xn
has the SLP with x1 + x2 + · · · + xn as a strong Lefschetz element using the fact that
it is isomorphic to the cohomology ring of a direct product of projective spaces
over the complex number field. We see that S/in(I) has the SLP. Thus, by [15,
Proposition 2.9] we conclude that R has the SLP.
Question 4.6. Let S = C[x1 , x2 , . . . , xn ] be a polynomial ring. Let k ∈ N. Is it true
that for the ideal I = hha , h2a , . . . , h(n−1)a,hnak i, one has the initial ideal
nak
in(I) = hxa1 , x2a
2 , . . . , xn i.
If the answer to the previous question is positive. Then again one can construct
more examples of Artinian ring having the SLP. In a joint work with Martino [8],
we explicitly derive several examples of complete intersection ideal generated by
complete symmetric polynomials. Again, one can ask similar question for those
complete intersection ideals.
We conclude the section with one remark from the recent paper of Fröberg
and Shapiro [5], where the authors established a connection between regular sequences of complete symmetric polynomials and the codimention of the Vandermonde variety. To an arbitrary pair (k; I), where k ≥ 2 is a positive integer and
I = {i0 < i1 < · · · < im−1 }, m ≥ k is a sequence of integers, Fröberg and Shapiro
A in [5]. Under the assumption i = 0 and
discuss the Vandermonde variety V dk;I
0
gcd(i1 , . . . , im−1 ) = 1, the authors asked [5, Problem 2], for which pairs (k; I), the
A has the expected codimention. In the first non-trivial case k = 3, m =
variety V dk;I
A has the expected codimention (equal
5, the authors conclude that the variety V d3;I
to 3) if and only if three complete symmetric polynomials hi2 −2 , hi3 −2 , hi4 −2 form a
regular sequence in C[x1 , x2 , x3 ]. The problem of when three complete symmetric
polynomials ha , hb , hc , form a regular sequence in C[x1 , x2 , x3 ] was considered in
[1, Conjecture 2.17]. In [5, Conjecture 13], in a special case, it is mentioned that
if (a, b, c) = (1, 4, 3k + 2), k ≥ 1, then ha , hb , hc neither is a regular sequence, nor
hc ∈ (ha , hb ). This will be clear from the following proposition:
Proposition 4.7. Let S = C[x1 , x2 , x3 ] be a polynomial ring. Then h1 , h4 , hn form a
regular sequence in S if and only if n = 3k, k ≥ 1.
Proof. By Newton’s formula (1), one has
k
e3 mod (h1 , h4 ),
hn = 0 mod (h1 , h4 ),
−(k + 1)e2 ek3 mod (h1 , h4 ),
Thus the claim follows from [8, Theorem 2.2].
if n = 3k;
if n = 3k + 1;
if n = 3k + 2.
14
NEERAJ KUMAR
R EFERENCES
[1] A. Conca, C. Krattenthaler, and J. Watanabe. Regular sequences of symmetric polynomials.
Rend. Semin. Math. Univ. Padova 121, 179–199 (2009).
[2] The CoCoATeam. CoCoA: a system for doing Computations in Commutative Algebra. Available at http://cocoa.dima.unige.it.
[3] R. Dvornicich and U. Zannier. Newton functions generating symmetric fields and irreducibility
of Schur polynomials. Adv. Math. 222, no. 6, 1982–2003 (2009).
[4] D. Eisenbud. Commutative Algebra. With a view toward algebraic geometry. Graduate Texts in
Mathematics, 150 Springer-Verlag, New York, xvi+785 pp. (1995).
[5] R. Fröberg and B. Shapiro. Vandermonde varities and relations among Schur polynomials.
arXiv:1302.1298v1 [math.AC].
[6] T. Harima and J. Watababe. The strong Lefschetz property for Artinian algebras with nonstandard grading. J. Algebra. 311, no. 2, 511–537 (2007).
[7] M. Hochster. Big Cohen-Macaulay modules and algebras and embeddability in rings of Witt
vectors. Queen’s Papers on Pure and Applied Math. 42, 106–195 (1975).
[8] N. Kumar and I. Martino. Regular sequences of power sums and complete symmetric polynomials. Matematiche (Catania). 67, no. 1, 103–117 (2012).
[9] T. Y. Lam and K.H. Leung. On Vanishing sums of roots of unity. Journal of algebra. 224 , no. 1,
91–109 (2000).
[10] I. G. Macdonald. Symmetric functions and Hall polynomials. Second edition. With contributions by A. Zelevinsky. Oxford Mathematical Monographs. Oxford Science Publications. The
Clarendon Press, Oxford University Press, New York, x+475 pp, (1995).
[11] H. B. Mann. On linear relations between the roots of unity. Mathematika. 12, 107–117 (1965).
[12] J. S. Milne. Algebraic geometry. Allied Publishers. (2005).
[13] B. Poonen and M. Rubinstein. The number of intersection points made by the diagonal of a
regular polygon. SIAM J. Discrete Math. 11, no. 1, 135–156 (1998).
[14] R. P. Stanley. Weyl groups, the hard Lefschetz theorem, and the Sperner property. SIAM J.
Algebraic Discrete Methods. 1, no. 2, 168–184 (1980).
[15] A. Wiebe. The Lefschetz property for componentwise linear ideals and Gotzmann ideals.
Comm. Algebra. 32, no. 12, 4601–4611 (2004).
D IPARTIMENTO DI M ATEMATICA , U NIVERSIT Á DI G ENOVA , V IA D ODECANESO 35, 16146
G ENOVA , I TALY
E-mail address: [email protected]
| 0math.AC
|
DRAFT.
1
Optimal control of linear systems with limited control actions:
threshold-based event-triggered control
arXiv:1701.04871v1 [cs.SY] 17 Jan 2017
Burak Demirel, Euhanna Ghadimi, Daniel E. Quevedo and Mikael Johansson
Abstract—We consider a finite-horizon linear-quadratic optimal control problem where only a limited number of control
messages are allowed for sending from the controller to the
actuator. To restrict the number of control actions computed
and transmitted by the controller, we employ a threshold-based
event-triggering mechanism that decides whether or not a control
message needs to be calculated and delivered. Due to the nature of
threshold-based event-triggering algorithms, finding the optimal
control sequence requires minimizing a quadratic cost function
over a non-convex domain. In this paper, we firstly provide
an exact solution to the non-convex problem mentioned above
by solving an exponential number of quadratic programs. To
reduce computational complexity, we, then, propose two efficient
heuristic algorithms based on greedy search and the Alternating
Direction Method of Multipliers (ADMM) method. Later, we
consider a receding horizon control strategy for linear systems
controlled by event-triggered controllers, and we also provide
a complete stability analysis of receding horizon control that
uses finite horizon optimization in the proposed class. Numerical
examples testify to the viability of the presented design technique.
Index terms — Optimal control; Linear systems; Eventtriggered control; Receding horizon control
I. I NTRODUCTION
The problem of determining optimal control policies for
discrete-time linear systems has been extensively investigated
in the literature; see, e.g., [1]–[3]. The standard optimal control
problem assumes that an unlimited number of control actions
is available at the actuator. Although this assumption is valid
for many applications, it does not hold for some specific
scenarios where communication and computation resources
become scarce; for example, (a) control systems with constrained actuation resources [4]–[6], (b) control systems with
shared processor resources [7], or (c) information exchange
over a shared communication channel [8].
To reduce the communication burden between the controller
and the actuator, Imer and Başar [9] introduced a constraint
on the number of control actions while designing optimal
control policies for linear scalar systems. Later, Bommannavar
and Başar [10] and Shi et al. [11] extended the work of [9]
to a class of higher-order systems. The problem, proposed
in [9], can be also formulated as a cardinality-constrained
linear-quadratic control problem. The introduction of cardinality constraint changes the standard linear-quadratic control
B. Demirel and D. E. Quevedo are with the Faculty of Electrical Engineering and Information Technology, The University of Paderborn, Warburger Str. 100, 33098 Paderborn, Germany (e-mail: [email protected];
[email protected]).
E. Ghadimi is with Huawei Technologies Sweden AB, Skalhogatan 9-11
box 54, SE-164 94, Kista, Sweden (e-mail: [email protected]).
M. Johansson is with ACCESS Linnaeus Center, School of Electrical
Engineering, KTH Royal Institute of Technology, Osquldas väg 10, SE 10044
Stockholm, Sweden (e-mail: [email protected]).
problem from a quadratic programming (QP) to a mixedinteger quadratic programming (MIQP) problem. As it has
been proved in [12], it is an NP-complete problem. Therefore,
Gao and Lie [13] provided an efficient branch and bound
algorithm to compute the optimal control sequence. Since the
cardinality constraint is highly non-convex, a convex relaxation of this problem, based on `1 -regularized `2 optimization,
was considered in the literature; see, e.g., [14]–[17]. Although
`1 -norm regularization of the cost function is an effective way
of promoting sparse solutions, no performance guarantees can
be provided in terms of the original cost function due to the
transformed cost function.
Event- and self-triggered control systems have been broadly
used in the literature to reduce the amount of communication
between the controller and the actuator while guaranteeing an
attainable control performance; see, e.g., [18]–[22]. As distinct
from sparse control techniques as mentioned earlier, the eventand self-triggered control require the design of both a feedback
controller that computes control actions and a triggering
mechanism that determines when the control input has to be
updated. A vast majority of the works in the literature first
designed a controller without considering sparsity constraint,
and then, in the subsequent design phase, they developed
the triggering mechanism for a fixed controller. In contrast,
another line of research concentrates on the design of feedback
control law while respecting a predefined triggering condition.
The design of event- and self-triggered algorithms can
be extremely beneficial in the context of model predictive
control (MPC) strategies; see, e.g., [23]–[31]. MPC is a control
scheme that solves a finite-horizon optimal control problem
at each sampling instant and only applies the first element
of the resulting optimal control input trajectories. The use of
event-triggered algorithms, therefore, reduces the frequency
of solving optimization problems and transmitting control
actions from the controller to the actuator, and, consequently,
saves computational and communication resources. Lehmann
et al. [26] proposed an event-based strategy for discrete-time
systems under additive, bounded disturbances. The controller
only computes a new control command whenever the difference between actual state and predictive state exceeds a
threshold. Sijs et al. [23] combined state estimation with MPC
in order to design an event-based estimation framework. The
authors of [24], [25] combined MPC with event-triggered
sampling strategies based on the ISS concept. As different
from the aforecited works, the authors of [19]–[21] studied
infinite horizon quadratic cost. All works discussed above
focus on discrete-time linear/non-linear systems; however,
there are also a substantial number of works, which considers
continuous-time systems, in the literature; see, e.g., [30]–
[32]. For instance, the authors of [30], [32] investigated the
DRAFT.
2
stability of event-based MPC algorithms for continuous-time
nonlinear systems yet they did not consider disturbance. Later,
Li and Shi [31] studied the MPC problem for continuous-time
nonlinear systems subject to bounded disturbances.
Contributions: In this paper, we formulate a finite-horizon
optimal event-triggered control problem where a thresholdbased event-triggering algorithm dictates the communication
between the controller and the actuator. Then, we propose
various algorithms to compute a control action sequence that
provides optimal and sub-optimal solutions for this problem,
which is, in general, hard to solve due to two main reasons: (a)
it has the combinatorial nature since the decisions are binary
variable (i.e., transmit or not transmit), and (b) introduction
of a threshold-based triggering condition leads to optimizing
a convex cost function over a non-convex domain. The main
contributions of this paper are threefold:
(i) we show that the optimal solution of the control problem,
mentioned above, can be determined via solving a set of
quadratic programming problems;
(ii) we provide an efficient heuristic algorithm based on
ADMM to design a control input sequence that provides a
sub-optimal solution for the finite-horizon optimal eventtriggered control problem;
(iii) we describe the receding horizon implementation of the
event-triggered control algorithm, including a proof of
practical stability.
Outline: The remainder of this paper is organized as follows. Section II formulates the event-triggered finite-horizon
LQ control problem and introduces assumptions. Section III
provides a simple procedure for designing optimal control laws
to minimize the linear-quadratic cost function while Section IV
presents two heuristic methods that usually achieves tolerable sub-optimal performance while significantly reducing the
computational complexity. In Section V, a receding horizon
control scheme with event-triggered algorithm is presented.
Section VI demonstrates the effectiveness and advantages of
the presented approach while Section VII concludes the paper.
Notation: We write N for the positive integers, N0 for
N ∪ {0}, and R for the real numbers. Let Rn0 denote the
set of non-negative real vectors of dimension n, and Rn be
the set of real vectors of dimension n. Vectors are written
in bold lower case letters (e.g., u and v) and matrices in
capital letters (e.g., A and B). If u and v are two vectors
in Rn , the notation u ≤ v corresponds to component-wise
inequality. The set of all real symmetric positive semi-definite
matrices of dimension n is denoted by Sn0 . We let 0n be the
n–dimensional column vectors of all zeros, 1n be the vectors
of all ones. The Kronecker product of two matrices (e.g., A
and B) is denoted by A ⊗ B. For any given x ∈ Rn , the `∞ –
norm is defined by k x k∞ = max |xi |. For a square matrix
1≤i≤n
A, λmax (A) denotes its maximum eigenvalue in terms of
magnitude. The notation{xk }k∈K stands for {x(k) : k ∈ K},
where K ⊆ N0 . The power set of any set N , written P(N ),
is the set of all subsets of N , including the empty set and N
itself. The cardinality of a set denoted by |N |.
II. F INITE H ORIZON O PTIMAL E VENT-T RIGGERED
C ONTROL
We consider the feedback control loop, depicted in Fig. 1.
The dynamics of the physical plant G can be described by the
discrete-time linear time-invariant system:
G:
x(t + 1) = Ax(t) + Bu(t) ,
x(0) = x0 ,
(1)
where x(t) ∈ Rn is the state variable at time instant t, u(t) ∈
Rm is the control input at time instant t, A ∈ Rn×n and
B ∈ Rn×m are system matrices of appropriate dimensions,
and x0 ∈ Rn is the given initial condition. The pair (A, B)
is assumed to be stabilizable, and A is not necessarily Schur
stable.
In this paper, we assume that the sensor S takes periodic
noise-free samples of the plant state x(t) and transmits these
samples to the controller node. The controller K is eventtriggered, and it computes new control commands and transmits them to the actuator A only at times when x(t) ∈ S ,
Rn \ C0 with
C0 , {x ∈ Rn :k x k∞ < ε} ,
(2)
for a given threshold ε > 0. In case x(t) ∈ C0 , the controller
does not compute any new control actions, and the actuator
input to the plant is set to zero:
u(t) = 0m ,
where
∀t ∈ T ,
T , {t ∈ N0 ; t < N : x(t) ∈ C0 } .
(3)
(4)
It is worth noting that T denotes a set of time instants at which
no control computations are needed and, in turn, the plant runs
open loop. Notice that T is not a given set of variables; in
contrast, it is generated by the initial state x0 and the tentative
control actions u(t) for all t ∈ {0, · · · , T − 1}. Hence, to
determine the set T , it is necessary to compute the tentative
control actions u(t). In other words, the set of time instants
T is strongly coupled to the tentative control inputs u(t).
Throughout this paper, we aim at designing an admissible
optimal control sequence π = {u(0), u(1), · · · , u(N − 1)} to
minimize the quadratic cost function:
J(x(0), π) = x| (N )P x(N )
+
N
−1
X
x| (t)Qx(t) + u| (t)Ru(t) , (5)
t=0
where the matrices Q and P are symmetric and positive semidefinite while R is symmetric and positive definite. Then, we
consider the constrained finite-time optimal control problem:
J ? (x(0)) : minimize
π
J(x(0), π)
subject to x(t + 1) = Ax(t) + Bu(t),
∀t ∈ {0, · · · , N − 1},
x(0) = x0 ,
u(t) = 0m , ∀t ∈ T .
(6)
Recall that the set of time instants, at which no computations
(and also transmissions) are needed, i.e., T , is defined in (4).
DRAFT.
3
A
G
x0
0
S
x(t) 2 Rn \ C0
x(t) 2 C0
u(t)
x(t)
1
K
K
Fig. 1. Event-triggered control system with the process G, the actuator A,
the sensor S and the controller K.
Remark 1 Note that the event-triggered control problem,
described in Section II, is a particular class of hybrid systems.
One may, therefore, convert it into an equivalent mixed logical
dynamical (MLD) system that represents the system by using a
blend of linear and binary constraints on the original variables
and some auxiliary variables; see, e.g., [33], [34]. In this
paper, instead of using the MLD system formulation, we will
exploit geometric properties of the problem at hand, which
allows us to propose an optimization problem for optimal
finite-horizon control and establish stability result for its
receding horizon implementation.
Discussion: Although the problem (6) resembles the
cardinality-constrained optimal control problem (proposed
in [9]–[17]), there is a fundamental difference between them,
which stems from the question whether scheduling is an
exogenous input or autonomously generated? While solving
the cardinality-constrained control problem, one needs to
optimize both control action and scheduling sequence, which
are strongly coupled to each other in the optimization process.
On the other hand, the event-triggered control systems are
switched systems with internally forced switchings. In these
problems, scheduling sequences are generated implicitly based
on the evolution of the state x(t) and the control signal
u(t). The major difficulty of this problem is that scheduling
depends on the particular initial condition x0 and the tentative
control input u(t), and cannot be explicitly determined unless
a specific control signal was given; see the survey paper [35]
and references therein.
III. C OMPUTATION OF OPTIMAL CONTROL ACTIONS
In this section, we concentrate on finding the control input
sequence π that solves the finite-horizon optimal control problem, proposed in (6). Therefore, we present a framework which
is based on dividing the non-convex domain into convex subdomains. This is a simple yet effective procedure to follow;
however, the number of convex optimization problems, which
needs to be solved, grows exponentially in the length of the
control horizon.
The optimal control problem (6) is hard to solve due to the
restriction on the control space. Nevertheless, without loss of
generality, it is possible to convert this problem to a set of
2
3
t
Fig. 2. A decision tree diagram. Denote T , {t ∈ N0 ; t < N : x(t) ∈ C0 }.
This means that T represents a set of time instances at which a new control
signal does not need to be computed and transmitted from the controller to
the actuator. In the figure, dark colored circles represent the case x(t) ∈ C0
for any t ∈ T whereas light colored circles represent the case x(t) ∈ Rn \
C0 for any t ∈
/T . For instance, when N = 4, the set T gets values in
P({1, 2, 3}) = ∅, {1}, {2}, {3}, {1, 2}, {1, 3}, {2, 3}, {1, 2, 3} .
optimization problems of the form
JT? (x(0)) : minimize
π
J(x(0), π)
subject to x(t + 1) = Ax(t) + Bu(t),
∀t ∈ {0, 1, · · · , N − 1},
x(t) ∈ C0 , ∀t ∈ T ,
x(t) ∈ Rn \ C0 , ∀t ∈
/ T,
u(t) = 0m , ∀t ∈ T ,
x(0) = x0 ,
(7)
where T ∈ P(N ) with N = {1, 2, · · · , N − 1}. To obtain the
optimal solution of the problem (6), it is necessary to solve the
problem (7) for all subsets of the power set of N , i.e., P(N ),
and select the one providing the lowest cost; see Fig. 2. More
precisely, this can be written as:
J ? (x(0)) : minimize JT? (x(0)) .
T ∈P(N )
(8)
The number of sub-problems, which are needed to be
solved, grows exponentially with the control horizon N (i.e.,
|P(N )| = 2N −1 ). Bear in mind that since these sub-problems
are independent of each other, they can be also solved in
parallel. In addition to its combinatorial nature, the problem (7)
requires the optimization of a convex function over a nonconvex domain. In the literature, there exist some available
tools to solve the convex objective non-convex optimization;
see, e.g., [36]–[38].
As the feasible region S can be expressed as a finite
union of polyhedra, the disjunctive formulation can be applied.
In particular, assuming n-dimensional problem instance (6),
its feasibility region can be divided into 2n + 1 disjunctive
polyhedra1 . Following the construction above, we write:
S,
2n
[
p=1
Cp ,
(9)
where n is the state dimension, and Cp ⊂ Rn are full1 See Fig. 3 for an example of two dimensional problem where each
disjunctive set is shown in different colors.
DRAFT.
4
dimensional polyhedra, i.e.,
Cp , x ∈ Rn : Tp x ≤ d ,
1
gives a sequence Σ = {σ(0), · · · , σ(N − 1)}. Then, for given
T and Σ, we rewrite the optimization problem (7) as
JΣ? (x(0)) : minimize
π
J(x(0), π)
subject to x(t + 1) = Ax(t) + Bu(t),
∀t ∈ {0, 1, · · · , N − 1},
x(t) ∈ Cσ(t) , ∀σ(t) ∈ Σ,
u(t) = 0m , ∀t ∈ T ,
x(0) = x0 .
(11)
It is necessary to solve all possible combination of problem (11) to determine the global optimal solution. Particularly,
the problem (6) can be solved as a sequence of (2n + 1)N −1
quadratic programming (QP) problems. Hence, we have the
following series of optimization problems:
J ? (x(0)) :
minimize
Σ∈{0,··· ,2n}N −1
JΣ? (x(0)) .
(12)
It is worth noting that even though the number of QP
sub-problems grows exponentially with N and n, all these
problems are independent of each other; therefore, one can
parallelize the problems and reduce the computation time.
Corollary 1 The global optimal solution of the control problem (6) can be computed by solving a set of convex quadratic
programming problems of the form (11), in parallel, and
selecting the solution that provides the lowest control cost
among all feasible solutions.
The following examples should give the reader a better
understanding of the problem.
Example 1 Let n = 2 and S = R2 \C0 with C0 = {(x1 , x2 ) ∈
R2 : −ε < x1 < ε, −ε < x2 < ε}. The non-convex set S can
be divided into four convex subsets:
C1 = {(x1 , x2 ) ∈ R2 : x1 ≥ x2 , x1 ≥ −x2 , x1 ≥ ε} ,
C2 = {(x1 , x2 ) ∈ R2 : x1 ≤ x2 , x1 ≥ −x2 , x2 ≥ ε} ,
C3 = {(x1 , x2 ) ∈ R2 : x1 ≤ x2 , x1 ≤ −x2 , x1 ≤ −ε} ,
C4 = {(x1 , x2 ) ∈ R2 : x1 ≥ x2 , x1 ≤ −x2 , x2 ≤ −ε} ,
which are illustrated by different colors in Fig. 3. The optimal
control inputs can be computed by solving quadratic programs
for different combinations of convex sets. In order to compute
C1
0.5
State variable x2
for some Tp ∈ R(2n−1)×n and d ∈ R2n−1 .
We define a piece-wise constant function of time σ(t),
which takes on values in {0, 1, · · · , 2n} and whose value
p determines, at each time t ∈ {0, · · · , T − 1}, the state
variable x(t) to belong to the interior of the polyhedron Cp .
We divide the non-convex set S into convex subsets Cp with
p ∈ {0, 1, · · · , 2n} and, at each time step t ∈ {0, · · · , T − 1},
we choose only one active set. The switching signal,
(
0 if t ∈ T ,
(10)
σ(t) =
p otherwise,
C2
0
C0
−0.5
C3
−1
−1
C4
−0.5
0
0.5
1
State variable x1
Fig. 3. The non-convex set S , R2 \ C0 can be divided into four convex
regions, which are illustrated by different colors. For a given scheduling
sequence, i.e., T = {6, 7}, the optimal control inputs can be computed by
solving the optimization problem (11) for all possible combinations of five
convex sets. For N = 7, finding optimal control actions requires solving
15, 625 problems, and selecting the minimum control cost among 2, 650
feasible solutions. In Fig. 3, the blue curve represents the state trajectory
for Σ1 = {4, 4, 4, 1, 1, 0, 0}, the red curve represents the state trajectory for
Σ2 = {4, 1, 1, 2, 2, 0, 0}, and the magenta curve denotes the state trajectory
for Σ3 = {4, 1, 2, 2, 3, 0, 0}.
optimal control inputs, it is necessary to solve 5N −1 convex
quadratic programming problems.
Example 2 Let us consider the second-order system:
0.9 0.2
0.6
x(t + 1) =
x(t) +
u(t) ,
0.8 1.5
0.8
(13)
|
with the initial condition x0 = [ 0 −1 ] . The event-triggered
controller transmits control messages if the `∞ -norm of the
state variable x(t) is greater than ε = 0.25. The performance
indices are given by Q = diag{2, 2} and R = 5. The
control horizon is chosen as N = 7. The state trajectories,
depicted in Fig. 3, can be obtained via solving the optimization problem (11) for given sets Σ1 = {4, 4, 4, 1, 1, 0, 0},
Σ2 = {4, 1, 1, 2, 2, 0, 0}, and Σ3 = {4, 1, 2, 2, 2, 0, 0}. Fig. 3
shows the state trajectories for two different sets. The set Σ1
leads to the global optimal solution.
Example 1 demonstrates that obtaining the optimal solution
to the finite-horizon optimal control problem, mentioned earlier, requires solving an exponential number of QPs. Therefore,
in the next section, we propose heuristic algorithms that significantly reduce computational complexity, by compromising
the optimality guarantees slightly.
IV. C OMPUTATION OF SUB - OPTIMAL CONTROL ACTIONS
In this section, we propose two heuristic algorithms that usually achieve a low control loss while significantly decreasing
the computation time. We, firstly, develop a simple algorithm
that takes the initial state x0 as input and finds a suitable
sequence Σ in a greedy manner by increasing the horizon
of the problem from 1 to N . Secondly, we apply an algorithm based on ADMM to compute the heuristic scheduling
sequence Σ. This sequence is then used to solve the disjunctive
problem (11) and obtain a good approximate control sequence
DRAFT.
1
2
3
4
5
6
7
Data: x0 , n, N ;
Compute σ(0) based on x0 ;
for Iteration count k = 2, . . . , N do
Obtain x̃ = x? and ũ = u? by solving (11) with
Σ = {σ(0), · · · , σ(k − 2), p} for all
p ∈ {0, 1, · · · , 2n};
Compute p? = argminΣ JΣ? (x(0));
Set σ(k − 1) = p? ;
end
return Σ, x̃, and ũ;
Algorithm 1: Greedy heuristic
π. Both techniques are of polynomial time complexity, hence,
reducing the computational efforts significantly. Numerical
investigations (see Section VI) indicate that these heuristic
algorithms usually provide close to optimal solutions.
A. Greedy heuristic method
We construct an intuitive greedy algorithm to solve the optimal control problem, proposed in (6), efficiently. As discussed
earlier, given a feasible switching sequence Σ, a corresponding
optimal control problem can be formulated as a quadratic
programming problem of the form (11). One, then, solves
the combinatorial problem (12) to find the optimal switch
sequence Σ? , thereby solving the optimal control problem (6).
A good way to deal with the combinatorial nature of this
type of algorithms and reduce the computational complexity is
to employ a greedy algorithm, which solves a global problem
by making a series of locally optimal decisions. Greedy algorithms do not necessarily provide optimal solutions; however,
they might determine local optimal solutions, which can be
used to estimate a globally optimal solution, in a reasonable
time.
Our proposed greedy algorithm works as follows. Since x0
is known, we can compute corresponding switching signal
σ(0). Then, we set Σ = {σ(0), p} for all p ∈ {0, 1, . . . , 2n}
and solve 2n + 1 number of QPs of the form (11) to
obtain the optimal switch signal σ(1). The same procedure
then repeats for k = 2, . . . , N − 1, where at each step k,
we solve 2n + 1 times the QP problem (11) to find the
best σ(k) keeping the previously found scheduling sequence
Σ = {σ(0), . . . , σ(k − 1)} unchanged.
Algorithm 1 provides the detailed steps of discussed greedy
heuristic method. Algorithm 1 executes (2n + 1)(N − 1)
number of QPs compared to (2n+1)N −1 as in optimal control
procedure. As a result, it runs in polynomial time.
B. A heuristic method based on ADMM
In this section, we develop a computationally efficient
method for solving (6). Our approach is based on, Alternating
Direction Method of Multipliers (ADMM), a well developed
technique to solve large-scale disciplined problems [39]. The
idea of utilizing ADMM as a heuristic to solve non-convex
problems has been recently considered in literature [38]–[41].
In particular, [38], [40] employ ADMM to approximately solve
problems with convex costs and non-convex constraints. The
approximate ADMM solutions are improved by multiplicity
5
of random initial starts and a number of local search methods
applied to ADMM solutions.
In this paper, using some of the techniques introduced
above, we first apply ADMM to original non-convex problem
to achieve an intermediate solution. This algorithm, for a
general non-convex problem, does not necessarily converge;
however, when it does, the corresponding solution gives a
good initial guess for finding the optimal solution to this
problem. In a second phase of our heuristic method, we
perform a polishing technique with the aim of achieving a
feasible and hopefully closer to the optimal solution (see [38]
for an overview of polishing techniques). Our polishing idea
is based on disjunctive programming problem (11) with the
fixed event index T and trajectory sequence Σ.
Initially developed to solve large-scale convex structured
problems, ADMM has been recently advocated, to a great
extent, for being effective approximation technique to solve
non-convex optimization problems (see [38]–[40] and references therein). We start with casting (6) to the ADMM form.
Essentially, we have
minimize
z
1 >
z F z + I(w)
2
Gz = h,
z = y,
(14)
where z collects the state and control components; i.e.,
z , [x(0)> , . . . , x(N )> , u(0)> , . . . , u(N − 1)> ]> , F represents the positive definite Hessian of the quadratic cost. That is
F , blkdiag(Qs , Rs ) with Qs , IN +1 ⊗Q and Rs , IN ⊗R.
Moreover, I(y) is the indicator function enforcing the eventtriggering control law; that is
∞ if (16) holds,
I(y) =
(15)
0
otherwise,
{∃t ∈ {0, . . . , N − 1}|y(t + 1) ∈ C0 and y(t + N + 2) 6= 0m },
(16)
where (16) detects the constraint violation in (6) in terms
of having x(t) ∈ C0 while corresponding u(t) 6= 0m .
The constraints in (14) are twofold. The first constraint
in (14) describes the LTI system dynamics (1). Here G ∈
RnN ×(n(N +1)+mN ) have the following i = {1, . . . , N }, j =
{1, . . . 2N + 1}- blocks,
i−1
A
for i = 1 . . . N, j = 1,
−In
j = i and j > 1,
Gij =
i+N −j
A
B
if
j ≥ N + 2 and i + N ≥ j,
0
otherwise.
> >
and h ∈ Rn(N +1)+mN , [x(0)> , 0>
n , . . . , 0m ] . Finally, the
last constraint in (14) is of consensus-type to ensure that nonconvex constraint in (6) is asymptotically satisfied.
After formulating the problem, each iteration of ADMM
algorithm consists of (see [39] for a complete treatment of the
DRAFT.
6
subject):
1 >
z Fz
2
z
!
2
ρ
G
0
h
k
k
+
z−
z −
+u
I
I
0
2
(17)
k
k+1/2
= Π(z
+ 0 I u )
G
0
h
k
k+1/2
k
z −
.
=u +
z
−
I
I
0
1
Here, k denotes the iteration counter, ρ > 0 is the step-size
(penalty) parameter, and Π denotes the projection onto nonconvex constraint and is given by
0
if i ≥ N + 2 and x(i − N − 1) ∈ C0
Πi (x) =
x(i) otherwise.
(18)
When the cost function in hand is convex and the constraints set is closed and convex then the ADMM algorithm converges to the optimal solution of the problem.
In our case, since I(y) is non-convex the ADMM algorithm (17) may not converge to optimal point. Moreover,
examples can be found in which (17) does not even converge into a feasible point. To further improve the ADMM
solution, we utilize an additional convex problem to polish the intermediate results. That is, we restrict the search
space to a convex set that includes the ADMM solution pattern. Let z̃ , [x̃(0)> , . . . , x̃(N )> , ũ(0)> , . . . , ũ(N − 1)> ]>
e
be the output of the ADMM algorithm and the sets Σ
e
and T , {t ∈ N0 ; t < N : x̃(t) ∈ C0 } denote the trajectory
> >
and the restriction
sequence associated with [x̃>
1 , . . . , x̃N ]
index of x̃ into C0 , respectively. Then our heuristic polishing
e Algoprocedure formulates (11) with T = Te and Σ = Σ.
rithm 2 provides a summary of our heuristic method. The input
variables include the error-tolerance threshold tol , randomized
initial state z 0 as well as zbest and fbest to store the final
solution.
A few comments related to Algorithm 2 are in order.
10
z k+1/2 = argmin
z
k+1
u
k+1
1) In contrast to the ADMM algorithms for convex problems that converge to the optimum for all positive range
of the step-size parameter ρ, the stability of the ADMM
algorithm for non-convex problems is sensitive to the
choice of ρ. Moreover, similar to the convex case,
the convergence speed is also affected by the choice
of ρ. A variety of techniques including proper stepsize selection, over relaxation, constraint matrix preconditioning and caching can be employed to accelerate
the ADMM procedure (17) (see [42], [43] for a reference
on the topic).
2) One can utilize an adaptive iteration count procedure
to improve the probability of finding feasible solution
to (6). The procedure works as the following. Run the
ADMM algorithm for a fixed number of iterations and
then check if the disjunctive QP problem has feasible
solution. If not, then increase the number of iterations
(e.g., double it) and repeat the procedure (run ADMM
and disjunctive QP) until finding a feasible solution or
reaching to a point where the current ADMM solution
2
3
4
5
6
7
8
9
Data: tol , z 0 ∼ N (0, σ 2 I), zbest = ∅, and fbest = ∞;
for Iteration count k = 1, 2, . . . , K do
update z from (17);
if kGz k − hk ≤ tol and (1/2)z k> F z k < fbest then
zbest = z k ;
end
end
if fbest < ∞ then
Obtain zbest from disjunctive problem (11);
end
return zbest ;
Algorithm 2: ADMM-Disjunctive heuristic
does not improve the previous one in terms of constraint
violations and quadratic cost.
3) The variable z 0 is initialized randomly using a normal
distribution. In general, repeating the ADMM algorithm
for a multiple of initial points may improve the approximate solution at the cost of extra computational
complexity (as suggested in e.g., [40]). However, in
our application we did not experience significant performance improvement by repeating the ADMM algorithm
(steps 1−6 in Algorithm 2) with different initializations.
One reason could be that in Algorithm 2, we employ the
e and Te based on the ADMM sosolution trajectories Σ
lution z̃ and not z̃ itself in the disjunctive problem (11).
4) Using caching and LDL> matrix factorization techniques in the first-step of (17) [39], each z-update in Al-
gorithm 2 costs on the order of O (n(N + 1) + mN )2
for dense P matrices. Furthermore, if we assume N ≥
max{n, m} then each ADMM iteration costs on the
order of O(N 2 ). This means that overall cost of the
ADMM algorithm is of O(KN 2 ) where K is the
number of ADMM iterations. In our application, with
proper algorithm parameter selection, the ADMM algorithm converges almost always within a few hundreds
of iterations (see the results presented in Section VI).
5) The computational complexity of the disjunctive
step in our heuristic method falls into the generic
complexity of convex QPs on the order of
2
O((|T |2n + (N − |T |)(2n − 1)) ) [44].
V. R ECEDING HORIZON CONTROL
Except a few special cases, finding a solution to an infinite
horizon optimal control problem is not possible. Receding
horizon control is an alternative scheme to infinite horizon
problem that repeatedly requires solving a constrained optimization problem. At each time instant t ≥ 0, starting at
the current state x(t) (assuming that a full measurement of
the state x(t) is available at the current time t ∈ N0 ), the
following cost function
J(x(t), πt?N −1 ) = Vf (x(t + N | t))
+
t+N
X−1
k=t
`(x(k | t), u(k | t))
subject to system dynamics and constraints involving states
and controls is minimized over a finite horizon N . Here, the
DRAFT.
7
1.2
R = 5 while the prediction horizon is set to N = 6. Fig. 4
shows the state trajectories of the receding horizon implementation of the event-triggered control system, described by (1)
– (4), for different initial conditions x0 . As can be seen in
the same figure, the number of transmitted control commands
depends on the initial condition.
State variable x2
0.6
0
Dµ
VI. N UMERICAL EXAMPLES
−0.6
−1.2
−1.2
A. Performance of sub-optimal event-triggered control actions
−0.6
0
0.6
1.2
State variable x1
Fig. 4. Evolution of the state trajectories of the system (13) under the receding
horizon strategy (19) and constraints (2)–(4) for a set of initial conditions x0 .
function ` defines the stage cost, and Vf defines the terminal
cost. Denote the minimizing control sequence, which is a
function of the current state x(t), by
πt?N −1 = u? (t | t)| , · · · , u? (t + N − 1 | t)| ,
then the control applied to the plant at time t ≥ 0 is the first
element of this sequence, that is,
u(t) = I 0 · · · 0 πt?N −1 .
Time is then stepped forward one instant, and the procedure
described above is repeated for another N -step ahead optimization horizon.
It is worth noting that, due to the fundamental limitation of
the optimization problem (6), the state x(t) never converges
to the origin if the system is unstable. It can only converge
into a set including the origin. The following result shows the
existence of this set:
Theorem 1 Suppose that Dµ , {x ∈ Rn : k x k∞ ≤ µ} is a
neighborhood of the origin, where
s
κλmax (A| P A + Q)nε2
µ,
(19)
λ2min (Q)
for some κ > 0. Then, the system (1) with event-triggering
constraints (3) and (4) is uniformly practically asymptotically
stable, i.e., lim k x(t) k∞ ≤ µ.
t→∞
Theorem 1 establishes practical stability of the system (1)
with event-triggering constraints (3) and (4). It shows that
if provided conditions are met, then the plant state will be
ultimately bounded in an `∞ -norm ball of radius µ. It is worth
noting that, as the other stability results, which use Lyapunov
techniques, this bound will not be tight. Next, we would like
to exemplify the convergence of the state x(t) to the set Dµ
from a set of initial conditions x0 .
Example 3 Consider the second-order system, given by (13),
|
πk
for all 0 ≤
with initial conditions x0 = 1.2 [ sin( πk
6 ) cos( 6 ) ]
k ≤ 12. The controller transmits a control message whenever
the event-triggering condition, i.e., k x(t) k∞ > 0.25, holds.
The performance indices are chosen as Q = diag{2, 2} and
To illustrate performance benefits of the proposed heuristic
algorithms, we consider a third-order plant with the following
state-space representation:
0.53 −2.17 0.62
0.4
x(t + 1) = 0.22 −0.06 0.51 x(t) + 0.7 u(t)
−0.92 −1.01 1.69
0.9
+ Bw w(t) . (20)
Assume that there is no disturbance acting on the plant, i.e.,
Bw = 0.
Our aim, here, is to minimize (5) with performance indices
Q = P = diag{2, 2, 2} and R = 5, and the horizon length
N = 8. The initial condition is chosen as
|
x0 = sin θ cos φ sin θ sin φ cos θ ,
where 0 ≤ θ ≤ π and 0 ≤ φ ≤ π. To perform simulations,
we selected 577 different data points, which are equidistant
over a half spherical surface. We evaluated the performance of
the greedy and the ADMM-based heuristic algorithms for all
these initial conditions. For each data point on the half sphere,
we run 823, 543 number of QPs to obtain the global optimal
solution. In total, we run more than 475 million number of
QPs in our evaluation. We used a linux machine with 32 cores
in Intel Core i7 Extreme Edition 980X to solve these QPs
in parallel threads. Within this computing resources, it takes
approximately 6 days to perform simulations for one problem
setup.
To compare the (true) optimal control performance with
our heuristic method, we performed Algorithm 2 on the
same problem set. In particular, we evaluated three sets
of problem with ε = 0.2, ε = 0.4, and ε = 0.6. It is
generally expected that the control problem becomes more
challenging with increasing the event-threshold ε. We run
the ADMM-based heuristic algorithm with adaptive iteration
count procedure described in Section IV-B. Moreover, we
picked ρ = 9.8, ρ = 5.8, and ρ = 6.9 respectively,
for the aforementioned problem scenarios. Under this setup,
in both cases (i.e., ε = 0.2 and ε = 0.4), the ADMM
method produced always feasible solutions for maximum 300
number of ADMM-iterations. However, when ε = 0.6 and
ρ = 6.9, the method created infeasible solutions for two
>
initial conditions: x0 = [0.3036, 0.2330, 0.9239] and x0 =
>
[0.6830, 0.1830, 0.7071] . To compute feasible solutions, we
needed to adjust ρ accordingly for these initial conditions.
Fig. 5 shows the outcome of the comparison. As it shows,
for ε = 0.2, the ADMM-based heuristic algorithm is able
to find near optimal solutions (with relative error less than
DRAFT.
8
ε = 0.4
ε = 0.2
ε = 0.6
0.8
0.5
0.4
0.4
0
0
0
0
0.1
0.2
0.3
(Jadmm − J ? )/J ?
0.4
0.6
1
0.8
0
0.1
0.2
0.3
(Jadmm − J ? )/J ?
x1 (t)
1
0.4
0.6
0
0.1
0.2
0.3
(Jadmm − J ? )/J ?
0
−1
−2
0.4
exhaustive search
ADMM
0
10
20
30
1
0.4
0.2
0.2
0
0
-2
-1
0
1
2
3
|Tadmm | − |T ? |
4
0
-2
-1
0
1
2
3
|Tadmm | − |T ? |
4
-2
-1
0
1
2
3
|Tadmm | − |T ? |
4
ε = 0.4
ε = 0.2
−0.5
20
30
40
50
0
−1
exhaustive search
ADMM
−2
−3
0.4
0.4
0.2
10
1
ε = 0.6
0.4
0
2
Fig. 5. Relative error (Jadmm − J ? )/J ? and control-action cardinality
difference of the (sub-optimal) ADMM-based heuristic solution from the
optimal cost for 577 problem instances of third order plant model (20).
0.6
0
x3 (t)
0.4
x2 (t)
0.5
0.2
50
exhaustive search
ADMM
0.6
0.4
40
0
10
20
30
40
50
0
10
20
30
40
50
0
10
20
30
40
50
2
0.2
0.2
0
0
0.1
0.2
0.3
(Jgreedy − J ? )/J ?
0.4
0.6
1
0
0
0.1
0.2
0.3
(Jgreedy − J ? )/J ?
0.4
0.4
0
0.1
0.2
0.3
(Jgreedy − J ? )/J ?
0.4
u(t)
0
0.4
0.4
0.2
0
−1
2
0.2
0
0
-5
-4
-3
-2 -1
0
|Tgreedy | − |T ? |
1
0
-5
-4
-3
-2 -1
0
|Tgreedy | − |T ? |
1
-5
-4
-3 -2 -1 0
|Tgreedy | − |T ? |
1
2
Fig. 6. Relative error (Jgreedy −
and control-action cardinality
difference of the greedy solution from the optimal cost for 577 problem
instances of third order plant model (20).
J ? )/J ?
5%) in almost all cases, and for ε = 0.4 and ε = 0.6, the
algorithm finds solutions with relative error less than 5% in
more than 80% of problem instances. The average optimality
gap (Jadmm − J ? )/J ? for the three cases were 0.0035, 0.0237,
and 0.0716, respectively. Finally, by comparing the cardinality
of control event index T for the optimal solutions and our
ADMM method, one concludes that the heuristic method in
more than 50% of cases finds similar-sparse solution as the
optimum ones, and in the rest of cases it tends to find more
sparse control actions.
We also compared the true optimal control performance
with a greedy heuristic method, shown in Algorithm 1, for
three sets of problems, mentioned above, with ε = 0.2,
ε = 0.4, and ε = 0.6. The greedy algorithm detected 22,
90, and 135 optimal solutions of the aforementioned 577
problems, respectively. Also, it computed the optimal solution
with relative error less than 5% in more than 65% of problems
at hand for all three problem setups. It is important to note
that the greedy heuristic method always provided feasible
solutions for all problem cases. The average optimality gaps
(Jgreedy − J ? )/J ? in all three cases were 0.0514, 0.0780,
and 0.0826, respectively. As compared to the ADMM-based
heuristic method, the optimality gaps became larger.
B. Receding horizon control implementation
We consider the open-loop unstable discrete-time system
represented by (20) with Bw = 0. The prediction horizon
uadmm (t)
0.2
1
0
−1
Fig. 7. States and input versus time for the event-triggered model predictive
controller presented in Section VI-B. In this figure, x1 (t), x2 (t) and x3 (t)
represent state variables whereas u(t) and uadmm (t) denote the optimal and
the heuristic control inputs computed based on the exhaustive search and the
ADMM algorithm, respectively.
is chosen to be N = 6 with performance indices Q =
P = diag{2, 2, 2} and R = 5. The event-threshold ε for
the event-triggered receding horizon implementation is set to
0.4. Fig. 7 demonstrates the time responses of the eventtriggered receding
horizonicontrol system for the initial conh
√
√
>
dition x0 = 0, 22 , − 22 . Here, two different approaches
for computing control input trajectories are considered: the
exhaustive search algorithm and the heuristic ADMM algorithm with ρ = 4.8. As can be seen in Fig. 7, while the state
trajectories associated to two control approaches deviate from
one another (except for the few first sampling instants that
they match together), the corresponding control inputs follow
a rather similar sparsity pattern. The event-triggered receding
horizon control via exhaustive search algorithm uses the communication channel 14 times over the 50 sampling instants,
which results in a reduction of 72% in communication and
computational burden, whereas the event-triggered receding
horizon control via ADMM algorithm uses the channel 16
times over the 50 sampling instants, which leads to a reduction
of 68% in communication and computational burden. For the
aforementioned initial value, the exhaustive search achieves a
better control cost: 65.42 compared to 77.72 as of the control
cost of the ADMM-based heuristic. However, this conclusion
does not hold in general. In particular, we experienced cases
DRAFT.
9
4
175
3
150
2
125
1
Control loss J∞
200
100
0.6
0.7
0.8
0.9
programming formulation to obtain the optimal solution via
solving an exponential number of quadratic programs. Then,
we proposed a heuristic algorithms to reduce the computational efforts while achieving an acceptable performance.
Later, we provided a complete stability analysis of receding
horizon control that uses a finite-horizon optimization in the
proposed class. Our numerical study confirmed the theory.
0
Communication rate π∞
Fig. 8. A comparison of the control performance and the communication frequency for different event-threshold values ε > 0 (shown in gray scale). The
marked curve with is obtained by averaging 2, 500 Monte Carlo simulations
for the horizon length 500 samples with the process noise {w(t)}t∈N0 and
the initial condition x0 generated randomly.
where starting from a different initial value, the receding horizon controller derived by exhaustive search performs worse
than the one based on ADMM-based heuristic.
C. Trade-off between communication rate and control performance
We consider the case where a random disturbance w(t) ∈
Rn , which is modeled by a discrete-time zero-mean Gaussian
white process with co-variance Σw , is acting on the plant
characterized by the injection matrix Bw = diag{1, 1, 1}. The
initial condition x0 is modeled as a random variable having a
normal distribution with zero mean and co-variance Σx0 . The
process noise w(t) is independent of the initial condition x0 .
The control performance is measured by a quadratic function:
J∞ =
499
1 X |
x (t)Qx(t) + u| (t)Ru(t) ,
500 t=0
(21)
while the communication rate is measured by
499
π∞ =
1 X
1
.
500 t=0 {kx(t)k∞ ≥0}
(22)
The visualization of the trade-off between the control loss
and the communication rate is demonstrated in Fig. 8. Different data points on the curve correspond to different eventthreshold values ranging from 0.5 to 4.0, and the control
loss (21) and the communication rate (22) are evaluated via
Monte Carlo simulations. Note that the communication rate
decreases dramatically with an increased control loss as the
threshold ε varies between 0 and 2.25 (dark colors). For
ε > 2.25 (lighter color), both quantities become less sensitive
to changes in the threshold value. We can also identify
0 ≤ ε ≤ 1.5 as a particularly attractive region where a large
decrease in the communication rate can be obtained for a small
loss in control performance.
VII. C ONCLUSION
In this paper, we considered a feedback control system
where an event-triggering rule dictates the communication
between the controller and the actuator. We investigated the
optimal control problem with an additional constraint on
the event-triggered communication. While this optimization
problem, in general, is non-convex, we used a disjunctive
VIII. A PPENDIX
We begin with the definition of the practical stability of
control systems.
Definition 1 (practical stability [45]) The system (1) is said
to be uniformly practically asymptotically stable in A ⊆ Rn
if A is a positively invariant set for (1) and if there exist a
KL-function β, and a nonnegative constant δ ≥ 0 such that
k x(t) k≤ β(k x0 k, t) + δ .
Definition 2 (practical-Lyapunov function [45]) A function
V : Rn → R≥0 is said to be a practical-Lyapunov function
in A for the system (1) if A is a positively invariant set and
if there exist a compact set, Ω ⊆ A, neighbourhood of the
origin, some K∞ -functions α1 , α2 and α3 , and some constants
d1 , d2 ≥ 0, such that
V (x) ≥ α1 (k x k), ∀x ∈ A ,
V (x) ≤ α2 (k x k) + d1 , ∀x ∈ Ω ,
V (f (x)) − V (x) ≤ −α3 (k x k) + d2 , ∀x ∈ A .
(23)
(24)
(25)
Proof of Theorem 1. To prove the practical stability of
the system (1) with a set of control actions π ? = {u? (t |
t), · · · , u? (t + N − 1 | t)}, which fulfills constraints (3)
and (4), we will analyze the value function,
VN? (x(t)) , min J(x(t), π),
π
where J(x(t), π) is as seen in (5). We borrow the shifted
sequence approach, described in [46], and we use a feasible
control sequence, i.e., π̃ = {u? (t + 1 | t), · · · , u? (t + N − 1 |
t), û}. By the optimality property, we obtain the following
bound (with x(t | t) = x(t) and u? (t | t) = u(t)):
VN? (x(t + 1)) − VN? (x(t)) ≤ VN (x(t + 1), π̃) − VN? (x(t))
= −`(x(t), u(t)) + Vf Ax(t + N | t) + B û
− Vf (x(t + N | t)) + `(x(t + N | t), û)
(26)
for all x(t) ∈ Rn . We now investigate the quantity of
∆Vf (t) + `(x(t + N | t), û) ,
where ∆Vf (t) , Vf Ax(t+N | t)+B û −Vf (x(t+N | t)).
Based on x(t + N | t), there are two possibilities:
• Suppose that k x(t + N | t) k∞ > ε, then for any feasible
control law, i.e., û = Kx(t + N | t), the following
inequality holds:
∆Vf (t) + `(x(t + N | t), û)
= x(t + N | t)| A|K P AK − P + Q∗ x(t + N | t) < 0 ,
(27)
DRAFT.
10
where AK = A+BK (i.e., Schur) and Q∗ = Q+K | RK.
From (26) and (27), it follows that
VN? (x(t
•
+ 1)) −
VN? (x(t))
<0.
which implies that
k x(t + 1) k22 ≤ γ
a3
η
k x(t) k22 + .
a2
a2
(33)
a
Since 0 < a2 ≤ a3 , it follows that γ = 1 − a2 , i.e., 0 < γ ≤
3
1. Therefore, by iterating (33), it is possible to exponentially
bound the state evolution via:
∆Vf (t)+`(x(t + N | t), û)
a
1 − γt η
k x(t) k22 ≤ γ t 3 k x(0) k22 +
.
|
|
= x(t + N | t) A P A − P + Q x(t + N | t)
a2
1 − γ a2
√
≤ x(t + N | t)| A| P A + Q x(t + N | t)
Using the norm inequality k v k∞ ≤k v k2 ≤ n k v k∞ for
|
2
n
≤ λmax A P A + Q k x(t + N | t) k2
any v ∈ R (see [47]), we get:
|
2
≤ λmax A P A + Q n k x(t + N | t) k∞
a
1 − γt η
2
k x(t) k2∞ ≤ γ t n 3 k x(0) k2∞ +
.
|
≤ λmax A P A + Q nε ,
a2
1 − γ a2
√
where we used the norm inequality k v k2 ≤ n k v k∞ Thus, lim k x(t) k∞ ≤ µ.
t→∞
n
for any v ∈ R ; see [47].
Suppose that k x(t + N | t) k∞ ≤ ε, then û = 0. Hence,
it follows that:
As a result, we conclude that, for any x(t + N | t) ∈ Rn , the
following inequality holds:
∆Vf (t) + `(x(t + N | t), û) ≤ η ,
(28)
2
|
where η = λmax A P A + Q nε . Inserting (28) and a1 k
x(t) k22 ≤ `(x(t), u(t)) where a1 , λmin (Q) into (26) yields:
VN? (x(t + 1)) − VN? (x(t)) ≤ −λmin (Q) k x(t) k22 +η .
(29)
Now, let the binary variable δ(t) ∈ {0, 1}, for each time
instant t, represent the transmission of the control action as
follows:
(
0 if k x(t) k∞ ≤ ε ,
δ(t) =
(30)
1 otherwise .
For any feasible control sequence, i.e., π 0
=
{Kx(t), · · · , Kx(t + N − 1
|
t)}, there exists
an
associated
scheduling
sequence,
denoted
by
∆(t) = {δ(t), · · · , δ(t + N − 1)}. Applying these feasible
control inputs, the upper bound of VN? (x) becomes:
VN? (x(t)) ≤ x| (t)S∆(t) x(t) ≤ λmax (S∆(t) ) k x(t) k22 ,
≤ a3 k x(t) k22 ,
(31)
where a3 , max λmax (S∆ ) and
∆∈S
S∆ = Qδ(0) +
N
−1 i−1
X
Y
i=1 j=0
A|δ(j) Qδ(j) Aδ(j) +
N
−1
Y
A|δ(k) QAδ(k) ,
k=0
with Qδ(i) , δ(i)Q∗ + (1 − δ(i))Q and Aδ(i) , δ(i)AK +
(1 − δ(i))A for all i ∈ {0, · · · , N − 1}. Here, S denotes a
set of all possible feasible schedules generated by a stabilizing
control sequence π 0 .
To obtain the lower bound, we have x| (t)Qx(t) ≤
VN? (x(t)) and hence a2 k x(t) k22 ≤ VN? (x(t)) where a2 ,
λmin (Q). From, (29) and (31) we establish the following
relation:
a2
?
VN (x(t + 1)) ≤ 1 −
V ? (x(t)) + η ,
(32)
a3 N
R EFERENCES
[1] B. D. O. Anderson and J. B. Moore, Optimal Control: Linear Quadratic
Methods. Englewood Cliffs, NJ: Prentice Hall, 1990.
[2] D. P. Bertsekas, Dynamic Programming and Optimal Control. Athena
Scientific, 2000.
[3] K. J. Åström, Introduction to Stochastic Control Theory.
Dover
Publications Inc., 2006.
[4] M. Gallieri and J. M. Maciejowski, “`asso MPC: Smart regulation of
overactuated systems,” in American Control Conference, 2012.
[5] E. N. Hartley, M. Gallieri, and J. M. Maciejowski, “Terminal spacecraft rendezvous and capture with LASSO model predictive control,”
International Journal of Control, vol. 86, no. 11, pp. 2104–2113, Aug.
2013.
[6] M. Chyba, S. Grammatico, V. T. Huynh, J. Marriott, B. Piccoli, and R. N.
Smith, “Reducing actuator switchings for motion control for autonomous
underwater vehicle,” in Proceeding of the American Control Conference,
2013.
[7] V. Gupta, “On a control algorithm for time-varying processor availability,” in Proceedings of the Hybrid Systems, Control and Computation
Conference, 2010.
[8] B. Demirel, V. Gupta, D. E. Quevedo, and M. Johansson, “Threshold optimization of event-triggered multi-loop control systems,” in Proceedings
of the 13th International Workshop on Discrete Event Systems, 2016.
[9] O. C. Imer and T. Başar, “Optimal control with limited controls,” in
Proceedings of the American Control Conference, 2006.
[10] P. Bommannavar and T. Başar, “Optimal control with limited control
actions and lossy transmissions,” in Proceedings of the 47th IEEE
Conference on Decision and Control, 2008.
[11] L. Shi, Y. Yuan, and J. Chen, “Finite horizon LQR control with limited
controller-system communication,” IEEE Transactions on Automatic
Control, vol. 58, no. 7, pp. 1815–1841, July 2013.
[12] J. Gao, “Cardinality constrained discrete-time linear-quadratic control,”
The Chinese University of Hong Kong, Master of Philosophy Thesis,
July 2005.
[13] J. Gao and D. Li, “Cardinality constrained linear-quadratic optimal
control,” IEEE Transactions on Automatic Control, vol. 56, no. 8, pp.
1936–1941, Aug. 2011.
[14] M. Nagahara and D. E. Quevedo, “Sparse representations for packetized
predictive networked control,” in IFAC World Congress, 2011.
[15] M. Gallieri and J. M. Maciejowski, “Stabilising terminal cost and
terminal controller for `asso -MPC: Enhanced optimality and region of
attraction,” in European Control Conference, 2013.
[16] M. Nagahara, D. E. Quevedo, and J. Østergaard, “Sparse packetized
predictive control networked control over erasure channels,” IEEE
Transactions on Automatic Control, vol. 59, no. 7, pp. 1899–1905, July
2014.
[17] R. P. Aguilera, R. Delgado, D. Dolz, and J. C. Agüero, “Quadratic
MPC with `0 -input constraint,” in Proceedings of the 19th IFAC World
Congress, 2014.
[18] K. J. Åström and B. M. Bernhardsson, “Comparison of riemann and
lebesgue sampling for first order stochastic systems,” in Proceedings of
the 41st IEEE Conference on Decision and Control, 2002.
DRAFT.
[19] D. Antunes and W. Heemels, “Rollout event-triggered control: Beyond
periodic control performance,” IEEE Transactions on Automatic Control,
vol. 59, no. 12, pp. 3296–3311, Dec. 2014.
[20] T. Gommans, D. Antunes, T. Donkers, P. Tabuada, and M. Heemels,
“Self-triggered linear quadratic control,” Automatica, vol. 50, pp. 1279–
1287, 2014.
[21] T. M. P. Gommans and W. P. M. H. Heemels, “Resource-aware MPC
for constrained nonlinear systems: A self-triggered control approach,”
Systems & Control Letters, vol. 79, pp. 59–67, May 2015.
[22] B. Demirel, V. Gupta, D. E. Quevedo, and M. Johansson, “On the tradeoff between communication and control cost in event-triggered dead-beat
control,” To Appear in IEEE Transactions on Automatic Control, 2016.
[23] J. Sijs, M. Lazar, and W. P. M. H. Heemels, “On integration of eventbased estimation and robust MPC in a feedback loop,” in Proceedings
of the 13th ACM International Conference on Hybrid Systems: Computation and Control, 2010.
[24] A. Eqtami, D. V. Dimarogonas, and K. J. Kyriakopoulos, “Eventtriggered control for discrete-time systems,” in Proceedings of American
Control Conference, 2010.
[25] ——, “Event-triggered strategies for decentralized model predictive
controllers,” in Proceedings of the 18th IFAC World Congress, 2011.
[26] D. Lehmann, E. Henriksson, and K. H. Johansson, “Event-triggered
model predictive control of discrete-time linear systems subject to
disturbances,” in Proceeding of the European Control Conference, 2013.
[27] E. Henriksson, D. E. Quevedo, E. G. W. Peters, H. Sandberg, and K. H.
Johansson, “Multiple-loop self-triggered model predictive control for
network scheduling and control,” IEEE Transactions on Control Systems
Technology, vol. 23, no. 6, pp. 2167–2181, Nov. 2015.
[28] N. He and D. Shi, “Event-based robust sampled-data model predictive
control: A non-monotonic Lyapunov function approach,” IEEE Transactions on Circuits and Systems I: Regular Papers, vol. 62, no. 10, pp.
2555–2564, Oct. 2015.
[29] F. D. Brunner, M. Heemels, and F. Allgöwer, “Robust self-triggered
MPC for constrained linear systems: A tube-based approach,” Automatica, vol. 72, pp. 73–83, Oct. 2016.
[30] P. Varutti, B. Kern, T. Faulwasser, and R. Findeisen, “Event-based model
predictive control for networked control systems,” in Proceedings of the
48th IEEE Conference on Decision and Control held jointly with the
28th Chinese Control Conference, 2009.
[31] H. Li and Y. Shi, “Event-triggered robust model predictive control of
continuous-time nonlinear systems,” Automatica, vol. 50, no. 5, pp.
1507–1513, May 2014.
[32] P. Varutti and R. Findeisen, “Event-based NMPC for networked control
11
[33]
[34]
[35]
[36]
[37]
[38]
[39]
[40]
[41]
[42]
[43]
[44]
[45]
[46]
[47]
systems over UDP-like communication channels,” in Proceedings of
American Control Conference, 2011.
A. Bemporad, G. Ferrari-Trecate, and M. Morari, “Observability and
controllability of piecewise affine and hybrid systems,” IEEE Transactions on Automatic Control, vol. 45, no. 10, pp. 1864–1876, Oct. 2000.
W. Heemels, B. D. Schutter, and A. Bemporad, “Equivalence of hybrid
dynamical models,” Automatica, vol. 37, no. 7, pp. 1085–1091, July
2001.
F. Zhu and P. J. Antsaklis, “Optimal control of hybrid switched systems:
A brief survey,” Discrete Event Dynamic Systems, vol. 25, no. 3, pp.
345–364, May 2014.
A. Michalka, “Cutting planes for convex objective nonconvex optimization,” Ph.D. dissertation, Columbia University, 2013.
D. Bienstock and A. Michalka, “Cutting-planes for optimization of
convex functions over nonconvex sets,” SIAM Journal on Optimization,
vol. 24, no. 2, pp. 643–677, 2014.
S. Diamond, R. Takapoui, and S. Boyd, “A general system for
heuristic solution of convex problems over nonconvex sets,” 2016,
arXiv:1601.07277.
S. Boyd, N. Parikh, E. Chu, B. Peleato, and J. Eckstein, “Distributed
optimization and statistical learning via the alternating direction method
of multipliers,” Foundations and Trends in Machine Learning, vol. 3,
no. 1, pp. 1–122, 2011.
R. Takapoui, N. Moehle, S. Boyd, and A. Bemporad, “A simple effective
heuristic for embedded mixed-integer quadratic programming,” in To
appear in Proceedings American Control Conference, 2016.
N. Derbinsky, J. Bento, V. Elser, and J. S. Yedidia., “An improved threeweight message-passing algorithm,” arXiv:1305.1961, 2013.
E. Ghadimi, A. Teixeira, I. Shames, and M. Johansson, “Optimal
parameter selection for the alternating direction method of multipliers
(ADMM): quadratic problems,” IEEE Transactions on Automatic Control, vol. 60, no. 3, pp. 644 – 658, March 2015.
P. Giselsson and S. Boyd, “Linear convergence and metric selection in douglas rachford splitting and admmlitting and ADMM,”
arXiv:1410.8479v5, 2016.
S. Boyd and L. Vandenberghe, Convex Optimization.
Cambridge
University Press, 2004.
Z.-P. Jiang and Y. Wang, “A converse Lyapunov theorem for discretetime systems with disturbances,” Systems & Control Letters, vol. 45, pp.
49–58, Aug. 2002.
J. B. Rawlings and D. Q. Mayne, Model Predictive Control: Theory and
Design. Nob Hill Publishing, 2009.
D. S. Bernstein, Matrix Mathematics. Princeton University Press, 2009.
| 3cs.SY
|
Variable screening with multiple studies
Tianzhou Ma1 , Zhao Ren2 and George C. Tseng1
1
Department of Biostatistics, University of Pittsburgh
2
Department of Statistics, University of Pittsburgh
arXiv:1710.03892v1 [stat.ME] 11 Oct 2017
Abstract
Advancement in technology has generated abundant high-dimensional data that allows integration of multiple relevant studies. Due to their huge computational advantage, variable screening methods based on marginal correlation have become promising
alternatives to the popular regularization methods for variable selection. However, all
these screening methods are limited to single study so far. In this paper, we consider
a general framework for variable screening with multiple related studies, and further
propose a novel two-step screening procedure using a self-normalized estimator for highdimensional regression analysis in this framework. Compared to the one-step procedure
and rank-based sure independence screening (SIS) procedure, our procedure greatly reduces false negative errors while keeping a low false positive rate. Theoretically, we
show that our procedure possesses the sure screening property with weaker assumptions on signal strengths and allows the number of features to grow at an exponential
rate of the sample size. In addition, we relax the commonly used normality assumption
and allow sub-Gaussian distributions. Simulations and a real transcriptomic application illustrate the advantage of our method as compared to the rank-based SIS method.
Key words and phrases: Multiple studies, Partial faithfulness, Self-normalized estimator, Sure screening property, Variable selection
1
Introduction
In many areas of scientific disciplines nowadays such as omics studies (including genomics,
transcriptomics, etc.), biomedical imaging and signal processing, high dimensional data with
much greater number of features than the sample size (i.e. p >> n) have become rule rather
than exception. For example, biologists may be interested in predicting certain clinical
outcome (e.g. survival) using the gene expression data where we have far more genes than
the number of samples. With the advancement of technologies and affordable prices in recent
biomedical research, more and more experiments have been performed on a related hypothesis
or to explore the same scientific question. Since the data from one study often have small
sample size with limited statistical power, effective information integration of multiple studies
can improve statistical power, estimation accuracy and reproducibility. Direct merging of the
data (a.k.a. “mega-analysis”) is usually less favored due to the inherent discrepancy among
the studies (Tseng et al., 2012). New statistical methodologies and theories are required to
solve issues in high-dimensional problem when integrating multiple related studies.
Various regularization methods have been developed in the past two decades and frequently used for feature selection in high-dimensional regression problems. Popular methods
include, but are not limited to, Lasso (Tibshirani, 1996), SCAD (Fan and Li, 2001), elastic
1
net (Zou and Hastie, 2005) and adaptive Lasso (Zou, 2006). When group structure exists
among the variables (for example, a set of gene features belonging to a pre-specified pathway), group version of regularization methods can be applied (Yuan and Lin, 2006; Meier
et al., 2008; Nardi et al., 2008). One can refer to Fan and Lv (2010) and Huang et al. (2012)
for a detailed overview of variable selection and group selection in high-dimensional models.
When the number of features grows significantly larger than the sample size, most regularization methods perform poorly due to the simultaneous challenges of computation expediency,
statistical accuracy and algorithmic stability (Fan et al., 2009). Variable screening methods
become a natural way to consider by first reducing to a lower or moderate dimensional problem and then performing variable regularization. Fan and Lv (2008) first proposed a sure
independent screening (SIS) method to select features based on their marginal correlations
with the response in the context of linear regression models and showed such fast selection
procedure enjoyed a “sure screening property”. Since the development of SIS, many screening methods have been proposed for generalized linear models (Fan et al., 2009, 2010; Chang
et al., 2013), nonparametric additive models or semiparametric models (Fan et al., 2011;
Chang et al., 2016), quantile linear regression (Ma et al., 2017), Gaussian graphical models
(Luo et al., 2014; Liang et al., 2015) or exploit more robust measures for sure screening (Zhu
et al., 2011; Li et al., 2012, 2017). However, all these screening methods are limited to single
study so far.
In this paper, we first propose a general framework for simultaneous variable screening
with multiple related studies. Compared to single study scenario, inclusion of multiple studies
gives us more evidence to reduce dimension and thus increases the accuracy and efficiency
of removing unimportant features during screening. To our knowledge, our paper is the first
to utilize multiple studies to help variable screening in high-dimensional linear regression
model. Such a framework provides a novel perspective to the screening problem and opens
a door to the development of methods using multiple studies to perform screening under
different types of models or with different marginal utilities. In this framework, it is natural
to apply a selected screening procedure to each individual study, respectively. However,
important features with weak signals in some studies may be incorrectly screened out if only
such a one-step screening is performed. To avoid such false negative errors and fully take
advantage of multiple studies, we further propose a two-step screening procedure, where one
additional step of combining studies with potential zero correlation is added to the one-step
procedure for a second check. This procedure has the potential to save those important
features with weak signals in individual studies but strong aggregate effect across studies
during the screening stage. Compared to the naive multiple study extension of SIS method,
our procedure greatly reduces the false negative errors while keeping a low false positive
rate. These merits are confirmed by our theoretical analysis. Specifically, we show that
our procedure possesses the sure screening property with weaker assumptions on signals and
allows the number of features to grow at an exponential rate of the sample size. Furthermore,
we only require the data to have sub-Gaussian distribution via using novel self-normalized
statistics. Thus our procedure can be applied to more general distribution family other than
Gaussian distribution, which is considered in Fan and Lv (2008) and Bühlmann et al. (2010)
for a related screening procedure under single study scenarios. After screening, we further
apply two general and applicable variable selection algorithms: the multiple study extension
of PC-simple algorithm proposed by Bühlmann et al. (2010) as well as a two-stage feature
selection method to choose the final model in a lower dimension.
2
The rest of the paper is organized as follows. In Section 2, we present a framework for
variable screening with multiple related studies as well as notations. Then we propose our
two-step screening procedure in Section 3. Section 4 provides the theoretical properties of our
procedure, and demonstrates the benefits of multiple related studies as well as the advantages
of our procedure. General algorithms for variable selection that can follow from our screening
procedure are discussed in Section 5. Section 6 and 7 include the simulation studies and
a real data application on three breast cancer transcriptomic studies, which illustrate the
advantage of our method in reducing false negative errors and retaining important features
as compared to the rank-based SIS method. We conclude and discuss possible extensions of
our procedure in Section 8. Section 9 provides technical proofs to the major theorems.
2
Model and Notation
Suppose we have data from K related studies, each has n observations. Consider a random
design linear model in each study k ∈ [K] ([K] = 1, . . . , K):
Y
(k)
=
p
X
(k)
(k)
βj Xj + (k) ,
(2.1)
j=1
(k)
(k)
(k)
where each Y (k) ∈ R, each X (k) = (X1 , . . . , Xp )T ∈ Rp with E(X (k) ) = µX and
(k)
cov(X (k) ) = ΣX , each (k) ∈ R with E((k) ) = 0 and var((k) ) = σ 2 such that (k) is
(k)
(k)
(k)
(k)
uncorrelated with X1 , . . . , Xp , and β (k) = (β1 , . . . , βp )T ∈ Rp . We assume implicitly
(k)
with E(Y (k)2 ) < ∞ and E{(Xj )2 } < ∞ for j ∈ [p] ([p] = 1, . . . , p).
When p is very large, we usually assume that only a small set of covariates are true
predictors that contribute to the response. In other words, we assume most of βj =
(1)
(K)
(βj , . . . , βj )T , where j ∈ [p], are equal to a zero vector. In addition, in this paper, we
(k)
assume βj ’s are either zero or non-zero in all K studies. This framework is partially motivated by a high-dimensional linear random effect model considered in literature (e.g.,Jiang
T
, 0T )T , where β(1) is the vector of the
et al. (2016)). More specifically, we can have β = (β(1)
first s0 non-zero components of β (1 ≤ s0 ≤ p). Consider a random effect model where only
(k)
the true predictors of each study are treated as the random effect, that is, β (k) = (β(1) , 0)T
(k)
and β(1) is distributed as N (β(1) , τ 2 Is0 ), where τ 2 is independent of and X. Consequently,
(k)
βj ’s are either zero or non-zero in all K studies with probability one. Such assumption
fits the reality well, for example, in a typical GWAS study, a very small pool of SNPs are
reported to be associated with a complex trait or disease among millions (Jiang et al., 2016).
With n i.i.d. observations from model (2.1), our purpose is to identify the non-zero β(1) ,
thus we define the following index sets for active and inactive predictors:
(k)
6= 0 for all k};
(k)
= 0 for all k},
A = {j ∈ [p]; βj 6= 0} = {j ∈ [p]; βj
AC = {j ∈ [p]; βj = 0} = {j ∈ [p]; βj
(2.2)
where A is our target. Clearly, under our setting, A and AC are complementary to each
other so that the identification of AC is equivalent to the identification of A. Let |A| = s0 ,
where | · | denotes the cardinality.
3
3
Screening procedure with multiple studies
3.1
Sure independence screening
For a single study (K = 1), Fan and Lv (2008) first proposed the variable screening method
called sure independence screening (SIS) which ranked the importance of variables according
to their marginal correlation with the response and showed its great power in preliminary
screening and dimension reduction for high-dimensional regression problems. Bühlmann
et al. (2010) later introduced the partial faithfulness condition that a zero partial correlation
for some separating set S implied a zero regression coefficient and showed that it held almost
surely for joint normal distribution. In the extreme case when S = ∅, it is equivalent to the
SIS method.
The purpose of sure screening is to identify a set of moderate size d (with d << p) that
will still contain the true set A. Equivalently, we can try to identify AC or subsets of AC
which contain unimportant features that need to be screened out. There are two potential
errors that may occur in any sure screening methods (Fan and Lv, 2010):
1. False Negative (FN): Important predictors that are marginally uncorrelated but
jointly correlated with the response fail to be selected.
2. False Positive (FP): Unimportant predictors that are highly correlated with the
important predictors can have higher priority to be selected than other relatively weaker
important predictors.
The current framework for variable screening with multiple studies is able to relieve us
from the FP errors significantly. Indeed, we have multiple studies in our model setting thus
we have more evidence to exclude noises and reduce FP errors than single study. In addition,
sure screening is used to reduce dimension at a first stage, so we can always include a second
stage variable selection methods such as Lasso or Dantzig selection to further refine the set
and reduce FP errors.
The FN errors occur when signals are falsely excluded after screening. Suppose ρj is
the marginal correlation of the jth feature with the response, with which we try to find the
set {j : ρj = 0} to screen out. Under the assumption of partial faithfulness (for explicit
definition, see Section 4.3), these variables have zero coefficients for sure so the FN errors
are guaranteed to be excluded. However, this might not be true for the empirical version
of marginal correlation. For a single study (K = 1), to rule out the FN errors in empirical
case, it is well-known that the signal-to-noise ratio has to be large (at least of an order of
(log p/n)1/2 after Bonferroni adjustment). In the current setting with multiple studies, the
requirement on strong signals remains the same if we naively perform one-step screening in
each individual study. As we will see next, we propose a novel two-step screening procedure
which allows weak signals in individual studies as long as the aggregate effect is strong
enough. Therefore our procedure is able to reduces FN errors in the framework with multiple
studies.
Before closing this section, it is worthwhile to mention that, to perform a screening test,
one usually applies Fisher’s z-transformation on the sample correlation (Bühlmann et al.,
2010). However, this will require the bivariate normality assumption. Alternatively, in this
paper, we propose to use the self-normalized estimator of correlation that works generally well
even for non-Gaussian data (Shao, 1999). Similar ideas have been applied in the estimation
of large covariance matrix (Cai and Liu, 2016).
4
3.2
Two-step screening procedure with multiple studies
(k)
In the presence of multiple studies, we have more evidence to reduce dimension and ρj = 0
for any k will imply a zero coefficient for that feature. On one hand, it is possible for features
(k)
with zero βj to have multiple non-zero ρj ’s. On the other hand, a non-zero βj will have
(k)
non-zero ρj ’s in all studies. Thus, we aim to identify the following two complementary sets
while performing screening with multiple studies:
(k)
A[0] = {j ∈ [p];
min |ρj | = 0},
A[1] = {j ∈ [p];
min |ρj | =
6 0}.
k
(k)
(3.1)
k
We know for sure that A[0] ⊆ AC and A ⊆ A[1] with the partial faithfulness assumption.
For j ∈ A[0] , the chance of detecting a zero marginal correlation in at least one study has
been greatly increased with increasing K, thus unimportant features will more likely be
screened out as compared to single study scenario.
(k)
One way to estimate A[1] is to test H0 : ρj = 0 of each k for each feature j. When any
of the K tests is not rejected for a feature, we will exclude this feature from Â[1] (we call it
the “One-Step Sure Independence Screening” procedure, or “OneStep-SIS” for short). This
can be viewed as an extension of the screening test to multiple study scenario. However, in
(k)
reality, it is possible for important features to have weak signals thus small |ρj |’s in at least
one study. These features might be incorrectly classified into Â[0] since weak signals can be
indistinguishable from null signals in individual testing. It will lead to the serious problem
of false exclusion of important features (FN) from the final set during screening.
This can be significantly improved by adding a second step to combine those studies
(k)
with potential zero correlation (i.e., fail to reject the null H0 : ρj = 0) identified in the
first step and perform another aggregate test. For the features with weak signals in multiple
studies, as long as their aggregate test statistics is large enough, they will be retained. Such
procedure will be more conservative in screening features as to the first step alone, but will
guarantee to reduce false negative errors.
(k)
(k)
For simplicity, we assume n i.i.d. observations (Xi , Yi ), i ∈ [n], are obtained from
all K studies. It is straightforward to extend the current procedure and analysis to the scenarios with different sample sizes across multiple studies, and thus omitted. Our proposed
“Two-Step Aggregation Sure Independence Screening” procedure (“TSA-SIS” for short) is
formally described below:
Step 1. Screening in each study
In the first step, we perform screening test in each study k ∈ [K] and obtain the estimate
of study set with potential zero correlations ˆlj for each j ∈ [p] as:
√ (k)
nσ̂
(k)
(k)
−1
ˆlj = {k; |T̂ | ≤ Φ (1 − α1 /2)} and T̂ = q j ,
(3.2)
j
j
(k)
θ̂j
P
P
(k)
(k)
(k)
(k)
(k)
(k)
where σ̂j = n1 ni=1 (Xij −X̄j )(Yi −Ȳ (k) ) is the sample covariance and θ̂j = n1 ni=1 [(Xij −
(k)
(k)
(k)
(k)
(k)
X̄j )(Yi − Ȳ (k) ) − σ̂j ]2 . T̂j is the self-normalized estimator of covariance between Xj
5
and Y (k) . Φ is the CDF of standard normal distribution and α1 the pre-specified significance
level.
(k)
In each study, we test if |T̂j | > Φ−1 (1 − α1 /2), if not, we will include study k into
ˆlj . This step does not screen out any variables, but instead separates potential zero and
non-zero study-specific correlations for preparation of the next step. Define the cardinality
of ˆlj as κ̂j = |ˆlj |. If κ̂j = 0 (i.e., no potential zero correlation), we will for sure retain feature
j and not consider it in step 2; Otherwise, we move on to step 2.
(k)
Remark 1. By the scaling property of T̂j , it is sufficient to impose assumptions on the stan(k)
(k)
(k)
dardized variables: W (k) = Y√ −E(Y(k) ) , Zj =
var(Y
)
(k)
(k)
Xj −E(Xj )
q
.
(k)
var(Xj )
(k)
Thus T̂j
(k)
can also be treated
(k)
as the self-normalized estimator of correlation. We thus can define θj = var(Zj W (k) ) and
(k)
(k)
(k)
σj = cov(Zj , W (k) ) = ρj .
Remark 2. In our analysis, the index set in (3.2) is shown to coincide with lj (j ∈ A[0] ) and
lj (j ∈ A[1] ) which will be introduced in more details in Section 4.
Step 2. Aggregate screening
In the second step, we wish to test whether the aggregate effect of potential zero correlations in ˆlj identified in step 1 is strong enough to be retained. Define the statistics
P (k) 2
L̂j =
(T̂j ) and this statistics will approximately follow a χ2κ̂j distribution with degree
k∈l̂j
of freedom κ̂j under null. Thus we can estimate Â[0] by:
Â[0] = {j ∈ [p]; L̂j ≤ ϕ−1
κ̂j (1 − α2 ) and κ̂j 6= 0},
(3.3)
or equivalently estimate Â[1] by:
Â[1] = {j ∈ [p]; L̂j > ϕ−1
κ̂j (1 − α2 ) or κ̂j = 0},
(3.4)
where ϕκ̂j is the CDF of chi-square distribution with degree of freedom equal to κ̂j and α2
the pre-specified significance level.
(k)
The second step takes the sum of squares of T̂j from studies with potential zero corP (k) 2
relation as the test statistics. For each feature j, we test if
(T̂j ) > ϕ−1
κ̂j (1 − α2 ). If
k∈l̂j
rejected, we conclude that the aggregate effect is strong and the feature needs to be retained,
otherwise, we will screen it out. This step performs a second check in addition to the individual testing in step 1 and potentially saves those important features with weak signals in
individual studies but strong aggregate effect.
In Table 1, we use a toy example to demonstrate our idea and compare the two approaches
(“OneStep-SIS” vs. “TSA-SIS”). In this example, suppose we have five studies (K = 5)
and three features (two signals and one noise). “S1” is a strong signal with β = 0.8 in all
studies, “S2” is a weak signal with β = 0.4 in all studies and “N1” is a noise with β = 0.
In hypothesis testing, both small β and zero β can give small marginal correlation and are
sometimes indistinguishable. Suppose T = 3.09 is used as the threshold (corresponding to
6
Table 1: Toy example to demonstrate the strength of two-step screening procedure.
k=1
k=2
k=3
k=4
k=5
ˆlj
κ̂j
TSA-SIS
L̂j
Â[0]
Â[1]
Â[0]
OneStep-SIS
Â[1]
S1 (signal)
S2 (signal)
N1 (noise)
(1)
(1)
(1)
|T̂1 | = 3.71
|T̂2 | = 3.70
|T̂3 | = 0.42
(2)
(2)
(2)
|T̂1 | = 3.16
|T̂2 | = 2.71
|T̂3 | = 0.54
(3)
(3)
(3)
|T̂1 | = 3.46
|T̂2 | = 2.65
|T̂3 | = 0.56
(4)
(4)
(4)
|T̂1 | = 3.63
|T̂2 | = 2.68
|T̂3 | = 0.12
(5)
(5)
(5)
|T̂1 | = 3.24
|T̂2 | = 1.94
|T̂3 | = 0.69
∅
{2, 3, 4, 5}
{1, 2, 3, 4, 5}
0
4
5
25.31 > ϕ4 (0.95) 1.27 < ϕ5 (0.95)
N
N
Y
Y
Y
N
N
Y
Y
Y
N (FN)
N
α1 = 0.001). For the strong signal “S1”, all studies have large marginal correlations, so both
“OneStep-SIS” and “TSA-SIS” procedures include it correctly. For the weak signal “S2”,
since in many studies it has small correlations, it is incorrectly screened out by “OneStepSIS” procedure (False Negative). However, the “TSA-SIS” procedure saves it in the second
step (with α2 = 0.05). For the noise “N1”, both methods tend to remove it after screening.
4
4.1
Theoretical properties
Assumptions and conditions
We impose the following conditions to establish the model selection consistency of our procedure:
(C1) (Sub-Gaussian Condition) There exist some constants M1 > 0 and η > 0 such that for
all |t| ≤ η, j ∈ [p], k ∈ [K]:
(k)2
E{exp(tZj
)} ≤ M1 ,
E{exp(tW (k)2 )} ≤ M1 .
(k)
In addition, there exist some τ0 > 0 such that min θj ≥ τ0 .
j,k
(C2) The number of studies K = O(pb ) for some constant b ≥ 0. The dimension satisfies:
log3 (p) = o(n) and κj log2 p = o(n), where κj is defined next.
(k)
(k)
(C3) For j ∈ A[0] , lj (j ∈ A[0] ) = {k; ρj = 0} and κj = |lj |. If k ∈
/ lj , then |ρj | ≥
q
q
(k)
log p
1.01θj , where C3 = 3(L + 1 + b).
C3
n
7
q
q
(k)
(k)
(C4) For j ∈ A[1] , lj (j ∈ A[1] ) = {k; |ρj | < C1 logn p 0.99θj } and κj = |lj |, where
q
q
(k)
(k)
C1 = L + 1 + b. If k ∈
/ lj , then |ρj | ≥ C3 logn p 1.01θj . In addition, we require
√
P (k) 2 C2 (log2 p+ κj log p)
|ρj | ≥
, where C2 is some large positive constant.
n
k∈lj
(k)
The first condition (C1) assumes that each standardized variable Zj or W (k) , j ∈ [p],
k ∈ [K], marginally follow a sub-Gaussian distribution in each study. This condition relaxes
the normality assumption in (Fan and Lv, 2008; Bühlmann et al., 2010). The second part
of (C1) assumes there always exist some positive τ0 not greater than the minimum variance
(k)
(k)
of Zj W (k) . In particular, if (Xj , Y (k) ) jointly follows a multivariate normal distribution,
(k)
(k)2
then θj = 1 + ρj ≥ 1, so we can always pick τ0 = 1.
The second condition (C2) allows the dimension p to grow at an exponential rate of
sample size n, which is a fairly standard assumption in high-dimensional analysis. Many
sure screening methods like “SIS”, “DC-SIS” and “TPC” have used this assumption (Fan
and Lv, 2008; Li et al., 2012, 2017). Though the PC-simple algorithm (Bühlmann et al.,
2010) assumes a polynomial growth of pn as a function of n, we notice that it can be readily
relaxed to an exponential of n level. Further, we require the product κj log2 p to be small,
which is used to control the errors in the second step of our screening procedure. It is always
true if K log2 p = o(n).
Conditions (C3) assumes a lower bound on non-zero correlation (i.e. k ∈
/ lj ) for features
(k)
[0]
from A . In other words, if the marginal correlation |ρj | is not zero, then it must have a
large enough marginal correlation to be detected. While this has been a key assumption for
a single study in many sure screening methods (Fan and Lv, 2008; Bühlmann et al., 2010;
Li et al., 2012, 2017), we only impose this assumption for j ∈ A[0] rather than all j ∈ [p].
This condition is used to control for type II error in step 1 for features from A[0] .
Condition (C4) gives assumptions on features from A[1] . We assume the correlations to
be small for those k ∈ lj and large for those k ∈
/ lj so that studies with strong or weak
signals can be well separated in the first step. This helps control the type II error in step
1 for features from A[1] . For those studies in lj , we further require their sum of squares of
correlations to be greater than a threshold, so that type II error can be controlled in step 2.
This condition is different from other methods with single study scenario, where they usually
assume a lower bound on each marginal correlation for features from A[1] just like (C3). We
relax this condition and only put restriction on their L2 norm. This allows features from A[1]
to have weak signals in each study but combined strong signal. To appreciate this relaxation,
we compare the minimal requirements with and without step 2. For each j ∈ A[1] , in order to
(k)
detect this feature, we need |ρj | ≥ C(log p/n)1/2 with some large constant C for all k ∈ lj ,
P (k) 2
and thus at least
|ρj | ≥ C 2 κj log p/n. In comparison, the assumption in (C4) is much
k∈lj
weaker in reasonable settings κj >> log p.
4.2
Consistency of the two-step screening procedure
We state the first theorem involving the consistency of screening in our step 1:
8
Theorem 1. Consider a sequence of linear models as in (2.1) which satisfy assumptions and
conditions (C1)-(C4), define the event A = {ˆlj = lj for √
all j ∈ [p]}, there exists a sequence
α1 = α1 (n, p) → 0 as (n, p) → ∞ where α1 = 2{1 − Φ(γ log p)} with γ = 2(L + 1 + b) such
that:
P (A) = 1 − O(p−L ) → 1 as (n, p) → ∞.
(4.1)
The proof of Theorem 1 can be found in Section 9. This theorem states that the screening
in our first step correctly identifies the set lj for features in both A[0] and A[1] (in which strong
and weak signals are well separated) and the chance of incorrect assignment is low. Given
the results in Theorem 1, we can now show the main theorem for the consistency of the
two-step screening procedure:
Theorem 2. Consider a sequence of linear models as in (2.1) which satisfy assumptions
and conditions (C1)-(C4), we know there exists a sequence
α1 = α1 (n, p) → 0 and α2 =
√
α2 (n, p) → 0 as (n, p) → ∞ where α1 = 2{1 −
pΦ(γ log p)} with γ = 2(L + 1 + b) and
α2 = 1 − ϕκj (γκj ) with γκj = κj + C4 (log2 p + κj log p) and some constant C4 > 0 such
that:
P {Â[1] (α1 , α2 ) = A[1] } = 1 − O(p−L ) → 1 as (n, p) → ∞.
(4.2)
The proof of Theorem 2 can be found in Section 9. The result shows that the two-step
screening procedure enjoys the model selection consistency and identifies the model specified
in (3.1) with high
√ probability. The choice of significance level that yields consistency is
α1 = 2{1 − Φ(γ log p)} and α2 = 1 − ϕκj (γκj ) .
4.3
Partial faithfulness and Sure screening property
Bühlmann et al. (2010) first came up with the partial faithfulness assumption which theoretically justified the use of marginal correlation or partial correlation in screening as follows:
ρj|S = 0 for some S ⊆ {j}C implies βj = 0,
(4.3)
where S is the set of variables conditioned on. For independence screening, S = ∅.
Under the two conditions: the positive definiteness of ΣX and non-zero regression coefficients being realization from some common absolutely continuous distribution, they showed
that partial faithfulness held almost surely (Theorem 1 in Bühlmann et al. (2010)). Since
the random effect model described in Section 2 also satisfies the two conditions, the partial
faithfulness holds almost surely in each study.
Thus, we can readily extend their Theorem 1 to a scenario with multiple studies:
Corollary 1. Consider a sequence of linear models as in (2.1) satisfying the partial faithfulness condition in each study and true active and inactive set defined in (2.2), then the
following holds for every j ∈ [p]:
(k)
ρj|S = 0 for some k for some S ⊆ {j}C implies βj = 0.
9
(4.4)
(k)
The proof is straightforward and thus omitted: if ρj|S = 0 for some study k, then with
(k)
partial faithfulness, we will have βj = 0 for that particular k. Since we only consider
(k)
features with zero or non-zero βj ’s in all studies in (2.2), we will have βj = 0. In the case
(k)
of independence screening (i.e. S = ∅), ρj = 0 for some k will imply a zero βj .
With the model selection consistency in Theorem 2 and the extended partial faithfulness
condition in Corollary 1, the sure screening property of our two-step screening procedure
immediately follows:
Corollary 2. Consider a sequence of linear models as in (2.1) which satisfy assumptions
and conditions (C1)-(C4) as well as the extended partial faithfulness condition in Corollary
1, there exists a sequence
√ α1 = α1 (n, p) → 0 and α2 = α2 (n, p) → 0 as (n, p) → ∞
where α1 = 2{1 − Φ(γ log p)} with γ = 2(L + 1 + b) and α2 = 1 − ϕκj (γκj ) with γκj =
p
κj + C4 (log2 p + κj log p) such that:
P {A ⊆ Â[1] (α1 , α2 )} = 1 − O(p−L ) → 1 as (n, p) → ∞.
(4.5)
The proof of this Corollary simply combines the results of Theorem 2 and the extended
partial faithfulness and is skipped here.
5
Algorithms for variable selection with multiple studies
Usually, performing sure screening once may not remove enough unimportant features. In our
case since there are multiple studies, we expect our two-step screening procedure to remove
many more unimportant features than in single study. If the dimension is still high after
applying our screening procedure, we can readily extend the two-step screening procedure
to an iterative variable selection algorithm by testing the partial correlation with gradually
increasing size of the conditional set S. Since such method is a multiple study extension of
the PC simple algorithm in Bühlmann et al. (2010), we call it “Multi-PC” algorithm (Section
5.1).
On the other hand, if the dimension has already been greatly reduced with the two-step
screening, we can simply add a second stage group-based feature selection techniques to
select the final set of variables (Section 5.2).
5.1
Multi-PC algorithm
We start from S = ∅, i.e., our two-step screening procedure and build a first set of candidate
active variables:
Â[1,1] = Â[1] = {j ∈ [p]; L̂j > ϕ−1
κ̂j (1 − α2 ) or κ̂j = 0}.
(5.1)
We call this set stage1 active set, where the first index in [, ] corresponds to the stage
of our algorithm and the second index corresponds to whether the set is for active variables
([, 1]) or inactive variables ([, 0]). If the dimensionality has already been decreased by a large
10
amount, we can directly apply group-based feature selection methods such as group lasso to
the remaining variables (to be introduced in Section 5.2).
However, if the dimension is still very high, we can further reduce dimension by increasing
the size of S and considering partial correlations given variables in Â[1,1] . We follow the
similar two-step procedure but now using partial correlation of order one instead of marginal
correlation and yield a smaller stage2 active set:
Â[2,1] = {j ∈ Â[1,1] ; L̂j|q > ϕκ̂−1j|q (1 − α2 ) or κ̂j|q = 0, for all q ∈ Â[1,1] \{j}},
(5.2)
where each self-normalized estimator of partial correlation can be computed by taking the
residuals from regressing over the variables in the conditional set.
We can continue screening high-order partial correlations, resulting in a nested sequence
of m active sets:
Â[m,1] ⊆ . . . ⊆ Â[2,1] ⊆ Â[1,1] .
(5.3)
Note that the active and inactive sets at each stage are non-overlapping and the union
of active and inactive sets at a stage m will be the active set in a previous stage m − 1, i.e.,
Â[m,1] ∪ Â[m,0] = Â[m−1,1] . This is very similar to the original PC-simple algorithm, but now
at each order-level, we perform the two-step procedure. The algorithm can stop at any stage
m when the dimension of Â[m,1] already drops to low to moderate level and other common
group-based feature selection techniques can be used to select the final set. Alternatively,
we can continue the algorithm until the candidate active set does not change anymore. The
algorithm can be summarized as follows:
Algorithm 1. Multi-PC algorithm for variable selection.
Step 1. Set m = 1, perform the two-step screening procedure to construct stage1 active set:
Â[1,1] = {j ∈ [p]; L̂j > ϕ−1
κ̂j (1 − α2 ) or κ̂j = 0}.
Step 2. Set m = m + 1. Construct the stagem active set:
Â[m,1] = {j ∈ Â[m−1,1] ; L̂j|S > ϕ−1
κ̂j|S (1 − α2 ) or κ̂j|S = 0,
for all S ⊆ Â[m−1,1] \{j} with |S| = m − 1}.
Step 3. Repeat Step 2 until m = m̂reach , where m̂reach = min{m : |Â[m,1] | ≤ m}.
11
5.2
Two-stage feature selection
As an alternative to “Multi-PC” algorithm for variable selection, we also introduce here a
two-stage feature selection algorithm by combining our two-step screening procedure and
other regular feature selection methods together. In single study, for example, Fan & Lv
(2008) performed sure independence screening in the first stage followed by model selection
techniques including Adaptive Lasso, Dantzig Selector and SCAD, etc., and named those
procedures as “SIS-AdaLasso”,“SIS-DS”, “SIS-SCAD” , accordingly.
In our case, since the feature selection is group-based, we adopt a model selection technique using group Lasso penalty in the second stage:
min
β
K
X
(k)
(k)
||y (k) − XÂ[1] βÂ[1] ||22 + λ
k=1
X
||βj ||2
,
(5.4)
j∈Â[1]
where Â[1] is the active set identified from our two-step screening procedure and the tuning
parameter λ can be chosen by cross-validation or BIC in practice just like for a regular group
Lasso problem. We call such two-stage feature selection algorithm as “TSA-SIS-groupLasso”.
In addition, at any stages of the “Multi-PC” algorithm when the dimension has already
been dropped to a moderate level, the group Lasso-based feature selection techniques can
always take over to select the final set of variables.
6
Numerical evidence
In this section, we demonstrate the advantage of TSA-SIS procedure in comparing to the
multiple study extension of SIS (named “Min-SIS”), which ranks the features by the minimum absolute correlation among all studies. We simulated data according to the linear
(k)
model in (2.1) including p covariates with zero mean and covariance matrix Σi,j = r|i−j|
(k)
(k)
where Σi,j denotes the (i, j)th entry of ΣX .
In the first part of simulation, we fixed the sample size n = 100, p = 1000, the number of
studies K = 5 and performed B = 1000 replications in each setting. We assumed that the
true active set consisted of only ten variables and all the other variables had zero coefficients
(i.e., s0 = 10). The indices of non-zero coefficients were evenly spaced between 1 and p.
The variance of the random error term in linear model was fixed to be 0.52 . We randomly
drew r from {0, 0.2, 0.4, 0.6} and allowed different r’s in different studies. We considered the
following four settings:
1. Homogeneous weak signals across all studies: nonzero βj generated from Unif(0.1, 0.3)
(1)
(2)
(K)
and βj = βj = . . . = βj = βj .
2. Homogeneous strong signals across all studies: nonzero βj generated from Unif(0.7, 1)
(1)
(2)
(K)
and βj = βj = . . . = βj = βj .
3. Heterogeneous weak signals across all studies: nonzero βj generated from Unif(0.1, 0.3)
(k)
and βj ∼ N (βj , 0.52 ).
12
Table 2: Sensitivity analysis on the choice of α1 and α2 in simulation (Sensitivity/Specificity)
Sensitivity/Specificity
α1 =0.01
0.001
0.0001
α2 = 0.15
0.05
0.01
0.001
0.793/0.901 0.525/0.984 0.210/0.999 0.142/1.000
0.947/0.826 0.864/0.943 0.691/0.990 0.373/0.999
0.966/0.816 0.922/0.932 0.840/0.985 0.681/0.998
Note: All value are based on average results from B = 1000 replications.
4. Heterogeneous strong signals across all studies: nonzero βj generated from Unif(0.7, 1)
(k)
and βj ∼ N (βj , 0.52 ).
We evaluated the performance of Min-SIS using receiver operating characteristic (ROC)
curves which measured the accuracy of variable selection independently from the issue of
choosing good tuning parameters (for Min-SIS, the tuning parameter is the top number of
features d). The OneStep-SIS procedure we mentioned above was actually one special case
of the Min-SIS procedure (by thresholding at α1 ). In presenting our TSA-SIS procedure,
we fixed α1 = 0.0001 and α2 = 0.05 so the result was just one point on the sensitivity vs.
1-specificity plot. We also performed some sensitivity analysis on the two cutoffs based on
the first simulation (see Table 2) and found the two values to be optimal since they had
both high sensitivity and high specificity. Thus we suggested fixing these two values in all
the simulations.
Figure 1 showed the results of simulation 1-4. When the signals were homogeneously weak
in all studies as in (1), TSA-SIS clearly outperformed the Min-SIS procedure (above its ROC
curve). It reached about 90% sensitivity with controlled false positive errors (specificity
∼ 95%). In order to reduce false negatives, Min-SIS had to sacrifice the specificity and
increased the false positives, which in the end lost the benefits of performing screening
(i.e. end up keeping too many features). When the signals became strong as in (2), both
procedures performed equally well. This fit our motivation and theory and showed the
strength of our two-step procedure in saving weak signals without much increase in false
positive rates. When the signals became heterogeneous as in (3) and (4), both procedures
performed worse than before. But the Min-SIS procedure never outperformed the TSASIS procedure since it only examined the minimum correlation among all studies while the
two-step procedure additionally considered the aggregate statistics.
7
Real data application
We next demonstrated our method in three microarray datasets of triple-negative breast
cancer (TNBC, sometimes a.k.a. basal-like), an aggressive subtype of breast cancer usually
with poor prognosis. Previous studies have shown that the tumor suppressor protein “p53”
played an important role in breast cancer prognosis and its expression was associated with
both disease-free survival and overall survival in TNBC (Yadav et al., 2015). Our purpose
was to identify the genes most relevant and predictive to the response - the expression level
of TP53 gene, which encodes p53 protein. The three datasets are publicly available on
authors’ website or at GEO repository including METABRIC (a large cohort consisting
of roughly 2000 primary breast tumours), GSE25066 and GSE76250 (Curtis et al., 2012;
13
Figure 1: Simulation results 1-4: the ROC curve is for Min-SIS, the black point is for our
TSA-SIS using α1 = 0.0001 and α2 = 0.05.
1.0
(2) Homogeneous strong
1.0
(1) Homogeneous weak
●
0.8
0.6
Sensitivity
0.0
0.2
0.4
0.6
0.4
0.0
0.2
Sensitivity
0.8
●
0.0
0.2
0.4
0.6
0.8
1.0
0.0
0.2
0.4
0.6
0.8
(3) Heterogeneous weak
(4) Heterogeneous strong
1.0
1.0
1−Specificity
1.0
1−Specificity
0.8
0.6
Sensitivity
0.0
0.2
0.4
0.6
0.4
0.0
0.2
Sensitivity
0.8
●
●
0.0
0.2
0.4
0.6
0.8
1.0
0.0
1−Specificity
0.2
0.4
0.6
0.8
1.0
1−Specificity
Itoh et al., 2014; Liu et al., 2016). We subset the data to focus on the TNBC cases only
and ended up with 275, 178 and 165 TNBC samples in each dataset, respectively. After
routine preprocessing and filtering by including genes sufficiently expressed and with enough
variation, a total of 3377 genes remained in common for the analysis.
We applied our Multi-PC algorithm and compared to the OneStep-SIS procedure as
well as the Min-SIS method by using d = n/ log(n) = 49 (as suggested by their paper).
We used α1 = 0.0001 and α2 = 0.05 (as determined by sensitivity analysis in simulation)
and the “Multi-PC” algorithm only ran up to the first order (i.e. m = 2) and stopped
with six features. This again showed the power of screening with multiple studies. After
feature selection, we fit the linear model in each study to obtain the coefficient estimates and
adjusted R2 . Table 3 showed the coefficient estimates and standard errors of the final set of
six genes selected by our procedure. We added three columns to indicate whether they were
also retained by the Min-SIS (and their relative rank) or OneStep-SIS procedures. As we
can see from the table, all the six genes selected by our procedure were missed by the other
14
Table 3: The six genes selected by our TSA-SIS procedure.
Gene
METABRIC
Est (SE)
Intercept
7.600 (1.502)
EXOC1
0.251 (0.081)∗∗
ITGB1BP1 -0.134 (0.045)∗∗
RBM23
0.168 (0.078)∗
SETD3
-0.166 (0.081)∗
SQSTM1 -0.114 (0.050)∗
TRIOBP -0.126 (0.062)∗
Adjusted-R2
0.151
GSE25066
Est (SE)
0.213 (0.553)
0.278 (0.157).
0.003 (0.111)
0.144 (0.167)
0.366 (0.184)∗
0.029 (0.099)
0.084 (0.118)
0.522
GSE76250 Min-SIS Rank in OneStep-SIS
Est (SE)
d=49 Min-SIS
|S|=25
-1.783 (0.971)
.
0.293 (0.167)
N
164
N
-0.178 (0.194)
N
123
N
0.367 (0.168)∗
N
152
N
-0.080 (0.175)
N
101
N
0.245 (0.183)
N
98
N
0.628 (0.261)∗
N
91
N
0.359
Note: “.” indicates significant level of 0.1, “∗” for level of 0.05, “∗∗” for level of 0.01.
methods. Those genes typically had weak signals in one or more studies thus were very likely
to be incorrectly excluded if only one step screening is performed. Since the METABRIC
study had a larger sample size, all the coefficients appeared to be more significant than the
other two studies.
The gene EXOC1 and p53 are both components of the Ras signaling pathway which
is responsible for cell growth and division and can ultimately lead to cancer (Rajalingam
et al., 2007). RBM23 encodes for an RNA-binding protein implicated in the regulation of
estrogen-mediated transcription and has been found to be associated with p53 indirectly via
a heat shock factor (Asano et al., 2016). ITGB1BP1 encodes for an integrin protein which is
essential for cell adhesion and other downstream signaling pathways that are also modulated
by p53 (Brakebusch et al., 2002).
8
Discussion
In this paper, we proposed a two-step screening procedure for high-dimensional regression
analysis with multiple related studies. In a fairly general framework with weaker assumptions
on the signal strength, we showed that our procedure possessed the sure screening property
for exponentially growing dimensionality without requiring the normality assumption. We
have shown through simulations that our procedure consistently outperformed the rankbased SIS procedure independent of their tuning parameter d. As far as we know, our paper
is the first proposed procedure to perform variable screening in high-dimensional regression
when there are multiple related studies. In addition, we also introduced two applicable
variable selection algorithms following the two-step screening procedure.
Variable selection in regression with multiple studies have been studied in a subfield of
machine learning called multi-task learning (MTL) before and the general procedure is to
apply regularization methods by putting group Lasso penalty, fused Lasso penalty or trace
norm penalty, etc. (Argyriou et al., 2007; Zhou et al., 2012; Ji and Ye, 2009). However, at
ultra-high dimension, such regularization methods usually fail due to challenges in computation expediency, statistical accuracy and algorithmic stability. Instead, sure screening can be
15
used as a fast algorithm for preliminary feature selection, and as long as it exhibits comparable statistical performance both theoretically and empirically, its computational advantages
make it a good choice in application (Genovese et al., 2012). Our method has provided an
alternative to target the high-dimensional multi-task learning problems.
The current two-step screening procedure is based on the linear models but relaxes the
Gaussian assumption to sub-Gaussian distribution. One can apply a modified Fisher’s ztransformation estimator rather than our self-normalized estimator to readily accommodate
general elliptical distribution families (Li et al., 2017). In biomedical applications, noncontinuous outcomes such as categorical, count or survival outcomes are more commonly
seen. Fan et al. (2010) extended SIS and proposed a more general independent learning
approach for generalized linear models by ranking the maximum marginal likelihood estimates. Fan et al. (2011) further extended the correlation learning to marginal nonparametric
learning for screening in ultra-high dimensional additive models. Other researchers exploited
more robust measure for the correlation screening (Zhu et al., 2011; Li et al., 2012; Balasubramanian et al., 2013). All these measures can be our potential extension by modifying the
marginal utility used in the screening procedure. Besides, the idea of performing screening
with multiple studies is quite general and is applicable to relevant statistical models other
than the regression model, for example, Gaussian graphical model with multiple studies. We
leave these interesting problems in future study.
9
Proofs
We start by introducing three technical lemmas that are essential for the proofs of the main
(k)
results. By the scaling property of T̂j and Remark 1, without loss of generality, we can
(k)
(k)
assume E(Xj ) = E(Y (k) ) = 0 and var(Xj ) = var(Y (k) ) = 1 for all k ∈ [K], j ∈ [p].
(k)
(k)
Therefore in the proof we do not distinguish between σj and ρj . The first lemma is on
(k)
the concentration inequalities of the self-normalized covariance and θ̂j .
Lemma 1. Under the assumptions (C1) and (C2), for any δ ≥ 2 and M > 0, we have:
q
(k)
(k)
σ̂ −σj
(i) P (max | j (k) 1/2
| ≥ δ logn p ) = O((log p)−1/2 p−δ+1+b ),
j,k
(θ̂j )
(k)
(k)
(ii) P (max |θ̂j − θj | ≥ Cθ
q
j,k
log p
)
n
= O(p−M ),
where Cθ is a positive constant depending on M1 , η and M only.
The second and third lemmas, which will be used in the proof of Theorem 2, describe
n
P
r (k)
(k)
(k)
(k)
(k)
√ (k)
√1
[(Xij −X̄j )(Yi −Ȳ (k) )−ρj ]
n
θ̂j
nρ
(k)
(k)
i=1
q
q j
the concentration behaviors of Ĥj :=
=
T̂
−
(k)
j
(k)
(k)
θj
n
P
(k)
and Ȟj
:=
(k) (k)
(k)
√1
(Xij Yi −ρj )
n
i=1
q
(k)
θj
.
16
θj
θj
Lemma 2. There exists some constant c > 0 such that,
P (|
X
(k)2
[Ȟj
− 1]| > t) ≤ 2 exp(−c min[
k∈lj
t2 1/2
, t ]),
κj
where c depends on M1 and η only.
Lemma 3. There exists some constant CH > 0 such that,
s
log2 p
(k)
(k)
P (max |Ȟj − Ĥj | > CH
) = O(p−M ),
j,k
n
s
log3 p
(k)2
(k)2
P (max |Ȟj − Ĥj | > CH
) = O(p−M ),
j,k
n
where CH depends on M1 , η, M and τ0 only.
The proofs of the three lemmas are provided in the Appendix.
Proof of Theorem 1. We first define the following error events:
[0]
I,A
Ej,k
[0]
II,A
Ej,k
[1]
I,A
Ej,k
[1]
II,A
Ej,k
(k)
= {|T̂j | > Φ−1 (1 − α1 /2) and j ∈ A[0] , k ∈ lj },
(k)
/ lj },
= {|T̂j | ≤ Φ−1 (1 − α1 /2) and j ∈ A[0] , k ∈
(k)
= {|T̂j | > Φ−1 (1 − α1 /2) and j ∈ A[1] , k ∈ lj },
(k)
/ lj }.
= {|T̂j | ≤ Φ−1 (1 − α1 /2) and j ∈ A[1] , k ∈
To show Theorem 1 that P (A) = 1 − O(p−L ), it suffices to show that,
[ I,A[0]
II,A[0]
P { (Ej,k
∪ Ej,k
)} = O(p−L ),
(9.1)
j,k
and
P{
[ I,A[1]
II,A[1]
(Ej,k ∪ Ej,k
)} = O(p−L ).
(9.2)
j,k
One√can apply Lemma 1 to bound each component in (9.1) and (9.2) with α1 = 2{1 −
Φ(γ log p)} and γ = 2(L + 1 + b). Specifically, we obtain that,
[ I,A[0]
P ( Ej,k
) = P ( max
j,k
j∈A[0] ,k∈lj
(k)
|T̂j | ≥ γ
p
1
log p) = O( √
p−γ+1+b ) = o(p−L ),
log p
17
(9.3)
(k)
where the second equality is due to Lemma 1 (i) with δ = γ, noting that σj
√ (k) q (k)
(k)
T̂j = nσ̂j / θ̂j . In addition, we have that,
[ I,A[1]
P ( Ej,k
) =P { max
j∈A[1] ,k∈lj
j,k
(k)
|T̂j | ≥ γ
(k)
≤P ( max
|
p
j∈A[1] ,k∈lj
log p}
(k)
σ̂j − ρj
r
| ≥ (γ − C1 )
(k)
(θ̂j )1/2
= 0 and
log p
) + O(p−L )
n
(9.4)
1
p−(γ−C1 )+1+b ) + O(p−L )
log p
=O(p−L ),
=O( √
where the inequality on the second line is due to assumption (C4) on lj for j ∈ A[1] , Lemma
(k)
(k)
(k)
1 (ii) with M = L, and assumption (C1) minj,k θj ≥ τ0 , i.e., θ̂j ≥ θj − Cθ (log p/n)1/2 ≥
(k)
0.99θj . The equality on the third line follows from Lemma 1 (i) where δ = γ −C1 = L+1+b.
In the end, we obtain that,
[ II,A[0]
p
(k)
II,A[1]
P { (Ej,k
∪ Ej,k
)} =P (max |T̂j | < γ log p)
j,k∈l
/j
j,k
(k)
≤P (max |
j,k∈l
/j
(k)
σ̂j − ρj
(k)
(θ̂j )1/2
r
| ≥ (C3 − γ)
log p
) + O(p−L )
n
(9.5)
1
p−(C3 −γ)+1+b ) + O(p−L )
log p
−L
=O(p ),
=O( √
where the inequality on the second line is due to assumptions (C3) and (C4) on lj , Lemma
(k)
(k)
1 (ii) with M = L and assumption (C1) on sub-Gaussian distributions, i.e., θ̂j ≤ θj +
(k)
(k)
Cθ (log p/n)1/2 ≤ 1.01θj . In particular, we have implicitly used the fact that maxj,l θj is
upper bounded by a constant depending on M1 and η only. The equality on the third line
follows from Lemma 1 (i) where δ = C3 − γ = L + 1 + b.
Finally, we complete the proof by combining (9.3)-(9.5) to show (9.1)-(9.2).
Proof of Theorem 2. We first define the following error events:
[0] ,2
EjA
[1] ,2
EjA
= {|L̂j | > ϕ−1 (1 − α2 ) or κ̂j = 0} for j ∈ A[0] ,
= {|L̂j | < ϕ−1 (1 − α2 ) and κ̂j 6= 0} for j ∈ A[1] .
To prove Theorem 2, we only need to show that,
[ A[0] ,2
[ A[1] ,2
Ej
Ej
P(
) = O(p−L ) and P (
) = O(p−L ),
j∈A[0]
with α2,κj := 1 − ϕκj [κj + C4 (log2 p +
j∈A[1]
p
κj log p)] := 1 − ϕκj (γκj ).
18
(9.6)
Recall the event A defined in Theorem 1. Thus we have that,
[
[0]
[1]
P {(∪j∈A[0] EjA ,2 ) (∪j∈A[1] EjA ,2 )}
X (k)2
X
≤P (AC ) + p max P (
T̂j > γκj ) + p max P (
T̂ (k)2 < γκj ).
j∈A[0]
j∈A[1] ,κj 6=0
k∈lj
k∈lj
Therefore, given the results in Theorem 1, it suffices to show,
X
P(
T̂ (k)2 > γκj ) = O(p−L−1 ) for any j ∈ A[0] ,
(9.7)
k∈lj
and
P(
X
T̂ (k)2 < γκj ) = O(p−L−1 ) for any j ∈ A[1] and κj > 0.
(9.8)
k∈lj
[0]
We first prove equation (9.7). Since j ∈ A , we have
P
(k)2
to bound the probability of k∈lj T̂j > γκj below.
P(
X
(k)2
T̂j
> γκj )
(k)2
Ĥj
Cθ
> (1 −
τ0
(k)
Ĥj
=
(k)
T̂j
r
(k)
θ̂j
(k)
θj
. We are ready
k∈lj
≤P (
X
k∈lj
r
log p
)γκj ) + O(p−L−1 )
n
r
s
log3 p
) + O(p−L−1 )
n
j
s
X (k)2
p
κ2j log p
C
θ
2
=P ( (Ȟj − 1) > κj + C4 (log p + κj log p) −
τ0
n
k∈lj
s
s
s
Cθ C4
log5 p
κj log2 p
log3 p
−
(
+
) − κj − κj CH
) + O(p−L−1 )
τ0
n
n
n
X (k)2
p
≤P ( (Ȟj − 1) > C20 (log2 p + κj log p)) + O(p−L−1 )
X (k)2
Cθ
≤P ( (Ȟj − 1) > (1 −
τ0
k∈l
log p
)γκj − κj − κj CH
n
k∈lj
=O(p−L−1 ).
(k)
The inequality on the second line is due to assumption (C1) that min θj
j,k
≥ τ0 > 0 and
Lemma 1 (ii) with M = L + 1. The inequality on the third line follows from Lemma 3 with
M = L + 1. The inequality on the fifth line is by the choice of γκj with a sufficiently large
C4 > 0 and the assumption (C2) that log3 p = o(n) and κj log2 p = o(n). The last equality
follows from Lemma 2.
19
Lastly, we prove (9.8) as follows,
X (k)2
P(
T̂j < γκj )
k∈lj
√ (k)
(k)
X (k)
nρj 2 θj
) (k) < γκj )
=P ( (Ĥj + q
(k)
θ̂j
k∈lj
θj
r
√ (k)
X (k)
nρj 2
Cθ log p
)γκj ) + O(p−L−1 )
≤P ( (Ĥj + q
) ≤ (1 +
(k)
τ
n
0
k∈lj
θj
s
r
X (k)2
X (k)2
log3 p
Cθ log p
≤P ( (Ȟj − 1) ≤ κj CH
− κj + (1 +
)γκj − Cm n
ρj
n
τ
n
0
k∈lj
k∈lj
s
√ (k)
√
(k)
X (k) nρj
log2 p X n|ρj |
q
Ȟj q
+ 2CH
) + O(p−L−1 ).
−2
(k)
(k)
n
k∈lj
k∈lj
θj
θj
(k)
The inequality on the third line is due to assumption (C1) that min θj
j,k
(9.9)
≥ τ0 > 0 and
Lemma 1 (ii) with M = L + 1. The inequality on the fourth line follows from Lemma 3
(k)
(both equations) and min(θj )−1 := Cm > 0, guaranteed by the sub-Gaussian assumption
j,k
in assumption (C1).
We can upper bound the term 2CH
s
2CH
q
log2 p
n
√ (k)
n|ρ |
q j
k∈lj
(k)
θj
P
in (9.9) as follow,
s
√ (k)
√
s X
sX
X
n|ρj |
log p
log2 p n √
(k)2
(k)2
q
ρj = o( n
≤ 2CH
κj
ρj ).
√
(k)
n k∈l
n
τ0
k∈lj
k∈lj
θj
j
2
(9.10)
The first inequality is by the Cauchy-Schwarz inequality and assumption (C1), and the
second equality by the assumption (C2) that κj log2 p = o(n).
√ (k)
P
(k)
(k) nρj
We next upper bound the term −2 k∈lj Ȟj q (k)
with high probability. Note that θj
θj
(k)
θj
(k)
−1
Cm
is bounded below and above, i.e., τ0 ≤
≤
by assumption (C1). In addition, Ȟj has
zero mean and is sub-exponential with bounded constants by assumption (C1). By Bernstein
inequality (Proposition 5.16 in Vershynin (2010)), we have with some constant c0 > 0,
P (|2
X
k∈lj
We pick t = CB
(k)
|Ȟj
√ (k)
n|ρ |
t2
t
q j | > t) ≤ 2 exp(−c0 min[ P (k)2 ],
).
√
(k)
(k)
n
ρj
max n|ρj |
θj
k∈lj
k∈lj
q P
(k)2
n k∈lj ρj log2 p with a large constant CB in the inequality above and
20
apply (9.10) to reduce (9.9) as follows,
X (k)2
P(
T̂j < γκj )
k∈lj
s X
X (k)2
X (k)2
(k)2
≤P ( (Ȟj − 1) ≤ −Cm n
ρj log2 p
ρj + 2CB n
k∈lj
k∈lj
k∈lj
p
+ 2C4 κj log p + 2C4 log2 p) + O(p−L−1 )
q
X (k)2
p
p
2
≤P ( (Ȟj − 1) ≤ −Cm C2 (log p + κj log p) + 2CB C2 log2 p(log2 p + κj log p)
k∈lj
p
+ 2C4 κj log p + 2C4 log2 p) + O(p−L−1 )
X (k)2
p
≤P ( (Ȟj − 1) ≤ −C20 (log2 p + κj log p)) + O(p−L−1 )
k∈lj
=O(p−L−1 ).
The inequality on the first line is obtained by the choice of γκj with the chosen C4 > 0
and the assumption (C2) that κj log2 p = o(n). The inequalities
√ on the second line and third
P
C2 (log2 p+ κj log p)
(k) 2
for a sufficiently large
line are by the assumption (C4) that k∈lj |ρj | ≥
n
C2 > 0. The last equality is by Lemma 2.
This completes the proof of (9.7) and (9.8), which further yields to
[
[0]
[1]
P {(∪j∈A[0] EjA ,2 ) (∪j∈A[1] EjA ,2 )} = O(p−L ),
with the results from Theorem 1. Therefore we complete the proof of Theorem 2.
Appendix
S1. Proof of Lemma 1
Proof. Part (i) immediately follows from Lemma 2 (i) equation (25) in Cai and Liu (2011).
To prove part (ii), we need to bound the three terms on the right side of the following
inequality,
(k)
(k)
(k)
(k)
(k)
(k)
(k)
(k)
max |θ̂j − θj | ≤ max |θ̂j − θ̃j | + max |θ̃j − θ̌j | + max |θ̌j − θj |,
j,k
(k)
where θ̃j
(k)
:=
j,k
1
n
n
P
(k)
(k)
(Xij Yi
j,k
(k)
(k)
− ρ̃j )2 with ρ̃j =
i=1
(k)
1
n
(A1)
j,k
n
P
(k)
(k)
(k)
Xij Yi , and θ̌j
i=1
(k)
:=
1
n
n
P
(k)
(k)
(Xij Yi
−
i=1
ρj )2 . Note that E(θ̌j ) = θj .
By the marginal sub-Gaussian distribution assumption in assumption (C1), we have that
(k) (k)
(k)
(k)
(Xij Yi − ρj )2 has mean θj and finite Orlicz ψ1/2 -norm (see, e.g., Adamczak et al.
(2011)). Thus we can apply equation (3.6) of Adamczak et al. (2011), i.e.,
√ (k)
t2
(k)
P (max n|θ̌j − θj | > t) ≤ 2 exp(−c min[ , t1/2 ]),
j,k
n
21
√
with t = (Cθ /3) n log p for a large enough constant Cθ > 0 depending on M1 , η and M only
to obtain that,
r
log p
(k)
(k)
P (max |θ̌j − θj | > (Cθ /3)
) = O(p−M ).
(A2)
j,k
n
2
We have used the assumption log p = o(n1/3 ) in assumption (C2) to make sure tn ≤ t1/2 .
By applying equation (1) in supplement of Cai and Liu (2011), we obtain that,
r
log p
(k)
(k)
P (max |θ̃j − θ̂j | > (Cθ /3)
) = O(p−M ).
(A3)
j,k
n
In addition, by a similar truncation argument as that in the proof of Lemma 2 in Cai and
Liu (2011) and equation (7) therein, we obtain that by picking a large enough Cθ > 0,
r
log p
(k)
(k)
P (max |θ̃j − θ̌j | > (Cθ /3)
) = O(p−M ).
(A4)
j,k
n
We complete the proof by combining (A1)-(A4) with a union bound argument.
S2. Proof of Lemma 2
(k)
(k)
Proof. It is easy to check that E(Ȟj ) = 0 and var(Ȟj ) = 1. The marginal sub-Gaussian
(k)
distribution assumption in assumption (C1) implies that Ȟj has finite Orlicz ψ1 -norm (i.e.,
(k)
sub-exponential distribution with finite constants). Therefore, (Ȟj )2 −1 is centered random
(k)
variable with finite Orlicz ψ1/2 -norm. Note that Ȟj are independent for k ∈ [K]. The result
follows from equation (3.6) of Adamczak et al. (2011).
S3. Proof of Lemma 3
√
(k) √
√ (k)
nX̄j
nȲ (k)
(k)
(k)
q
Proof. Note that Ȟj −Ĥj =
. By assumption (C1), we have that E( nX̄j ) =
(k)
nθj
√ (k)
√ (k)
√
√ (k)
√
E( nȲ ) = 0, var( nX̄j ) = var( nȲ (k) ) = 1, and both nX̄j and nȲ (k) are subGaussian with bounded constants. Therefore, the first equation follows from Bernstein
inequality (e.g., Definition 5.13 in Vershynin (2010)) applied to centered sub-exponential
√ (k) √
(k)
variable nX̄j · nȲ (k) , noting θj ≥ τ0 by assumption (C1). The second equation follows
from the first one, log3 p = o(n), and a Bernstein inequality (e.g., Corollary 5.17 in Vershynin
(k)
(2010)) applied to the sum of centered sub-exponential variables Ȟj .
References
Adamczak, R., Litvak, A. E., Pajor, A., and Tomczak-Jaegermann, N. (2011). Restricted
isometry property of matrices with independent columns and neighborly polytopes by
random sampling. Constructive Approximation, 34(1):61–88.
22
Argyriou, A., Evgeniou, T., and Pontil, M. (2007). Multi-task feature learning. In Advances
in neural information processing systems, pages 41–48.
Asano, Y., Kawase, T., Okabe, A., Tsutsumi, S., Ichikawa, H., Tatebe, S., Kitabayashi,
I., Tashiro, F., Namiki, H., Kondo, T., et al. (2016). Ier5 generates a novel hypophosphorylated active form of hsf1 and contributes to tumorigenesis. Scientific reports,
6.
Balasubramanian, K., Sriperumbudur, B., and Lebanon, G. (2013). Ultrahigh dimensional
feature screening via rkhs embeddings. In Artificial Intelligence and Statistics, pages 126–
134.
Brakebusch, C., Bouvard, D., Stanchi, F., Sakai, T., and Fassler, R. (2002). Integrins in
invasive growth. The Journal of clinical investigation, 109(8):999.
Bühlmann, P., Kalisch, M., and Maathuis, M. H. (2010). Variable selection in highdimensional linear models: partially faithful distributions and the pc-simple algorithm.
Biometrika, 97(2):261–278.
Cai, T. and Liu, W. (2011). Adaptive thresholding for sparse covariance matrix estimation.
Journal of the American Statistical Association, 106(494):672–684.
Cai, T. T. and Liu, W. (2016). Large-scale multiple testing of correlations. Journal of the
American Statistical Association, 111(513):229–240.
Chang, J., Tang, C. Y., and Wu, Y. (2013). Marginal empirical likelihood and sure independence feature screening. Annals of statistics, 41(4).
Chang, J., Tang, C. Y., and Wu, Y. (2016). Local independence feature screening for
nonparametric and semiparametric models by marginal empirical likelihood. Annals of
statistics, 44(2):515.
Curtis, C., Shah, S. P., Chin, S.-F., Turashvili, G., Rueda, O. M., Dunning, M. J., Speed, D.,
Lynch, A. G., Samarajiwa, S., Yuan, Y., et al. (2012). The genomic and transcriptomic
architecture of 2,000 breast tumours reveals novel subgroups. Nature, 486(7403):346–352.
Fan, J., Feng, Y., and Song, R. (2011). Nonparametric independence screening in sparse
ultra-high-dimensional additive models. Journal of the American Statistical Association,
106(494):544–557.
Fan, J. and Li, R. (2001). Variable selection via nonconcave penalized likelihood and its
oracle properties. Journal of the American statistical Association, 96(456):1348–1360.
Fan, J. and Lv, J. (2008). Sure independence screening for ultrahigh dimensional feature space. Journal of the Royal Statistical Society: Series B (Statistical Methodology),
70(5):849–911.
Fan, J. and Lv, J. (2010). A selective overview of variable selection in high dimensional
feature space. Statistica Sinica, 20(1):101.
23
Fan, J., Samworth, R., and Wu, Y. (2009). Ultrahigh dimensional feature selection: beyond
the linear model. Journal of Machine Learning Research, 10(Sep):2013–2038.
Fan, J., Song, R., et al. (2010). Sure independence screening in generalized linear models
with np-dimensionality. The Annals of Statistics, 38(6):3567–3604.
Genovese, C. R., Jin, J., Wasserman, L., and Yao, Z. (2012). A comparison of the lasso and
marginal regression. Journal of Machine Learning Research, 13(Jun):2107–2143.
Huang, J., Breheny, P., and Ma, S. (2012). A selective review of group selection in highdimensional models. Statistical science: a review journal of the Institute of Mathematical
Statistics, 27(4).
Itoh, M., Iwamoto, T., Matsuoka, J., Nogami, T., Motoki, T., Shien, T., Taira, N., Niikura, N., Hayashi, N., Ohtani, S., et al. (2014). Estrogen receptor (er) mrna expression
and molecular subtype distribution in er-negative/progesterone receptor-positive breast
cancers. Breast cancer research and treatment, 143(2):403–409.
Ji, S. and Ye, J. (2009). An accelerated gradient method for trace norm minimization.
In Proceedings of the 26th annual international conference on machine learning, pages
457–464. ACM.
Jiang, J., Li, C., Paul, D., Yang, C., Zhao, H., et al. (2016). On high-dimensional misspecified mixed model analysis in genome-wide association study. The Annals of Statistics,
44(5):2127–2160.
Li, R., Liu, J., and Lou, L. (2017). Variable selection via partial correlation. Statistica
Sinica, 27(3):983.
Li, R., Zhong, W., and Zhu, L. (2012). Feature screening via distance correlation learning.
Journal of the American Statistical Association, 107(499):1129–1139.
Liang, F., Song, Q., and Qiu, P. (2015). An equivalent measure of partial correlation coefficients for high-dimensional gaussian graphical models. Journal of the American Statistical
Association, 110(511):1248–1265.
Liu, Y.-R., Jiang, Y.-Z., Xu, X.-E., Hu, X., Yu, K.-D., and Shao, Z.-M. (2016). Comprehensive transcriptome profiling reveals multigene signatures in triple-negative breast cancer.
Clinical cancer research, 22(7):1653–1662.
Luo, S., Song, R., and Witten, D. (2014). Sure screening for gaussian graphical models.
arXiv preprint arXiv:1407.7819.
Ma, S., Li, R., and Tsai, C.-L. (2017). Variable screening via quantile partial correlation.
Journal of the American Statistical Association, pages 1–14.
Meier, L., Van De Geer, S., and Bühlmann, P. (2008). The group lasso for logistic regression.
Journal of the Royal Statistical Society: Series B (Statistical Methodology), 70(1):53–71.
Nardi, Y., Rinaldo, A., et al. (2008). On the asymptotic properties of the group lasso
estimator for linear models. Electronic Journal of Statistics, 2:605–633.
24
Rajalingam, K., Schreck, R., Rapp, U. R., and Albert, S. (2007). Ras oncogenes and
their downstream targets. Biochimica et Biophysica Acta (BBA)-Molecular Cell Research,
1773(8):1177–1195.
Shao, Q.-M. (1999). A cramér type large deviation result for student’s t-statistic. Journal
of Theoretical Probability, 12(2):385–398.
Tibshirani, R. (1996). Regression shrinkage and selection via the lasso. Journal of the Royal
Statistical Society. Series B (Methodological), pages 267–288.
Tseng, G. C., Ghosh, D., and Feingold, E. (2012). Comprehensive literature review and
statistical considerations for microarray meta-analysis. Nucleic acids research, 40(9):3785–
3799.
Vershynin, R. (2010). Introduction to the non-asymptotic analysis of random matrices. arXiv
preprint arXiv:1011.3027.
Yadav, B. S., Chanana, P., and Jhamb, S. (2015). Biomarkers in triple negative breast
cancer: A review. World journal of clinical oncology, 6(6):252.
Yuan, M. and Lin, Y. (2006). Model selection and estimation in regression with grouped
variables. Journal of the Royal Statistical Society: Series B (Statistical Methodology),
68(1):49–67.
Zhou, J., Liu, J., Narayan, V. A., and Ye, J. (2012). Modeling disease progression via fused
sparse group lasso. In Proceedings of the 18th ACM SIGKDD international conference on
Knowledge discovery and data mining, pages 1095–1103. ACM.
Zhu, L.-P., Li, L., Li, R., and Zhu, L.-X. (2011). Model-free feature screening for ultrahighdimensional data. Journal of the American Statistical Association, 106(496):1464–1475.
Zou, H. (2006). The adaptive lasso and its oracle properties. Journal of the American
statistical association, 101(476):1418–1429.
Zou, H. and Hastie, T. (2005). Regularization and variable selection via the elastic net.
Journal of the Royal Statistical Society: Series B (Statistical Methodology), 67(2):301–
320.
25
| 10math.ST
|
Solving the Resource Constrained Project Scheduling Problem Using the
Parallel Tabu Search Designed for the CUDA Platform1
Libor Bukata∗, Přemysl Šůcha, Zdeněk Hanzálek
arXiv:1711.04556v1 [cs.DC] 13 Nov 2017
Department of Control Engineering, the Czech Technical University in Prague,
Karlovo náměstı́ 13, 121 35 Prague 2, the Czech Republic
Abstract
In the paper, a parallel Tabu Search algorithm for the Resource Constrained Project Scheduling Problem is proposed. To deal with this NP-hard combinatorial problem many optimizations have been
performed. For example, a resource evaluation algorithm is selected by a heuristic and an effective
Tabu List was designed. In addition to that, a capacity-indexed resource evaluation algorithm was
proposed and the GPU (Graphics Processing Unit) version uses a homogeneous model to reduce the
required communication bandwidth. According to the experiments, the GPU version outperforms the
optimized parallel CPU version with respect to the computational time and the quality of solutions.
In comparison with other existing heuristics, the proposed solution often gives better quality solutions.
Cite as: Libor Bukata, Premysl Sucha, Zdenek Hanzalek, Solving the Resource Constrained Project
Scheduling Problem using the parallel Tabu Search designed for the CUDA platform, Journal of Parallel
and Distributed Computing, Volume 77, March 2015, Pages 58-68, ISSN 0743-7315, http://dx.doi.
org/10.1016/j.jpdc.2014.11.005.
Source code: https://github.com/CTU-IIG/RCPSPCpu, https://github.com/CTU-IIG/RCPSPGpu
Keywords: Resource Constrained Project Scheduling Problem, Parallel Tabu Search, CUDA,
homogeneous model, GPU
1. Introduction
The Resource Constrained Project Scheduling
Problem (RCPSP), which has a wide range of applications in logistics, manufacturing and project
management [1], is a universal and well-known
problem in the operations research domain. The
problem can be briefly described using a set of
∗
The corresponding author.
Email addresses: [email protected] (Libor
Bukata), [email protected] (Přemysl Šůcha),
[email protected] (Zdeněk Hanzálek)
1 c
2017. This manuscript version is made available under the CC-BY-NC-ND 4.0 license http://
creativecommons.org/licenses/by-nc-nd/4.0/.
Preprint submitted to J. Parallel Distrib. Comput.
activities and a set of precedence constraints describing the relationships among activities. Each
activity requires a defined amount of the resources
and every resource has a limited capacity. The
objective is to find the best feasible schedule according to a criterion. The RCPSP was proved
to be NP-hard in the strong sense when the criterion is makespan [2]. For that reason only small
instances (approximately up to 30 activities) can
be reliably solved by exact methods like Branch &
Bound [3], therefore a heuristic or a meta-heuristic
is required to solve the problem satisfactorily.
In recent times, there is an increased interest
in using graphics cards to solve difficult combinatorial problems (e.g. [4, 5, 6]), since a modern
March 12, 2018
graphics card is usually much more powerful than
a current multi-core CPU. Although the graphics
cards have some restrictions (e.g. a high-latency
global memory access), the new GPU architectures like Kepler and Fermi can significantly reduce these bottlenecks. As a consequence modern
GPUs are applicable to the problems which were
solvable only on CPUs previously.
Not only the high computational power makes
graphics cards attractive to researchers and practitioners, but also the mature Nvidia CUDA framework which enables us to create GPU programs
in an effective and relatively easy way since it extends standard languages like C/C++ by adding
GPU specific functions and language keywords.
Nevertheless, the CUDA is only designed for the
Nvidia graphics cards.
From the implementation point of view, there
are two models. The first one is called a homogeneous model where all required data structures
are stored in a GPU at the beginning of an algorithm and the results are read at the end of the
algorithm. There is no communication between
the CPU and the GPU during the computations.
The second approach is a heterogeneous model.
The main logic of an algorithm runs on the CPU
and the GPU is used only for the most computationally intensive tasks. The disadvantage of the
heterogeneous model is the frequent communication during computations, therefore, the communication bandwidth can state a bottleneck. However, the heterogeneous model is usually simpler
to implement.
cores). Every location (i.e. an index of a thread)
has different parameters (a tabu tenure, stopping
criteria). At the beginning of the search each
thread initializes its own location by a short TS
operator, i.e. a modified version of the Taillard’s
robust tabu search. Then the asynchronous parallel tabu search is started. Every thread independently reads a solution and parameters from the
location, possibly makes a diversification, runs the
TS operator on the solution, and writes back and
sets an UPDATE flag if an improving solution is
found. After that the thread location index is circularly incremented. Diversification takes place if
the read solution does not have the UPDATE flag
set. Every best global solution is copied to half of
the locations of the circular buffer to propagate
elite solutions. Since the circular buffer is shared
by many threads, the access has to be as short as
possible and the locations have to be protected by
critical sections.
Relatively many authors try to use a GPU
for solving combinatorial problems. For example,
Czapiński and Barnes [10] implemented a GPU
version of TS to solve the Flowshop Scheduling
Problem (FSP). The success of the implementation illustrates an achieved speedup against the
CPU version. The GPU version was up to 89.01
times faster than the CPU (Intel Xeon 3.0 GHz,
2 GB memory, Nvidia Tesla C1060 GPU). Nevertheless, the quality of solutions was not investigated.
The Flowshop Scheduling Problem was also
solved by Zajı́ček and Šůcha [11]. The authors
implemented a GPU version of an island based
genetic algorithm. Islands are used for migration
of individuals among sub-populations, where each
sub-population is a subset of solutions and an individual corresponds to a specific solution. Subpopulation can be evaluated, mutated and crossed
over independently of other sub-populations, therefore, huge parallelization can be achieved. It should
be noted that a homogeneous model was used.
The maximal speedup against the CPU was 110
for 100 activities and 5 machines (AMD Phenom
II X4 945 3.0 GHz, Nvidia Tesla C1060).
Czapiński [5] proposed a Parallel Multi-start
Tabu Search for the Quadratic Assignment Prob-
1.1. Related works
The Tabu Search meta-heuristic was proposed
by Glover in 1986 [7]. Hundreds of publications
have been written since that time. The basic concept of the TS meta-heuristic is clarified in Gendreau [8]. The author has described the basic
terms of the TS, as a Tabu List (TL), aspiration
criteria, diversification, intensification, etc.
From the Tabu Search parallelization point of
view James et al. [9] proposed a sophisticated solution. The authors use a circular buffer where
the size of the buffer is equivalent to the number
of the started threads (often the number of CPUs
2
lem. The main idea is to start several Parallel
Tabu Search instances with different parameters
and initial solutions. All Tabu Search instances
should terminate approximately at the same time
since the synchronization is required to get the
most promising solutions. When a stop criterion
is met, the modified solutions are read back and
the most promising solutions are used as the initial solutions in the next run. The Tabu Search
instance runs entirely on the GPU, therefore communication overheads are reduced to minimum.
The achieved results reveal the effectiveness of the
implementation since Nvidia GTX 480 is up to 70
times faster than a six-core Intel Core i7-980x 3.33
GHz.
Hofmann et al. [12] investigated the suitability
of graphics cards for genetic algorithms. The authors selected two problems to solve, namely the
Weierstrass function minimization and the Traveling Salesman Problem. The first problem can
be very effectively implemented on the GPU since
the Weierstrass function is comprised of floatingpoint operations and trigonometric functions that
are directly supported by the GPU hardware. In
contrast to the first problem the second problem
was not tailor-made for the GPU, therefore, the
multi-core CPU was able to compete with the
GPU. The authors suggest that all parts of a genetic algorithm should be performed on the Fermi
or newer GPUs (i.e. homogeneous model).
Boyer et al. [13] used dynamic programming
to solve the knapsack problem on a GPU. An effective data compression was proposed to reduce
memory occupancy. The achieved results show
that the Nvidia GTX 260 graphics card was up
to 26 faster than Intel Xeon 3.0 GHz.
The above mentioned combinatorial problems
have something in common. The solution evaluation is quite simple since it is usually a “simple sum”. On the other hand, the RCPSP requires much more complicated schedule evaluation methods and data structures.
CPU version in both performance speedup and
the quality of solutions. This is possible thanks
to an effective schedule evaluation and a GPUoptimized Simple Tabu List. In addition, the required data transfers are reduced to minimum due
to the homogeneous model. Our Parallel Tabu
Search is able to outperform other Tabu Search
implementations in the quality of the resulting solutions.
The paper is structured as follows: The following section introduces the RCPSP mathematical formulation and notation. The Tabu Search
meta-heuristic is briefly described in Section 3.
Our proposed Parallel Tabu Search algorithm for
the CUDA platform is described in detail in Sections 4, 5, and 6. The performed experiments are
located in Section 7 and the last section concludes
the work.
2. Problem Statement
According to the standard notation, classification for the RCPSP is PS|prec|Cmax [14]. A
project can be described as follows: There is a set
of activities V = {0, . . . , N − 1} with durations
D = {d0 , . . . , dN −1 } where N is the number of
activities. There are two dummy activities 0 and
N −1 such that d0 = dN −1 = 0. Activity 0 is a predecessor of all other activities and activity N − 1
is the end activity of a project. A schedule of the
RCPSP can be represented as a vector of activities’ start times S = {s0 , . . . , sN −1 } where si ∈ N.
Alternatively, a schedule can be expressed as an
order of activities W = {w0 , . . . , wN −1 } ∈ W
where wu is the u-th activity of the schedule and
W is a set of all feasible solutions.
The RCPSP can be represented as Direct Acyclic
Graph G(V, E) where nodes V are activities and
edges E are precedence relations. If there is edge
(i, j) ∈ E then si + di ≤ sj since activity j has to
be scheduled after activity i.
Each activity requires some amount of renewable resources. The number of project resources is
denoted as M and a set of resources capacities is
R = {R0 , . . . , RM −1 } where Rk ∈ N. Maximal re−1
source capacity Rmax is equal to maxM
k=0 Rk . Activity resource requirement ri,k ∈ N means that
1.2. Contribution and Paper Outline
The proposed solution is the first known GPU
algorithm for the RCPSP. The performed experiments revealed that the GPU outperforms the
3
Activity i
0
1
2
3
4
5
6
7
8
9
10
11
activity i requires ri,k ≤ Rk resource units of resource k during its execution. As si and di values
are positive integers, the resulting schedule length
Cmax (i.e. the project makespan) will also be an
integer as well.
A lower bound of the project makespan can be
found by neglecting resources. For each activity
i ∈ V all outgoing edges (i, j) ∈ E are weighted
by its duration di . The longest path from 0 to
N −1 in graph G(V, E) corresponds to the critical
path. Its length is equal to the optimal project
makespan on the condition that all resources have
an unlimited capacity.
∀i∈V
sj ≥ si + di
∀(i, j) ∈ E
!
X
Cmax
max
ri,k ≤ Rk
t=0
ri,0
0
5
2
3
2
3
4
2
4
1
2
0
ri,1
0
3
1
2
3
4
1
2
5
2
2
0
Successors
{1, 2}
{3, 6}
{4, 5}
{5, 10}
{7}
{8, 9}
{7, 9}
{8, 10}
{11}
{11}
{11}
{}
Table 1: Data of an example instance.
2.1. Mathematical Formulation
minimize Cmax
s.t. Cmax = max (si + di ) = sN −1
di
0
4
3
5
5
3
2
4
2
3
4
0
3. Brief description of the Tabu Search metaheuristic
(1)
(2)
To move through solution space W, a transformation of the current solution to a neighborhood
solution is required. This transformation is called
a move which can be seen as a light solution modification like a swap of two elements in an order,
etc.
The Tabu Search meta-heuristic was proposed
by Glover [7] as an improvement of the local search
technique [8]. A local search algorithm starts
from the initial solution and iteratively improves
this solution by applying the best neighborhood
moves until a local optimum is reached, whereas
the Tabu Search introduces a short-term memory called Tabu List which reduces the probability of getting stuck in a local optimum or plateau
by forbidding the previously visited solutions. As
a consequence, not only improving solutions are
permitted and the search process is able to climb
(3)
(4)
i∈Ft
∀k ∈ {0, . . . , M − 1}
Ft = {i ∈ V |si ≤ t < si + di }
The objective of the RCPSP is to find a feasible schedule W with the minimal schedule length
Cmax . The schedule length is the latest finish
time of any activity (Equations (1),(2)). Equation (3) ensures that all precedence relations are
satisfied. A schedule is feasible if all precedence
relations are satisfied and the resources are not
overloaded, i.e. the activities requirements do not
exceed the capacity of any resource at any time
(Equation (4)).
2.2. Instance Example
The data of an example instance are showed in
Table 1. In the project there are 10 non-dummy
activities and 2 renewable resources with maximal
capacity 6. The corresponding graph of precedences is shown in Figure 1. The critical path
is highlighted by bold lines and its length is 16.
One of the feasible solutions of the instance is the
activity order W = {0, 1, 2, 3, 4, 6, 5, 7, 9, 10, 8, 11}
with Cmax = 22. The resource utilization for this
order is depicted in Figure 2.
3
i
10
5
di
4
1
6
0
4
2
0
2
4
3
5
7
4
8
11
2
0
5
3
9
3
Figure 1: Graph of precedences for the example instance.
4
R1
resource utilization
6
5
5
4
4
9
3
2
8
2
1
3
1
0
0
1
2
3
4
5
6
7
6
7
8
9
10
11
12
13
14
10
15
16
17
18
19
20
21
t [s]
R0
6
resource utilization
22
5
6
4
3
5
3
1
9
8
2
2
1
0
0
1
2
3
4
5
4
6
7
8
9
7
10
11
12
13
14
10
15
16
17
18
19
20
21
22
t [s]
Figure 2: Utilization of resources for the example instance.
to the hills in the search space W if it is necessary.
Due to efficiency, the Tabu List usually contains only parts of solutions or several previously
applied moves. As moves or parts of solutions do
not have to be unique, it is possible that a forbidden move leads to the best solution. In this case
it is reasonable to permit the move since the resulting solution was not visited before. In general,
exceptions allowing forbidden moves are called aspiration criteria [8].
The quality of the resulting solutions can be
further improved by a suitable search strategy.
For example, if a current location in the solution space is promising, i.e. the best solution
was found recently, then a more thorough search
is performed – intensification. It can be accomplished by concentrating more computational power
to this locality of the space. Opposite to that, if
a current location is unpromising, i.e. only poor
solutions were found, then diversification is performed. The diversification moves the current
search location to another one where better solutions could be found. It is often realized by
applying a few random moves.
The Tabu Search process is stopped if a stop
criterion is met. The stop criterion can be the
5
number of iterations, achieved quality of the best
found solution, the maximal number of iterations
since the last best solution was found, etc.
4. Exploration of the Solution Space
4.1. Creating Initial Activity Order
Our Tabu Search algorithm starts from initial
feasible solution W init ∈ W which is created in
the following way: First of all, the longest paths
in graph G from the start activity 0 to all other
activities are found. The weight of each graph
edge (i, j) ∈ E is set to 1. Activities with the
same maximal distance from the start activity are
grouped to levels. The level lk corresponds to all
activities with maximal distance k from the start
activity, therefore, activity 0 is at level l0 and activity N − 1 is at level lmax where subscript max
corresponds with the last level number. The final
feasible schedule can be created from levels such
that W = {{l0 }, . . . , {lmax }}. Alternative feasible
schedules can be created by shuffling the activities
on the same level.
4.2. Move Transformation
Schedule order W is changed in our Tabu Search
algorithm by a swap move. A simple example is
illustrated in Figure 3. Two dummy activities (0
and N − 1) cannot be swapped due to precedence
constraints, therefore the activity at w0 is always
0 and the activity at wN −1 is always N − 1. The
swap move is defined as swap(u, v) where u and
v are swapped indices. As swap(u, v) modifies a
schedule in the same way as swap(v, u) only swaps
with u < v are taken into account without loss of
generality.
reduced neighborhood, moves are restricted to all
swap(u, v), where u < v and |v − u| ≤ δ. Value δ
is the maximal distance between the swapped activities in order W . The size of the neighborhood
|Nreduced (W )| is parametrized by δ.
There are two reasons why only feasible moves
are applied. The neighborhood size is reduced
without noticeable deterioration of the project makespan and there is no need to check the feasibility
of schedules.
swap(3, 6)
0
3
7
1
2
8
4
6
5
9
0
3
7
4
2
8
1
6
5
9
4.4. Filtering Infeasible Moves
In order to saturate a GPU, feasible schedules in Nreduced should be evaluated in a parallel way by dividing the schedules equally among
the threads. Since the evaluation of a schedule is
much more time-consuming than checking whether
a move is feasible it is advantageous to filter out
all infeasible moves before the neighborhood evaluation. It reduces the branch divergency of warps,
hence the overall performance of the resources
evaluation is improved.
In Algorithm 1 is shown how infeasible moves
are filtered out from the neighborhood. The filter
works in two phases since it was discovered that
it is more effective due to the lower branch divergency than to filter out all the infeasible moves
at once. In the end, only part of the array with
feasible moves is taken into account in the neighborhood evaluation.
Figure 3: Example of the swap move.
Let W ∈ W be a feasible activity order. A
feasible order means that there is not a violated
precedence relation. A feasible move is a move
which does not violate any precedence relation,
therefore if this move is applied to a feasible schedule then the modified schedule will be feasible as
well. Move swap(u, v) is feasible if the following
equations are satisfied.
(wu , wx ) ∈
/E
(wx , wv ) ∈
/E
∀x ∈ {u + 1, . . . , v}
∀x ∈ {u, . . . , v − 1}
(5)
(6)
The First Equation (5) means that there are
no edges from activity wu to the activities at indices from u+1 to v. If there is any edge, then activity wu cannot be moved to position v without
a precedence violation. In a similar way, Equation (6) states that activity wv cannot be moved
to index u if there is a precedence relation that
becomes violated.
4.3. Neighborhood Generation
Full neighborhood Nf ull (W ) ⊆ W of schedule
W is a set of schedules obtained by applying all
feasible moves. Since the full neighborhood is usually too large to be evaluated in a reasonable time
only a subset of the neighborhood is usually taken
into account. Such a subset will be called as a reduced neighborhood denoted Nreduced (W ). In the
Algorithm 1 Removing infeasible moves from the reduced neighborhood.
Require: Nreduced (W )
Ensure: It filters out infeasible moves from the reduced
neighborhood.
1: Let MovesArray be an array containing all potential
swaps in Nreduced (W ).
2: All moves not satisfying Equation (5) are removed, i.e.
set empty.
3: Reorder MovesArray such that all empty moves are in
the end of the array.
4: Remove moves that do not satisfy Equation (6).
5: Move all feasible moves to the beginning of
MovesArray.
4.5. Simple Tabu List and Cache
The tabu list in [10] is not suitable for a GPU
since it is necessary to go through all moves in
6
5. Schedule Evaluation
the tabu list to decide whether a move is in the
tabu list or not. As a consequence it places higher
demands on the device memory bandwidth. To
avoid this a simple and efficient Simple Tabu List
(STL) with constant algorithmic complexity is proposed. Access to the STL is performed like an
access to a circular buffer. Its size is fixed and is
equal to |tabuList|. In our case, the STL stores
swap moves. Each swap(u, v) is stored to the STL
as a pair of swapped indices (u, v). A special value
is used for an empty move, e.g. swap(0, 0). At
each iteration of the TS algorithm one move is
added (see Algorithm 3) and the oldest one is removed if the STL is full.
During the evaluation of W , precedence relations and resource constraints have to be taken
into account to calculate activities start times si
and Cmax . The precedence earliest start time esprec
i
of activity i can be calculated as max∀(j,i)∈E (sj + dj ),
where j are predecessors of activity i. The resources earliest start time esres
can be computed
i
using either a time-indexed or capacity-indexed resources evaluation algorithm. The capacity-indexed
algorithm is a completely new approach to the
best of our knowledge, whereas the time-indexed
algorithm is well-known [15]. The names of the
algorithms were selected with respect to the indexed unit of a resource state array. According
to a heuristic the probable faster resources evaluation algorithm is selected in the schedule evaluation procedure. Having considered both precedence and resource constraints the final earliest
start time is esi = max(esprec
, esres
i ).
i
Algorithm 2 Check if a move is in the STL.
Require: tabuCache − STL cache.
Require: (u, v) − Swap move indices.
Ensure: It returns true if the move is in the STL, otherwise false.
1: return tabuCache[u, v]
The Tabu Cache (TC) was proposed for effective checking if the move is in the STL. It is illustrated in Algorithm 2. Checking if a swap is in the
STL occurs much more often than adding a new
move since a move is added only once per iteration and check if the move is in the STL occurs for
every neighborhood schedule. The TC is implemented as a 2-dimensional N × N boolean array
which is synchronized with the STL. A check if a
move is in the STL requires one read operation,
thus, the required memory bandwidth is very low.
It is obvious that a check if the move is in the STL
has O(1) algorithmic complexity.
5.1. Capacity-indexed resources evaluation
5.1.1. Required Data Structures
The most difficult part during the project makespan evaluation is computation of the activities’
start times with respect to the resource capacities.
In our approach, the evaluation of resources requires one array ck with length Rk per resource k.
Value ck [Rk − ri,k ] corresponds to the earliest resource start time of activity i with resource requirement ri,k > 0 on resource k. At the start of
the evaluation, all the resources arrays are set to
zeros. After that, activities are added one by one
to a schedule according to W and arrays are updated with respect to the activity requirements
and precedences. The resources arrays are ordered descendly, i.e. ck [Rk − l] ≤ ck [Rk − l −
1] | ∀l ∈ {1, . . . , Rk − 1}. The state of resources
is represented as a set C = {c0 , . . . , cM −1 }.
Algorithm 3 Add a move to the STL.
Require: tabuList − Fixed size array.
Require: tabuCache − STL cache.
Require: writeIndex − Current write position.
Require: (u, v) − Swap move indices.
Ensure: Add move to STL and update TC.
1: (uold , vold ) = tabuList[writeIndex ]
2: tabuCache[uold , vold ] = false
3: tabuList[writeIndex ]= (u, v)
4: tabuCache[u, v] = true
5: writeIndex = (writeIndex + 1) % |tabuList|
5.1.2. The Earliest Resources Start Time
∈ N of activResource earliest start time esres
i
ity i with respect to an occupation of resources
can be calculated using Equation (7). It is guaranteed that resources are not overloaded if activity i start time si ≥ esres
i . Final activity start
7
time si can be more delayed due to the precedence
relations.
esres
=
i
max ck [Rk − ri,k ]
k∈{0,...,M −1}: ri,k >0
0
∃ri,k > 0
to be decremented to zero (lines 2, 13, 19) for each
resource k.
It is performed by setting ri,k elements of ck
to the activity finish time si + di after the last
ck ≥ si + di . Until the variable requiredEffort is
zero, the shifted (right shift about ri,k ) copy of
the original resource array ck (original values are
stored in copy auxiliary variable) is made. The
complexity of the algorithm is O(M · Rmax ).
(7)
otherwise
5.1.3. Update of the Resources Arrays
If activity i is added into the schedule, resources arrays C have to be updated by Algorithm 4. Each resource array ck is updated indi-
resource availability - ck
si
Algorithm 4 Method updates state of resources after
adding activity i.
Require: ri,k , di , C, R
Require: copy − Auxiliary array with length Rmax ).
Require: si − Scheduled start time of activity i.
Ensure: Update C - activity i is added.
1: for (k = 0; k < M ; ++k) do
2:
requiredEffort = ri,k · di
3:
if (requiredEffort > 0) then
4:
resIdx = 0; copyIdx = 0
5:
newTime = si + di
6:
while (requiredEffort > 0 AND resIdx < Rk )
do
7:
if (ck [resIdx ] < newTime) then
8:
if (copyIdx ≥ ri,k ) then
9:
newTime = copy[copyIdx − ri,k ]
10:
end if
11:
timeDiff = newTime − max(ck [resIdx ], si )
12:
if (requiredEffort − timeDiff > 0) then
13:
requiredEffort -= timeDiff
14:
copy[copyIdx ++] = ck [resIdx ]
15:
ck [resIdx ] = newTime
16:
else
17:
ck [resIdx ] = max(ck [resIdx ], si )
18:
ck [resIdx ] += requiredEffort
19:
requiredEffort = 0
20:
end if
21:
end if
22:
resIdx = resIdx + 1
23:
end while
24:
end if
25: end for
si + di
7
+1
6
+1
5
Rmax
+3
4
+2
3
+2
activity
required
effort
2
9 units
1
4
5
6
7
8
9
10
resource earliest start time
Figure 4: An example of the resource state update.
The algorithm is illustrated on an example in
Figure 4. There is one resource with maximal
capacity 7. Added activity i requires 3 resource
units (i.e. ri,k = 3) and its duration di is 3. The
activity was scheduled at si = 5. The solid line
corresponds to the original resource state ck =
{7, 7, 5, 5, 5, 5, 4} and the dotted line corresponds
to the updated resource c0k = {8, 8, 8, 7, 7, 5, 4}.
Activity required effort is depicted by a square
with dashed border. The positive numbers between solid and dotted lines are effort contributions when old start time (solid line) will be changed
to the new start time (dotted line). It should be
noticed, that the sum of all contributions is the
requiredEffort for a given activity.
vidually (line 1). Value requiredEffort = ri,k · di
will be called Required Resource Effort. Activity
i can be added to the schedule if and only if each
resource is able to provide its Required Resource
Effort. In other words, variable requiredEffort has
8
5.2. Time-indexed resources evaluation
5.2.1. Required Data Structures
In the time-indexed evaluation algorithm the
state of each resource k is stored in array τk . Each
element τk [t] corresponds to the number of available resource units that resource k is able to provide at time t ∈ {0, . . . , UBCmax }, where UBCmax
is the upper boundPof the makespan which can be
calculated as e.g. ∀i∈V di . Each τk array has initialized all its elements to the Rk value before the
start of the evaluation. The state of all resources
will be denoted as T = {τ0 , . . . , τM −1 }.
5.2.3. Update of the Resources Arrays
The state of resources is updated as is shown
in Algorithm 6. Having scheduled activity i at si
the τk arrays have to be updated in the [si , si +
di ) interval. For each resource k, values in the
interval are decreased by ri,k units.
Algorithm 6 Updating of resources after adding activity i.
Require: di , T, ri,k
Require: si − Scheduled start time of activity i.
Ensure: It updates state of resources T .
for (k = 0; k < M − 1; ++k) do
for (t = si ; t < si + di ; ++t) do
τk [t] -= ri,k
end for
end for
5.2.2. The Earliest Resources Start Time
Algorithm 5 Algorithm calculates the earliest start time
of activity i.
Require: ri,k , di , R, UBCmax , T
Require: esprec
− The precedence earliest start time.
i
Ensure: Calculate the earliest start time esi of activity i.
1: loadTime = 0;
2: for (t = esprec
; t < UBCmax AND loadTime < di ;
i
++t) do
3:
sufficientCapacity = true
4:
for (k = 0; k < M − 1; ++k) do
5:
if (τk [t] < ri,k ) then
6:
loadTime = 0
7:
sufficientCapacity = false
8:
end if
9:
end for
10:
if (sufficientCapacity == true) then
11:
++loadTime
12:
end if
13: end for
14: return t − loadTime
5.3. Schedule Evaluation Procedure
The schedule evaluation procedure is shown
in Algorithm 7. Activities are read one by one
from the activities order W . For each activity wu ,
all its predecessors are found and the precedence
relations are used to update activity wu ’s precedence earliest start time esprec
wu ∈ N (see lines 3–6).
Then the resources restrictions are checked and
res
the start time is adjusted to swu = max(esprec
wu , eswu ).
Project makespan Cmax is the finish time of activAlgorithm 7 Complete schedule evaluation.
Require: W, C, T, E
Ensure: Calculate Cmax and the activities’ start times.
1: Cmax = 0
2: for (u = 0; u < N ; ++u) do
3:
esprec
wu = 0
4:
for all ((j, wu ) ∈ E) do
prec
5:
esprec
wu = max(eswu , sj + dj )
6:
end for
prec
7:
esres
wu = getEarliestResourcesTime(activity wu , eswu )
prec
res
8:
swu = max(eswu , eswu )
9:
updateResources(activity wu , swu )
10:
Mark current activity wu as scheduled.
11:
Cmax = max(Cmax , swu + dwu )
12: end for
13: return Cmax
The earliest start time of activity i can be calculated using Algorithm 5. In the algorithm, the
loadTime variable corresponds with the number
of consecutive time units in which resources are
able to meet resource requirements of activity i. If
loadTime = di then a time interval into which activity i can be scheduled was found. Having considered variable t as a finish time of a candidate
interval, the resulting interval is the first interval
[t − loadTime, t) ∩ N such that loadTime = di .
The final earliest start time is the lower endpoint
of the interval.
ity N − 1. As only feasible moves are allowed, an
infeasibility test of the resulting schedules is not
required.
9
5.4. Heuristic Selection of Resources Evaluation
Algorithms
Before the search is started on the GPU, the
probable faster resources evaluation algorithm is
heuristically selected by decision rules and the required resources arrays are allocated. To create
the rules, the JRip classifier from the Weka datamining tool [16] was learned using pre-calculated
attributes shown in Table 2.
Min. resource capacity:
Avg. resource capacity:
Max. resource capacity:
Avg. activity duration:
Avg. branch factor:
Critical path length:
Evaluation algorithm:
Min. resource capacity ≥ 29
true
TIME
false
Avg. resource capacity ≥ 29
Avg. branch factor ≥ 2.1
true
TIME
min Rk
∀k∈{0,...,M −1}
X
1
Rk
M
∀k∈{0,...,M −1}
false
Min. resource capacity ≥ 25
Max. resource capacity ≥ 42
true
TIME
max Rk
∀k∈{0,...,M −1}
X
1
di
N
∀i∈V
|E|
N
false
CAPACITY
Figure 5: An example of the decision tree.
see Section 2
CAPACITY/TIME
Table 2: Attributes used for learning.
Attribute “Evaluation algorithm” determines
the class, i.e. the time-indexed or capacity-indexed
evaluation algorithm, to which the classifier should
classify. As the final class dependends on the
hardware and instance parameters it is necessary
to determine the probable correct class by measuring — each evaluation algorithm was selected
for a small number of iterations and the faster
one was selected as the desired one. The resulting rules heuristically decide which of the two
algorithms should be more effective for a given
instance. The rules can be transformed into a
decision tree as is shown in Figure 5. Once the
rules are created they can be applied to other similar instances without any measuring overhead as
well. To show the effectivity and usefulness of
the heuristic the experiments were performed in
Section 7.
6. Parallel Tabu Search for CUDA platform
Our Parallel Tabu Search for GPU (PTSG) is
proposed with respect to the maximal degree of
parallelization since thousands of CUDA threads
10
are required to be fully loaded to exploit the graphics card power. In our approach, the parallelization is carried out in two ways. The first one
is a parallelization performed within the scope of
a block, for example the parallel filter (see Section 4.4), the parallel neighborhood evaluation,
and other parallel reductions. The second one is
a parallelization introduced by launching many
blocks on the multi-processors simultaneously.
The basic steps of the PTSG are described in
Figure 6. First of all, an instance is read and the
initial solutions are created in accordance with
Section 4.1. After that, every second solution
is improved by using the forward-backward improvement method. The method is iteratively
shaking a schedule from the left to the right in
order to make a resource profile straight as long
as the schedule is getting shorter. To get more details about the method, refer to the original article
by Li and Willis [17]. The created solutions are
copied into a working set, i.e. a set of shared solutions. The best solution in the working set will
be called the global best solution and its make∗
span will be denoted as Cmax
. Furthermore, the
block’s Tabu Lists and Tabu Caches are initialized and auxiliary arrays such as τk , and ck are
allocated. Having had prepared required datastructures, the host is ready to launch the kernel.
In the GPU part, every block is an independent Tabu Search instance communicating with
the others through the global memory (see Section 6.1). At the beginning, every block reads
nication conditions are satisfied (see Section 6.1).
The search is stopped if the specified number of
∗
is equal to the
iterations was achieved or Cmax
length of a critical path.
After the termination of the kernel, the best
global solution is copied from the global memory
to the host memory. The solution is printed and
all allocated data-structures are freed.
create initial solutions
apply forward-backward improvement method for every second solution
CPU
6.1. Co-operation Among CUDA Blocks and Iterations Distribution
To assure the high quality solutions, the cooperation among Tabu Search instances is accomplished by exchanging solutions through the working set F that has a fixed number of solutions
|F |. Each solution k ∈ F consists of the ork
, the tabu
der of activities W k , makespan Cmax
k
list and iterations counter IC . The solutions exchange takes place if the last read solution has not
been improved for more than Iassigned iterations
or the block found an improvement of the last
read solution. The block writes the best found
solution to F if it improves the last read solution
and reads the next solution from F . Since the
working set could be accessed by many blocks at
the same time, it is necessary to use read/write
locks in order to maintain data integrity. The
co-operation among Tabu Search instances was
inspired by James et al. [9].
After the block has read a solution from the
working set, it is checked whether the solution
was not read more than Φmax times without being
improved. If it is the case, a small number Φsteps
of random feasible swaps is applied to randomize
the read solution — diversification. After that,
the read solution k ∈ F has assigned the number
of iterations Iassigned according to the following
equation.
initialize and fill GPU structures
launch kernel
read an initial solution
generate Nreduced and filter
out all infeasible moves
GPU
(a block)
main loop
evaluate the neighborhood
and select the best move m∗
apply m∗ and add it to STL
no
exchange data?
yes
co-operation
with other blocks
no
stop condition?
yes
load data to CPU
CPU
print the best GPU solution
Figure 6: Parallel Tabu Search for the CUDA platform.
an initial solution from the working set. After
that, the search is started for a specified number
of iterations of the main loop. In the main loop,
the neighborhood is generated, evaluated and the
best move m∗ is selected, applied and added into
the STL. Move m∗ leads to the best criterion improvement or to the smallest criterion deterioration. This move cannot be in the STL with one
exception — the move leads to the global best solution. At the end of an iteration the solutions are
exchanged through the working set if the commu-
quality
}|
{
k
Cmax
−100
−1
∗
quantity
Iassigned
$ z }| {
1 Iblock
=
5 Itotal
z
Cmax
0.8e
z
intactness
}|
{ !%
IC k
−4
+ 0.2e
11
Iblock
(8)
Iblock is the number of iterations assigned to
each Tabu Search instance and Itotal is the total
number of iterations calculated as Iblock B, where
B is the number of launched blocks. The part
denoted as quantity corresponds to the maximal
number of iterations which can be assigned to
read solution k. It is ensured, that at least 5 solutions are read from the working set by each block.
The quality part takes into account the quality of
read solution k. It is obvious that the high quality
solutions are preferred to poor ones — intensification. And the last part intactness guarantees
that each solution k ∈ F has been given some
iterations to prove the quality.
CUDA toolkit (version 5.0.35) and Microsoft Visual Studio 2010.
The sequential CPU version of the algorithm
corresponds to one Tabu Search instance with the
exception that solutions are not interchanged (|F | =
1 and B = 1). Instead of using the selection
heuristic (see Section 5.4) the faster evaluation
algorithm was selected dynamically by periodic
measuring every 1000 iterations. The parallel CPU
version differs from the sequential version in the
neighborhood evaluation. The feasible schedules
in the neighborhood are divided among CPU threads
to reduce evaluation time. Both the CPU and
GPU versions were fully optimized with respect
to memory access patterns and hardware architecture (cache sizes). To fully saturate the GPU the
maximal number of available registers per CUDA
thread was limited to 32 due to possibility to
launch 4 blocks on a multiprocessor at once (altogether 16 blocks on the GPU), where each block
has 512 CUDA threads.
6.2. Memory Model
The placement of the data-structures is a crucial task highly influencing the effectiveness of the
GPU program, therefore, each decision should be
considered thoroughly with respect to the access
pattern, required bandwidth and data visibility
(local or shared data). In the shared memory current block order Wblock , durations of activities D,
and auxiliary arrays are stored. Although D is
a read-only array which could be located in the
constants memory, it was moved to the shared
memory due to the higher bandwidth. The texture memory is used for storing read-only data as
ri,k values and predecessors of the activities. In
the local memory, private data-structures of each
thread are located, i.e. resources arrays ck , τk ,
and start times of activities S. The long latency
of the memory is compensated by using a partial coalescing since the arrays are often accessed
at the same relative indices as the majority of
threads evaluate similar schedules (Wblock + swap
move). Finally, the global memory is employed to
store the working set F .
J30
J60
J90 J120
N
30+2 60+2 90+2 120+2
M
4
|dataSet|
480
480
480
600
δ
30
60
60
60
|tabuList|
60
250
600
800
Φsteps
20
Φmax
3
|F |
16
Table 3: PTSG parameters and data-sets information.
To evaluate the performance and the quality of
resulting solutions the benchmark using the wellknown J30, J60, J90 and J120 data-sets was performed. The number of instances in a data-set
will be denoted as |dataSet|. The selected PTSG
parameters and data-sets information are stated
in Table 3.
The results for the J30 data-set are shown in
Tables 4 and 6. The CPM dev and OPT dev values are the average percentage distance from the
critical path length and the average percentage
distance from the optimal makespan respectively.
Best sol states the number of optimal solutions
which have been proved to be optimal according
7. Experimental Results
Experiments were performed on the AMD Phenom(tm) II X4 945 server (4 cores, 8 GB memory) equipped with a mid-range Nvidia Geforce
GTX 650 Ti (1 GB, 768 cuda cores, 4 multiprocessors) graphics card. The testing environment
was the Windows Server 2008 with an installed
12
CPU
OPT dev
0.04
–
Itotal CPM dev
10000
13.43
20000
–
Best sol
471
–
CPM dev
13.41
13.38
GPU
OPT dev
0.02
0.01
Best sol
473
478
Table 4: Quality of solutions — J30.
Itotal CPM dev
10000
11.13
20000
–
30000
–
50000
–
CPU
UB dev
0.51
–
–
–
Best sol
380
–
–
–
CPM dev
11.22
11.08
10.99
10.91
GPU
UB dev
0.57
0.47
0.41
0.36
Best sol
375
388
394
394
Table 5: Quality of solutions — J60.
CPU seq.
CPU par.
GPU
GPU
Itotal Comp time
10000
1255
10000
343
10000
176
20000
306
Sched sec
126400
478300
985543
1120900
CPU seq.
CPU par.
GPU
GPU
GPU
GPU
Speedup
1.00
3.65
7.12
–
Table 6: Performance comparison — J30.
Itotal CPM dev
10000
10.81
20000
–
30000
–
50000
–
CPU
UB dev
0.92
–
–
–
Itotal Comp time
10000
7094
10000
1732
10000
257
20000
485
30000
709
50000
1164
Sched sec
59700
248700
1733800
1818600
1861900
1879400
Speedup
1.00
4.10
27.60
–
–
–
Table 7: Performance comparison — J60.
Best sol
367
–
–
–
CPM dev
11.04
10.82
10.73
10.56
GPU
UB dev
1.09
0.93
0.86
0.74
Best sol
365
371
373
375
Table 8: Quality of solutions — J90.
Itotal CPM dev
10000
33.41
20000
–
30000
–
50000
–
CPU
UB dev
2.70
–
–
–
Best sol
215
–
–
–
CPM dev
34.67
34.04
33.66
33.54
GPU
UB dev
3.50
3.11
2.85
2.76
Best sol
194
208
213
222
Table 9: Quality of solutions — J120.
CPU seq.
CPU par.
GPU
GPU
GPU
GPU
Itotal Comp time
10000
20294
10000
5001
10000
475
20000
923
30000
1348
50000
2221
Sched sec
36000
148300
1599600
1632000
1660700
1674000
Speedup
1.00
4.06
42.70
–
–
–
CPU seq.
CPU par.
GPU
GPU
GPU
GPU
Table 10: Performance comparison — J90.
Itotal Comp time
10000
148170
10000
35812
10000
2938
20000
5742
30000
8513
50000
14160
Sched sec
25700
107200
1340400
1351900
1353300
1347800
Speedup
1.00
4.14
50.40
–
–
–
Table 11: Performance comparison — J120.
13
Algorithm and reference
Genetic Algorithm - Gonçalves et al. [18]
This work - Nvidia Geforce GTX 650 Ti
This work - AMD Phenom(tm) II X4 945
CARA algorithm - Valls et al. [19]
Ant Colony Optimization - Zhou et al. [20]
Tabu Search - Artigues et al. [21]
Simulated Annealing - Bouleimen and Lecocq [22]
J30
13.38
13.38
13.43
13.46
-
CPM dev
J60
J90 J120
10.49
30.08
10.91 10.56 33.54
11.13 10.81 33.41
11.45 11.12 34.53
11.42
35.11
12.05
36.16
11.90
37.68
Table 12: Comparison with other heuristics.
to the results in the PSPLIB homepage — http:
//www.om-db.wi.tum.de/psplib/. Comp time
is the total run-time stated in seconds and the
Sched sec is the number of evaluated schedules
per second. It is obvious that the GPU version
is able to achieve a similar quality of solutions in
terms of CPM dev as the CPU version. Having
had Itotal doubled, the GPU version found 478
optimal solutions from the 480 solutions in the
data-set. From the performance point of view Table 6 reveals a significant improvement in computational time if parallelization is performed. For
example, the parallel CPU version is 3.65 times
faster than the sequential CPU version and the
GPU is almost 2 times faster than the parallel
CPU version. If Itotal is increased to 20000 the
GPU is still slightly faster and achieves better
quality solutions.
For the J60 data-set the results are shown in
Tables 5 and 7, where UB dev is the average percentage distance from the best currently known
upper bounds. The CPU version gives slightly
better solutions for 10000 iterations, but on the
other hand if the GPU is given 20000 iterations
the quality of solutions is comparable with the
CPU version and the GPU is still 3.56 times faster
than the parallel CPU version. The lower quality
of GPU solutions for the same Itotal is probably
caused by wasting work when many Parallel Tabu
Search instances have read the same solution from
the working set and only one writes the best improvement. It can be noted, that the parallel CPU
version is more than 4 times faster than the sequential one. The reason of that is either better
cache utilization or the AMD True Core Scalability technology.
The results in Tables 8 and 10 for the J90 dataset show that the GPU is better utilized for bigger instances and the GPU is more than 10 times
faster than the parallel CPU version for the same
number of iterations. The same quality of solutions is achieved 5.4 times faster on the GPU.
Results for the J120 data-set are shown in Tables 9 and 11. It can be noted that the quality
of GPU solutions is substantially lower for 10000
iterations. The GPU requires about 50000 iterations to achieve the quality of the CPU solutions. On the other hand, the GPU is able to
compete with the CPU since 50000 iterations is
performed 2.5 times faster than 10000 iterations
for the parallel CPU version. The GPU evaluates more than one million schedules per second,
whereas the CPU evaluates one hundred thousand.
The quality of the solutions is compared with
the existing solutions for the RCPSP in Table 12.
Our proposed PTSG outperforms other Tabu Search
implementations with respect to the quality of
solutions. For example, Artigues’ Tabu Search
[21] has been given at least 11000 iterations for
the J120 data-set and achieves 36.16 % CPM dev.
Having had 10000 iterations the proposed PTSG
reaches 33.41 % and 34.67 % for the CPU and
GPU respectively. In addition, the proposed PTSG
can be just as good as other heuristic approaches
like Ant Colony Optimization and Simulated Annealing. On the other hand, the state of the art
random-key genetic algorithms give even better
solutions than the PTSG.
From the performance point of view it is difficult to compare since the different algorithms and
hardware architectures were used for experiments.
14
dependent on the characteristics of instances and
it cannot be generally determined which evaluation algorithm is faster. The capacity-indexed
evaluation algorithm is usually faster for long schedules with low resource capacities in contrast to the
time-indexed algorithm which usually performs
better for short schedules with high resource capacities.
For example, Artigues’s Tabu Search requires 67 s
per J120 instance on average. The testing configuration was not stated. The PTSG requires 4.9 s
(10000 iterations) on the mid-range GPU with the
substantially higher quality of solutions. The genetic algorithm by Gonçalves et al. [18] takes
180 s per J120 instance on average on the Intel
Core 2 Duo 2.4 GHz processor.
7.1. Evaluation of the Selection Heuristic
The heuristic (see Section 5.4) is using the
JRip classifier from the Weka data mining tool [16]
to decide which resources evaluation algorithm
should be faster. To get data for the learning,
the Progen generator [23] was used to generate
4 data-sets with 30, 60, 90, and 120 activities
respectively. The parameters of the generated
data-sets were set the same as for J30, J60, J90,
and J120 data-sets with the exception that different random seeds were used. For each generated data-set the classifier was learned using
weka.classifiers.rules.JRip -F 3 -N 2.0
-O 10 -S 0 command and tested on the corresponding standard data-set with the same number of activities. The achieved results in Table 13
reveal that the accuracy is decreasing with the
number of activities. The reason for this behavior
can be the smaller ratio of the resources evaluation time to the total run-time.
J30
72.3 %
J60
85.8 %
J90
91.9 %
time-indexed
capacity-indexed
heuristic
J30 J60 J90 J120
1.02 1.10 1.09 1.23
1.27 1.34 1.96 1.73
1
1
1
1
Table 14: Effect of the heuristic on the PTSG performance.
7.2. Demonstration of Convergence
To demonstrate that co-operation among blocks
is beneficial the graph of convergency (in Figure 7)
was created for j1206_4.sm instance. It can be
seen that the quality of solutions is getting better
with the increasing number of launched blocks,
therefore, it is obvious that co-operation leads to
the better solutions. To ensure the smoothness of
J120
96.3 %
Table 13: Accuracy of the Selection Heuristic — the percentage of correctly classified.
To prove that the proposed heuristic also improves the PTSG performance the run-time was
measured for each evaluation algorithm and normalized with respect to the reference run-time, i.e.
the run-time achieved by using the heuristic. The
results in Table 14 show that the heuristic accelerates the PTSG up to 2 times and its effect is decreasing as the evaluation of schedules becomes a
less time-consuming part of the PTSG. The timeindexed algorithm seems to be faster than the
capacity-indexed algorithm on the standard datasets. On the other hand, the achieved speedup is
Figure 7: Graph of convergence for the GPU version.
the graph each point was averaged over 50 measurements.
8. Conclusion
The first known GPU algorithm dealing with
the Resource Constrained Project Scheduling Prob15
lem has been proposed. The performed experiments on the standard benchmark instances reveal the merits of the proposed solution. The
achieved quality of solutions is very good and outperforms the other Tabu Search implementations
to the best of our knowledge. In addition to this,
the GPU algorithm design has proved to be very
effective since the mid-range GPU was substantially faster than the optimized parallel CPU version. The Nvidia Geforce GTX 650 Ti GPU is
able to evaluate more than one million schedules
per second for the J120 data-set on average. The
achieved performance boost could not be reached
without effective structures and auxiliary algorithms. The Simple Tabu List implementation is
adapted to the features of the GPU, the capacityindexed evaluation algorithm was proposed and
many parallel reductions were applied. In addition to this, the homogeneous model reduces the
required communication bandwidth between the
CPU and the GPU.
In spite of the fact that GPUs are not primarily designed for solving combinatorial problems
the rising interest about these solutions can be
seen [24]. The reason for this is the high computational power of graphics cards and the relatively
user friendly programming API that the CUDA
offers. So it can be expected that GPUs will be
more and more used in operations research in the
future.
[5]
[6]
[7]
[8]
[9]
[10]
[11]
[12]
[13]
[14]
References
[1] N. Fu, H. C. Lau, P. Varakantham, F. Xiao, Robust
Local Search for Solving RCPSP/max with Durational Uncertainty, J. Artif. Int. Res. 43 (1) (2012)
43–86.
[2] J. Blazewicz, J. Lenstra, A. Kan, Scheduling subject
to resource constraints: classification and complexity,
Discrete Applied Mathematics 5 (1) (1983) 11–24.
[3] A. S. Chaleshtarti, S. Shadrokh, Branch and
Bound Algorithms for Resource Constrained Project
Scheduling Problem Subject to Cumulative Resources, in: Proceedings of the 2011 International
Conference on Information Management, Innovation
Management and Industrial Engineering - Volume 01,
ICIII ’11, IEEE Computer Society, Washington, DC,
USA, 2011, pp. 147–152.
[4] A. Delvacq, P. Delisle, M. Gravel, M. Krajecki, Parallel Ant Colony Optimization on Graphics Processing
[15]
[16]
[17]
[18]
16
Units, Journal of Parallel and Distributed Computing
73 (1) (2013) 52–61.
M. Czapiński, An effective Parallel Multistart Tabu
Search for Quadratic Assignment Problem on CUDA
platform, Journal of Parallel and Distributed Computing (2012).
W. Božejko, Z. Hejducki, M. Uchroński, M. Wodecki,
Solving the Flexible Job Shop Problem on MultiGPU, Proceedings of the International Conference on
Computational Science, ICCS 2012 9 (0) (2012) 2020–
2023.
F. Glover, Future Paths for Integer Programming and
Links to Artificial Intelligence, Comput. Oper. Res.
13 (5) (1986) 533–549.
M. Gendreau, An Introduction to Tabu Search, in:
F. Glover, G. Kochenberger (Eds.), Handbook of
Metaheuristics, Vol. 57 of International Series in Operations Research & Management Science, Springer
New York, 2003, pp. 37–54.
T. James, C. Rego, F. Glover, A cooperative parallel
tabu search algorithm for the quadratic assignment
problem, European Journal of Operational Research
195 (3) (2009) 810–826.
M. Czapiński, S. Barnes, Tabu Search with two approaches to parallel flowshop evaluation on CUDA
platform, J. Parallel Distrib. Comput. 71 (2011) 802–
811.
T. Zajı́ček, P. Šůcha, Accelerating a Flow Shop
Scheduling Algorithm on the GPU, Workshop on
Models and Algorithms for Planning and Scheduling
Problems (MAPSP).
J. Hofmann, S. Limmer, D. Fey, Performance investigations of genetic algorithms on graphics cards,
Swarm and Evolutionary Computation (2013).
V. Boyer, D. El-Baz, M. Elkihel, Solving knapsack
problems on gpu, Computers & Operations Research
39 (1) (2012) 42–47.
P. Brucker, A. Drexl, R. Möhring, K. Neumann,
E. Pesch, Resource-constrained project scheduling:
Notation, classification, models, and methods, European Journal of Operational Research 112 (1) (1999)
3–41.
J. E. Kelley, 1963. The critical-path method:
Resources planning and scheduling. In: Muth,
J.F., Thompson, G.L. (Eds.), Industrial Scheduling.
Prentice-Hall, Englewood Cliffs, NJ, pp. 347–365.
M. Hall, E. Frank, G. Holmes, B. Pfahringer,
P. Reutemann, I. H. Witten, The WEKA Data Mining Software: An Update, SIGKDD Explor. Newsl.
11 (1) (2009) 10–18.
K. Li, R. Willis, An iterative scheduling technique
for resource-constrained project scheduling, European Journal of Operational Research 56 (3) (1992)
370–379.
J. F. Gonçalves, M. G. Resende, J. J. Mendes, A Biased Random-Key Genetic Algorithm with Forward-
[19]
[20]
[21]
[22]
[23]
[24]
Backward Improvement for the Resource Constrained
Project Scheduling Problem, Journal of Heuristics
17 (5) (2011) 467–486.
V. Valls, S. Quintanilla, F. Ballestı́n, Resourceconstrained project scheduling: A critical activity reordering heuristic, European Journal of Operational
Research 149 (2) (2003) 282–301.
L. Zhou, D. Wang, W. Peng, An ACO for Solving
RCPSP, in: Computer Science and Computational
Technology, 2008. ISCSCT ’08. International Symposium on, Vol. 2, 2008, pp. 250–253.
C. Artigues, P. Michelon, S. Reusser, Insertion techniques for static and dynamic resource-constrained
project scheduling, European Journal of Operational
Research 149 (2) (2003) 249–267.
K. Bouleimen, H. Lecocq, A new efficient simulated
annealing algorithm for the resource-constrained
project scheduling problem and its multiple mode
version, European Journal of Operational Research
149 (2) (2003) 268–281.
R. Kolisch, C. Schwindt, A. Sprecher, Benchmark Instances for Project Scheduling Problems, in: Handbook on Recent Advances in Project Scheduling,
Kluwer, 1998, pp. 197–212.
A. R. Brodtkorb, T. R. Hagen, M. L. Sætra, Graphics
processing unit (GPU) programming strategies and
trends in GPU computing, Journal of Parallel and
Distributed Computing 73 (1) (2013) 4–13.
17
| 2cs.AI
|
Optimal Resource Allocation in Distributed
Broadband Wireless Communication Systems
Yao Yao, Mustafa Mehmet-Ali, Shahin Vakilinia
Department of Electrical & Computer Engineering, Concordia University
Montreal, Canada
E-mail: {y_yao11,s_vakili}@encs.concordia.ca, [email protected],
Abstract—This paper is concerned with optimization of
distributed broadband wireless communication (BWC) systems.
BWC systems contain a distributed antenna system (DAS)
connected to a base station with optical fiber. Distributed BWC
systems have been proposed as a solution to the power constraint
problem in traditional cellular networks. So far, the research on
BWC systems have advanced on two separate tracks, design of
the system to meet the quality of service requirements (QoS) and
optimization of location of the DAS. In this paper, we consider a
combined optimization of BWC systems. We consider uplink
communications in distributed BWC systems with multiple levels
of priority traffic with arrivals and departures forming renewal
processes. We develop an analysis that determines packet delay
violation probability for each priority level as a function of the
outage probability of the DAS through the application of results
from renewal theory. Then, we determine the optimal locations of
the antennas that minimize the antenna outage probability. We
also study the trade off between the packet delay violation
probability and packet loss probability. This work will be helpful
in the designing of the distributed BWC systems.
Index Terms— Queuing delay, multiple levels of priority
traffic, distributed antenna system (DAS), outage probability,
antenna placement.
I. INTRODUCTION
The future services in wireless networks will be requiring
higher transmission rates. This rate is expected to rise to
1Gbits/s for the next generation communication systems
according to the objective of IMT-Advanced (International
Mobile Telecommunications) system, which is set by ITU-R
(International
Telecommunication
UnionRadiocommunication Standardization Sector) [1-3]. However,
the constraint on the transmit power limits the transmission
rate, especially in the uplink communications from user to
base station. One promising solution to this problem is the
distributed broadband wireless communications (BWC)
systems [2].
A distributed BWC system is made up of a distributed
antenna system (DAS) and the radio over fiber (RoF)
technology [2]. In DAS, antennas are placed at geographically
separate locations in the cell [2], which is a contrary to the so
called centralized antenna system (CAS) that having an
antenna set at the center of a cell. The RoF technology, which
is very reliable and has very small delay, is responsible for
providing communication between these antennas and a
central processor where they are jointly processed [2]. In this
way, a user is very likely to have an antenna nearby compared
to CAS, which will in turn reduce the transmission power
requirement of that user. What’s more, it’s possible to place
many distributed antennas in each cell due to the relatively
low cost of distributed antennas [4].
By adopting DAS, a BWC system can solve the power
constraint problem naturally instead of shrinking the cell size
as in conventional CAS, which results in larger overhead and
delay due to higher frequency of handovers between cells.
There has been a number of research works on distributed
BWC systems. These work have dealt with two main
problems, which are determining the antenna set that meet
QoS metrics such as delay and packet loss and the other
optimal placement of the antennas.
[5] considers the downlink transmission, from base station
to the user, and assumes that the central server at the base
station maintains a separate queue for each of the users. This
work considers selection of a subset of the available
distributed antennas such that the probability of packet
queuing delay is less than a threshold. [6] considers uplink
transmission in a distributed BWC system. Again, the
objective has been selection of a subset of distributed antennas
at each user to keep packet loss probability below a threshold
value. In both [5] and [6], locations of the antennas are fixed
and the objective function has not been optimized with respect
to (wrt) the location of the antennas.
Optimal placement of antennas has been considered in other
works but independent of QoS metrics. In [4, 7], the objective
has been to optimize the sum-rate uplink capacity of a single
cell with multiple users. It is assumed that the antennas are
placed on a circle centered at the cell center and their locations
on the circle are uniformly distributed independent of each
other. The optimal radius of the antenna circle has been
determined, however, this work is more appropriate to the
CDMA systems since all the users in the cell will be
transmitting simultaneously. Further, the paper considers the
effects of inter-cell interference only by simulation.
In [8], the optimal location of antennas in a cellular setting
has been studied for the downlink transmission based on the
performance metrics of capacity and power efficiency. This
work takes the inter-cell interference into account. Though no
closed form results are obtained, their numerical results show
that for some certain cases, the optimal layout consists of
placement of one antenna at center and remaining ones on a
circle around it, and that circle shrinks as the interference
coefficient increases. This work does not provide any results
on the positions of the antennas on the circle.
In this paper, we consider the uplink communications and
assume that each user has multiple levels of priority traffic
Fig. 1.
Topology of the distributed BWC system which has
a cellular network architecture
with different QoS requirements. We determine the packet
delay violation probability through the use of results from
renewal theory for each type of traffic as a function of outage
probability. Then, we determine the optimal location of
antennas that minimizes the antenna outage probability. We
note that our work is more general compared to [5, 6], because
it considers multiple types of traffic with different QoS
requirements. In relation to optimal placement of antennas,
our work is closer to that in [8], but we consider the uplink as
opposed to downlink communications, which is more critical
due to limited transmit power of the users.
The outline of the remainder of the paper is as follows.
Section Ⅱ. describes our system model. Section Ⅲ presents
the queuing delay analysis for multiple types of traffic flows.
Section Ⅳ gives the application of these results to a system
with two types of traffic flows. Section V. extends the
application to a system with four types of traffic flows. Next,
section VI. presents the work on the optimal placement of the
antennas. Section VII presents the numerical and simulation
results regarding the analysis. And finally section VIII gives
the conclusions of the paper.
II. SYSTEM MODEL
A. Network Topology
We consider a cellular network architecture with distributed
antenna system. Each cell together with its neighbors are
referred to as a cluster. The cells in a cluster will be numbered
as 𝐴𝑗 , 0≤j≤F-1, where F is the cluster size. 𝐴0 will denote
the cell at the center of the cluster and it will be referred as the
target cell. Fig. 1 shows an example network with a cluster of
seven cells. Each cell contains a central processor and M
distributed antennas, which are placed at different
geographical locations and are linked to the central processor
through RoF technology. In the figure only the antennas in the
target cell 𝐴0 have been shown to prevent crowding, however,
each of the other cells also contains M distributed antennas.
It will be assumed that all the antennas have the same height
𝒽, and the users are all at the ground level. In the target cell,
locations of the M antennas wrt the cell center will be denoted
by (𝐿𝑚 , 𝛩𝑚 , 𝒽) in polar coordinates, for 1≤m≤M. We
number the antennas according to their increasing polar angles,
thus, 𝜃𝑚 > 𝜃𝑚−1 . Define antennas location vector 𝑳 =
((𝐿1 , 𝛩1 , 𝒽), (𝐿2 , 𝛩2 , 𝒽), … (𝐿𝑀 , 𝛩𝑀 , 𝒽))
B. Frequency Reuse Model
We will assume that all the frequencies will be available in
every cell, thus the reuse factor will be one as in [8]. The
available spectrum will be divided into a number of channels
and each user will be assigned a single channel at any time.
We will also assume that the system is saturated, so that all the
channels will always be busy, and the users using the same
channel in the neighboring cells will be interfering with each
other. This corresponds to the worst case scenario of
co-channel interference which is the major cause of an outage
in a cellular system [9]. From Fig. 1, a user in the target cell
will be considered as a target user and the users in the
neighboring cells using the same channel are considered as the
interferers.
Thus there will be a single user in each cell using the same
channel in a cluster. Let user i refer to the user in cell 𝐴i , who
is using a particular channel in the cluster. We will let
(ℓ𝑖 , 𝜃𝑖 , 0) denote the location of user i in polar coordinates
relative to the center of its home cell and (𝑥𝑖 , 𝑦𝑖 , 0) denote its
location in Cartesian coordinates relative to the center of cell
𝐴0 . Then, the latter may be expressed in terms of the former as
follows,
(𝑥𝑖 , 𝑦𝑖 , 0) = (𝑥𝑖𝐶 + ℓ𝑖 𝑐𝑜𝑠 𝜃𝑖 , 𝑦𝑖𝐶 +ℓ𝑖 𝑠𝑖𝑛 𝜃𝑖 , 0)for 0≤i ≤F-1.
C. Transmission Model
The time-axis is slotted and the transmission of a packet
always starts at the beginning of a slot and it takes one slot. A
user will broadcast its packet which will be received by the
antennas in its cell, the received signals will be locally
decoded and the results will be forwarded to the central
processor through fiber lines, respectively. If at least one of
the antennas can decode the packet successfully, the central
processor will receive the transmitted packet correctly. On the
other hand, if none of the antennas is able to decode the packet
successfully, then the central processor will drop this packet or
ask for retransmissions, depending on the type of the packet.
III. QUEUING DELAY
In this section, we will determine the queuing delay bounds
for a user with multiple types of traffic with different QoS
requirements. We will assume that a user has N types of traffic
flows, numbered as 1≤n≤N. The packets of flows are stored
in separate queues. Each traffic flow is assigned a priority
level, which increases with decreasing flow number. A packet
of a traffic flow is only served if the queues of higher priority
traffic flows are empty. Let 𝑥𝑛 , 𝑦𝑛 denote independent and
identically distributed inter-arrival time and service time of the
packets of traffic flow n respectively, if the system was
serving only this traffic flow. Let 𝜇𝑥𝑛 , 𝜎𝑥2𝑛 and 𝜇𝑦𝑛 , 𝜎𝑦2𝑛
denote the mean and variance of the random variables 𝑥𝑛 , 𝑦𝑛
respectively. We note that the service time of a packet begins
when it reaches to the head of its queue and completed either
following its successful transmission or discarding by the
system.
In this section, we will determine the queuing delay bounds
of packets of different flows. We will use the delay bound
violation probability from the theory of effective bandwidth in
our analysis [10, 11]. The same metric had also been used in
[5]. The delay-bound violation probability is defined as the
probability that the packet delay exceeds a certain value
𝑛
𝑃𝑟𝑜𝑏(𝐷𝑛 > 𝐷𝑡ℎ
)
(1)
𝑛
for the traffic flow n, where 𝐷𝑛 and 𝐷𝑡ℎ
represent the
queuing delay and a queuing delay threshold respectively. The
delay bound is determined through the utilization of energy
function, which is defined as the asymptotic log moment
generating function of arrival and service processes [10].
Initially, we will assume that the system is operating with a
single type of traffic, 𝑛𝑡ℎ traffic flow, and later on the results
will be extended to multiple types of traffic. Let 𝐴𝑛 (𝑡)
denote the number of packet arrivals to the queue during the
time interval [0, t). Defining 𝐶𝑛 (t) as the number of
departures from the queue during the interval [0, t) assuming
that it is under saturation. Let 𝛬𝑛 (𝜑) and 𝛤𝑛 (𝜑), (𝜑 > 0),
denote the energy functions of the arrival and service
processes [10],
1
𝛬𝑛 (𝜑) = 𝑙𝑖𝑚𝑡→∞ 𝑙𝑜𝑔 𝐸(𝑒 𝜑𝐴𝑛 (𝑡) )
Λ𝑛 (𝜑) =
𝜑
𝜇𝑥 𝑛
+
𝜑2 𝜎2𝑥𝑛
2𝜇3𝑥𝑛
𝛤𝑛 (𝜑) = 𝑙𝑖𝑚𝑡→∞ 𝑙𝑜𝑔 𝐸(𝑒 𝜑𝐶𝑛 (𝑡) )
(3)
𝑡
From [5], we have the delay-bound violation probability for
the packets of 𝑛𝑡ℎ traffic flow,
∗ )𝐷𝑛
𝑡ℎ
(4)
𝛤𝑛 (𝜑) =
,
𝜑
𝜇𝑦
+
𝑛
𝜑2 𝜎2𝑦
2𝜇3𝑦
𝑛
(5)
𝑛
Next, we will extend the results to a user with multiple types
of traffic. In this scenario, the 𝑛𝑡ℎ traffic flow will have
priority level n. The arrival process of this traffic flow will
remain as before but its service process will change because of
the services given to the packets of higher priority flows. Let
𝐶̃𝑛 (t) denote the number of departures from the 𝑛𝑡ℎ flow
during the interval [0, t),
𝐶̃𝑛 (𝑡) = 𝐶𝑛 (𝑡 − ∑𝑛−1
𝑗=1 𝑌𝑗 (𝑡))
(6)
𝐴𝑗 (𝑡)
∑𝑖=1
𝑦𝑗,𝑖
where,
𝑌𝑗 (𝑡) =
with 𝑦𝑗,𝑖 denoting the service time of the 𝑖 𝑡ℎ packet from the
𝑗𝑡ℎ flow. Then the above may be written as,
𝐶̃𝑛 (𝑡) = 𝐶𝑛 (𝑡) − ∑𝑛−1
𝑗=1 𝐶𝑛 (𝑌𝑗 (𝑡))
(7)
Next letting 𝛤̃𝑛 (𝜑) denote the energy function of the
service process 𝐶̃𝑛 (t),
𝑛−1
1
(𝑡)−𝜑 ∑𝑗=1 𝐶𝑛 (𝑌𝑗 (𝑡))
𝛤̃𝑛 (𝜑) = 𝑙𝑖𝑚𝑡→∞ 𝑙𝑜𝑔 𝐸 [𝑒 𝜑𝐶𝑛
]
𝑡
(8)
We will assume that the system is stable, thus none of the
queues are starving for service. Then, we have 𝑡 > ∑𝑛−1
𝑗=1 𝑌𝑗 (𝑡),
as a result, the random variables 𝑌𝑗 (𝑡) with different j will be
independent of each other. Thus,
1
−𝜑𝐶𝑛 (𝑌𝑗 (𝑡))
𝛤̃𝑛 (𝜑) = 𝑙𝑖𝑚𝑡→∞ 𝑙𝑜𝑔 {𝐸[𝑒 𝜑𝐶𝑛 (𝑡) ] ∏𝑛−1
]} (9)
𝑗=1 𝐸 [𝑒
𝑡
Expressing the above as sum of two logarithms and
substituting from (3),
(2)
𝑡
1
𝑛
𝑃𝑟𝑜𝑏(𝐷𝑛 > 𝐷𝑡ℎ
) ≈ 𝑒 −𝛬𝑛 (𝜑
infinity where 𝜇 and 𝜎 2 are the mean and variance of the
renewal interval [12]. The application of this result gives the
energy function of the arrival and service processes as follows,
1
𝑛−1
𝛤̃𝑛 (𝜑) = 𝛤𝑛 (𝜑) + 𝑙𝑖𝑚 {∑ 𝑙𝑜𝑔𝐸 [𝐸[𝑒−𝜑𝐶𝑛(𝑘) |𝑌𝑗 (𝑡)]]} (10)
𝑡→∞
𝑡
𝑗=1
Again, application of the central limit theorem for the
renewal processes to the conditional expectation gives,
𝜑2 𝜎2
𝑦𝑛
)
𝑦𝑛 2𝜇3
𝑦𝑛 ]
𝜑
𝑌𝑗 (𝑡)(−𝜇
𝐸 [𝐸[𝑒 −𝜑𝐶𝑛 (𝑘) |𝑌𝑗 (𝑡)]] = 𝐸 [𝑒
+
(11)
∗
where 𝜑 is the unique solution of 𝛬𝑛 (𝜑) + 𝛤𝑛 (−𝜑) = 0.
In general, it may not be possible to find the energy
functions defined in (2-3), which limits the applicability of
this method. In that case, we propose to use the asymptotic
central limit theorem for renewal processes in their
evaluations, which states that in a renewal process, the number
of renewal points asymptotically approaches to a Gaussian
distribution with mean
𝑡
𝜇
, and variance
𝜎2𝑡
𝜇3
as 𝑡 goes to
From its definition in (6), 𝑌𝑗 (𝑡) is a random sum with mean
and variance given by,
𝜇𝑌𝑗 (𝑡) =
𝜇𝑦𝑗 𝑡
𝜇 𝑥𝑗
,
𝜎𝑌2𝑗 (𝑡) = 𝜇𝑦2𝑗
𝜎𝑥2𝑗 𝑡
3
𝜇𝑥
𝑗
+
𝜎𝑦2𝑗 𝑡
𝜇 𝑥𝑗
(12)
From the general central limit theorem, 𝑌𝑗 (𝑡) approaches
to a normal distribution for large values of t with mean and
variance given in the above. Then, the expectation on the RHS
of (11) corresponds to the moment generating function of a
normal random variable, as a result (10) becomes,
𝜙2 𝜎𝑌2
𝑛−1
1
𝜙𝜇
+
∑ 𝑙𝑜𝑔𝑒 𝑌𝑗 (𝑡)
𝑡→∞ 𝑡
𝛤̃𝑛 (𝜑) = 𝛤𝑛 (𝜑) + 𝑙𝑖𝑚
𝑗 (𝑡)
2
𝑗=1
𝜙=−
{
Substituting from (5,12) in the above,
𝑛
2𝜇𝑦
𝑛
𝜇𝑥
𝑗
{
2
2 2
𝜑 𝜑 𝜎𝑦𝑛
𝜇𝑦
𝑛
𝜇𝑦
2
𝜎2𝑦𝑗
𝜎2𝑥𝑗
𝜑2 𝜎2𝑦
𝜑
𝜙
𝑗
𝛤̃𝑛 (𝜑) =
+ 3 𝑛 + ∑𝑛−1
+ (𝜇2𝑦 3 + )|
𝑗=1 𝜙
𝜇𝑦
|
𝑗
𝜇𝑥
𝑗
𝜇𝑥
𝑗
+
2𝜇3𝑦
𝑛
𝜑2 𝜎2
𝜑
𝑦𝑛
+
𝑦𝑛 2𝜇3
𝑦𝑛
𝜙=−𝜇
}
(13)
}
This completes the derivation of the energy function of the
service process of the 𝑛𝑡ℎ flow.
Next, we will demonstrate the accuracy of the asymptotic
central limit theorem for renewal processes through an
example. Let us consider the Binomial process, whose exact
energy function is known [11]. The exact and asymptotic
energy functions of a Binomial process are given by,
respectively. The voice and data packets will be stored in
separate queues and voice will be given higher priority wrt
data traffic. It will be assumed that a transmitted voice packet
that can not be decoded will be dropped since voice is delay
sensitive. On the other hand, the failed data packets will be
retransmitted and will be lost only after L unsuccessful
transmissions. In the following, we will study the delay and
packet losses of data flow since voice flow is redundant and is
given higher priority. Assume that both voice and data arrive
according to independent Poisson processes with rates 𝜆𝑉 and
𝜆𝐷 packets/sec respectively. The exact energy function of a
Poisson process is known, [11], giving the energy functions of
voice and data as,
𝛬1 (𝜑) = 𝜆𝑉 (𝑒 𝜑 − 1),
𝛬2 (𝜑) = 𝜆𝐷 (𝑒 𝜑 − 1)
From (3) and (13), the energy functions of the voice and
data departure processes are respectively,
1
𝛤(𝜑) = 𝑙𝑜𝑔(𝑞 + (1 − 𝑞)𝑒 𝜑 ), 𝛤 ∗ (𝜑) = (1 − 𝑞)𝜑(1 + 𝑞𝜑)
2
where 1-q is the success probability. These two results have
been plotted in Fig. 2 as a function of 𝜑. It may be seen that
they are very close to each other, which gives confidence to
the use of asymptotic energy function.
2 2
−𝜑 𝜑 𝜎𝑦
2
+
𝜑 2 𝜎𝑦22
𝜑
3
𝜇
𝛤1 (𝜑) = 𝛤̃1 (𝜑) = 𝜑; 𝛤̃2 (𝜑) =
+
+ 𝜆𝑉 (𝑒 𝑦2 2𝜇𝑦2 − 1)
3
𝜇𝑦2
2𝜇𝑦2
Next, we determine 𝜇𝑦2 , 𝜎𝑦22 which are the mean and
variance of the service time of data packets. The service time
of the data packets has a truncated geometric distribution,
{
𝑃𝑟𝑜𝑏(𝑦2 = 𝑖 𝑠𝑙𝑜𝑡𝑠) = (1 − 𝑝)𝑝𝑖−1 𝑓𝑜𝑟 𝑖 < 𝐿
𝑃𝑟𝑜𝑏(𝑦2 = 𝑖 𝑠𝑙𝑜𝑡𝑠) = 𝑝𝐿−1 𝑓𝑜𝑟 𝑖 = 𝐿
where p is the probability that transmission of a packet will
result in an outage. Let 𝑀𝐷 (𝑧) denote the PGF of the above
service time distribution,
𝑀𝐷 (𝑧) = 𝐸[𝑧 𝑦2 ] = (1 − 𝑝)𝑧
1−(𝑧𝑝)𝐿−1
1−𝑧𝑝
+ 𝑧 𝐿 𝑝𝐿−1
(14)
From the above PGF,
Fig.2. Comparison of asymptotic and exact energy functions of Binomial
process with success probability of 1-q for q=0.1
1−𝑝𝐿
1−𝑝
, 𝜎𝑦22 =
−𝑝2𝐿 −𝑝𝐿+1 (1+2𝐿)+𝑝𝐿 (1−2𝐿)+2𝑝
(1−𝑝)2
Finally, the delay violation probability of data packets is given
by,
∗
1
1
𝜇𝑦2 =
2
2
2
𝑃𝑟𝑜𝑏(𝐷2 > 𝐷𝑡ℎ
) ≈ 𝑒 −Λ2(𝜑 )𝐷𝑡ℎ
∗
where 𝜑 is the unique solution of the equation 𝛬2 (𝜑) +
𝛤̃2 (−𝜑) = 0. The packet loss probability of data packets is
given by,
𝑃𝑟𝑜𝑏(𝑝𝑎𝑐𝑘𝑒𝑡 𝑙𝑜𝑠𝑠) = 𝑝𝐿
2
Fig. 3: A two-state Markov fluid. 𝛾 𝑖 is the transition rate from state i to
state the other state. 𝜋 𝑖 is the probability that the system will be at state
i.
IV. A SYSTEM WITH TWO TYPES OF TRAFFIC FLOWS
Next, we will apply the above results to a system with two
types of traffic flows, which will be referred as voice and data
V. A SYSTEM WITH FOUR TYPES OF TRAFFIC FLOWS
In this section, we will extend the analysis to a system
with four types of traffic flows by adding a multimedia A and
a multimedia B traffic flow into the formal system model.
In the new system, multimedia A is given a lower
priority than the voice traffic but higher than the multimedia B,
while multimedia B is given a higher priority than the data
traffic. We assume multimedia A is a two-state Markov fluid
which is characterized by four parameters: λ1A , λ2A , γ1A and
γ2A . Where λ1A , λ2A refers to the Poisson packet arrival rate at
state 1 and 2, and γ1A , γ2A refers to the transition rate from
state 1 to 2 and from 2 to 1 respectively. Similar thing is for
the multimedia B traffic with suffix B. We also assume that
service of multimedia packets will be the same as the service
of the voice packets.
Let 𝑔2 (𝑡) denote the pdf of the inter-arrival time of the
packets of the 2nd traffic flow, the multimedia A, then, it is
given by,
1
2
𝑔𝑥2 (𝑡) = 𝜋𝐴1 𝜆𝐴1 𝑒 −𝜆𝐴 𝑡 + 𝜋𝐴2 𝜆𝐴2 𝑒 −𝜆𝐴 𝑡
(15)
where 𝜋𝐴1 , 𝜋𝐴2 are the steady-state probabilities of the
two-state Markov source being in states 1 and 2 respectively.
These probabilities are given by 𝜋𝐴1 =
2
𝛾𝐴
1
2
𝛾𝐴 +𝛾𝐴
, 𝜋𝐴2 = 1 − 𝜋𝐴1 .
Next we can determine the mean and variance of 𝑥2 by
obtaining the first and second moment of the Laplace
transform of the above expression. Then we have,
𝜋𝐴1 𝜋𝐴2
𝜇𝑥2 = 1 + 2
𝜆𝐴 𝜆𝐴
2
2
1
1
2
2
𝜎𝑥22 = ( 1 ) (2𝜋𝐴1 − 𝜋𝐴1 ) + ( 2 ) (2𝜋𝐴2 − 𝜋𝐴2 )
𝜆𝐴
𝜆𝐴
Same thing is for the 3rd traffic flow, the multimedia B.
And thus we can have the following equation set:
1
𝜇𝑥1 = ,
𝜇𝑦1 = 1
𝜆𝑉
1
𝜎𝑥21 = 2 , 𝜎𝑦21 = 0
𝜆𝑉
𝜋𝐴1 𝜋𝐴2
𝜇𝑥2 = 1 + 2 ,
𝜇𝑦2 = 1
𝜆𝐴 𝜆𝐴
2
2
2
2
1
1
2
2
𝜎𝑥22 = ( 1 ) (2𝜋𝐴1 − 𝜋𝐴1 ) + ( 2 ) (2𝜋𝐴2 − 𝜋𝐴2 ) , 𝜎𝑦22 = 0
𝜆𝐴
𝜆𝐴
𝜋𝐵1 𝜋𝐵2
𝜇𝑥3 = 1 + 2 ,
𝜇𝑦3 = 1
𝜆𝐵 𝜆𝐵
1
1
1
12
2
22
2
1 ) (2𝜋𝐵 − 𝜋𝐵 ) + ( 2 ) (2𝜋𝐵 − 𝜋𝐵 ) , 𝜎𝑦3 = 0
𝜆𝐵
𝜆𝐵
1
1 − 𝑝𝐿
𝜇𝑥4 =
,
𝜇𝑦4 =
𝜆𝐷
1−𝑝
2𝐿
𝐿+1 (1
1
−𝑝
−
𝑝
+
2𝐿)
+ 𝑝𝐿 (1 − 2𝐿) + 2𝑝
= 2 , 𝜎𝑦24 =
(1 − 𝑝)2
𝜆𝐷
(16)
𝜎𝑥23 = (
{
𝜎𝑥24
Submitting (16) into (13), we could have 𝛤̃4 (𝜑). And
finally the delay violation probability of data packets is given
by,
4
𝑃𝑟𝑜𝑏(𝐷4 > 𝐷𝑡ℎ
) ≈ 𝑒 −𝛬4(𝜑
∗ )𝐷4
𝑡ℎ
where 𝜑 ∗ is the unique solution of the equation
𝛬4 (𝜑) + 𝛤̃4 (−𝜑) = 0.
VI. OPTIMAL DISTRIBUTED ANTENNA PLACEMENT
In this section, we will determine optimal locations of
antennas by minimizing system’s outage probability seen by a
target user. First, we will determine the signal to noise ratio of
the target user, and then we will derive the probability of
system outage.
The distance between user i and the 𝑚𝑡ℎ antenna is given
by,
𝜌𝑚,𝑖 = √(𝑥𝑖 − 𝐿𝑚 𝑐𝑜𝑠(𝛩𝑚 ))2 + (𝑦𝑖 − 𝐿𝑚 𝑠𝑖𝑛(𝛩𝑚 ))2 + 𝒽2 (17)
0 ≤ 𝑖 ≤ 𝐹 − 1 and 1 ≤ 𝑚 ≤ 𝑀.
In the above, 2λ is the path loss exponent. hm,i are
independent identically distributed (i.i.d) complex Gaussian
random variables of Rayleigh fading channels. Since |hm,i |
2
has a Rayleigh distribution, |hm,i | has an exponential
distribution with mean equal to one.
In the following, we treat interferences from other
co-channel users as noises, which are assumed to be the only
source of noise. We consider that ach channel is time slotted
such that the neighbor cell users communicate with in a
synchronous manner. Meanwhile, a user, in a main cell is
imposed by interference of the neighbor-cells users. In this
paper, we mainly focus on the scenarios where user utilizes
the channel used by set of neighbor-cells users. This implies
that channel interference level sensed by the user is a random
variable dependent on the other users’ transmission states. We
model each channel as an ON-OFF source alternating between
state ON (active) and state OFF (inactive). An ON/OFF
channel usage model specifies a time slot in which the
neighbor user signals is or isn’t transmitting over the channel.
When more neighbor users are in the OFF time slot, main user
transmission is affected less interference. Suppose that each
channel changes its state independently. Let α𝑖 be the
probability that the 𝑖𝑡ℎ neighbor user channel exists in state
ON, Then, the channel state can be characterized by a
two-state Markov chain.
Let W denote the received signal strength at an antenna
when a user is transmitting at a distance 1 from that antenna.
Then 𝑆𝑚,𝑖 , the received signal strength at the antenna m from
user i, can be written as,
ì
2
ï W (rm,i )-2 l hm,i
Sm,i = í
ïî 0
with probability a ,0 ≤ 𝑖 ≤ 𝐹 − 1
with probability 1-a
(18)
As a result the instantaneous SNR, Γm , at the antenna m is
given by,
𝑆
𝑚,0
𝛤𝑚 = ∑𝐹−1
𝑖=1 𝑆𝑚,𝑖
=
𝕏𝑚,0
𝕐𝑚
(19)
𝑆𝑚,𝑖
𝕏𝑚,𝑖 =
where,
𝕐𝑚 = ∑𝐹−1
𝑖=1 𝕏𝑚,𝑖
,
𝑊
In general, the outage probability at the antenna m can be
expressed as [13],
𝑃𝑟𝑜𝑏𝑚 (𝑜𝑢𝑡𝑎𝑔𝑒) = 𝑃𝑟𝑜𝑏(𝐼𝑚 < 𝑅)
(20)
where 𝐼𝑚 denotes the mutual information between the
transmitter and receiver antenna m, and R is the required
transmission rate or spectrum efficiency in bits/sec/Hz [6].
From [14], 𝐼𝑚 can be expressed as,
𝐼𝑚 = 𝑙𝑜𝑔 (1 + 𝛤𝑚 )
(21)
where,
1
𝑗
𝑏𝑚,𝑛 = (𝑗−1)!
(22)
ℤ𝑚 = 𝐾𝕐𝑚 − 𝕏𝑚,0
Then, substituting (19) in (22), 𝛲𝑟𝑜𝑏𝑚 (outage) may be
expressed as,
𝛲𝑟𝑜𝑏𝑚 (𝑜𝑢𝑡𝑎𝑔𝑒) = 𝑃𝑟𝑜𝑏[ℤ𝑚 > 0]
(23)
We continue our analysis by setting α = 1 . For Let
𝒻ℤm (x) denote the pdf of the random variable (rv) ℤ𝑚 and
ℤ𝑚 (𝑠) its Laplace transform,
ℤ𝑚 (𝑠) = 𝕐𝑚 (𝐾𝑠)𝕏𝑚,0 (−𝑠) = 𝕐𝑚 (𝐾𝑠)𝕏𝑚,0 (−𝑠)
𝕐𝑚 (𝑠) = ∏𝐹−1
𝑖=1 𝕏𝑚,𝑖 (𝑠)
𝕏𝑚,𝑖 (𝑠) = 𝐸[𝑒 −𝑠𝕏𝑚,𝑖 ] =
𝑁
2𝜆
−𝑠+(𝜌𝑚,0 )
∏𝐹−1
𝑖=1
(25)
𝑠+(𝜌𝑚,𝑖 )
2𝜆
𝛼+1−𝛼 ,
2𝜆
(𝜌𝑚,𝑖 )
𝐾𝑠+(𝜌𝑚,𝑖 )
2𝜆
= 𝛨𝑚 ℱ𝑚 (𝑠)
2𝜆
−𝐹+1
𝛨𝑚 = − ∏𝐹−1
, ℱ𝑚 (𝑠) = ∏𝑁
𝑛=1
𝑖=0 (𝜌𝑚,𝑖 ) (𝐾)
where
𝑢[𝑥] = {
2𝜆
𝐾
𝑏𝑚,𝑛
𝑥 𝑗−1 𝑒 −𝑐𝑚,𝑛 𝑥 𝑢[𝑥]}
(𝑗 − 1)!
0 , 𝑥<0
From (23), the probability of outage of antenna m is given by,
1
𝑁
2𝜆
… , (𝜌𝑚,𝐹−1 ) }
𝐾
(26)
1
𝑘
(𝑠+𝑐𝑚,𝑛 ) 𝑛
where
2𝜆
1 ≤ 𝑛 ≤ 𝑁 with 1 ≤ 𝑁 ≤ 𝐹, and 𝑐𝑚,1 = −(𝜌𝑚,0 ) .
Now, expanding the above in a partial fraction,
𝑘𝑛
𝛲𝑟𝑜𝑏𝑚 (𝑜𝑢𝑡𝑎𝑔𝑒) = ∫ 𝒻ℤ𝑚 (𝑥)𝑑𝑥 = 𝛨𝑚 ∑ ∑
𝑗
𝑏𝑚,𝑛
𝑛=2 𝑗=1 (𝑐𝑚,𝑛 )
0
𝑗
(29)
ℙ(𝑜𝑢𝑡𝑎𝑔𝑒) = ∏𝑀
𝑚=1 𝛲𝑟𝑜𝑏𝑚 (𝑜𝑢𝑡𝑎𝑔𝑒)
(30)
Next, unconditioning the above result wrt the location
vector
of
the
users
(𝓵, 𝜽, 𝟎) = {(ℓ0 , 𝜃0 , 0), (ℓ1 , 𝜃1 , 0), … , (ℓ𝐹−1 , 𝜃𝐹−1 , 0)} , we
have,
1
1
𝛦(ℙ) = ∫ …
∫
2𝜋
2𝜋
ℓ𝐹−1 =0 𝜃0 =0
𝐹−1
∫ ℙ(𝑜𝑢𝑡𝑎𝑔𝑒) ∏ 𝑝(ℓ𝑖 ) 𝑝(𝜃𝑖 )𝑑ℓ𝑖 𝑑𝜃𝑖
∫…
𝑖=0
𝜃𝐹−1 =0
(31)
where 𝑝(ℓ𝑖 ), 𝑝(𝜃𝑖 ) are the marginal pdfs of the coordinates
of user i, ℓ𝑖 and 𝜃𝑖 respectively. Since we are assuming that
user locations are uniformly distributed in a cell we have,
𝑃𝑟𝑜𝑏(ℓ < ℓ𝑖 ) =
have,
ℓ𝑖 2
𝑟2
, and taking into account that r=1, we
p(θi ) =
1
2π
, for 0 ≤ θ < 2π (32)
for
(27)
with 𝑐𝑚,𝑛 , 𝑘𝑛 denoting a distinct value and its frequency in
2𝜆 1
𝑗
𝑛=2 𝑗=1
1 , 𝑥≥0
p(ℓi ) = 2ℓi , for 0 ≤ ℓ ≤ 1,
2𝜆
(𝜌𝑚,𝑖 )
where,
the set {−(𝜌𝑚,0 ) , (𝜌𝑚,1 )
𝑘𝑛
1
𝒻ℤ𝑚 (𝑥) = 𝐻𝑚 { 𝑏𝑚,1
𝑒 −𝑐𝑚,1 𝑥 𝑢[−𝑥] + ∑ ∑
ℓ0 =0
0≤𝑖 ≤𝐹−1
Substituting (25) in (24) gives,
2𝜆
| (𝑠 = −𝑐𝑚,𝑛 )
(24)
where 𝕐𝑚 (𝑠) and 𝕏𝑚,0 (𝑠) are Laplace transform of the rvs
𝕐𝑚 and 𝕏𝑚,0 . From (19), 𝕐m is the sum of F-1 independent
exponentially distributed rvs, we have,
(𝜌𝑚,0 )
𝑑𝑠 𝑗−1
We assume that outages of the antennas are independent of
each other, therefore, the system outage probability,
ℙ(𝑜𝑢𝑡𝑎𝑔𝑒), for a fixed set of user locations is given by,,
Next let us define,
ℤ𝑚 (𝑠) =
𝑑 𝑗−1 [(𝑠+𝑐𝑚,𝑛)𝑗 ℱ𝑚 (𝑠)]
+∞
𝛲𝑟𝑜𝑏𝑚 (𝑜𝑢𝑡𝑎𝑔𝑒) = 𝑃(𝛤𝑚 < 2𝑅 − 1)
where
(28)
𝑗
(𝑠+𝑐𝑚,𝑛 )
A Taking the inverse transform of the above gives the pdf
of the rv ℤ𝑚 ,
Substituting (21) in the (20),
𝐾 = 2𝑅 − 1 ,
𝑗
𝑏𝑚,𝑛
𝑘
𝑛
ℤ𝑚 (𝑠) = 𝐻𝑚 ∑𝑁
𝑛=1 ∑𝑗=1
We define the antenna locations that minimize the expected
outage probability of the system in (31) as being optimal.
Unfortunately, it has not been possible to obtain the minimum
of the expected outage probability of the system analytically
due to complexity of this function. As a result, we have
determined minimum value of the expected outage probability
of the system numerically by applying the Robbins-Monro
stochastic approximation method as in [8].
VII. NUMERICAL RESULTS
In this section, we present some numerical results regarding
the analysis in the paper and simulation results to verify the
accuracy of the analysis.
First, we present numerical and simulation results regarding
the a system with two types of traffics which consist a voice
traffic and a data traffic. It will be assumed that a transmitted
voice packet that can not be decoded will be dropped since
voice is delay sensitive. On the other hand, the failed data
packets will be retransmitted and a data packet will be
dropped only after L unsuccessful transmissions. Dropped
packets will be treated as packet loss. Both voice and data
arrive according to independent Poisson processes with rates
𝜆𝑉 and 𝜆𝐷 packets/sec respectively.
Fig.s 3, 4 present the delay violation probabilities of data
packets as a function of the delay threshold for the data traffic
with λV , λD , p and L as parameters. From Fig. 3. it may be
seen that increasing λV increases the delay violation
probability for a fixed value of λD . Similarly, from Fig. 4,
increasing λD increases the delay violation probability for a
fixed value of λV .
In Fig 5, we plot numerical and simulation results of delay
violation probability to show the accuracy of our analysis. As
may be seen, there is a very good match between numerical
and simulation results. This increases the confidence in the
validity of the analysis, specially, in the use of asymptotic
results from the renewal theory.
Fig. 6 presents the delay violation probabilities of data
packets as a function of the data delay threshold with total
number of transmissions, L, as a parameter. As may be seen,
delay violation probability increases, while packet loss
probability decreases as L increases, which shows the tradeoff
between delay violation and the packet loss probabilities.
Fig. 7,8,9 present the similar delay violation probabilities
of data packets with respect to different parameters in a
system with four types of traffics. These three numerical
results show the same trend as in the system with two types of
traffics.
Fig. 3. Probability of queuing delay being greater than a threshold value
2
𝐷𝑡ℎ
, for 𝜆𝐷 = 0.5 packets/slot, 𝐿 = 4, wrt different values of 𝜆𝑣 and p
2
Fig. 4. Probability of queuing delay being greater than a threshold value 𝐷𝑡ℎ
,
for 𝜆𝑉 = 0.1 packets/slot, 𝑝 = 0.1 wrt to different values of 𝜆𝐷 and L
Fig. 5. Simulation and numerical results of probability of queuing delay being
2
greater than a threshold value 𝐷𝑡ℎ
, for
𝜆𝐷 = 0.6 packets/slot, 𝜆𝑉 = 0.2 packets/slot, 𝑝 = 0.1, 𝐿 = 4
Fig. 6. Probability of queuing delay being greater than a threshold value
2
𝐷𝑡ℎ
, for λD = 0.7 packets/slot, λV = 0.1 packets/slot , 𝑝 = 0.2 wrt
different values of 𝐿
a.
b.
c.
d.
Fig. 7. Probability of queuing delay being greater than a threshold value
4
𝐷𝑡ℎ
, for 𝜆𝑉 = 0.4packets/slot, 𝜆𝐴1 = 0.1packets/slot, 𝜆𝐴2 = 0.2packets/
slot, 𝜋𝐴1 = 0.4, 𝜋𝐴2 = 0.6, 𝜆1B = 0.3packets/slot, 𝜆2𝐵 = 0.2packets/slot,
𝜋𝐵1 = 0.7, 𝜋𝐵2 = 0.3, 𝑝 = 0.1 wrt different values of 𝜆𝐷 and L
Fig. 8. Probability of queuing delay being greater than a threshold
4
value𝐷𝑡ℎ
, for 𝜆𝐴1 = 0.1packets/slot, 𝜆𝐴2 = 0.2packets/slot, 𝜋𝐴1 =
2
0.4, 𝜋𝐴 = 0.6, 𝜆1B = 0.3packets/slot, 𝜆2𝐵 = 0.2packets/slot, 𝜋𝐵1 =
0.7, 𝜋𝐵2 = 0.3, 𝜆𝐷 = 0.4packets/slot, 𝐿 = 4, wrt different values of 𝜆𝑉
and p
Fig. 9. Probability of queuing delay being greater than a threshold value
4
𝐷𝑡ℎ
, for 𝜆𝑉 = 0.2packets/slot, 𝜆1𝐵 = 0.3𝑝𝑎𝑐𝑘𝑒𝑡𝑠/𝑠𝑙𝑜𝑡, 𝜆2𝐵 = 0.2𝑝𝑎𝑐𝑘𝑒𝑡𝑠/
𝑠𝑙𝑜𝑡, 𝜋𝐵1 = 0.7, 𝜋𝐵2 = 0.3, 𝜆𝐷 = 0.2𝑝𝑎𝑐𝑘𝑒𝑡𝑠/𝑠𝑙𝑜𝑡, 𝐿 = 4, 𝑝 = 0.1 wrt
different values of 𝜆𝐴1 , 𝜆𝐴2 , 𝜋𝐴1 and 𝜋𝐴2
Next, we present numerical and simulation results about the
analysis of the antenna placement part in the paper as well as
simulation results. The numerical results for the optimal
location of the antennas have been obtained through the
application of the Robbins-Monro algorithm [15], the steps of
which are given below,
Initialize the antennas location vector 𝑳 and let this
vector to be 𝑳(𝟏)
Randomly generate a user location vector (𝓵, 𝛉, 𝟎)
according to their uniform distributions in the cells.
Update the optimum locations of the antennas by applying
the equation below
𝜕ℙ(𝑜𝑢𝑡𝑎𝑔𝑒)
𝑳(𝒏 + 𝟏) = 𝑳(𝒏) + 𝑐𝑛 (
)| 𝑳 = 𝑳(𝒏)
𝜕𝑳
Let n=n+1 and iteratively run steps b, c, d until ̅̅̅̅̅̅
𝑳(𝒏)
1
𝑛−1
̅̅̅̅̅̅
∑
converges,
where
𝑳(𝒏) =
𝑳(𝒊) . ̅̅̅̅̅̅
𝑳(𝒏)
𝑛−1 𝑖=1
converges to the optimum antenna location vector 𝑳′ .
In this implementation of the Robbins-Monro algorithm, we
chose the sequence of positive 𝑐𝑛 as 𝑐𝑛 = 15𝑛−0.75 , which
satisfies the Polyak and Juditsky constraints in [16].
We compared delay performance of the proposed algorithm
with the case algorithm, where bandwidth slots are distributed
between the data and voice traffics with different ratios in the
case algorithm. Considering a TDMA system, the entire slot
has 20 sub-slots and each sub-slot has a frame for potential
voice or data transmission. The entire slot has both data and
voice frames. The different ratios (Figures 12, 13) represent
number of frames of data within the entire slot. Let r represent
this ratio. When there is no data or voice arrival, the TDMA
system will transmit empty frames. Using the case algorithm,
in spite of sensitivity to the data delay (as is shown in Figure
12), voice delay increases significantly (as is shown in Figure
13). However, the proposed algorithm has the lowest voice
delay and doesn’t have much data delay sensitivity.
Also, to reduce the amount of computations for the
execution of the algorithm, we assumed that the optimal
locations of the antennas will be a circle centered at the cell
center and the antennas will be spaced evenly on the circle.
The reason for this assumption is that the cells are symmetric
in every aspect. The number of antennas is assumed to be M=4
and the cluster size to be F=7, R=1 bits/sec/Hz. In figures 5-10
wireless channel parameters are set as follows: 𝒽 = 0.05
and path loss exponent 2λ = 4.
The application of the
Robbins-Monro procedure gives the optimum locations of
π
π
antennas as (𝐿𝑚 , 𝛩𝑚 , 𝒽)=(0.58, 𝑚 − ,0.05), for m=1,…4,
2
2
which results in the minimum expected value of outage
probability of, 𝛦(ℙ) = 0.0087, see Table 1.
In fig 11-13 wireless channel parameters are set as follows :
𝒽 = 0.05 and path loss exponent 2λ = 2 . In this case,
Robbins-Monro procedure gives the optimum locations of
π
π
antennas as (𝐿𝑚 , 𝛩𝑚 , 𝒽) = (0.416, 𝑚 − , 0. 05), for
2
2
m=1,…4, which results in the minimum expected value of
outage probability of, 𝛦(ℙ) = 0.1081, see Table 2.
In Fig. 10, we plotted simulation results for the expected
value of the system outage probability as a function for the
optimal radius of the antenna circle for different values of
neighbor users’ activation probability. From the figure it
may be seen that there is an optimal radius of the antenna
circle, 0.42, which results (for α = 1 ) in the minimum
expected value of outage probability of Ε(ℙ) = 0.106
Comparing to the case that all the antennas are at the center,
the optimal location provides almost 33% improvements in
system’s average outage probability. Simulation and
numerical results show a very good match.
Fig. 10: Simulation result for system’s average outage probability 𝛦(ℙ) as a
function of the first antenna’s location. (different curves represents for different
π 2π
8π
value of 𝛩1 = (0 , …,
)
18
18
18
𝑛
̅̅̅
𝐿1
𝑛
̅̅̅
𝐿1
1
0.00
2
0.90
0.013
35
0.55
0.0095
0.091
36
0.55
3
0.63
0.0094
0.011
37
0.55
4
0.0094
0.54
0.009
38
0.55
0.0094
5
0.50
0.0099
39
0.55
0.0094
6
0.47
0.010
40
0.55
0.0094
7
0.45
0.010
41
0.55
0.0094
8
0.46
0.010
42
0.56
0.0094
9
0.48
0.010
43
0.56
0.0094
10
0.49
0.010
44
0.56
0.0093
11
0.50
0.0099
45
0.56
0.0093
12
0.50
0.0099
46
0.56
0.0093
13
0.50
0.0099
47
0.56
0.0093
14
0.51
0.0098
48
0.56
0.0093
15
0.51
0.0097
49
0.57
0.0093
16
0.51
0.0097
50
0.57
0.0093
17
0.51
0.0096
51
0.57
0.0093
18
0.51
0.0096
52
0.57
0.0092
19
0.52
0.0096
53
0.57
0.0092
20
0.52
0.0096
54
0.57
0.0092
21
0.52
0.0095
55
0.57
0.0092
22
0.52
0.0095
56
0.57
0.0092
23
0.53
0.0095
57
0.58
0.0091
24
0.53
0.0095
58
0.58
0.0091
25
0.53
0.0095
59
0.58
0.0091
26
0.54
0.0095
60
0.58
0.0090
27
0.54
0.0095
61
0.58
0.0090
28
0.54
0.0095
62
0.58
0.0090
29
0.54
0.0095
63
0.58
0.0090
30
0.54
0.0095
64
0.58
0.0090
31
0.54
0.0095
65
0.58
0.0090
32
0.54
0.0095
66
0.58
0.0090
33
0.54
0.0095
67
0.58
0.0090
34
0.54
0.0095
68
0.58
0.0090
E(p)
E(p)
Table 1. Robbins-Monro outcome of the optimum location of the
first antenna as. ( here we only show the optimum antenna radius
𝐿1 (𝑛) since angles 𝛩1 (𝑛) changes slightly around 0 as n increases.)
Fig. 11: Probability of queuing delay being greater than a threshold value
𝐷𝑡ℎ
,
for 𝜆𝑉 = 0.2packets/slot, 𝜆𝐴1 = 0.2𝑝𝑎𝑐𝑘𝑒𝑡𝑠/𝑠𝑙𝑜𝑡, 𝜆𝐴2 =
0.3𝑝𝑎𝑐𝑘𝑒𝑡𝑠/𝑠𝑙𝑜𝑡, 𝜋𝐴1 = 0.8, 𝜋𝐴2 = 0.2, 𝜆1𝐵 = 0.3𝑝𝑎𝑐𝑘𝑒𝑡𝑠/𝑠𝑙𝑜𝑡, 𝜆2𝐵 =
0.2𝑝𝑎𝑐𝑘𝑒𝑡𝑠/𝑠𝑙𝑜𝑡, 𝜋𝐵1 = 0.7, 𝜋𝐵2 = 0.3, 𝜆𝐷 = 0.2𝑝𝑎𝑐𝑘𝑒𝑡𝑠/𝑠𝑙𝑜𝑡, 𝐿 =
4, 𝑝 = 0.1
𝑛
̅̅̅
𝐿1
E(p)
𝑛
̅̅̅
𝐿1
E(p)
𝑛
̅̅̅
𝐿1
E(p)
1
0. 9
0.293
18
0.37
0.114
35
0.40
0.11
2
0.89
0.293
19
0.37
0.114
36
0.40
0.11
3
0.71
0.173
20
0.37
0.113
37
0.40
0.11
4
0.54
0.119
21
0.37
0.113
38
0.40
0.11
5
0.44
0.109
22
0.38
0.113
39
0.40
0.11
6
0.35
0.115
23
0.38
0.113
40
0.40
0.109
7
0.34
0.115
24
0.38
0.113
41
0.40
0.109
8
0.35
0.115
25
0.38
0.113
42
0.40
0.109
9
0.35
0.115
26
0.38
0.113
43
0.40
0.109
10
0.35
0.115
27
0.38
0.113
44
0.41
0.108
11
0.36
0.114
28
0.39
0.113
45
0.41
0.108
12
0.36
0.114
29
0.39
0.113
46
0.41
0.108
13
0.36
0.114
30
0.39
0.113
47
0.41
0.108
14
0.36
0.114
31
0.39
0.113
48
0.41
0.108
15
0.36
0.114
32
0.39
0.112
49
0.41
0.108
16
0.37
0.114
33
0.39
0.112
50
0.41
0.108
17
0.37
0.114
34
0.39
0.112
51
0.41
0.108
Table 2. Robbins-Monro outcome of the optimum location of the first
antenna as. ( here we only show the optimum antenna radius 𝐿1 (𝑛) since
angles 𝛩1 (𝑛) changes slightly around 0 as n increases.)
Fig. 12: Simulation result for system’s average outage probability 𝛦(ℙ)
as a function of the first antenna’s location. (different curves represents
for different value of α= 1, 0.8, 0.5)
Fig. 14: Probability of voice queuing delay being greater than a threshold
1
value 𝐷𝑡ℎ
, for 𝜆𝐷 = 0.5 packets/slot, 𝑝 = 0.1
VIII. CONCLUSION
Presently, distributed broadband wireless communication
(BWC) system is a very promising wireless technology due to
its low cost and high performance. This paper considers uplink
Fig. 13: Probability of data queuing delay being greater than a threshold communications with multiple levels of priority traffic having
2
value 𝐷𝑡ℎ
, for 𝜆𝑉 = 0.1 packets/slot, 𝑝 = 0.1 𝑎𝑛𝑑 𝑟 = 0.85
any renewal arrival and departure processes. We develop an
analysis that determines packet delay violation probability for
each priority level as a function of the outage probability of
the distributed antenna system through the application of
results from the renewal theory. Then, we determine the
optimal locations of the antennas that minimize the antenna
outage probability. We also present simulation results that
show the accuracy of the analysis. The results are very easy to
use and they will be useful in the design of BWC systems.
Finally, the packet delay violation probability result may also
be applicable in any other communication systems with
multiple types of priority traffic.
REFERENCES
[1]
[2]
[3]
[4]
[5]
[6]
[7]
[8]
[9]
[10]
[11]
[12]
[13]
[14]
[15]
[16]
[17]
[18]
[19]
[20]
[21]
[22]
W. Mohr, J.F. Monserrat, A. Osseiran, M. Werner “IMT-advanced
and next-generation mobile networks [Guest Editorial]” , IEEE
Communications Magazine, Vol. 49, Issue 2, pp. 82 – 83, 2011.
J. Z. Wang, F. Adachi, A. Gameiro, N.J. Gomes, K.I. Kitayama, P.
Monteiro, M. Sawahashi, X.G. Xia “Guest Editorial: Distributed
broadband wireless communications”, IEEE Journal on Selected Areas
in Communications, Vol. 29 , No. 6, pp. 1121 – 1122, 2011.
J.W. Lee, D. Toumpakaris, Yu Wei “Interference mitigation via joint
detection” , IEEE Journal on Selected Areas in Communications, Vol.
29 , Issue 6, pp: 1172 – 1184, 2011.
J.S. Gao, Y.Z. Li, S.D. Zhou, J. Wang “On sum rate of multi-user
distributed antenna system with circular antenna layout”, Proceedings of
the IEEE Vehicular Technology Conference , VTC-66, pp: 596 – 600,
Fall 2007.
Q.H Du, X. Zhang “QoS-aware base-station selections for distributed
MIMO links in broadband wireless networks”, IEEE Journal on
Selected Areas in Communications, Vol. 29, No.6, pp.1123 – 1138,
2011.
D. Zennaro, S. Tomasin, L. Vangelista “Base station selection in uplink
macro diversity cellular Systems with hybrid ARQ” , IEEE Journal on
Selected Areas in Communications, Vol. 29, Issue 6, pp: 1249 – 1259,
2011.
W. Feng, X.B. Xu, S.D. Zhou, J. Wang, M.H. Xia “Sum rate
characterization of distributed antenna systems with circular antenna
layout” Proceedings of the IEEE Vehicular Technology Conference,
VTC-69, pp: 1 – 5, Spring 2009.
S. Firouzabadi, A. Goldsmith “ Optimal placement of distributed
antennas in cellular systems”, Proceedings of the IEEE International
Workshop on Signal Processing Advances in Wireless Communications,
SPAWC-12, pp. 461 – 465, 2011.
J.-P.M.G Linnartz “Exact analysis of the outage probability in
multiple-user mobile radios”, IEEE Transactions on Communications,
Vol. 40 , Issue 1, pp. 20 – 23, 1992.
C.S. Chang , J.A. Thomas “Effective bandwidth in high-speed digital
networks”, IEEE Journal on Selected Areas in Communications, Vol.
13, No. 6, pp. 1091 – 1100, 1995.
P. Rabinovitch, D. De Vleeschauwer “What is Effective Bandwidth”
Available http://www3.sympatico.ca/peter.rabinovitch/eb.pdf
W. Feller, An Introduction to Probability Theory and Its Applications,
Volume 2, Chapter Ⅺ,John Wiley & Sons, Inc.
Y. Zhao, R. Adve, T.J. Lim “Outage probability at arbitrary SNR with
cooperative diversity networks” IEEE Communications Letters, Vol.9 ,
Issue 8, pp.: 700 – 702, 2005.
T.M. Cover, J.A.Thomas, Elemenst of Information Theory, John
Wiley,1991
H.J. Kushner and D.S. Clark “Stochastic Approximation Methods for
Constrained and Unconstrained Systems”, Chapter 1, Spinger-Verlag,
New York, Heidelberg, Berlin
B. T. Polyak and A. B. Juditsky “Acceleration of Stochastic
Approximation by Averaging” SIAM Jorunal of control and
optimization Vol. 30, No. 4, pp. 838-855, July 1992
Shoja M, Taheri H, Vakilinia S. A new approach to prevent black hole
attack in AODV. International Journal of Computer Science and
Information Security. 2011 Jan;9(1):24-9.
Kiskani MK, Khalaj BH, Vakilinia S. Delay qos provisioning in
cognitive radio systems using adaptive modulation. InProceedings of the
6th ACM workshop on QoS and security for wireless and mobile
networks 2010 Oct 20 (pp. 49-54). ACM.
Vakilinia S, Khalaj BH, Froushani MH. Qos-aware water-filling for
real-time transmission in ofdma systems. InTelecommunications (ICT),
2011 18th International Conference on 2011 May 8 (pp. 84-89). IEEE.
Rastegar, S.H., Vakilinia, S. and Khalaj, B.H., 2012, June. Rate
adaptation and power allocation for time-correlated MISO Rayleigh
fading channel with delay-limited HARQ. In Communications (ICC),
2012 IEEE International Conference on (pp. 3915-3919). IEEE.
Hosseini M, Lee B, Vakilinia S. Energy performance of cool roofs under
the impact of actual weather data. Energy and Buildings. 2017 Jun
15;145:284-92.
Vakilinia S, Alvandi M, Shoja MK, Vakilinia I. Cross-Layered Secure
and QoS Aware Design of VOIP over Wireless Ad-Hoc Networks.
International Journal of Business Data Communications and Networking
(IJBDCN). 2013 Oct 1;9(4):23-45.
[23] Vakilinia, Shahin, Xinyao Zhang, and Dongyu Qiu. "Analysis and
optimization of big-data stream processing." In Global Communications
Conference (GLOBECOM), 2016 IEEE, pp. 1-6. IEEE, 2016.
[24] Vakilinia, Shahin, and Iman Vakilinia. "QoS aware energy efficient
resource allocation in wireless cooperative OFDMA relay networks." In
Wireless and Mobile Networking Conference (WMNC), 2013 6th Joint
IFIP, pp. 1-4. IEEE, 2013.
[25] Vakilinia, Shahin, Behdad Heidarpour, and Mohamed Cheriet. "Energy
Efficient Resource Allocation in Cloud Computing Environments."
IEEE Access 4 (2016): 8544-8557.
[26] Zemmouri, Samy, Shahin Vakilinia, and Mohamed Cheriet. "Let's adapt
to network change: Towards energy saving with rate adaptation in
SDN." In Network and Service Management (CNSM), 2016 12th
International Conference on, pp. 272-276. IEEE, 2016.
[27] Vakilinia, S., Cheriet, M., & Rajkumar, J. (2016, October). Dynamic
resource allocation of smart home workloads in the cloud. In Network
and Service Management (CNSM), 2016 12th International Conference
on (pp. 367-370). IEEE.
[28] Heidarpour, Behdad, Zbigniew Dziong, and Shahin Vakilinia.
"Volume-based pricing in Stackelberg duopoly wireless markets."
Telecommunications Network Strategy and Planning Symposium
(Networks), 2016 17th International. IEEE, 2016.
| 1cs.CV
|
Optimal locally repairable codes of distance 3 and 4 via cyclic codes
arXiv:1801.03623v1 [cs.IT] 11 Jan 2018
Yuan Luo∗, Chaoping Xing and Chen Yuan †
Abstract
Like classical block codes, a locally repairable code also obeys the Singleton-type bound (we call
a locally repairable code optimal if it achieves the Singleton-type bound). In the breakthrough work of
[14], several classes of optimal locally repairable codes were constructed via subcodes of Reed-Solomon
codes. Thus, the lengths of the codes given in [14] are upper bounded by the code alphabet size q.
Recently, it was proved through extension of construction in [14] that length of q-ary optimal locally
repairable codes can be q + 1 in [7]. Surprisingly, [2] presented a few examples of q-ary optimal locally
repairable codes of small distance and locality with code length achieving roughly q 2 . Very recently,
it was further shown in [8] that there exist q-ary optimal locally repairable codes with length bigger
than q + 1 and distance propositional to n. Thus, it becomes an interesting and challenging problem to
construct new families of q-ary optimal locally repairable codes of length bigger than q + 1.
In this paper, we construct a class of optimal locally repairable codes of distance 3 and 4 with unbounded length (i.e., length of the codes is independent of the code alphabet size). Our technique is
through cyclic codes with particular generator and parity-check polynomials that are carefully chosen.
1 Introduction
Due to applications to distributed storage systems, locally repairable codes have recently attracted great
attention of researchers [5, 4, 12, 13, 7, 8, 3, 11, 14, 15, 1]. A local repairable code is nothing but a
block code with an additional parameter called locality. For a locally repairable code C of length n with
k information symbols and locality r (see the definition of locally repairable codes in Section 2.1), it was
proved in [4] that the minimum distance d(C) of C is upper bounded by
k
d(C) 6 n − k −
+ 2.
(1)
r
The bound (1) is called the Singleton-type bound for locally repairable codes and was proved by extending
the arguments in the proof of the classical Singleton bound on codes. In this paper, we refer an optimal
locally repairable code to a block code achieving the bound (1).
1.1 Known results
The early constructions of optimal locally repairable codes gave codes with alphabet size that is exponential
in code length (see [6, 13]. There was also an earlier construction of optimal locally repairable codes given
∗
Department of Computer Sciences and Engineering, Shanghai Jiaotong University, Shanghai 200240, P. R. China. (email:
[email protected])
†
The authors are with Division of Mathematical Sciences, School of Physical and Mathematical Sciences, Nanyang Technological University, Singapore 637371, Republic of Singapore (email: {xingcp,YUAN0064}@ntu.edu.sg).
1
in [12] with alphabet size comparable to
length. However, the construction in [12] only produced a
kcode
specific value of the length n, i.e., n = r (r + 1). Thus, the rate of the code is very close to 1. There are
also some existence results given in [12] and [14] with less restriction on locality r. But the results in both
the papers require large alphabet size which is an exponential function of the code length.
A recent breakthrough construction given in [14] makes use of subcodes of Reed-Solomon codes. This
construction produces optimal locally repairable codes with length linear in alphabet size although the length
of codes is upper bounded by alphabet size. The construction in [14] was extended via automorphism group
of rational function fields by Jin, Ma and Xing [7] and it turns out that there are more flexibility on locality
and the code length can go up to q + 1, where q is the alphabet size.
Based on the classical MDS conjecture, one should wonder if q-ary optimal locally repairable codes
can have length bigger than q + 1. Surprisingly, it was shown in [2] that there exist q-ary optimal locally
repairable codes of length exceeding q + 1. Although [2] produced a few optimal locally repairable codes
with specific parameters, it paves a road for people to continue search for such optimal codes. Very recently,
it was shown in [8] via elliptic curves that there exist q-ary optimal locally repairable codes with length n
bigger than q + 1 and distance proportional to length n.
1.2 Our result
In this paper, by carefully choosing generator polynomials, we can show that there exist optimal locally
repairable codes of distance 3 and 4 that are cyclic. One feature of these codes is that their lengths are
unbounded, i.e., lengths are independent of the code alphabet sizes. More precisely, we have the following
main result in this paper.
Theorem 1.1. Let q be a prime power. Assume that n is a positive integer with gcd(n, q) = 1. Then
i
h
n
, 3 cyclic locally repairable code with locality r if
(i) there is a q-ary optimal n, k = n − 1 + r+1
r > 2 and gcd(n, q − 1) ≡ 0 (mod r + 1); and
i
h
n
, 4 cyclic locally repairable code with locality r if
(ii) there is a q-ary optimal n, k = n − 2 + r+1
n
r > 3, gcd(n, q − 1) ≡ 0 (mod r + 1) and gcd r+1
, r + 1 divides 2.
Remark 1.
(i) The conditions in Theorem 1.1 are easy to be satisfied and the length n is unbounded.
(ii) Indeed, parameters given in Theorem 1.1 satisfy the equality of (1). Let us verify this only for part (i)
of Theorem 1.1. We have
n
n − 1 − r+1
n
1
n
n
k
−
−
−
+2=1+
+2 =3+
= 3.
n−k−
r
r+1
r
r+1
r+1 r
1.3 Open problems
In view of the known results and our result in this paper, all known optimal q-ary locally repairable codes
with length n and minimum distance satisfy either (i) d is small and n is unbounded; or (ii) d is proportional
to n and q is linear in n. One natural question is
Open problem 1. Are there optimal q-ary locally repairable codes with length n much bigger than q (for
instance n = Ω(q 2 )) and minimum distance proportional to n?
2
The other open problem is the following.
Open problem 2. Are there optimal q-ary locally repairable codes with unbounded length n for every constant d.
1.4 Organization of the paper
The paper is organized as follows. In Section 2, we provide some preliminaries on locally repairable codes
and cyclic codes. In Section 3, we present proof of Theorem 1.1. In addition, we also give a construction of
optimal q-ary locally repairable codes of length 2(q − 1) and distance 4.
2 Preliminaries
In this section, we present some preliminaries on locally repairable codes and cyclic codes.
2.1 Locally repairable codes
Informally speaking, a block code is said with locality r if every coordinate of a given codeword can be
recovered by accessing at most r other coordinates of this codeword. Precisely speaking, a locally repairable
code with locality r is given as follows.
Definition 1. Let C ⊆ Fnq be a q-ary block code of length n. For each α ∈ Fq and i ∈ {1, 2, · · · , n},
define C(i, α) := {c = (c1 , . . . , cn ) ∈ C | ci = α}. For a subset I ⊆ {1, 2, · · · , n} \ {i}, we denote by
CI (i, α) the projection of C(i, α) on I. Then C is called a locally repairable code with locality r if, for
every i ∈ {1, 2, · · · , n}, there exists a subset Ii ⊆ {1, 2, · · · , n} \ {i} with |Ii | 6 r such that CIi (i, α) and
CIi (i, β) are disjoint for any α 6= β ∈ Fq .
Apart from the usual parameters: length, rate and minimum distance, the locality of a locally repairable
code plays a crucial role. In this paper, we always consider locally repairable codes that are linear over Fq .
Thus, a q-ary locally repairable code of length n, dimension k, minimum distance d and locality r is said to
be an [n, k, d]q -locally repairable code with locality r.
2.2 Cyclic codes
Cyclic codes are well known and understood in the coding community. We briefly introduce them and list
some useful facts on cyclic codes in this subsection. A q-ary cyclic code C is identified with an ideal of
the ring Fq [x]/(xn − 1). As every ideal of Fq [x]/(xn − 1) is a principal ideal generated by a divisor g(x)
of xn − 1, we can just write C = hg(x)i. The dimension of C is n − deg(g(x)). Furthermore, c(x) is a
codeword of C if and only if c(x) is divisible by g(x). In other words, if α1 , . . . , αn−k ∈ F̄q are all n − k
roots of g(x) (where F̄q stands for algebraic closure of Fq ), then c(x) belongs to C if and only if c(αi ) = 0
for all i = 1, 2, . . . , n − k.
n −1
. Let h̃(x) be the reciprocal polynomial of
Let C = hg(x)i for a divisor g(x) of xn − 1. Put h(x) = xg(x)
h(x). Then the Euclidean dual code of C is C ⊥ = hh̃(x)i. We recall some facts about cyclic codes without
giving proof. The reader may refer to the books [10, 9].
Lemma 2.1.
(i) Let h(x) be a divisor of xn − 1, then the codes hh(x)i and hh̃(x)i are equivalent.
3
(ii) Let C = hg(x)i a q-ary cyclic code of length n for a divisorP
g(x) of xn −1 and assume that α1 , . . . , αt
are roots of g(x). If C has a nonzero codeword c(x) = tj=1 cij xij of weight at most t, then the
determinant
i
det αℓj
16ℓ6t,16j6t
is zero.
(iii) Let C = hg(x)i be a q-ary cyclic code of length n for a divisor g(x) of xn − 1. Let β ∈ F̄q be an nth
primitive root of unity. If there exist an integer t and a positive integer δ such that β t , β t+1 , . . . , β t+δ−2
are roots of g(x), then C has minimum distance at least δ.
Let Sn denote the symmetric group of length n. An automorphism of a q-ary block code C is a permutation σ ∈ Sn satisfying that (cσ(1) , cσ(2) , . . . , cσ(n) ) ∈ C whenever (c1 , c2 , . . . , cn ) ∈ C. All automorphisms
C form a subgroup of Sn , denoted by Aut(C). It is called the automorphism group of C. The code C
is called transitive if, for any i, j ∈ {1, 2, . . . , n}, there exists an automorphism σ ∈ Aut(C) such that
σ(i) = j.
If C is the cyclic code, then Aut(C) contains the subgroup generated by the cyclic shift (12 · · · n). Thus,
C is transitive.
3 Constructions
In this section, we first give a general result on locality for transitive codes and then apply it to the proof of
Theorem 1.1. In addition, we also present a construction of q-ary [2(q − 1), 4] locally repairable code with
locality r.
3.1 A general result
Lemma 3.1. If C is a q-ary transitive linear code with dual distance d⊥ 6 r + 1, then C has locality r.
Proof. Let a = (a1 , a2 , . . . , an ) be a codeword of C ⊥ with the Hamming weight d⊥ . Let I denote the
support of a. Then |I| = d⊥ 6 r + 1. Let J be a subset of {1, 2, . . . , n} such that |J| = r + P
1 and I ⊆ J.
Choose an element i ∈ I. Then for every codeword c = (c1 , c2 , . . . , cn ) ∈ C, we have 0 = j∈I aj cj =
P
−1 P
j∈J aj cj . This gives ci = −ai
j∈J\{i} aj cj . This implies that ci can be repaired by {cj }j∈J\{i} . Now
for any ℓ ∈ {1, 2, . . . , n}, there exists an automorphism σ ∈ Sn such that σ(i) = ℓ. This implies that
X
X
0=
aj cσ(j) =
aj cσ(j) .
j∈I
Hence, cℓ = cσ(i) = −a−1
i
P
j∈J\{i} aj cσ(j) .
j∈J
The proof is completed.
Example 3.2. Let q be a prime power. Let r, n be positive integers with r > 2, n|(q − 1) and (r + 1)|n. We
present a q-ary optimal cyclic locally repairable code of length n minimum distance d and locality r for any
1 6 d 6 n. Although such a code was already given in [14], we provide different view for this code. Such
an observation finally leads to discovery of optimal locally repairable codes with unbounded length given
in Theorem 1.1. Let β be an nth primitive root of unity. For a positive integer d with 1 6 d 6 n, d can be
uniquely written as d = (r + 1)a + b for some integer a > 0 and integer b ∈ {0, 1, . . . , r}.
4
Case 1. b ∈ {2, 3, 4, . . . , r}.
Let
g(x) =
n
−1
r+1
d−2
Y
i=0
g(x)|(xn
Y
(x − β i ) ×
(x − β (r+1)j+b−2 ).
j=a+1
It is obvious that
− 1) since all roots of g(x) are the nth roots of unity and they are distinct. We
denote by C the cyclic code with generator polynomial g(x). Due to the fact that g(x) contains the d − 1
roots 1, β, . . . , β d−2 , it follows from Lemma 2.1(i) that C has the minimum distance at least d. Moreover,
rn
n
− (d − a − 2) = r+1
− ar − b + 2. Thus, we have
the dimension of C is k := n − deg(g(x)) = n − r+1
k
n
−b + 2
n
n−k−
+ ar + b − 2 −
−a+
+2=
+ 2 = a(r + 1) + b = d.
r
r+1
r+1
r
Thus, to show that C satisfies the Singleton-type bound (1), it remains to show that the locality of C is r.
By Lemma 3.1, it is sufficient to the dual distance of C is at most r + 1. By Lemma 2.1(i), it is sufficient to
n −1
. Observe that
show that the cyclic code hh(x)i has minimum distance at most r + 1, where h(x) = xg(x)
n
−1
r+1
Y
n
(x − β (r+1)j+b−2 ) = x r+1 − β
(b−2)n
r+1
.
j=0
Thus, we write
n
g(x) = (x r+1 − β
(b−2)n
r+1
d−2
Y
)
(x − β i ).
i=0,i6=b−2 mod r+1
Then
h(x)
d−2
Y
xn − 1
(x − β i ) =
x
i=0,i6=b−2 mod r+1
n
r+1
−β
(b−2)n
r+1
=
r
X
β
(b−2)in
r+1
x
(r−i)n
r+1
i=0
is a codeword of hh(x)i and it has Hamming weight r + 1. This implies that C ⊥ has minimum distance at
most r + 1.
Case 2. b ∈ {0, 1}.
In this case, we define the cyclic code C generated by
g(x) =
d−2
Y
(x − β i ) ×
i=0
n
r+1
Y
(x − β (r+1)j+b−2 ).
j=a+1
Compared with Case 1, the degree of g(x) increases by 1 due to fact that b − 2 is negative. We can mimic
the proof of Case 1 and skip the detail.
Example 3.3. Let q be a prime power. Let r, n be positive integers with r > 2, n|(q + 1) and (r + 1)|n.
We present a q-ary optimal cyclic locally repairable code of length n, locality r and minimum distance
d = a(r + 1) + b with 2|a and b ∈ {2, 4, 6, . . . , 2⌈ r−1
2 ⌉}. The codes in this example were founded in [7]
already, but we provide a new view on the construction.
2
Since n|(q + 1), we have (xn − 1)|(xq −1 − 1). This implies that each irreducible factor of xn − 1 is
either linear or quadratic. If an irreducible factor f (x) of xn − 1 is quadratic and γ is a root of f (x), then
5
the other root is γ q . As n|(q + 1) and γ n = 1, we have γ q+1 = 1. This gives γ q = γ −1 . In conclusion, for
each nth root of unity γ, (x − γ)(x − γ −1 ) is a polynomial over Fq .
Let β be an nth primitive root of unity. Let
n
− a+2
r+1
2
d−2
g(x) =
2
Y
Y
(x − β i ) ×
(x − β (r+1)j ).
j= a+2
2
i=− d−2
2
d−2
We first show that g(x) is defined over Fq . If β i is a root of g(x), then either − d−2
2 6 i 6 2 or (r + 1)|i.
This implies that both β i and β −i are the roots of g(x). Thus, g(x) is the product of x − 1 with some
polynomials of the form (x − β i )(x − β −i ). We conclude that g(x) is defined over Fq . It is clear that
g(x)|(xn − 1) since all roots of g(x) are nth roots of unity and they are distinct. Denote by C the cyclic
d−2
d−2
code with generator polynomial g(x). Due to the fact that g(x) contains the d − 1 roots β − 2 , . . . , β 2 ,
it follows from Lemma 2.1(i) that C has the minimum distance at least d. Moreover, the dimension of C is
n
rn
k := n − deg(g(x)) = n − r+1
− (d − a − 2) = r+1
− ar − b + 2. Thus, we have
n−k−
n
−b + 2
n
k
+ ar + b − 2 −
−a+
+2=
+ 2 = a(r + 1) + b = d.
r
r+1
r+1
r
To show that C satisfies the Singleton-type bound (1), it remains to show that the locality of C is r. By
Lemma 3.1, it is sufficient to show that the dual distance of C is at most r + 1. By Lemma 2.1(i), it is
n −1
.
sufficient to show that the cyclic code hh(x)i has minimum distance at most r + 1, where h(x) = xg(x)
Observe that
d−2
2
Y
n
g(x) = (x r+1 − 1)
(x − β i ).
i= 2−d
,(r+1)∤i
2
Then
d−2
h(x)
2
Y
(x − β i ) =
,(r+1)∤i
i= 2−d
2
xn − 1
x
n
r+1
−1
=
r
X
x
(r−i)n
r+1
i=0
is a codeword of hh(x)i and it has Hamming weight r + 1. This implies that C ⊥ has minimum distance at
most r + 1.
3.2
Proof of part (i) of Theorem 1.1
.
q−1
n
Proof. Let β ∈ F̄q be an nth primitive root of unity. Put α = β r+1 . Then α 6= 1 and αq−1 = (β n ) r+1 = 1,
n
i.e., α is an element of Fq \ {0, 1}. Let g(x) = (x − 1)(x r+1 − α).
Note that
!
r
X
n
in
r−i
(x r+1 − α)
α x r+1 = xn − αr+1 = xn − 1.
i=0
n
r+1
n
This means that x
− α is a factor of xn − 1. Furthermore, since gcd(x − 1, x r+1 − α) = 1, we have
n
g(x)|(x − 1). Let C be the cyclic code generated by g(x).
6
Let us first show that the locality of C is r. By Lemma 3.1, it is sufficient to show that the Hamming
n −1
distance of C ⊥ is at most r + 1. Put h(x) = xg(x)
. By Lemma 2.1(i), it is sufficient to show that the
Hamming distance of hh(x)i is at most r + 1 since hh(x)i and C ⊥ are equivalent. Consider the codeword
in
P
(x − 1)h(x) of hh(x)i. As (x − 1)h(x) = ri=0 αr−i x r+1 , the Hamming weight of (x − 1)h(x) is r + 1.
Hence, the Hamming distance of hh(x)i is upper bounded by r + 1.
Finally, as 1, β are roots of g(x), the minimum distance is at least 3 by Lemma 2.1(iii). By the bound
(1), the minimum distance of C is upper bounded by 3. Thus, the minimum distance of C is exactly 3. This
completes the proof.
3.3
Proof of part (ii) of Theorem 1.1
.
q−1
n
Proof. Let β ∈ F̄q be an nth primitive root of unity. Put α = β r+1 . Then α 6= 1 and αq−1 = (β n ) r+1 =
1, i.e.,
α is an element of Fq \ {0, 1}. Furthermore, α is a (r + 1)th primitive root of unity. Since
gcd
Then
n
r+1 , r
+ 1 divides 2, there exist integers a, b such that a ×
n
n
r+1
+ b(r + 1) = 2. Put γ = αa ∈ Fq .
an
γ r+1 = α r+1 = α2−b(r+1) = α2 .
As r > 3, we must have γ ∈ Fq \ {0, 1}. Since γ is a (r + 1)th root of unity, the linear polynomial (x − γ)
n
n
2 − α 6= 0, i.e, x − γ is not a factor of x r+1 − α. This
−
α
=
α
divides xn − 1. Furthermore, we have γ r+1
n
implies that g(x) := (x − 1)(x − γ)(x r+1 − α) is a divisor of xn − 1. Let C be the q-ary cyclic code of
length n generated by g(x).
By Lemma 3.1, to show that the locality of C is r, it is sufficient to show that the Hamming distance of
n −1
⊥
. Then by Lemma 2.1(i), it is sufficient to show that the Hamming
C is at most r + 1. Put h(x) = xg(x)
distance of hh(x)i is at most r + 1 since hh(x)i and C ⊥ are equivalent. Consider the codeword (x − 1)(x −
in
P
γ)h(x) of hh(x)i. Since (x−1)(x−γ)h(x) = ri=0 αr−i x r+1 , the Hamming weight of (x−1)(x−γ)h(x)
is r + 1. Hence, the Hamming distance of hh(x)i is upper bounded by r + 1.
Now, we claim that the cyclic code C generated by g(x) has minimum distance at least 4. We prove
this claim by contradiction. Suppose that the minimum distance of C were at most 3. Then there exists a
nonzero polynomial c(x) = c0 + c1 xj + c2 xk ∈ C with 0 < j < k 6 n − 1 divisible by g(x). Since
n
β, β r+2 , β 1+2(r+1) are roots of x r+1 − α and hence roots of g(x), by Lemma 2.1(ii) we have
1
βj
βk
0 = det 1 β j β (r+1)j β k β (r+1)k = β j+k (β (r+1)j − 1)(β (r+1)k − 1)(β (r+1)j − β (r+1)k ). (2)
1 β j β 2(r+1)j β k β 2(r+1)k
This forces that n|j(r + 1) or n|(k − j)(r + 1) or n|k(r + 1) since β ∈ F̄q is an nth primitive root of unity.
We claim that each of these three divisibilities leads to both n|j(r + 1) and n|k(r + 1).
Case 1. n|j(r + 1).
Since 1, β, β r+2 are also the roots of c(x), by Lemma 2.1(ii) we have
1
1
1
= (β j − 1)(β k(r+2) − 1) − (β k − 1)(β j(r+2) − 1) = 0.
βj
βk
det 1
(3)
(r+2)j
(r+2)k
1 β
β
7
By using the condition β j(r+1) = 1, Equation (3) is simplified to
0 = (β j − 1)(β k(r+2) − 1) − (β k − 1)(β j(r+2) − 1)
= (β j − 1)(β k(r+2) − 1) − (β k − 1)(β j − 1)
= (β j − 1)(β k(r+2) − β k ) = β k (β j − 1)(β k(r+1) − 1).
Observe that β is a primitive nth root of unity and 0 < j < n. It follows that β k(r+1) = 1, or equivalently
n|k(r + 1).
Case 2. n|(k − j)(r + 1).
By using the condition β (k−j)(r+1) = 1, i.e., β j(r+1) = β k(r+1) , Equation (3) is simplified to
0 = (β j − 1)(β k(r+2) − 1) − (β k − 1)(β j(r+2) − 1)
= (β j − 1)(β j(r+1)+k − 1) − (β k − 1)(β j(r+2) − 1)
= (β k − β j )(1 − β j(r+1) ) = β j (β k−j − 1)(1 − β j(r+1) ).
Since 0 < k 6= j < n, we must have n|j(r + 1). This gives β k(r+1) = β j(r+1) = 1. Therefore, we have
n|k(r + 1) as well.
Case 3. n|k(r + 1).
In this case, we can mimic the proof of Case 1 by swapping j and k.
All three cases lead to the conclusion that n|k(r + 1) and n|j(r + 1). Set h1 =
Then h1 6= h2 ∈ {1, 2, . . . , r}. Moreover, we have the following four identities.
n
αh1 = β r+1 ×h1 = β j ;
n
n
αh2 = β r+1 ×h2 = β k ;
α2h1 = γ r+1 ×h1 = γ j ;
j(r+1)
n
and h2 =
k(r+1)
n .
n
α2h2 = γ r+1 ×h2 = γ k .
Finally, we notice that 1, β, γ are the roots of c(x). By Lemma 2.1(ii), this gives
1 1 1
1
1
1
0 = det 1 β j β k = det 1 αh1 αh2 = (αh1 − 1)(αh2 − 1)(αh1 − αh2 ).
1 γj γk
1 α2h1 α2h2
This is a contradiction due to fact that α is a (r + 1)th primitive root of unity and h1 6= h2 ∈ {1, 2, . . . , r}.
The minimum distance of C is at most 4 by the bound (1). The proof is completed.
3.4 A locally reparable codes of length 2(q − 1) and distance 4
Finally, we present a q-ary optimal locally repairable code of length n = 2(q − 1) and minimum distance 4.
Theorem 3.4. If r + 1 divides 2(q − 1), then there exists a q-ary optimal [n = 2(q − 1), n −
locally repairable code with locality r + 1.
n
r+1
− 2, 4]
n
Proof. Let β ∈ F̄q be a nth primitive root of unity. Then β 2 is an element of Fq . Put α = β r+1 . Then α 6= 1
q−1
n
and αq−1 = (β n ) r+1 = 1, i.e., α is an element of Fq \ {0, 1}. Let g(x) = (x − 1)(x − β 2 )(x r+1 − α). It is
straightforward to verify that g(x) is a divisor of xn − 1. Let C be the cyclic code with generator polynomial
g(x).
By the similar augments in the proof of Theorem 1.1(i), one can show that dual code of C has minimum
distance at most r + 1.
Since 1, β, β 2 are roots of g(x), C has distance at least 4.
It is easy to check that code C meets the Singleton-type bound (1). The proof is completed.
8
References
[1] A. Barg, I. Tamo, and S. Vlăduţ, Locally recoverable codes on algebraic curves, IEEE Trans. Inform.
Theory 63(8)(2017), 4928–4939. 1
[2] A. Barg, K. Haymaker, E. Howe, G. Matthews, and A. Vrilly-Alvarado, Locally recoverable codes
from algebraic curves and surfaces, in Algebraic Geometry for Coding Theory and Cryptography,
E.W. Howe, K.E. Lauter, and J.L. Walker, Editors, Springer, 2017, pp. 95–126. 1, 2
[3] M. Forbes and S. Yekhanin, On the locality of codeword symbols in non-linear codes, Discrete Mathematics 324(6)(2014), 78–84. 1
[4] P. Gopalan, C. Huang, H. Simitci and S. Yekhanin, On the locality of codeword symbols, IEEE Trans.
Inf. Theory 58(11)(2012), 6925–6934. 1
[5] J. Han and L. A. Lastras-Montano, Reliable memories with subline accesses, Proc. IEEE Internat.
Sympos. Inform. Theory, 2007, 2531–2535. 1
[6] C. Huang, M. Chen, and J. Li, Pyramid codes: Flexible schemes to trade space for access efficiency
in reliable data storage systems, Sixth IEEE International Symposium on Network Computing and
Applications, 2007, 79–86. 1
[7] L. Jin, L. Ma and C, Xing, Construction of optimal locally repairable codes via automorphism groups
of rational function fields, https://arxiv.org/abs/1710.09638. 1, 2, 5
[8] X. Li, L. Ma and C. Xing, Optimal locally repairable codes via elliptic curves,
https://arxiv.org/abs/1712.03744. 1, 2
[9] S. Ling and C. Xing, “Coding Theory–a first course”, Cambridge, 2004. 3
[10] F. J. MacWilliams and N. J. A. Sloane, “The Theory of Error-Correcting Codes”, North-Holland, 2006.
3
[11] D. S. Papailiopoulos and A.G. Dimakis, Locally repairable codes, IEEE Trans. Inf. Theory
60(10)(2014), 5843–5855. 1
[12] N. Prakash, G.M. Kamath, V. Lalitha and P.V. Kumar, Optimal linear codes with a local-errorcorrection property, Proc. 2012 IEEE Int. Symp. Inform. Theory, 2012, 2776–2780. 1, 2
[13] N. Silberstein, A.S. Rawat, O.O. Koyluoglu and S. Vichwanath, Optimal locally repairable codes via
rank-matric codes, Proc. IEEE Int. Symp. Inf. Theory, 2013, 1819–1823. 1
[14] I. Tamo and A. Barg, A family of optimal locally recoverable codes, IEEE Trans. Inform. Theory
60(8)(2014), 4661–4676. 1, 2, 4
[15] I. Tamo, D.S. Papailiopoulos and A.G. Dimakis, Optimal locally repairable codes and connections to
matroid theory, IEEE Trans. Inform. Theory 62(12)(2016), 6661–6671. 1
9
| 7cs.IT
|
arXiv:1707.05021v2 [math.ST] 16 Nov 2017
Strong Local Nondeterminism of Spherical
Fractional Brownian Motion
Xiaohong Lan ∗
School of Mathematical Sciences
University of Science and Technology of China
Yimin Xiao †
Department of Statistics and Probability
Michigan State University
November 17, 2017
Abstract
Let B = B (x) , x ∈ S2 be the fractional Brownian motion indexed
by the unit sphere S2 with index 0 < H ≤ 21 , introduced by Istas [12].
We establish optimal upper and lower bounds for its angular power spectrum {dℓ , ℓ = 0, 1, 2, . . .}, and then exploit its high-frequency behavior to
establish the property of its strong local nondeterminism of B.
Key words: Angular power spectrum, Karhunen-Loève expansion, Spherical fractional Brownian motion, Strong local nondeterminism.
2010 Mathematics Subject Classification: 60G60, 60G22, 60G17.
1
Introduction
The spherical fractional Brownian motion (SFBM, for brevity) was introduced
by Istas in 2005 [12], as an extension of the spherical Brownian motion of Lévy
[18] as well as a spherical analogue of fractional Brownian motion indexed by the
Euclidean spaces. Later Istas [13, 14] established the Karhunen-Loève expansion
and studied quadratic variations of the spherical fractional Brownian motion.
The purpose of this paper is to investigate the property of strong local nondeterminism (SLND) for the spherical fractional Brownian motion. This is motivated by studies of sample path properties of Gaussian random fields indexed
by the Euclidean space RN and by the currently increasing interest in stochastic
∗ Corresponding author. Research of X. Lan is supported by NSFC grants 11501538. Email: [email protected]
† Research of Y. Xiao is partially supported by NSF grants DMS-1612885 and DMS1607089. E-mail: [email protected]
1
modeling of spherical data in statistics, cosmology and other applied areas (see
below).
The concept of local nondeterminism (LND) of a Gaussian process was first
introduced by Berman [3] to unify and extend his methods for studying the
existence and joint continuity of local times of real-valued Gaussian processes.
Roughly speaking, a Gaussian process is said to have the LND property if has
locally approximately independent increments, see [3, Lemma 2.3] for precise
description. This property allowed Berman to overcome some difficulties caused
by the complex dependence structure of a non-Markovian Gaussian process
for studying its local times. Pitt [24] and Cuzick [6] extended Berman’s LND
to Gaussian random fields. However, the property of LND is not enough for
establishing fine regularity properties such as the law of the iterated logarithm
and the uniform modulus of continuity for the local times or self-intersection
local times of Gaussian random fields. For studying these and many other
problems on Gaussian random fields, the appropriate properties of strong local
nondeterminism (SLND) have proven to be more powerful. Instead of recalling
definitions of various forms of strong local nondeterminism for (isotropic or
anisotropic) Gaussian random fields indexed by RN and their applications, we
refer to Xiao [26, 27, 28] for more information.
Recently, Lan, Marinucci and Xiao [16] have studied the SLND property of
a class of Gaussian random fields indexed by the unit sphere S2 , which are also
called spherical Gaussian random fields. The main difference between [16] and
the aforementioned work for Gaussian fields indexed by the Euclidean space is
that [16] takes the spherical geometry of S2 into full consideration and its method
relies on harmonic analysis on the sphere. More specifically, Lan, Marinucci and
Xiao have considered a centered isotropic Gaussian random field T = {T (x), x ∈
S2 }. That is, T satisfies
E T (x)T (y) = E T (gx)T (gy)
(1)
for all g ∈ SO(3) which is the group of rotations in R3 . See [20] for a systematic
account on random fields on S2 . By applying harmonic analytic tools on the
sphere, Lan, Marinucci and Xiao [16] have proved that the SLND property of an
isotropic Gaussian field T on S2 is determined by the high-frequency behavior of
its angular power spectrum. Moreover, by applying SLND, they have established
exact uniform modulus of continuity for a class of isotropic Gaussian fields on
S2 .
Since SFBM B = B (x) , x ∈ S2 is not isotropic in the sense of (1), the
results on SLND in [16] are not directly applicable. In our approach we will
make use of the Karhunen-Loève expansion of the spherical fractional Brownian motion obtained by Istas [13] and derive optimal upper and lower bounds
for the coefficients {dℓ , ℓ = 0, 1, . . .} (see (7) below for the definition). These
bounds for {dℓ } correct the last part of Theorem 1 in [13] and will be useful for
studying the dependence structures and sample path properties of SFBM. This
paper provides an important step towards this direction. More specifically, we
demonstrate that the coefficients {dℓ } play the same role as the angular power
2
spectrum of an isotropic Gaussian field on S2 in [16] and their high frequency
behavior determines the property of strong local nondeterminism of SFBM. For
this reason, we will also call the sequence {dℓ , ℓ = 0, 1, . . .} the angular power
spectrum of SFBM.
Similarly to the cases of Gaussian random fields indexed by the Euclidean
space RN (cf. [26, 27, 28]), we expect that the SLND property in Theorem
3.2 will be useful for studying regularity (e.g., the exact modulus of continuity,
exact modulus of nondifferentiability, etc) and fractal properties of SFBM. This
will be carried out in a subsequent paper [17].
Our analysis on SFBM and other spherical Gaussian random fields is strongly
motivated by applications in a number of scientific areas, such as geophysics, astrophysics, cosmology, and atmospheric sciences (see e.g. [2, 4, 7, 8]). Huge data
sets from satellite missions such as the Wilkinson Microwave Anisotropy Probe
(WMAP ) of NASA (see http://map.gsfc.nasa.gov/) and the Planck mission of
the European Space Agency (see http://sci.esa.int/planck/53103-planck-cosmology/)
have been collected and made publicly available. Spherical random fields (usually assumed to be Gaussian) have been proposed for modeling such data sets.
Related to aforementioned aspects, we also mention that in probability and
statistics literature various isotropic or anisotropic Gaussian random fields on
S2 have been constructed and studied (see e.g. [9, 10, 11, 15, 23]). Excursion
probabilities and topological properties of excursion sets of isotropic Gaussian
random fields on S2 have been studied in [5, 21]. Many interesting questions
on probabilistic and statistical properties of anisotropic Gaussian random fields
on the sphere can be raised. In order to study these problems. it would be
interesting to establish appropriate properties of strong local nondeterminism
for anisotropic Gaussian random fields on S2 .
The rest of the paper is organized as follows. In Section 2, we recall briefly
some background material on SFBM, including its Karhunen-Loève expansion
from [13], and an analysis of its random coefficients. Our main result in this
section is Theorem 2.2, which provides optimal upper and lower bounds for the
angular power spectrum {dℓ , ℓ = 0, 1, . . .} of SFBM. In Section 3, we combine
the high frequency behavior of {dℓ } and Proposition 7 in [16] to establish the
property of strong local nondeterminism for SFBM.
2
SFBM and asymptotic behavior of its angular
power spectrum
Let N be the North pole on S2 , and dS2 (x, y) the geodesic distance between x
and y on S2 . Recall from Istas [12] the definition of SFBM.
Definition 2.1 The SFBM B = B (x) , x ∈ S2 is a centered Gaussian random field with
B (N ) = 0, a.s.
(2)
and
2
2H
E [B (x) − B (y)] = dS2 (x, y)
3
,
(3)
for any x, y ∈ S2 and 0 < H ≤ 21 .
It is well known that fractional Brownian motion indexed by Rd can be
defined for every 0 < H ≤ 1. This is different from the case when the index
set is S2 . Istas [12] proved that SFBM exists if and only if the Hurst index
H ∈ (0, 12 ]. When H = 12 , it is the classical Lévy spherical Brownian motion,
see [18, 22].
Our work is based on the following Karhunen-Loève expansion of B (x) ,
x ∈ S2 proved by Istas [13, Theorem 1]:
∞ X
ℓ
X
p
i πdℓ εℓm (Yℓm (x) − Yℓm (N )) .
B (x) =
(4)
ℓ=0 m=−ℓ
Here {Yℓm : ℓ = 0, 1, ..., m = −ℓ, ..., ℓ} are the spherical harmonic functions on
S2 ; that is, they are eigenfunctions of the spherical Laplacian:
1 ∂
∂
1 ∂2
∆S2 =
,
sin θ
+
sin θ ∂θ
∂θ
sin2 θ ∂φ2
where (θ, φ) ∈ [0, π] × [0, 2π) denotes the spherical coordinates of x ∈ S2 . More
precisely, {Yℓm : ℓ = 0, 1, ..., m = −ℓ, ..., ℓ} satisfy
∆S2 Yℓm (θ, φ) = −ℓ(ℓ + 1)Yℓm (θ, φ).
It is known (cf. [20]) that {Yℓm
: ℓ = 0, 1, ..., m = −ℓ, ..., ℓ} form an orthonormal
basis for the space L2 S2 , dσ , where dσ = sin θdθdφ is the Lebesgue measure
on S2 . An explicit form for spherical harmonics is given by
s
2ℓ + 1 (ℓ − m)!
Pℓm (cos θ) eimφ , for m ≥ 0,
Yℓm (θ, φ) =
4π (ℓ + m)!
(5)
m
Yℓm (θ, φ) = (−1) Y ℓ,−m (θ, φ) ,
for m < 0,
where z denotes the complex conjugate of z ∈ C, and Pℓm (cos ϑ) are the associated Legendre functions (cf. [20, pp.315–316]) defined in terms of the Legendre
polynomials {Pℓ , ℓ = 0, 1, . . .} as
Pℓm (x) = (−1)m (1 − x2 )m/2
dm
Pℓ (x)
dxm
for m = 0, 1, . . . , ℓ, and
Pℓm (x) = (−1)m
(ℓ − m)!
Pℓ,−m (x)
(ℓ + m)!
for m negative. Moreover, the following orthonormality property holds:
Z
′
′
Ylm (x) Y l′ m′ (x) dσ (x) = δ ll δ m
m .
S2
4
(6)
by
Now recall from [13] that, for all ℓ ≥ 0, the coefficients dℓ in (4) are defined
Z
1
2H
dℓ =
dS2 (x, N ) Pℓ (hx, N i) dσ (x) ,
(7)
2π S2
where h·, ·i is the usual inner product in R3 . Similarly to the angular power spectrum of an isotropic Gaussian random field on S2 , the sequence {dℓ , ℓ = 0, 1, 2, ...}
plays an important role in determining the dependence structure and other probabilistic properties of the SFBM {B (x) , x ∈ S2 }. Hence we will also refer to
{dℓ , ℓ = 0, 1, 2, ...} as the angular power spectrum of SFBM.
Moreover, it follows from Theorem 1 ofIstas [13] that the Karhunen-Loève
expansion (4) holds in L2 S2 × Ω, dσ ⊗ P sense and in L2 (P) sense for every fixed x ∈ S2 . Hence one can represent the random coefficients εℓm , ℓ =
0, 1, ..., m = −ℓ, ..., ℓ in (4) by
Z
−1/2
εℓm = −i (πdℓ )
B(x)Yℓm (x)dσ(x),
S2
where the equality holds in L2 (P) sense. Notice that {εℓm }lm is a set of complexvalued Gaussian random variables. Obviously, E (εℓm ) = 0 for all ℓ, m, due to
the zero-mean property of B(x). Moreover, recall (5) and (6), we can verify that
m
1
εℓm = (−i) εℓ,−m and E (εℓ1 m2 εℓ2 m2 ) = δ ℓℓ12 δ m
m2 .
(8)
Therefore, {εℓm , ℓ ≥ 0, m = 0, 1, ..., ℓ} is a set of i.i.d. standard complex Gaussian random variables.
The following is the main result of this section. The bounds in (9) correct
the last part of Theorem 1 in [13]. The high-frequency behavior of dℓ in (9) is
essential for proving the SLND property of B in Theorem 3.2 below. Together,
they will allow us to study precise analytic and geometric properties of the
sample functions of SFBM. See [17] for further information.
Theorem 2.2 Let B(x), x ∈ S2 be the spherical fractional Brownian motion
of index H ∈ (0, 1/2]. There exists a uniform constant K1 such that
K1−1 ≤ d0 ≤ K1
K1−1 ℓ−(2H+2) ≤ dℓ ≤ K1 ℓ−(2H+2)
for all ℓ = 1, 2, ....
(9)
Proof. We work in spherical coordinates (θ, φ) with θ ∈ [0, π] and φ ∈ [0, 2π] .
By the definition of dℓ in (7) and a change of variable x = cos θ we obtain
Z π
dℓ =
θ2H Pℓ (cos θ) sin θdθ.
(10)
0
Recall from [13] the Dirichlet-Mehler representation for Pℓ (cos θ) (0 < θ < π),
√ Z π
sin (ℓ + 12 )ϕ
2
√
Pℓ (cos θ) =
dϕ,
π θ
cos θ − cos ϕ
5
we see that (10) becomes
√ Z π
Z π
sin ℓ + 12 ϕ
2
√
dℓ =
dϕdθ
θ 2H sin θ
π 0
cos θ − cos ϕ
θ
#
√ Z π
"Z ϕ
1
θ 2H sin θ
2
√
ϕ
sin
ℓ+
=
dθ dϕ.
π 0
2
cos θ − cos ϕ
0
From the elementary fact that sin θ2 ≤
0 ≤ θ ≤ π, it follow that
θ
sin
2
2H
h
≤ θ 1−
2H
2H
θ
θ
2H
≤
≤ 2 · sin
,
2
2
Hence, we have
Z
0
1
6
θ 2
2
i
≤ 2 sin 2θ for all
0 ≤ θ ≤ π.
deℓ ≤ dℓ ≤ 22H deℓ ,
where
22H+1/2
deℓ =
π
θ
2
π
(11)
2H
Z ϕ
sin θ2
sin θ
1
√
dθdϕ.
ϕ
sin
ℓ+
2
cos θ − cos ϕ
0
Meanwhile, by a change of variable u = sin 2θ in the inside integral, we have
Z
0
ϕ
2H
Z sin ϕ2
sin 2θ
sin θ
u2H+1
3
2
√
q
du
dθ = 2
cos θ − cos ϕ
0
sin2 ϕ2 − u2
Z
ϕ 2H+1 1 v H
1
2
√
= 2 sin
dv
2
1−v
0
1
1
ϕ 2H+1
sin
= 2 2 B H + 1,
,
2
2
where B (·, ·) is the Beta function defined on R+ × R+ . Consequently,
Z π
22H+1
ϕ 2H+1
1
1
e
dℓ =
ϕ sin
dϕ.
B H + 1,
sin
ℓ+
π
2
2
2
0
For ℓ = 0,
(12)
Z π
1
ϕ 2H+2
22H+1
dϕ,
B H + 1,
sin
de0 =
π
2
2
0
which is readily seen that de0 ≤ 8. For ℓ ≥ 1, we will make use of some
techniques from complex analysis. Let z = eiϕ/2 , then the integral in (12) can
be written as
Z π
Z
1
ϕ 2H+1
sin
ℓ+
ϕ
sin
(13)
dϕ = 2 Im fℓ (z) dz,
2
2
0
C
6
where C = z : z = eit , 0 ≤ t ≤ π/2 is the unit circle in the first quadrant with
the direction counterclockwise and fℓ (z) is the principal branch of the complex
function
2H+1
z 2ℓ
1
Fℓ (z) = 2H 2H+1 z −
, z ∈ C,
2 i
z
Obviously, fℓ (z) is analytic in C\ {0, ±1}. Moreover, let ǫ > 0 be an arbitary
small value, L1 and L2 the two line segments defined by
L1 = {z : z = i (1 − t) , ǫ ≤ t ≤ 1}
and
L2 = {z : z = t, ǫ ≤ t ≤ 1 − ǫ} ;
Meantime, let C1 and C2 be the circles with radius ǫ in the first and second
quadrants, defined respectively by
C1 = z : z = ǫeit , 0 ≤ t ≤ π/2
and
n
o
C2 = z : z = ǫei(π−t) + 1, 0 ≤ t ≤ π/2 .
Now consider the following integrals
Z
Z
h
i2H+1
i2ℓ−2H−1 1
2
2ℓ−2H−1
−
(1
−
t)
−
1
(1
−
t)
dt
fℓ (z) dz =
22H i2H+1 ǫ
L1
Z
2H+1
(−1)ℓ 1 2ℓ−2H−1 2
=
dt,
u
u +1
2H
2
ǫ
Z
fℓ (z) dz
Z 1
2H+1
1
dt
t2ℓ−2H−1 t2 − 1
2H
2H+1
2 i
ǫ
Z
i2H+1 1
ℓ−H−1/2 2H+1
(1 − v)
v
dt,
22H ǫ
=
L2
=
and
Z
fℓ (z) dz =
C1
Z
C2
fℓ (z) dz =
1
2H
2 i2H+1
1
22H i2H+1
Z
0
Z
0
π
2
π
2
2ℓ−2H−1
2H+1
ǫ2 e2i(π−t) − 1
dt,
ǫei(π−t)
2H+1
2ℓ−2H−1
dt.
ǫ2 e2i(π−t) + 2ǫei(π−t)
1 + ǫei(π−t)
Careful calculations show that
Z
Z
fℓ (z) dz = 0, lim
Im
L1
ǫ→0
Cj
7
fℓ (z) dz = 0, j = 1, 2
(14)
and
lim
ǫ→0
Z
fℓ (z) dz =
L2
i2H+1
B (2H + 2, ℓ − H + 1/2)
22H
(15)
Thus, by the Cauchy integral theorem in complex analysis, we have
Z
Z
fℓ (z) dz = − lim
fℓ (z) dz
C
ǫ→0
= − lim
ǫ→0
Z
L1 +C1 +L2 +C2
L1
fℓ (z) dz −
i2H+1
1
B
2H
+
2,
ℓ
−
H
+
22H
2
which leads to that
Z
1
1
sin H +
π .
Im fℓ (z) dz = 2−2H−1 B 2H + 2, ℓ − H +
2
2
C
(16)
in view of the equalities (14) and (15). Therefore, by combining the equalities
(16) and (12), (13) above, we obtain that
1
1
2
1
B 2H + 2, ℓ − H +
sin H +
π .
deℓ = B H + 1,
π
2
2
2
Here, recall the formula
B (a, b) =
Γ (a) Γ (b)
, a > 0, b > 0
Γ (a + b)
with Gamma function Γ (y) = (y − 1) Γ (y − 1) for any y > 1 and the stirling’s
approximation
y y
p
1
1+O
, as y → ∞,
Γ (y) = 2πy
e
y
where the error term O y1 being the same order as y1 , we derive the following
estimates: there exists a uniform constant C1 such that for any ℓ ≥ 1,
C1−1 ℓ−(2H+2) ≤ deℓ ≤ C1 ℓ−(2H+2) .
Let K1 = 2C1 , then the inequalities in (9) are derived in view of (11).
3
Strong local nondeterminism
In order to prove the strong local nondeterminism property of B (x) , x ∈ S2 ,
we first recall the following lemma, which is a consequence of Proposition 7 in
[16].
8
Lemma 3.1 Assume a sequence {dℓ , ℓ = 0, 1, ...} satisfies the condition (9).
Then there exists a constant C2 > 0 depending on H only, such that for all
choices of n ∈ N, all x, x1 , ..., xn ∈ S2 , and γ j ∈ R, j = 1, 2, ..., n, we have
∞ X
ℓ
X
ℓ=0 m=−ℓ
2
n
X
γ j Yℓm (xj ) ≥ C2 ε2H ,
dℓ Yℓm (x) −
(17)
j=1
where ε = min {dS2 (x, xk ) , k = 1, ..., n}.
Now we are ready to state and prove the following theorem. Its conclusion
is referred to as the property of strong local nondeterminism of SFBM
B (x) , x ∈ S2 .
Theorem 3.2 For a SFBM B (x) , x ∈ S2 , there exists a constant K2 > 0
depending only on the Hurst index H ∈ (0, 1/2], such that for all integers n ≥ 1
and all x, x1 , ..., xn ∈ S2 , we have
Var (B (x) |B (x1 ) , ..., B (xn )) ≥ K2 min dS2 (x, xk )
0≤k≤n
2H
,
(18)
where Var (B (x) |B (x1 ) , ..., B (xn )) denotes the conditional variance of B(x)
given B (x1 ) , ..., B(xn ), and x0 = N .
Remark 3.3 Note that the strong local nondeterminism (18) here is slightly
different from the SLND property proved in [16] for isotropic Gaussian random
fields: the minimum on the right hand side of (18) is not only taken over
x1 , ..., xn but also over x0 = N . This is because of the assumption B (N ) = 0 in
the definition of SFBM. From statistics viewpoint, Var (B (x) |B (x1 ) , ..., B (xn ))
is the squared error of predicting the value B(x), given observations of B at
locations x1 , . . . , xn . Since we already know that B (N ) = 0, this information
may reduce the prediction error.
Proof of Theorem 3.2. It is known that for a Gaussian random field B,
n
2
X
γ j B(xj )
,
Var (B(x)|B (x1 ) , ..., B (xn )) = inf E B(x) −
j=1
where the infimum is taken over all (γ 1 , ..., γ n ) ∈ Rn . Hence, in order to establish
(18), it is sufficient to prove that there exists a positive constant C3 such that
n
2
X
γ j B(xj )
≥ C3 ε2H
E B(x) −
(19)
j=1
holds for all γ 1 , ..., γ n ∈ R, where ε = min {dS2 (x, xk ) , k = 0, ..., n} . Let γ 0 =
n
X
γ j , then it follows from the Karhunen-Loève expansion (4) of B (x) and
1−
j=1
9
the properties of random coefficients {εℓm }lm in (8) that the left hand side of
(19) is equal to
( ∞ ℓ
)
n
2
X X p
X
E
γ j Yℓm (xj )
i πdℓ εℓm Yℓm (x) −
j=0
ℓ=0 m=−ℓ
=
∞ X
ℓ
X
ℓ=0 m=−ℓ
=π
πdℓ E |εℓm |
∞ X
ℓ
X
ℓ=0 m=−ℓ
2
dℓ Yℓm (x) −
Yℓm (x) −
n
X
n
X
2
γ j Yℓm (xj )
(20)
j=0
2
γ j Yℓm (xj ) ,
j=0
where x0 = N. Therefore (19) is an immediate consequence of Lemma 3.1 in
view of the equalities (20). This completes the proof of (18).
References
[1] Adler, R. J. and Taylor, J. E. (2007). Random Fields and Geometry.
Springer, New York.
[2] Alkhaled, A. A., Michalak, A. M., Kawa, S. R., Olsen, S. C. and Wang, J.W. (2008). A global evaluation of the regional spatial variability of column
integrated CO2 distributions. J Geophys Res 113: D20303.
[3] Berman, S. M. (1973), Local nondeterminism and local times of Gaussian
processes. Indiana Univ. Math. J. 23, 69–94.
[4] Cabella, P. and Marinucci, D. (2009). Statistical challenges in the analysis
of Cosmic Microwave Background radiation. Ann. Appl. Statist. 2, 61–95.
[5] Cheng, D. and Xiao, Y. (2016). Excursion probability of Gaussian random
fields on sphere. Bernoulli 22, 1113–1130.
[6] Cuzick, J. (1982), Multiple points of a Gaussian vector field. Z. Wahrsch.
Verw. Gebiete 61, 431–436.
[7] de Boyer Montégut, C., Madec, G., Fischer, A. S., Lazar, A. and Iudicone,
D. (2004). Mixed layer depth over the global ocean: an examination of
profile data and a profile-based climatology. J. Geophys. Res. 109, C12003.
[8] Dodelson, S. (2003). Modern Cosmology. Academic Press.
[9] Estrade, A. and Istas, J. (2010). Ball throwing on spheres. Bernoulli 16,
953–970.
[10] Huang, C., Zhang, H. and Robeson S. M. (2011). On the validity of
commonly used covariance and variogram functions on the sphere. Math.
Geosci. 43, 721–733.
10
[11] Huang, C., Zhang, H. and Robeson S. M. (2012). A simplified representation of the covariance structure of axially symmetric processes on the
sphere. Stat. Prob. Lett. 82, 1346–1351.
[12] Istas, J. (2005). Spherical and hyperbolic fractional Brownian motion.
Elect. Comm. Probab. 10, 254–262.
[13] Istas, J. (2006). Karhunen-Loève expansion of spherical fractional Brownian
motions. Statist. Probab. Lett. 76, 1578–1583.
[14] Istas, J. (2007). Quadratic variations of spherical fractional Brownian motions. Stoch. Process. Appl. 117, 476–486.
[15] Jun, M. and Stein, M. L. (2008). Nonstationary covariance models for global
data. Ann. Appl. Statist. 2, 1271–1289.
[16] Lan, X., Marinucci, D. and Xiao, Y. (2016). Strong local nondeterminism and exact modulus of continuity for spherical Gaussian fields, Stoch.
Process. Appl., accepted, arXiv:1603.03642.
[17] Lan, X. and Xiao, Y. (2017). Exact moduli of continuity and nonwhere
differentiability of spherical fractional Brownian motion. In Preparation.
[18] Lévy, P. (1965). Processus Stochastiques et Mouvement Brownien,
Gauthier-Villars.
[19] Marcus, M. B. and Rosen J. (2006). Markov Processes, Gaussian Processes,
and Local Times, Cambridge University Press, Cambridge.
[20] Marinucci, D. and Peccati, G. (2011). Random Fields on the Sphere. Representation, Limit Theorem and Cosmological Applications, Cambridge University Press, Cambridge.
[21] Marinucci, D. and Vadlamani, S. (2016). High-frequency asymptotics for
Lipschitz-Killing curvatures of excursion sets on the sphere. Ann. Appl.
Probab. 26, 462–506.
[22] Noda, A. (1987). Generalized Radon transform and Lévy’s Brownian motion. Nagoya Math. J. 105, 71–87.
[23] Panchenko, D. and Talagrand, M. (2007). On the overlap in the multiple
spherical SK models. Ann. Probab. 35, 2321–2355.
[24] Pitt, L. D. (1978), Local times for Gaussian vector fields. Indiana Univ.
Math. J. 27, 309–330.
[25] Stein, E. M. and Weiss, G. (1971). Introduction to Fourier Analysis on
Euclidean Spaces. Princeton University Press.
11
[26] Xiao, Y. (2007). Strong local nondeterminism and sample path properties
of Gaussian random fields. Asymptotic Theory in Probability and Statistics
with Applications (Tze Leung Lai, Qiman Shao, Lianfen Qian, editors), pp.
136–176, Higher Education Press, Beijing.
[27] Xiao, Y. (2009). Sample path properties of anisotropic Gaussian random
fields. In: A Minicourse on Stochastic Partial Differential Equations, (D.
Khoshnevisan and F. Rassoul-Agha, editors), Lecture Notes in Math. 1962,
pp. 145–212. Springer, New York.
[28] Xiao, Y. (2013). Recent developments on fractal properties of Gaussian
random fields. In: Further Developments in Fractals and Related Fields,
(J. Barral and S. Seuret, eds.) pp.255–288, Springer, New York.
12
| 10math.ST
|
arXiv:1205.6010v1 [q-bio.GN] 27 May 2012
The Chromatin Organization of an Eukaryotic Genome : Sequence
Specific+ Statistical=Combinatorial
(Extended Abstract)∗
Davide Corona†
Valeria Di Benedetto‡
Raffaele Giancarlo§
Filippo Utro¶
January 22, 2018
Abstract
Nucleosome organization in eukaryotic genomes has a deep impact on gene function. Although
progress has been recently made in the identification of various concurring factors influencing nucleosome
positioning, it is still unclear whether nucleosome positions are sequence dictated or determined by a
random process. It has been postulated for a long time that, in the proximity of TSS, a barrier determines
the position of the +1 nucleosome and then geometric constraints alter the random positioning process
determining nucleosomal phasing. Such a pattern fades out as one moves away from the barrier to become
again a random positioning process. Although this statistical model is widely accepted, the molecular
nature of the barrier is still unknown. Moreover, we are far from the identification of a set of sequence
rules able: to account for the genome-wide nucleosome organization; to explain the nature of the barriers
on which the statistical mechanism hinges; to allow for a smooth transition from sequence-dictated to
statistical positioning and back. Here we show that sequence complexity, quantified via various methods,
can be the rule able to at least partially account for all the above. In particular, we have conducted our
analyses on four high resolution nucleosomal maps of the model eukaryotes S.cerevisiae, C. elegans and
D.melanogaster, and found that nucleosome depleted regions can be well distinguished from nucleosome
enriched regions by sequence complexity measures. In particular, the depleted regions are less complex
than the enriched ones. Moreover, around TSS, complexity measures alone are in striking agreement
with in vivo nucleosome occupancy, in particular precisely indicating the positions of the +1 and -1
nucleosomes. Those findings indicate that the intrinsic richness of subsequences within sequences plays a
role in nucleosomal formation in genomes, and that sequence complexity constitutes the molecular nature
of nucleosome barrier.
1
Background
It is well known that chromatin organization in Eukaryotic genomes has a deep impact on gene regulation
and function, e.g., [1]. Therefore, it comes as no surprise that a substantial amount of research has been
devoted to this fascinating topic, with particular focus on the identification of nucleosome positions within
the genomes of model organisms and of mechanisms influencing that positioning. Although quite a bit
of progress has been made on the identification of various concurring factors that influence nucleosome
positioning [10], an answer to the question raised by Kornberg in 1981 [4] as to which extend nuclesome
positions are “sequence dictated” or determined by a random process has remained elusive. As a matter
∗ Work presented at the 8th SIBBM Seminar (Annual Conference Meeting of the Italian Biophysics and Molecular Biology
Society)- May 24-26 2012, Palermo, Italy
† Dulbecco Telethon Institute c/o Università di Palermo, Dipartimento di Biologia Cellulare e dello Sviluppo, Palermo, Italy;
[email protected]
‡ Università di Palermo, Dipartimento di Matematica ed Informatica, Palermo, Italy; [email protected]
§ Università di Palermo, Dipartimento di Matematica ed Informatica, Palermo, Italy; [email protected]
¶ Computational Genomics Group, IBM T.J. Watson Research Center, Yorktown Heights, USA; [email protected]
1
of fact, it has been the object of intense debate, in particular in view of a result by Widom et al. [9]
claiming that there exists a genomic code for nucleosome positioning. In mathematical terms, a code
is a very specific and constrained object, even when it is degenarate as the genetic code. Therefore, it
seems that a verbatim use of the term code in the context of chromatin studies would be quite misleading
and actually very easy to challenge, e.g., [11, 12]. Indeed, the focus has shifted from sequence dictating
to sequence influencing chromatin organization, a finding that is much less amenable to challenge and
dismissal [3]. In contrast to the “sequence dictating-influencing” debate just outlined, the statistical
positioning mechanism has had very little challenge. It is worth recalling that it is based on two main
ingredients: the existence of a barrier and of a statistical law governing the positioning of the nucleosomes
[5]. Generalizing the just mentioned earlier results by Kornberg and Stryer, Möbius and Gerland [8] have
shown that such a mechanism can be described and quantified by the Tonks model of statistical physics.
In particular, in the proximity of TSS [6, 8], the barrier determines the position of the +1 nucleosome and
then geometric constraints alter the random positioning process determining the well known nucleosomal
pattern in DNA. Such a pattern fades out as one moves away from the barrier to become again a random
positioning process. An analogous behaviour is followed by the -1 nucleosome.
2
Statement of Results
All of the above studies leave open several questions. Indeed, given the poor results in finding a consistent
and concise set of “sequence-rules” , e.g., [11, 12], that explain how the sequence influences nucleosome
positioning, it is a challenging problem to show that such a set of rules exists. Analogously, although
the statistical model is widely accepted [10, 12], the molecular nature of the barrier is unknown [12].
Moreover, the interplay between sequence-dictated vs statistical positioning has been hardly explored,
although it is to be expected that the true cellular state is probably a combination of both machanisms
[2]. A study by Mavrich et al. [6], performed on DNA regions around 4799 TSS in Yeast, addresses in
part this latter problem indicating that the sequence dictates the positioning of the so-called +1 and -1
nucleosomes, leaving the rest to the statistical positioning mechanism. In that study, an effort is also
made to identify some of the compositional properties of the nucleosome-free region (NFR) responsible for
the formation of the barriers. However, it is also pointed out that the identified sequence biases, mostly
for dinucleotides, may be special cases of more general, and yet unknown, properties of the genomic
sequences up and downstram of a TSS. In a nutshell, we are far from the identification of a concise set
of sequence rules able: (a) to account for the genome-wide nucleosome organization in a genome; (b) to
explaing the nature of the barriers on which the statistical mechanism hinges; (c) to allow for a smooth
transition from sequence-dictated to statistical positioning and back.
As said before, a code for nucleosome positioning seems to be too stringent. Fortunately, there are
other intrinsic properties of sequences, in particular the ones measured by complexity measures, that
may play a role in influencing nucleosomal organization in a eukaryotic genome. It is worth recalling
that complexity measures usually quantify the “intrinsic richness” of distinct subsequences within a given
sequence. Here we study whether sequence complexity, quantified via various methods, can be the rule
able to at least partially account for all (a)-(c) above. Towards this end, we have the following results.
(1) We have conducted extensive studies on four high resolution nucleosomal maps of three model
orgamisms, i.e., Yeast, C. Elegans and Drosophila Melanogaster [3, 7, 12], to obtain the following
results. Nucleosome depleted regions in each map can be well distinguished from nuclesome enriched
regions in each map by sequence complexity measures. In particular, the depleted regions are less
complex than the enriched ones. Such a finding indicates that the intrinsic richness of subsequences
within sequences plays a role in influencing nucleosomal formation in genomes, addressing point (a)
above.
(2) We have applied our methodology to the same TSS dataset of Mavrich et al. [6], and also accounted
for the additional insights by Möbius and Gerland [8]. We find that the NFR is characterized by
an area of lower complexity with respect to the two regions flanking it, with the TSS being placed
in the proximity of the absolute minimum of each complexity curve we have computed. The two
barriers are towards the local maxima at the left and right of the TSS. Particularly striking is the
complexity curve obtained with linguistic complexity, based on a dictionary of words of length up to
seventeen (therefore much larger than the one considered by Mavrich et al.). Indeed, the positions
of the +1 and -1 nucleosomes, respectively, are in the proximity of the first maximum to the right
2
and to the left, respectively, of the TSS on the complexity curve. In view of point (b) above, such a
finding suggests, at least as far as TSS are concerned, that the combinatorial properties of the NFR
subsequence are strongly associated with the creation of the barriers, highlighting that its nature
is at least in part combinatorial, i.e., dictated by intrinsic properties of sequences.
(3) Since the complexity of subsequences within a sequence can be modulated, points (1) and (2) above
suggest that, at least for TSS, such a modulation can smoothly accomodate for both sequencedictated and statistical positioning, in a continuum that requires no other particular arrangement
for switching between the two “states” of interest. Indeed, the low complexity area characterizing
the NFR indicated where nucleosomes should not be, giving also an indication that the +1 and -1
nucleosomes must be placed towards the local maxima of the complexity curve at the end of the
NFR. Such an interpretation sheds some light on point (c) above.
References
[1] G. Felsenfeld and M. Groudine. Controlling the double helix. Nature, 421:448–453, 2003.
[2] C. Jiang and B.F. Pugh. Nucleosome positioning and gene regulation: advances through genomics.
Nature Genetics, 10:161–172, 2010.
[3] N. Kaplan, I.K. Moore, Y. Fondufe-Mittendorf, A. J. Gossett, D. Tillo, Y. Field, E. M. LeProust,
T.R. Hughes, J. D. Lieb, J. Widom, and E. Segal. The DNA-encoded nucleosome organization of a
eukaryotic genome. Nature, 2008.
[4] R. D. Kornberg. The locations of nucleosomes in chromatin:specific or statistical? Nature, 292:579–
580, 1981.
[5] R.D. Kornberg and L. Stryer. Statistical distributions of nucleosomes: nonrandom locations by a
stochastic mechanism. Nucleic Acids Research, 16:6677–6690, 1988.
[6] T.N. Mavrich, I.P. Ioshikhes, B.J. Venters, C. Jiang, L.P. Tomsho, J. Qi, S.C. Schuster, I. Albert,
and B. F. Pugh. A barrier nucleosome model for statistical positioning of nucleosomes throughout
the yeast genome. Genome Research, 2008.
[7] T.N. Mavrich, C. Jiang, I.P. Ioshikhes, X. Li, B.J. Venters, S.J. Zanton, L.P. Tomsho, J. Qi, R.L.
Glaser, S.C. Schuster, D.S. Gilmour, I. Albert, and B.F. Pugh. Nucleosome organization in the
Drosophila genome. Nature, 453:358–364, 2008.
[8] W. Möbius and U. Gerland. Quantitative test of the barrier nucleosome model for statistical positioning of nucleosomes up- and downstream of transcription start sites. PLoS Computational
Biology, 6:e891, 2010.
[9] E. Segal, Y. Fondufe-Mittendorf, L. Chen, A. Thastrom, Y. Field, J.K. Moore, J.P. Wang, and
J. Widom. A genomic code for nucleosome positioning. Nature, 442:772–778, 2006.
[10] E. Segal and J. Widom. What controls nucleosome positions? Trends in Genetics, 746:1–9, 2009.
[11] A. Valouev, J. Ichikawa, T. Thaisan, J. Stuart, R. Swati, H. Peckham, K. Zeng, J.A. Malek, G. Costa,
K. McKernan, A. Sidow, A. Fire, and S. M. Johnson. A high-resolution, nucleosome position map of
c. elegans reveals a lack of universal sequence-dictated positioning. Genome Research, 18:1051–1063,
2008.
[12] Y. Zhang, Z. Moqtaderi, B. P. Rattner, G. Euskirchen, M. Snyder, J.T. Kadonaga, X . Shirley Liu,
and K. Struhl. Intrinsic histone-DNA interactions are not the major determinant of nucleosome
positions in vivo. Nature Structural and Molecular Biology, 16:847–852, 2009.
3
| 5cs.CE
|
Submitted to the Annals of Statistics
arXiv: arXiv:0000.0000
ACCURACY ASSESSMENT FOR HIGH-DIMENSIONAL
LINEAR REGRESSION∗
arXiv:1603.03474v2 [math.ST] 24 Sep 2016
By T. Tony Cai, and Zijian Guo
University of Pennsylvania
This paper considers point and interval estimation of the `q loss
of an estimator in high-dimensional linear regression with random
design. We establish the minimax rate for estimating the `q loss and
the minimax expected length of confidence intervals for the `q loss of
rate-optimal estimators of the regression vector, including commonly
used estimators such as Lasso, scaled Lasso, square-root Lasso and
Dantzig Selector. Adaptivity of the confidence intervals for the `q loss
is also studied. Both the setting of known identity design covariance
matrix and known noise level and the setting of unknown design
covariance matrix and unknown noise level are studied. The results
reveal interesting and significant differences between estimating the
`2 loss and `q loss with 1 ≤ q < 2 as well as between the two settings.
New technical tools are developed to establish rate sharp lower
bounds for the minimax estimation error and the expected length of
minimax and adaptive confidence intervals for the `q loss. A significant difference between loss estimation and the traditional parameter
estimation is that for loss estimation the constraint is on the performance of the estimator of the regression vector, but the lower bounds
are on the difficulty of estimating its `q loss. The technical tools developed in this paper can also be of independent interest.
1. Introduction. In many applications, the goal of statistical inference
is not only to construct a good estimator, but also to provide a measure of
accuracy for this estimator. In classical statistics, when the parameter of
interest is one-dimensional, this is achieved in the form of a standard error
or a confidence interval. A prototypical example is the inference for a binomial proportion, where often not only an estimate of the proportion but also
its margin of error are given. Accuracy measures of an estimation procedure
have also been used as a tool for the empirical selection of tuning parameters.
A well known example is Stein’s Unbiased Risk Estimate (SURE), which has
been an effective tool for the construction of data-driven adaptive estimators in normal means estimation, nonparametric signal recovery, covariance
∗
The research was supported in part by NSF Grants DMS-1208982 and DMS-1403708,
and NIH Grant R01 CA127334.
MSC 2010 subject classifications: Primary 62G15; secondary 62C20, 62H35
Keywords and phrases: Accuracy assessment, adaptivity, confidence interval, highdimensional linear regression, loss estimation, minimax lower bound, minimaxity, sparsity.
1
2
T. T. CAI AND Z. GUO
matrix estimation, and other problems. See, for instance, [25, 21, 15, 11, 32].
The commonly used cross-validation methods can also be viewed as a useful
tool based on the idea of empirical assessment of accuracy.
In this paper, we consider the problem of estimating the loss of a given
estimator in the setting of high-dimensional linear regression, where one
observes (X, y) with X ∈ Rn×p and y ∈ Rn , and for 1 ≤ i ≤ n,
yi = Xi· β + i .
iid
Here β ∈ Rp is the regression vector, Xi· ∼ Np (0, Σ) are the rows of X,
iid
and the errors i ∼ N (0, σ 2 ) are independent of X. This high-dimensional
linear model has been well studied in the literature, with the main focus
on estimation of β. Several penalized/constrained `1 minimization methods,
including Lasso [28], Dantzig selector [12], scaled Lasso [26] and square-root
Lasso [3], have been proposed. These methods have been shown to work well
in applications and produce interpretable estimates of β when β is assumed
to be sparse. Theoretically, with a properly chosen tuning parameter, these
estimators achieve the optimal rate of convergence over collections of sparse
parameter spaces. See, for example, [12, 26, 3, 23, 4, 5, 30].
b the `q loss kβb − βk2 with 1 ≤ q ≤ 2 is comFor a given estimator β,
q
b
monly used as a metric of accuracy for β. We consider in the present paper
b
both point and interval estimation of the `q loss kβb − βk2q for a given β.
Note that the loss kβb − βk2q is a random quantity, depending on both the
estimator βb and the parameter β. For such a random quantity, prediction
and prediction interval are ususally used for point and interval estimation,
respectively. However, we slightly abuse the terminologies in the present paper by using estimation and confidence interval to represent the point and
interval estimators of the loss kβb − βk2q . Since the `q loss depends on the
b it is necessary to specify the estimator in the discussion of loss
estimator β,
estimation. Throughout this paper, we restrict our attention to a broad collection of estimators βb that perform well at least at one interior point or a
small subset of the parameter space. This collection of estimators includes
most state-of-art estimators such as Lasso, Dantzig selector, scaled Lasso
and square-root Lasso.
High-dimensional linear regression has been well studied in two settings.
One is the setting with known design covariance matrix Σ = I and known
noise level σ = σ0 and sparse β. See for example, [16, 2, 22, 30, 27, 20, 7, 1,
19]. Another commonly considered setting is sparse β with unknown Σ and
σ. We study point and interval estimation of the `q loss kβb − βk2q in both
settings. Specifically, we consider the parameter space Θ0 (k) introduced in
HIGH-DIMENSIONAL ACCURACY ASSESSMENT
3
(2.3), which consists of k-sparse signals β with known design covariance
matrix Σ = I and known noise level σ = σ0 , and Θ(k) defined in (2.4),
which consists of k-sparse signals with unknown Σ and σ.
1.1. Our contributions. The present paper studies the minimax and adaptive estimation of the loss kβb − βk2q for a given estimator βb and the minimax
expected length and adaptivity of confidence intervals for the loss. A major
step in our analysis is to establish rate sharp lower bounds for the minimax
estimation error and the minimax expected length of confidence intervals for
the `q loss over Θ0 (k) and Θ(k) for a broad class of estimators of β, which
contains the subclass of rate-optimal estimators. We then focus on the estimation of the loss of rate-optimal estimators and take the Lasso and scaled
Lasso estimators as generic examples. For these rate-optimal estimators, we
propose procedures for point estimation as well as confidence intervals for
their `q losses. It is shown that the proposed procedures achieve the corresponding lower bounds up to a constant factor. These results together
establish the minimax rates for estimating the `q loss of rate-optimal estimators over Θ0 (k) and Θ(k). The analysis shows interesting and significant
differences between estimating the `2 loss and `q loss with 1 ≤ q < 2 as well
as between the two parameter spaces Θ(k) and Θ0 (k).
n
o
2 over Θ (k) is min √1 , k log p
b
• The minimax rate for estimating kβ−βk
0
2
n
n
and over Θ(k) is k logn p . So loss estimation is much easier with the prior
√
information Σ = I and σ = σ0 when lognp k . logn p .
• The minimax rate for estimating kβb − βk2q with 1 ≤ q < 2 over both
2
Θ0 (k) and Θ(k) is k q logn p .
√
In the regime lognp k . logn p , a practical loss estimator is proposed
for estimating the `2 loss and shown to achieve the optimal convergence
rate √1n adaptively over Θ0 (k). We say estimation of loss is impossible if
the minimax rate can be achieved by the trivial estimator 0, which means
that the estimation accuracy of the loss is at least of the same order as the
loss itself. In all other considered cases, estimation of loss is shown to be
impossible. These results indicate that loss estimation is difficult.
We then turn to the construction of confidence intervals for the `q loss.
A confidence interval for the loss is useful even when it is “impossible”
to estimate the loss, as a confidence interval can provide non-trivial upper
and lower bounds for the loss. In terms of convergence rate over Θ0 (k)
or Θ(k), the minimax rate of the expected length of confidence intervals
for the `q loss, kβb − βk2q , of any rate-optimal estimator βb coincides with
4
T. T. CAI AND Z. GUO
the minimax estimation rate. We also consider the adaptivity of confidence
b (The framework for
intervals for the `q loss of any rate-optimal estimator β.
adaptive confidence intervals is discussed in detail in Section 3.1.) Regarding
confidence intervals for the `2 loss in the case of known Σ = I and σ = σ0 ,
a procedure is proposed and is shown to achieve the optimal length √1n
√
adaptively over Θ0 (k) for lognp . k . logn p . Furthermore, it is shown that
this is the only regime where adaptive confidence intervals
exist, even over
√
n
two given parameter spaces. For example, when k1 log p and k1 k2 , it is
impossible to construct a confidence interval for the `2 loss with guaranteed
coverage probability over Θ0 (k2 ) (consequently also over Θ0 (k1 )) and with
the expected length automatically adjusted to the sparsity. Similarly, for
the `q loss with 1 ≤ q < 2, adaptive confidence intervals is impossible over
Θ0 (k1 ) and Θ0 (k2 ) for k1 k2 . logn p . Regarding confidence intervals for
the `q loss with 1 ≤ q ≤ 2 in the case of unknown Σ and σ, the impossibility
of adaptivity also holds over Θ(k1 ) and Θ(k2 ) for k1 k2 . logn p .
Establishing rate-optimal lower bounds requires the development of new
technical tools. One main difference between loss estimation and the traditional parameter estimation is that for loss estimation the constraint is
on the performance of the estimator βb of the regression vector β, but the
lower bound is on the difficulty of estimating its loss kβb − βk2q . We introduce
useful new lower bound techniques for the minimax estimation error and
the expected length of adaptive confidence intervals for the loss kβb − βk2q .
In several important cases, it is necessary to test a composite null against
a composite alternative in order to establish rate sharp lower bounds. The
technical tools developed in this paper can also be of independent interest.
In addition to Θ0 (k) and Θ(k), we also study an intermediate parameter
space where the noise level σ is known and the design covariance matrix Σ is
unknown but of certain structure. Lower bounds for the expected length of
minimax and adaptive confidence intervals for kβb − βk2q over this parameter
space are established for a broad collection of estimators βb and are shown to
be rate sharp for the class of rate-optimal estimators. Furthermore, the lower
bounds developed in this paper have wider implications. In particular, it is
shown that they lead immediately to minimax lower bounds for estimating
kβk2q and the expected length of confidence intervals for kβk2q with 1 ≤ q ≤ 2.
1.2. Comparison with other works. Statistical inference on the loss of
specific estimators of β has been considered in the recent literature. The
papers [16, 2] established, in the setting Σ = I and n/p → δ ∈ (0, ∞), the
b
b
limit of the normalized loss p1 kβ(λ)
− βk22 where β(λ)
is the Lasso estimator with a pre-specified tuning parameter λ. Although [16, 2] provided an
HIGH-DIMENSIONAL ACCURACY ASSESSMENT
5
exact asymptotic expression of the normalized loss, the limit itself depends
on the unknown β. In a similar setting, the paper [27] established the limit
of a normalized `2 loss of the square-root Lasso estimator. These limits of
the normalized losses help understand the properties of the corresponding
estimators of β, but they do not lead to an estimate of the loss. Our results imply that although these normalized losses have a limit under some
regularity conditions, such losses cannot be estimated well in most settings.
A recent paper, [20], constructed a confidence interval for kβb − βk22 in
the case of known Σ = I, unknown noise level σ, and moderate dimension
where n/p → ξ ∈ (0, 1) and no sparsity is assumed on β. While no sparsity
assumption on β is imposed, their method requires the assumption of Σ = I
and n/p → ξ ∈ (0, 1). In contrast, in this paper, we consider both unknown
Σ and known Σ = I settings, while allowing p n and assuming sparse β.
Honest adaptive inference has been studied in the nonparametric function
estimation literature, including [8] for adaptive confidence intervals for linear
functionals, [18, 10] for adaptive confidence bands, and [9, 24] for adaptive
confidence balls, and in the high-dimensional linear regression literature,
including [22] for adaptive confidence set and [7] for adaptive confidence
interval for linear functionals. In this paper, we develop new lower bound
tools, Theorems 8 and 9, to establish the possibility of adaptive confidence
intervals for kβb − βk2q . The connection between `2 loss considered in the
current paper and the work [22] is discussed in more detail in Section 3.2.
1.3. Organization. Section 2 establishes the minimax lower bounds of
estimating the loss kβb − βk2q with 1 ≤ q ≤ 2 over both Θ0 (k) and Θ(k)
and shows that these bounds are rate sharp for the Lasso and scaled Lasso
2 . Secb
estimators, respectively. We then turn to interval estimation of kβ−βk
q
tions 3 and 4 present the minimax and adaptive minimax lower bounds for
the expected length of confidence intervals for kβb−βk2q over Θ0 (k) and Θ(k).
For Lasso and scaled Lasso estimators, we show that the lower bounds can
be achieved and investigate the possibility of adaptivity. Section 5 considers
the rate-optimal estimators and establishes the minimax convergence rate
of estimating their `q losses. Section 6 presents new minimax lower bound
techniques for estimating the loss kβb −βk2q . Section 7 discusses the minimaxity and adaptivity in another setting, where the noise level σ is known and
the design covariance matrix Σ is unknown but of certain structure. Section
8 applies the newly developed lower bounds to establish lower bounds for a
related problem, that of estimating kβk2q . Section 9 proves the main results
and additional proofs are given in the supplemental material [6].
6
T. T. CAI AND Z. GUO
1.4. Notation. For a matrix X ∈ Rn×p , Xi· , X·j , and Xi,j denote respectively the i-th row, j-th column, and (i, j) entry of the matrix X. For
a subset J ⊂ {1, 2, · · · , p}, |J| denotes the cardinality of J, J c denotes
the complement {1, 2, · · · , p}\J, XJ denotes the submatrix of X consisting of columns X·j with j ∈ J and for a vector x ∈ Rp , xJ is the subvector of x with indices in J. For a vector x ∈ Rp , supp(x) denotes the
1
P
support of x and the `q norm of x is defined as kxkq = ( pi=1 |xi |q ) q
for q ≥ 0 with kxk0 = |supp(x)| and kxk∞ = max1≤j≤p |xj |. For a ∈ R,
a+ = max {a, 0}. We use max kX·j k2 as a shorthand for max1≤j≤p kX·j k2
and min kX·j k2 as a shorthand for min1≤j≤p kX·j k2 . For a matrix A, we define the spectral norm
Pp kAk2 = supkxk2 =1 kAxk2 and the matrix `1 norm
kAkL1 = sup1≤j≤p i=1 |Aij |; For a symmetric matrix A, λmin (A) and
λmax (A) denote respectively the smallest and largest eigenvalue of A. We
use c and C to denote generic positive constants that may vary from place
to place. For two positive sequences an and bn , an . bn means an ≤ Cbn
for all n and an & bn if bn . an and an bn if an . bn and bn . an , and
an bn if lim supn→∞ abnn = 0 and an bn if bn an .
2. Minimax estimation of the `q loss. We begin by presenting the
minimax framework for estimating the `q loss, kβb − βk2q , of a given estimator
b and then establish the minimax lower bounds for the estimation error for
β,
b We also show that such minimax lower
a broad collection of estimators β.
bounds can be achieved for the Lasso and scaled Lasso estimators.
2.1. Problem formulation. Recall the high-dimensional linear model,
(2.1)
yn×1 = Xn×p βp×1 + n×1 ,
∼ Nn (0, σ 2 I).
iid
We focus on the random design with Xi· ∼ N (0, Σ) and Xi· and i are
independent. Let Z = (X, y) denote the observed data and βb be a given
b q (Z) any estimator of the loss kβb − βk2q , the
estimator of β. Denoting by L
minimax rate of convergence for estimating kβb − βk2q over a parameter space
Θ is defined as the largest quantity γβ,`
b q (Θ) such that
(2.2)
b q (Z) − kβb − βk2 | ≥ γ b (Θ) ≥ δ,
inf sup Pθ |L
q
β,`q
b q θ∈Θ
L
b q for
for some constant δ > 0 not depending on n or p. We shall write L
b
Lq (Z) when there is no confusion.
We denote the parameter by θ = (β, Σ, σ), which consists of the signal β, the design covariance matrix Σ and the noise level σ. For a given
HIGH-DIMENSIONAL ACCURACY ASSESSMENT
7
θ = (β, Σ, σ), we use β(θ) to denote the corresponding β. Two settings are
considered: The first is known design covariance matrix Σ = I and known
noise level σ = σ0 and the other is unknown Σ and σ. In the first setting,
we consider the following parameter space that consists of k-sparse signals,
Θ0 (k) = {(β, I, σ0 ) : kβk0 ≤ k} ,
(2.3)
and in the second setting, we consider
(2.4)
1
Θ(k) = (β, Σ, σ) : kβk0 ≤ k,
≤ λmin (Σ) ≤ λmax (Σ) ≤ M1 , 0 < σ ≤ M2 ,
M1
where M1 ≥ 1 and M2 > 0 are constants. The parameter space Θ0 (k) is a
subset of Θ(k), which consists of k-sparse signals with unknown Σ and σ.
2
b
The minimax rate γβ,`
b q (Θ) for estimating kβ − βkq also depends on the
b Different estimators βb could lead to different losses kβb − βk2
estimator β.
q
b We
and in general the difficulty of estimating the loss kβb −βk2 varies with β.
q
first recall the properties of some state-of-art estimators and then specify
the collection of estimators on which we focus in this paper. As shown in
[12, 4, 3, 26], Lasso, Dantzig Selector, scaled Lasso and square-root Lasso
satisfy the following property if the tuning parameter is properly chosen,
2 log p
2
(2.5)
sup Pθ kβb − βkq ≥ Ck q
→ 0,
n
θ∈Θ(k)
where C > 0 is a constant. The minimax lower bounds established in [30, 23,
2
31] imply that k q logn p is the optimal rate for estimating β over the parameter
space Θ(k). It should be stressed that all of these algorithms do not require
knowledge of the sparsity k and is thus adaptive to the sparsity provided
k . logn p . We consider a broad collection of estimators βb satisfying one of
the following two assumptions.
(A1) The estimator βb satisfies, for some θ0 = (β ∗ , I, σ0 ) ∈ Θ0 (k),
2
∗ 2
∗
∗ q log p 2
b
σ ≤ α0 ,
(2.6)
Pθ0 kβ − β kq ≥ C kβ k0
n 0
where 0 ≤ α0 < 14 and C ∗ > 0 are constants.
(A2) The estimator βb satisfies
2
∗ 2
∗
∗ q log p 2
b
(2.7)
sup
Pθ kβ − β kq ≥ C kβ k0
σ ≤ α0 ,
n
{θ=(β ∗ ,I,σ):σ≤2σ0 }
where 0 ≤ α0 <
1
4
and C ∗ > 0 are constants and σ0 > 0 is given.
8
T. T. CAI AND Z. GUO
In view of the minimax rate given in (2.5), Assumption (A1) requires βb
to be a good estimator of β at at least one point θ0 ∈ Θ0 (k). Assumption
(A2) is slightly stronger than (A1) and requires βb to estimate β well for a
single β ∗ but over a range of noise levels σ ≤ 2σ0 while Σ = I. Of course,
any estimator βb satisfying (2.5) satisfies both (A1) and (A2). In addition
to Assumptions (A1) and (A2), we also introduce the following sparsity
assumptions that will be used in various theorems.
(B1) Let c0 be the constant defined in (9.14). The sparsity levels k and
k0 satisfy k ≤ c0 min{pγ , logn p } for some constant 0 ≤ γ < 21 and
√
k0 ≤ c0 min{k, lognp }.
(B2) The sparsity levels k1 , k2 and k0 satisfy k1 ≤ k2 ≤ c0 min{pγ , logn p } for
some constant 0 ≤ γ <
1
2
√
and c0 > 0 and k0 ≤ c0 min{k1 , lognp }.
2.2. Minimax estimation of the `q loss over Θ0 (k). The following theorem establishes the minimax lower bounds for estimating the loss kβb − βk2q
over the parameter space Θ0 (k).
Theorem 1. Suppose that the sparsity levels k and k0 satisfy Assumption (B1). For any estimator βb satisfying Assumption (A1) with kβ ∗ k0 ≤ k0 ,
log p 1
2
b
b
√
(2.8)
inf sup Pθ |L2 − kβ − βk2 | ≥ c min k
,
σ02 ≥ δ.
n
n
b
L2 θ∈Θ0 (k)
For any estimator βb satisfying Assumption (A2) with kβ ∗ k0 ≤ k0 ,
2 log p
2
2
b
b
(2.9) inf sup Pθ |Lq − kβ − βkq | ≥ ck q
σ ≥ δ, for 1 ≤ q < 2,
n 0
b q θ∈Θ0 (k)
L
where δ > 0 and c > 0 are constants.
Remark 1. Assumption (A1) restricts our focus to estimators that can
perform well at at least one point (β ∗ , I, σ0 ) ∈ Θ0 (k). This weak condition
makes the established lower bounds widely applicable as the benchmark for
evaluating estimators of the `q loss of any βb that performs well at a proper
subset, or even a single point of the whole parameter space.
In this paper, we focus on estimating the loss kβb − βk2q with 1 ≤ q ≤ 2.
q
b
Similar results can be established for the loss in the form of kβ−βk
q with 1 ≤
q ≤ 2; Under the same assumptions as those in Theorem 1, the lower bounds
for estimating the loss kβb − βkqq hold with replacing the convergence rates
with their 2q power; that is, (2.8) remains the same while the convergence
HIGH-DIMENSIONAL ACCURACY ASSESSMENT
9
p
2 p
rate k q ( log p/nσ0 )2 in (2.9) is replaced by k( log p/nσ0 )q . Similarly, all
the results established in the rest of the paper for kβb − βk2q hold for kβb − βkqq
with corresponding convergence rates replaced by their 2q power.
Theorem 1 establishes the minimax lower bounds for estimating the `2
loss kβb − βk22 of any estimator βb satisfying Assumption (A1) and the `q
loss kβb − βk2q with 1 ≤ q < 2 of any estimator βb satisfying Assumption
(A2). We will take the Lasso estimator as an example and demonstrate
the implications of the abovetheorem. We randomly split Z = (y, X) into
subsamples Z (1) = y (1) , X (1) and Z (2) = y (2) , X (2) with sample sizes n1
and n2 , respectively.
The Lasso estimator βbL based on the first subsample
Z (1) = y (1) , X (1) is defined as
(1)
p
X
kX·j k2
ky (1) − X (1) βk22
βbL = arg minp
+λ
|βj |,
√
β∈R
n1
n1
(2.10)
j=1
p
√
where λ = A log p/n1 σ0 with A > 2 being a pre-specified constant.
Without loss of generality, we assume n1 n2 . For the case 1 ≤ q < 2,
(2.5) and (2.9) together imply that the estimation of the `q loss kβbL − βk2q
is impossible since the lower bound can
be achieved by the trivial
estimator
2
log
p
L
2
b
of the loss, 0. That is, supθ∈Θ0 (k) Pθ |0 − kβ − βkq | ≥ Ck q n
→ 0.
√
p
For the case q = 2, in the regime k lognp , the lower bound k log
in
n
(2.8) can be achieved by the zero estimator and hence estimation √of the loss
kβbL − βk22 is impossible. However, the interesting case is when lognp . k .
n
e 2 proposed in (2.11) achieves the minimax lower
, the loss estimator L
log p
bound
√1
n
in (2.8), which cannot be achieved by the zero estimator. We now
e 2 . Based on the second half
detail the construction of the loss estimator L
(2)
(2)
(2)
sample Z = y , X
, we propose the following estimator,
2
1
2
(2)
(2) bL
e
− σ0
(2.11)
L2 =
y −X β
.
n2
2
+
Note that the first subsample Z (1) = y (1) , X (1) is used to produce the
Lasso estimator βbL in (2.10) and the second subsample Z (2) = y (2) , X (2)
is retained to evaluate the loss kβbL − βk22 . Such sample splitting technique
is similar to cross-validation and has been used in [22] for constructing confidence sets for β and in [20] for confidence intervals for the `2 loss.
e 2 achieves the
The following proposition establishes that the√ estimator L
n
n
minimax lower bound of (2.8) over the regime log p . k . log p .
10
T. T. CAI AND Z. GUO
Proposition 1. Suppose that k . logn p and βbL is the Lasso estimator
√
defined in (2.10) with A > 2, then the estimator of loss proposed in (2.11)
satisfies, for any sequence δn,p → ∞,
1
L
2
b
e
(2.12)
lim sup sup Pθ L2 − kβ − βk2 ≥ δn,p √
= 0.
n
n,p→∞ θ∈Θ0 (k)
2.3. Minimax estimation of the `q loss over Θ(k). We now turn to the
case of unknown Σ and σ and establish the minimax lower bound for estimating the `q loss over the parameter space Θ(k).
Theorem 2. Suppose that the sparsity levels k and k0 satisfy Assumption (B1). For any estimator βb satisfying Assumption (A1) with kβ ∗ k0 ≤ k0 ,
2 log p
2
b q − kβb − βk | ≥ ck q
≥ δ, 1 ≤ q ≤ 2,
(2.13)
inf sup Pθ |L
q
n
b q θ∈Θ(k)
L
where δ > 0 and c > 0 are constants.
Theorem 2 provides a minimax lower bound for estimating the `q loss
of any estimator βb satisfying Assumption (A1), including the scaled Lasso
estimator defined as
p
(2.14)
{βbSL , σ̂} = arg
min
β∈Rp ,σ∈R+
X kX·j k2
ky − Xβk22 σ
√ |βj |,
+ + λ0
2nσ
2
n
j=1
p
√
where λ0 = A log p/n with A > 2. Note that for the scaled Lasso estimator, the lower bound in (2.13)
trivial
can be achieved by the
loss estimator
2
log
p
SL
2
b
0 in the sense, supθ∈Θ(k) Pθ |0 − kβ − βkq | ≥ Ck q n
→ 0, and hence
estimation of loss is impossible in this case.
3. Minimaxity and adaptivity of confidence intervals over Θ0 (k).
We focused in the last section on point estimation of the `q loss and showed
the impossibility of loss estimation except for one regime. The results naturally lead to another question: Is it possible to construct “useful” confidence
intervals for kβb − βk2q that can provide non-trivial upper and lower bounds
for the loss? In this section, after introducing the framework for minimaxity
and adaptivity of confidence intervals, we consider the case of known Σ = I
and σ = σ0 and establish the minimaxity and adaptivity lower bounds for
the expected length of confidence intervals for the `q loss of a broad collection of estimators over the parameter space Θ0 (k). We also show that such
HIGH-DIMENSIONAL ACCURACY ASSESSMENT
11
minimax lower bounds can be achieved for the Lasso estimator and then discuss the possibility of adaptivity using the Lasso estimator as an example.
The case of unknown Σ and σ will be the focus of the next section.
3.1. Framework for minimaxity and adaptivity of confidence intervals. In
this section, we introduce the following decision theoretical framework for
confidence intervals of the loss kβb −βk2q . Given 0 < α < 1 and the parameter
b `q the set of all (1 − α)
space Θ and the loss kβb − βk2 , denote by Iα Θ, β,
q
level confidence intervals for kβb − βk2q over Θ,
(3.1)
2
b `q = CIα β,
b `q , Z = [l (Z) , u (Z)] : inf Pθ kβb − β(θ)k ∈ CIα β,
b `q , Z
Iα Θ, β,
≥1−α .
q
θ∈Θ
b `q , Z when there is no confusion. For any
We will write CIα for CIα β,
b `q , Z = [l (Z) , u (Z)], its length is denoted by
confidence interval CIα β,
b `q , Z
L CIα β,
= u (Z) − l (Z) and the maximum expected length over
a parameter space Θ1 is defined as
b `q , Z , Θ1 = sup Eθ L CIα β,
b `q , Z .
(3.2)
L CIα β,
θ∈Θ1
b `q ,
For two nested parameter spaces Θ1 ⊆ Θ2 , we define the benchmark L∗α Θ1 , Θ2 , β,
measuring the degree of adaptivity over the nested spaces Θ1 ⊂ Θ2 ,
(3.3)
b `q =
b `q , Z .
sup Eθ L CIα β,
L∗α Θ1 , Θ2 , β,
inf
b q ,Z )∈Iα (Θ2 ,β,`
b q ) θ∈Θ1
CIα (β,`
b `q , which is the minimax
b `q for L∗ Θ1 , Θ1 , β,
We will write L∗α Θ1 , β,
α
expected length of confidence intervals of kβb − βk2q over Θ1 . The bench
b `q is the infimum of the maximum expected length
mark L∗α Θ1 , Θ2 , β,
over Θ1 among
all (1 − α)-level confidence intervals over Θ2 . In contrast,
∗
b
Lα Θ1 , β, `q is considering all (1 − α)-level confidence intervals over Θ1 .
In words, if there is priorinformation
that the parameter lies in the smaller
b `q measures the benchmark length of conparameter space Θ1 , L∗α Θ1 , β,
fidence intervals over the parameter space Θ1 , which is illustrated in the left
of Figure 1; however, if there is only prior
information
that the parameter
∗
b
lies in the larger parameter space Θ2 , Lα Θ1 , Θ2 , β, `q measures the benchmark length of confidence intervals over the parameter space Θ1 , which is
illustrated in the right of Figure 1.
12
T. T. CAI AND Z. GUO
Θ
𝐋∗ (Θ , 𝛽, ℓ𝓁 )
Θ
𝐋∗ (Θ , Θ , 𝛽, ℓ𝓁 )
ΘΩ
b `q and L∗α Θ1 , Θ2 , β,
b `q .
Fig 1. The plot demonstrates the definition of L∗α Θ1 , β,
∗
Rigorously, we define a confidence
interval
CI to be simultaneously adapb `q ,
tive over Θ1 and Θ2 if CI∗ ∈ Iα Θ2 , β,
b `q , and L (CI∗ , Θ2 ) L∗ Θ2 , β,
b `q .
(3.4) L (CI∗ , Θ1 ) L∗α Θ1 , β,
α
The condition (3.4) means that the confidence interval CI∗ has coverage over
the larger parameter space Θ2 and achieves
the minimax
rate
over both Θ1
∗
∗
∗
b
b `q
and Θ2 . Note that L (CI , Θ1 ) ≥ Lα Θ1 , Θ2 , β, `q . If Lα Θ1 , Θ2 , β,
b `q , then the rate-optimal adaptation (3.4) is impossible to achieve
L∗α Θ1 , β,
for Θ1 ⊂ Θ2 . Otherwise, it is possible to construct confidence intervals simultaneously adaptive over parameter spaces Θ1 and Θ2 . The possibility of
adaptation over parameter spaces Θ1 and
by in
Θ2 can thus be answered
∗
∗
b `q .
b `q and L Θ1 , Θ2 , β,
vestigating the benchmark quantities Lα Θ1 , β,
α
Such framework has already been introduced in [7], which studies the minimaxity and adaptivity of confidence intervals for linear functionals in highdimensional linear regression.
We will adopt the minimax and adaptationframework
discussed above
∗
b
and establish the minimax expected length Lα Θ0 (k), β, `q and the adap
b `q . In terms of the minimax extation benchmark L∗α Θ0 (k1 ), Θ0 (k2 ), β,
pected length and the adaptivity behavior, there exist fundamental differences between the case q = 2 and 1 ≤ q < 2 . We will discuss them separately
in the following two sections.
3.2. Confidence intervals for the `2 loss over Θ0 (k). The following theorem establishes the minimax lower bound for the expected length of confidence intervals of kβb − βk22 over the parameter space Θ0 (k).
Theorem 3. Suppose that 0 < α < 14 and the sparsity levels k and k0
satisfy Assumption (B1). For any estimator βb satisfying Assumption (A1)
13
HIGH-DIMENSIONAL ACCURACY ASSESSMENT
with kβ ∗ k0 ≤ k0 , then there is some constant c > 0 such that
k
log
p
1
∗
b `2 ≥ c min
(3.5)
Lα Θ0 (k), β,
,√
σ02 .
n
n
√
In particular, if βbL is the Lasso estimator defined in (2.10) with A > 2,
then the minimax expected length for (1 − α) level confidence intervals of
kβbL − βk22 over Θ0 (k) is
k log p 1
∗
L
b
σ02 .
(3.6)
Lα Θ0 (k), β , `2 min
,√
n
n
We now consider adaptivity of confidence intervals for the`2 loss. The fol
b `2 .
lowing theorem gives the lower bound for the benchmark L∗α Θ0 (k1 ), Θ0 (k2 ), β,
We will then discuss Theorems 3 and 4 together.
Theorem 4. Suppose that 0 < α < 41 and the sparsity levels k1 , k2 and
k0 satisfy Assumption (B2). For any estimator βb satisfying Assumption (A1)
with kβ ∗ k0 ≤ k0 , then there is some constant c > 0 such that
k2 log p 1
∗
b
(3.7)
Lα Θ0 (k1 ), Θ0 (k2 ), β, `2 ≥ c min
,√
σ02 .
n
n
In particular, if βbL is the Lasso estimator defined in (2.10) with A >
the above lower bound can be achieved.
√
2,
The lower bound established in Theorem 4 implies that of Theorem 3 and
both lower bounds hold for a general class of estimators satisfying Assumption(A1). There is a phase
bound of the benchmark
transition for the lower
√
n
∗
b
Lα Θ0 (k1 ), Θ0 (k2 ), β, `2 . In the regime k2 log p , the lower bound in (3.7)
√
when lognp . k2 . logn p , the lower bound in (3.7) is √1n σ02 . For the
p 2
Lasso estimator βbL defined in (2.10), the lower bound k log
n σ0 in (3.5) and
k2 log p 2
0
n σ0 in (3.7) can be achieved by the confidence intervals CIα (Z, k, 2)
0
and CIα (Z, k2 , 2) defined in (3.15), respectively. Such an interval estimator
is also used for the `q loss with 1 ≤ q < 2. The minimax lower bound √1n σ02
in (3.6) and (3.7) can be achieved by the following confidence interval,
!
!
ψ (Z)
ψ (Z)
(3.8)
CI1α (Z) = 1 2
, 1 2
− σ02
− σ02 ,
χ
χ
α (n2 )
α (n2 )
n2 1−
n2
is
k2 log p 2
n σ0 ;
2
+
2
+
14
T. T. CAI AND Z. GUO
where χ21− α (n2 ) and χ2α (n2 ) are the 1 − α2 and α2 quantiles of χ2 random
2
2
variable with n2 degrees of freedom, respectively, and
2
1
2
(2)
(2) bL
, σ0 log p .
y −X β
(3.9)
ψ (Z) = min
n2
2
Note that the two-sided confidence interval (3.8) is simply based on the
observed data Z, not depending on any prior knowledge of the sparsity k.
Furthermore, it is a two-sided confidence interval, which tells not only just an
upper bound, but also a lower bound for the loss. The coverage property and
the expected length of CI1α (Z) are established in the following proposition.
Proposition 2. Suppose k . logn p and βbL is the estimator defined in
√
(2.10) with A > 2. Then CI1α (Z) defined in (3.8) satisfies,
(3.10)
lim inf inf P kβbL − βk22 ∈ CI1α (Z) ≥ 1 − α,
n,p→∞ θ∈Θ0 (k)
and
1
L CI1α (Z) , Θ0 (k) . √ σ02 .
n
(3.11)
Θ0 𝑘1
Θ0 𝑘 2
𝑘1 log 𝑝
𝑛
𝑘1 log 𝑝
𝑛
Θ0 𝑘1
𝑘2 log 𝑝
𝑛
Θ0Ω𝑘1
Θ0 𝑘2
1
𝑛
1
𝑛
Θ0 𝑘1
Θ0 𝑘 2
1
𝑛
Θ0 Ω𝑘1
Θ0Ω𝑘1
Fig 2. Illustration of L∗α Θ0 (k1 ), βbL , `2 (top) and L∗α Θ0 (k1 ), Θ0 (k2 ), βbL , `2 (bottom)
over regimes k1 ≤ k2 .
k2 . logn p (rightmost).
√
n
log p
(leftmost), k1 .
√
n
log p
. k2 .
n
(middle)
log p
and
√
n
log p
. k1 ≤
Regarding the Lasso estimator βbL defined in (2.10), we will discuss the
possibility of adaptivity of confidence intervals for kβbL −βk22 . The adaptivity
behavior of confidence intervals for kβbL − βk22 is demonstrated in
Figure 2.
√
n
As illustrated in the rightmost plot of Figure 2, in the regime log p . k1 ≤
k2 . n , we obtain L∗ Θ0 (k2 ), Θ0 (k2 ), βbL , `2 L∗ Θ0 (k1 ), βbL , `2
log p
α
α
15
HIGH-DIMENSIONAL ACCURACY ASSESSMENT
√1 ,
n
which implies that adaptation is possible over this regime. As shown
in Proposition 2, the confidence
interval CI1α (Z) defined in (3.8) is fully
√
n
adaptive over the regime log p . k . logn p in the sense of (3.4).
Illustrated in the leftmost and middle plots of Figure 2, it is impossible
to construct an adaptive confidence interval for kβbL − βk22 over regimes k1 ≤
√
√
k2 . lognp and k1 lognp . k2 . logn p since L∗α Θ0 (k1 ), Θ0 (k2 ), βbL , `2
√
L∗α Θ0 (k1 ), βbL , `2 if k1 lognp and k1 k2 . To sum up, adaptive confidence intervals for kβbL −βk22 is only possible over the regime
√
n
log p
.k.
n
log p .
Comparison with confidence balls. We should note that the problem of
constructing confidence intervals for kβb − βk22 is related to but different from
that of constructingnconfidence sets for β itself.
o Confidence balls constructed
2
b
in [22] are of form β : kβ − βk2 ≤ un (Z) , where βb can be the Lasso estimator and un (Z) is a data dependent squared radius. See [22] for further
details. A naive application of this confidence ball leads to a one-sided confidence interval for the loss kβb − βk22 ,
n
o
(3.12)
CIinduced
(Z) = kβb − βk22 : kβb − βk22 ≤ un (Z) .
α
Due to the reason that confidence sets
n for β were sought foroin Theorem
b 2 ≤ un (Z) will suffice
1 in [22], confidence sets in the form β : kβ − βk
2
to achieve the optimal length. However, since our goal is to characterize
kβb − βk22 , we apply the unbiased risk estimation discussed in Theorem 1 of
[22] and construct the two-sided confidence interval in (3.8). Such a twosided confidence interval is more informative than the one-sided confidence
interval (3.12) since the one-sided confidence interval does not contain the
information whether the loss is close to zero or not. Furthermore, as shown in
[22], the length of confidence interval CIinduced
(Z) over the parameter space
α
k log p
1
√
Θ0 (k) is of order n + n . The two-sided confidence interval CI1α (Z)
constructed in (3.8) is of expected length
√1 + k log p
n
n
√
n
log p .
√1 ,
n
which is much shorter than
in the regime k
That is, the two-sided confidence interval
(3.8) provides a more accurate interval estimator of the `2 loss. This is
illustrated in Figure 3.
The lower bound technique developed in the literature of adaptive confidence sets [22] can also be used to establish some of the lower bound
results for the case q = 2 given in the present paper. However, new techniques are needed in order to establish the rate√sharp lower bounds for the
minimax estimation error (2.9) in the region lognp ≤ k . logn p and for the
16
T. T. CAI AND Z. GUO
CI (Z)
0
∥𝛽−𝛽 ∥
CI
(Z)
Fig 3. Comparison of the two-sided confidence interval CI1α (Z) with the one-sided confidence interval CIinduced
(Z).
α
expected
length of the confidence intervals (3.18) and (7.3) in the region
√
n
n
log p . k1 ≤ k2 . log p , where it is necessary to test a composite null against
a composite alternative in order to establish rate sharp lower bounds.
3.3. Confidence intervals for the `q loss with 1 ≤ q < 2 over Θ0 (k).
We now consider the case 1 ≤ q < 2 and investigate the minimax expected
length and adaptivity of confidence intervals for kβb −βk2q over the parameter
space Θ0 (k). The following theorem characterizes the minimax convergence
rate for the expected length of confidence intervals.
Theorem 5. Suppose that 0 < α < 41 , 1 ≤ q < 2 and the sparsity levels k
and k0 satisfy Assumption (B1). For any estimator βb satisfying Assumption
(A2) with kβ ∗ k0 ≤ k0 , then there is some constant c > 0 such that
2
b `q ≥ ck q log p σ 2 .
(3.13)
L∗α Θ0 (k), β,
n 0
√
In particular, if βbL is the Lasso estimator defined in (2.10) with A > 4 2,
then the minimax expected length for (1 − α) level confidence intervals of
kβbL − βk2q over Θ0 (k) is
2 log p
(3.14)
L∗α Θ0 (k), βbL , `q k q
σ2.
n 0
We now construct the confidence interval achieving the minimax convergence rate of (3.14),
2 log p
0
∗
q
(3.15)
CIα (Z, k, q) = 0, C (A, k)k
,
n
(
)
2
3η
where C ∗ (A, k) = max
√
√
0 Aσ
0
η0 +1
(22Aσ0 )2
4 ,
4
q
q
2k log p
1
1
−42
−(9+11η0 ) 2knlog p
4
n
4
1
with η0 =
1
√2 . The following proposition establishes the coverage property and
1.01 √A+
A− 2
the expected length of CI0α (Z, k, q).
HIGH-DIMENSIONAL ACCURACY ASSESSMENT
17
Proposition 3. Suppose k . logn p and βbL is the estimator defined in
√
(2.10) with A > 4 2. For 1 ≤ q ≤ 2, the confidence interval CI0α (Z, k, q)
defined in (3.15) satisfies
(3.16)
lim inf inf Pθ kβb − βk2q ∈ CI0α (Z, k, q) = 1,
n,p→∞ θ∈Θ0 (k)
and
(3.17)
2 log p
L CI0α (Z, k, q) , Θ0 (k) . k q
σ2.
n 0
In particular, for the case q = 2,√(3.16) and (3.17) also hold for the estimator
βbL defined in (2.10) with A > 2.
This result shows that the confidence interval CI0α (Z, k, q) achieves the
minimax rate given in (3.14). In contrast to the `2 loss where the two-sided
confidence interval (3.8) is significantly shorter than
the one-sided interval
√
and achieves the optimal rate over the regime lognp . k . logn p , for the `q
loss with 1 ≤ q < 2, the one-sided confidence interval achieves the optimal
rate given in (3.14).
We now consider adaptivity of confidence
intervals. Thefollowing theorem
b `q with 1 ≤ q < 2.
establishes the lower bound for L∗α Θ0 (k1 ), Θ0 (k2 ), β,
Theorem 6. Suppose 0 < α < 41 , 1 ≤ q < 2 and the sparsity levels k1 , k2
and k0 satisfy Assumption (B2). For any estimator βb satisfying Assumption
(A2) with kβ ∗ k0 ≤ k0 , then there is some constant c > 0 such that
(3.18)
2
√
n
q log p 2
σ
if
k
≤
k
.
ck
1
2
2 n 0
log p ;
2
√
n
b `q ≥ ck q −1 √1 σ 2
n
L∗α Θ0 (k1 ), Θ0 (k2 ), β,
if
k
.
1
2
log p . k2 . log p ;
n 0
√
2q −1 log p 2
ck2 k1 n σ0 if lognp . k1 ≤ k2 . logn p .
In particular,
if p ≥ n and βbL is the Lasso estimator defined in (2.10) with
√
A > 4 2, the above lower bounds can be achieved.
The lower bounds of Theorem 6 imply that of Theorem 5 and both lower
bounds hold for a general class of estimators satisfying Assumption (A2).
However, the lower bound (3.18) in Theorem 6 has a significantly different meaning from (3.13) in Theorem 5 where (3.18) quantifies the cost
of adaptation without knowing the sparsity level. For the Lasso estimator
18
T. T. CAI AND Z. GUO
βbL defined in (2.10), by comparing
Theorem 5 and
Theorem 6, we obtain
∗
L
∗
L
b
b
Lα Θ0 (k1 ), Θ0 (k2 ), β , `q Lα Θ0 (k1 ), β , `q if k1 k2 , which implies
the impossibility of constructing adaptive confidence intervals for the case
1 ≤ q < 2. There exists marked difference between the case 1 ≤ q < 2 and
the case q = 2, where
it is possible to construct adaptive confidence intervals
√
n
over the regime log p . k . logn p .
For the Lasso estimator βbL defined in (2.10), it is shown in Proposition 3
that the confidence interval CI0α (Z, k2 , q) defined in (3.15) achieves the lower
2
2
2
−1
−1
bound k2q logn p σ02 of (3.18). The lower bounds k2q k1 logn p σ02 and k2q √1n σ02
of (3.18) can be achieved by the following proposed confidence interval,
(3.19)
!
!
2
ψ
(Z)
ψ
(Z)
−1
CI2α (Z, k2 , q) = 1 2
− σ02
, (16k2 ) q
− σ02 ,
1 2
(n
)
(n
)
χ
χ
2
2
n2 1− α
n2 α
+
2
+
2
where ψ (Z) is given in (3.9). The following result verifies the above claim.
Proposition 4. Suppose p ≥ n, k1 ≤ k2 . logn p and βbL is defined in
√
(2.10) with A > 4 2. Then CI2α (Z, k2 , q) defined in (3.19) satisfies,
(3.20)
lim inf inf Pθ kβb − βk2q ∈ CI2α (Z, k2 , q) ≥ 1 − α,
n,p→∞ θ∈Θ0 (k2 )
and
(3.21)
L
CI2α (Z, k2 , q) , Θ0 (k1 )
2
−1
q
. k2
1
log p
+√
k1
n
n
σ02 .
4. Minimaxity and adaptivity of confidence intervals over Θ(k).
In this section, we focus on the case of unknown Σ and σ and establish the
rates of convergence for the minimax expected length of confidence intervals
for kβb − βk2q with 1 ≤ q ≤ 2 over Θ(k) defined in (2.4). We also study
the possibility of adaptivity of confidence intervals for kβb − βk2q . The following
establishes the lower bounds
for the benchmark
quantities
theorem
∗
∗
b
b
Lα Θ (ki ) , β, `q with i = 1, 2 and Lα Θ (k1 ) , Θ (k2 ) , β, `q .
Theorem 7. Suppose that 0 < α < 14 , 1 ≤ q ≤ 2 and the sparsity
levels k1 , k2 and k0 satisfy Assumption (B2). For any estimator βb satisfying
Assumption (A1) at θ0 = (β ∗ , I, σ0 ) with kβ ∗ k0 ≤ k0 , there is a constant
c > 0 such that
2
b `q ≥ ck q log p , for i = 1, 2;
(4.1)
L∗α Θ (ki ) , β,
i
n
HIGH-DIMENSIONAL ACCURACY ASSESSMENT
(4.2)
19
2
b `q ≥ ck q log p .
L∗α {θ0 } , Θ (k2 ) , β,
2
n
In particular,
if βbSL is the scaled Lasso estimator defined in (2.14) with
√
A > 2 2, then the above lower bounds can be achieved.
The lower bounds (4.1) and (4.2) hold for any βb satisfying Assumption
(A1) at an interior point θ0 , including the scaled Lasso estimator as a special case. We demonstrate the impossibility of adaptivity of confidence inbSL defined in (2.14).
tervals for the `q loss of the scaled
Lassoestimator β
Since L∗α Θ (k1 ) , Θ (k2 ) , βbSL , `q ≥ L∗α {θ0 } , Θ (k2 ) , βbSL , `q , by (4.2),
we have L∗α Θ (k1 ) , Θ (k2 ) , βbSL , `q L∗α Θ (k1 ) , βbSL , `q if k1 k2 .
The comparison of L∗α Θ (k1 ) , βbSL , `q and L∗α Θ (k1 ) , Θ (k2 ) , βbSL , `q is
illustrated in Figure 4. Referring to the adaptivity defined in (3.4), it is
impossible to construct adaptive confidence intervals for kβbSL − βk2q .
Θ 𝑘
Θ 𝑘
𝑘 log 𝑝
𝑛
ΘΩ
𝑘
𝑘 log 𝑝
𝑛
Fig 4. Illustration of L∗α Θ (k1 ) , βbSL , `q (left) and L∗α Θ (k1 ) , Θ (k2 ) , βbSL , `q (right).
b `q , Z for the
Theorem 7 shows that for any confidence interval CIα β,
b
loss of any estimator
Assumption
(A1),
under the coverage
β satisfying
b
b
constraint that CIα β, `q , Z ∈ Iα Θ (k2 ) , β, `q , its expected length at
2
any given θ0 = (β ∗ , I, σ) ∈ Θ (k0 ) must be of order k2q logn p . In contrast to
Theorem 4 and 6, Theorem 7 demonstrates that confidence intervals must
be long at a large subset of points in the parameter space, not just at a small
number of “unlucky” points. Therefore, the lack of adaptivity for confidence
intervals is not due to the conservativeness of the minimax framework.
In the following, we detail the construction of confidence intervals for
b
kβ SL −βk2q . The construction of confidence intervals is based on the following
20
T. T. CAI AND Z. GUO
definition of restricted eigenvalue, which is introduced in [4],
(4.3)
κ(X, k, s, α0 ) =
min
√
min
δ6=0,
J0 ⊂{1,··· ,p},
kδJ c k1 ≤α0 kδJ0 k1
|J0 |≤k
0
kXδk2
,
nkδJ01 k2
where J1 denotes the subset corresponding to the s largest in absolute value
coordinates of δ outside of J0 and J01 = J0 ∪ J1 . Define the event B =
{σ̂ ≤ log p} . The confidence interval for kβbSL − βk2q is defined as
(4.4)
CIα (Z, k, q) =
on B
on B c ,
[0, ϕ (Z, k, q)]
{0}
where
2
2
2 log p
2 log p
16Amax kX·j k2 σ
b
k q
ϕ (Z, k, q) = min
, kq
log p σ
b2 .
max
kX
k
·j 2
nκ2 X, k, k, 3
n
n
min kX·j k2
Properties of CIα (Z, k, q) are established as follows.
Proposition 5. Suppose k . logn p and βbSL is the estimator defined in
√
(2.14) with A > 2 2. For 1 ≤ q ≤ 2, then CIα (Z, k, q) defined in (4.4)
satisfies the following properties,
(4.5)
lim inf inf Pθ kβb − βk2q ∈ CIα (Z, k, q) = 1,
n,p→∞ θ∈Θ(k)
and
(4.6)
2
L (CIα (Z, k, q) , Θ (k)) . k q
log p
.
n
Proposition 5 shows that the confidence interval CIα (Z, ki , q) defined in
(4.4) achieves the lower bound in (4.1), for i = 1, 2, and the confidence
interval CIα (Z, k2 , q) defined in (4.4) achieves the lower bound in (4.2).
5. Estimation of the `q loss of rate-optimal estimators. We have
established the minimax lower bounds for the estimation accuracy of the
loss of a broad class of estimators βb satisfying the weak assumptions (A1)
or (A2) and also demonstrated that such minimax lower bounds are sharp
for the Lasso estimator or scaled Lasso estimator. In this section, we will
show that the minimax lower bounds are sharp for the class of rate-optimal
estimators satisfying the following Assumption (A).
HIGH-DIMENSIONAL ACCURACY ASSESSMENT
(A) The estimator βb satisfies, for k
(5.1)
21
n
log p ,
2
q log p
2
∗
b
≤ Cp−δ ,
sup Pθ kβ − βkq ≥ C kβk0
n
θ∈Θ(k)
for constants δ > 0, C ∗ > 0 and C > 0.
We say an estimator βb is rate-optimal if it satisfies Assumption (A). As
shown in [12, 4, 3, 26], Lasso, Dantzig Selector, scaled Lasso and squareroot Lasso are rate-optimal when the tuning parameter is chosen properly.
We shall stress that Assumption (A) implies Assumptions (A1) and (A2).
Assumption (A) requires the estimator βb to perform well over the whole
parameter space Θ(k) while Assumptions (A1) and (A2) only require βb
to perform well at a single point or over a proper subset. The following
proposition shows that the minimax lower bounds established in Theorem
1 to Theorem 7 can be achieved for the class of rate-optimal estimators.
Proposition 6.
Let βb be an estimator satisfying Assumption (A).
1. There exist estimators of the loss kβb − βk2q with 1 ≤ q < 2 achieving,
up to a constant factor, the minimax lower bounds (2.9) in Theorem 1
2 with 1 ≤ q ≤ 2
b
and (3.13) in Theorem 5 and estimators of loss kβ−βk
q
achieving, up to a constant factor, the minimax lower bounds (2.13)
in Theorem 2 and (4.1) in Theorem 7.
2. Suppose that the estimator
βb is constructed based on the subsample
(1)
(1)
(1)
Z
= y ,X
, then there exist estimators of the loss kβb − βk22
achieving, up to a constant factor, the minimax lower bounds (2.8) in
Theorem 1, (3.5) in Theorem 3 and (3.7) in Theorem 4.
(1)
b
3. Suppose the
estimator β is constructed based on the subsample Z =
(1)
(1)
y ,X
and it satisfies Assumption (A) with δ > 2 and the assumption k(βb − β)S c k1 ≤ c∗ k(βb − β)S k1 where S = supp(β). Then
for p ≥ n there exist estimators of the loss kβb − βk2q with 1 ≤ q < 2
achieving the minimax lower bounds (3.18) in Theorem 6.
For reasons of space, we do not discuss the detailed construction of confidence intervals achieving these minimax lower bounds here and postpone
the construction to the proof of Proposition 6.
Remark 2. Sample splitting has been widely used in the literature. For
(1)
b
example, the
condition that β is constructed based on the subsample Z =
(1)
(1)
y ,X
has been introduced in [22] for constructing confidence sets for
β and in [20] for constructing confidence intervals for the `2 loss. Such a
22
T. T. CAI AND Z. GUO
condition is imposed purely for technical reasons to create the independence
between the estimator βb and the subsample Z (2) = y (2) , X (2) , which is used
b As shown in [4], the assumption
to evaluate the `q loss of the estimator β.
k(βb − β)S c k1 ≤ c∗ k(βb − β)S k1 is satisfied for Lasso and Dantzig Selector.
6. General tools for minimax lower bounds. A major step in our
analysis is to establish rate sharp lower bounds for the estimation error and
expected length of confidence intervals for the `q loss. We introduce in this
section new technical tools that are needed to establish these lower bounds.
A significant distinction of the lower bound results given in the previous
sections from those for the traditional parameter estimation problems is that
the constraint is on the performance of the estimator βb of the regression
vector β, but the lower bounds are on the difficulty of estimating its loss
kβb − βk2q . It is necessary to develop new lower bound techniques to establish
rate-optimal lower bounds for the estimation error and the expected length
of confidence intervals for the loss kβb − βk2q . These technical tools may also
be of independent interest.
We begin with notation. Let Z denote a random variable whose distribution is indexed by some parameter θ ∈ Θ and let π denote a prior
on the parameter space Θ. We will use fθ (z) to denote the density of Z
given θ and fπ (z) to denote the marginal density of Z under the prior π.
Let
R Pπ denote the distribution of Z corresponding to fπ (z), i.e., Pπ (A) =
1z∈A fπ (z) dz, where 1z∈A is the indicator function. For a function g, we
write
Eπ (g(Z)) for the expectation
R
R under fπ . More specifically, fπ (z) =
fθ (z) π (θ) dθ and Eπ (g(Z)) = g (z) fπ (z) dz. The L1 distance between
two
R probability distributions with densities f0 and f1 is given by L1 (f1 , f0 ) =
|f1 (z) − f0 (z)| dz. The following theorem establishes the minimax lower
bounds for the estimation error and the expected length of confidence intervals for the `q loss, under the constraint that βb is a good estimator at at
least one interior point.
Theorem 8. Suppose 0 < α, α0 < 41 , 1 ≤ q ≤ 2, Σ0 is positive definite,
θ0 = (β ∗ , Σ0 , σ0 ) ∈ Θ, and F ⊂ Θ. Define d = minθ∈F kβ (θ) − β ∗ kq . Let π
denote a prior over the parameter space F. If an estimator βb satisfies
1 2
∗ 2
b
(6.1)
Pθ0 kβ − β kq ≤ d ≥ 1 − α0 ,
16
then
(6.2)
1 2
2
b
b
inf sup Pθ |Lq − kβ − βkq | ≥ d ≥ c̄1 ,
4
b q θ∈{θ0 }∪F
L
HIGH-DIMENSIONAL ACCURACY ASSESSMENT
23
and
(6.3)
b `q =
L∗α {θ0 } , Θ, β,
b `q , Z
inf
Eθ0 L CIα β,
≥ c∗2 d2 ,
b
b
CIα (β,`q ,Z )∈Iα (Θ,β,`q )
n
o
1
9
where c̄1 = min 10
, 10
− α0 − L1 (fπ , fθ0 ) + and c∗2 = 12 (1 − 2α − α0 − 2L1 (fπ , fθ0 ))+ .
Remark 3. The minimax lower bound (6.2) for the estimation error
and (6.3) for the expected length of confidence intervals hold as long as the
estimator βb estimates β well at an interior point θ0 . Besides Condition (6.1),
another key ingredient for the lower bounds (6.2) and (6.3) is to construct
the least favorable space F with the prior π such that the marginal distributions fπ and fθ0 are non-distinguishable. For the estimation lower bound
(6.2), constraining that kβb − β ∗ k2q can be well estimated at θ0 , due to the
non-distinguishability between fπ and fθ0 , we can establish that the loss
kβb − βk2q cannot be estimated well over F. For the lower bound (6.3), by
Condition (6.1) and the non-distinguishability between fπ and fθ0 , we will
show that kβb − βk2q over F is much larger than kβb − β ∗ k2q and hence the
honest confidence intervals must be sufficiently long.
Theorem 8 is used to establish the minimax lower bounds for both the
estimation error and the expected length of confidence intervals of the `q loss
over Θ(k). By taking θ0 ∈ Θ(k0 ) and Θ = Θ(k), Theorem 2 follows from
(6.2) with a properly constructed subset F ⊂ Θ(k). By taking θ0 ∈ Θ(k0 )
and Θ = Θ(k2 ), Theorem 7 follows from (6.3) with a properly constructed
F ⊂ Θ(k2 ). In both cases, Assumption (A1) implies Condition (6.1).
Several minimax lower bounds over Θ0 (k) can also be implied by Theorem
8. For the estimation
error, the minimax lower bounds (2.8) and (2.9) over
√
n
the regime k . log p in Theorem 1 follow from (6.2). For the expected length
of confidence intervals, the minimax
lower bounds
(3.7) in Theorem 4 and
√
√
n
n
(3.18) in the regions k1 ≤ k2 . log p and k1 . log p . k2 . logn p in Theorem
6 follow from (6.3). In these cases, Assumption (A1) or (A2) can guarantee
that Condition (6.1) is satisfied. However,
the minimax lower bound for esti√
n
mation error (2.9) in the region log p ≤ k . logn p and for the expected length
√
of confidence intervals (3.18) in the region lognp . k1 ≤ k2 . logn p cannot
be established using the above theorem. The following theorem, which requires testing a composite null against a composite alternative, establishes
the refined minimax lower bounds over Θ0 (k).
24
T. T. CAI AND Z. GUO
Theorem 9. Let 0 < α, α0 < 41 , 1 ≤ q ≤ 2, and θ0 = (β ∗ , Σ0 , σ0 ) where
Σ0 is a positive definite matrix. Let k1 and k2 be two sparsity levels. Assume
that for i = 1, 2 there exist parameter spaces Fi ⊂ {(β, Σ0 , σ0 ) : kβk0 ≤ ki }
such that for given disti and di
kΣ0 (β (θ) − β ∗ )k2 = disti
and
kβ (θ) − β ∗ kq = di ,
for all θ ∈ Fi .
Let πi denote a prior over the parameter space Fi for i = 1, 2. Suppose
that for θ1 = β ∗ , Σ0 , σ02 + dist21 and θ2 = β ∗ , Σ0 , σ02 + dist22 , there exist
constants c1 , c2 > 0 such that
∗ 2
2 2
b
(6.4)
Pθi kβ − β kq ≤ ci di ≥ 1 − α0 , for i = 1, 2.
Then we have
(6.5)
inf
sup
b q θ∈F1 ∪F2
L
b q − kβb − βk2q | ≥ c∗3 d22 ≥ c̄3 ,
Pθ |L
and
(6.6)
b `q ≥ c∗ (1 − c2 )2 d2 − (1 + c1 )2 d2 ,
L∗α Θ0 (k1 ) , Θ0 (k2 ) , β,
4
2
1
+
P
d2
, c∗4 = 1 − 2α0 − 2α − 2i=1 L1 (fπi , fθi ) − 2L1 (fπ2 , fπ1 ) +
(1 − c2 )2 − 41 − (1 + c1 )2 d12
2 +
n
o
P
1
9
and c̄3 = min 10
, 10
− 2α0 − 2i=1 L1 (fπi , fθi ) − 2L1 (fπ2 , fπ1 ) + .
where c∗3 = min
1
,
4
Remark 4. As long as the estimator βb performs well at two points, θ1
and θ2 , the minimax lower bounds (6.5) for the estimation error and (6.6) for
the expected length of confidence intervals hold. Note that θi in the above
theorem does not belong to the parameter space {(β, Σ0 , σ0 ) : kβk0 ≤ ki }, for
i = 1, 2. In contrast to Theorem 8, Theorem 9 compares composite hypotheses F1 and F2 , which will lead to a sharper lower bound than comparing the
simple null {θ0 } with the composite alternative F. For simplicity, we construct least favorable parameter spaces Fi such that the points in Fi is of
fixed `2 distance and fixed `q distance to β ∗ , for i = 1, 2, respectively. More
importantly, we construct F1 with the prior π1 and F2 with the prior π2 such
that fπ1 and fπ2 are not distinguishable, where θ1 and θ2 are introduced to
facilitate the comparison. By Condition (6.4) and the construction of F1
and F2 , we establish that the `q loss cannot be simultaneously estimated
well over F1 and F2 . For the lower bound (6.6), under the same conditions,
it is shown that the `q loss over F1 and F2 are far apart and any confidence interval with guaranteed coverage probability over F1 ∪ F2 must be
HIGH-DIMENSIONAL ACCURACY ASSESSMENT
25
sufficiently long. Due to the prior information Σ = I and σ = σ0 , the lower
bound construction over Θ0 (k) is more involved than that over Θ(k). We
shall stress that the construction of F1 and F2 and the comparison between
composite hypotheses are of independent interest.
The minimax lower bound (2.9) in the region
√
n
log p
n
follows from
√log p
n
log p . k1 ≤ k2 .
.k.
(6.5) and the minimax lower bound (3.18) in the region
n
log p for the expected length of confidence intervals follows from (6.6). In
these cases, Σ0 is taken as I and Assumption (A2) implies Condition (6.4).
7. An intermediate setting with known σ = σ0 and unknown Σ.
The results given in Sections 3 and 4 show the significant difference between
Θ0 (k) and Θ(k) in terms of minimaxity and adaptivity of confidence intervals
for kβb − βk2q . Θ0 (k) is for the simple setting with known design covariance
matrix Σ = I and known noise level σ = σ0 , and the Θ(k) is for unknown
Σ and σ. In this section, we further consider minimaxity and adaptivity of
confidence intervals for kβb − βk2q in an intermediate setting where the noise
level σ = σ0 is known and Σ is unknown but of certain structure. Specifically,
we consider the following parameter space,
(7.1)
1
kβk0 ≤ k,
≤ λmin (Σ) ≤ λmax (Σ) ≤ M1
M1
Θσ0 (k, s) = (β, Σ, σ0 ) :
,
kΣ−1 kL1 ≤ M, max k Σ−1 i· k0 ≤ s
1≤i≤p
for some constants M1 ≥ 1 and M > 0. Θσ0 (k, s) basically assumes known
noise level σ and imposes sparsity conditions on the precision matrix of the
random design. This parameter space is similar to those used in the literature
of sparse linear regression with random design [29, 13, 14]. Θσ0 (k, s) has two
sparsity parameters where k represents the sparsity of β and s represents
the maximum row sparsity of the precision matrix Σ−1 . Note that Θ0 (k) ⊂
Θσ0 (k, s) ⊂ Θ(k) and Θ0 (k) is a p
special case of Θσ0 (k, s) with M1 = 1.
n/ log p, the minimaxity and adaptivity
Under the assumption s
lower bounds for the expected length of confidence intervals for kβb − βk2q
with 1 ≤ q < 2 over Θσ0 (k, s) are the same as those over Θ0 (k). That
is, Theorems 5 and 6 hold with Θ0 (k1 ), Θ0 (k2 ), and Θ0 (k) replaced by
Θσ0 (k1 , s), Θσ0 (k2 , s), and Θσ0 (k, s), respectively. For the case q = 2, the
following theorem establishes the minimaxity and adaptivity lower bounds
for the expected length of confidence intervals for kβb − βk22 over Θσ0 (k, s).
26
T. T. CAI AND Z. GUO
p
Theorem 10. Suppose 0 < α, α0 < 1/4, M1 > 1, s n/log p and
the sparsity levels k1 , k2 and k0 satisfy Assumption (B2) with the constant
c0 replaced by c∗0 defined in (9.14). For any estimator βb satisfying
2
∗ 2
∗
∗ q log p 2
b
(7.2)
sup Pθ kβ − β kq ≥ C kβ k0
σ ≤ α0 ,
n
θ∈Θ(k0 )
with a constant C ∗ > 0, then there is some constant c > 0 such that
(7.3)
log p
log p 1
∗
b
√
, max k1
,
σ02
Lα Θσ0 (k1 , s), Θσ0 (k2 , s), β, `2 ≥ c min k2
n
n
n
and
b `2 ≥ c ki log p σ 2
L∗α Θσ0 (ki , s), β,
0
n
(7.4)
and
i = 1, 2.
In particular,
if p ≥ n and βb is constructed based on the subsample Z (1) =
y (1) , X (1) and satisfies Assumption (A) with δ > 2, the above lower bounds
can be attained.
In contrast to Theorems 3 and 4, the lower bounds for the case q = 2
change in the absence of the prior knowledge Σ = I but the possibility of
adaptivity of confidence intervals over Θσ0 (k, s) is similar to that
√ over Θ0 (k).
L
b
Since the Lasso estimator β defined in (2.10) with A > 4 2 satisfies Assumption (A) with δ > 2, by Theorem 10, the minimax lower
bounds (7.3)
√
n
L
L
b
b
and (7.4) can be attained for β . For β , only when log p . k1 ≤ k2 .
n
b `2 k1 log p
, L∗ Θσ (k1 , s), βbL , `2 L∗ Θσ (k1 , s), Θσ (k2 , s), β,
log p
α
0
α
0
0
n
and adaptation between
s) is possible. In other regimes,
Θσ0 (k1 , s) andΘσ0 (k2 ,
∗
L
∗
b
b `2
if k1 k2 , then Lα Θσ0 (k1 , s), β , `2 Lα Θσ0 (k1 , s), Θσ0 (k2 , s), β,
and adaptation between Θσ0 (k1 , s) and Θσ0 (k2 , s) is impossible. For reasons
of space, more discussion on Θσ0 (k, s), including
the construction of adap√
n
tive confidence intervals over the regime log p . k1 ≤ k2 . logn p , is postponed
to the supplement [6].
8. Minimax lower bounds for estimating kβk2q with 1 ≤ q ≤
2. The lower bounds developed in this paper have broader implications.
In particular, the established results imply the minimax lower bounds for
estimating kβk2q and the expected length of confidence intervals for kβk2q
with 1 ≤ q ≤ 2. To build the connection, it is sufficient to note that the
trivial estimator βb = 0 satisfies Assumptions (A1) and (A2) with β ∗ = 0.
HIGH-DIMENSIONAL ACCURACY ASSESSMENT
27
Then we can apply the lower bounds (2.8), (2.9) and (2.13) to the estimator
βb = 0 and establish the minimax lower bounds of estimating kβk2q ,
(8.1)
log p 1
2
b
σ02 ≥ δ;
inf sup Pθ |L2 − kβk2 | ≥ c min k
,√
n
n
b 2 θ∈Θ0 (k)
L
(8.2)
2 log p
2
2
b
q
inf sup Pθ |Lq − kβkq | ≥ ck
σ ≥ δ,
n 0
b q θ∈Θ0 (k)
L
(8.3)
2 log p
2
b
q
inf sup Pθ |Lq − kβkq | ≥ ck
≥ δ,
n
b q θ∈Θ(k)
L
for 1 ≤ q < 2,
for 1 ≤ q ≤ 2,
for some constants δ > 0 and c > 0. Similarly, all the lower bounds for the
expected length of confidence intervals for kβb−βk2q established in Theorem 3
to Theorem 7 imply corresponding lower bounds for kβk2q . The lower bound
min{k logn p , √1n }σ02 in (8.1) is the same as the detection boundary in the
sparse linear regression for the case Σ = I and σ = 1; See [19] and [1] for
more details. Estimation of kβk22 in high-dimensional linear regression has
been considered in [17] under the general setting where Σ and σ are unknown
and the lower bound (8.3) with q = 2 leads to one key component of the
lower bound ck logn p for estimating kβk22 .
9. Proofs. In this section, we present the proofs of the lower bound
results. In Section 9.1, we establish the general lower bound result, Theorem
8. By applying Theorem 8 and Theorem 9, we establish Theorems 4 and 6
in Section 9.2. For reasons of space, the proofs of Theorems 1, 2, 3, 5, 7, 9,
10, the upper bound results, including Propositions 1, 2, 3, 4, 5, 6 and the
proofs of technical lemmas are postponed to the supplement [6].
We define the χ2 distance between two density functions f1 and f0 by
2
R
R f 2 (z)
0 (z))
χ2 (f1 , f0 ) = (f1 (z)−f
dz = f10 (z) dz − 1, and it is well known that
f0 (z)
(9.1)
L1 (f1 , f0 ) ≤
p
χ2 (f1 , f0 ).
Let PZ,θ∼π denote the joint probability of Z and θ with the joint density
function f (θ, z) = fθ (z) π (θ) . We introduce the following lemma, which is
used in the proofs of Theorem 8 and Theorem 9. The proof of this lemma
can be found in the supplement [6].
28
T. T. CAI AND Z. GUO
Lemma 1.
For any event A, we have
(9.2)
Pπ (Z ∈ A) = PZ,θ∼π (Z ∈ A) ,
(9.3)
|Pπ1 (Z ∈ A) − Pπ2 (Z ∈ A)| ≤ L1 (fπ2 , fπ1 ) .
We will write Pπ (A) and PZ,θ∼π (A) for Pπ (Z ∈ A) and PZ,θ∼π (Z ∈ A)
b q (Z) denotes a data-dependent loss estimator and
respectively. Recall that L
β(θ) denotes the corresponding β of the parameter θ.
1
9.1. Proof of Theorem 8. We set c0 = 14 and α1 = 10
.
Proof of (6.2)
We assume
1 2
∗ 2
b
b
(9.4)
Pθ0 Lq (Z) − kβ(Z) − β kq ≤ d ≥ 1 − α1 .
4
Otherwise, we have
1
2
∗
2
b
b q (Z) − kβ(Z)
(9.5)
Pθ0 L
− β kq ≥ d ≥ α1 .
4
and hence (6.2) follows. Define the event
1
∗
2
2
2
∗
2
2
b −β k ≤c d , L
b −β k ≤ d .
b q (z) − kβ(z)
(9.6) A0 = z : kβ(z)
q
0
q
4
By (6.1) and (9.4), we have Pθ0 (A0 ) ≥ 1 − α0 − α1 . By (9.3), we obtain
Z
(9.7)
Pπ (A0 ) ≥ 1 − α0 − α1 − |fθ0 (z) − fπ (z)| dz.
For z ∈ A0 and θ ∈ F, by triangle inequality,
(9.8)
b − β(θ)kq ≥ kβ(θ) − β ∗ kq − kβ(z)
b − β ∗ kq ≥ (1 − c0 ) d.
kβ(z)
b − β (θ) k2 ≥ kβ(z)
b − β (θ) k2 − kβ(z)
b − β ∗ k2 −
b q (z) − kβ(z)
For z ∈ A0 and θ ∈ F, then L
q
q
q
b − β ∗ k2 ≥ (1 − 2c0 − 1 )d2 , where the first inequality follows
b q (z) − kβ(z)
L
q
4
from triangle inequality and the last inequality follows from (9.6) and (9.8).
Hence, for z ∈ A0 , we obtain
(9.9)
b − β (θ) k2 ≥ (1 − 2c0 − 1 )d2 .
b q (z) − kβ(z)
inf L
q
θ∈F
4
HIGH-DIMENSIONAL ACCURACY ASSESSMENT
29
b
b q (Z) − kβ(Z)
Note that supθ∈F Pθ L
− β (θ) k2q ≥ (1 − 2c0 − 14 )d2 ≥
b
b q (Z) − kβ(Z)
supθ∈F Pθ inf θ∈F L
− β (θ) k2q ≥ (1 − 2c0 − 14 )d2 . Since the
max risk is lower bounded
by the Bayesian risk, we can further lower bound
b
b q (Z) − kβ(Z)
the last term by Pπ inf θ∈F L
− β (θ) k2q ≥ (1 − 2c0 − 14 )d2 .
Combined with (9.9), we establish
1 2
2
b
b
(9.10) sup Pθ Lq (Z) − kβ(Z) − β (θ) kq ≥ (1 − 2c0 − )d ≥ Pπ (A0 ).
4
θ∈F
Combining (9.5), (9.7) and (9.10), we establish (6.2).
Proof of (6.3)
b `q , Z ∈ Iα Θ, β,
b `q , we have
For CIα β,
(9.11)
b
b `q , Z
− β (θ) k2q ∈ CIα β,
inf Pθ kβ(Z)
≥ 1 − α.
θ∈Θ
n
o
b − β ∗ kq < c0 d, kβ(z)
b − β ∗ k2 ∈ CIα β,
b L, z .
Define the event A = z : kβ(z)
q
By (6.1) and (9.11), we have Pθ0 (A) ≥ 1 − α − α0 . (9.2) and (9.3) imply
PZ,θ∼π (A) = Pπ (A) ≥ 1 − α − α0 − L1 (fπ , fθ0 ) .
n
o
b − β (θ) k2 ∈ CIα β,
b `q , z
Define the event Bθ = z : kβ(z)
and M =
q
∪θ∈F Bθ . By (9.11), we have
Z Z
Z Z
PZ,θ∼π (M) =
1z∈M fθ (z)dz π (θ) dθ ≥
1z∈Bθ fθ (z)dz π (θ) dθ ≥ 1−α.
(9.12)
Combined with (9.12), we have PZ,θ∼π (A ∩ M) ≥ 1 − 2α − α0 − L1(fπ , fθ0) .
b − β(θ̄)k2 ∈ CIα β,
b `q , z ;
For z ∈ M, there exists θ̄ ∈ F such that kβ(z)
q
b − β ∗ k2 ∈ CIα β,
b `q , z and kβ(z)
b − β ∗ kq < c0 d.
For z ∈ A, we have kβ(z)
q
b `q , z
b − β(θ̄)k2 , kβ(z)
b − β ∗ k2 ∈ CIα β,
Hence, for z ∈ A ∩ M, we have kβ(z)
q
q
∗
∗
b
b
and kβ(z) − β(θ̄)kq ≥ kβ(θ̄) − β kq − kβ(z) − β kq ≥ (1 − c0 ) d and hence
(9.13)
b `q , z
L CIα β,
≥ (1 − 2c0 ) d2 .
n
o
b `q , z
≥ (1 − 2c0 ) d2 . By (9.13), we
Define the event C = z : L CIα β,
have Pπ (C) = PZ,θ∼π (C) ≥ PZ,θ∼π (A ∩ M) ≥ 1 − 2α − α0 − L1 (fπ , fθ0 ) . By
(9.3), we establish Pθ0 (C) ≥ 1 − 2α − α0 − 2L1 (fπ , fθ0 ) and hence (6.3).
30
T. T. CAI AND Z. GUO
9.2. Proof of Theorems 4 and 6. We first specify some constants used in
0
and
the proof. Let C ∗ be given in (2.6). Define 1 = 1−2α−2α
12
(9.14)
(
c0 = min
2
1
, 32 log 1 + 21 ,
2
3
q
log(1 + 21 ),
1 − 2γ
,
16C ∗
1 − 2γ
16C ∗
2 )
, c∗0 = min c0 ,
√
M1 − 1
√
C ∗ M1 + M1 − 1
Theorems 4 and 6 follow from Theorem 11 below.
Theorem 11. Suppose 0 < α < 14 , 1 ≤ q ≤ 2 and the sparsity levels
k1 , k2 and k0 satisfy Assumption (B2). Suppose that βb satisfies Assumption
(A2) with kβ ∗ k0 ≤ k0 .
1. If k2 .
√
n
log p ,
then there is some constant c > 0 such that
2
b `q ≥ ck q log p σ 2 .
L∗α Θ0 (k1 ) , Θ0 (k2 ) , β,
2
n 0
(9.15)
√
2. If lognp . k2 . logn p , then there is some constant c > 0 such that
(9.16)
b `q
L∗α Θ0 (k1 ) , Θ0 (k2 ) , β,
2
−1
q
2
2
−1
log p
log p
k2
σ02 ,
≥ c max
− (1 + c1 )2 k1q
, √
(1 − c2 )2 k2q k1
n
n
n
+
1
where c1 =
C ∗ M1 k0q
1
(k1 −k0 ) q
1
and c2 =
C ∗ k0q
1−1
1
2 (k1 −k0 ) 2
.
M1 (k2 −k0 ) q
2 −1
k2q
√
In particular, the minimax lower bound (9.15) and the term n σ02 in (9.16)
can be established under the weaker assumption (A1) with kβ ∗ k0 ≤ k0 .
By Theorem 11, we
establish (3.7) in Theorem 4 and (3.18) in Theorem 6.
√
n
In the regime k2 . log p , the lower bound (3.7) for q = 2 and (3.18) for 1 ≤
√
q < 2 follow from (9.15). For the case q = 2, in the regime lognp . k2 . logn p ,
the first term of the right hand side of (9.16) is 0 while the second term is √1n ,
which leads to (3.7). For 1 ≤ q < 2, let k1∗ = min{k
1 , ζ0 k2 } for some constant
∗
b `q ≥
0 < ζ0 < 1, an application of (9.16) leads to Lα Θ0 (k1∗ ) , Θ0 (k2 ) , β,
(
)
2
2
c max k2q
−1
−1
k1∗ logn p ,
k2q
√
n
σ02 . By this result, if k1 ≤ ζ0 k2 , the lower bounds
√
(3.18) in the regions k1 .
n
log p
. k2 .
n
log p
√
and
n
log p
. k1 ≤ k2 .
n
log p
.
HIGH-DIMENSIONAL ACCURACY ASSESSMENT
31
b `q ≥
follow; if ζ0 k2 < k1 ≤ k2 , by the fact that L∗α Θ0 (k1 ) , Θ0 (k2 ) , β,
b `q and k ∗ = ζ0 k2 ≥ ζ0 k1 , the lower bounds (3.18)
L∗α Θ0 (k1∗ ) , Θ0 (k2 ) , β,
1
√
n
log p
. k2 .
n
log p
√
n
log p
. k1 ≤ k2 . logn p follow.
The following
lemma shows that (3.7) holds for βbL defined in (2.10) with
√
bL
A > 2 by verifying
√ Assumption (A1) and (3.18) holds for β defined in
(2.10) with A > 4 2 by verifying Assumption (A2). Its proof can be found
in the supplement [6].
over the regions k1 .
and
√
If A > 4 2, then we have
2
log p 2
inf
Pθ kβbL − β ∗ k2q ≤ Ckβ ∗ k0q
σ ≥ 1−c exp −c0 n −p−c .
n
{θ=(β ∗ ,I,σ):σ≤2σ0 }
√
In particular, the above result holds for q = 2 under the assumption A > 2.
Lemma 2.
References.
[1] Ery Arias-Castro, Emmanuel J Candès, and Yaniv Plan. Global testing under sparse
alternatives: Anova, multiple comparisons and the higher criticism. The Annals of
Statistics, 39(5):2533–2556, 2011.
[2] Mohsen Bayati and Andrea Montanari. The lasso risk for gaussian matrices. Information Theory, IEEE Transactions on, 58(4):1997–2017, 2012.
[3] Alexandre Belloni, Victor Chernozhukov, and Lie Wang. Square-root lasso: pivotal
recovery of sparse signals via conic programming. Biometrika, 98(4):791–806, 2011.
[4] Peter J Bickel, Ya’acov Ritov, and Alexandre B Tsybakov. Simultaneous analysis of
lasso and dantzig selector. The Annals of Statistics, 37(4):1705–1732, 2009.
[5] Peter Bühlmann and Sara Van De Geer. Statistics for high-dimensional data: methods, theory and applications. Springer Science & Business Media, 2011.
[6] T Tony Cai and Zijian Guo. Supplement to “accuracy assessment for highdimensional linear regression”. 2016.
[7] T Tony Cai and Zijian Guo. Confidence intervals for high-dimensional linear regression: Minimax rates and adaptivity. The Annals of Statistics, to appear.
[8] T Tony Cai and Mark G Low. An adaptation theory for nonparametric confidence
intervals. The Annals of statistics, 32(5):1805–1840, 2005.
[9] T Tony Cai and Mark G Low. Adaptive confidence balls. The Annals of Statistics,
34(1):202–228, 2006.
[10] T Tony Cai, Mark G Low, and Zongming Ma. Adaptive confidence bands for nonparametric regression functions. Journal of the American Statistical Association,
109:1054–1070, 2014.
[11] T. Tony Cai and Harrison H Zhou. A data-driven block thresholding approach to
wavelet estimation. The Annals of Statistics, 37(2):569–595, 2009.
[12] Emmanuel Candès and Terence Tao. The dantzig selector: statistical estimation when
p is much larger than n. The Annals of Statistics, 35(6):2313–2351, 2007.
[13] Victor Chernozhukov, Christian Hansen, and Martin Spindler. Post-selection and
post-regularization inference in linear models with many controls and instruments.
2015.
32
T. T. CAI AND Z. GUO
[14] Victor Chernozhukov, Christian Hansen, and Martin Spindler. Valid post-selection
and post-regularization inference: An elementary, general approach. arXiv preprint
arXiv:1501.03430, 2015.
[15] David L Donoho and Iain M Johnstone. Adapting to unknown smoothness via wavelet
shrinkage. Journal of the American Statistical Association, 90(432):1200–1224, 1995.
[16] David L Donoho, Arian Maleki, and Andrea Montanari. The noise-sensitivity
phase transition in compressed sensing. Information Theory, IEEE Transactions
on, 57(10):6920–6941, 2011.
[17] Zijian Guo, Wanjie Wang, T Tony Cai, and Hongzhe Li. Optimal estimation of
co-heritability in high-dimensional linear models. arXiv preprint arXiv:1605.07244,
2016.
[18] Marc Hoffmann and Richard Nickl. On adaptive inference and confidence bands. The
Annals of Statistics, 39(5):2383–2409, 2011.
[19] Yuri I Ingster, Alexandre B Tsybakov, and Nicolas Verzelen. Detection boundary in
sparse regression. Electronic Journal of Statistics, 4:1476–1526, 2010.
[20] Lucas Janson, Rina Foygel Barber, and Emmanuel Candès. Eigenprism: Inference
for high-dimensional signal-to-noise ratios. arXiv preprint arXiv:1505.02097, 2015.
[21] Ker-Chau Li. From stein’s unbiased risk estimates to the method of generalized cross
validation. The Annals of Statistics, 13(4):1352–1377, 1985.
[22] Richard Nickl and Sara van de Geer. Confidence sets in sparse regression. The Annals
of Statistics, 41(6):2852–2876, 2013.
[23] Garvesh Raskutti, Martin J Wainwright, and Bin Yu. Minimax rates of estimation
for high-dimensional linear regression over-balls. Information Theory, IEEE Transactions on, 57(10):6976–6994, 2011.
[24] James Robins and Aad Van Der Vaart. Adaptive nonparametric confidence sets. The
Annals of Statistics, 34(1):229–253, 2006.
[25] Charles M Stein. Estimation of the mean of a multivariate normal distribution. The
Annals of Statistics, 9(6):1135–1151, 1981.
[26] Tingni Sun and Cun-Hui Zhang. Scaled sparse linear regression. Biometrika,
101(2):269–284, 2012.
[27] Christos Thrampoulidis, Ashkan Panahi, and Babak Hassibi. Asymptotically exact
error analysis for the generalized `22 -lasso. arXiv preprint arXiv:1502.06287, 2015.
[28] Robert Tibshirani. Regression shrinkage and selection via the lasso. Journal of the
Royal Statistical Society. Series B (Methodological), 58(1):267–288, 1996.
[29] Sara van de Geer, Peter Bühlmann, Yaacov Ritov, and Ruben Dezeure. On asymptotically optimal confidence regions and tests for high-dimensional models. The Annals
of Statistics, 42(3):1166–1202, 2014.
[30] Nicolas Verzelen. Minimax risks for sparse regressions: Ultra-high dimensional phenomenons. Electronic Journal of Statistics, 6:38–90, 2012.
[31] Fei Ye and Cun-Hui Zhang. Rate minimaxity of the lasso and dantzig selector for the
`q loss in `r balls. The Journal of Machine Learning Research, 11:3519–3540, 2010.
[32] Feng Yi and Hui Zou. SURE-tuned tapering estimation of large covariance matrices.
Computational Statistics & Data Analysis, 58:339–351, 2013.
HIGH-DIMENSIONAL ACCURACY ASSESSMENT
DEPARTMENT OF STATISTICS
THE WHARTON SCHOOL
UNIVERSITY OF PENNSYLVANIA
PHILADELPHIA, PENNSYLVANIA 19104
USA
E-mail: [email protected]
[email protected]
URL: http://www-stat.wharton.upenn.edu/∼tcai/
33
| 10math.ST
|
arXiv:1502.05445v3 [math.GR] 30 May 2017
Effective Separability of Finitely Generated Nilpotent Groups
1
Effective Separability of
Finitely Generated Nilpotent Groups
Mark Pengitore∗
June 1, 2017
Abstract
We give effective proofs of residual finiteness and conjugacy separability for finitely generated
nilpotent groups. In particular, we give precise asymptotic bounds for a function introduced
by Bou-Rabee that measures how large the quotients that are needed to separate non-identity
elements of bounded length from the identity which improves the work of Bou-Rabee. Similarly, we give polynomial upper and lower bounds for an analogous function introduced by
Lawton, Louder, and McReynolds that measures how large the quotients that are needed to
separate pairs of distinct conjugacy classes of bounded word length using work of Blackburn
and Mal’tsev.
1
Introduction
We say that Γ is residually finite if for each γ ∈ Γ − {1} there exists a surjective homomorphism
to a finite group ϕ : Γ → Q such that ϕ (γ ) 6= 1. Mal’tsev [29] proved that if Γ is a residually
finite finitely presentable group, then there exists a solution to the word problem of Γ. We say
that Γ is conjugacy separable if for each non-conjugate pair γ , η in Γ there exists a surjective
homomorphism to a finite group ϕ : Γ → Q such that ϕ (γ ) and ϕ (η ) are not conjugate. Mal’tsev
[29] also proved that if Γ is a conjugacy separable finitely presentable group, then there exists a
solution to the conjugacy problem of Γ
Residual finiteness, conjugacy separability, subgroup separability, and other residual properties
have been extensively studied and used to great effect in resolving important conjectures in geometry, such as the work of Agol on the Virtual Haken conjecture. Much of the work in the literature
has been to understand which groups satisfy various residual properties. For example, free groups,
polycyclic groups, surface groups, and fundamental groups of compact, orientable 3-manifolds
have all been shown to be residually finite and conjugacy separable [2, 13, 17, 18, 34, 36]. Recently, there have been several papers that have made effective these separability properties for
certain classes of groups. The main purpose of this article is to improve on the effective residual
∗ Purdue
University, West Lafayette, IN. E-mail: [email protected]
Effective Separability of Finitely Generated Nilpotent Groups
2
finiteness results of [4] and establish effective conjugacy separability results, both for the class of
finitely generated nilpotent groups.
1.1
Residual Finiteness
For a finitely generated group Γ with a finite generating subset S, [4] (see also [33]) introduced a
function FΓ,S (n) on the natural numbers that quantifies residual finiteness. Specifically, the value of
FΓ,S (n) is the maximum order of a finite group needed to distinguish a non-identity element from
the identity as one varies over non-identity elements in the n-ball. Numerous authors have studied
the asymptotic behavior of FΓ,S (n) for a wide collection of groups Γ (see [3, 4, 5, 8, 9, 20, 31, 33]).
To state our results, we require some notation. For two non-decreasing functions f , g : N → N,
we write f g if there exists a C ∈ N such that f (n) ≤ Cg(Cn) for all n ∈ N. We write f ≈ g
when f g and g f . For a finitely generated nilpotent group Γ, we denote T (Γ) to be the
normal subgroup of finite order elements. As we will see in Subsection 2.2.1, the dependence of
FΓ,S (n) is mild; subsequently, we will suppress the dependence of FΓ on the generating subset in
this subsection.
For finitely generated nilpotent groups, Bou-Rabee [4, Thm 0.2] proved that FΓ (n) (log(n))h(Γ)
where h(Γ) is the Hirsch length of Γ. Our first result establishes the precise asymptotic behavior
of FΓ (n).
Theorem 1.1. Let Γ be an infinite, finitely generated nilpotent group. There exists a ψRF (Γ) ∈
N such that FΓ (n) ≈ (log(n))ψRF (Γ) . Additionally, one can compute ψRF (Γ) given a basis for
(Γ/T (Γ))c where c is the step length of Γ/T (Γ).
The proof of Theorem 1.1 is done in two steps. To establish the upper bound, we appeal to
some structural properties of finitely generated nilpotent groups. To establish the lower bound, we
construct a sequence {γi } ⊆ Γ such that the order of the minimal finite group that separates γi from
the identity is bounded below by C(log(Ckγi kS ))ψRF (Γ) for some C ∈ N.
The following is a consequence of the proof of Theorem 1.1.
Corollary 1.2. Let Γ be a finitely generated nilpotent group. Then FΓ (n) ≈ (log(n))h(Γ) if and
only if h(Z(Γ/T (Γ))) = 1.
We now introduce some terminology. Suppose that G is a connected, simply connected nilpotent
Lie group with Lie algebra g. We say that G is Q-defined if g admits a basis with rational structure constants. The Mal’tsev completion of a torsion free, finitely generated nilpotent group Γ
is a connected, simply connected, Q-defined nilpotent Lie group G such that Γ embeds into as a
cocompact lattice.
The next theorem demonstrates that the asymptotic behavior of FΓ (n) is an invariant of the Mal’tsev
completion of Γ/T (Γ).
Theorem 1.3. Suppose that Γ1 and Γ2 are two infinite, finitely generated nilpotent groups such
that Γ1 /T (Γ1 ) and Γ2 /T (Γ2 ) have isomorphic Mal’tsev completions. Then FΓ1 (n) ≈ FΓ2 (n).
Effective Separability of Finitely Generated Nilpotent Groups
3
The proof of theorem 1.3 follows by an examination of a cyclic series that comes from a refinement
of the upper central series and its interaction with the topology of the Mal’tsev completion.
Since the 3-dimensional integral Heisenberg group embeds into every infinite, non-abelian nilpotent group, Theorem 1.1, Theorem 1.3, [4, Thm 2.2], and [4, Cor 2.3] allow us to characterize
Rd within the collection of connected, simply connected Q-defined nilpotent Lie groups by the
asymptotic behavior of residual finiteness of a cocompact lattice.
Corollary 1.4. Let G be a connected, simply connected, Q-defined nilpotent Lie group. Then G is
Lie isomorphic to Rdim(G) if and only if FΓ (n) (log(n))3 where Γ ⊆ G is any cocompact lattice.
For the last result of this section, we need some terminology. We say that a group Γ is irreducible
if there is no non-trivial splitting of Γ as a direct product. For a function of the form f (n) =
(log(n))m , we call m the polynomial in logarithm degree of growth for f (n).
Theorem 1.5.
(i) For c ∈ N, there exists m(c) ∈ N satisfying the following. For each ℓ ∈ N there exists an
irreducible, torsion free, finitely generated nilpotent group Γ of step length c and h(Γ) ≥ ℓ
such that FΓ (n) (log(n))m(c) .
(ii) Every natural number not equal to 2 can be realized as the polynomial in logarithm degree of
growth for FΓ (n) where Γ is an irreducible, torsion free, finitely generated nilpotent group.
(iii) Suppose 2 ≤ c1 < c2 are natural numbers. For each ℓ ∈ N, there exist irreducible, torsion
free, finitely generated nilpotent groups Γℓ and ∆ℓ of step lengths c1 and c2 , respectively,
such that FΓℓ (n) ≈ F∆ℓ (n).
(iv) For natural numbers c > 1 and m ≥ 1, there exists an irreducible, torsion free, finitely
generated nilpotent group Γ of step length c such that (log(n))m FΓ (n).
For Theorem 1.5(i), we consider free nilpotent groups of a fixed step length and increasing rank.
We make use of central products of filiform nilpotent groups for Theorem 1.5(ii) - (iv).
Using Theorem 1.5, we are able to relate the polynomial in logarithm degree of growth of the
residual finiteness function with well known invariants of the class of finitely generated nilpotent
groups. Theorem 1.5(i) implies the polynomial in logarithm degree of growth of FΓ (n) does not
depend on the Hirsch length of Γ. Similarly, Theorem 1.5(iv) implies there is no upper bound in
terms of step length of Γ for the polynomial in logarithm degree of growth of FΓ (n). On the other
hand, the step size of Γ is not determined by the polynomial in logarithm growth of FΓ (n) as seen
in Theorem 1.5(iii).
1.2
Conjugacy Separability
We now turn our attention to effective conjugacy separability. Lawton–Louder–McReynolds [24]
introduced a function ConjΓ,S (n) on the natural numbers that quantifies conjugacy separability. To
Effective Separability of Finitely Generated Nilpotent Groups
4
be precise, the value of ConjΓ,S (n) is the maximum order of the minimal finite quotient needed to
separate a pair of non-conjugate elements as one varies over non-conjugate pairs of elements in
the n-ball. Since the dependence of ConjΓ,S (n) on S is mild (see Lemma 2.1), we will suppress the
generating subset throughout this subsection.
To the author’s knowledge, the only previous work on the asymptotic behavior of ConjΓ (n) is due
to Lawton–Louder–McReynolds [24]. They demonstrate that if Γ is a surface group or a finite
2
rank free group, then ConjΓ (n) nn [24, Cor 1.7]. In this subsection, we initiate the study of the
asymptotic behavior of ConjΓ (n) for the collection of finitely generated nilpotent groups.
Our first result is the precise asymptotic behavior of ConjH2m+1 (Z) (n) where H2m+1 (Z) is the (2m +
1)-dimensional integral Heisenberg group.
Theorem 1.6. ConjH2m+1 (Z) (n) ≈ n2m+1 .
For general nilpotent groups, we establish the following upper bound for ConjΓ (n).
Theorem 1.7. Let Γ be a finitely generated nilpotent group. Then ConjΓ (n) nk for some k ∈ N.
Blackburn [2] was the first to prove conjugacy separability of finitely generated nilpotent groups.
Our strategy for proving Theorem 1.7 is to effectivize [2].
For the same class of groups, we have the following lower bound which allows us to characterize
virtually abelian groups within the class of finitely generated nilpotent groups. Moreoever, we
obtain the first example of a class of groups for which the asymptotic behavior of FΓ (n) and
ConjΓ (n) are shown to be dramatically different.
Theorem 1.8. Let Γ be a finitely generated nilpotent group.
(i) If Γ contains a normal abelian subgroup of index m, then log(n) ConjΓ (n) (log(n))m .
(ii) Suppose that Γ is not virtually abelian. Then there exists a ψLower (Γ) ∈ N such that nψLower (Γ)
ConjΓ (n). Additionally, one can compute ψLower (Γ) given a basis for (Γ/T (Γ))c where c is
the step length of Γ/T (Γ).
The proof of Theorem 1.8(i) is elementary. We prove Theorem 1.8(ii) by finding an infinite sequence of non-conjugate elements {γi , ηi } such that the order of the minimal finite group that
ψ
(Γ)
separates the conjugacy classes of γi and ηi is bounded below by C ni Lower for some C ∈ N
where kγi kS , kηi kS ≈ ni for some finite generating subset S.
We have the following theorem which is similar in nature to Theorem 1.3.
Theorem 1.9. Let Γ and ∆ be infinite, finitely generated nilpotent groups of step size greater than
or equal to 2, and suppose that Γ/T (Γ) and ∆/T (∆) have isomorphic Mal’tsev completions. Then
nψLower (Γ) Conj∆ (n) and nψLower (∆) ConjΓ (n).
We apply Theorem 1.8 to construct nilpotent groups that help demonstrate the various asymptotic
behaviors that the growth of conjugacy separability may exhibit.
Effective Separability of Finitely Generated Nilpotent Groups
5
Theorem 1.10. For natural numbers c > 1 and k ≥ 1, there exists an irreducible, torsion free,
finitely generated nilpotent group Γ of step length c such that nk ConjΓ (n).
Theorem 1.10 implies that the conjugacy separability function does not depend of the step length
of the nilpotent group. We consider central products of filiform nilpotent groups for Theorem 1.10.
1.3
Acknowledgements
I would like to thank my advisor Ben McReynolds for all his help and my mentor Priyam Patel for
her guidance. I am indebted to Karel Dekimpe for his suggestions for Proposition 11.1. I also want
to thank Khalid Bou-Rabee for looking at earlier drafts of this article and for his many insightful
comments. Finally, I want to thank Alex Barrios, Rachel Davis, Jonas Deré, Artur Jackson, Gijey
Gilliam, Brooke Magiera, and Nick Miller for conversations on this article.
2
Background
We will assume the reader is familiar with finitely generated groups, Lie groups and Lie algebras.
2.1
Notation and conventions
We let lcm {r1 , . . . , rm } be the lowest common multiple of {r1 , · · · , rm } ⊆ Z with the convention
that lcm(a) = |a| and lcm(a, 0) = 0. We let gcd(r1 , · · · , rm ) be the greatest common multiple of
{r1 , · · · , rm } ⊆ Z with the convention that gcd(a, 0) = |a|.
We denote kγ kS as the word length of γ with respect to S and denote the identity of Γ as 1. We
denote the order of γ as an element of Γ as OrdΓ (γ ). We write γ ∼ η when there exists a g ∈ Γ such
that g−1 γ g = η . For a normal subgroup ∆ E Γ, we set π∆ : Γ → Γ/∆ to be the natural projection
and write γ̄ = π∆ (γ ) when ∆ is clear from context. For a subset X ⊆ Γ, we denote hX i to be the
subgroup generated by X .
We define the commutator of γ and η as [γ , η ] = γ −1 η −1 γ η . We denote the m-fold commutator
of {γi }m
i=1 ⊆ Γ as [γ1 , . . . , γm ] with the convention that [γ1 , . . . , γm ] = [[γ1 , , . . . , γm−1 ], γm ].
We denote the center of Γ as Z(Γ) and the centralizer of γ in Γ as CΓ (γ ). We define Γi to be the
i-th term of the lower central series and Z i (Γ) to be the i-th term of the upper central series. For
γ ∈ Γ − {1}, we denote Height(γ ) as the minimal j ∈ N such that πZ j−1 (Γ) (γ ) 6= 1.
We define the abelianization of Γ as Γab with the associated projection given by πab = π[Γ,Γ] . For
m ∈ N, we define Γm ∼
= hγ m | γ ∈ Γi and denote the associated projection as σm = πΓm .
dim (g)
dim (g)
dim (g)
When given a basis X = {Xi }i=1R for g, we denote k ∑i=1R αi Xi kX = ∑i=1R |αi |. For a Lie
algebra g with a Lie ideal h, we define πh : g → g/h to be the natural Lie projection.
For a R-Lie algebra g, we denote Z(g) to the center of g, gi to be the i-th term of the lower central
series, and Z i (g) to be the i-th term of the upper central series.
Effective Separability of Finitely Generated Nilpotent Groups
6
For A ∈ g, we define the map adA : g → g to be given by adA (B) = [A, B]. We denote the m-fold Lie
bracket of {Ai }m
i=1 ⊆ g as [A1 , · · · Am ] with the convention that [A1 , · · · , Am ] = [[A1 , · · · , Am−1 ], Am ].
2.2
2.2.1
Finitely generated groups and separability
Residually finite groups
Following [4] (see also [33]), we define the depth function of Γ as DΓ : Γ − {1} → N ∪ {∞} for
finitely generated groups to be given by
def
DΓ (γ ) = min {|Q| | ϕ : Γ → Q, |Q| < ∞, and ϕ (γ ) 6= 1} .
def
We define FΓ,S : N → N by FΓ,S (n) = max {DΓ (γ ) | kγ kS ≤ n and γ 6= 1} . When Γ is a residually
finite group, then FΓ,S (n) < ∞ for all n ∈ N. For any two finite generating subsets S1 and S2 ,
we have FΓ,S1 (n) ≈ FΓ,S2 (n) (see [4, Lem 1.1]). Therefore, we will suppress the choice of finite
generating subset.
2.2.2
Conjugacy separable groups
Following [24], we define the conjugacy depth function of Γ as CDΓ : Γ × Γ − {(γ , η ) | γ ∼ η } →
N ∪ {∞} to be given by
def
CDΓ (γ , η ) = min {|Q| | ϕ : Γ → Q, |Q| < ∞, and ϕ (γ ) ≁ ϕ (η )} .
def
We define ConjΓ,S (n) : N → N as ConjΓ,S (n) = max {CDΓ (γ , η ) | γ ≁ η and kγ kS , kη kS ≤ n} .
When Γ is a conjugacy separable group, then ConjΓ,S (n) < ∞ for all n ∈ N.
Lemma 2.1. If S1 , S2 are two finite generating subsets of Γ, then ConjΓ,S1 (n) ≈ ConjΓ,S2 (n).
The proof is similar to [4, Lem 1.1] (see also [24, Lem 2.1]). As before, we will suppress the
choice of finite generating subset.
[24, Lem 2.1] implies that the order of minimal finite group required to separate a non-identity
element γ ∈ Γ from the identity is bounded above by order of the minimal finite group required
to separate the conjugacy class of γ from the identity. Thus, FΓ (n) ConjΓ (n) for all conjugacy
separable groups. In particular, if Γ is conjugacy separable, then Γ is residually finite.
2.3
Nilpotent groups and nilpotent Lie groups
See [12, 16, 22, 35] for a more thorough account of the material in this subsection. Let Γ be a
def
non-trivial, finitely generated group. The i-th term of the lower central series is defined by Γ1 = Γ
def
def
and for i > 1 as Γi = [Γi−1 , Γ]. The i-term of the upper central series is defined by Γ0 = {1} and
def
i−1 (Γ))) for i > 1.
Z i (Γ) = πZ−1
i−1 (Γ) (Z(Γ/Z
Effective Separability of Finitely Generated Nilpotent Groups
7
Definition 2.2. We say that Γ is a nilpotent group of step size c if c is the minimal natural number
such that Γc+1 = {1}, or equivalently, Z c (Γ) = Γ. If the step size is unspecified, we simply say
that Γ is a nilpotent group. We say that finitely generated nilpotent group is an admissible group.
For an admissible group Γ, the set of finite order elements of Γ, denoted as T (Γ), is a finite order
characteristic subgroup. Moreover, when |Γ| = ∞, then Γ/T (Γ) is torsion free.
Let g be a non-trivial, finite dimensional R-Lie algebra. The i-th term of the lower central series
def
def
of g is defined by g1 = g and for i > 1 as gi = [gi−1 , g]. We define the i-th term of the upper
def
def
central series as Z 0 (g) = {0} and Z i(g) = πZ i−1 (g) (Z(g/Z i−1 (g))) for i > 1.
Definition 2.3. We say that g is a nilpotent Lie algebra of step length c if c is the minimal natural
number satisfying gc = g, or equivalently, gc+1 = {0}. If the step size is unspecified, we simply
say that g is a nilpotent Lie algebra.
For a connected, simply connected nilpotent Lie group G of step length c with Lie algebra g, the
exponential map, written as exp : g → G, is a diffeomorphism [22, Thm 1.127] whose inverse is
formally denoted as Log. The Baker-Campbell-Hausdorff formula [12, (1.3)] implies that every
A, B ∈ g satisfies
∞
1
def
def
A ∗ B = Log(exp A · exp B) = A + B + [A, B] + ∑ CBm (A, B)
2
m=3
(1)
where CBm (A, B) is as rational linear combination of m-fold Lie brackets of A and B. By assumption, CBm (A, B) = 0 for m > c. For {Ai }m
i=1 in g, we may more generally write
c
A1 ∗ · · · ∗ Am = Log(exp A1 · · · exp Am ) = ∑ CBi (A1 , · · · , Am )
(2)
i=1
ℓ ⊆ {A }m
where CBi (A1 , · · · , Am ) is a rational linear combination of i-fold Lie brackets of {A jt }t=1
i i=1
via repeated applications of the Baker-Campbell-Hausdorff formula.
We define the adjoint representation Ad : G → Aut(g) of G as Ad(g)(X ) = (dΨg )1 (X ) where
Ψg (x) = g x g−1 . By [22, 1.92], we may write for γ ∈ Γ and A ∈ g
i
c (ad
1
Log(γ ) ) (A)
Ad(γ )(A) = A + [Log(γ ), A] + ∑
.
2
i!
i=3
(3)
By [28], a connected, simply connected nilpotent Lie group G with Lie algebra g admits a codim(G)
compact lattice Γ if and only if g admits a basis {Xi }i=1 with rational structure constants (see
[27, Thm 7] for more details). We say G is Q-defined if it admits a cocompact lattice. For any
torsion free admissible group Γ, [27, Thm 6] implies that there exists a Q-defined group unique
up to isomorphism in which Γ embeds as a cocompact lattice.
Definition 2.4. We call this Q-defined group the Mal’tsev completion of Γ. When given a Qdefined group G, the tangent space at the identity with the Lie bracket of vector fields is a finite
dimensional, nilpotent Lie algebra. We say that a connected, simply connected nilpotent Lie group
is an admissible Lie group.
Effective Separability of Finitely Generated Nilpotent Groups
2.4
8
Polycyclic groups
See [19, 32, 35] for the material contained in the following subsection.
Definition 2.5. A group Γ is polycyclic if there exists an ascending chain of subgroups {∆i }m
i=1
such that ∆1 is cyclic, ∆i E ∆i+1 , and ∆i+1 /∆i is cyclic for all i. We call {∆i }m
i=1 a cyclic series
m
for Γ. We say {ξi }m
i=1 is a compatible generating subset with respect to {∆i }i=1 if hξ1 i = ∆1 and
hξi+1 , ∆i i = ∆i+1 for i > 1. We define the Hirsch length of Γ, denoted as h(Γ), as the number of
indices i such that |∆i+1 : ∆i | = ∞.
For a general polycyclic group, there may be infinitely many different cyclic series of arbitrary
length (see [19, Ex 8.2]). However, the Hirsch length of Γ is independent of the choice of cyclic
series. With respect to the compatible generating subset {ξi }m
i=1 , [19, Lem 8.3] implies that we may
αi
m
represent every γ ∈ Γ uniquely as γ = ∏i=1 ξi where αi ∈ Z if |∆i+1 : ∆i | = ∞ and 0 ≤ αi < ri if
|∆i+1 : ∆i | = ri . If |Γ| < ∞, then the second paragraph after [19, Defn 8.2] implies that |Γ| = ∏m
i=1 ri .
Definition 2.6. We call the collection of such m-tuples a Mal’tsev basis for Γ with respect to the
m
compatible generating subset {ξi }m
i=1 . We call (αi )i=1 the Mal’tsev coordinates of γ .
For an admissible group Γ, we may refine the upper central series of Γ to obtain a cyclic series and
compatible generating subset. First assume that Γ is abelian. We may write Γ ∼
= Zm ⊕T (Γ), and we
h(Γ)
let {ξi }i=1 be a free basis for Zm . Since T (Γ) is a finite abelian group, there exists an isomorphism
ϕ : T (Γ) → ⊕ℓi=1 Z/pki i Z. If xi generates Z/pki i Z in ⊕ℓi=1 Z/pki i Z, we then set ξi = ϕ −1 (xi−h(Γ) ) for
h(Γ)+ℓ
h(Γ) + 1 ≤ i ≤ h(Γ) + ℓ. Thus, the groups {∆i }i=1
h(Γ)+ℓ
Γ with a compatible generating subset {ξi }i=1 .
i
form a cyclic series for
given by ∆i = hξt it=1
h(Γ)
Now assume that Γ has step length c > 1. There exists a generating basis {zi }i=1 for Z(Γ) and
h(Γ )
h(Γ)
integers {ti }i=1 such that ztii i=1c is a basis for Γc and for each i there exist xi ∈ Γc−1 and
h(Γ)
yi ∈ Γ such that ztii = [xi , yi ]. We may choose a cyclic series {Hi }i=1 such that Hi = hzs iis=1 .
Induction implies that there exists cyclic series {Λi }ki=1 and compatible generating subset {λi }ki=1
−1
(Λi−ℓ ). For
for Γ/Z(Γ). For 1 ≤ i ≤ ℓ, we set ∆i = Hi , and for ℓ + 1 ≤ i ≤ ℓ + k, we set ∆i = πZ(Γ)
1 ≤ i ≤ ℓ, we set ξi = zi . For ℓ + 1 ≤ i ≤ ℓ + k, we choose a ξi such that πZ(Γ) (ξi ) = λi−ℓ . It then
ℓ+k
follows that {∆i }ℓ+k
i=1 is a cyclic series for Γ with a compatible generating subset given by {ξi }i=1 .
c
i
i−1
Moreover, the given construction implies that h(Γ) = ∑i=1 rankZ (Z (Γ)/Z (Γ)). Whenever Γ is
an admissible group, we choose the cyclic series and compatible generating subset this way.
Definition 2.7. An admissible Γ with a cyclic series {∆i } and a compatible generating subset {ξi }
is called an admissible 3-tuple and is denoted as {Γ, ∆i , ξi }.
Let {Γ, ∆i , ξi } be an admissible 3-tuple such that Γ is torsion free. [16, Thm 6.5] implies that
multiplication of γ , η ∈ Γ can be expressed as polynomials in terms of the Mal’tsev basis ash(Γ)
h(Γ)
sociated to the cyclic series {∆i }i=1 and the compatible generating subset {ξi }i=1 . Specifih(Γ) b
h(Γ)
h(Γ)
cally, we may write γ η = (∏i=1 ξiai ) · (∏ j=1 ξ j j ) = ∏s=1 ξsds where each ds can be expressed
as a polynomial in the Mal’stev coordinates of γ and η , respectively. Similarly, we may write
Effective Separability of Finitely Generated Nilpotent Groups
h(Γ)
9
h(Γ)
γ ℓ = (∏i=1 ξiai )ℓ = ∏ j=1 ξ e j where each e j can be expressed as a polynomial in the Mal’tsev
coordinates of γ and the integer ℓ.
The polynomials that define the group product and group power operation of Γ with respect to
the given cyclic series and compatible generating subset have unique extensions to Rh(Γ) . That
implies G is diffeomorphic to Rh(Γ) (see [16, Thm 6.5], [22, Cor 1.126]) where G is the Mal’tsev
completion of Γ. Consequently, the dimension and step length of G are equal to the Hirsch length
and step length of Γ, respectively. Thus, we may write h(Γ) = dim(G). Furthermore, we may
identify Γ with its image in GΓ which is the set Zh(Γ) .
The following definition will be of use for the last lemma of this subsection.
Definition 2.8. Let Γ be a torsion free admissible group, and let ∆ ≤ Γ be a subgroup. We define
the isolator of ∆ in Γ as the subgroup given by
√
Γ
∆ = {γ ∈ Γ | there exists k ∈ N such that γ k ∈ ∆} ∪ {1} .
We make some
observations. By the paragraph
proceeding exercise 8 of [35, Ch 8] and [16,
√
√
Γ
∆ is a subgroup such that | Γ ∆ : ∆| < ∞. If Γ √
is abelian, then we may write Γ =
Thm√4.5], √
Γ
Γ
(Γ/ ∆) ⊕ ∆. We also have for any subgroup ∆ E Γ that Γ/ Γ ∆ is torsion free.
We finish this section with the following result. When given a torsion free admissible group Γ, the
following lemma relates the word length of element γ in Γ to the Mal’tsev coordinates of γ with
respect to a choice of a cyclic series and a compatible generating subset.
Lemma 2.9. Let {Γ, ∆i , ξi } be an admissible 3-tuple such that Γ has step length c. If γ ∈ Γ such
that kγ kS ≤ n, then |αi | ≤ C nc where (αi ) are the Mal’tsev coordinates of γ for some C ∈ N.
Proof. If Γ is finite, then the statement is evident. Thus, we may assume that |Γ| = ∞. We proceed
by induction on step length, and observe that the base case of abelian groups is clear. Now suppose
that Γ has step length c > 1 and that kγ kS ≤ n. Since kπΓi (γ )kπΓi (S) ≤ n, the inductive hypothesis
implies there exists a constant C0 > 0 such that |αt | ≤ C0 nt when πΓt (ξi ) 6= 1. We let m be the
smallest integer such that ξi ∈
/ Γc when i > m and ξi ∈ Γc for i ≤ m, and let k be the length of the
cyclic series ∆i .
Suppose that ξi ∈
/ Γt+1 , but ξi ∈ Γt where t < c. We have by assumption that |αi | ≤ C0 nt . By
[15, 3.B2], we have kξiαi kS ≈ |αi |1/t . Thus, there exists a constant C1 such that kξiαi kS ≤ C1 |αi |1/t
1/t
when OrdΓ (ξi ) = ∞. Therefore, we may write kξiαi kS ≤ C0 C1 n. Thus, for ξi ∈
/ Γc such that
OrdΓ (ξi ) = ∞, we have kξiαi kS ≤ C2 n for some C2 ∈ N. By increasing C2 , we may assume that
|C2 | > |T (Γ)|, and thus, we cover the cases where OrdΓ (ξi ) < ∞.
Letting ζ = (∏ki=m+1 ξiαi )−1 where ξi ∈ Γc , it is evident that kζ kS ≤ C3 n for some C3 ∈ N. Thus,
αi
if η = ∏m
i=1 ξi , we may write kη kS = kγ ζ kS ≤ kγ kS + kζ kS ≤ 2C0 n. Hence, we may assume
that γ ∈ Γc .
αi
− αi
m
, we may write kξtαt kS =
We may write γ = ∏m
i=1 ξi where ξi ∈ Γc . If we let λt = ∏i=1,i6=t ξi
kζ λt kS ≤ kζ kS + kλt kS ≤ 4C0 n. Thus, we need only consider when γ = ξiαi for ξi ∈ Γc . Since
ξi ∈ Γc , [15, 3.B2] implies that |αi | ≤ C3 nc for some C3 ∈ N.
Effective Separability of Finitely Generated Nilpotent Groups
3
10
Torsion-free Quotients with One Dimensional Centers
In the following subsection, we define what a choice of an admissible quotient with respect to a
central non-trivial element is, what a choice of a maximal admissible quotient is, and define the
constants ψRF (Γ) and ψLower (Γ) for an infinite admissible group Γ.
3.1
Existence of torsion free quotients with one dimensional center
The following proposition will be useful throughout this article.
Proposition 3.1. Let Γ be a torsion free admissible group, and suppose that γ is a central, nontrivial element. There exists a normal subgroup Λ in Γ such that Γ/Λ is an irreducible, torsion
free, admissible group such that hπΛ (γ )i is a finite index subgroup of Z(Γ/Λ). If γ is primitive,
then Z(Γ/Λ) ∼
= hπΛ (γ )i.
Proof. We construct Λ by induction on Hirsch length, and since the base case is trivial, we may
assume that h(Γ) > 1. If Z(Γ) ∼
= Z, then the proposition is now evident by letting Λ = {1}.
h(Z(Γ))
for Z(Γ) such that zk1 = γ for some
Now assume that h(Z(Γ)) ≥ 2. There exists a basis {zi }i=1
h(Z(Γ))
k ∈ Z − {0}. Letting K = hzi ii=2 , we note that K E Γ and πK (γ ) 6= 1. Additionally, it follows
that Γ/K is a torsion free, admissible group. If h(Z(Γ/K)) = 1, then our proposition is evident by
defining Λ ∼
= K.
Now suppose that h(Z(Γ/K)) ≥ 2. Since h(Γ/K) < h(Γ), the inductive hypothesis implies there
exists a subgroup Λ1 such that Λ1 E Γ/K and where (Γ/K)/Λ1 is a torsion free admissible
group. Letting ρ : Γ/K → (Γ/K)/Λ1 be the natural projection, induction additionally implies that
Z((Γ/K)/Λ1 ) ∼
= hρ (πK (γ ))i. Taking Λ2 = πK−1 (Λ1 ), we note that Λ2 /K ∼
= Λ1 and that K ≤ Λ2 .
Thus, the third isomorphism theorem implies that (Γ/K)/(Λ2 /K) ∼
= Γ/Λ2 . Hence, Γ/Λ2 is a
torsion free admissible, and by construction, hπΛ2 (γ )i is a finite index subgroup of Z(Γ/Λ2 ).
Letting Λ satisfy the hypothesis of the proposition for γ , we now demonstrate that Γ/Λ is irreducible. Suppose for a contradiction there exists a pair of non-trivial admissible groups ∆1 and ∆2
such that Γ/Λ ∼
= ∆1 × ∆2 . Since Γ/Λ is torsion free, ∆1 and ∆2 are torsion free. By selection, it
follows that Z(∆1 ) and Z(∆2 ) are torsion free, finitely generated abelian groups. Hence, Z2 is isomorphic to a subgroup of Z(Γ/Λ). Subsequently, h(Z(Γ/Λ)) ≥ 2 which is a contradiction. Thus,
either ∆1 ∼
= {1} or ∆2 ∼
= {1}, and subsequently, Γ/Λ is irreducible.
Definition 3.2. Let γ ∈ Γ be a central, non-trivial element, and let J be the set of subgroups of
Γ that satisfy Proposition 3.1 for γ . Since the set {h(Γ/Λ) | Λ ∈ J } is bounded below by 1, there
exists an Ω ∈ J such that h(Γ/Ω) = min{h(Γ/Λ) | Λ ∈ J }. We say Ω is a choice of a admissible
quotient of Γ with respect to γ .
For a primitive γ ∈ Z(Γ) − 1, we let Γ/Λ1 and Γ/Λ2 be two different choices of admissible quotients of Γ with respect to γ . In general, Γ/Λ1 ≇ Γ/Λ2 . On the other hand, we have, by definition,
that h(Γ/Λ1 ) = h(Γ/Λ2 ). Subsequently, the Hirsch length of a choice of an admissible quotient
Effective Separability of Finitely Generated Nilpotent Groups
11
with respect to γ is a natural invariant of Γ associated to γ . Such a quotient corresponds to a torsion
free quotient of Γ of minimal Hirsch length such that γ has a non-trivial image that generates the
center. That will be useful in finding the smallest finite quotient in which γ has a non-trivial image.
Definition 3.3. Let Γ be a torsion free admissible group of step length c. For each γ ∈ Z(Γ) − {1},
we let Γ/Λγ be a choice of an admissible quotient of Γ with respect to γ . Let J be the set of
γ ∈ Z(Γ) − {1} such that there exists a k such that γ k = [a, b] where a ∈ Γc−1 and b ∈ Γ. Observe
that the set h(Γ/Λ
γ | γ ∈ J ) is bounded above by h(Γ). Thus, there exists an η ∈ J such that
h(Γ/Λη ) = max h(Γ/Λγ ) | γ ∈ J . We say that Γ/Λη corresponds to a choice of a maximal
admissible quotient of Γ.
We now define ψRF (Γ) and ψLower (Γ) when Γ is an infinite admissible group.
Definition 3.4. Let Γ be an infinite admissible group, and let (Γ/T (Γ))/Λ be a choice of a maximal admissible quotient of Γ/T (Γ). We set ψRF (Γ) = h((Γ/T (Γ))/Λ). When Γ is not virtually
abelian, we define ψLower (Γ) = ψRF (Γ) (c − 1) where c is the step length of Γ/T (Γ).
Suppose that Γ/Λ1 and Γ/Λ2 are two choices of a maximal admissible quotients of Γ when Γ
is torsion free. In general, Γ/Λ1 ≇ Γ/Λ2 . However, h(Γ/Λ1 ) = h(Γ/Λ2 ) = ψRF (Γ) by definition; hence, ψRF (Γ) is a well defined invariant of Γ. The value ψRF (Γ) is important because it
corresponds to the polynomial in logarithm degree of growth for FΓ (n). Similarly, we have that
ψLower (Γ) is a well defined invariant of admissible groups that are not virtually abelian. Moreover,
the value ψLower (Γ) will correspond to the polynomial degree of growth of an asymptotic lower
bound for ConjΓ (n).
A natural observation is that if h(Z(Γ/T (Γ))) = 1, then ψRF (Γ) = h(Γ). Additionally, if Γ is an
infinite, finitely generated abelian group, then ψRF (Γ) = 1.
Let Γ be a torsion free admissible group with a primitive γ ∈ Z(Γ) − {1}, and let Γ/Λ be a choice
of an admissible quotient Γ/Λ with respect to γ . The next proposition demonstrates that we may
choose a cyclic series and a compatible generating subset such that a subset of the compatible
generating subset generates Λ.
Proposition 3.5. Let Γ be a torsion free admissible group, and let γ be a primitive, central nontrivial element. Let Γ/Λ be a choice of an admissible quotient with respect to γ . There exists a
h(Γ)
h(Γ)
cyclic series {∆i }i=1 and a compatible generating subset {ξi }i=1 such that Γ/Λ is a choice of
an admissible quotient with respect to ξ1 where γ = ξ1 . Moreover, there exists a subset, possibly
h(Λ)
empty, ξi j j=1 of the compatible generating subset satisfying the following. The subgroups
t
h(Λ)
Wt = ξi j j=1 form a cyclic series for Λ with a compatible generating subset given by ξi j j=1 .
Proof. We proceed by induction on h(Γ), and note that the base case of h(Γ) = 1 is evident. Thus,
we may assume that h(Γ) > 1. If h(Z(Γ)) = 1, then Λ ∼
= {1}; hence, we may take any cyclic series
and compatible generating subset. Therefore, we may also assume that h(Z(Γ)) > 1.
h(Z(Γ)
h(Z(Γ))
There exists a generating basis {zi }i=1 for Z(Γ) such that z1 = γ . Letting K = hzi ii=2 , we
note that K ≤ Λ. By passing to the quotient Γ/K, we note that (Γ/K)/(Λ/K) is a choice of an
Effective Separability of Finitely Generated Nilpotent Groups
12
admissible quotient with respect to πK (γ ) = 1. Induction implies that there exists a cyclic series
h(Γ/K)
h(Γ/K)
such that there exists a subset
{∆i /K}i=1
and a compatible generating subset {πK (ξi )}i=1
t
h(Γ/K)
πK (ξi j ) j=1 satisfying the following. The subgroups Wt /K = ξi j j=1 form a cyclic series for
h(Λ/K)
Λ/K with a compatible generating subset πK (ξi j ) j=1 . We let Hi = hzs iis=1 for 1 ≤ i ≤ h(Z(Γ))
D
E
i−h(K)
and for i > h(Z(Γ)), we let Hi = {K} ∪ {ξt }t=1
. We also take ηi = zi for 1 ≤ i ≤ h(Z(Γ)) and
h(Γ)
for i > h(Z(Γ)), we take ηi = ξi−h(Z(Γ)) . Thus, {Hi }i=1 is cyclic series for Γ with a compatible
h(Γ)
generating subset {ηi }i=1 .
h(Λ)
Consider the subset ηi j j=1 where ηi j = z j+1 for 1 ≤ j ≤ h(K) and where ηi j = ξi j−h(K) for
h(Λ)
j > h(K). Thus, by selection, ηi j j=1 is the required subset.
For the next two propositions, we establish some notation. Let Γ be a torsion free admissible nilpotent group. For each primitive element γ ∈ Z(Γ) − {1}, we let Γ/Λγ be a choice of an admissible
quotient with respect to γ .
We demonstrate that we may calculate ψRF (Γ) for Γ when given a generating basis for (Γ/T (Γ))c
where c is the step length of Γ/T (Γ).
h(Z(Γ))
Proposition 3.6. Let Γ be a torsion free admissible group, and let {zi }i=1
be a basis of
ti h(Γc )
h(Γc )
Z(Γ). Moreover, assume there exist integers {ti }i=1 such that zi i=1 is a basis of Γc and
√
that there exist ai ∈ Γc−1 and bi ∈ Γ such that ztii = [ai , bi ]. For each γ ∈ Z(Γ) Γc , there exists an
i0 ∈ {1, · · · , h(Γc )} such that Γ/Λzi0 is a choice of an admissible quotient with respect to γ . More
h(Z(Γ)
generally, if {zi }i=1 is any basis of Z(Γ), then there exists i0 ∈ {1, · · · , h(Z(Γ))} such that Γ/Λi0
is a choice of an admissible quotient with respect to γ .
√
h(M)
Proof. Letting M = Z(Γ) Γc , we may write γ = ∏i=1 zαi i . There exist indices 1 ≤ i1 < · · · < iℓ ≤
h(M) such that αi j 6= 0 for 1 ≤ j ≤ ℓ and αi = 0, otherwise. We observe that Γ/Λzit satisfies the
conditions of Proposition 3.1 for γ for each 1 ≤ t ≤ ℓ. Therefore, h(Γ/Λγ ) ≤ min{h(Γ/Λzit ) | 1 ≤
t ≤ ℓ}.
Since πΛγ (γ ) 6= 1, there exists it0 such that πΛγ (zit0 ) 6= 1. Thus, Γ/Λγ satisfies the conditions of
Proposition 3.1 for zit0 . Thus, h(Γ/Λzit ) ≤ h(Γ/Λγ ). In particular, min{h(Γ/Λzit ) | 1 ≤ t ≤ ℓ} ≤
0
h(Γ/Λγ ). Therefore, h(Γ/Λγ ) = min{h(Γ/Λzit ) | 1 ≤ i ≤ ℓ}. The last statement follows using
similar reasoning.
The following proposition demonstrates that ψRF (Γ) can always be realized as the Hirsch length
of a choice of an admissible quotient with respect to a central element of a fixed basis of Γc .
h(Z(Γ))
Proposition 3.7. Let Γ be a torsion free admissible group of step length c with a basis {zi }i=1
h(Γ )
h(Γ )
for Z(Γ). Moreover, assume there exist integers {ti }i=1c such that ztii i=1c is a basis of Γc
and that there exist elements ai ∈ Γc−1 and bi ∈ Γ such that ztii = [ai , bi ]. There exists an i0 ∈
Effective Separability of Finitely Generated Nilpotent Groups
13
{1, · · · , h(Γc )} such that ψRF (Γ) = h(Γ/Λzi0 ). Hence, ψRF (Γ) = max {h(Γ/Λzi ) | 1 ≤ i ≤ h(Γc )} .
More generally, if {zi } is any basis of Z(Γ), then ψRF (Γ) = max {h(Γ/Λzi ) | 1 ≤ i ≤ h(Γ)}
Proof. Let J be the set of central non-trivial γ such that there exists a k where γ k is a c-fold
commutator bracket. Given that the set {h(Γ/Λγ ) |γ ∈ J } is bounded above by h(Γ), there exists
a non-trivial element η ∈ J such that h(Γ/Λη ) = max{h(Γ/Λγ )|γ ∈ J }. Proposition 3.6 implies
there exists an i0 ∈ {1, · · · , h(Z(Γ))} such that h(Γ/Λη ) = h(Γ/Λzi0 ). By the definition of ψRF (Γ),
it follows ψRF (Γ) = max{h(Γ/Λzi0 ) | 1 ≤ i ≤ h(Γc )}. The last statement follows using similar
reasoning.
3.2
Properties of admissible quotients of Γ
We demonstrate conditions for a choice of an admissible quotient of Γ with respect to some primitive, central, non-trivial element to have the same step length as Γ. For the next proposition, if Γ
is an admissible group,then we fix its step length as c(Γ).
p
Proposition 3.8. Let Γ be a torsion free admissible group. If we let γ ∈ Z(Γ) Γc(Γ) − {1} be a
primitive element with a choice of admissible quotient Γ/Λ with respect to γ , then c(Γ/Λ) = c(Γ).
In particular, if Γ/Λ is a choice of a maximal admissible quotient of Γ, then c(Γ/Λ) = c(Γ). If
c(Γ) > 1, then h(Γ/Λ) ≥ 3.
Proof. By definition, there exists k ∈ Z − {0} such that γ k ∈ Γc(Γ) Suppose for a contradiction that
c(Γ/Λ) < c(Γ). That implies Γc(Γ) ≤ ker(πΛ ), and hence, πΛ (γ k ) = 1. Since Γ/Λ is torsion free,
it follows that πΛ (γ ) = 1. That contradicts the construction of Γ/Λ, and thus, c(Γ/Λ) = c(Γ).
Since every irreducible, torsion free admissible group Γ where c(Γ) ≥ 2 contains a subgroup
isomorphic to the 3-dimensional integral Heisenberg group, we have h(Γ/Λ) ≥ 3.
The following proposition relates the value ψRF (Γ) to the value ψRF (Λ) when Λ is a torsion-free
quotient of Γ of lower step length.
√
Proposition 3.9. Let Γ be a torsion free admissible group of step length c > 1. If M = Z(Γ) Γc ,
then ψRF (Γ) ≥ ψRF (Γ/M).
h(N)
h(Z(Γ/M))
and integers {ti }i=1 satisfying the following. The
Proof. There exist elements {zi }i=1
h(Z(Γ/M))
set {πM (zi )}i=1
generates Z(Γ/M)
ai ∈ (Γ/M)c−2 and bi ∈ Γ/M such that
E
D and there exist
p
h(N)
ti
t
π ([ai , bi ]) = πM (zi ). Finally, the set {πM (zi ) i }i=1 generates Z(Γ/M) (Γ/M)c−1 .
h(Γ )
h(M)
There also exist γi ∈ Γ such that the {[zi , γi ]}i=1c generate Γc . Finally, there exist elements {yi }i=1
h(M)
in Z(Γ) and integers {si }i=1 such that ysi i = [zi , γi ]. For each i ∈ {1, · · · , h(M)}, we let Γ/Λi be a
choice of an admissible quotient with respect to yi .
−1
Let (Γ/M)/Ωi be a choice of an admissible with respect to πM (zi ). It is evident that Λ ≤ πM
(Ωi ).
−1
Thus, it follows that h(Γ/Λi ) ≥ h(Γ/πM (Ωi )) = h((Γ/M)/Ωi ). Proposition 3.7 implies ψRF (Γ) ≥
h((Γ/M)/Ωi ). Applying Proposition 3.7 again, we have that ψRF (Γ) ≥ ψRF (Γ/M).
Effective Separability of Finitely Generated Nilpotent Groups
14
This last proposition demonstrates that the definition of ψRF (Γ) is the maximum value over all
possible Hirsch lengths of choice of admissible quotients with respect to primitive, central nontrivial elements of Γ.
Proposition 3.10. Let Γ be a torsion free admissible group of step length c, and for each primitive
γ ∈ Z(Γ)
− {1}, let Γ/Λγ be a choice of an admissible quotient with respect to γ . Then ψRF (Γ) =
max h(Γ/Λγ ) | γ ∈ Z(Γ) − {1} .
√
Proof. The statement is evident for torsion free, finitely generated, abelian groups since Z(Γ) Γc ∼
=
Γ. Thus, we may assume that c > 1.
√
Let M = Z(Γ) Γc , and let γ ∈ Z(Γ) − {1}. There exists a basis {zi }h(Z(Γ)) for Z(Γ) and integers
h(Γ)
h(Γ )
{ti }i=1c such that ztii i=1 is a basis for Γc . Moreover, there exist ai ∈ Γc−1 and bi ∈ Γ such
that ztii = [ai , bi ]. If γ ∈ M, then by definition of ψRF (Γ) and Proposition 3.7, we have h(Γ/Λγ ) ≤
ψRF (Γ). Thus, we may assume that γ ∈
/ M.
/ M, πM (γ ) 6= 1. Hence, it is evident that (Γ/M)/πM (Λγ ) satisfies Proposition 3.1 for
Since γ ∈
πM (γ ). Thus, if (Γ/M)/Ω is a choice of an admissible quotient with respect to πM (γ ), we note
that Γ/πK−1 (Ω) satisfies Proposition 3.1 for γ . Thus, by definition,
h(Γ/Λγ ) ≤ h(Γ/πK−1 (Ω)) ≤ h((Γ/M)/Ω) ≤ ψRF (Γ/M).
Proposition 3.9 implies that ψRF (Γ/M) ≤ ψRF (Γ). Thus, h(Γ/Λγ ) ≤ ψRF (Γ) giving the result.
Definition 3.11. A choice of a torsion free admissible group Γ with a choice of a maximal admish(Γ)
sible quotient Γ/Λ, and cyclic series {∆i }i=1 a compatible generating subset {ξi }m
i=1 that satisfy
Proposition 3.5 is called an admissible 4-tuple and is denoted as {Γ, Λ, ∆i , ξi }.
Whenever we are given an admissible 4-tuple {Γ, Λ, ∆i , ξi }, we take the Mal’tsev completion to be
constructed as defined in §2.4. We observe that the vectors vi = Log(ξi ) span g the Lie algebra of
h(Γ)
the Mal’tsev completion. We call the subset {vi }i=1 an induced basis for g.
Definition 3.12. An admissible 4-tuple {Γ, Λ, ∆i , ξi }, Mal’tsev completion G with Lie algebra g,
h(Γ)
and induced basis {vi }i=1 is called an admissible 7-tuple and is denoted as {Γ, Λ, ∆i , ξi , G, g, νi }.
h(Γ)
h(Γ)
We always take S = {ξi }i=1 as a generating subset of Γ and X = {νi }i=1 as a basis of g.
If we are given a admissible 4-tuple {Γ, Λ, ∆i , ξi }, then we have a natural choice of an admissible
3-tuple given by {Γ, ∆i , ξi }. Whenever we are given an admissible 7-tuple {Γ, Λ, ∆i , ξi , G, g, νi },
then {Γ, Λ, ∆i , ξi } is an admissible 4-tuple and {Γ, ∆i , ξi } is an admissible 3-tuple.
4
Commutator geometry and lower bounds for residual finiteness
The following definitions and propositions will be important in the construction of the lower
bounds found in the proof of Theorem 1.1.
Effective Separability of Finitely Generated Nilpotent Groups
4.1
15
Finite Index Subgroups and Cyclic Series
The following proposition tells us how to view finite index subgroups in light of a choice of a
cyclic series and compatible generating subset.
Proposition 4.1. Let {Γ, ∆i , ξi } be an admissible 3-tuple. Suppose Γ is torsion free, and let K ≤ Γ
h(Γ)
be a finite index subgroup. Then there exist natural numbers {ti }i=1 satisfying the following. The
h(Γ)
subgroups {Hi }i=1 given by Hi = hξsts iis=1 form a cyclic series for K with a compatible generating
h(Γ)
subset given by {ξiti }i=1 .
Proof. We proceed by induction on Hirsch length. For the base case, assume that h(Γ) = 1. In this
case, we have that Γ ∼
= Z and that K ∼
= tZ for some t ≥ 1. Now the statement of the proposition is
evident by choosing H1 = K and the compatible generating subset is given by t.
Thus, we may assume h(Γ) > 1. Observing that ∆h(Γ)−1 ∩ K is a finite index subgroup of ∆h(Γ)−1
and that h(∆h(Γ)−1 ) = h(Γ) − 1, the inductive hypothesis implies there exist natural numbers
h(Γ)−1
h(Γ)
given by Hi = hξsts iis=1 form a cyclic
{ti }i=1 satisfying the following. The groups {Hi }i=1
h(Γ)−1
series for ∆h(Γ)−1 ∩ K with a compatible generating subset given by {ξiti }i=1 . We also have that
π∆h(Γ)−1 (K) is a finite index subgroup of Γ/∆h(Γ)−1 . Thus, there exists a natural number th(Γ) such
D
E
E
D
t
th(Γ)
h(Γ)
that K/∆h(Γ)−1 ∼
, then the groups {Hi }
(ξ h(Γ) ) . If we set Hh(Γ) ∼
= π∆
= Hh(Γ)−1 , ξ
h(Γ)−1
h(Γ)
i=1
h(Γ)
form a cyclic series for K with a compatible generating subset given by
h(Γ)
{ξiti }i=1 .
We now apply Proposition 4.1 to give a description of the subgroups of Γ of the form Γm for m ∈ N.
Corollary 4.2. Let {Γ, ∆i , ξi } be an admissible 3-tuple such that Γ is torsion free, and let m ∈ N.
The subgroups Hi = hξsm iis=1 form a cyclic series for Γm with a compatible generating subset given
h(Γ)
by {ξim }i=1 . In particular, |Γ/Γm | = mh(Γ) .
h(Γ)
h(Γ)
Proof. Proposition 4.1 implies there exist natural numbers {ti }i=1 such that the subgroups {Hi }i=1
given by Hi = hξsts iis=1 form a cyclic series for Γm with a compatible generating subset given by
ti h(Γ)
h(Γ)
ξi i=1 . We observe that (Γ/∆1 )m ∼
= (Γm /∆1 ). It is also evident that the series {Hi /∆1 }i=2 is a
h(Γ)
cyclic series for (Γ/∆1 )m with a compatible generating subset given by π∆1 (ξiti ) i=2 . Thus,
the inductive hypothesis implies that ti = m for all 2 ≤ i ≤ h(Γ). To finish, we observe that
Γm ∩ ∆ 1 ∼
= ∆m
1 . Thus, t1 = m as desired.
Let Γ be a torsion free admissible group, and let K ≤ Γ be a finite index subgroup. The following
proposition allows us to understand how K intersects a fixed choice of an admissible quotient of Γ
with respect to a primitive central element.
Proposition 4.3. Let {Γ, Λ, ∆i , ξi } be an admissible 4-tuple. Let K be a finite index subgroup of Γ.
ThereDexistEindices 1 ≤ i1 < i2 < · · · iℓ ≤ h(Γ) with natural numbers {ts }ℓs=1 such that the subgroups
t
Hs = ξi jj
s
j=1
form a cyclic series for K ∩ Λ with a compatible generating subset {ξitss }ℓs=1 .
Effective Separability of Finitely Generated Nilpotent Groups
16
h(Γ)
h(Γ)
Proof. By assumption, the cyclic series {∆i }i=1 and compatible generating subset {ξi }i=1 satisfy
the conditions of Proposition 3.5 for Λ. Thus, there exists indices 1 ≤ i1 < i2 < · · · iℓ such that the
j
subgroups M j = hξis is=1
form a cyclic series for Λ with a compatible generating subset given by
h(Λ)
{ξis }s=1 . Applying Proposition 4.1 to the admissible 3-tuple {Γ, ∆i , ξi }, we have that there exist
h(Γ)
natural numbers {ti }i=1 such that the subgroups given by Wi = hξsts iis=1 form a cyclic series for K
h(Γ)
with a compatible generating subset given by ξiti i=1 . Since K ∩ Λ is a finite index subgroup of
D ti Es
Λ, the groups given by Hs = ξi j j
form the desired cyclic series for K ∩ Λ with a compatible
t
j=1
generating subset given by {ξisis }ℓs=1 . Therefore, {tis }ℓs=1 are the desired integers.
4.2
Reduction of Complexity for Residual Finiteness
We first demonstrate that we may assume that Γ is torsion free when calculating FΓ (n).
Proposition 4.4. Let Γ be an infinite admissible group. Then FΓ (n) ≈ FΓ/T (Γ) (n).
Proof. We proceed by induction on |T (Γ)|, and observe that the base case is evident. Thus,
we may assume that |T (Γ)| > 1. Note that πZ(T (Γ)) : Γ → Γ/Z(T (Γ)) is surjective and that
ker(πZ(T (Γ)) ) = Z(T (Γ)) is a finite central subgroup. Since admissible groups are linear, [5, Lem
2.4] implies that FΓ (n) ≈ FΓ/T (Z(Γ)) (n). Since (Γ/Z(T (Γ)))/T (Γ/Z(T (Γ))) ∼
= Γ/T (Γ), the inductive hypothesis implies that FΓ (n) ≈ FΓ/T (Γ) (n).
The following proposition implies that we may pass to a choice of a maximal admissible quotient
of Γ when computing the lower bounds of FΓ (n) when Γ is a torsion free admissible group.
Proposition 4.5. Let {Γ, Λ, ∆i , ξi } be an admissible 4-tuple. If ϕ : Γ → Q is a surjective homomorphism to a finite group, then ϕ (ξ1m ) 6= 1 if and only if πϕ (Λ) (ϕ (ξ1m )) 6= 1 where m ∈ N.
Proof. If Λ ∼
= {1}, then there is nothing to prove. Thus, we may assume that Λ ≇ {1}. Proposition
3.5 implies that ξ1 ∈
/ Λ and that there exists a collection of elements of the Mal’tsev basis {ξis }ℓs=1
h(Λ)
s
such that Λ ∼
= hξis is=1 . Moreover, we have that Hs = hξit it=1 is cyclic series for Λ with a compatℓ
ible generating subset given by {ξis }ℓs=1 . Proposition
D E4.3 implies there exist {ts }s=1 ⊆ N such that
t
h(Λ)
the series of subgroups {Ws }s=1 given by Ws = ξi jj
s
j=1
forms a cyclic series for ker(ϕ ) ∩ Λ with
h(Λ)
a compatible generating subset given by {ξitss }s=1
Since the backwards direction is clear, we proceed with forward direction. To be more specific,
we demonstrate that if ϕ (ξ1m ) 6= 1, then πϕ (Λ) (ξ1m ) 6= 1. We proceed by induction on |ϕ (Λ)|, and
observe that the base case is clear. Thus, we may assume that |ϕ (Λ)| > 1. In order to apply the
inductive hypothesis, we find a non-trivial normal subgroup M ≤ Z(Q) such that ϕ (ξ1m ) ∈
/ M.
We first observe that if ϕ (ξi0 ) 6= 1 for some i0 ∈ {2, · · · , h(Z(Γ))}, we may set M = hϕ (ξi0 )i. It is
/ M. Thus, we may assume that ξi ∈ ker(ϕ ) for
straightforward to see that M 6= {1} and that ϕ (z) ∈
i ∈ {2, · · · , h(Z(Γ))}.
Effective Separability of Finitely Generated Nilpotent Groups
17
In this next paragraph, we prove that there exists an element of the compatible generating subset,
/ ker(ϕ ), and ϕ (ξi0 ) ∈ Z(Q). To that end, we note that if |tis | = 1,
say ξi0 , such that ξi0 ∈ Λ, ξi0 ∈
then ξis ∈ ker(ϕ ). Since |ϕ (Λ)| > 1, the set E = {ξis | |tis | =
6 1} is non empty. Given that E is
a finite set, there exists ξis0 ∈ E such that Height(ξis0 ) = min{Height(ξis ) | ξis ∈ E}. We claim
that ϕ (ξis0 ) is central in Q, and since we are assuming that ϕ (ξi ) = 1 for i ∈ {2, · · · , h(Z(Γ))},
we may assume that Height(ξis0 ) > 1. Since Height([ξis0 , ξt ]) < Height(ξis0 ) for any ξt and that
ϕ (Λ) E Q, it follows [ϕ (ξis0 ), ϕ (ξt )] ∈ ϕ (Λ). Thus, [ϕ (ξis0 ), ϕ (ξt )] is a product of ϕ (ξis j ) where
Height(ξis j ) < Height(ξis0 ). Since ξis j ∈ ϕ (Λ) and Height(ξis j ) < Height(ξis0 ), the definition of
E and the choice of ξi0 implies that tis j = 1. Thus, ξis j ∈ ker(ϕ ), and subsequently, ϕ (ξis j ) = 1.
Hence, [ϕ (ξis0 ), ϕ (ξt )] = 1, and thus, ϕ (ξi0 ) ∈ Z(Q) − {1}.
Since ϕ (ξis0 ) is central in Q, the group M = hϕ (ξi0 )i is a normal subgroup of Q. By selection,
ϕ (ξ1m ) ∈
/ M, and since |πM (ϕ (Λ))| < |ϕ (Λ)|, we may apply the inductive hypothesis to the surjective homomorphism πM ◦ ϕ : Γ → Q/M. Letting N = πM ◦ ϕ (Λ), we have that πN (πM (ϕ (ξ1m ))) 6= 1.
Thus, πϕ (Λ)(ξ1m ) 6= 1.
4.3
Rank and Step Estimates
Definition 4.6. Let {Γ, ∆i , ξi } be an admissible 3-tuple such that Γ is an infinite admissible group
of step size c, and let ℓ ≤ h(Γ). Let ~a = (ai )ℓi=1 where 1 ≤ ai ≤ h(Γ) for all i. We write [ξ~a ] =
[ξa1 , . . . , ξaℓ ]. We call [ξ~a ] a simple commutator of weight ℓ with respect to ~a. Let Wk (Γ) be the set
of non-trivial simple commutators of weight k. Since Γ is a nilpotent group of step size c, Wk (Γ) is
an empty set for k ≥ c + 1. Thus, the set of non-trivial simple commutators of any weight, denoted
as W (Γ), is finite.
When considering a surjective homomorphism to a finite group ϕ : Γ → Q, we need to ensure that
the step length of Q is equal to the step length of Γ. We do that by assuming that ϕ ([ξ~a ]) 6= 1 for
all [ξ~a ] ∈ W (Γ) ∩ Z(Γ).
Proposition 4.7. Let {Γ, ∆i , ξi } be an admissible 3-tuple, and assume that Γ is torsion free. Let
ϕ : Γ → Q be a surjective homomorphism to a finite group such that if [ξ~a ] ∈ W (Γ) ∩ Z(Γ), then
ϕ ([ξ~a ]) 6= 1. Then ϕ ([ξ~a ]) 6= 1 for all [ξ~a ] ∈ W (Γ). If c1 and c2 are the step sizes of Γ and Q,
respectively, then c1 = c2 .
Proof. We first demonstrate that ϕ ([ξ~a ]) 6= 1 for all [ξ~a ] ∈ W (Γ) by induction on Height([ξ~a ]).
Observe that if [ξ~a ] ∈ Wk (Γ), then Height([ξ~a ]) = c(Γ) − k + 1. Thus, if [ξ~a ] ∈ Wc(Γ) (Γ), then
Height([ξ~a ]) = 1. Hence, the base case follows from assumption.
Now consider [ξ~a ] ∈ W (Γ) where Height([ξ~a ]) = ℓ > 1. If [ξ~a ] ∈ Z(Γ), then the assumptions
of the proposition imply ϕ ([ξ~a ]) 6= 1. Thus, we may assume there exists an element ξi0 of the
Mal’tsev basis such that [[ξ~a ], ξi0 ] 6= 1. The induction hypothesis implies that ϕ ([[ξ~a ], ξi0 ]) 6= 1
since [[ξ~a ], ξi0 ] is a simple commutator of Height([[ξ~a ], ξi0 ]) ≤ ℓ − 1. Thus, ϕ ([ξ~a ]) 6= 1. Therefore,
for each [ξ~a ] ∈ W (Γ), it follows that ϕ ([ξ~a ]) 6= 1.
Effective Separability of Finitely Generated Nilpotent Groups
18
If c1 < c2 , then ϕ factors through Γ/Γc1 , and subsequently Wc1 ⊆ ker(ϕ ). Since Wc1 ⊆ W (Γ) ∩
Z(Γ), we have a contradiction, and subsequently, c1 = c2 .
For a surjective homomorphism to a finite p-group ϕ : Γ → Q, the following proposition gives
conditions for |Q| ≥ ph(Γ) . To be more specific, if ϕ is an injective map when restricted to the set
of central simple commutators and is an injective map when restricted to a central elements of a
fixed compatible generating subset, then ϕ is an injection when restricted to that same compatible
generating subset.
Proposition 4.8. Let {Γ, ∆i , ξi } be an admissible 3-tuple such that Γ is torsion free. Let ϕ : Γ → Q
be a surjective homomorphism to a finite p-group. Suppose that ϕ ([ξ~a ]) 6= 1 for all [ξ~a ] ∈ W (Γ) ∩
Z(Γ). Also, suppose that ϕ (ξi ) 6= 1 for ξi ∈ Z(Γ) and ϕ (ξi ) 6= ϕ (ξ j ) for ξi , ξ j ∈ Z(Γ) where i 6= j.
Then ϕ (ξi ) 6= 1 for 1 ≤ i ≤ h(Γ) and ϕ (ξi ) 6= ϕ (ξ j ) for 1 ≤ j1 < j2 ≤ h(Γ).
/ Z(Γ). By selection, there exists ξ j2 such that [ξ j1 , ξ j2 ] 6= 1. Since [ξ j1 , ξ j2 ] is a
Proof. Let ξ j1 ∈
simple commutator of weight 2, Proposition 4.7 implies that ϕ ([ξ j1 , ξ j2 ]) 6= 1. Thus, ϕ (ξ j1 ) 6= 1.
We now demonstrate that ϕ (ξi ) 6= ϕ (ξ j ) for all i < j. Assume for a contradiction that ϕ (ξi ) =
ϕ (ξ j ). That implies ξi ξ j−1 ∈ ker(ϕ ). Since ker(ϕ ) is a normal finite index subgroup, Proposition
h(Γ)
h(Γ)
4.1 implies that there exist natural numbers {ti }i=1 such that the subgroups {Hi }i=1 given by Hi ∼
=
ti h(Γ)
i
t
s
hξs is=1 form a cyclic series for ker(ϕ ) with a compatible generating subset given by ξi i=1 .
/ ker(ϕ ) for all i, each ti must be a non-zero power of p. We may
Since Q is a p-group and ξi ∈
h(Γ) sℓ tℓ
−1
write ξi ξ j = ∏ℓ=1 ξℓ for sℓ ∈ Z. We observe that si ti is divisible by p for all i, and since each
element can be represented uniquely in terms of its Mal’tsev coordinates, it follows that sℓ = 0 for
s t
all ℓ 6= i, j. Thus, we may write ξi ξ j−1 = ξisi ti ξ j j j . That implies 1 = si ti and −1 = s j t j . Since p
divides si ti , we have a contradiction. Thus, ϕ (ξi ) 6= ϕ (ξ j ).
h(Γ)
Thus, {ϕ (ξi )}i=1 is a generating subset of Q where OrdQ (ϕ (ξi )) ≥ p for all i. [16, Thm 1.10]
implies that |Q| divides some power of ph(Γ) . Hence, |Q| ≥ ph(Γ) .
The following definition will be important in the proofs of Theorem 1.1 and Theorem 1.8.
Definition 4.9. Let {Γ, ∆i , ξi } be an admissible 3-tuple such that Γ is torsion free and h(Z(Γ)) = 1.
k([ξ ])
For [ξ~a ] ∈ W (Γ) ∩ Z(Γ), we let k ([ξ~a ]) satisfy ξ1 ~a = [ξ~a ]. Let B(Γ) = lcm{|k ([ξ~a ]) | | [ξ~a ] ∈
W (Γ) ∩ Z(Γ)}.
Proposition 4.10. Let {Γ, ∆i , ξi } be a admissible 3-tuple where Γ is torsion free and h(Z(Γ)) = 1.
Suppose ϕ : Γ → Q is a surjective homomorphism to a finite p-group such that p > B(Γ), and
suppose that ϕ (ξ1 ) 6= 1. Then c1 = c2 , Z(Q) = hϕ (ξ1 )i and |Q| ≥ ph(Γ) where c1 and c2 are the
step lengths of Q and Γ, respectively.
Proof. Since ϕ (ξ1 ) 6= 1 and Q is p-group, we have OrdQ (ϕ (ξ1 )) ≥ p. We claim that if [ξ~a ] ∈
W (Γ) ∩ Z(Γ), then ϕ ([ξ~a ]) 6= 1. Suppose for a contradiction that ϕ ([ξ~a ]) = 1 for some [ξ~a ] ∈
B(Γ)
B(Γ)
W (Γ) ∩ Z(Γ). Since ϕ (ξ1 ) is a power of ϕ ([ξ~a ]) by definition, we have ϕ (ξ1 ) = 1. Thus,
ϕ (ξ1 ) has order strictly less than p which is a contradiction.
Effective Separability of Finitely Generated Nilpotent Groups
19
Since ϕ ([ξ~a ]) 6= 1 for [ξ~a ] ∈ W (Γ) ∩ Z(Γ), Proposition 4.7 implies that c1 = c2 . On the other
hand, Proposition 4.8 implies that ϕ (ξi ) 6= 1 for all 1 ≤ i ≤ h(Γ) and ϕ (ξ j1 ) 6= ϕ (ξ j2 ) for all
h(Γ)
h(Γ)
1 ≤ j1 < j2 ≤ h(Γ). Thus, {ϕ (∆i )}i=1 is a cyclic series for Q and {ϕ (ξi )}i=1 is a compatible
generating subset for Q. Since Q is a p-group, |ϕ (∆i ) : ϕ (∆i−1 )| ≥ p for each 1 ≤ i ≤ h(Γ)
with the convention that ∆0 = {1}. Hence, the second paragraph after [19, Defn 8.2] implies
h(Γ)
|Q| = ∏i=1 |∆i : ∆i−1 | ≥ ph(Γ) .
h(Γ)
We finish by demonstrating Z(Q) = hϕ (ξ1 )i. Since {ϕ (∆i )}i=1 is an ascending central series that
is a refinement of the upper central series, there exists i0 such that ϕ (∆i0 ) = Z(Q). For t > 1, there
exists j 6= t such that [ξt , ξ j ] 6= 1. Since [ξt , ξ j ] is a simple commutator of weight 2, Proposition
/ Z(Q).
4.7 implies that ϕ ([ξt , ξ j ]) 6= 1. Given that ϕ ([ξt , ξ j ]) = [ϕ (ξt ), ϕ (ξ j )], it follows ϕ (ξt ) ∈
That implies ϕ (∆t ) Z(Q) for all t > 1. Hence, Z(Q) = hϕ (ξ1 )i.
If {Γ, Λ, ∆i , ξi } is an admissible 4-tuple with a surjective homomorphism to a finite group ϕ : Γ →
Q and m ∈ Z − {0}, then the following proposition gives conditions such that Q has no proper
quotients in which ϕ (ξ1m ) 6= 1.
Proposition 4.11. Let {Γ, Λ, ∆i , ξi } be an admissible 4-tuple. Suppose that ϕ : Γ → Q is a surjective homomorphism to a finite p-group where ϕ (Λ) ∼
= {1}, p > B(Γ/Λ), and |Q| ≤ pψRF (Γ) . If
m
ψ
(Γ)
ϕ (ξ1 ) 6= 1 for some m ∈ Z, then |Q| = p RF . Additionally, if N is a proper quotient of Q, then
ρ (ϕ (ξ1m )) = 1 where ρ : Q → N is the natural projection. Finally, Z(Q) ∼
= Z/pZ.
Proof. Let us first demonstrate that |Q| = pψRF (Γ) . Since Λ ≤ ker(ϕ ), we have an induced homomorphism ϕe : Γ/Λ → Q such that ϕe ◦ πΛ = ϕ . Since ϕe : Γ/Λ → Q is a surjective homomorphism
to a finite p-group where p > B(Γ/Λ), h(Z(Γ/Λ)) = 1, and ϕ (ξ1 ) 6= 1, Proposition 4.10 implies
that ϕ (Z(Γ/Λ)) ∼
= Z(Q) and |Q| ≥ pψRF (Γ) . Therefore, |Q| = pψRF (Γ) .
We now demonstrate that Z(Q) ∼
= Z/pZ. Since ϕ (∆i /Λ) is a cyclic series for Q with a compatible
generating given by {ϕ (ξi ) | ξi ∈
/ Λ}, it follows that |Q| = ∏ξi ∈Λ
/ OrdQ (ϕ (ξi )) (see the second paragraph after [19, Defn 8.2]). Thus, we must have that OrdQ (ϕ (ξi )) ≤ p. Since OrdQ (ϕ (ξ1 )) ≥ p.
we have OrdQ (ϕ (ξ1 )) = p. Since Z(Q) ∼
= hϕ (ξ1 )i, it follows that Z(Q) ∼
= Z/pZ.
Since Z(Q) ∼
= Z/pZ, there are no proper, non-trivial subgroups of Z(Q). Given that ker(ρ ) E Q,
we have Z(Q) ∩ ker(ρ ) = Z(Q); hence, ρ (ϕ (ξim )) = 1 because ϕ (ξ1m ) ∈ Z(Q) ≤ ker(ρ ).
5
Some Examples of Precise Residual Finiteness
To demonstrate the techniques used in the proof of Theorem 1.1, we make a precise calculation of
FH2m+1 (Z) (n) where H2m+1 (Z) is the (2m + 1)-dimensional integral Heisenberg group.
Effective Separability of Finitely Generated Nilpotent Groups
5.1
20
Integral Heisenberg Group Basics
We start by introducing basic facts about the (2m + 1)-dimensional integral Heisenberg group
which will be useful in the calculation of FH2m+1 (Z) (n) and in Section 9. We may write
1 ~x z
H2m+1 (Z) = ~0 Im ~y z ∈ Z, ~x,~yT ∈ Zm
0 ~0 1
where Im is the m × m identity matrix. If γ ∈ H2k+1 (Z), we write
1 ~xγ zγ
γ = ~0 Im ~yγ
0 ~0 1
where ~xγ = [xγ ,1 , . . . , xγ ,m ] and ~yTγ = [yγ ,1 , . . . , yγ ,m ].
m
We let E = {~ei }m
i=1 be the standard basis of Z and then choose a generating subset for H2m+1 (Z)
given by S = {α1 , . . . , αm , β1 , . . . , βm , λ } where
~
~
1 ~ei 0
1 0 0
1 0 1
αi = ~0 Im ~0 , βi = ~0 Im ~eTi , and λ = ~0 Im ~0 .
0 ~0 1
0 ~0 1
0 ~0 1
Thus, if γ ∈ BH2m+1 (Z),S (n), then ~xγ ,~xη ,~yTγ ,~yTη ∈ BZm ,E (C0 n) and |zγ | ≤ C0 n2 for some C0 ∈ N [15,
3.B2]. Lastly, we obtain a finite presentation for H2m+1 (Z) written as
H2m+1 (Z) = κ , µi , ν j for 1 ≤ i, j ≤ m | [µt , νt ] = κ for 1 ≤ t ≤ m
(4)
with all other commutators being trivial.
5.2
Residual Finiteness of H2m+1 (Z)
Since the upper and lower asymptotic bounds for FH2m+1 (Z) (n) require different strategies, we
approach them separately. We start with the upper bound as it is more straight forward.
Before we begin with the upper bound, we collect some basic facts. We take presentation given in
t−1
Equation D4 with S = {µi , ν j , κ |1
E ≤ i, j ≤ m}. Let ∆1 = hκ i, ∆i = h{κ } ∪ {µs }is=1 for 2 ≤ i ≤ m+1,
i−m−1
2m+1
and ∆i = {∆m+1 } , {νt }t=1
for m+2 ≤ i ≤ 2m+1. One can see that {∆i }i=1
is a cyclic series
for H2m+1 (Z) and that S is a compatible generating subset.
Proposition 5.1. FH2m+1 (Z) (n) (log(n))2m+1 .
Proof. Let kγ kS ≤ n. We seek to construct a surjective homomorphism ϕ : H2m+1 (Z) → Q to
a finite group such that ϕ (γ ) 6= 1. Moreover, we want to construct this finite group such that
|Q| ≤ C0 (log(C0 n))2m+1 for some C0 > 0.
Effective Separability of Finitely Generated Nilpotent Groups
21
λj
βi
m
Via the Mal’tsev basis, we may write γ = κ α ∏m
µ
ν
. We proceed based on
∏
j=1 j
i=1 i
whether γ has a trivial image in the abelianization or not.
Suppose that πab (γ ) 6= 1. Since γ 6= 1, either βi 6= 0 for some i, or λ j 6= 0 for some j. Without
loss of generality, we may assume that there exists some i0 such that βi0 6= 0. The Prime Number
Theorem [37, 1.2] implies there exists a prime p such that p ∤ |βi0 | and where p ≤ C2 log(C2 |βi0 |) ≤
C2 log(C1 C2 n2 ). Consider the map ρ : H2m+1 (Z) → Z/pZ given by
κα
m
βi
∏ µi
i=1
!
m
λj
∏ νj
j=1
!
−→ (β1 , · · · , βm , λ1 , · · · , λm ) −→ βi0 −→ βi0 ( mod p ).
Here, the first arrow is the abelianization map, the second arrow is the projection to the βi0 coordinate, and the last arrow is the natural projection from Z to Z/pZ. By construction, ρ (γ ) 6= 1 and
|Z/pZ| ≤ C1 C2 log(C1 C2 n2 ). Thus, DH2m+1 (Z) (γ ) ≤ C3 log(C3 n) for some C3 > 0.
Now suppose that πab (γ ) = 1. That implies βi , λ j = 0 for all i, j. As before, the Prime Number Theorem [37, 1.2] implies there exists a prime p such that p ∤ |α | and p ≤ C4 log(C4 n) for
some C4 ∈ N. We have that σ p (γ ) = σ p (κ α1 ) 6= 1. The second paragraph after [19, Defn 8.2]
implies that | H2m+1 (Z)/(H2m+1 (Z)) p | = p2m+1 since |∆i+1 : ∆i | = p for 1 ≤ i ≤ 2m + 1. Hence,
| H2m+1 (Z)/(H2m+1 (Z)) p | ≤ (C4 )2m+1 (log(C4 n))2m+1 . Thus, DH2m+1 (Z) (γ ) ≤ C4 (log(C4 n))2m+1 ,
and therefore, FH2m+1 (Z) (n) (log(n))2m+1 .
Proposition 5.2. (log(n))2m+1 FH2m+1 (Z) (n).
Proof. In order to demonstrate that (log(n))2m+1 FH2m+1 (Z) (n), we construct a sequence of elements {γi } such that C1 (log(C1 kγi kS ))2m+1 ≤ DH2m+1 (Z) (γi ) for some C1 > 0. The proof of Proposition 5.1 implies that when γ ∈
/ Z(H2m+1 (Z)) that DH2m+1 (Z) (γ ) ≤ C1 log(C1 kγ kS ) for some C1 ∈ N.
Thus, the elements we are looking for will be central elements.
Let {pi } be an enumeration of the primes, and let αi = (lcm{1, 2, · · · , pi − 1})2m+2 . We claim for
all i that DH2m+1 (Z) (κiα ) ≈ log(kκ αi kS ))2m+1 . It is clear that κ̄ αi 6= 1 in H2m+1 (Z)/(H2m+1 (Z)) pi .
p
[15, 3.B2] implies that kκ αi kS ≈ |αi |, and the Prime Number Theorem [37, 1.2] implies that
log(|αi |) ≈ pi . Subsequently, log(kκ αi kS ) ≈ pi , and thus, (log(kκ αi kS ))2m+1 ≈ pi2m+1 . Given
that | H2m+1 (Z)/ (H2m+1 (Z)) pi | = pi2m+1 , we establish that (log(kγi kS ))2m+1 ≈ DH2m+1 (Z) (γi ) by
demonstrating for all surjective homomorphisms ϕ : H2m+1 (Z) → Q to finite groups satisfying
|Q| < pi2m+1 that ϕ (κ )αi = 1.
[16, Thm 2.7] implies that we may assume that |Q| = qβ where q is a prime. Since ϕ (κiα ) = 1 when
ϕ (κ ) = 1, we may assume that ϕ (κ ) 6= 1. Since [µi , νi ] = κ for all i, it follows that ϕ (νi ), ϕ (µ j ) 6= 1
for all i, j and that |Q| ≥ q2m+1 (see the second paragraph after [19, Defn 8.2]).
Suppose Q is a pi -group. If ϕ (γi ) 6= 1, then Proposition 4.11 implies that |Q| = pi2m+1 and that
there are no proper quotients of Q where the image of ϕ (γi ) does not vanish. In particular, there
are no proper quotients of H2m+1 (Z)/(H2m+1 (Z)) pi where σ pi (γi ) does not vanish. Thus, we may
assume that q 6= pi .
Effective Separability of Finitely Generated Nilpotent Groups
22
If q > pi , then we have OrdQ (ϕ (νi )), OrdQ (ϕ (µ j )) ≥ pi for all i, j. That implies |∆i : ∆i−1 | > pi .
Thus, the second paragraph after [19, Defn 8.2] implies that |Q| > pi2m+1 ; hence, we may disregard
this possibility. We now assume that Q is a q-group where q < pi . If qβ < p, then |Q| | αi . Since
the order of an element of a finite group divides the order of the group, we have λ | αi where
λ = OrdQ (ϕ (κ )). Thus, ϕ (γi ) = 1.
Hence, we may assume that Q is a q-group where q < pi and pi < qβ < pi2m+1 . There exists
v such that q(2m+1)v < pi2m+1 < p(2m+1)(v+1) . Thus, we may write β = vt + r where t ≤ 2m + 1
and 0 ≤ r < t. By construction, q(2m+1)t+r ≤ αi , and since q < pi , it follows that q(2m+1)t+r | αi .
Subsequently, λ | αi and ϕ (κ αi ) = 1 as desired.
Corollary 5.3. Let H2m+1 (Z) be the integral Heisenberg group. Then FH2m+1 (Z) (n) ≈ (log(n))2m+1 .
6
Proof of Theorem 1.1
Our goal for Theorem 1.1 is to demonstrate that FΓ (n) ≈ (log(n))ψRF (Γ) . Proposition 4.4 implies
that we may assume that Γ is torsion free. We proceed with the proofs of the upper and lower
asymptotic bounds for FΓ (n) separately since they require different strategies. We start with the
upper bound as its proof is simpler.
For the upper bound, our task is to prove for any non-identity element γ ∈ Γ there exists a surjective
ψRF (Γ)
ϕ : Γ → Q such that ϕ (γ ) 6= 1 and |Q| ≤ C0 (log(C
homomorphism to a finite groupp
p 0 kγ kS ))
Z(Γ)
Z(Γ)
/
Γc(Γ) , we pass to thep
quotient given by Γ/
Γc(Γ) and then
for some C0 ∈ N. When γ ∈
appeal to induction on step length. Otherwise, for γ ∈ Z(Γ) Γc(Γ) , we find a choice of an admissible
quotient of Γ with respect to some primitive central element in which γ has a non-trivial image.
Proposition 6.1. Let Γ be a torsion free admissible group. Then FΓ (n) (log(n))ψRF (Γ) .
h(Γ)
h(Γ)
Proof. Let {∆i }i=1 be a cyclic series with a compatible generating subset {ξi }i=1 . By assumption,
h(Γ )
h(Γ )
there exist integers {ti }i=1c such that ξiti i=1c is a basis for Γc and there exist ai ∈ Γc−1 and bi ∈ Γ
such that ξiti = [ai , bi ] for 1 ≤ i ≤ h(Γc ).
Suppose γ ∈ Γ such that kγ kS ≤ n. Using the Mal’tsev coordinates of γ , we may write γ =
h(Γ) α
∏i=1 ξi i , and Lemma 2.9 implies that |αi | ≤ C1 nc(Γ) for some C1 ∈ N for all i. We construct
a surjective homomorphism ϕ : Γ → Q to a finite group where ϕ (γ ) 6= 1 and |Q| ≤ C2 (kγ kS )ψRF (Γ)
for some C2 > 0.
p
Letting M = Z(Γ) Γc(Γ) , suppose that πM (γ ) 6= 1. Passing to the group Γ/M, the inductive hypothesis implies there exists a surjective homomorphism ϕ : Γ/M → Q such that ϕ (γ̄ ) 6= 1, and DΓ (γ ) ≤
C3 (log(C3 n))ψRF (Γ/M) for some C3 ∈ N. Proposition 3.9 implies that ψRF (Γ/M) ≤ ψRF (Γ), and
thus, DΓ (γ ) ≤ C3 (log(C3 n))ψRF (Γ) .
h(Γ )
Otherwise, we may assume that γ ∈ M. Therefore, we may write γ = ∏i=1c ξiαi , and since γ 6= 1,
there exists 1 ≤ j ≤ h(Γc ) such that α j 6= 0. The Prime Number Theorem [37, 1.2] implies that
there exists a prime p such that p ∤ |α j | and p ≤ C4 log(C4 |α j |) for some C4 ∈ N. If Γ/Λ j is a choice
Effective Separability of Finitely Generated Nilpotent Groups
23
of an admissible quotient with respect to ξ j , then γ̄ 6= 1 in Γ/Λ j · Γ p . Corollary 4.2 implies |Γ/Λ j ·
h(Γ/Λ )
j
Γ p | ≤ C4
· (log(C4 |α j |))h(Γ/Λ j ) . Proposition 3.7 implies that h(Γ/Λ j ) ≤ ψRF (Γ). Thus, we
have DΓ (γ ) ≤ C5 (log(C5 n))ψRF (Γ) for some C5 ∈ N. Hence, FΓ (n) (log(n))ψRF (Γ) .
In order to demonstrate that (log(n))ψRF (Γ) FΓ (n), we require a sequence of elements {γ j } ⊆
Γ such that C1 (log(C1 kγ j kS ))ψRF (Γ) ≤ DΓ (γ j ) for some C1 ∈ N independent of j. That entails
finding elements that are of high complexity with respect to residual finiteness, i.e. non-identity
elements that have relatively short word length in comparison to the order of the minimal finite
group required to separate them from the identity.
Proposition 6.2. Let Γ be torsion free admissible group. Then (log(n))ψRF (Γ) FΓ (n).
Proof. Let Γ/Λ be a choice of a maximal admissible quotient of Γ. There exists a g ∈ Z(Γ) − {1}
such that Γ/Λ is an admissible quotient with respect to γ . Moreover, there exists a k ∈ Z − {0},
a ∈ Γc−1 , and b ∈ Γ such that gk = [a, b]. If g is not trivial, then there exists a primitive xΛ ∈ Z(Γ)
such that xs = g for some s. In particular, xΛ is a primitive, central non-trivial element such that
xsΛk = [a, b].
h(Γ)
h(Γ)
Let {∆i,Λ }t=1 be a choice of cyclic series with a compatible generating subset given by {ξi,Λ }i=1
that satisfy Proposition 3.5 for Λ such that ξ1,Λ = xΛ . Proposition 3.5 implies that Γ/Λ is a choice
of an admissible quotient with respect to ξ1,Λ . Let α j,Λ = (lcm{1, 2, · · · , p j,Λ − 1})ψRF (Γ)+1 where
α
{p j,Λ } is an enumeration of primes greater than B(Γ/Λ). Letting γ j,Λ = ξ1 j,Λ , we claim that {γ j,Λ }
is our desired sequence.
Before continuing, we make some remarks. The value B(Γ/Λ) depends on the choice of a maximal
admissible quotient of Γ. To be more specific, if Γ/Ω is another choice of a maximal admissible
quotient of Γ, then, in general, Γ/Λ ≇ Γ/Ω, and subsequently, B(Γ/Λ) 6= B(Γ/Ω). As a natural
consequence, the sequence of elements {γi,Λ } depends on the choice of a maximal admissible
quotient of Γ. However, we will demonstrate that the given construction will work for any choice
of a maximal admissible quotient we take.
We claim for all j that DΓ (γ j,Λ ) ≈ (log(p j,Λ ))ψRF (Γ) . It is evident that γ j,Λ 6= 1 in Γ/Λ · Γ p j,Λ , and
ψRF (Γ)
Proposition 4.2 implies that |Γ/Λ · Γ p j,Λ | = p j,Λ
. To proceed, we demonstrate for all surjective
ψ (Γ)
RF
homomorphisms ϕ : Γ → Q to finite groups where |Q| < p j,Λ
that ϕ (γ ) = 1.
[16, Thm 2.7] implies that we may assume that |Q| = qβ where q is a prime. If ξ1,Λ ∈ ker(ϕ ), then
ϕ (γ j,Λ ) = 1. Hence, we may assume that ϕ (ξ1,Λ ) 6= 1. Proposition 4.5 implies that ϕ (γ j,Λ ) 6= 1 if
and only if πϕ (Λ) (ϕ (γ j,Λ )) 6= 1. Thus, we may restrict our attention to surjective homomorphisms
that factor through Γ/Λ, i.e. homomorphisms ϕ : Γ → Q where ϕ (Λ) ∼
= {1}.
Suppose that q = p j,Λ . If ϕ (γ j,Λ ) = 1, then there is nothing to prove. So we may assume that
ψRF (Γ)
ψRF (Γ)
, Proposition 4.11 implies that |Q| = p j,Λ
and that if N is a
ϕ (γ j,Λ ) 6= 1. Since |Q| ≤ p j,Λ
proper quotient of Q with natural projection given by ρ : Q → N, then ρ (ϕ (γ j,Λ )) = 1. We have
two natural consequences. There are no proper quotients of Γ/Λ · Γ p j,Λ where ϕ (γ j,Λ ) has nontrivial image. Additionally, If ϕ : Γ → Q is a surjective homomorphism to a finite p j,Λ -group
Effective Separability of Finitely Generated Nilpotent Groups
ψ (Γ)
RF
where |Q| < p j,Λ
24
then ϕ (γ j,Λ ) = 1. Thus, we may assume that q 6= p j,Λ .
Suppose that q > p j,Λ . Since ϕe : Γ/Λ → Q is a surjective homomorphism to a finite q-group where
ψRF (Γ)
q > B(Γ/Λ), Proposition 4.10 implies that |Q| > p j,Λ
. Hence, we may assume that q < p j,Λ .
Now suppose that Q is a q-group where |Q| < p j,Λ . By selection, it follows that |Q| divides α j,Λ .
Since the order of an element divides the order of the group, we have that OrdQ (ϕ (ξ1,Λ )) divides
α j,Λ . That implies ϕ (γ j,Λ ) = 1.
Now suppose that Q is a q-group where q < p j,Λ and qβ > p j,Λ . Thus, there exists ν ∈ N such
ψRF (Γ)
that qν ψRF (Γ) < p j,Λ
< q(ν +1) ψRF (Γ) . Subsequently, we may write β = ν t + r where t ≤ ψRF (Γ)
and 0 ≤ r < ν . By construction, qvt+r ≤ α j,Λ , and since q < p j,Λ , it follows that qβ = qvt+r | α j .
Given that the order of any element in a finite group divides the order of the group, it follows that
ψRF (Γ)
OrdQ (ϕ (ξ1,Λ )) divides α j,Λ . Thus, ϕ (γΛ, j ) = 1, and therefore, DΓ (γ j,Λ ) = p j,Λ
.
Since γ j,Λ ∈ Γc where c is the step length fo Γ, [15, 3.B2] implies (kγ j,Λ kS ) ≈ (|α j,Λ |)1/c , and
ψ (Γ)
the Prime Number Theorem [37, 1.2] implies log(|α j,Λ |) ≈ p j,Λ . Hence, log(kγ j,Λ kS ) RF ≈
ψ (Γ)
ψRF (Γ)
p j,Λ
. Thus, DΓ (γ j,Λ ) ≈ log(kγ j,Λ kS ) RF , and subsequently, (log(n))ψRF (Γ) FΓ (n).
Proof. Theorem 1.1
Let Γ be an infinite admissible group. Proposition 4.4 implies that FΓ (n) ≈ FΓ/T (Γ) (n). Proposition 6.1 implies that FΓ/T (Γ) (n) (log(n))ψRF (Γ) , and Proposition 6.2 implies (log(n))ψRF (Γ)
FΓ/T (Γ) (n). Thus, FΓ (n) ≈ (log(n))ψRF (Γ) .
7
Cyclic series, lattices in nilpotent Lie groups, and Theorem 1.3
Let Γ be a torsion free admissible group, and let G be its Mal’tsev completion. Let Γ/Λ be a
choice of a maximal admissible quotient. The main task of this section is the demonstration that
the value h(Γ/Λ) is a well-defined invariant of the Mal’tsev completion of Γ. Thus, we need
to establish some properties of cocompact lattices in admissible Lie groups. We start with the
following lemma that relates the Hirsch lengths of centers of cocompact lattices within the same
admissible Lie group.
Lemma 7.1. Let G be an admissible Lie group with two cocompact lattices Γ1 and Γ2 . Then
dim(G) = h(Z(Γ1 )) = h(Z(Γ2 )).
Proof. This proof is a straightforward application of [12, Lem 1.2.5].
We now introduce the notion of one parameter families of group elements of a Lie group.
Definition 7.2. Let G be a connected, simply connected Lie group. We call a map f : R → G a
one parameter family of group elements of G if f is an injective group homomorphism from the
real line with addition.
Effective Separability of Finitely Generated Nilpotent Groups
25
Let {Γ, ∆i , ξi } be an admissible 3-tuple such that Γ is torsion free. Via the exponential map and [12,
Lem 1.2.5], the maps given by fΓ,i (t) = exp(t vi ) are one parameter families of group elements.
The discussion below [12, Thm 1.2.4 Pg 9] implies that we may uniquely write each g ∈ G as
h(Γ)
g = ∏i=1 fΓ,i (ti ) where G is the Mal’tsev completion of Γ.
Definition 7.3. We say that the one parameter families fΓ,i are associated to {Γ, ∆i , ξi }.
We characterize when a discrete subgroup of an admissible Lie group is a cocompact lattice based
on how it intersects a collection of one parameter families of group elements.
Proposition 7.4. Let G be a Q-defined admissible Lie group, and suppose that Γ is a discrete
subgroup of G. Suppose there exists a collection of one parameter families of group elements of
dim(G)
G, written as fi : R → G for 1 ≤ i ≤ dim(G), such that G is homeomorphic to ∏i=1 fi (R). Then
Γ is a cocompact lattice in G if and only if Γ ∩ fi(R) ∼
= Z for all i.
Proof. Let ρ : G → G/Γ be the natural projection onto the space of cosets. Suppose that there
exists i0 such that fi0 (R) ∩ Γ ≇ Z. Since Γ is discrete in Γ, Γ ∩ fi0 (R) is a discrete subset of fi0 (R).
Given that Γ ∩ fi0 (R) is discrete and not infinite cyclic, we have Γ ∩ fi0 (R) ∼
= {1}. Hence, each
element of the sequence { fi0 (t)}t∈N projects to a unique element of G/Γ. Thus, {ρ ( fi0 (t))}t∈N is
an infinite sequence in G/Γ with no convergent subsequence. Hence, Γ is not a cocompact lattice
of G
Now suppose that fi (R) ∩ Γ ∼
= Z for all i. That implies for each i ∈ {1, · · · , h(Γ)} there exists ti > 0
dim(G)
∼
such that Γ ∩ fi (R) = { fi (n ti ) | n ∈ Z}. Letting E = ∏i=1 fi ([0,ti ]), we claim that E is compact
and that ρ (E) ∼
= G/Γ.
dim(G)
Let f : Rdim(G) → G be the continuous map given by f ( ai , · · · , adim(G) ) = ∏i=1 fi (ai ). Since
dim(G)
∏i=1 [0,ti ] is a closed and bounded subset of Rdim(G) , the Heine-Borel theorem implies that
dim(G)
∏i=1 [0,ti ] is compact. Since f is continuous, E is compact.
dim(G)
We now claim that each coset of Γ in G has a representative in E. Let g = ∏i=1 fi (ℓi ). For
each i, there exists si ∈ Z such that si ti ≤ ℓi ≤ (si + 1) ti . Let ki = ℓi − si ti and write h ∈ E to be
dim(G)
given by h = ∏i=1 f (ki ). By construction, ρ (h) = ρ (g), and subsequently, ρ (g) ∈ ρ (E). Thus,
ρ (E) = ρ (G). Since G/Γ is the image of a compact set under a continuous map, G/Γ is compact.
[32, Thm 2.1] implies that Γ is a cocompact lattice of G.
These next two propositions give some structural information needed about the Mal’tsev completion of a torsion free admissible group and some structural information of choices of an admissible
quotients with respect to some primitive, central non-trivial element.
Proposition 7.5. Let Γ be a torsion free admissible group. Let γ ∈ Z(Γ) − {1} be a primitive
element, and let Γ/Λ be a choice of an admissible quotient with respect to γ . Suppose that G is
the Mal’tsev completion of Γ, and let H be the Mal’tsev completion of Λ. Then H is isomorphic to
a closed, connected, normal subgroup of G.
Effective Separability of Finitely Generated Nilpotent Groups
26
h(Γ)
h(Γ)
Proof. Proposition 3.5 there exists a cyclic series {∆i }i=1 and compatible generating subset {ξi }i=1
h(Λ)
h(Γ)
s
satisfying the following. There exists a subset {ξis }s=1 ⊆ {ξi }i=1 such that if Ws = hξit it=1
, then
h(Λ)
h(Λ)
{Ws }s=1 is a cyclic series for Λ with a compatible generating subset given by {ξis }s=1 where
h(Γ)
ξ1 = γ . Let { fΓ,i }i=1 be the one parameter families of group elements of G associated to the
admissible 3-tuple {Γ, ∆i , ξi }.
[12, Thm 1.2.3] implies that we may view H as a connected subgroup of G. We proceed by
induction on h(Γ) to demonstrate that H is a closed and normal subgroup of G. If h(Γ) = 1, then
Γ = Z. It then follows that G is Lie isomorphic to R and that H ∼
= {1}. Now our claim is evident.
Now suppose that h(Γ) > 1. If h(Z(Γ)) = 1, then we may take Λ = {1} which implies H ∼
= {1}.
h(Z(Γ))
Thus, our claims are evident. Now suppose that h(Z(Γ)) > 1. Let Ω = hξi ii=1 , and let K be
the Mal’tsev completion of Ω. By [12, Lem 1.2.5], it follows that K ≤ Z(G). Thus, K is a closed,
connected, normal subgroup of G.
h(Λ)
We claim that H/K is Mal’tsev completion of πK (Λ). We may write H = ∏s=1 fΓ,is (R). Since Λ
is a cocompact lattice of H, Proposition 7.4 implies that Λ ∩ fΓ,is (R) ∼
= Z for all 1 ≤ s ≤ ℓ. By
Proposition 7.4 again, we have that K ∩ Λ is a cocompact lattice of K. [11, Prop 5.1.4] implies that
πK (Λ) is a cocompact lattice in H/K.
Observe that πK (Λ) ∼
= Λ/Ω. We have that Λ/Ω satisfies Proposition 3.1 for πK (ξ1 ). Thus, the
inductive hypothesis implies that H/K is a closed normal subgroup of G/K. Since H isomorphic
to the pullback of the closed normal subgroup of G/K, H is a closed normal subgroup of G.
The next proposition that G is the Mal’tsev completion of Γ with a choice of an admissible quotient
Γ/Λ with respect to a primitive, central non-trivial element of Γ, then the Mal’tsev completion of
H intersects any cocompact as a cocompact lattice.
Proposition 7.6. Let Γ be a torsion free admissible group. Let γ ∈ Z(Γ) − {1} be a primitive
element, and let Γ/Λ be a choice of an admissible quotient with respect to γ , G be the Mal’tsev
completion of Γ, and let H ≤ G be the Mal’tsev completion of Λ. If Ω ≤ G is another cocompact
lattice of G, then Ω ∩ H is a cocompact lattice of H.
h(Γ)
Proof. Proposition 3.5 implies that there exists a cyclic series {∆i }i=1 and a compatible generat
h(Λ)
h(Γ)
ing subset {ξi }i=1 satisfying the following. There exists a subset ξi j j=1 such that the groups
i
{Wi } where Wi ∼
= ξi j j=1 form a cyclic series for Λ with a compatible generating subset given by
h(Λ)
h(Γ)
ξi j j=1 . Let { fi }i=1 be the associated one parameter families of group elements of the Mal’tsev
∼ ∏h(Λ) fi (R). Proposi∼ ∏h(Γ) fi (R). By construction, H =
completion G of Γ. We may write G =
j=1 j
i=1
tion 7.4 implies that Ω ∩ fi,Γ ∼
= Z for all i. In particular, Ω ∩ fi ,Γ (R) ∼
= Z for all j. Proposition 7.4
j
implies that Ω ∩ H is a cocompact lattice in H as desired.
The following lemma demonstrates that you can select a cyclic series and compatible generating subset for a cocompact lattice in an admissible Lie group by intersecting the lattice with a
collection of one parameter families of group elements.
Effective Separability of Finitely Generated Nilpotent Groups
27
Lemma 7.7. Let G be a Q-defined admissible Lie group with a cocompact lattice Γ. Let fi be a coldim(G)
lection of one parameter families of elements of G such that G is homeomorphic to ∏i=1 fi (R).
i
Let hξi i ∼
= fi (R) ∩ Γ. Then the groups given by ∆i = hξt it=1 form a cyclic series for Γ with a
h(Γ)
compatible generating subset given by {ξi }i=1 .
Proof. We proceed by induction on the dimension of G. If dim(G) = 1, then our statement is
dim(G)−1
evident. Now suppose that dim(G) > 1. If we let H ∼
fi (R), then Proposition 7.4
= ∏i=1
implies that Γ ∩ H is a cocompact lattice in H. The inductive hypothesis implies that the elet
ments ξi given by hξi i ∼
= fi (R) ∩ Γ satisfy the following. The groups given by ∆t = hξi ii=1 for
1 ≤ t ≤ dim(G) − 1 form a cyclic series for Γ ∩ H with a compatible generating subset given by
dim(G)−1
. Since Γ is a cocompact lattice in G, Proposition 7.4 implies that fdim(G) (R) ∩ Γ ∼
{ξi }i=1
= Z.
dim(G)
Letting ∆dim(G) = ∆dim(G)−1 , ξdim(G) , we have that the groups given by {∆i }i=1
series for Γ with a compatible generating subset given by
form a cyclic
dim(G)
{ξi }i=1 .
Let Γ be a torsion free admissible group. We now demonstrate that the value ψRF (Γ) is a welldefined invariant of the Mal’tsev completion of Γ.
Proposition 7.8. Let G be a Q-defined admissible Lie group, and suppose that Γ1 and Γ2 are two
cocompact lattices of G. Then ψRF (Γ1 ) = ψRF (Γ2 ).
Proof. If h(Z(Γ1 )) = 1, then Proposition 7.1 implies that h(Z(Γ2 )) = 1. It then follows from the
definition of ψRF (Γ1 ) and ψRF (Γ2 ) that ψRF (Γ1 ) = h(Γ) = ψRF (Γ2 ).
Therefore, we may assume that h(Z(Γ1 )), h(Z(Γ2 )) ≥ 2. In this case, we demonstrate the equality
by showing that ψRF (Γ1 ) ≤ ψRF (Γ2 ) and ψRF (Γ2 ) ≤ ψRF (Γ1 ).
h(Γ)
Let G be the Mal’tsev completion of G. Let {∆i }i=1 be a cyclic series for Γ with a compatible
h(Γ)
generating subset {ξi }i=1 , and let fi be the associated one parameter families of group elements.
h(Γ)
We may write G ∼
= ∏i=1 fi (R).
h(Γ )
Let {ηi }i=12 ⊆ Γ2 such that hηi i ∼
= Γ2 ∩ fi (R). If we let Wi ∼
= ηj
h(Γ )
that {Wi }i=12 is a
Let ξi be a central
i
, then Proposition
j=1
7.7 implies
h(Γ )
cyclic series for Γ2 with a compatible generating subset given by {ηi }i=12 .
element of the compatible generating subset of Γ, and let Γ1 /Λ be a choice
of an admissible quotient with respect to ξi . Let H be the Mal’tsev completion of Λ. Since
πΛ (ξi ) ∼
= Z(Γ/H), it is evident that hπH ( fi (R))i ∼
= Z(Γ/H). In particular, πH (ηi ) 6= 1. Proposition
7.6 implies that H ∩ Ω is a compact lattice of H and Proposition 7.5 implies that H is a closed
connected normal subgroup of G. Thus, [11, Prop 5.1.4] implies that πH (Ω) is a compact lattice
in G/H. Proposition 7.1 implies that h(Γ2 /Λ) ∼
= h(πH (Γ2 )), it follows that πH (Γ2 ) satisfies the
conditions of Proposition 3.1. If we let Γ2 /Ω be a choice of an admissible quotient with respect
to ηi , it follows that h(Γ/Ω) ≤ πH (Γ2 ) ≤ h(Γ/Λ). By Proposition 3.7, h(Γ2 /Ω) ≤ ψRF (Γ1 ). [12,
Lem 1.2.5] implies that ηi ∈ Z(Γ2 ), and thus, the above inequality holds for each element of the
compatible generating subset of Γ2 in Z(Γ2 ). Therefore, Proposition 3.7 implies that ψRF (Γ2 ) ≤
ψRF (Γ1 ). By interchanging Γ1 and Γ2 , we have ψRF (Γ1 ) ≤ ψRF (Γ2 ).
Effective Separability of Finitely Generated Nilpotent Groups
28
We now come to the main result of this section.
Proof. Theorem 1.3
Suppose that Γ1 and Γ2 are two infinite admissible groups such that Γ1 /T (Γ1 ) and Γ2 /T (Γ2 ) have
isomorphic Mal’tsev completions. Proposition 4.4 implies that FΓ1 (n) ≈ FΓ1 /T (Γ1 ) (n) and FΓ2 (n) ≈
FΓ2 /T (Γ2 ) (n). Theorem 1.1 implies FΓ1 (n) ≈ (log(n))ψRF (Γ1 ) and FΓ2 /T (Γ2 ) (n) ≈ (log(n))ψRF (Γ2 ) .
Proposition 7.8 implies ψRF (Γ1 /T (Γ1 )) = ψRF (Γ2 /T (Γ2 )). Thus, FΓ1 (n) ≈ FΓ2 (n).
8
8.1
Some Examples and the Proof of Theorem 1.5
Free nilpotent groups and Theorem 1.5(i)
Definition 8.1. Let F(X ) be the free group of rank m generated by X . We define N(X , c, m) =
F(X )/(F(X ))c+1 as the free nilpotent group of step size c and rank m on the set X .
Following [21, Sec 2.7], we construct a cyclic series for N(X , c, m) and a compatible generating
subset using iterated commutators in the set X .
Definition 8.2. We call elements of X basic commutators of weight 1 of N(X , c, m), and we choose
an arbitrary linear order for weight 1 basic commutators. If γ1 and γ2 are basic commutators of
weight i1 and i2 , respectively, then [γ1 , γ2 ] is a basic commutator of weight i1 + i2 of N(X , c, m) if
γ1 > γ2 and if γ1 = [γ1,1 , γ1,2 ] where γ1,1 and γ1,2 are basic commutators such that γ1,2 ≤ γ2 .
Basic commutators of higher weight are greater with respect to the linear order than basic commutators of lower weight. Moreover, we choose an arbitrary linear order for commutators of the
same weight.
For xi0 ∈ X , we say that a 1-fold commutator γ contains xi0 if γ = xi0 . Inductively, we say that a
j-fold commutator [γ1 , γ2 ] contains xi0 if either γ1 contains xi0 or γ2 contains xi0 .
Note that any basic commutator of weight greater or equal to 2 must contain two distinct commutators of weight 1 but not necessarily more than 2. Additionally, if γ is a basic commutator of
weight k, then γ can contain at most k distinct basic commutators of weight 1.
It is well known that the number of basic commutators of N(X , c, m) is equal to the Hirsch length
of N(X , c, m). Letting µ be the Möbius function, we may write
!
c
r
1
h(N(X , c, m)) = ∑
∑ µ (d)m d .
r=1 r d|r
h(N(X,c,m))
We label the basic commutators as {ξi }i=1
with respect to the given linear order.
h(N(X,c,m))
i
where ∆i = hξt it=1
Definition 8.3. One can see that the subgroup series {∆i }i=1
is a cyclic
h(N(X,c,m))
is a compatible generating subseries for N(X , c, m), and [21, Cor 2.7.3] implies that {ξi }i=1
Effective Separability of Finitely Generated Nilpotent Groups
29
h(N(X,c,m))
h(N(X,c,m))
the cyclic series of basic commutators for N(X , c, m) and {ξi }i=1
set. We call {∆i }i=1
the compatible generating subset of basic commutators for N(X , c, m).
Proposition 8.4. Let N(X , c, m) be the free nilpotent group of step size c and rank m on the set X =
{xi }m
i=1 . Let γ be a basic commutator of weight c in the set X that contains only Y ( X . There exists
a normal subgroup Ω such that N(X , c, m)/Ω is torsion free where hπΩ (γ )i ∼
= Z(N(X , c, m)/Ω).
Additionally, if η is a j-fold commutator that contains elements of X /Y , then πΩ (η ) = 1.
h(N(X,c,m))
h(N(X,c,m))
be
be the cyclic series of basic commutators, and let {ξi }i=1
Proof. Let {∆i }i=1
the compatible generating subset of basis commutators. By assumption, there exists an i0 ∈
{1, · · · , h(Z(N(X , c, m)))} such that ξi0 = γ . Without loss of generality, we may assume that ξ1 = γ .
c
We will construct a normal descending series {Kt }t=1
such that N(X , c, m)/Kt is torsion free for
each t, πKt (ξ1 ) 6= 1 for each t, and if η is a i-fold commutator that contains only elements of X /Y
where i ≥ t, then πKt (η ) = 1. We will also have that Kt is generated by basis commutators of
weight greater than or equal to t, and finally, we will have hπK1 (ξ1 )i ∼
= Z(N(X , c, m)/K1 ). We
proceed by induction on t.
h(Z(N(X,c,m)))
. Observe that if η is a c-fold commutator
Consider the subgroup given by Kc = hξi ii=2
such that η contains only elements of X /Y , then it follows by construction that πK1 (η ) = 1. Thus,
we have the base case.
Now suppose that subgroup Kt has been constructed for t < c, and let η be a (t − 1)-fold commutator bracket that contains elements of X /Y . It then follows that [η , xi ] contains elements of
X /Y . Thus, πKt ([η , xi ]) = 1 by assumption. Since that is true for all 1 ≤ i ≤ m, we have that
πKt (η ) ∈ Z(N(X , c, m)/Kt ). Let W be the set of basic commutator brackets ξi such that ξi 6= ξ1
/ hπKt (W )i and if η is a ℓ-fold commutator bracket
and πKt (ξi ) is central. By construction, πKt (ξ1 ) ∈
that contains elements of X /Y where ℓ ≥ t − 1, then πKt (η ) ∈ hπKt (W )i. We set Kt−1 ∼
= hKt ,W i,
and suppose that η is a ℓ-fold commutator that contains elements of X /Y and where ℓ ≥ t − 1.
(hπKt (W )i), we have that Kt−1 is
By construction, we have that πKt−1 (η ) = 1. Since Kt−1 ∼
= πK−1
t
normal in N(X , c, m) and Kt ≤ Kt−1 . Finally, it is evident that N(X , c, m)/Kt is torsion free. Hence,
induction gives the construction of Kt for all i. Additionally, the construction of K1 implies that
Z(N(X , c, m)/K1 ) ∼
= K1 , we have our proposition.
= hπK1 (ξ1 )i. Thus, by taking Ω ∼
Proof. Theorem 1.5(i)
Let c ≥ 1 and ℓ ≥ 2, and let Xℓ = {xi }ℓi=1 . Let N(Xℓ , c, ℓ) to be the free nilpotent group of step size
c and rank ℓ on the set Xℓ . Theorem 1.1 implies that there exists a natural number ψRF (N (Xℓ , c, ℓ))
such that FN(Xℓ ,c,ℓ) (n) ≈ (log(n))ψRF (N(Xℓ ,c,ℓ)) . We will demonstrate that (log(n))ψRF (N(Xℓ ,c,ℓ))
(log(n))ψRF (N(Xc ,c,c)) for each ℓ > c, and since N (Xℓ , c, ℓ) is a nilpotent group of step size c and
Hirsch length greater than ℓ, we will have our desired result.
h(N(X ,c,ℓ))
h(N(X ,c,ℓ))
be the compatbe the cyclic series of basic commutators and {ξi }i=1 ℓ
Let {∆i }i=1 ℓ
ible generating subset of basic commutators for N(Xℓ , c, ℓ). For each ξi ∈ Z(N(Xℓ , c, ℓ)), let
N(Xℓ , c, ℓ)/Λi be a choice of an admissible quotient with respect to ξi . Proposition 3.7 implies
there exists an i0 ∈ {1, · · · , h(Z(N(Xℓ , c, ℓ)))} such that h(Γ/Λi0 ) = ψRF (N(Xℓ , c, ℓ)).
Effective Separability of Finitely Generated Nilpotent Groups
30
For each ξi ∈ Z(N(Xℓ , c, ℓ)), there exists a subset Yi ⊆ X such that ξi is a basic commutator of
weight c that contains only elements of Yi . Proposition 8.4 implies that there exists a subgroup Ωi
such that N(Xℓ , c, ℓ)/Ωi satisfies Proposition 3.1 with respect to ξi . Moreover, elements of X −Yi
are contained in Ωi .
There is a natural surjective homomorphism ρi : N(Xℓ , c, ℓ) → N(Yi , c, |Yi |) given by sending elements of X −Yi to the identity. Therefore, we have an induced map ϕ : N(Yi , c, |Yi |) → N(Xℓ , c, ℓ)/Ωi
such that πΩi = ϕ ◦ ρi . In particular, N(Xℓ , c, ℓ)/Ωi ∼
= N(Yi , c, |Yi |)/ρi (Ωi ). Thus, N(Xℓ , c, ℓ)/Ωi satisfies the conditions of Proposition 3.1 for ρi (ξi ). Proposition 3.7 implies that h(N(Xℓ , c, ℓ)/Λi ) ≤
ψRF (N(Yi , c, |Yi |)). Since N(Xℓ , c, ℓ) has step size c, we have that |Yi | ≤ c for any ξi ∈ Z(N(Xℓ , c, ℓ)).
Additionally, we have that N(Yi , c, |Yi |) ∼
= N(X j , c, j) when |Yi | = j. In particular, ψRF (N(Yi , c, |Yi |)) =
ψRF (N(X j , c, j)). By setting m(c) = max{ψRF (N(X j , c, j)) | 1 ≤ j ≤ c}, Proposition 3.7 implies
FN(Xℓ ,c,ℓ) (n) (log(n))m(c) .
8.2
Central products and applications
The examples we contruct for Theorem 1.5(ii), (iii) and (iv) arise as iterated central products of
torsion free admissible groups whose centers have Hirsch length 1. In the given context, Corollary
1.2 allows us to compute the precise residually finiteness function in terms of the Hirsch length of
the torsion free admissible groups of whom we take the central product.
Definition 8.5. Let Γ and ∆ be finitely generated groups, and let θ : Z(Γ) → Z(∆) be an isomorphism. We define the central product of Γ and ∆ with respect to θ as Γ ◦θ ∆ = (Γ × ∆)/K where
K = {(z, θ (z)−1 ) | z ∈ Z(Γ)}. We define the central product of the groups {Γi }ℓi=1 with respect to
the automorphism θi : Z(Gi) → Z(Gi+1 ) for 1 ≤ i ≤ ℓ − 1 inductively. Assuming that (Γi ◦θi )ℓi=1
has already been defined, we define (Γi ◦θi )ℓi=1 as the central product of (Γi ◦θi )ℓ−1
i=1 and Γℓ with
ℓ−1
respect to the induced isomorphisms θ̄ℓ−1 : Z((Γi ◦θi )i=1 ) → Z(Γℓ ). When Γ = Γi and θ = θi for
all i, we write the central product as (Γ◦θ )ℓi=1 .
Suppose that Γ ◦θ ∆ is a central product of two nilpotent groups. Since products and quotients of
nilpotent groups are nilpotent, it follows that Γ ◦θ ∆ is a nilpotent group. However, the isomorphism type of Γ ◦θ ∆ is dependent on θ .
Proposition 8.6. Let {Γi }ℓi=1 be a collection of torsion free admissible groups where h(Z(Γi )) = 1
for all i. Let Z(Γi ) = hzi i, and let θi : Z(Γi ) → Z(Γi+1 ) be the isomorphism given by θ (zi ) = zi+1
for 1 ≤ i ≤ ℓ − 1. Then h((Γi ◦θi )ℓi=1 ) = ∑ℓi=1 h(Γi ) − ℓ + 1 and h(Z(Γi ◦θi )ℓi−1 )) = 1
Proof. We may assume that ℓ = 2. First note that if Γ is a torsion free admissible group with
a normal subgroup ∆ E Γ such that Γ/∆ is torsion free, then h(Γ) = h(∆) + h(Γ/∆). Observe
that Γ1 ◦θ Γ2 /Z(Γ1 ◦θ Γ2 ) ∼
= Γ1 /Z(Γ1 ) × Γ2 /Z(Γ2 ). Since h(Z(Γ1 ◦θ Γ2 )) = 1, we may write
h(Γ1 /Z(Γ1 )) + h(Γ2 /Z(Γ2 )) + 1 = h(Γ1 ◦θ Γ2 ). Thus, h(Γ1 ◦θ Γ2 ) = h(Γ1 ) − 1 + h(Γ2 ) − 1 + 1 =
h(Γ1 ) + h(Γ2 ) − 1.
Definition 8.7. For ℓ ≥ 3, we define Λℓ to be the torsion free admissible group generated by the
set Sℓ = {xi }ℓi=1 with relations consisting of commutator brackets of the form [x1 , xi ] = xi+1 for
2 ≤ i ≤ ℓ − 1 and all other commutators being trivial.
Effective Separability of Finitely Generated Nilpotent Groups
31
Λℓ is an example of a Filiform nilpotent group. It has Hirsch length ℓ and has step length ℓ −
1. Defining ∆i = hxs iℓs=m−i+1 , it follows that {∆i }ℓi=1 is a cyclic series for Λℓ and {ξi }ℓi=1 is a
compatible generating subset where ξi = xℓ−i+1 . Additionally, h(Z(Λℓ )) = 1.
Proof. Theorem 1.5(ii), (iii) and (iv)
Assume that ℓ ≥ 3. By construction, Λℓ is a torsion free admissible group of Hirsch length ℓ such
that h(Z(Γℓ )) = 1 . Corollary 1.2 implies that FΛm (n) ≈ (log(n))ℓ which gives Theorem 1.5(ii).
For 2 ≤ c1 < c2 and ℓ ≥ 1, there exist natural numbers jℓ and ιℓ satisfying ( jℓ − 1) (c1 + 1) =
jℓ
ℓ lcm(c1 + 1, c2 + 1) and (ιℓ − 1) (c2 + 1) = ℓ lcm(c1 + 1, c2 + 1), respectively. Let Γℓ = (Λi ◦θΓ )i=1
ιℓ
where θΓ : Z(Λc1 +1 ) → Z(Λc1 +1 ) and θ∆ : Z(Λc2 +1 ) → Z(Λc2 +1 ) are the
and ∆ℓ = (Λi ◦θ∆ )i=1
identity isomorphisms, respectively. Proposition 8.6 implies that h(Γℓ ) = h(∆ℓ ) and h(Z(Γℓ )) =
h(Z(∆ℓ )) = 1, and thus, Corollary 1.2 implies that FΓℓ (n) ≈ F∆ℓ (n).
m with finite generating
Lastly, let c > 1 and m ≥ 1, and consider the group Γc m = (Λc+1 ◦θ )ci=1
2
subset Scm . Proposition 8.6 implies that h(Γcm ) = c m + c m − 1, and since c m2 + c m − 1 ≥ m,
Corollary 1.2 implies that (log(n))m FΓc m (n) as desired.
9
A review of Blackburn and a proof of Theorem 1.6
We start with a review of Blackburn’s proof of conjugacy separability for infinite admissible
groups. This section provides motivation for estimates in the following sections and how one
obtains an upper bound for ConjΓ (n).
Let {Γ, ∆i , ξi } be an admissible 3-tuple such that Γ is torsion free, and let γ , η ∈ Γ such that γ ≁ η .
In order to construct a surjective homomorphism to a finite group that separates the conjugacy
classes of γ and η , we proceed by induction on h(Γ). Since the base case is evident, we may
assume that h(Γ) > 1. When π∆1 (γ ) ≁ π∆1 (η ), induction implies there exists a surjective homomorphism to a finite group ϕ : Γ → Q such that ϕ (γ ) ≁ ϕ (η ). Otherwise, we may assume that
η = γ ξ1t for some t ∈ Z − {0}. The following integer is of particular importance.
h(Γ)
Definition 9.1. Let Γ be a torsion free admissible group with a cyclic series {∆i }i=1 and a comh(Γ)
(CΓ/∆1 (γ̄ )) → ∆1 be given by
patible generating subset {ξi }i=1 . Let γ ∈ Γ. If we let ϕ : π∆−1
1
D
E
τ (γ ) ∼
ϕ (η ) = [γ , η ], we then define τ (Γ, ∆i , ξi , γ ) = τ (γ ) ∈ N such that ξ
= Im(ϕ ).
1
We choose a prime power pα such that pα | τ (Γ, ∆i , ξi , γ ) and pα ∤ t. We then find a w ∈ N such
α
α −w
that if α ≥ w, then for each γ ∈ Γ p there exists η ∈ Γ satisfying η p
= γ (see [2, Lem 2]).
Consider the following definition (see [2, Lem 3]).
Definition 9.2. Let {Γ, ∆i , ξi } be an admissible 3-tuple such that Γ is torsion free, and let γ ∈ Γ.
We define e(Γ, ∆i , ξi , γ ) = e(γ ) ∈ N to be the smallest natural number such that if α ≥ e(γ ), then
α −e(γ )
).
CΓ/Γ pα (γ̄ ) ⊆ σ pα (CΓ (γ ) · Γ p
Letting e′ (γ̄ ) = e(Γ/∆1 , ∆i /∆1 , ξ̄i , γ̄ ), we set ω = α + w + e′ (γ̄ ). Blackburn then proves that
σ pβ (γ ) ≁ σ pβ (η ) (see §12 and [2]). However, as a consequence of the choice of a cyclic series
Effective Separability of Finitely Generated Nilpotent Groups
32
and a compatible generating subset, it becomes evident that the integer w is unnecessary. When
Γ has finite order elements, Blackburn inducts on |T (Γ)|. Thus, it suffices to bound pe(Γ,∆i ,ξi ,γ )
and τ (Γ, ∆i , ξi , γ ) in terms of kγ kS and kη kS . Following Blackburn’s method, we calculate the
asymptotic upper bound for ConjH2m+1 (Z) (n). We then demonstrate that the upper bound is sharp.
Before starting, we make the following observations for H2m+1 (Z). Using the cyclic series and
compatible generating subset given in the second paragraph of Subsection 5.2, we have τ (γ ) =
τ (H2m+1 (Z), ∆i , ξi , γ ) = gcd{xγ ,i , yγ , j |1 ≤ i, j ≤ m}. Thus, τ (γ ) ≤ C0 kγ kS for some C0 ∈ N. Moreover, via Subsection 5.1 we may write the conjugacy class of γ as
1 ~xγ τ (γ ) β + zγ
β ∈Z .
~0 Im
~yγ
(5)
0 ~0
1
Proposition 9.3. ConjH2m+1 (Z) (n) n2m+1 .
Proof. Let γ , η ∈ Γ such that kγ kS , kη kS ≤ n and γ ≁ η . We need to construct a surjective homomorphism ϕ : H2m+1 (Z) → Q to a finite group such that ϕ (γ ) ≁ ϕ (η ) and |Q| ≤ C n2m+1 for some
C ∈ N. We proceed based on whether γ and η have equal images in (H2m+1 (Z))ab . Corollary 1.4
(see also [4, Cor 2.3]) implies that there exists a surjective homomorphism ϕ : Z2m → Q such that
ϕ (πab (γ η −1 )) 6= 1 and |Q| ≤ C1 log(C1 n) for some C1 ∈ N. Since γ and η are non-equal central
elements in Q, it follows that ϕ (πab (γ )) ≁ ϕ (πab (η )), and thus, CDH2m+1 (Z) (γ , η ) ≤ C1 log(C1 n).
Thus, we may assume that πab (γ ) = πab (η ). In particular, we may write η = γ λ t where |t| ≤ C0 n2 .
Let pω be a prime power that divides τ (γ ) but not t. We claim that σ pω (γ ) ≁ σ pω (γ λ t ), and for a
contradiction, suppose otherwise. That implies
there exists x ∈ H2m+1 (Z) such that σ pω ([γ , x]) =
t
ω
σ p (λ ). Equation (5) implies that zη ∈ ℓγ β + zγ : β ∈ Z ( mod pω ) . Therefore, there exist
a, b ∈ Z such that t = a τ (γ ) + b pω . Thus, pω | t, a contradiction. Hence, σ pω (γ ) ≁ σ pω (η ).
When τ (γ ) 6= 0, we have that pω ≤ τ (γ ) ≤ C0 n. Hence, CDH2m+1 (Z) (γ , η ) ≤ C02m+1 n2m+1 . When
τ (γ ) = 0, the Prime Number Theorem [37, 1.2] implies that there exists a prime p such that p ∤ t
where p ≤ C2 log(C2 |t|) for some C2 ∈ N. Hence, p ≤ C3 log(C3 n) for some C3 ∈ N, and thus,
CDH2m+1 (Z) (γ , η ) ≤ C3 (log(C3 n))2m+1 . Hence, ConjH2m+1 (Z) (n) n2m+1 .
The following proposition finishes the proof of Theorem 1.6.
Proposition 9.4. n2m+1 ConjH2m+1 (Z) (n).
Proof. We will construct a sequence of non-conjugate pairs {γi , ηi } such that CDH2m+1 (Z) (γi , ηi ) =
ni2m+1 where kγi k, kηi k ≈ ni . Let {pi } be an enumeration of the primes. Writing pi · e1 as the scalar
product, we consider the following pair of elements:
1 pi ·~e1 1
1 pi ·~e1 2
Im ~0 and ηi = ~0
Im ~0 .
γi = ~0
~0
~0
0
1
0
1
Effective Separability of Finitely Generated Nilpotent Groups
33
Equation (5) implies that we may write the conjugacy class of γi as
1 pi ·~e1 t pi + 1
~0 t ∈ Z .
~0
Im
~0
0
1
(6)
Since σ pi (γi ) and σ pi (ηi ) are non-equal central elements of H2m+1 (Z)/(H2m+1 (Z)) pi , it follows
that γi ≁ ηi for all i. Moreover, we have kγi kS , kηi kS ≈ pi . Given that | H2m+1 (Z)/(H2m+1 (Z)) pi | =
pi2m+1 , we claim CDH2m+1 (Z) (γi , ηi ) = pi2m+1 . In order to demonstrate our claim, we show for all
surjective homomorphisms to a finite group ϕ : H2m+1 (Z) → Q where |Q| < pi2m+1 that ϕ (γi ) ∼
ϕ (ηi ). [16, Thm 2.7] implies that we may assume that |Q| = qµ . Since ϕ (γi ) = ϕ (ηi ) when
ϕ (λ ) = 1, we may assume that ϕ (λ ) 6= 1.
Suppose first that q = pi . We demonstrate that if Q is a group where ϕ (γi ) ≁ ϕ (ηi ), then there
exists no proper quotient of Q such that the images of ϕ (γi ) and ϕ (ηi ) are non-conjugate. Since
B(H2m+1 (Z)) = 1, Proposition 4.10 implies that |Q| = pi2m+1 . Since every choice of an admissible
quotient with respect to any primitive, central, non-trivial element is isomorphic to the trivial
subgroup, Proposition 4.11 implies that there exist no proper quotients of Q such that the image of
ϕ (λ 2 ) is non-trivial Thus, if N is a proper quotient of Q with natural projection ρ : Q → N, then
ker(ρ ) ∩ Z(Q) ∼
= Z(Q) since Z(Q) ∼
= Z/p j Z by Proposition 4.11. Thus, ρ (ϕ (γi )) = ρ (ϕ (ηi ));
hence, ρ (ϕ (γi )) ∼ ρ (ϕ (ηi )). In particular, if Q is a p-group where |Q| < pi2m+1 , then ϕ (γi ) ∼
ϕ (ηi ). Thus, we may assume that q 6= pi .
If q > pi , then Proposition 4.10 implies that p2m+1 > qµ . Thus, we may assume that q < pi .
Since Proposition 4.10 implies that Z/qν Z ∼
= Z(Q), Equation 6 implies that if 1 ≡ p t ( mod qν Z)
for some t ∈ Z , then ϕ (γ p ) ∼ ϕ (η p ). The smallest qν where this fails is qν = pi since p̄i is a
unit in Z/qν Z if and only if gcd(pi , qν ) = 1. Therefore, ϕ (γi ) ∼ ϕ (ηi ) when qµ < pi . Hence,
n2m+1 ConjH2m+1 (Z) (n).
The following corollary will be useful for the proof of Theorem 1.8.
Corollary 9.5. Let H3 (Z) be the 3-dimensional Heisenberg group with the presentation given by
hκ , µ , ν : [µ , ν ] = κ , κ central i, and let p be a prime. Suppose ϕ : H3 (Z) → Q is a surjective
homomorphism such that Q is a q-group where q is a distinct from p and where ϕ (κ ) 6= 1. Then
ϕ (µ p κ ) ∼ ϕ (µ p κ 2 ).
Proof. We may write the conjugacy class of µ p κ as {µ p κ t p+1 | t ∈ Z}. Proposition 4.10 implies
that Z(Q) ∼
= hϕ (κ )i. Hence, Z(Q) ∼
= Z/mZ where m = OrdQ (ϕ (κ )). Since Q is a q-group, it
β
follows that m = q . Given that gcd(p, qβ ) = 1, there exists integers r, s such that r p+sqβ = 1. We
β
have that µ p κ r p+1 ∼ µ p κ . We may write ϕ (µ p κ r p+1 ) = ϕ (µ p κ 1−s q +1 ) = ϕ (µ p κ 2 ). Therefore,
ϕ (µ p κ ) ∼ ϕ (µ p κ 2 ) as desired.
Effective Separability of Finitely Generated Nilpotent Groups
34
10 Relating complexity in groups and Lie algebras
Let Γ be a torsion free admissible group with finite generating subset S, Mal’tsev completion G,
and Lie algebra g of G. The overall goal of this section is to provide a bound of k Log(γ )kLog(S) in
terms of kγ kS where Log(S) gives a norm for the additive structure of g.
Proposition 10.1. Let {Γ, ∆i , ξi , G, g, νi } be an admissible 6-tuple, and suppose that Γ has step
2
length c. Let γ ∈ Γ. Then there exists a C ∈ N such that k Log(γ )kX ≤ C (kγ kS )c .
Proof. Using the Mal’tsev coordinates of γ , we may write γ = ∏i=1 ξiαi . Lemma 2.9 implies
that there exists a C1 ∈ N such that |αi | ≤ C1 (kγ kS )c for all i. A straightforward application of
the Baker-Campbell-Hausdorff formula (2) implies that Log(ξiαi ) = αi νi . Writing Ai = αi νi , it
follows that kAi kX ≤ C1 (kγ kS )c . Equation (2) implies that we may write
h(Γ)
c
k Log(γ )kX ≤ ∑ kCBi (A1 , · · · , Ah(Γ) )kX
i=1
where CBi (A1 , · · · , Ah(Γ) ) is a rational linear combination of i-fold Lie brackets in {A js }ts=1 ⊆
h(Γ)
h(Γ)
{Ai }i=1 . Let {A js }ts=1 ⊂ {Ai }i=1 where [A j1 , · · · A jt ] 6= 0. Via induction on the length of the
iterated Lie bracket, one can see that there exists Ct ∈ N such that [A j1 , · · · , A jt ] ≤ Ct ∏ts=1 kA js kX ≤
h(Γ)
Ct C1 (kγ kS )t c . By maximizing over all possible t-fold Lie brackets of elements of {Ai }i=1 , there
2
exists a Di ∈ N such that kCBi (A1 , · · · , Ah(Γ) )kX ≤ Di (kγ kS )t c . Hence, k Log(γ )kX ≤ C (kγ kS )c
for some C ∈ N.
An immediate application of Proposition 10.1 is that the adjoint representation of Γ has matrix
coefficients bounded by a polynomial in terms of word length.
Proposition 10.2. Let {Γ, Λ, ∆i , ξi , G, g, νi } be an admissible 7-tuple. Let γ ∈ Γ, and let (µi, j ) be
the matrix representative of Ad(γ ) with respect to X . Then |µi, j | ≤ C (kγ kS )c for some C ∈ N where
c is the step length of Γ.
2
Proof. Proposition 10.1 implies there exists a C1 ∈ N such that k Log(γ )kX ≤ C1 (kγ kS )c . Via
3
induction on the length of the Lie bracket and Equation (3), we have k Ad(γ )(vi )kX ≤ C2 (kγ kS )c
for some C2 ∈ N.
11 Preliminary estimates for Theorem 1.7
Let {Γ, ∆i , ξi } be an admissible 3-tuple such that Γ is torsion free. Let γ be a non-trivial element
of Γ, and let p be some prime. In the following section, we demonstrate the construction of the
integer e(γ ) = e(Γ, ∆i , ξi , γ ) and give an asymptotic bound for pe(γ ) in terms of kγ kS independent
of the prime p. We first provide a bound for τ (Γ, ∆i , ξi , γ ) in terms of kγ kS .
Effective Separability of Finitely Generated Nilpotent Groups
35
Proposition 11.1. Let {Γ, Λ, ∆i , ξi , G, g, νi } be an admissible 7-tuple, and let γ ∈ Γ. There exists
k,C ∈ N such that |τ (Γ, ∆i , ξi , γ )| ≤ C (kγ kS )k .
Proof. Before we start, we make some simplifying notation by letting τ (γ ) = τ (Γ, ∆i , ξi , γ ). Conτ (Γ,γ )
sider the smooth map Φ : G → G given by Φ(g) = [γ , g]. Suppose η ∈ Γ satisfies Φ(η ) = ξ1
.
−1
The commutative diagram (1.2) on [12, Pg 7] implies that we may write (I − Ad(γ ))(Log(η )) =
τ (Γ,γ )
Log(ξ1
) where (dΦγ )1 = I − Ad(γ −1 ). Proposition 10.2 implies that I − Ad(γ −1 ) is a strictly
3
upper triangular matrix whose coefficients are bounded by C (kγ kS )(c(Γ)) for some C ∈ N. Since
τ (γ )
it is evident that Log(ξ1 ) = τ (γ )ν1 , backwards substitution gives our result.
The first statement of the following proposition is originally found in [2, Lem 3]. We reproduce
its proof so that we may provide estimates for the value e(Γ, ∆i , ξi , γ ) in terms of kγ kS .
Proposition 11.2. Let {Γ, Λ, ∆i , ξi , G, g, νi } be an admissible 7-tuple. Let p be prime and γ ∈ Γ.
Then there exists e(Γ, ∆i , ξi , γ ) = e(γ ) ∈ N such that if α ≥ e(γ ), then CΓ/Γ pα (γ̄ ) ⊆ σ pα (CΓ (γ ) ·
α −e(γ )
Γp
). Moreover, pe(γ ) ≤ C(kγ kS )k for some C ∈ N and k ∈ N.
Proof. We proceed by induction on Hirsch length, and given that the statement is clear for Z by
setting e(γ ) = 0 for all γ , we may assume that h(Γ) > 1.
We construct e(γ ) based on the value of τ (γ ) = τ (Γ, ∆i , ξi , γ ) (see Definition 9.1). By induction,
we may assume that we have already constructed e′ (γ̄ ) = e(Γ/∆1 , ∆i /∆1 , ξ̄i , γ¯). When τ (γ ) = 0,
we set e(γ ) = e′ (γ̄ ). Suppose α ≥ e(γ ) and that η̄ ∈ CΓ/Γ pα (γ̄ ) for some η ∈ Γ. By selection,
α −e(γ )
(CΓ/∆1 (γ̄ )) · Γ p
η̄ ∈ CΓ/Γ pα ·∆1 (γ̄ ). Thus, we may write η ∈ π∆−1
1
it follows that η̄ ∈ σ pα (CΓ (γ ) · Γ
pα −e(γ )
. Since π∆−1
(CΓ/∆1 (γ̄ )) = CΓ (γ ),
1
α −e(γ )
). Thus, CΓ/Γ pα (γ̄ ) ⊆ σ pα (CΓ (γ ) · Γ p
).
When τ (γ ) 6= 0, we let β be the largest power of p such that pβ | τ (γ ), and set e(Γ) = e′ (γ̄ )+ β . Let
α ≥ e(γ ), and let η ∈ Γ satisfy η̄ ∈ CΓ/Γ pα (γ̄ ). Thus, η̄ ∈ CΓ/Γ pα ·∆1 (γ̄ ), and subsequently, induction
α −e(γ )+β
implies η̄ ∈ πΓ pα ·∆1 (CΓ/∆1 (γ̄ ) · Γ p
α −e(γ )+β
τ (γ )
). Thus, we may write η = µ ε a λ where µ ∈ CΓ (γ ), λ ∈
α −e(γ )+β
α −e′ (γ̄ )
Γp
, and ϕγ (ε ) = ξ1 . Hence, we have [γ , η ] = [γ , ε a ] ∈ Γ p
. Since [γ , ε a ] ∈ Γ p
a
α
−e(
γ
)+
β
β
and [γ , ε ] ∈ ∆1 , we have that p
| a τ (γ ). By definition of p , it follows that pα −e(γ ) | a,
α
−e(
γ
)
α −e(γ )
). Hence, CΓ/Γ pα (γ̄ ) ⊆ σ pα (CΓ (γ ) · Γ p
).
and thus, η̄ ∈ σ pα (CΓ (γ ) · Γ p
We proceed by induction on Hirsch length to demonstrate the asymptotic upper bound, and since
the base case is clear, we assume that h(Γ) > 1. Let γ ∈ Γ, and suppose that τ (γ ) = 0. By
′
construction, e(γ ) = e′ (γ̄ ), and thus, induction implies that there exist C1 , k1 ∈ N such that pe (γ̄ ) ≤
C1 (kγ̄ kS̄ )k1 . When τ (γ ) 6= 0, it follows that e(γ ) = e′ (γ̄ ) + β where β is the largest power of
p that divides τ (γ ). Proposition 11.1 implies there exist k2 ,C2 ∈ N such that pβ ≤ C2 (kγ kS )k2 .
Consequently, pe(γ ) ≤ C1 C2 (kγ kS )k1 +k2 .
Effective Separability of Finitely Generated Nilpotent Groups
36
12 Proof of Theorem 1.7
Let Γ be an infinite admissible group. In order to demonstrate that there exists k1 ∈ N such that
ConjΓ (n) nk1 , we need to show for any γ , η ∈ Γ where γ ≁ η and kγ kS , kη kS ≤ n there exists
a prime power pω ≤ C nk2 such that σ pω (γ ) ≁ σ pω (η ) for some C, k2 ∈ N. It then follows that
CDΓ (γ , η ) ≤ Ch(Γ) nh(Γ) k2 . We first specialize to torsion free admissible groups.
Proposition 12.1. Let {Γ, Λ, ∆i , ξi , g, νi } be an admissible 7-tuple. Then there exists k ∈ N such
that ConjΓ (n) nk .
Proof. Let γ , η ∈ Γ such that kγ kS , kη kS ≤ n and where γ ≁ η . We demonstrate that there exists a
k0 ∈ N such that CDΓ (γ , η ) ≤ C0 nk0 for some C0 ∈ N by induction on h(Γ), and since the base case
is clear, we may assume that h(Γ) > 1. If π∆1 (γ ) ≁ π∆1 (η ), then the inductive hypothesis implies
that there exists a surjective homomorphism to a finite group ϕ : Γ/∆1 → Q such that ϕ (γ ) ≁ ϕ (η )
and where |Q| ≤ C1 nk1 for some C1 , k1 ∈ N. Thus, CDΓ (γ , η ) ≤ C1 nk1 . Otherwise, we may assume
that η = γ ξ1t , and Lemma 2.9 implies that |t| ≤ C2 nc(Γ) for a constant C2 ∈ N.
For notational simplicity, let τ (γ ) = τ (Γ, ∆i , ξi , γ ) and e′ (γ ) = e(Γ, ∆i /∆1 , ξ̄i , γ¯). Since γ ≁ γ ξ1t ,
there exists a prime power pα such that pα | τ (γ ) but pα ∤ t. We set ω = α +e′ (γ̄ ), and suppose for a
contradiction there exists x ∈ Γ such that σ pω (x−1 γ x) = σ pω (γ ξ1 )t . That implies x̄ ∈ CΓ/Γ pω ·∆1 (γ̄ ),
α
and thus, x̄ ∈ πΓ pω ·∆1 (CΓ/∆1 (γ ) · Γ p ) by Proposition 11.2. Subsequently, x = g µ for some g ∈
q τ (Γ,γ )
α
(CΓ/∆1 (γ̄ )) and µ ∈ Γ p . Hence, σ pω ([γ , g]) = σ pω (ξ1 )t , and since [γ , g] = ξ1
π∆−1
1
q ∈ Z, it follows that
σ pω (γ ) ≁ σ pω (η ).
t−q τ (Γ,γ )
ξ1
α +e(Γ/∆1 ,γ̄ )
∈ Γp
.
That implies
pα
for some
| t, which is a contradiction. Hence,
′
Proposition 11.2 implies that pe (γ̄ ) ≤ C3 nk2 for C3 , k2 ∈ N. When τ (γ ) = 0, the Prime Number
Theorem [37, 1.2] implies that we may choose p such that |p| ≤ C4 log(C4 n) for some C4 ∈ N.
Hence, CDΓ (γ , η ) ≤ C5 (log(C5 n))h(Γ) k2 for some C5 ∈ N. When τ (γ ) 6= 0, Proposition 11.1
implies that τ (γ ) ≤ C6 nk3 for some C6 , k3 ∈ N. Thus, pω ≤ C3 C6 nk2 +k3 . Therefore, CDΓ (γ , η ) ≤
(C3 C6 )h(Γ) nh(Γ)(k2 +k3 ) . Hence, ConjΓ (n) nk3 where k3 = max{k1 , h(Γ)(k2 + k3 )}.
Proof. Theorem 1.7
Let Γ be an infinite admissible group Γ with a choice of a cyclic series {∆i }m
i=1 and a compatible
m
generating subset {ξi }i=1 . Let k1 be the natural number from Proposition 12.1 and k2 be the
natural number from Proposition 11.2, both for Γ/T (Γ). Letting k3 = h(Γ) · max{k1 , k2 }, we
claim that ConjΓ (n) nk3 . Let γ , η ∈ Γ satisfy γ ≁ η and kγ kS , kη kS ≤ n. In order to show that
CDΓ (γ , η ) ≤ C0 nk3 where C0 ∈ N, we construct a surjective homomorphism to a finite group that
distinguishes the conjugacy classes of γ and η via induction on |T (Γ)|. To simplify the following
arguments, we let e(γ̄ ) = e(Γ/T (Γ), ∆i /T (Γ), ξ̄i , γ̄ ).
Proposition 12.1 implies that we may assume that there exists a subgroup P ⊆ Z(Γ) of prime order
p. If πP (γ ) ≁ πP (η ), then induction implies that there exists a surjective homomorphism to a
finite group ϕ : Γ/P → N such that ϕ (γ ) ≁ ϕ (η ) and where |N| ≤ C1 nk3 for some C1 ∈ N. Thus,
CDΓ (γ , η ) ≤ C1 nk3 . Otherwise, we may assume that η = γ µ where hµ i = P.
Effective Separability of Finitely Generated Nilpotent Groups
37
Suppose there exists Q ⊆ Z(Γ) such that |Q| = q where q is a prime distinct from p. Suppose for a
contradiction that there exists x ∈ Γ such that x−1 γ x = γ µ λ where Q = hλ i. Since [γ , x] ∈ Z(Γ) and
OrdΓ (λ ) = q, basic commutator properties imply that [γ , xq ] = µ q . Given that ps+qr = 1 for some
r, s ∈ Z, it follows that [γ , xq r ] = γ µ 1−p s = γ µ which is a contradiction. Hence, induction implies
there exists a surjective homomorphism to a finite group θ : Γ/Q → M such that θ (γ ) ≁ θ (γ µ )
and where |M| ≤ C2 nk3 for some C2 ∈ N. Thus, CDΓ (γ , η ) ≤ C2 nk3 .
Suppose that T (Γ) is a p-group with exponent pm . We set ω = m + e′ (γ̄ ), and suppose for a contradiction there exists x ∈ Γ such that σ pω (x−1 γ x) ∼ σ pω (γ µ ). Thus, x̄ ∈ CΓ/T (Γ)·Γ pω (γ̄ ), and subm
sequently, Proposition 11.2 implies that x̄ ∈ πT (Γ)·Γ pω (CΓ/T (Γ) (γ̄ ) · Γ p ). Therefore, we may write
m
−1 ∈ Γ pm . Moreover,
x = g λ where λ ∈ Γ p and g ∈ πT−1
(Γ) (CΓ/T (Γ) (γ̄ )). Subsequently, [γ , g] µ
m
since [γ , g] ∈ T (Γ) and T (Γ) ∩ Γ p = 1, it follows that [γ , x] = µ which is a contradiction. Propo′
h(Γ)
sition 11.2 imply that pe (γ̄ ) ≤ C3 nk2 for some C3 ∈ N. Thus, CDΓ (γ , η ) ≤ C3 |T (Γ)| nh(Γ) k2 , and
k
subsequently, ConjΓ (n) n 3 .
13 Proofs of Theorem 1.8 and Theorem 1.9
Let Γ be an infinite admissible group with a choice of a cyclic series {∆i }m
i=1 and a compatible
m
generating subset {ξi }i=1 . Since the proofs of Theorem 1.8(i) and Theorem 1.8(ii) require different strategies, we approach them separately. We start with Theorem 1.8(i) since it only requires
elementary methods.
We assume that Γ contains an infinite, finitely generated abelian group K of index ℓ. We want to
demonstrate that log(n) ConjΓ (n) (log(n))ℓ . Since FΓ,S (n) ConjΓ (n), Corollary 1.4 (see also
[4, Cor 2.3]) implies that log(n) ConjΓ (n). Thus, we need only to demonstrate that ConjΓ (n)
(log(n))ℓ . For any two non-conjugate elements γ , η ∈ Γ such that kγ kS , kη kS ≤ n we want to
construct a surjective homomorphism ϕ : Γ → Q such that ϕ (γ ) ≁ ϕ (η ) and |Q| ≤ C (log(C n))ℓ
for some C ∈ N.
Proof. Theorem 1.8(i)
Let S1 be a finite generating subset for K, and let {υi }ℓi=1 be a set of coset representatives of
K in Γ. We take S = S1 ∪ {υi }ℓi=1 as the generating subset for Γ. If kγ kS ≤ n, we may write
γ = gγ υγ where kgγ kS1 ≤ C1 n for some C1 ∈ N and υγ ∈ {υi }ℓi=1 . Conjugation in Γ induces a
map ϕ : Γ/K → Aut(K) given by ϕ (πK (υi )) = ϕi . Thus, we may write the conjugacy class of γ as
ℓ
ϕi (gγ ) (υi−1 υγ υi ) i=1 . Finally, there exists C2 ∈ N such that if kγ kS1 ≤ n, then kϕi (γ )kS ≤ C2 n
for all i.
Suppose γ , η ∈ Γ are two non-conjugate elements such that kγ kS , kη kS ≤ n. If πK (γ ) ≁ πK (η ),
then by taking the map πK : Γ → Γ/K, it follows that CDΓ (γ , η ) ≤ ℓ. Otherwise, we may assume
that η = gη υγ . By Corollary 1.4 (see also [4, Cor 2.3]), there exists a surjective homomorphism
−1
−1
fi : Γ → Qi such that fi (g−1
γ υγ ϕi (gη ) (υi υη υi )) 6= 1 and |Qi | ≤ C3 log(2C2 C3 n) for some C3 ∈
N. By letting H = ∩ℓi=1 ker( fi ), it follows that πH (γ ) ≁ πH (η ) and |Γ/H| ≤ C3ℓ (log(2C2 C3 n))ℓ .
Therefore, ConjΓ (n) (log(n))ℓ , and subsequently, log(n) ConjΓ (n) (log(n))ℓ .
Effective Separability of Finitely Generated Nilpotent Groups
38
For Theorem 1.8(ii), suppose that Γ does not contain an abelian group of finite index. In order
to demonstrate that nψRF (Γ)(c(Γ/T (Γ))−1) ConjΓ (n), we desire a sequence of non-conjugate pairs
ψ (Γ)(c(Γ/T (Γ))−1)
where kγi kS , kηi kS ≈ ni . In particular, we must
{γi , ηi } such that CDΓ (γi , ηi ) = ni RF
find non-conjugate elements whose conjugacy classes are difficult to separate i.e. non-cojugate
elements that have relatively short word length in comparison to the order of the minimal finite
group required to separate their conjugacy classes.
We first reduce to the calculation of the lower asymptotic bounds for ConjΓ (n) to torsion free
admissible groups by appealing to the conjugacy separability of two elements within a finite index
subgroup.
Proposition 13.1. Let Γ be an infinite admissible group, and let ∆ be a subgroup. Suppose there
exist γ , η ∈ ∆ such that γ ≁ η in Γ. Then CD∆ (γ , η ) ≤ CDΓ (γ , η ).
Proof. We first remark that since Γ and ∆ are admissible groups, Theorem 1.7 implies CDΓ (γ , η ) <
∞ and CD∆ (γ , η ) < ∞. Suppose that ϕ : Γ → Q is surjective homomorphism to a finite group such
that |Q| = CDΓ (γ , η ). If we let ι : ∆ → Γ be the inclusion, then we have a surjective map ϕ ◦ ι :
∆ → ϕ (∆) to a finite group where ϕ (ι (γ )) ≁ ϕ (ι (η )). By definition, CD∆ (γ , η ) ≤ |ϕ (∆)| ≤ |Q|.
Thus, CDΓk (γ , η ) ≤ CDΓ (γ , η ).
Proof. Theorem 1.8(ii)
We first assume that Γ is torsion free. Let Γ/Λ be a choice of a maximal admissible quotient of Γ.
There exists a g ∈ Z(Γ) − {1} such that Γ/Λ is a choice of an admissible quotient with respect to
g. Moreover, there exists a k such that gk = [y, z] for some y ∈ Γc−1 and z ∈ Γ. If g is not primitive,
then there exists a xΛ ∈ Z(Γ) − {1} such that xsΛ = g for some s ∈ Z − {0}. In particular, xtΛ = [y, z]
where t = s k.
t B(Γ/Λ)
. Equation 4 implies that HΛ =
There exists aΛ ∈ Γc(Γ)−1 and bΛ ∈ Γ such that [aΛ , bΛ ] = xΛ
E
D
t B(Γ/Λ) ∼
aΛ , bΛ , xΛ
= H3 (Z). Let p j,Λ be an enumeration of primes greater than B(Γ/Λ), and
p
t B(Γ/Λ)
p
2 t B(Γ/Λ)
and η j,Λ = aΛj,Λ xΛ
. Since the images of γ j,Λ and η j,Λ are non-equal
let γ j,Λ = aΛj,Λ xΛ
p
j,Λ
central elements of Γ/Λ · Γ , it follows that γ j,Λ ≁ η j,Λ for all j.
We claim that γ j,Λ and η j,Λ are our desired non-conjugate elements. In particular, we will demonψ (Γ)(c(Γ)−1)
and that kγ j,Λ kS , kη j,Λ kS ≈ ni . By construction, we have
strate CDΓ (γ j,Λ , η j,Λ ) ≈ ni RF
that γ j,Λ , η j,Λ ∈ Γc(Γ)−1 and kγ j,Λ kS′ , kη j,Λ kS′ ≈ p j,Λ where S′ = S ∩ Γ2 . [15, 3.B2] implies that
1/(c(Γ)−1)
ψ (Γ)
RF
kγ j,Λ kS , kη j,Λ kS ≈ p j,Λ
. Hence,
. Therefore, (kγ j,Λ kS )ψLower (Γ) , (kη j,Λ kS )ψLower (Γ) ≈ p j,Λ
we need to demonstrate for all surjective homomorphisms to finite groups ϕ : Γ → Q where
ψRF (Γ)
|Q| < p j,Λ
that ϕ (γ j,Λ ) ∼ ϕ (η j,Λ ).
[16, Thm 2.7] implies that we may assume that |Q| = qβ where q is prime. Since ϕ (γ j,Λ ) =
t B(Γ/Λ)
t B(Γ/Λ)
) = 1, we may also assume that ϕ (xΛ
) 6= 1. Suppose that q >
ϕ (η j,Λ ) when ϕ (xΛ
p j,Λ . Consider the homomorphism given by ρ ◦ ϕ : Γ → Q/ϕ (Λ) where ρ : Q → Q/ϕ (Λ) is the
◦ ϕ : Γ/Λ →
natural projection. Since Λ ≤ ker(ρ ◦ ϕ ), we have an induced homomorphism ρ]
Q/ϕ (Λ). Since h(Z(Γ/Λ)) = 1, ρ]
◦ ϕ (xt B(Γ/Λ) ) 6= 1, and q > B(Γ/Λ), Proposition 4.10 implies
ψRF (Γ)
ψRF (Γ)
. Hence, |Q| > p j,Λ
. Thus, we may assume that q ≤ p j,Λ
that |Q/ϕ (Λ)| > p j,Λ
Effective Separability of Finitely Generated Nilpotent Groups
39
Now assume that q = p j,Λ . Suppose that ϕ (Λ) is a non-trivial subgroup of Q. As before, we have
ψRF (Γ)
◦ ϕ : Γ/Λ → Q/ϕ (Λ). Since |Q/ϕ (Q)| ≤ p j,Λ
an induced homomorphism ρ]
, Proposition 4.11
ψ (Γ)
RF
implies that |Q/ϕ (Q)| = p j,Λ
ϕ (Λ) = {1}.
ψ (Γ)
. Thus, we have that |Q| > p j RF
. Hence, we may assume that
Suppose that q = p j,Λ . If ϕ (γ j,Λ ) ∼ ϕ (η j,Λ ), then there is nothing to prove. Thus, we may assume
that ϕ (γ j,Λ ) ≁ ϕ (η j,Λ ). Proposition 4.11 implies that if N is a proper quotient of Q with natural
projection θ : Q → N, then ker(θ ) ∩ Z(Q) ∼
= Z(Q). Thus, we have that θ (ϕ (γ j,Λ )) = θ (ϕ (η j,Λ ))
ψRF (Γ)
since θ (ϕ (xΛ )) = 1. In particular, if Q is a p j,Λ -group where ϕ (Λ) ∼
, then
= {1} and |Q| < p j,Λ
ϕ (γ j,Λ ) ∼ ϕ (η j,Λ ). Hence, we may assume that q 6= p j,Λ .
Since ϕ (HΛ ) is a q-group where q 6= p j,Λ , Corollary 9.5 implies that there exists g ∈ HΛ such that
ϕ (g−1 γ j,Λ g) = ϕ (η j,Λ ) as elements of ϕ (HΛ ). Thus, ϕ (γ j,Λ ) ∼ ϕ (η j,Λ ). Since we have exhausted
ψRF (Γ)
all possibilities, it follows that CDΓ (γ j , η j ) = p j,Λ
. Hence, nψRF (Γ)(c(Γ)−1) ConjΓ,S (n).
Now suppose that Γ is an infinite admissible group where |T (Γ)| > 1. There exists a finite index,
torsion free subgroup of Γ which we denote as ∆. Let ∆/Λ be a choice of a maximal admissible
quotient of ∆. Using above reasoning, there exists xΛ ∈ ∆ such that ∆/Λ is a choice of an admissible
quotient with respect to xΛ where xtΛ = [y, z] for some y ∈ ∆c−1 and z ∈ ∆.
Let {p j,Λ } be an enumeration of primes greater than B(∆/Λ). There exist an aΛ ∈ ∆c(Γ)−1
t B(∆/Λ)
p
t B(∆/Λ)
p
2 t B(∆/Λ)
and η j,Λ = aΛj xΛ
be
. Let γ j,Λ = aΛj xΛ
and bΛ ∈ ∆ such that [aΛ , bΛ ] = xΛ
p
the elements from the above construction for ∆. Let ρ : Γ → Γ/T (Γ) · Γ be the natural projection. We have that ρ (γ j ) 6= ρ (η j ) and ρ (γ j ), ρ (η j ) 6= 1 by construction. Additionally, we
have that πT (Γ) (∆) is a finite index subgroup of Γ/T (Γ). Thus, [16, Lem 4.8(c)] implies that
Z(πT (Γ) (∆)) = πT (Γ) (∆) ∩ Z(Γ/Z(Γ)). Hence, πT (Γ) (xΛ ) ∈ Z(Γ/T (Γ)). Since ρ (γ j ) and ρ (η j ) are
non-equal central elements of Γ/T (Γ) · Γ p , we have that γ j ≁ η j .
Proposition 13.1 implies that CD∆ (γ j,Λ , η j,Λ ) ≤ CDΓ (γ j,Λ , η j,Λ ). By the above construction, we
ψ (∆)(c(∆)−1)
have n j RF
≤ CDΓ (γ j,Λ , η j,Λ ) where kγ j,Λ kS , kη j,Λ kS ≈ n j . If S′ is a finite generating
subset of Γ, then kγ j,Λ kS ≈ kγ j,Λ kS′ and kη j,Λ kS ≈ kη j,Λ kS′ . Hence, kγ j,Λ kS′ , kη j,Λ kS′ ≈ n j ,
ψ (∆)(c(∆)−1)
and n j RF
CDΓ (γ j,Λ , η j,Λ ). Since the projection to the torsion free quotient πT (Γ) :
Γ → Γ/T (Γ) is injective when restricted to ∆, ∆ is isomorphic to a finite index subgroup of
Γ/T (Γ), and thus, Theorem 1.3 implies that ψRF (Γℓ ) = ψRF (Γ). Since c(Γ/T (Γ)) = c(∆), we
have nψRF (Γ)(c(Γ/T (Γ))−1) ConjΓ (n).
Proof. Theorem 1.9
Suppose that Γ and ∆ are two infinite admissible of step size 2 or greater such that Γ/T (Γ)
and ∆/T (∆) has isomorphic Mal’tsev completions. Proposition 7.8 implies that ψRF (Γ/T (Γ)) =
ψRF (∆/T (∆)). By definition of ψRF (Γ) and ψRF (∆), we have ψRF (Γ) = ψRF (∆). Since c(Γ/T (Γ)) =
c(∆/T (∆)), our theorem is now evident.
Effective Separability of Finitely Generated Nilpotent Groups
40
14 Proof of Theorem 1.10
Proof. For s ∈ N, let Λs be the group given in Definition 8.7, and let θ : Z(Λs ) → Z(Λs ) be
the identity morphism. Let c > 1 and m ≥ 1, and consider the group Γcm = (Λc+1 ◦θ )m
i=1 with
a finite generating subset Scm . Proposition 8.6 implies that h(Γcm ) = c m2 + c m − 1, and since
c2 m2 + c2 m − 1 ≥ m, Theorem 1.8(ii) implies that nm ConjΓcm (n) as desired.
References
[1] H. Bass, The degree of polynomial growth of finitely generated nilpotent groups, Proc. London Math. Soc. 25
(1972), 603–614.
[2] N. Blackburn, Conjugacy in nilpotent groups, Proc. Amer. Math. Soc. 16 (1965), 143–148.
[3] K. Bou-Rabee, Approximating a group by its solvable quotients, New York J. Math. 17 (2011), 699–712.
[4] K. Bou-Rabee, Quantifying residual finiteness, J. Algebra 323 (2010), 729–737.
[5] K. Bou-Rabee, T. Kaletha, Quantifying residual finiteness of arithmetic groups, Compos. Math. 148 (2012),
907–920.
[6] K. Bou-Rabee, D. B. McReynolds, Asymptotic growth and least common multiples in groups, Bull. Lond.
Math. Soc. 43 (2011), 1059–1068.
[7] K. Bou-Rabee, D. B. McReynolds, Extremal behavior of divisibility functions, Geom. Dedicata 2 (2014), 1–9
[8] K. Bou-Rabee, M. Hagen, P. Patel, Residual finiteness growths of virtually special groups, to appear in Math.
Zeit., ArXiv (2014).
[9] N. V. Buskin, Efficient separability in free groups, Sib. Math. J. 50 (2009), 603–608.
[10] J. S. Chahal, Solution of the congruence subgroup problem for solvable algebraic groups, Nagoya Math. J.
79 (1980), 141-144
[11] L. Corwin, F. Greenleaf, Representations of nilpotent Lie groups and their applications. Part 1. Basic theory
and Examples, Cambridge Studies in Advanced Mathematics, 18. Cambridge University Press, Cambridge.
1990, viii+269 pp.
[12] K. Dekimpe, Almost-Beiberbach groups: Affine and polynomial structures, Springer–Verlag, 1996.
[13] E. Formanek, Conjugate separability in polycyclic groups, J. Algebra 42 (1976), 1–10.
[14] V. V. Gorbacevic̆, Discrete subgroups of solvable Lie groups of type (E). Math. USSR Sbornik, 1971, 14 N◦ ,
pp. 233-251.
[15] M. Gromov, Asymptotic invariants of infinite groups, Geometric group theory, Vol. 2 (Sussex, 1991), 1-295,
London Math. Soc. Lecture Note Ser., 182, Cambridge Univ. Press, 1993.
[16] P. Hall, The Edmonton notes on nilpotent groups, Queen Mary College Math. Notes. Math. Dept., Queen
Mary College, London 1969.
[17] E. Hamilton, H. Wilton, and P. A. Zalesskii, Separability of double cosets and conjugacy classes in 3-manifold
groups, J. Lond. Math. Soc. (2) 87 (2013), no. 1, 269–288
Effective Separability of Finitely Generated Nilpotent Groups
41
[18] J. Hempel, Residual finiteness for 3-manifolds, Combinatorial group theory and topology (Alta, Utah, 1984),
379–396, Ann. of Math. Study., 111, Princeton Univ. Press, Princeton, NJ, 1987
[19] D. F. Holt, B. Eick, E. A. O’Brien, Handbook of computational group theory, Discrete Mathematics and its
Applications (Boca Ralton). Chapman & Hall/CRC, Boca Raton, FL, 2005., xvi+514 pp.
[20] M. Kassabov, F. Matucci, Bounding the residual finiteness of free groups, Proc. Amer. Math. Soc. 139 (2011),
2281–2286.
[21] E. Khukhro, Nilpotent Groups and their automorphisms, de Gruyter Expositions in Mathematics, 8. Walter
de Gruyter & Co., Berlin, 1993.
[22] A. W. Knapp, Lie groups beyond an introduction, Birkhäuser, 2002.
[23] G. Kozma, A. Thom, Divisibility and laws in finite simple groups, ArXiv (2014).
[24] S. Lawton, L. Louder, D. B. McReynolds, Decision problems, complexity, traces, and representations, ArXiv
(2013).
[25] C. R. Leedham-Green, S. McKay, The structure of groups of prime power order, London Mathematical Society Monographs. New Series, 27. Oxford Science Publications, Oxford University Press, Oxford, 2002.
xii+334 pp.
[26] R. C. Lyndon, P. E. Schupp, Combinatorial group theory, Springer–Verlag, 2001.
[27] A. I. Mal’tsev. On a class of homogeneous spaces, Amer. Math. Soc. Translation (1951), 33 pp.
[28] A. I. Mal’tsev, On homomorphisms onto finite groups, Uchen. Zap. Ivanovskogo Gos. Ped. Inst. 18 (1958),
49–60.
[29] A. I. Mal’tsev, On isomorphic matrix representations of infinite groups, Amer. Math. Soc. Transl. (2) 45
(1965), 1–18.
[30] D. V. Osin, Subgroup distortions in nilpotent groups, Comm. Algebra 29 (2001), 5439–5463.
[31] P. Patel, On a theorem of Peter Scott, Proc. Amer. Math. Soc. 142 (2014), 2891–2906.
[32] M. S. Raghunathan, Discrete subgroups of Lie groups, Math. Student 2007, Centenary Volume, 59–70 (2008).
[33] I. Rivin, Geodesics with one self-intersection, and other stories, Adv. Math. 231 (2012), 2391–2412.
[34] V. N. Remeslennikov, Finite approximability of groups with respect to conjugacy, Siberian Math. J. 23 (1971)
783–792.
[35] D. Segal, Polycyclic Groups, Cambridge University Press, 1983.
[36] P. Stebe, Conjugacy separability of groups of integer matrices, Proc. Amer. Math. Soc. 32 (1972), 1–7.
[37] G. Tenenbaum, Introduction to analytic and probabilistic number theory, Cambridge University Press, 1995
[38] A. Thom, About the length of laws for finite groups, ArXiv (2015)
| 4math.GR
|
1
Pilot Reuse Among D2D Users in D2D Underlaid
Massive MIMO Systems
arXiv:1708.01358v1 [cs.IT] 4 Aug 2017
Hao Xu, Student Member, IEEE, Wei Xu, Senior Member, IEEE, Zhaohui Yang, Student Member, IEEE, Jianfeng
Shi, Student Member, IEEE, and Ming Chen, Member, IEEE
Abstract—In a device-to-device (D2D) underlaid massive
MIMO system, D2D transmitters reuse the uplink spectrum
of cellular users (CUs), leading to cochannel interference. To
decrease pilot overhead, we assume pilot reuse (PR) among
D2D pairs. We first derive the minimum-mean-square-error
(MMSE) estimation of all channels and give a lower bound
on the ergodic achievable rate of both cellular and D2D links.
To mitigate pilot contamination caused by PR, we then propose
a pilot scheduling and pilot power control algorithm based on
the criterion of minimizing the sum mean-square-error (MSE)
of channel estimation of D2D links. We show that, with an
appropriate PR ratio and a well designed pilot scheduling scheme,
each D2D transmitter could transmit its pilot with maximum
power. In addition, we also maximize the sum rate of all D2D
links while guaranteeing the quality of service (QoS) of CUs, and
develop an iterative algorithm to obtain a suboptimal solution.
Simulation results show that the effect of pilot contamination can
be greatly decreased by the proposed pilot scheduling algorithm,
and the PR scheme provides significant performance gains over
the conventional orthogonal training scheme in terms of system
spectral efficiency.
I. I NTRODUCTION
With the increasing demand on broadband wireless communications, the problem of spectrum insufficiency has become
a major factor limiting the wireless system performance [1].
Massive multiple-input multiple-output (MIMO) transmission
was proposed in [2] and has triggered considerable research
interest recently due to its great gains in spectral efficiency
(SE) and energy efficiency (EE) [3]–[5]. Besides, device-todevice (D2D) communication has also been proven promising
in enhancing the SE of the traditional cellular systems and
has drawn great attention recently [6]–[8]. Different from the
conventional cellular communication where all traffic is routed
via base station (BS), D2D communication allows two closely
located users to communicate directly, and thus have distinct
Copyright (c) 2015 IEEE. Personal use of this material is permitted. However, permission to use this material for any other purposes must be obtained
from the IEEE by sending a request to [email protected]. This work
was in part supported by the NSFC (Nos. 61372106, 61471114, & 61221002),
NSTMP under 2016ZX03001016-003, the Six Talent Peaks project in Jiangsu
Province under GDZB-005, Science and Technology Project of Guangdong
Province under Grant 2014B010119001, the Scholarship from the China
Scholarship Council (No. 201606090039), Program Sponsored for Scientific
Innovation Research of College Graduate in Jiangsu Province under Grant
KYLX16 0221, and the Scientific Research Foundation of Graduate School
of Southeast University under Grant YBJJ1651. (Corresponding author: Hao
Xu, Wei Xu.)
H. Xu, W. Xu, Z. Yang, J. Shi and M. Chen are with the National
Mobile Communications Research Laboratory, Southeast University, Nanjing
210096, China (Email: {xuhao2013, wxu, yangzhaohui, shijianfeng and
chenming}@seu.edu.cn).
advantages such as high SE, short packet delay, low energy
consumption and increased safety.
There has been extensive research on design and analysis of
massive MIMO systems [9]–[13]. In [9], the uplink capacity
bounds were derived under both perfect and imperfect channel
state information (CSI), and the tradeoff between SE and EE
was studied. Ref. [10] compared two most prominent linear
precoders with respect to (w.r.t.) SE and radiated EE in a
massive MIMO system. Unlike [9] and [10], which considered
simplified single-cell scenarios, [11]–[13] studied multi-cell
massive MIMO systems. As for underlaid D2D communication, a great challenge to the existing cellular architecture is the
cochannel interference due to spectrum reuse. There has been
a lot of literature working on interference mitigation for D2D
underlaid systems [14]–[20]. In [14]–[17], resource allocation
and power control algorithms were proposed to maximize the
SE of D2D users (DUs), and in [18]–[20], extended algorithms
were carried out to maximize the EE of DUs.
Though massive MIMO and D2D communication have been
widely studied, only a few papers investigated the interplay
between massive MIMO and D2D communication [21], [22].
In [22], the SE of cellular and D2D links was investigated
under both perfect and imperfect CSI, but the overhead for
acquiring CSI was not considered. In massive MIMO systems,
orthogonal pilots are transmitted by cellular users (CUs) to
obtain CSI. When D2D communication is introduced and
orthogonal pilots are used at each D2D transmitter (D2D-Tx)
for channel estimation, the pilot overhead is large which will
significantly affect the system performance. Furthermore, as a
multi-user transmission strategy, massive MIMO is designed
to support multiple users transmitting on the same timefrequency block. Though D2D-to-cellular interference can be
greatly reduced by a large antenna array at BS, cellular-to-D2D
interference still persists and may be worse than a conventional
single-input single-output (SISO) D2D underlaid system.
In order to shorten pilot overhead, an effective strategy
is allowing orthogonal pilots to be reused among different
users. Most of the existing works with pilot reuse (PR) mainly
focus on multiple-cell scenarios, i.e., mobile users in the same
cell use orthogonal pilots, and users in different cells reuse
the same set of pilots [23]–[27]. To the best of the authors’
knowledge, only a few works have considered the strategy
of PR within a cell [28]–[31]. In [28], the authors analyzed
the feasibility of PR over spatially correlated massive MIMO
channels with constrained channel angular spreads. Authors of
[29] and [30] allowed D2D-Txs to reuse the pilots of CUs and
proposed an interference-aided minimum-mean-square-error
2
(MMSE) detector to suppress the D2D-to-cellular interference.
[31] also studied a D2D underlaid massive MIMO system with
PR, but the performance of CUs was left out of consideration
for simplicity. In contrast to these existing works, our work
analyzes the achievable rate of both cellular and D2D links
under PR, and proposes pilot scheduling as well as power
control algorithms to optimize the system performance. The
main contributions of this paper are summarized as follows:
• We assume that CUs use orthogonal pilots while all D2DTxs reuse another set of pilots for channel estimation. The
motivation of PR stems from that D2D pairs usually locate
dispersively and use low power for short-distance transmission. Hence, letting several D2D pairs which are far away
from each other use the same pilot for channel estimation
would cause endurable pilot contamination. Under PR, we
first derive the expression of MMSE estimate of all channels.
With the obtained imperfect CSI, all receivers apply the partial
zero forcing (PZF) receive filters studied in [32] for signal
detection. Then, we derive the effective signal-to-interferenceplus-noise ratio (SINR) and a lower bound on the ergodic
achievable rate of each user.
• Different from the estimation of cellular channel vectors
which is only affected by noise, the channel estimation of D2D
links experiences effect from both noise and pilot contamination due to PR. To mitigate pilot contamination, we develop a
pilot scheduling and pilot power control algorithm under the
criterion of minimizing the sum mean-square-error (MSE) of
channel estimation of D2D links. We first show that with an
appropriate number of orthogonal pilots available for DUs and
a well designed pilot scheduling scheme, each D2D-Tx should
transmit its pilot using the maximum power. Then, we develop
a heuristic pilot scheduling scheme to allocate pilots to DUs,
and show that the sum MSE of channel estimation of D2D
links can be decreased significantly.
• We study the sum SE maximization for D2D links while
guaranteeing the quality of service (QoS) of CUs. Efficient
power control algorithms often play an important role in reducing cochannel interference and reaping the potential benefits of
D2D communication. However, these algorithms are usually
carried out based on the knowledge of instantaneous CSI of
all links [14]–[17]. Apart from the computational complexity,
such algorithms require BS to gather instantaneous CSI of all
links, which is difficult for implementation. Therefore, in this
paper, we consider performing the power control algorithm
periodically at a coarser frame level granularity based on
large-scale fading coefficients which vary slowly. Simulation
results show that the proposed algorithm converges rapidly and
obtains a much higher sum SE of D2D links compared to the
typical orthogonal training scheme.
Note that since PR among D2D pairs in a D2D underlaid
massive MIMO system has been less well researched, we
consider a simplified single-input multiple-output (SIMO)
transmission for D2D communication as [22], and mainly
focus on analyzing the effect of PR on the system performance.
As for MIMO transmission, similar results can be obtained
by simply modifying the analysis and optimization of this
manuscript. On the other hand, our analysis in this paper
focuses on a single-cell scenario for the sake of clarity. Re-
garding the multi-cell massive MIMO system, there has been
a lot of literature working on mitigating pilot contamination
[33]–[35]. As a result, for a multi-cell D2D underlaid massive
MIMO system, where PR among D2D pairs persists, we can
first use the algorithms developed in [33]–[35] to allocate
pilots to CUs if CSI can be exchanged among cells, and then
straightforwardly extend the proposed algorithms to improve
system performance.
In this paper, we follow the common notations. N, R and
C denote the set of natural numbers, the real space and the
complex space, respectively. The boldface upper (lower) case
letters are used to denote matrices (vectors). IM stands for
the M × M dimensional identity matrix and 0 denotes the
all-zero vector or matrix. “ \ ” represents the set subtraction
operation. Superscript (·)H denotes the conjugated-transpose
operation and E{·} denotes the expectation operation. We use
kak2 to denote the Euclidean norm of a. a 0(a ≻ 0)
means that each element in a is positive (nonnegative).
The rest of this paper is organized as follows. In Section
II, a D2D underlaid massive MIMO system is introduced. In
Section III, we present the MMSE estimate of all channels
under PR and show how PR affects the channel estimation.
The achievable rate of both cellular and D2D links is analyzed
in Section IV. In Section V, we aim to minimize the sum MSE
of channel estimation of D2D links and maximize the sum SE
of all D2D links. Finally, numerical verifications are presented
in Section VI before concluding remarks in Section VII.
II. S YSTEM M ODEL
Consider the uplink of a D2D underlaid massive MIMO
system with one BS, N CUs and K D2D pairs. The set
of CUs and D2D pairs are denoted by N = {1, · · · , N }1
and K = {1, · · · , K}, respectively. The BS is equipped with
B antennas and each CU has one transmit antenna. As for
the D2D communication, we assume SIMO transmission, i.e.,
each D2D-Tx is equipped with one antenna and each D2D
receiver (D2D-Rx) is equipped with M antennas as in [22].
In this system, all transmitters use the same time-frequency
resource block to transmit signals, leading to cochannel interference. The B × 1 dimensional received data vector at BS
is
y (c) =
N q
X
n=1
(c)
(c)
qs,n un h(c)
n xn +
K q
X
(d) (d) (d)
ps,i ui hi xi + z, (1)
i=1
(c)
where qs,n is the data transmit power of CU n, and xn is the
(c)
zero-mean unit-variance data symbol of CU n. un is the realvalued large-scale fading coefficient from CU n to BS and is
(c)
assumed to be known as a priori. hn ∼ CN (0, IB ) denotes
(d)
the fast fading vector channel from CU n to BS. ps,i , xi ,
(d)
(d)
ui and hi are similarly defined for D2D-Tx i. z ∈ CB is
the complex Gaussian noise at BS with covariance N0 IB .
1 Here we misuse the notation N while avoiding possible ambiguity with
the N in the complex normal distribution sign CN .
3
Analogously, the M × 1 dimensional received data vector
at D2D-Rx k is given by
(d)
yk =
N q
K q
X
X
(d) (d) (d)
(c) (c)
ps,i vik gik xi +
qs,n vnk gnk x(c)
n + nk , (2)
n=1
i=1
(d)
vik
(d)
gik
where
and
∼ CN (0, IM ) denote the real-valued
large-scale fading coefficient and the fast fading vector channel
(c)
(c)
from D2D-Tx i to D2D-Rx k, respectively. vnk and gnk
are similarly defined for the link from CU n to D2D-Rx k.
nk ∈ CM is the complex Gaussian noise at D2D-Rx k with
covariance N0 IM .
(d)
We then analogously derive the MMSE estimate of hk as
follows
q
(d)
pp,k uk
(d)
ĥk = P
Y (c) λk , ∀k ∈ K,
(7)
(d)
pp,i ui + N0
i∈Xk
where Xk is the set of all D2D pairs using the same pilot as
(d)
(d)
(d)
(d)
(d)
D2D pair k. Denote hk = ĥk + h̃k . Then, ĥk and h̃k
are statistically independent satisfying
(d)
A. Channel Estimation at BS
Similar as the uplink data transmission in (1), the B × τ
dimensional received signal matrix of pilot transmission at BS
is
K q
N q
X
X
(c) (c) H
(d) (d)
(c)
qp,n un hn ωn +
pp,i ui hi λH
Y =
i +Z, (3)
i=1
where qp,n and pp,i denote the pilot transmit powers of CU n
and D2D-Tx i, respectively. λi ∈ {ωN +1 , · · · , ωτ } is the pilot
allocated to D2D pair i. Z is the noise matrix which consists
of independently and identically distributed (i.i.d.) Gaussian
elements with zero mean and variance N0 . Then, the MMSE
(c)
estimate of hn is given by [36]
q
(c)
qp,n un
(c)
Y (c) ωn , ∀n ∈ N .
(4)
ĥn =
(c)
qp,n un + N0
(c)
Given the channel estimate vector ĥn , we can express the
(c)
(c)
(c)
(c)
true channel vector hn as hn = ĥn + h̃n , where the error
(c)
vector h̃n represents the CSI uncertainty. Due to the property
(c)
of MMSE estimation [36], ĥn is statistically independent of
(c)
h̃n , and they follow
(c)
(c)
(c)
ĥ(c)
n ∼ CN (0, δn IB ), h̃n ∼ CN (0, εn IB ),
(5)
where
(c)
δn(c)
=
qp,n un
(c)
qp,n un + N0
, ε(c)
n
=1−
δn(c) .
(6)
(d)
(d)
pp,k uk
(d)
Orthogonal pilots are usually adopted to obtain the CSI of
all links. In a D2D underlaid system, to reduce pilot overhead,
we assume that CUs use orthogonal pilots while all D2DTxs reuse another set of pilots for channel estimation. Denote
Ω = (ω1 , · · · , ωN , ωN +1 , · · · , ωτ ) ∈ Cτ ×τ as the pilot
matrix with orthogonal column vectors (i.e., ΩH Ω = Iτ ). τ
(N < τ ≤ N + K) is the length of the pilots and is also
the number of pilots available for channel estimation (this is
the smallest amount of pilots that are required). Then, without
loss of generality, we assume that pilot ωn (1 ≤ n ≤ N ) is
allocated to CU n and the remaining pilots {ωN +1 , · · · , ωτ }
are reused among all D2D pairs.
(d)
(8)
where
δk = P
III. C HANNEL E STIMATION
n=1
(d)
ĥk ∼ CN (0, δk IB ), h̃k ∼ CN (0, εk IB ),
(d)
pp,i ui
i∈Xk
(d)
+ N0
(d)
, ε k = 1 − δk .
(9)
B. Channel Estimation at D2D-Rxs
The M ×τ dimensional received pilot signal matrix at D2DRx k can be written as
N q
K q
X
X
(d)
(c) (c)
(d) (d)
Yk =
qp,n vnk gnk ωnH +
pp,i vik gik λH
i +Nk ,
n=1
i=1
(10)
where Nk is the noise matrix consisting of i.i.d. Gaussian
elements with zero mean and variance N0 . Then, the MMSE
(d)
estimate of gik is given by
q
(d)
pp,i vik
(d)
(d)
Yk λi , ∀i, k ∈ K,
(11)
ĝik = P
(d)
pp,j vjk + N0
j∈Xk
(d)
(d)
(d)
(d)
Denote gik = ĝik + g̃ik . Then, as mentioned above, ĝik is
(d)
statistically independent of g̃ik , and the distributions of them
are given by
(d)
(d)
(d)
(d)
ĝik ∼ CN (0, µik IM ), g̃ik ∼ CN (0, ǫik IM ),
(12)
where
(d)
(d)
µik = P
j∈Xk
pp,i vik
(d)
pp,j vjk
(d)
+ N0
(d)
, ǫik = 1 − µik .
(c)
Similarly, we have the MMSE estimate of gnk as follows
q
(c)
qp,n vnk
(c)
(d)
ĝnk =
Yk ωn , ∀n ∈ N , k ∈ K.
(13)
(c)
qp,n vnk + N0
(c)
(c)
Denote the channel estimation error vector by g̃nk , then, ĝnk
(c)
and g̃nk are statistically independent and satisfy
(c)
(c)
(c)
(c)
ĝnk ∼ CN (0, µnk IM ), g̃nk ∼ CN (0, ǫnk IM ),
where
(c)
µnk =
(c)
qp,n vnk
(c)
qp,n vnk
+ N0
(c)
(c)
, ǫnk = 1 − µnk .
(14)
(15)
From (4) and (13), it can be observed that the estimation
(c)
(c)
of hn and gnk is only affected by pilot noise. The pilot
interference from other CUs and DUs disappear completely
4
due to the orthogonality of the pilots. As for the estimation of
(d)
(d)
hk and gik , it is clear from (7) and (11) that apart from the
effect of pilot noise, it is also affected by pilot contamination
due to PR among D2D pairs.
IV. ACHIEVABLE R ATE A NALYSIS
In this section, we analyze the achievable rate of both
(c)
cellular and D2D links under PR. Let βn denote the unit
norm receive filter used by BS for detecting the signal of CU
(d)
n, and βk denote the unit norm receive filter adopted by
D2D-Rx k for detecting the signal from D2D-Tx k. Since the
receive filter can be used either to boost the desired signal
power or to eliminate interference signal, the SINR of each
link critically depends on the receive filter that is used. In this
paper, we adopt PZF receivers, which use part of degrees of
freedom for signal enhancement and the remaining degrees of
freedom for interference suppression, for signal detection at
both BS and D2D-Rxs.
We assume that BS uses bc and bd degrees of freedom to
cancel the interference from the nearest bc cellular interferers
and the nearest bd D2D interferers using different orthogonal
pilots. From (7), we can obtain the following relationship
v
u
u pp,k u(d) (d)
(d)
k
(16)
ĥ , ∀k ∈ K, i ∈ Xk \ k,
ĥk = t
(d) i
pp,i ui
which indicates that the estimation of channels from two
different D2D-Txs using the same pilot to BS are in the same
direction. As a result, the interference from D2D interferers
applying the same pilot can be eliminated simultaneously by
using one degree of freedom. Since τ − N orthogonal pilots
are reused by K D2D pairs, we divide all D2D pairs into
τ − N sets with D2D pairs in each set using the same pilot for
channel estimation. Then, we know that BS is able to cancel
the interference from bd sets of D2D interferers. The feasible
set of (bc , bd ) is given by
{(bc , bd ) ∈ N×N | bc ≤ N −1, bd ≤ τ −N, bc +bd ≤ B −1} .
(17)
(c)
The PZF filter βn can be obtained by normalizing the
(c)
projection of channel estimation ĥn on the nullspace of
channel estimation vectors of cancelled interferers (refer to
(c)
(45) in Appendix A). For the sake of convenience, let Cn
(c)
denote the set of uncancelled CUs when detecting xn , and
D(c) denote the set of uncancelled DUs when detecting cellular
signals.
Similarly, each D2D-Rx uses mc and md degrees of freedom to cancel the interference from the nearest mc cellular
interferers and the nearest md D2D interferers using different
orthogonal pilots. (mc , md ) should be in the following set
Compared with MMSE receivers which optimally balance
signal enhancement and interference suppression, PZF receivers are suboptimal. However, we adopt PZF receivers in
this paper due to the following advantages. First, PZF receivers
take both signal boosting and interference cancellation into
account, which is similar as MMSE receivers in concept. It has
been shown that the performance of PZF receivers in terms of
system throughput can approach that of MMSE receivers and
is much better than that of maximum ratio combining (MRC)
or fully zero forcing (ZF) receivers [32]. Second, the simple
structure of PZF receivers makes the analysis of the system
more tractable, and allows us to evaluate the performance of
the D2D underlaid massive MIMO system in a more explicit
way. In addition, PZF receivers can be simplified as MRC
receivers by letting (bc , bd ) = (0, 0) or (mc , md ) = (0, 0),
and can also be reduced to fully ZF receivers by letting
(bc , bd ) = (N −1, τ −N ) or (mc , md ) = (N, τ −N −1), which
can make the analysis of this paper more general. In order to
cancel the nearest interferers and obtain PZF receivers, BS
and D2D-Rxs have to know the positions of all transmitters.
To relax this requirement and make it more practical, we can
obtain PZF receivers by cancelling the interferers with the
largest large-scale fading coefficients.
A. A lower bound on achievable rate of cellular links
Since BS only has the information of estimated channel
vectors (4) and (7), which are treated as the true CSI, using
(c)
(c)
PZF receiver βn for detecting xn , we can write the postprocessing received signal associated with CU n at BS as
H
rn(c) = βn(c)
y (c)
q
H
(c)
(c)
= qs,n un βn(c) ĥ(c)
n xn
H
X q
Xq
(c)
(d) (d) (d)
(c)
qs,a ua ĥ(c)
ps,i ui ĥi xi
+ βn(c)
x
+
a
a
(c)
i∈D (c)
a∈Cn \n
+
N q
X
(c)
(c)
qs,a ua h̃(c)
a xa
+
a=1
K q
X
(d) (d) (d)
ps,i ui h̃i xi
!
+ z , (19)
i=1
where only the first term of the second equality is the desired
information, while the other terms represent the cochannel
interference, channel estimation error and noise, respectively.
Thus, the effective SINR of cellular link n can be expressed
as
(c)
{(mc , md ) ∈ N×N | mc ≤ N, md ≤ τ −N −1, mc +md ≤ M−1}.
(18)
(d)
The PZF filter βk can be obtained by first projecting chan(d)
nel estimation ĝkk onto the nullspace of channel estimation
vectors of cancelled interferers, and then normalizing the
(d)
(d)
projection. Let Ck and Dk respectively denote the sets of
(d)
uncancelled CU and DUs when detecting xk .
ηn(c) =
Sn
(c→c)
In
(c)
(c)
where Sn = qs,n un
(d→c)
+ In
(c→c)
signal from CU n, In
(c)
βn
H
(c)
+ α(c) βn
(c)
(20)
2
2
ĥn
(d→c)
and In
2,
represents the desired
respectively denote the
5
cochannel interference from all uncancelled cellular and D2D
interferers, and they are given by
2
H
X
ĥa(c) ,
In(c→c) =
qs,a ua(c) βn(c)
(c)
a∈Cn \n
X
In(d→c) =
(d)
ps,i ui
i∈D (c)
H
(d)
ĥi
βn(c)
2
.
(21)
α(c) characterizes the effect of both channel estimation error
and noise experienced by CU n, and can be formulated as
α(c) =
N
X
(c)
qs,a u(c)
a εa +
a=1
K
X
(d) (d)
ps,i ui εi + N0 .
(22)
i=1
In [9] and [22], the asymptotic uplink rate of cellular links in
a massive MIMO (or D2D underlaid massive MIMO) system
has been studied. As for the system considered in this paper,
we can also obtain a similar result about the asymptotic
uplink rate of CUs as shown in the following corollary. Since
corollary 1 can be analogously verified as that in [22], we omit
the proof process for brevity.
Corollary 1: With fixed transmit powers at all transmitters,
using linear filters for signal detection at BS, the asymptotic
uplink rate of each cellular link grows unboundedly as B goes
to infinity.
In fact, the number of antennas at BS is usually finite due
to multiple practical constraints. Hence, in the following of
this paper, we consider a more practical scenario where BS is
equipped with large but finite numbers of antennas. Consider
the block fading model, where all channels remain unchanged
over the coherence interval with length T . Then, based on
(5), (8) and (20), we can derive a lower bound on the ergodic
achievable rate of cellular links as shown in the following
theorem.
Theorem 1: Given the SINR formula in (20), the ergodic
achievable rate of cellular link n is lower bounded by
τ
Rn(c,lb) = 1 −
(23)
log2 1 + ηn(c,lb) , ∀n ∈ N ,
T
where
(c)
qs,n φn
ηn(c,lb) = N
,
(24)
P
(c)
(c)
qs,a ϕan + σ
a=1
and
(c)
φ(c)
− bc − bd − 1)u(c)
n = (B
n δn ,
(
(c)
(c)
ua ,
a ∈ Cn \ n
ϕ(c)
an =
(c) (c)
(c) ,
ua εa , a = n or a ∈ N \ Cn
σ (c) =
K
X
(d)
ϕi
=
(d)
To detect xk , the received signal at D2D-Rx k after using
(d)
PZF receiver βk can be written as
H
(d)
(d)
(d)
yk
rk = βk
q
H
(d)
(d)
(d) (d)
= ps,k vkk βk
ĝkk xk
q
H
X q
X
(d) (d) (d)
(c) (c)
(d)
ps,i vik ĝik xi +
qs,n vnk ĝnk x(c)
+ βk
n
(d)
(d)
i∈Dk \k
n∈Ck
!
N q
K q
X
X
(d) (d) (d)
(c) (c) (c)
ps,i vik g̃ik xi +
qs,n vnk g̃nk xn + nk . (26)
+
n=1
i=1
where only the first term of the second equality is the desired
signal, while the other terms respectively denote the cochannel
interference, channel estimation error and noise. Then, the
effective SINR of D2D link k is
(d)
Sk
(d)
ηk =
(d→d)
Ik
(d)
where Sk
(d)
= ps,k vkk
(c→d)
+ Ik
(d)
βk
H
(d→d)
Ik
,
(d)
+ αk
(d)
2,
(d)
βk
(27)
2
2
ĝkk
denotes the desired
(c→d)
Ik
signal from D2D-Rx k.
respectively denote
(d)
D2D and cellular cochannel interference, αk characterizes the
effect of both channel estimation error and noise experienced
by D2D-Rx k, and they are given by
2
H
X
(d)
(d→d)
(d)
(d)
Ik
=
ps,i vik βk
ĝik ,
(d)
(c→d)
Ik
=
X
(c)
qs,n vnk
(d)
n∈Ck
(d)
ps,i ϕi + N0 ,
(d)
αk =
K
X
i=1
(d)
i ∈ D(c)
ui ,
.
(d) (d)
ui εi , i ∈ K \ D(c)
B. A lower bound on achievable rate of D2D links
i∈Dk \k
i=1
(
(c)
affected by channel estimation error of cellular link n, i.e., εn ,
and channel estimation errors of bc cancelled cellular links,
(c)
(c)
i.e., εa , a ∈ N \ Cn . Since CUs use orthogonal pilots for
(c)
channel estimation, the value of τ has no effect on εn and
(c)
(c)
(c,lb)
εa , a ∈ N \Cn . Hence, for fixed pilot transmit power, Rn
decreases monotonically w.r.t. τ . In contrast, when bd > 0,
(c)
(c)
(c)
(c,lb)
except the effect of εn and εa , a ∈ N \ Cn , ηn
is
also influenced by the estimation errors of channels from bd
(d)
cancelled D2D-Txs to BS, i.e., εi , i ∈ K \ D(c) . Due to PR,
(d)
increasing τ results in smaller εi , i ∈ K \ D(c) , and thereby
(c,lb)
helps increase ηn . Therefore, it would be hard to determine
(c,lb)
the monotonicity of Rn
w.r.t. τ .
(25)
Proof: See Appendix A.
Remark 1: When considering the effect of pilot length τ
(c,lb)
(c,lb)
is
on Rn , from (24), we can find that when bd = 0, ηn
(d) (d)
ps,i vik ǫik +
H
(d)
(c)
βk
ĝnk
N
X
2
(c) (c)
,
qs,n vnk ǫnk + N0 .
(28)
n=1
Similarly as the cellular uplink case, we can also derive a
lower bound on the ergodic achievable rate of D2D links.
Theorem 2: Given the SINR formula in (27), the ergodic
achievable rate of D2D link k is lower bounded by
τ
(d,lb)
(d,lb)
Rk
= 1−
log2 1 + ηk
, ∀k ∈ K,
(29)
T
6
where
(d,lb)
ηk
(d)
=
ps,k φk
K
P
i=1
and
(d)
φk
(d)
ψik
(d)
σk
(d)
ps,i ψik
,
+
(30)
(d)
σk
(d) (d)
= (M − mc − md − 1)vkk µkk ,
(d)
(d)
vik , i ∈ Dk \ Xk
(d) (d)
(d)
= vik
,
ǫik , i = k or K \ Dk
(d) (d)
(d) (d)
(M −mc −md −1)vik µik +vik ǫik , i ∈ Xk \ k
X
X
(c)
(c) (c)
=
qs,n vnk +
qs,n vnk ǫnk + N0 .
(31)
(d)
Since orthogonal pilots {ωN +1 , · · · , ωτ } are reused among
K D2D pairs, denote the PR pattern by O ∈ R(τ −N )×K
with each element in O being binary-valued. If D2D pair k is
assigned pilot ωt ∈ {ωN +1 , · · · , ωτ }, we have ot−N,k = 1,
otherwise, we have ot−N,k = 0. Denote the pilot transmit
T
power vector of DUs by pp , [pp,1 , · · · , pp,K ] . Then, aiming
at minimizing (32), we arrive at the following problem
min
O,pp
s.t.
(d)
n∈N \Ck
n∈Ck
Proof: See Appendix B.
Remark 2: From (30), it can be found that for any (mc , md )
(d,lb)
in feasible set (18), ηk
is an implicit function of τ due to
PR. Increasing τ results in more accurate channel estimations
(d,lb)
of D2D links, and thereby helps increase ηk . However,
as τ increases, the number of symbols available for data
transmission becomes smaller. In Section VI, we show by
(d,lb)
simulation results that Rk
first increases and then decreases
w.r.t. τ .
In the following, we focus on pilot scheduling and power
control design based on (23) and (29). Since large-scale fading
coefficients vary slowly, the proposed pilot scheduling and
power control algorithms can be performed periodically at a
coarser frame level granularity, which will greatly decrease the
computational complexity of BS. As a result, Theorem 1 and
Theorem 2 are helpful for the following analysis.
V. P ILOT S CHEDULING
AND
P OWER C ONTROL
Up to now, we have investigated the cannel estimation as
well as achievable rate of both cellular and D2D links in a
D2D underlaid massive MIMO system with PR. Based on
the above analysis, we focus on two problems in this section.
The first problem aims to minimize the sum MSE of channel
estimation of D2D links, and the second problem aims to
maximize the sum rate of all D2D links while guaranteeing
the QoS requirements of CUs.
A. Pilot Power Control and Pilot Scheduling
As mentioned in Section III, the channel estimation of
cellular links is only affected by additive noise. Therefore, we
assume that each CU transmits pilot signal with the maximum
power. As for D2D links, apart from the effect of additive
noise, channel estimation is also influenced by pilot contamination. Due to the location dispersion of D2D pairs and shortdistance D2D transmission, it should be preferred that the pilot
contamination can be greatly reduced by using an effective
pilot scheduling and pilot power control algorithm. According
to (12) and (13), the sum MSE of channel estimation of D2D
links can be written as
X
K
K
X
2
(d)
(d)
E g̃kk
M ǫkk .
(32)
=
k=1
2
k=1
K
X
(d)
M ǫkk
(33a)
k=1
0 ≤ pp,k ≤ τ Pk , ∀k ∈ K,
τ
X
ot−N,k = 1, ∀k ∈ K,
(33b)
(33c)
t=N +1
where Pk is the maximum data transmit power of D2D-Tx k.
Constraints (33c) indicate that each D2D pair can be allocated
(d)
only one pilot. Note that for simplicity, ǫkk is formulated
as a function of O in an implicit way in (13). We can also
equivalently rewrite it in an explicit way as follows
(d)
ǫkk =
τ
X
ot−N,k 1 − K
P
t=N +1
j=1
(d)
pp,k vkk
(d)
, ∀k ∈ K.
ot−N,j pp,j vjk +N0
(34)
To solve problem (33), we first give the optimal condition
for pp in the following theorem.
Theorem 3: There always exists τ ≤ N + K such that when
the optimal pilot scheduling matrix O opt has been determined,
opt
the optimal popt
p satisfies pp,k = τ Pk , ∀k ∈ K.
Proof: See Appendix C.
The above theorem
indicates that with a proper τ and the optimal O opt , constraints
(33b) are always active. We can explain Theorem 3 in an
intuitive way as follows. When the number of orthogonal pilots
available for DUs is appropriate (i.e., with a relatively low PR
ratio) and these pilots are allocated to D2D pairs by using
the optimal pilot scheduling scheme (a special case is τ =
N + K and all D2D pairs use different orthogonal pilots for
channel estimation), ignorable pilot contamination would be
caused due to the dispersive positions of D2D pairs. Hence, all
D2D-Txs should transmit their pilot signals in the maximum
power to increase the estimation accuracy. In contrast, with
a small τ (i.e., with a relatively high PR ratio), even using
the optimal pilot scheduling scheme, pilot contamination may
still influence the estimation accuracy greatly and solving (59)
may yield popt
p,k = 0. In this case, we need to enlarge the set
of pilots for DUs to decrease PR ratio.
Based on Theorem 3, in the following, we assume that
D2D-Txs always transmit pilots in the maximum power. Then,
problem (33) becomes
min
O
s.t.
K
X
(d)
M ǫkk
k=1
τ
X
t=N +1
ot−N,k = 1, ∀k ∈ K,
(35a)
(35b)
which is a mixed integer programming problem. The optimal
O can be obtained through exhaustive search (ES). Recalling
7
(13), the number of scalar multiplication required to compute
the objective function in (33) is O(K). Thus, obtaining the optimal O through ES involves a complexity of O(K(τ −N )K ).
Due to the exponential complexity, it will be impractical to
run ES when the number of D2D pairs is large. Therefore, we
propose a low complexity pilot scheduling algorithm.
To mitigate pilot contamination in a multi-cell massive
MIMO system, [33] proposed the GCPA scheme, in which
a metric is defined to indicate the interference strength among
CUs and a binary matrix is used to describe the connections
of CUs. However, to obtain the binary matrix, a suboptimal
threshold needs to be found by applying iterative grid search.
In this paper, we define a continuous-valued metric χik to
evaluate the potential interference strength between D2D pair
i and k
∀k ∈ K, i = k
0,
(d) 2 (d) 2 !
χik =
.
v
vik
+ ki
, ∀k ∈ K, i ∈ K \ k
(d)
(d)
ln 1+ vkk
vii
(36)
Denote Λ as the set of D2D pairs which have been allocated
pilots, then, we summarize the pilot scheduling algorithm in
Algorithm 1.
Denote the data transmit power vectors of CUs and DUs
T
T
by qs , [qs,1 , · · · , qs,N ] and ps , [ps,1 , · · · , ps,K ] , respectively. Then, we arrive at the following problem
max
qs ,ps
k∈K\Λ i∈K
′
2: t = arg
min
P
t∈{N +1,··· ,τ } {i|λi =ωt }
χik′ ,
′
3: λk′ = ωt′ , ot′ −N,k′ = 1, Λ = Λ ∪ k .
end for
The basic idea of the PSA algorithm is that the D2D pair
experiencing larger pilot contamination possesses a higher
priority for pilot allocation. The main steps in each iteration
′
can be explained as follows. First, D2D pair k ∈ K \ Λ
experiencing the largest potential interference from other DUs
is selected. Then, the pilot causing the least interference to
′
′
k is chosen. Finally, pilot ωt′ is assigned to D2D pair k ,
′
i.e., ot′ −N,k′ = 1, and Λ is updated by Λ = Λ ∪ k . The
algorithm will be carried out for K times until all D2D pairs
are allocated with pilots.
B. Data Power Control
In this subsection, we aim to maximize the sum rate of
all DUs while guaranteeing the QoS
n requirements
o
n of oCUs.
(c)
(d)
Since the exact expressions of E Rn
and E Rk
are
unapproachable, we use their lower bounds (23) and (29) for
replacement. Simulation results show that the gap between
the ergodic achievable rate and its lower bound is marginal,
verifying the feasibility of the approximation.
(d,lb)
Rk
k=1
ηn(c,lb)
s.t.
(37a)
≥ γn , ∀n ∈ N ,
(37b)
0 ≤ qs,n ≤ Qn , ∀n ∈ N ,
0 ≤ ps,k ≤ Pk , ∀k ∈ K,
(37c)
(37d)
where γn and Qn respectively denote the target SINR and
the maximum data transmit power of CU n. From (24) and
(29), we can see that qs and ps are coupled in the expressions
(c,lb)
(d,lb)
of ηn
and Rk . Moreover, the fractional structure of
SINR expressions and the log(·) operation make (37) nonconvex. Therefore, it is difficult to obtain the optimal solution
of (37). In the following, we divide the optimization into two
consecutive parts. In the first part, we optimize qs with ps
fixed, and vice versa for the other part.
1) Data Power Control for Cellular Links: For given ps ,
problem (37) can be rewritten as
max
qs
Algorithm 1 Pilot Scheduling Algorithm (PSA)
Initialization:
Λ = ∅, O = 0. Calculate χik , ∀i, k ∈ K.
Pilot Allocation:
for j = 1, · · · , K do
P
′
χik ,
1: k = arg max
K
X
K
X
(d,lb)
Rk
k=1
ηn(c,lb)
s.t.
(38a)
≥ γn , ∀n ∈ N ,
(38b)
0 ≤ qs,n ≤ Qn , ∀n ∈ N .
(38c)
Based on (24), we can write SINR constraints (38b) in a vector
form as
qs F qs + θ , ∆(qs ),
(39)
where
F=
θ=
"
...
γ1 ϕ N 1
..
.
..
..
.
(c)
φ1
(c)
φN
γ1 σ (c)
(c)
φ1
(c)
φ1
.
,
(c)
(c)
γN ϕ1N
(c)
(c)
γ1 ϕ11
...
,··· ,
γN ϕ N N
(c)
φN
γN σ (c)
(c)
φN
#T
,
T
∆(qs ) = [∆1 (qs ), · · · , ∆N (qs )] .
(40)
In (40), θ consists of scaled cochannel interference from D2D
interferers and additive noise, and ∆(qs ) can be seen as an
interference function [37].
Remark 3: If problem (38) is feasible, the SINR constraint
opt
of each CU would hold with equality for the optimal qs .
opt
Equivalently, for the vector form (39), we have qs
=
F qsopt + θ. This can be readily verified by reductio. Assume
(c,lb)
opt
that ηn
> γn for qs , then, we can further increase the
opt
objective function by decreasing qs,n
slightly.
According to [37], if the spectral radius of F is less than 1,
opt
−1
the optimal power vector has the form qs = (IN − F ) θ .
opt
To obtain qs , matrix inversion and spectral radius calculation
are required, resulting in high complexity. As a result, we
obtain qsopt by using the following low complexity iterative
scheme
qs (l + 1) = Λ(qs (l)),
(41)
8
where l represents the time instant, and Λ(qs (l)) =
{min {Q1 , ∆1 (qs (l))} , · · · , min {QN , ∆N (qs (l))}}. We denote this data power control algorithm for cellular links by
DPCC. By proving that Λ(qs ) is standard [38], we can readily
verify that the DPCC algorithm converges to the optimal
solution for any initial power vector qs 0. The detailed
proof is given in Appendix D.
2) Data Power Control for D2D Links: For fixed qs , we
arrive at the following problem. For notational simplicity, the
constant coefficient 1 − Tτ in (29) is omitted.
K
X
max
ps
k=1
K
X
s.t.
k=1
(d)
ps,k φk
K
P
i=1
(d)
ps,i ψik
+
(d)
σk
ps,k ϕk ≤ ζ,
(42a)
(42b)
(42c)
where ζ = min {ζ1 , · · · , ζN }, and ζn =
N
P
(c)
qs,a ϕan − N0 .
(c)
1
γn qs,n φn
−
a=1
It can be directly seen that problem (42) is non-convex due
(d,lb)
to the fractional expression of ηk
and the log(·) operation in
(42a). In order to solve (42), we transform it into the equivalent
form in (60) which admits suboptimal solutions using the wellknown WMMSE approach in [39]. The details are provided in
Appendix E. We summarize the data power control algorithm
for D2D links (labeled as ‘DPCD’) in Algorithm 2.
Algorithm 2 Data Power Control Algorithm for D2D Links
(DPCD)
√
1: Initialize fk = Pk , wk = 1, ∀k ∈ K, and accuracy ρ1 .
2: repeat
3:
wk′ = wk , ∀k ∈ K,
4:
νk =
5:
−1
q
(d)
wk = 1 − νk fk φk
, ∀k ∈ K,
q
(d)
fk φk
, ∀k
K
P
(d)
(d)
(d)
fk2 φk +
fi2 ψik +σk
∈ K,
i=1
7:
Since we propose to solve the original problem (37) using
the JDPC algorithm, which operates in an iterative mechanism, it is necessary to characterize its convergence. In each
iteration, the optimal qs is first obtained by using the DPCC
algorithm with fixed ps . Then, for determined qs , the DPCD
algorithm outputs a suboptimal ps . As a result, the sum SE
of D2D links increases in each iteration. Noting the fact that
the sum SE of D2D links is always upper bounded, we can
conclude that the JDPC algorithm converges to a suboptimal
solution of problem (37).
(d)
log2
1 +
0 ≤ ps,k ≤ Pk , ∀k ∈ K,
6:
C. Convergence Analysis for the JDPC Algorithm
Obtain λ by using the bisection method,
q
√
(d)
wk νk φk
Pk ,
, ∀k ∈ K,
fk = min
K
P
(d)
(d)
(d)
wk ν 2 φ +
wi νi2 ψ +λϕ
k
k
ki
D. Implementation and Complexity Analysis
Based on the above analysis, within a coherence interval,
each transmitter first transmits its pilot with the maximum
power for channel estimation, and then transmits its signal
using the power obtained by solving (37). To successfully
implement the proposed PSA and JDPC algorithms, BS
requires the information of large-scale fading coefficients
(c)
(d) (c) (d)
(c)
(d)
un , uk , vnk , vik , ∀n ∈ N , ∀i, k ∈ K. un , uk , ∀n ∈ N
(c) (d)
can be directly estimated by BS, whereas vnk , vik , ∀n ∈
N , ∀i, k ∈ K need to be estimated at all D2D-Rxs and fed
back to BS. After collecting these information, BS runs the
pilot scheduling and power control algorithms, and then sends
the results to all users. Since large-scale fading coefficients
vary slowly, the proposed algorithms can be carried out at a
coarser frame level granularity.
In the following, we analyze the computational complexity
of the PSA and JDPC algorithms. Since the PSA algorithm
has a complexity of O(K 2 ) per iteration and it is carried out
K times, the total complexity is O(K 3 ). As for the JDPC
algorithm, assume that it is carried out L1 times, and in
each loop L2 and L3 iterations are required for the DPCC
and DPCD algorithms to converge. The major complexity of
the DPCC algorithm lies in computing F qs , which involves
a complexity of O(N 2 ). Hence, the total complexity of the
DPCC algorithm is O(L2 K 2 ). Since the complexity of obtaining the Lagrange multipliers using the bisection method for
accuracy ρ is O(log2 (1/ρ)), the overall complexity of solving
problem (42) using the DPCD algorithm is O(L3 log2 (1/ρ)).
Therefore, the JDPC algorithm has a total complexity of
O(L1 (L2 K 2 + L3 log2 (1/ρ))). Simulation results show that
L1 , L2 and L3 are small, so the proposed algorithms involve
low complexity for efficient solutions.
k
i=1
8:
until
K
P
|ln wk − ln wk′ | ≤ ρ1 ,
k=1
9: ps,k = fk2 , ∀k
∈ K.
Based on the above analysis, we can alternatively optimize
qs and ps by iteratively applying the proposed DPCC and
DPCD algorithms, and a suboptimal solution of problem (37)
can be obtained. Let JDPC denote the joint data power control
algorithm which iteratively carries out DPCC and DPCD until
convergence. For brevity, we omit the detailed description of
the JDPC Algorithm.
VI. S IMULATION R ESULTS
In this section, we present simulation results to evaluate
the performance of the D2D underlaid massive MIMO system
with PR. We consider a network in a 1000 m × 1000 m square
area and all transmitters are located uniformly in this cell. The
distance between a D2D-Tx and its associated receiver is uniformly distributed in the range of [0 m, Dmax m]. For brevity,
we assume equal minimum SINR requirement for CUs, i.e.,
γn = γ, ∀n ∈ N , and equal maximum power constraint
for all transmitters, i.e., Qn = Pk = P, ∀n ∈ N , k ∈ K.
According to Theorem 3, with a relatively low PR ratio, all
9
100
17 dBm
-100 dBm
3.7
8 dB
10−3
45
40
Sum SE of CUs (bps/Hz)
τ=25
35
30
70
60
50
30
6
τ=10
15
10
Fully ZF, Simulation
Fully ZF, Lower bound
MRC, Simulation
MRC, Lower bound
τ=25
5
2000
3000
4000
10
12
14
16
18
20
22
24
Fig. 2. Simulated sum SE of D2D links and its lower bound versus pilot
length with qs,n = ps,k = P, ∀n ∈ N , k ∈ K, N = 5, K = 20, B = 1024,
T = 50 and Dmax = 100.
τ=15
1000
8
Pilot length
τ=10
20
0
80
40
τ=15
25
M=8, PZF, Simulation
M=8, PZF, Lower bound
M=8, MRC, Simulation
M=8, MRC, Lower bound
M=4, PZF, Simulation
M=4, PZF, Lower bound
M=4, MRC, Simulation
M=4, MRC, Lower bound
90
Sum SE of DUs (bps/Hz)
TABLE I
S IMULATION PARAMETERS
Maximum data transmit power P
Additive noise power N0
Path loss exponent
Standard deviation of log-normal shadowing fading
Accuracy ρ, ρ1
5000
6000
7000
8000
9000 10000
Number of BS Antennas
Fig. 1. Simulated cellular sum SE and its lower bound versus the number
of BS antennas with qs,n = ps,k = P, ∀n ∈ N , k ∈ K, N = 5, K = 20,
M = 8, T = 50 and Dmax = 100.
D2D-Txs can transmit their pilot signals in the maximum
power to increase the estimation accuracy. Therefore, in the
following, we assume that qp,n = pp,k = τ P, ∀n ∈ N , k ∈ K.
Unless otherwise specified, the other system parameters are
summarized in Table I. All simulation results are obtained
by averaging over 104 channel realizations, and each channel
realization is obtained by generating a random user distribution
as well as a random set of fading coefficients.
A. Spectral Efficiency Versus the Corresponding Lower Bound
In this subsection, we evaluate the tightness between the
simulated SE and the corresponding lower bound of both
cellular and D2D links.
First, we compare the simulated cellular sum SE with the
corresponding lower bound under different values of pilot
length and different receive filters at BS in Fig. 1. It can
be seen that cellular sum SE increases with the number of
BS antennas for all considered configurations, and fully ZF
receivers (i.e., (bc , bd ) = (N − 1, τ − N )) greatly outperform
MRC receivers (i.e., (bc , bd ) = (0, 0)) in terms of cellular SE.
As pilot length τ grows, cellular sum SE decreases for the
MRC case, while increases for the fully ZF case. This has
been explained in Remark 1. Moreover, Fig. 1 also shows
that the lower bound of cellular sum SE closely matches the
simulation for the MRC case, and almost coincides with the
simulation for the fully ZF case.
Next, in Fig. 2, we compare the simulated sum SE of
D2D links with the corresponding lower bound under different
numbers of D2D-Rx antennas and different receive filters.
Note that when D2D-Rxs adopt PZF receivers for signal
detection, as mentioned in Section IV, (mc , md ) should be
in feasible set (18). Therefore, in Fig. 2, we assume that
(mc , md ) = (1, min{τ − 6, 1}) when M = 4, and (mc , md ) =
(1, min{τ − 6, 2}) when M = 8. It can be seen from this
figure that PZF receivers greatly outperform MRC receivers
in terms of D2D sum SE. When M = 4, the lower bound on
the sum SE of DUs obtained by PZF receivers is approximate
to the simulated SE obtained by MRC receivers. The gap
between the simulated SE and its lower bound is small for
all considered configurations, indicating that it is feasible to
solve the data power control problem in Section V-B based
on lower bounds (24) and (29). Fig. 2 also shows that when
τ ≤ 9, D2D sum SE increases with pilot length, while after
that D2D sum SE decreases almost linearly with pilot length.
This is because only a few orthogonal pilots are reused among
D2D pairs for a small τ . In this case, the channel estimation is
significantly influenced by pilot contamination. Therefore, the
SE of D2D links can be enhanced by increasing τ . However,
as τ becomes large enough, the channel estimation accuracy
can be hardly improved by further enlarging τ . Counterproductively, increasing pilot length reduces the number of symbols
available for data transmission, and thereby decreases the SE
of D2D links.
In the following simulation, we assume that BS applies fully
ZF receivers for signal detection. As for D2D communication,
since the number of D2D-Rx antennas may be smaller than the
number of all transmitters in the cell, we assume that D2DRxs apply PZF receivers for signal detection. In addition,we
set B = 1024 and M = 8.
B. Performance of the PSA Algorithm
In this subsection, we evaluate the performance of the
proposed PSA algorithm. For comparison, in Fig. 3, we plot
sum MSE (32) versus pilot length for different pilot scheduling
algorithms: the proposed PSA scheme, the GCPA algorithm
proposed in [33] and the random pilot scheduling (RPS).
As a benchmark, we also depict the lower bound on the
sum MSE, which is obtained when all users apply different
orthogonal pilots for channel estimation. It can be observed
from Fig. 3 that the sum MSE decreases with pilot length for
all algorithms, which is consistent with intuition. Moreover,
10
2
10
110
100
1
Sum SE of DUs (bps/Hz)
Sum MSE
10
0
10
Proposed PSA
GCPA in [33]
RPS
Lower bound
−1
10
−2
90
80
70
60
(m , m )=(1, 2), τ=10
c
10
−3
10
6
8
10
12
14
16
18
20
22
40
1
24
c
2
3
4
90
80
70
(mc, md)=(1, 2), τ=10
(mc, md)=(1, 2), OT
MRC, τ=10
MRC, OT
50
180
80
70
60
(mc, md)=(1, 2), τ=10
(mc, md)=(1, 2), OT
50
4
6
8
Number of iterations
7
8
9
10
Fig. 5. Convergence behaviors of the JDPC algorithm with N = 5, K = 20,
T = 50, γ = 5 dB and Dmax = 100.
10
40
MRC, τ=10
MRC, OT
10
20
30
40
50
Number of iterations
Fig. 4. Convergence behaviors of the inner updates in the first outer iteration
of the JDPC algorithm with N = 5, K = 20, T = 50, γ = 5 dB and
Dmax = 100.
the proposed PSA scheme approaches the lower bound quickly
as τ increases, and outperforms the other two algorithms
significantly in terms of sum MSE.
C. Performance of the JDPC Algorithm
In this subsection, we evaluate the performance of the
proposed JDPC algorithm. Denote the orthogonal training
scheme (i.e., τ = N + K) by ‘ OT ’.
Fig. 4 and Fig. 5 illustrate the convergence behaviors of
the proposed power control algorithm under different configurations. Specifically, the left and right panels of Fig. 4
respectively correspond to the DPCC algorithm and the DPCD
algorithm, while Fig. 5 corresponds to the JDPC algorithm.
It can be seen from Fig. 4 that the sum SE of D2D links
monotonically increases during the iterative procedure and
converges rapidly for both DPCC and DPCD algorithms. By
iteratively carrying out these two algorithms to optimize the
data transmit power of CUs and DUs, the sum SE of D2D
links can be effectively increased. Fig. 5 shows that the JDPC
algorithm converges after only a few iterations (within 3
iterations for all considered configurations). This makes the
proposed data power control algorithm suitable for practical
applications.
Sum SE of the system (bps/Hz)
90
Sum SE of DUs (bps/Hz)
Sum SE of DUs (bps/Hz)
100
2
6
DPCD algorithm
100
40
5
Number of iterations
Fig. 3. Performances of different pilot scheduling algorithms versus pilot
length with N = 5, K = 20 and Dmax = 100.
60
d
MRC, τ=10
MRC, OT
Pilot length
DPCC algorithm
d
(m , m )=(1, 2), OT
50
Proposed scheme, τ=10
Proposed scheme, τ=15
Proposed scheme, OT
Classical massive MIMO
ICAPC in [40]
160
140
120
100
80
60
50
60
70
80
90
100
110
120
130
140
150
Maximum distance between a D2D pair (m)
Fig. 6. Sum SE of the system versus the maximum distance between a D2D
pair with (mc , md ) = (1, 2), N = 5, K = 20, T = 50 and γ = 5 dB.
In Fig. 6, the sum SE of the system versus the maximum
distance between a D2D pair is depicted. We compare the
proposed scheme with the iterative channel allocation and
power control (ICAPC) algorithm proposed in [40]. Note
that [40] considered a D2D underlaid cellular system with
each transceiver equipped with one antenna. Therefore, to be
fair, we set the same system configurations when simulating
the ICAPC algorithm. We also include the sum SE curve
obtained by using classical massive MIMO communication
as a benchmark in Fig. 6. To obtain this benchmark, instead
of applying direct communication between D2D pairs, data
signals of D2D-Txs are first forwarded to BS and then sent
(c)
to D2D-Rxs. Let Rn denote the uplink SE of cellular link n,
(d)
and Rk denote the SE of the transmission from D2D-Tx k to
D2D-Rx k with BS working as the relay. Then, the benchmark
N
K
P
P
(c)
(d)
Rn +
can be obtained by maximizing R =
Rk .
n=1
k=1
When fully ZF receivers are applied at the BS, the maximum R
can be obtained by letting all transmitters using the maximum
power for signal transmission. As expected, Fig. 6 shows
that the sum SE of the system decreases with Dmax when
D2D communication is adopted. Since the ICAPC algorithm
assumed perfect CSI and aimed to maximize the sum SE
11
300
110
N=5
N=10
N=15
105
Sum SE of DUs (bps/Hz)
Sum SE of DUs (bps/Hz)
250
200
150
100
100
95
(m , m )=(0,0)
c
90
d
(m , m )=(1,1)
c
d
(m , m )=(1,2)
c
85
d
(m , m )=(2,2)
c
d
(m , m )=(2,3)
c
d
(m , m )=(2,4)
50
10 20
40
60
80
100
120
140
160
180
Number of D2D pairs
180
Sum SE of DUs (bps/Hz)
160
140
120
100
80
Proposed scheme, K=20, τ=10
Proposed scheme, K=40, τ=10
Proposed scheme, K=20, OT
Proposed scheme, K=40, OT
ICAPC in [40], K=20
ICAPC in [40], K=40
40
20
0
50
60
70
80
90
100
110
c
−5
d
0
5
10
15
20
Target SINR of CUs (dB)
Fig. 7. Sum SE of D2D links versus the number of D2D pairs under different
values of coherence block length with (mc , md ) = (1, 2), τ = 10, T = 50,
γ = 5 dB and Dmax = 100.
60
80
−10
120
130
140
150
Coherence block length
Fig. 8. Sum SE of D2D links versus the coherence block length with
(mc , md ) = (1, 2), N = 5, γ = 5 dB and Dmax = 100.
of DUs based on instantaneous CSI, the ICAPC algorithm
outperforms the proposed scheme slightly in terms of system
throughput when OT is adopted. However, with PR among
DUs, the proposed scheme can obviously increase the system
SE. Moreover, we can also conclude from Fig. 6 that due to the
property of short-distance transmission, D2D communication
can help improve the sum SE of a massive MIMO system
significantly especially when Dmax is small.
Fig. 7 depicts the sum SE of D2D links versus the number of
D2D pairs under different values of N . It can be seen that the
sum SE of DUs first increases with K and then approaches
a saturation point when K > 150. This is because we set
τ = 10, as K grows, PR ratio increases, resulting in more
pilot contamination for D2D channel estimations. Moreover,
since a larger N results in more cochannel interference and
requires longer pilot overhead, as Fig. 7 shows, the sum SE
of DUs decreases with N .
The impact of the coherence block length T on the sum SE
of DUs is investigated in Fig. 8. As expected, the sum SE of
DUs increases with T for all considered configurations. For
both the proposed scheme with OT and the ICAPC algorithm,
as T grows, the sum SE of DUs with K = 40 is first lower
and then higher than that of the K = 20 case. This is because
Fig. 9. Sum SE of D2D links versus the target SINR of CUs under different
(mc , md ) with N = 5, K = 20 , τ = 10, T = 50 and Dmax = 100.
with OT, in the small T regime, the sum SE of DUs is mainly
influenced by τ , while in the high T regime, it is mainly
affected by K. Moreover, Fig. 8 also shows that compared
with OT, the sum SE of DUs can be greatly increased by PR
especially when T is small.
Fig. 9 illustrates how the target SINR of CUs and PZF
parameters (mc , md ) affect the sum SE of D2D links. Several
observations can be made from this figure. First, the sum SE
of D2D links decreases with γ, and the loss in D2D sum SE
resulted from the increase of γ reduces as mc grows from 0
to 2. Second, PZF receivers (i.e., mc + md > 0) outperform
the MRC receiver (i.e., mc = md = 0) in terms of D2D
sum SE for all different choices of (mc , md ). Moreover, for
the considered configuration, the maximum sum SE of D2D
links is obtained when (mc , md ) = (1, 2). Further increasing
the degrees of freedom for interference suppression results in
decrease of D2D sum SE since less degrees of freedom are
left for signal enhancement.
VII. C ONCLUSIONS
In this paper, we consider a D2D underlaid massive MIMO
system. Due to the fact that D2D pairs are usually located
dispersively and conduct short-distance transmission in low
power, letting several D2D pairs far from each other use
the same pilot for channel estimation would be feasible and
beneficial. Hence, we allow PR among DUs to reduce pilot
overhead. Based on this setup, we first investigate the channel
estimation under PR and derive a lower bound on the ergodic
achievable rate of each link. To mitigate pilot contamination
caused by PR, we develop a pilot scheduling algorithm under
the criterion of minimizing the sum MSE of channel estimation
of D2D links. In addition, we also maximize the sum rate of all
D2D links based on the large-scale fading coefficients instead
of the instantaneous CSI. An iterative power control algorithm
is proposed to obtain a suboptimal solution. Simulation results
show that the effect of pilot contamination can be decreased
greatly by exploiting the proposed pilot scheduling algorithm,
and the PR scheme can provide significant performance gains
over the conventional orthogonal training scheme in terms of
system SE.
12
A PPENDIX A
P ROOF OF T HEOREM 1
H
H
H
(c)
(c)
(c)
(c)
(c)
(d)
By definition, βn
ĥn , βn
ĥa and βn
ĥi
are respectively the projections of channel estimation vectors
(c)
(c)
(c)
(d)
ĥn , ĥa and ĥi onto PZF receiver βn . As mentioned in
Section III, channel estimation of cellular links is only affected
(d)
(c)
(c)
by additive noise. Hence, ĥn is independent of ĥa and ĥi ,
∀n ∈ N , a ∈ N \ n, i ∈ K. Then, from (20), we can find that the
(c)
desired signal of CU n Sn doesn’t dependent on cochannel
(c→c)
(d→c)
cellular and D2D interference
In
and In
. Using the
convexity of log2 1 + x1 (∀x > 0) and applying the Jensen’s
inequality, we have
E Rn(c) ≥ Rn(c,lb)
)!−1
(
τ
1
, ∀n ∈ N , (43)
, 1−
log2 1 + E
(c)
T
ηn
(c)
(c,lb)
where Rn and Rn
denote the instantaneous rate and a
lower bound on the ergodic achievable rate of cellular link n,
respectively. Based on (20) and the above analysis, we have
(
)
n
o
n
o
1
1
(c→c)
(d→c)
(c)
E
=
E
E
I
+
E
I
+
α
.
n
n
(c)
(c)
ηn
Sn
(44)
In order to obtain
expression
o we
o
n of (44),
n
n ano explicit
(d→c)
(c→c)
. For
and E In
need to calculate E 1(c) , E In
Sn
convenience, denote G = [G1 , G2 ] ∈ CB×B , where the
columns of G form an orthogonal basis of the B dimensional
space. Specifically, the columns of G2 ∈ CB×(bc +bd ) form
an orthogonal basis of the subspace spanned by channel
estimation vectors of cancelled interferers, and each column
of G1 ∈ CB×(B−bc −bd ) is orthogonal to G2 . Then, the unit
(c)
norm PZF receiver βn can be chosen as
(c)
βn(c) =
G1 GH
1 ĥn
(c)
G1 GH
1 ĥn
.
(45)
(c)
(c)
=
(c)
GH
1 ĥn
2
,
1
(c)
2
n
E In(c→c)
o
=
H
ĥ(c)
βn(c)
a
qs,a u(c)
a E
X
(c)
qs,a u(c)
a δa ,
(c)
a∈Cn \n
=
(
X
2
)
(c)
a∈Cn \n
o
n
X
(d)
ps,i ui E
E In(d→c) =
i∈D (c)
=
X
(
(d) (d)
βn(c)
H
(d)
ĥi
2
ps,i ui δi .
)
(48)
i∈D (c)
Substituting (47) and (48) into (43) yields lower bound (23).
A PPENDIX B
P ROOF OF T HEOREM 2
Unlike the channel estimation of cellular links which is only
affected by additive noise, pilot contamination exists when
estimating D2D channels due to PR. As a result, the estimation
(d)
of gkk is not only affected by noise, but also influenced by
pilot contamination caused by DUs i ∈ Xk \ k. From (11), we
can obtain the following relationship
v
u (d)
uµ
(d)
(d)
ĝik = t ik
ĝ , ∀i ∈ Xk \ k.
(49)
(d) kk
µkk
Denote
(d→d)
(d→d)
Iˆk
(d→d)
(d→d)
= Iˆk
+ I˜k
,
2
H
X
(d)
(d)
(d)
=
ps,i vik βk
ĝik ,
(d)
i∈Dk \Xk
(d→d)
I˜k
=
(d)
X
ps,i vik
i∈Xk \k
(46)
(c)
projection of ĥn on G1 . Obviously, ĥn is independent
(c)
of each column of G1 . Hence, each element of GH
1 ĥn
(c)
follows i.i.d. CN (0, δn ). According to [41, Section 2.1.6],
GH
∼ W1 (B − bc − bd , 1) is a central complex
1 ĥn
2
Wishart random variable with B − bc − bd degrees of freedom.
Then, we have
1
1
=
.
(47)
E
(c)
(c)
(c)
Sn
qs,n un (B − bc − bd − 1)δn
(c)
δn
δi
(d)
βk
H
(d)
ĝik
2
.
(50)
Then, from (49), we have
2
H
(c)
(c)
which indicates that βn
ĥn equals to the norm of the
(c)
(c)
complex Gaussian random variables and follows CN (0, δa ).
2
H
(c)
(c)
Hence, βn
follows exponential distribution, i.e.,
ĥa
2
H
(c)
(c)
(c)
1
βn
∼ Exp(1), ∀a ∈ Cn \ n. Similarly,
ĥa
(c)
δa
2
H
(c)
(d)
1
∼ Exp(1), ∀i ∈ D(c) . Then, we have
βn
ĥi
(d)
Ik
2
By expressing ĥn as the sum of projections of ĥn onto G1
and G2 , we have
H
H
(c)
(c)
H (c)
ĥ(c)
G1 GH
βn(c)
1 ĥn + G2 G2 ĥn
n = βn
(c)
= G1 GH
1 ĥn
(c)
Since βn is an unit
norm vector and is independent of
H
(c)
(c)
(c)
(c)
ĥa , ∀a ∈ Cn \ n, βn
ĥa is the linear combination of
(d→d)
I˜k
(d)
Sk
=
P
i∈Xk \k
(d) (d)
ps,i vik µik
(d) (d)
ps,k vkk µkk
.
(51)
Analogous to the proof process in Appendix A, we can
(d)
(d→d)
(c→d)
easily verify that Sk is independent of Iˆk
and Ik
.
Hence, the relationship between the ergodic achievable rate of
D2D link k and its lower bound can be expressed as
(d)
(d,lb)
E Rk ≥ Rk
)!−1
(
τ
1
, ∀k ∈ K, (52)
, 1−
log2 1+ E
(d)
T
η
k
13
where
(
E
Meanwhile, the optimal values of (κopt , ξ opt ) should satisfy
1
(d)
ηk
)
= E
+
(
1
(d)
Sk
(d→d)
I˜k
(d)
Sk
)
n
o
n
o
(d→d)
(c→d)
(d)
E Iˆk
+ E Ik
+ αk
.
(53)
Besides, we also have
(
)
1
1
E
=
,
(d)
(d)
(d)
Sk
ps,k vkk (M − mc − md − 1)µkk
n
o
X
(c→d)
(c) (c)
E Ik
=
qs,n vnk µnk ,
(d)
n∈Ck
n
o
1
, ∀k ∈ K,
Vk (popt
p )
opt
opt
Uk (popt
p ) − ξk Vk (pp ) = 0, ∀k ∈ K.
κopt
k =
(d→d)
E Iˆk
=
X
(d) (d)
ps,i vik µik .
(54)
(d)
i∈Dk \Xk
Substituting (51) and (54) into (52) yields lower bound (29).
(58)
opt
opt
From Lemma 1, we can obtain (popt
p , ξ , κ ) by iteratively
carrying out the following two steps until convergence: 1)
update (ξ, κ) based on (58) for given pp ; 2) update pp
for given (ξ, κ) by solving (57). In the last iteration of
the parametric algorithm, assume that (ξ opt , κopt ) has been
obtained. Then, we find the final popt
p by solving (57), which
can be further reformulated as
!
!
K
X opt opt (d)
X
opt (d)
opt opt
max
κi ξi vki −κk ξk N0
pp,k κk vkk −
pp
k=1
i∈Xk
(59a)
A PPENDIX C
P ROOF OF T HEOREM 3
s.t.
When the optimal pilot scheduling scheme has been determined to allocate pilots {ωN +1 , · · · , ωτ } to DUs, problem
(33) can be equivalently written as
max
pp
s.t.
K
X
Uk (pp )
k=1
Vk (pp )
=
K
X
k=1
(d)
pp,k vkk
P
i∈Xk
(d)
pp,i vik + N0
0 ≤ pp,k ≤ τ Pk , ∀k ∈ K.
(55a)
(55b)
Obviously, problem (55) is a non-convex sum-of-ratios optimization, which aims to maximize the summation of fractional
functions. The parametric algorithm is often adopted to solve
this kind of problem if the numerator of each summation term
is concave and the denominator of each summation term is
convex [42], [43]. Since Uk (pp ) and Vk (pp ) are both affine
w.r.t. pp for any k, we can optimally solve problem (55)
using the parametric algorithm. According to [42], (55) can
be equivalently transformed to the following problem
max
pp ,ξ
s.t.
K
X
ξk
pp
s.t.
K
X
k=1
κopt
Uk (pp ) − ξkopt Vk (pp )
k
0 ≤ pp,k ≤ τ Pk , ∀k ∈ K.
Since the above problem in (59) is linear w.r.t. pp , it is
P opt opt (d)
opt (d)
κi ξi vki ;
directly known that popt
p,k = τ Pk if κk vkk ≥
i∈Xk
opt
Otherwise, popt
p,k = 0. Notice that we have to guarantee pp,k > 0
opt
to estimate the channel vector of D2D link k. When pp,k = 0,
we can always change the pilot assigned to D2D pair k by
P opt opt (d)
(d)
κi ξi vki . The worst
increasing τ so that κopt
k vkk ≥
i∈Xk
case is that τ = N + K and all D2D pairs use different
orthogonal pilots for channel estimation. In this case, pilot
contamination vanishes and each D2D-Tx would transmit
its pilot with the maximum power to increase the channel
estimation accuracy. Therefore, Theorem 3 is proven.
A PPENDIX D
C ONVERGENCE P ROOF OF OF THE DPCC A LGORITHM
(57a)
(57b)
A PPENDIX E
(56b)
(56c)
where ξ = (ξ1 , · · · , ξK )T . In order to solve the problem in (56),
we resort to alternating optimization by using the following
Lemma. Applying [42, Lemma 2.1], we obtain
opt
Lemma 1: If (popt
p , ξ ) is the optimal solution of the
above maximization problem in (56), then there exists κopt =
opt T
opt
(κopt
1 , · · · , κK ) such that pp is the optimal solution to the
following problem
max
(59b)
As discussed in [38, Theorem 2], the iterative process in
(41) converges to the optimal solution for any initial power
vector qs 0 if function Λ(qs ) is standard. Moreover, the
sufficient condition for Λ(qs ) to be standard is that ∆(qs )
is standard. Therefore, we only need to prove that ∆(qs ) is
standard. Since the elements in F are nonnegative, qs 0
and θ ≻ 0, we have
• Positivity: ∆(qs ) = F qs + θ ≻ 0;
′
′
• Monotonicity: If qs qs , then ∆(qs ) − ∆(qs ) = F (qs −
′
′
qs ) 0, i.e. ∆(qs ) ∆(qs );
• Scalability: ∀ς > 1, ς∆(qs ) = ςF qs + ςθ ≻ ςF qs + θ =
∆(ςqs ).
As a result, Λ(qs ) is standard and the iterative process in
(41) converges to the optimal solution for any initial power
vector qs 0.
(56a)
k=1
Uk (pp )
≥ ξk , ∀k ∈ K,
Vk (pp )
0 ≤ pp,k ≤ τ Pk , ∀k ∈ K,
0 ≤ pp,k ≤ τ Pk , ∀k ∈ K.
14
S OLVING (42) VIA WMMSE A LGORITHM
√
For notational convenience, we denote fk = ps,k , ∀k ∈ K.
Then, (42) can be equivalently transformed into the following
weighted sum-MSE minimization problem
min
w,ν,f
s.t.
K
X
k=1
K
X
k=1
(wk ek − ln wk )
(d)
fk2 ϕk ≤ ζ,
i=1
(62)
Plugging (62) in (61) and simplifying (60), we have
s.t.
K
X
k=1
(d)
log2
1 +
(d)
fk2 φk
K
P
i=1
(d)
fk2 ϕk ≤ ζ,
0 ≤ fk ≤
(d)
fi2 ψik + σk
(63a)
(63b)
p
Pk , ∀k ∈ K,
(63c)
which is equivalent to (42). The equivalence implies that
a suboptimal solution of problem (42) can be obtained by
solving (60), which is easier to handle since the objective
function is convex w.r.t. each variable when the other variables
are fixed. The optimal w (or ν) can be obtained based on (62)
when ν and f (w and f ) are fixed. For given w and ν, to
get the optimal f , we attach Lagrange multiplier λ to the first
constraint of (60) and obtain the Lagrange function as follows
!
K
K
X
X
2 (d)
L(f , λ) ,
(wk ek − ln wk ) + λ
fk ϕk − ζ . (64)
k=1
k=1
From [44], the first-order optimality condition of L w.r.t. fk
yields
q
(d)
p
wk νk φk
.
Pk ,
fk (λ) = min
K
P
(d)
(d)
(d)
2
2
wi νi ψki + λϕk
wk νk φk +
i=1
= ζ. Since for any k ∈ K,
R EFERENCES
(60c)
1
, ∀k ∈ K.
=
ek
k=1
(d)
fk2 ϕk
(60b)
To prove the equivalence between (42) and (60), we derive
the optimal ν and w by checking the first-order optimality
condition of problem (60)
q
(d)
fk φk
opt
νk =
, ∀k ∈ K,
K
P
(d)
(d)
(d)
fi2 ψik + σk
fk2 φk +
f
K
P
fk (λ) strictly decreases with λ, we can obtain λopt using the
bisection method. Then, substituting λopt into (65), we get fkopt .
i=1
max
have λopt > 0 and
(60a)
where wk is a positive weight variable, and ek is the meansquare estimation error
!
2
q
K
X
(d)
(d)
2 (d)
2
fi ψik + σk
. (61)
ek = νk fk φk − 1 + νk
K
X
k=1
k=1
p
0 ≤ fk ≤ Pk , ∀k ∈ K,
wkopt
According to the complementary slackness condition, if λopt =
K
P
(d)
fk2 ϕk < ζ, then, fkopt = fk (0). Otherwise, we
0 yields
(65)
[1] FCC Spectrum Policy Task Force, “Report of the spectrum efficiency
working group,” Nov. 2002.
[2] T. L. Marzetta, “Noncooperative cellular wireless with unlimited numbers of base station antennas,” IEEE Trans. Wireless Commun., vol. 9,
no. 11, pp. 3590–3600, Nov. 2010.
[3] F. Rusek, D. Persson, B. K. Lau, E. G. Larsson, T. L. Marzetta,
O. Edfors, and F. Tufvesson, “Scaling up MIMO: Opportunities and
challenges with very large arrays,” IEEE Signal Process. Mag., vol. 30,
no. 1, pp. 40–60, Jan. 2013.
[4] E. G. Larsson, O. Edfors, F. Tufvesson, and T. L. Marzetta, “Massive
MIMO for next generation wireless systems,” IEEE Commun. Mag.,
vol. 52, no. 2, pp. 186–195, Feb. 2014.
[5] H. Xie, B. Wang, F. Gao, and S. Jin, “A full-space spectrum-sharing
strategy for massive MIMO cognitive radio systems,” IEEE J. Sel. Areas
Commun., vol. 34, no. 10, pp. 2537–2549, Oct. 2016.
[6] J. Liu, N. Kato, J. Ma, and N. Kadowaki, “Device-to-device communication in LTE-advanced networks: A survey,” IEEE Commun. Surveys
Tuts., vol. 17, no. 4, pp. 1923–1940, Dec. 2015.
[7] A. Asadi, Q. Wang, and V. Mancuso, “A survey on device-to-device
communication in cellular networks,” IEEE Commun. Surveys Tuts.,
vol. 16, no. 4, pp. 1801–1819, Apr. 2014.
[8] X. Lin, J. Andrews, A. Ghosh, and R. Ratasuk, “An overview of 3GPP
device-to-device proximity services,” IEEE Commun. Mag., vol. 52,
no. 4, pp. 40–48, Mar. 2014.
[9] H. Q. Ngo, E. G. Larsson, and T. L. Marzetta, “Energy and spectral efficiency of very large multiuser MIMO systems,” IEEE Trans. Commun.,
vol. 61, no. 4, pp. 1436–1449, Apr. 2013.
[10] H. Yang and T. L. Marzetta, “Performance of conjugate and zeroforcing beamforming in large-scale antenna systems,” IEEE J. Sel. Areas
Commun., vol. 31, no. 2, pp. 172–179, Feb. 2013.
[11] J. Hoydis, S. Ten Brink, and M. Debbah, “Massive MIMO in the UL/DL
of cellular networks: How many antennas do we need?” IEEE J. Sel.
Areas Commun., vol. 31, no. 2, pp. 160–171, Feb. 2013.
[12] N. Liang, W. Zhang, and C. Shen, “An uplink interference analysis for
massive MIMO systems with MRC and ZF receivers,” in Proc. IEEE
Wireless Commun. Netw. Conf. (WCNC), 2015, pp. 310–315.
[13] T. Bai and R. W. Heath Jr, “Asymptotic coverage probability and rate
in massive MIMO networks,” Proc. IEEE GlobalSIP, pp. 1–5, 2014.
[14] D. Zhu, J. Wang, A. L. Swindlehurst, and C. Zhao, “Downlink resource reuse for device-to-device communications underlaying cellular
networks,” IEEE Sig. Process. Lett., vol. 21, no. 5, pp. 531–534, May
2014.
[15] S. Maghsudi and S. Stanczak, “Joint channel allocation and power
control for underlay D2D transmission,” in Proc. IEEE Int. Conf.
Commun. (ICC), London, UK., 2015, pp. 2091–2096.
[16] D. Feng, L. Lu, Y. Yuan-Wu, G. Y. Li, G. Feng, and S. Li, “Deviceto-device communications underlaying cellular networks,” IEEE Trans.
Commun., vol. 61, no. 8, pp. 3541–3551, Aug. 2013.
[17] W. Zhao and S. Wang, “Resource allocation for device-to-device communication underlaying cellular networks: An alternating optimization
method,” IEEE Commun. Lett., vol. 19, no. 8, pp. 1398–1401, Aug.
2015.
[18] Y. Jiang, Q. Liu, F. Zheng, X. Gao, and X. You, “Energy efficient joint
resource allocation and power control for D2D communications,” IEEE
Trans. Veh. Tech., vol. 65, no. 8, pp. 6119 – 6127, Aug. 2016.
[19] F. Wang, C. Xu, L. Song, and Z. Han, “Energy-efficient resource
allocation for device-to-device underlay communication,” IEEE Trans.
Wireless Commun., vol. 14, no. 4, pp. 2082–2092, Apr. 2015.
[20] T. D. Hoang, L. B. Le, and T. Le-Ngoc, “Energy-efficient resource
allocation for D2D communications in cellular networks,” in Proc. IEEE
Int. Conf. Commun., London, UK., 2015, pp. 2251–2256.
[21] X. Lin, R. W. Heath, and J. G. Andrews, “Spectral efficiency of massive
MIMO systems with D2D underlay,” in Proc. IEEE Int. Conf. Commun.
(ICC), London, UK., 2015, pp. 4345–4350.
15
[22] X. Lin, R. Heath, and J. Andrews, “The interplay between massive
MIMO and underlaid D2D networking,” IEEE Trans. Wireless Commun.,
vol. 14, no. 6, pp. 3337–3351, June 2015.
[23] J. Jose, A. Ashikhmin, T. L. Marzetta, and S. Vishwanath, “Pilot
contamination and precoding in multi-cell TDD systems,” IEEE Trans.
Wireless Commun., vol. 10, no. 8, pp. 2640–2651, Aug. 2011.
[24] H. Yin, D. Gesbert, M. Filippou, and Y. Liu, “A coordinated approach
to channel estimation in large-scale multiple-antenna systems,” IEEE J.
Sel. Areas Commun., vol. 31, no. 2, pp. 264–273, Feb. 2013.
[25] A. Ashikhmin and T. Marzetta, “Pilot contamination precoding in multicell large scale antenna systems,” in Proc. IEEE ISIT, Cambridge, MA,
USA, 2012, pp. 1137–1141.
[26] F. Fernandes, A. Ashikhmin, and T. L. Marzetta, “Inter-cell interference
in noncooperative TDD large scale antenna systems,” IEEE J. Sel. Areas
Commun., vol. 31, no. 2, pp. 192–201, Feb. 2013.
[27] H. Q. Ngo and E. G. Larsson, “Evd-based channel estimation in multicell
multiuser MIMO systems with very large antenna arrays,” in Proc. IEEE
ICASSP, Kyoto, Japan, 2012, pp. 3249–3252.
[28] L. You, X. Gao, X.-G. Xia, N. Ma, and Y. Peng, “Pilot reuse for massive
MIMO transmission over spatially correlated rayleigh fading channels,”
IEEE Trans. Wireless Commun., vol. 14, no. 6, pp. 3352–3366, Feb.
2015.
[29] X. Liu, Q. He, Y. Li, L. Xiao, and J. Wang, “Pilot reuse for deviceto-device underlay massive MIMO systems,” in IEEE VTC-Fall, Sep.
2015, pp. 1–5.
[30] X. Liu, Y. Li, X. Li, L. Xiao, and J. Wang, “Pilot reuse and interferenceaided MMSE detection for D2D underlay massive MIMO,” IEEE Trans.
Veh. Tech., vol. 66, no. 4, pp. 3116 – 3130, Apr. 2017.
[31] H. Xu, N. Huang, Z. Yang, J. Shi, B. Wu, and M. Chen, “Pilot allocation
and power control in D2D underlay massive MIMO systems,” IEEE
Commun. Lett., vol. 21, no. 1, pp. 112–115, Jan. 2017.
[32] N. Jindal, J. G. Andrews, and S. Weber, “Multi-antenna communication
in ad hoc networks:achieving MIMO gains with SIMO transmission,”
IEEE Trans. Commun., vol. 59, no. 2, pp. 529–540, Feb. 2011.
[33] X. Zhu, L. Dai, and Z. Wang, “Graph coloring based pilot allocation
to mitigate pilot contamination for multi-cell massive MIMO systems,”
IEEE Commun. Lett., vol. 19, no. 10, pp. 1842–1845, Aug. 2015.
[34] M. Li, S. Jin, and X. Gao, “Spatial orthogonality-based pilot reuse for
multi-cell massive MIMO transmission,” in Proc. IEEE Int. Conf. WCSP,
Oct. 2013, pp. 1–6.
[35] R. Mochaourab, E. Björnson, and M. Bengtsson, “Adaptive pilot clustering in heterogeneous massive mimo networks,” IEEE Trans. Wireless
Commun., vol. 15, no. 8, pp. 5555–5568, Aug. 2016.
[36] T. Kailath, A. H. Sayed, and B. Hassibi, Linear Estimation. Prentice
Hall Upper Saddle River, NJ, 2000, vol. 1.
[37] R. Chen, J. G. Andrews, R. W. Heath, and A. Ghosh, “Uplink power
control in multi-cell spatial multiplexing wireless systems,” IEEE Trans.
Wireless Commun., vol. 6, no. 7, pp. 2700–2711, July 2007.
[38] R. D. Yates, “A framework for uplink power control in cellular radio
systems,” IEEE J. Sel. Areas Commun., vol. 13, no. 7, pp. 1341–1347,
Sep. 1995.
[39] Q. Shi, M. Razaviyayn, Z.-Q. Luo, and C. He, “An iteratively weighted
MMSE approach to distributed sum-utility maximization for a MIMO
interfering broadcast channel,” IEEE Trans. Signal Process., vol. 59,
no. 9, pp. 4331–4340, Sep. 2011.
[40] H. Xu, Z. Yang, N. Huang, J.-Y. Wang, J. Shi, and M. Chen, “Channel
allocation and power control in D2D uplink underlaid cellular networks,”
in Proc. IEEE Globecom Workshops (GC Wkshps), Washington, DC,
2016, pp. 1–6.
[41] A. M. Tulino and S. Verdú, Random Matrix Theory and Wireless
Communications. Now Publishers Inc, 2004, vol. 1.
[42] Y. Jong, “An efficient global optimization algorithm for nonlinear sumof-ratios problem,” online: www. optimizationonline. org, 2012.
[43] L. Xu, G. Yu, and Y. Jiang, “Energy-efficient resource allocation in
single-cell OFDM systems: Multi-objective approach,” IEEE Trans.
Wireless Commun., vol. 14, no. 10, pp. 5848–5858, Oct. 2015.
[44] S. Boyd and L. Vandenberghe, Convex Optimization.
Cambridge
University Press, 2004.
| 7cs.IT
|
A note on the penalty parameter in Nitsche’s method for unfitted boundary
value problems
Frits de Prentera,∗, Christoph Lehrenfeldb , André Massingc
a Department
of Mechanical Engineering, Eindhoven University of Technology, The Netherlands
for Numerical and Applied Mathematics, University of Göttingen, Germany
c Department of Mathematics and Mathematical Statistics, Umeå University, Sweden
arXiv:1709.05832v1 [math.NA] 18 Sep 2017
b Institute
Abstract
Nitsche’s method is a popular approach to implement Dirichlet-type boundary conditions in situations where
a strong imposition is either inconvenient or simply not feasible. The method is widely applied in the context
of unfitted finite element methods. From the classical (symmetric) Nitsche’s method it is well-known that
the stabilization parameter in the method has to be chosen sufficiently large to obtain unique solvability
of discrete systems. In this short note we discuss an often used strategy to set the stabilization parameter
and describe a possible problem that can arise from this. We show that in specific situations error bounds
can deteriorate and give examples of computations where Nitsche’s method yields large and even diverging
discretization errors.
Keywords: Nitsche’s method, unfitted/immersed finite element methods, penalty/stabilization parameter,
accuracy, stability, error analysis
1. Introduction
We consider the discretization of an unfitted Poisson problem. The problem domain is described separately from the encapsulating mesh, on which a finite element basis is defined. We consider the restriction of this finite element space with respect to the problem domain. Such an approach is used in many
methods which are similar in virtue, e.g., the fictitious domain method [1], the cut finite element method
(CutFEM) [2–5], the finite cell method (FCM) [6–12], immersogeometric analysis [13–15], the immersed
boundary method [16], the unfitted discontinuous Galerkin method (UDG) [17, 18], the extended finite element method (XFEM) [19–24], and several others. To impose essential boundary conditions, many of these
methods apply some version of the classical Nitsche’s method [25].
This requires stabilization by a penalization parameter to preserve the coercivity of the bilinear operator.
Without additional stabilization of cut elements with e.g., ghost penalty terms [2, 26] – which is customary
for methods referred to as CutFEM – this penalty parameter depends on the shape and size of the cut
elements, i.e., it depends on the position of the geometry relative to the computational mesh. A typical
choice which is sufficient to provide unique solvability of discrete problems, is to choose the stabilization
parameter as an element wise constant and proportional to the ratio between the surface measure of the
intersection between an element and the boundary and the volume measure of the intersection of the same
element and the domain. As this ratio can become arbitrarily large, the stabilization parameter is not
generally bounded.
In the mathematical literature, e.g., [2, 3], unfitted Nitsche formulations are generally supplemented by
the aforementioned additional stabilization to bound the stabilization parameter in order to prove properties
∗ Corresponding
author. Tel.: +31 61 516 2599
Email addresses: [email protected] (Frits de Prenter), [email protected] (Christoph Lehrenfeld),
[email protected] (André Massing)
Preprint submitted to Computers & Mathematics with Applications
September 19, 2017
of the method. In our opinion, the importance of this has not sufficiently been addressed in the literature
however. Moreover, the method has effectively (and successfully) been applied without additional stabilization in the engineering literature, see e.g., [10–12]. With this note, we therefore detailedly treat the possible
problem that can occur when no additional measures are taken to bound the stabilization parameter and aim
to provide a disclaimer for directly applying the classical form of Nitsche’s method to immersed problems.
To this end, we discuss and analyze the method and demonstrate that it can lead to poor results in the
discretization error when the geometry intersects the computational domain such that large values of the
stabilization parameter are required. We also investigate different situations where degenerated geometry
configurations yield satisfactory and unsatisfactory results and provide a possible interpretation. We further
review alternative formulations and possible modifications and stabilizations of the method.
Section 2 of this note introduces a model problem, followed by the Nitsche’s method under consideration
in Section 3. In Section 4 we carry out a simple error analysis of the method in a norm which is natural to the
formulation. This analysis reveals the possible large discretization errors for unfortunate cut configurations.
In Section 5 we give examples of bad cut configurations and show how these can lead to large and even
degenerating discretization errors. Finally, we list alternative approaches to impose boundary conditions
and variants of the classical Nitsche’s method to circumvent this possible problem in Section 6.
2. The model problem
As a model for more general elliptic boundary value problems, we consider the Poisson problem with
inhomogeneous Dirichlet boundary data posed on an open and bounded domain Ω ⊂ Rd with Lipschitz
boundary ∂Ω:
−∆u = f
u= g
in Ω,
(1a)
on ∂Ω.
(1b)
We assume that the properties of the domain Ω imply that a shift theorem of the form:
kukH 2 (Ω) . kf kL2 (Ω) + kgk
1
H 2 (∂Ω)
,
(2)
1
holds whenever f ∈ L2 (Ω) and g ∈ H 2 (∂Ω), e.g., when ∂Ω is smooth or Ω ⊂ R2 is a convex domain
with piecewise C 2 boundary [27]. In this note . is used to denote an inequality with a constant that is
independent of the mesh. This problem has the standard well-posed weak formulation: Find u ∈ Hg1 (Ω),
such that:
Z
Z
∇u∇v dx =
f v dx for all v ∈ H01 (Ω).
(3)
Ω
H01 (Ω)
Ω
Hg1 (Ω)
are the standard Sobolev spaces with corresponding homogeneous
In this formulation
and
and inhomogeneous boundary data. For simplicity, we assume that Dirichlet boundary conditions are posed
on the whole boundary of Ω. However, the results presented in this work extend to the case where Dirichlet
boundary data is only prescribed on a part of the boundary as demonstrated in Section 5.
3. A Nitsche-based unfitted finite element method
We consider a (triangular or rectilinear) shape regular background discretization Te which encapsulates
the domain Ω. To each element T ∈ Te we associate the local mesh size hT = diam(T ) and define the global
mesh size by h = maxT ∈Te hT . We define the active mesh T by:
T = {T ∈ Te : T ∩ Ω 6= ∅},
(4)
e =S
and set Ω
T ∈T T , denoting the union of all elements from the active mesh. The set of geometric entities
are illustrated in Figure 1. On the active mesh T we define the finite dimensional function space:
2
Figure 1: Example of a domain Ω with background mesh Te and active mesh T .
e : v|T ∈ Pk (T ), T ∈ T },
Vh = {v ∈ C(Ω)
(5)
consisting of continuous piecewise polynomials of order k. A simple prototype formulation for a Nitschebased unfitted finite element method is to find uh ∈ Vh , such that:
ah (uh , vh ) = lh (vh )
∀ vh ∈ Vh ,
(6)
where:
Z
ah (uh , vh ) :=
ZΩ
lh (vh ) :=
Z
Z
∇uh ∇vh dx +
(−∂n uh )vh ds +
∂Ω
Z
f vh dx +
(−∂n vh + λT vh )g ds.
Ω
Z
(−∂n vh )uh ds +
∂Ω
λT uh vh ds,
(7)
∂Ω
(8)
∂Ω
Here, ∂n denotes the partial derivative in normal direction of the boundary of Ω and λT denotes a elementwise stabilization parameter. The choice of the stabilization parameter is crucial. As the review of the error
analysis in Section 3 reveals, the local stabilization parameter λT has to be chosen sufficiently large in order
to guarantee the discrete coercivity of the bilinear form ah (·, ·). A common – see e.g., [10–12] – strategy is
to solve the eigenvalue problem:
Z
∂Ω∩T
Find (uµ , µT ) ∈ Vh |0T × R, such that:
Z
(∂n uµ )(∂n vh ) ds = µT
∇uµ ∇vh ds for all vh ∈ Vh |0T ,
(9)
Ω∩T
on each element T which has a non-empty intersection with ∂Ω, see for instance [28]. Here, Vh |0T is the
space of functions in Vh |T that are L2 (T )-orthogonal to constants. Let NT = dim(Vh |T ), then (9) is an
(NT − 1)-dimensional eigenvalue problem with only non-negative eigenvalues. Motivated by the forthcoming
stability analysis, a suitable choice of the stabilization parameter is to set:
R
(∂n uh )2 ds
> 0.
(10)
λT = 2 max{µT } = 2 max 0 R∂Ω∩T
|∇uh |2 dx
uh ∈Vh |T
Ω∩T
3
ε
|T ∩ Ω| ε→0
−→ 0
|T |
ε
|T ∩ Ω| ε→0
−→ 0
|T |
|T ∩ ∂Ω| ε→0
−→ 0
|∂T |
|T ∩ ∂Ω| ε→0
−→ c > 0
|∂T |
(a)
(b)
Figure 2: Different cut configurations. In (a) the part of the boundary in the element vanishes for vanishing volume, while for
the sliver case (b) the measure of the boundary part stays bounded from below as the volume goes to zero.
Remark 1. With techniques as in [29], it can be shown that λT scales with |T ∩ ∂Ω|d−1 /|T ∩ Ω|d . Thus,
shape-regularity implies that for a Nitsche-based formulation on a fitted and shape-regular mesh λT ∼ |T ∩
∂Ω|d−1 /|T ∩ Ω|d ∼ h−1 . In contrast, in an unfitted method, the ratio |T ∩ ∂Ω|d−1 /|T ∩ Ω|d depends not only
on the cut element size but also highly on the cut configuration (see Figure 2). In particular, a sliver cut can
lead to arbitrarily large values of λT on relatively large parts of the boundary, depending on the thickness of
the sliver. In Section 5 it is demonstrated how cut elements of approximately this shape can result in poor
discretization errors.
Remark 2. To generate and solve the discrete linear system associated with problem (6) several issues have
to be resolved, i.e., sufficiently accurate numerical integration on T ∩ Ω and T ∩ ∂Ω has to be provided, e.g.,
[30–39], and a particular tuning of the linear algebra solver might be necessary, e.g., [40, 41]. All these
aspects are important, but in this work we focus only of the choice of λT and its effect on the numerical
solution, assuming that numerical integration can be performed sufficiently accurate and applying a direct
solver to solve the linear systems exactly.
4. A simple error analysis
In this section we perform an error analysis for the variational form in (6). We first define a meshdependent norm in which a best approximation property holds in Section 4.1. In Section 4.2 we discuss
the approximability in the mesh-dependent norm and the consequences for the discretization error in the
H 1 (Ω)-norm.
4.1. Best approximation in the mesh-dependent energy norm
With the choice of the stabilization parameter as in (10), the stability analysis is natural in the following
mesh-dependent norm which we will refer to as the energy norm:
1
−1
|||v|||2 := k∇vk2Ω + kλT 2 ∂n vk2∂Ω + kλT2 vk2∂Ω ,
v ∈ Vh ⊕ H 2 (Ω).
(11)
We note that the use of this norm in the error analysis is essentially dictated by Nitsche’s method and not
by the model problem. We obtain coercivity and continuity in the following two lemmas.
Lemma 3 (Coercivity). There holds:
ah (u, u) ≥
1
|||u|||2 for all u ∈ Vh .
5
4
(12)
Proof. Exploiting the eigenvalue characterization in the definition of λT we obtain for arbitrary γ > 0:
Z
1
1
−1
λT 2 ∂n u λT2 u ds
ah (u, u) ≥ k∇uk2Ω + kλT2 uk2∂Ω − 2
∂Ω
1 − 12
kλ ∂n uk2∂Ω
γ T
1
1
−1
≥ (1 − γ)k∇uk2Ω + (1 − γ)kλT2 uk2∂Ω + (2γ − )kλT 2 ∂n uk2∂Ω ≥ c|||u|||2 ,
γ
≥
k∇uk2Ω
+ (1 −
1
2
γ)kλT uk2∂Ω
with c = 1 − γ ∗ for γ ∗ > 0 such that 1 − γ ∗ = 2γ ∗ −
that c > 15 .
−
1
γ∗ ,
(13)
which has only the positive solution γ ∗ =
√
1+ 13
6
so
Lemma 4 (Continuity). There holds:
ah (u, v) ≤ 2|||u||||||v||| for all u, v ∈ Vh ⊕ H 2 (Ω).
(14)
Proof. With Cauchy-Schwarz we directly obtain:
−1
1
−1
1
1
1
ah (u, v) ≤ k∇ukΩ k∇vkΩ +kλT 2∂n uk∂Ω kλT2 vk∂Ω +kλT 2∂n vk∂Ω kλT2 uk∂Ω +kλT2 uk∂Ω kλT2 vk∂Ω
1
1
−1
−1
≤ k∇ukΩ k∇vkΩ + kλT 2∂n uk∂Ω + kλT2 uk∂Ω kλT 2∂n vk∂Ω + kλT2 vk∂Ω
1
1
2 12
2 12
−1
−1
k∇vk2Ω + kλT 2∂n vk∂Ω + kλT2 vk∂Ω
≤ k∇uk2Ω + kλT 2∂n uk∂Ω + kλT2 uk∂Ω
12
21
1
1
−1
−1
≤ k∇uk2Ω + 2kλT 2∂n uk2∂Ω + 2kλT2 uk2∂Ω
k∇vk2Ω + 2kλT 2∂n vk2∂Ω + 2kλT2 vk2∂Ω .
(15)
Applying these lemmas and exploiting that the Nitsche formulation is consistent in the sense of Galerkin
orthogonality, we obtain Céa’s Lemma [42]:
Lemma 5 (Céa’s Lemma). Let u ∈ H 2 (Ω) be the solution to (1) and uh ∈ Vh be the solution to (6). Then
there holds:
|||u − uh ||| ≤ 11 inf |||u − vh |||.
(16)
vh ∈Vh
Proof. For an arbitrary vh ∈ Vh there holds |||u − uh ||| ≤ |||u − vh ||| + |||vh − uh ||| and:
|||vh − uh |||2 ≤ 5ah (vh − uh , vh − uh ) = 5ah (vh − u, vh − uh ) ≤ 10|||vh − uh ||||||vh − u|||,
(17)
s.t. the claim directly follows.
We note that in the proof we could not use coercivity for u − uh , but only for discrete functions. To
this end we applied the triangle inequality and were able to apply the coercivity estimate on the discrete
function vh − uh , which was then combined with Galerkin orthogonality and continuity in the usual way.
This version of Céa’s lemma suggests that Nitsche’s method provides a robust and quasi-optimal method.
The best approximation is, however, only provided in the mesh-dependent norm ||| · |||. In the next paragraph
we discuss the discretization error of the best approximation in this norm.
4.2. Approximability of the solution in the mesh-dependent energy norm
We consider the approximability of the solution u with the finite element space Vh in the energy norm, i.e.,
inf vh ∈Vh |||u − vh |||, which is the r.h.s. in (16). The aim of this paragraph is to show that the approximability
in the discrete energy norm is not robust with respect to the position of the domain boundary. To this end,
5
we split the energy norm, cf. (11), into two parts with one part that is robust and one part which is not
robust:
o
n
1
−1
inf |||u − vh |||2 ≥ inf k∇(u − vh )k2Ω + kλT 2 ∂n (u − vh )k2∂Ω + inf kλT2 (u − vh )k2∂Ω .
(18)
vh ∈Vh
vh ∈Vh
vh ∈Vh
k+1
We assume that the solution u is smooth, u ∈ H
(Ω), with k ≥ 1 denoting the order of the discretization.
In the sequel we use the notation & for an inequality with a constant that is independent of the cut
configuration. For the first part of the r.h.s. in (18) we take the idea from [21] to show that the solution can
be optimally approximated in Vh with respect to these parts of the norm independent of the cut position.
Afterwards, we show that the remaining part is the crucial problem which results in the missing robustness.
e be a continuous extension of u to Ω
e (the embedding domain, i.e., the union of all
Let ũ ∈ H k+1 (Ω)
active elements), then we have:
k∇(u − vh )kΩ = k∇(ũ − vh )kΩ . k∇(ũ − vh )kΩ
e,
∀ vh ∈ Vh .
To bound the latter part we can apply a standard best approximation result for Vh on the fitted mesh
Te . This estimate is robust in the position of the boundary, i.e., there is a vh ∈ Vh s.t. k∇(u − vh )kΩ .
−1
k
hk kũkH k+1 (Ω
e) . h kukH k+1 (Ω) . Further, due to λT . hT and a trace inequality, we have:
−1
kλT 2 ∂n vk∂Ω∩T . kvkH 1 (Ω∩T ) + hT kvkH 2 (Ω∩T ) ,
∀ v ∈ H 2 (Ω ∩ T ).
−1
Applying this to u−vh gives similar results (independent of the position of the boundary), hence kλT 2 ∂n (u−
vh )k∂ Ω . hk kukH k+1 (Ω) .
We now consider the second term in (18) which is the crucial problem for the approximation of u in the
energy norm:
X
1
kλT2 (u − vh ) k2∂Ω =
λT ku − vh k2T ∩∂Ω .
T ∈T
Let hΩ∩T := diam(Ω ∩ T ) and h∂Ω∩T := diam(∂Ω ∩ T ) be the characteristic lengths of the cut elements and
their parts on the boundary. Furthermore, let ṽh be the best approximation of u in Vh in the energy norm.
2k+2+(d−1)
We have that ku − ṽh k2T ∩∂Ω is typically not zero and scales with h∂Ω∩T
, d − 1 denoting the dimension of
the boundary. In the good case, if the cut element T ∩ Ω is shape regular, there holds h∂Ω∩T ∼ hΩ∩T ∼ λ−1
T
2k+1+(d−1)
such that we would have λT ku − ṽh k2T ∩∂Ω ≤ ch∂Ω∩T
≤ ch2k+d
for some (bounded) constant c > 0
T
which is a robust bound for the element contribution of T . In the bad case, if the cut element is not shape
regular however, such as for a sliver cut element in Figure 2, it can occur that λT h−1
∂Ω∩T , such that the
contribution of a single element λT ku − ṽh k2T ∩∂Ω becomes unbounded.
We conclude that the last term in (18) is unbounded and thus the same holds for the left hand side.
Therefore the discretization error in the mesh-dependent energy norm, regardless of the best approximation
property, can not be bounded from above. This missing robustness is essentially due to the possible existence
of unfortunate cut configurations such as sliver cuts, which introduce values of the penalization parameter,
λT , that are not bounded in terms of the reciprocal of the length of the boundary intersecting element T ,
h∂Ω∩T .
The energy norm is very natural with respect to the (symmetric) Nitsche formulation. Nevertheless,
one is typically not interested in this norm but rather in the H 1 (Ω)-norm of the error. To obtain a priori
estimates for the error in the H 1 (Ω)-norm we use a Poincaré-type inequality and combine it with the result
in Céa’s lemma:
ku − uh kH 1 (Ω) . |||u − uh ||| . inf |||u − vh |||.
(19)
vh ∈Vh
Whilst the approximability in the energy norm is not robust with respect to the boundary position, these
inequalities are independent of the boundary position† . Therefore we can not give an optimal a priori error
estimate for the H 1 (Ω)-norm.
1
† By
e.g., Lemma B.63 in [43] we have |||u|||2 ≥ k∇uk2Ω + kλT2 uk2∂Ω ≥ k∇uk2Ω + λ̃kuk2∂Ω ≥ δ 2 kvk2H 1 (Ω) , with λ̃ ∼ 1/h the
lower bound for λT .
6
Besides showing that the traditional method to show optimality does not hold for Nitsche’s method, we
can also give an intuitive derivation of why Nitche’s method can yield bad approximations in the H 1 (Ω)norm. To do this, we compare the energy and Sobolev norms. At first sight these norms are not even
1
equivalent, as kλ− 2 ∂n vk∂Ω is not bounded by the H 1 (Ω)-norm. This holds for all (even unsymmetric or
fitted) Nitsche-type formulations however, and is not related to the problem of unbounded stabilization
parameters we are targeting here. When we restrict ourselves to the space Wh = Vh ⊕ hui, both norms are
equivalent:
ckvh kH 1 (Ω) ≤ |||vh ||| ≤ Ckvh kH 1 (Ω) , ∀vh ∈ Wh .
(20)
To investigate the strength of this equivalence, we examine the constants c and C in (20). The first constant
0 < δ † < c < 1 is bounded from below and above independent of u or the cut configuration. The upper
bound of c follows from:
|||vh |||
|vh |H 1
|||vh |||
. min
= min
< 1,
(21)
c = min
⊥
⊥
vh ∈Wh kvh kH 1 (Ω)
vh ∈ Vh |∂Ω kvh kH 1 (Ω)
vh ∈ Vh |∂Ω kvh kH 1 (Ω)
⊥
with Vh |∂Ω the restriction of Vh to functions that do not intersect the boundary, i.e.,
⊥
Vh |∂Ω = {vh ∈ Vh |vh = ∂n vh = 0 on ∂Ω} .
The second constant in (20), C, does depend on the cut configuration:
C = max
vh ∈Wh
|||vh |||
|||vh |||
≥ max
.
vh ∈Vh kvh kH 1 (Ω)
kvh kH 1 (Ω)
(22)
Due to the finite dimensionality of Wh we have C < ∞, but, in contrast to c, C clearly depends on λ and
−1
therefore on the cut configuration. The inequality in (22) disregards the possibility that kλT 2 vh k∂Ω
kvh kH 1 (Ω) for vh ∈ Wh due to possibly large normal derivatives of solution u, i.e., the general limitation
for Nitsche-type methods we mentioned before. But also when this is not the case, C will be large for large
values of λ however. Therefore, depending on the cut configuration, the equivalence between the energy
norm and the H 1 (Ω)-norm can degenerate, such that the projections in the different norms can deviate from
each other. With the solution to Nitsche’s method closely resembling the projection in the energy norm,
this may deviate substantially from an optimal approximation in H 1 (Ω).
Remark 6. Using c and C we can also extend (19):
kuh − vh kH 1 (Ω) ≤
1
1
C
|||uh − vh ||| ≤ 10 |||u − vh ||| ≤ 10 ku − vh kH 1 (Ω) ,
c
c
c
∀vh ∈ Vh ,
(23)
such that:
C
10 + 1 inf ku − vh kH 1 (Ω) .
vh ∈Vh
c
This seems like a satisfying result, but note that C/c 1 when λ 1.
ku − uh kH 1 (Ω) ≤
(24)
5. Numerical examples
In this section we will demonstrate some examples of situations where the previously discussed discretization leads to large values for the stabilization parameter. The aim of this section is to illustrate the problem
described in Section 4 and to show that it is not just a problem in the a priori analysis, but really yields
large discretization errors in practice.
The numerical integration on elements that only partially intersect the domain is performed using the
bisection-based tessellation scheme proposed in [44] with a maximal refinement depth of two. This leads
to an exact representation of the domain for the first three examples and a reasonable good geometric
approximation for the other examples. To make sure the results of these last examples are not affected by
geometrical errors, we adapt the boundary conditions to the approximated geometry. On the approximated
geometry we apply Gauss quadrature such that functions of order 2k + 1 are integrated exactly, in order to
achieve an accurate integration of the force terms and the boundary conditions.
7
5.1. Example 1: An overlapping square on a triangular and quadrilateral mesh
The simplest example of a discretization with sliver cuts as described in the Section 3 is the squared
domain Ω = (−1 − ε, 1 + ε)2 with 0 < ε 1 on a structured grid that aligns with Ω for ε = 0. The domain
and the grid are shown in Figure 3a (for the triangular mesh), in which it is clearly visible that the cut
elements become of arbitrarily large aspect ratios for ε → 0.
Triangular grid
We use a triangular mesh of isosceles right triangles with legs of length h = 1/K with K = 16 and
continuous piecewise linear basis functions. We apply an element local stabilization parameter by the
procedure described in Section 3. We set up the problem such that the analytical solution equals u =
sin(πx) + sin(πy) and impose Dirichlet conditions through the previously discussed Nitsche method on all
boundaries.
|||u − uh |||
ku − uh kH 1 (Ω)
ku − uh kL2 (Ω)
104
102
100
10−2
ε
10−10
ε
(a) Mesh and domain Ω.
10−8
10−6
ε
10−4
10−2
(b) Discretization errors.
Figure 3: Example 1 with a triangular grid.
The resulting discretization errors in the mesh-dependent energy norm, the H 1 (Ω)-norm and the L2 (Ω)norm are given in Figure 3b. From the results it is clearly visible that for ε → 0 the discretization error
degenerates in both the energy as the H 1 (Ω)-norm.
Quadrilateral grid
We repeat the example on a quadrilateral grid with the same nodal positions as the triangular grid, i.e.,
the mesh size h = 1/K with K = 16, see Figure 4a. Piecewise bilinear elements are used which lead to the
same number of degrees of freedom. The results are shown in Figure 4b, from which we can observe the
same tendencies for the convergence ε → 0.
We observe that the impact of a small ε is less severe on the quadrilateral grid compared to the triangular
grid. The error is several orders of magnitude smaller and the diverging effect only starts at a smaller value
of ε. This difference is explained in the next paragraph, which is not a rigorous proof, but gives an intuitive
illustration of the nature of the difference. It further serves as a motivation for the design of Examples 2
and 3.
Discussion of the difference between triangular and quadrilateral grids
In Section 4.1 we have seen that the numerical solution closely approximates the energy norm projection
of the analytical solution on the finite dimensional function space. For large values of λT , the contribution
of normal derivatives to the energy norm diminishes and we can consider the energy norm to be a weighted
combination of an L2 (∂Ω)-norm on the boundary and an H01 (Ω)-seminorm in the interior. For increasing
8
|||u − uh |||
ku − uh kH 1 (Ω)
ku − uh kL2 (Ω)
104
102
100
10−2
ε
10−10
ε
(a) Mesh and domain Ω.
10−8
10−6
ε
10−4
10−2
(b) Discretization errors.
Figure 4: Example 1 with a quadrilateral grid.
λT ∝ ε−1 , the balance between these weighted norms is lost, and the energy norm projection basically
consists of an L2 (∂Ω)-norm projection for all d.o.f.’s supported on the boundary and an H01 (Ω)-seminorm
projection in the interior for the remaining d.o.f.’s. Hence, in the limit of ε → 0, the numerical solution uh
satisfies:
1
ε→0
uh |∂Ω −→ u∗h := argmin kλT2 (vh − g)k2L2 (∂Ω) .
(25)
vh ∈Vh |∂Ω
Hence M := dim(Vh |∂Ω ) d.o.f.’s are fixed independently of the remaining part of the variational formulation.
Note that M = 16K +7 for the triangular grid (corresponding to the number of dots minus one in Figure 5a)
and that M = 8K + 8 for the quadrilateral grid (corresponding directly to the number of dots in Figure 5b).
Besides constraining the boundary values, the L2 (∂Ω)-norm projection on the boundary can also affect the
normal derivatives or even the complete numerical solution in the cut elements however. Let N denote
the number of basis functions with support on ∂Ω (squares in Figure 5), then we have N = 16K + 8 for
both meshes. Note that N is also equal to the number of basis functions supported on cut elements. For
the triangular grid we have M = N − 1, which implies that in the limit of ε → 0 only one d.o.f. in all
cut elements together is not constrained by the boundary condition. For the quadrilateral mesh we have
M = 1/2N + 4, such that only approximately half of the d.o.f.’s supported on the boundary is constrained
by the boundary condition. Along the straight parts of the boundary, the normal derivatives can therefore
be set by the H01 (Ω)-seminorm projection in the interior independent of the boundary condition. Only in
the corners the normal derivatives are constrained by the boundary condition. As the problematic area is
much smaller compared to the case with the triangular mesh, the impact of large stabilization parameters
is – although asymptotically the same – less severe.
(a) Triangular grid.
(b) Quadrilateral grid.
Figure 5: Sketch of d.o.f.’s of Vh (squares) and d.o.f.’s of Vh |∂Ω (dots).
From Figures 3b and 4b it is clear that small values of ε do not only yield bad approximations, but
actually cause diverging discretization errors. To explain this effect, we note that the boundary data g has
9
magnitudes ∼ 1 whilst the basis functions that are only supported on cut elements have values of magnitude
∼ ε on the boundary. The L2 (∂Ω)-norm projection of the boundary condition therefore constrains the
coefficients of these basis functions to a value of magnitude ∼ ε−1 . This results in gradients in the cut
elements that scale with ε−1 , which has also been observed by peak stresses in immersed elasticity problems
1
[45, 46]. These large gradients introduce an O(ε− 2 ) error in the H 1 (Ω)-norm, i.e., the sum of the volumes of
the cut elements has a magnitude of ∼ ε, whilst the error in the gradient squared has a value of magnitude
∼ ε−2 . We notice that the L2 (Ω)-norm error seems to be unaffected by the sliver cuts.
5.2. Example 2: An overlapping square with mixed boundary conditions
In this example we reuse the quadrilateral mesh and try to remove the problematic corner regions. To
this end, we pose Neumann conditions on the left and right boundaries, ∂ΩN = {|x| = 1 + ε}, and Dirichlet
conditions on the lower and upper boundaries, ∂ΩD = {|y| = 1 + ε}. Our hypothesis is that this will
yield good approximation results in the H 1 (Ω)-norm for the quadrilateral grid, because in that case an
L2 (∂ΩD )-norm projection on the Dirichlet boundary will constrain M = 4K + 6 d.o.f.’s whilst the number of
basis functions supported on the Dirichlet boundary equals N = 2M . On all elements, including the corner
elements, there is exactly one constraint for every two degrees of freedom. This allows to approximate the
boundary values without affecting the normal derivatives.
The results are visible in Figure 6 and indeed show that in this situation the error in the H 1 (Ω)-norm
1
is robust with respect to ε, even though the error in the mesh-dependent energy norm still scales with ε− 2 .
This is because also the best possible approximation of the boundary conditions has an error, such that the
energy norm error increases for increasing
λ. Furthermore, we notice that the errors in the energy norm are
√
reduced by approximately a factor 2, which is accounted for by the reduction of the size of the Dirichlet
boundary by a factor 2.
|||u − uh |||
ku − uh kH 1 (Ω)
ku − uh kL2 (Ω)
102
100
10−2
10−10
10−8
10−6
ε
10−4
10−2
Figure 6: Discretization errors for Example 2.
5.3. Example 3: A tilted square with mixed boundary conditions
In this example we demonstrate that Example 2, be it useful for academic purposes, does not represent
a general case for straight boundaries and tensor product grids. The reason that the second example yields
robust approximation errors is that the Dirichlet boundary is parallel to a grid line, such that N = 2M ,
i.e., the number of d.o.f.’s constrained by the L2 (∂ΩD )-norm projection is exactly right for linear bases.
This is not generally the case for immersed methods however. In this third example we do not consider the
square with boundaries ∂ΩN = {|x| = 1 + ε} and ∂ΩD = {|y| = 1 + ε}, but consider a kinked domain with
boundaries ∂ΩN = {|x − εy| = 1} and ∂ΩD = {|y − εx| = 1} as can be seen in Figure 7a. The (quadrilateral)
grid and the analytical solution are the same as in the previous examples.
In this configuration we have M = 8K + 6 constraints in the L2 (∂ΩD )-norm projection on the Dirichlet
boundary and N = 8K +10 basis functions that are supported on the Dirichlet boundary. Even though there
is a kernel of dimension 4 in the projection of the boundary conditions on the nodes that are supported on
the Dirichlet boundary, basis functions with values of magnitude ∼ on the boundary will be constrained
10
103
|||u − uh |||
ku − uh kH 1 (Ω)
ku − uh kL2 (Ω)
2
10
101
100
10−1
ε
1−ε
10−2
ε
1+ε
ε
1−ε
ε
1+ε
10−10
ε
1+ε
10−8
(a) Mesh and domain Ω.
10−6
ε
10−4
10−2
(b) Discretization errors.
Figure 7: Example 3.
to nodal values of magnitude ∼ ε−1 . The results in Figure 7b indeed show similar behavior to the first
example.
5.4. Example 4: A rounded square (k · k8 -norm unit ball)
The previous examples consisted of domains with piecewise straight boundaries. Because unfitted methods are generally used for complex domains, this example involves a curved boundary that creates sliver-like
cuts which can occur in many realistic applications. The boundary of the domain is defined by the function
x8 + y 8 = (1 + ε)8 and Dirichlet conditions are imposed on the complete boundary. A sketch of the domain
is visible in Figure 8a.
|||u − uh |||
ku − uh kH 1 (Ω)
ku − uh kL2 (Ω)
102
101
100
10−1
10−2
ε
10−10
ε
10−8
10−6
ε
10−4
10−2
(b) Discretization errors.
(a) Mesh and domain Ω.
Figure 8: Example 4.
The results are visible in Figure 8b. From the results we note that for this example not even the error in
the L2 (Ω)-norm is robust with respect to the cut configuration. Furthermore, a peculiar effect is observed
as the errors don’t simply increase for decreasing ε, but peak at certain values. This occurs because a
smaller value of ε doesn’t simply imply thinner cut elements for this test case. The zones of the domain
that constitute the thin (or sliver) cut elements – the parts of the domain for which |x| > 1 or |y| > 1 – do
11
not only get thinner but also get shorter for decreasing ε. The intersections between the domain boundary
and the grid lines x = ±1 and y = ±1 form the boundaries of these zones, and move towards the points
x = ±1, y = 0 and x = 0, y = ±1 for ε → 0. As a result, the cut elements become thinner for decreasing
ε until these intersections shift through a vertex, removing the thinnest cut element from the system. At
this point the error decreases, after which it again increases until the intersection shifts through the next
vertex. Counting the number of degrees of freedom in the system confirms that the errors peak at values
of ε where the intersection between the boundary and the grid lines x = ±1 and y = ±1 shift through a
vertex, because at these values also the number of degrees of freedom is reduced, which is indicated by the
vertical gray dash dotted lines in the figure.
5.5. Higher order basis functions and three-dimensional domains
The same effects as observed in the previous examples occur for higher order basis functions and threedimensional domains. The fourth example has also been used to investigate the behavior of a second order
discretization. The results are visible in Figure 9. Note that this also requires a larger stabilization parameter
as the space Vh |0T in (10) is larger for higher order bases. In Figure 10 the results of example 4 in a threedimensional setting are visible. The domain is defined as x8 + y 8 + z 8 < (1 + ε)8 , Dirichlet conditions are
imposed on the complete boundary and the approximated solution equals u = sin(πx) + sin(πy) + sin(πz).
For this three-dimensional computation we have used a first order trilinear basis on a grid with size h = 1/8
instead of h = 1/16 and the maximal refinement depth has been reduced to 0 to save computation time.
Also in three dimensions the same effect is observed.
102
|||u − uh |||
ku − uh kH 1 (Ω)
ku − uh kL2 (Ω)
102
100
|||u − uh |||
ku − uh kH 1 (Ω)
ku − uh kL2 (Ω)
101
100
10−2
10−1
10−2
10−4
10−10
10−8
10−6
ε
10−4
10−10
10−2
Figure 9: Second order B-splines.
10−8
10−6
ε
10−4
10−2
Figure 10: Three-dimensional example.
5.6. Summary of observations
In the previous examples we observed that sliver-like cut configurations (on Dirichlet boundaries) can
cause severe problems in the discretization. The discretization error measured in the H 1 (Ω)-norm and
the mesh-dependent energy norm can not be bounded independently of ε, the parameter that controls the
volume ratios in the sliver cut elements. While the L2 (Ω)-norm error seems to be robust for the simple
cases, it is not for the more complex examples.
We tried to give an interpretation of the origins that trigger the large errors for small ε and were able to
characterize a difference in the relative number of d.o.f.’s involved in the projection of the Dirichlet conditions
by the penalty terms. As a result, in Example 1 the effect is more severe for the triangular discretization
than for the quadrilateral discretization. On specific domains, the instabilities due to the sliver cases are
even not triggered at all, as demonstrated in Example 2. This is similar to penalty methods for boundary
fitting discretizations, where exactly the right number of d.o.f.’s is involved in the penalty, such that very
large penalties can be applied [47]. However, small deviations immediately display the missing robustness
12
with respect to ε and lead to large errors, cf. Example 3. We have obtained similar results for a more
complicated domain with curved boundaries, for quadratic basis functions and in three dimensions, such
that the results do not only apply for piecewise (bi-)linear discretizations and two-dimensional problems on
simple domains.
6. Conclusion and resolutions of the problem
From Section 4 and 5 it is evident that the stabilization parameter needs to be bounded in order to
guarantee optimal approximation properties. In fact, several modifications of Nitsche’s method that preclude
the previously discussed problems already exist in the literature. We divide these in three groups. Firstly,
we discuss alternatives to the symmetric Nitsche formulation. Secondly, there exist approaches to preclude
degrees of freedom that have very small supports in the problem domain, such that the required stabilization
parameter is bounded. Thirdly, we discuss a stabilization mechanism called the ghost penalty, which achieves
a similar effect by penalizing jumps in normal derivatives along boundaries of cut elements.
6.1. Alternative formulations
The most obvious alternative for the symmetric Nitsche formulation is the unsymmetric Nitsche formulation, e.g., [48–50], which however is not adjoint consistent. Other methods to weakly impose Dirichlet
conditions are the penalty method and the use of Lagrange multipliers and related methods, e.g., [47, 51–53].
Another approach with unfitted grids is to manipulate the function space such that it allows for the strong
imposition of boundary conditions, i.e., [54–56].
6.2. Precluding d.o.f.’s with small supports in the problem domain
When the system does not contain functions with very small supports in the problem domain, it follows
from (10) that the required stabilization parameter is reduced. This is an additional advantage of methods
that eliminate functions with very small supports from the system, which were initially introduced in order
to improve the conditioning of the system matrix. The simplest way to achieve this is by simply excluding
these basis functions from the basis, e.g., [28, 44, 56, 57], which however may decrease the accuracy. In
[54, 55, 58–60], functions with small supports in the problem domain are constrained to geometrically
nearby functions, which decreases the required stabilization parameter without compromising the accuracy.
Another approach that precludes basis functions from having a very small stiffness (having the same effect
as precluding functions with a very small support) is using a fictitious domain stiffness, e.g., [12], which is
thoroughly analyzed in [61].
6.3. Ghost penalty
The ghost penalty, see [2, 26], is a consistent penalty term that can be added to the bilinear form and
which is customary in methods referred to as CutFEM, see e.g., [3]. The ghost penalty introduces a penalty
in the vicinity of the boundary that weakly bounds polynomials of neighboring elements to coincide. This
is done by adding terms that penalize jumps in the normal derivatives along element interfaces or local
projection type penalties [62]. This weakly enforces higher order continuity along element interfaces near
the boundary, and therefore relates functions with small support in the boundary region to interior functions.
Similar to strongly enforcing this higher order continuity by constraining basis functions with small supports
to geometrically nearby functions, this effectively controls and bounds the stabilization parameter. Besides
bounding the stabilization parameter, penalizing the jumps in the normal derivatives gives control over the
gradients in cut elements and therefore impede gradients of magnitude ∼ ε−1 , which we observed in the cut
elements.
Remark 7. Inspired by the observations in the experiments, we also suggest a quick fix that is simple to
implement. To alleviate the problem of unbounded stabilization parameters, we modify Nitsche’s method to
a hybrid Nitsche-penalty method. First, we manually set an upper bound for the stabilization parameter
Cλ > 0. We then use Nitsche’s method on the elements where the required stabilization parameter (computed
13
as in (10)) is smaller than this upper bound and use a penalty method ( i.e., no flux terms) with penalty
parameter Cλ on the elements where Nitsche’s method would require a larger value than this manually set
upper bound. Bounding the penalty parameter is easily motivated from an accuracy point of view, when
considering the results and analysis previously presented in this manuscript. It is noted that a pure penalty
method is not consistent with the initial problem (9). This inconsistency can be argued to be acceptable
however, considering that this generally only occurs on very small parts of the domain and that on these
parts of the boundary the flux terms are negligibly small compared to the penalty terms anyway.
This hybrid Nitsche-penalty method yields good approximations in the H 1 (Ω)-norm for the fourth test
case, with Cλ = 16·103 . The results are visible in Figure 11 and show that the quality of the approximation is
virtually independent of ε. This suggests that this could be a simple and straightforward fix for this problem.
We consider this a preliminary result however, as this adaptation of the method has not undergone rigorous
testing. Also setting the value of Cλ needs a more thorough analysis and could require different values for
different mesh sizes, higher order discretizations and other problems than Poisson’s problem.
102
|||u − uh |||
ku − uh kH 1 (Ω)
ku − uh kL2 (Ω)
100
10−2
10−10
10−8
10−6
ε
10−4
10−2
Figure 11: Discretization errors of the modification of Nitche’s method for Example 4. The energy norm error for this hybrid
−1
1
1
method is defined as: |||v|||2 := k∇vk2Ω + kλT 2 ∂n vk2ΓNitsche + kλT2 vk2ΓNitsche + kCλ2 vk2Γpenalty .
Acknowledgement
The research of F. de Prenter was funded by the NWO under the Graduate Program Fluid & Solid Mechanics.
C. Lehrenfeld gratefully acknowledges funding by the German Science Foundation (DFG) within the project
LE 3726/1-1.
References
[1] R. Glowinski, T. Pan, J. Periaux, A fictitious domain method for Dirichlet problem and applications, Computer Methods
in Applied Mechanics and Engineering 111 (3–4) (1994) 283–303.
[2] E. Burman, P. Hansbo, Fictitious domain finite element methods using cut elements: II. A stabilized Nitsche method,
Applied Numerical Mathematics 62 (4) (2012) 328–341.
[3] E. Burman, S. Claus, P. Hansbo, M. Larson, A. Massing, CutFEM: Discretizing geometry and partial differential equations,
International Journal for Numerical Methods in Engineering 104 (7) (2014) 472–501.
[4] A. Massing, M. Larson, A. Logg, M. Rognes, A stabilized Nitsche fictitious domain method for the Stokes problem, Journal
of Scientific Computing 61 (3) (2014) 604–628.
[5] E. Burman, S. Claus, A. Massing, A Stabilized Cut Finite Element Method for the Three Field Stokes Problem, SIAM
Journal on Scientific Computing 37 (4) (2015) A1705–A1726.
[6] J. Parvizian, A. Düster, E. Rank, Finite cell method, Computational Mechanics 41 (1) (2007) 121–133.
[7] A. Düster, J. Parvizian, Z. Yang, E. Rank, The finite cell method for three-dimensional problems of solid mechanics,
Computer Methods in Applied Mechanics and Engineering 197 (45) (2008) 3768–3782.
[8] D. Schillinger, L. Dede, M. Scott, J. Evans, M. Borden, E. Rank, T. Hughes, An isogeometric design-through-analysis
methodology based on adaptive hierarchical refinement of NURBS, immersed boundary methods, and T-spline CAD
surfaces, Computer Methods in Applied Mechanics and Engineering 249 (2012) 116–150.
14
[9] E. Rank, M. Ruess, S. Kollmannsberger, D. Schillinger, A. Düster, Geometric modeling, isogeometric analysis and the
finite cell method, Computer Methods in Applied Mechanics and Engineering 249 (2012) 104–115.
[10] M. Ruess, D. Schillinger, Y. Bazilevs, V. Varduhn, E. Rank, Weakly enforced essential boundary conditions for NURBSembedded and trimmed NURBS geometries on the basis of the finite cell methodod, International Journal for Numerical
Methods in Engineering 95 (10) (2013) 811–846.
[11] M. Ruess, D. Schillinger, A. Özcan, E. Rank, Weak coupling for isogeometric analysis of non-matching and trimmed
multi-patch geometries, Computer Methods in Applied Mechanics and Engineering 269 (2014) 46–71.
[12] D. Schillinger, M. Ruess, The Finite Cell Method: A review in the context of higher-order structural analysis of CAD and
image-based geometric models, Archives of Computational Methods in Engineering (2015) 1–65.
[13] D. Kamensky, M.-C. Hsu, D. Schillinger, J. Evans, A. Aggarwal, Y. Bazilevs, M. Sacks, T. Hughes, An immersogeometric
variational framework for fluid–structure interaction: Application to bioprosthetic heart valves, Computer Methods in
Applied Mechanics and Engineering 284 (2015) 1005–1053.
[14] V. Varduh, M.-C. Hsu, M. Ruess, D. Schillinger, The tetrahedral finite cell method: Higher-order immersogeometric
analysis on adaptive non-boundary-fitted meshes, International Journal for Numerical Methods in Engineering 107 (12)
(2016) 1054–1079.
[15] F. Xu, D. Schillinger, D. Kamensky, V. Varduhn, C. Wang, M.-C. Hsu, The tetrahedral finite cell method for fluids:
Immersogeometric analysis of turbulent flow around complex geometries, Computers & Fluids 141 (2016) 135–154.
[16] C. Peskin, The immersed boundary method, Acta Numerica 11 (2002) 479–517.
[17] P. Bastian, C. Engwer, An unfitted finite element method using discontinuous Galerkin, International Journal for Numerical
Methods in Engineering 79 (12) (2009) 1557–1576.
[18] R. Massjung, An Unfitted Discontinuous Galerkin Method Applied to Elliptic Interface Problems, SIAM Journal on
Numerical Analysis 50 (6) (2012) 3134–3162.
[19] T. Belytschko, N. Moes, S. Usui, C. Parimi, Arbitrary discontinuities in finite elements, International Journal for Numerical
Methods in Engineering 50 (4) (2001) 993–1013.
[20] T.-P. Fries, T. Belytschko, The extended/generalized finite element method: an overview of the method and its applications, International Journal for Numerical Methods in Engineering 84 (3) (2010) 253–304.
[21] A. Hansbo, P. Hansbo, An unfitted finite element method, based on Nitsche’s method, for elliptic interface problems,
Computer Methods in Applied Mechanics and Engineering 191 (47–48) (2002) 5537 – 5552.
[22] S. Groß, V. Reichelt, A. Reusken, A finite element based level set method for two-phase incompressible flows, Computing
and Visualization in Science 9 (4) (2006) 239–257.
[23] R. Becker, E. Burman, P. Hansbo, A Nitsche extended finite element method for incompressible elasticity with discontinuous modulus of elasticity, Computer Methods in Applied Mechanics and Engineering 198 (41) (2009) 3352 – 3360.
[24] R. Becker, E. Burman, P. Hansbo, A hierarchical NXFEM for fictitious domain simulations, International Journal for
Numerical Methods in Engineering 86 (4) (2011) 549–559.
[25] J. Nitsche, Über ein Variations zur Lösung von Dirichlet-Problemen bei Verwendung von Teilräumen die keinen Randbedingungen unterworfen sind, Abhandlungen aus dem mathematischen Seminar der Universität Hamburg 36 (1) (1971)
9–15.
[26] E. Burman, Ghost penalty, Comptes Rendus Mathematique 348 (21) (2010) 1217–1220.
[27] P. Grisvard, Elliptic problems in nonsmooth domains, Pitman, 1985.
[28] A. Embar, J. Dolbow, I. Harari, Imposing Dirichlet boundary conditions with Nitsche’s method and spline based finite
elements, International Journal for Numerical Methods in Engineering 83 (7) (2010) 877–898.
[29] T. Warburton, J. Hesthaven, On the constants in hp-finite element trace inverse inequalities, Computer Methods in
Applied Mechanics and Engineering 192 (25) (2003) 2765–2773.
[30] Z. Yang, M. Ruess, S. Kollmannsberger, A. Düster, E. Rank, An efficient integration technique for the voxel-based finite
cell method, International Journal for Numerical Methods in Engineering 91 (5) (2012) 457–471.
[31] A. Abedian, J. Parvizian, A. Düster, H. Khademyzadeh, E. Rank, Performance of different integration schemes in facing
discontinuities in the finite cell method, International Journal of Computational Methods 10 (3).
[32] S. Duczek, U. Gabbert, Efficient integration method for fictitious domain approaches, Computational Mechanics 56 (4)
(2015) 725–738.
[33] T.-P. Fries, S. Omerović, Higher-order accurate integration of implicit geometries, International Journal for Numerical
Methods in Engineering 106 (5) (2015) 323–371.
[34] C. Lehrenfeld, High order unfitted finite element methods on level set domains using isoparametric mappings, Computer
Methods in Applied Mechanics and Engineering 300 (2016) 716–733.
[35] M. Joulaian, S. Hubrich, A. Düster, Numerical integration of discontinuities on arbitrary domains based on moment fitting,
Computational Mechanics 57 (6) (2016) 979–999.
[36] L. Kudela, N. Zander, S. Kollmannsberger, E. Rank, Smart octrees: Accurately integrating discontinuous functions in 3D,
Computer Methods in Applied Mechanics and Engineering 306 (2016) 406–426.
[37] V. Thiagarajan, V. Shapiro, Adaptively Weighted Numerical Integration in the Finite Cell Method, Computer Methods
in Applied Mechanics and Engineering 311 (2016) 250–279.
[38] A. Stavrev, L. Nguyen, R. Shen, V. Varduhn, M. Behr, S. Elgeti, D. Schillinger, Geometrically accurate, efficient, and flexible quadrature techniques for the tetrahedral finite cell method, Computer Methods in Applied Mechanics and Engineering
310 (2016) 646–673.
[39] T.-P. Fries, S. Omerović, D. Schöllhammer, J. Steidl, Higher-order meshing of implicit geometries – Part i: Integration
and interpolation in cut elements, Computer Methods in Applied Mechanics and Engineering 313 (2017) 759–784.
[40] C. Lehrenfeld, A. Reusken, Optimal preconditioners for Nitsche-XFEM discretizations of interface problems, Numerische
15
Mathematik 135 (2) (2016) 313–332.
[41] F. de Prenter, C. Verhoosel, G. van Zwieten, E. van Brummelen, Condition number analysis and preconditioning for the
finite cell method, Computer Methods in Applied Mechanics and Engineering 316 (2017) 297–327.
[42] J. Céa, Approximation variationnelle des problèmes aux limites, Ph.D. thesis (1964).
[43] A. Ern, J. Guermond, Theory and Practice of Finite Elements, Springer, 2004.
[44] C. Verhoosel, G. Van Zwieten, B. Van Rietbergen, R. De Borst, Image-based goal-oriented adaptive isogeometric analysis
with application to the micro-mechanical modeling of trabecular bone, Computer Methods in Applied Mechanics and
Engineering 284 (2015) 138–164.
[45] L. van Miegroet, P. Duysinx, Stress concentration minimization of 2D filets using X-FEM and level set description,
Structural and Multidisciplinary Optimization 33 (4) (2007) 425–438.
[46] W. Jiang, C. Annavarapu, J. Dolbow, I. Harari, A robust Nitsche’s formulation for interface problems with spline-based
finite elements, International Journal for Numerical Methods in Engineering 104 (7) (2015) 676–696.
[47] I. Babuška, The finite element method with penalty, Mathematics of Computation 27 (122) (1973) 221–228.
[48] E. Burman, A penalty-free nonsymmetric Nitsche-type method for the weak imposition of boundary conditions, SIAM
Journal on Numerical Analysis 50 (4) (2012) 1959–1981.
[49] T. Boiveau, E. Burman, A penalty-free Nitsche method for the weak imposition of boundary conditions in compressible
and incompressible elasticity, IMA Journal of Numerical Analysis 36 (2) (2016) 770–795.
[50] D. Schillinger, I. Harari, M.-C. Hsu, D. Kamensky, S. Stoter, Y. Yu, Y. Zhao, The non-symmetric Nitsche method for the
parameter-free imposition of weak boundary and coupling conditions in immersed finite elements, Computer Methods in
Applied Mechanics and Engineering 309 (2016) 625–652.
[51] I. Babuška, The finite element method with Lagrangian multipliers, Numerische Mathematik 20 (3) (1973) 179–192.
[52] S. Fernández-Méndez, A. Huerta, Imposing essential boundary conditions in mesh-free methods, Computer Methods in
Applied Mechanics and Engineering 193 (12) (2004) 1257–1275.
[53] J. Baiges, R. Codina, F. Henke, S. Shahmiri, W. Wall, A symmetric method for weakly imposing Dirichlet boundary
conditions in embedded finite element meshes, International Journal for Numerical Methods in Engineering 90 (5) (2012)
636–658.
[54] K. Höllig, U. Reif, J. Wipper, Weighted extended B-spline approximation of Dirichlet problems, SIAM Journal on Numerical Analysis 39 (2) (2001) 442–462.
[55] K. Höllig, C. Apprich, A. Streit, Introduction to the WEB–method and its applications, Advances in Computational
Mathematics 23 (1) (2005) 215–237.
[56] R. Sanches, P. Bornemann, F. Cirak, Immersed b-spline (i-spline) finite element method for geometrically complex domains,
Computer Methods in Applied Mechanics and Engineering 200 (13) (2011) 1432–1445.
[57] A. Reusken, Analysis of an extended pressure finite element space for two-phase incompressible flows, Computing and
visualization in science 11 (4) (2008) 293–305.
[58] T. Rüberg, F. Cirak, Subdivision–stabilised immersed b–spline finite elements for moving boundary flows, Computer
Methods in Applied Mechanics and Engineering 209 (2012) 266–283.
[59] T. Rüberg, F. Cirak, A fixed-grid b-spline finite element technique for fluid–structure interaction, International Journal
for Numerical Methods in Fluids 74 (9) (2014) 623–660.
[60] T. Rüberg, F. Cirak, J. Garcı́a Aznar, An unstructured immersed finite element method for nonlinear solid mechanics,
Advanced Modeling and Simulation in Engineering Sciences 3 (1) (2016) 623–660.
[61] M. Dauge, A. Düster, E. Rank, Theoretical and Numerical Investigation of the Finite Cell Method, Journal of Scientific
Computing 65 (3) (2015) 1039–1064.
[62] R. Becker, M. Braack, A finite element pressure gradient stabilization for the Stokes equations based on local projections,
Calcolo 38 (4) (2001) 173–199.
16
| 5cs.CE
|
Solution of the square lid-driven cavity flow of a Bingham plastic using the
finite volume method
Alexandros Syrakosa , Georgios C. Georgioub,a,∗, Andreas N. Alexandrouc
a
Oceanography Centre, University of Cyprus, PO Box 20537, 1678 Nicosia, Cyprus
Department of Mathematics and Statistics, University of Cyprus, PO Box 20537, 1678 Nicosia, Cyprus
c
Department of Mechanical and Manufacturing Engineering, University of Cyprus, PO Box 20537, 1678 Nicosia, Cyprus
arXiv:1604.08461v1 [physics.comp-ph] 28 Apr 2016
b
Abstract
We investigate the performance of the finite volume method in solving viscoplastic flows. The creeping
square lid-driven cavity flow of a Bingham plastic is chosen as the test case and the constitutive equation
is regularised as proposed by Papanastasiou [J. Rheology 31 (1987) 385-404]. It is shown that the
convergence rate of the standard SIMPLE pressure-correction algorithm, which is used to solve the
algebraic equation system that is produced by the finite volume discretisation, severely deteriorates as
the Bingham number increases, with a corresponding increase in the non-linearity of the equations. It
is shown that using the SIMPLE algorithm in a multigrid context dramatically improves convergence,
although the multigrid convergence rates are much worse than for Newtonian flows. The numerical
results obtained for Bingham numbers as high as 1000 compare favorably with reported results of other
methods.
Keywords: Bingham plastic, Papanastasiou regularisation, lid-driven cavity, finite volume method,
SIMPLE, multigrid
This is the accepted version of the article published in: Journal of Non-Newtonian Fluid Mechanics
195 (2013) 19–31, doi:10.1016/j.jnnfm.2012.12.008
c 2016. This manuscript version is made available under the CC-BY-NC-ND 4.0 license http:
//creativecommons.org/licenses/by-nc-nd/4.0/
1. Introduction
Viscoplastic flows constitute an important branch of non-Newtonian fluid mechanics, as many materials of industrial, geophysical, and biological importance are known to exhibit yield stress. In general,
yield-stress fluids are suspensions of particles or macromolecules, such as pastes, gels, foams, drilling
fluids, food products, and nanocomposites. A comprehensive review of viscoplasticity has been carried
out by Barnes [1]. Such materials behave as (elastic or inelastic) solids, below a certain critical shear
stress level, i.e. the yield stress, and as liquids otherwise. The flow field is thus divided into unyielded
(rigid) and yielded (fluid) regions.
∗
Corresponding author
Email addresses: [email protected] (Alexandros Syrakos), [email protected] (Georgios C.
Georgiou), [email protected] (Andreas N. Alexandrou)
Preprint submitted to Journal of Non-Newtonian Fluid Mechanics
April 29, 2016
The simplest constitutive equation describing viscoplasticity is that proposed by Bingham [2]:
τ ≤ τy
γ̇ = 0
,
τy
+ µ γ̇ , τ > τy
τ =
γ̇
(1)
where τy is the yield stress, µ is the plastic viscosity, τ is the stress tensor, γ̇ is the rate of strain tensor,
γ̇ ≡ ∇u + (∇u)T
(2)
u is the velocity vector, and the superscript T denotes the transpose of the velocity-gradient tensor ∇u.
The symbols τ and γ̇ denote the magnitudes of the stress and rate-of-strain tensors, respectively:
τ ≡
1
τ :τ
2
12
,
γ̇ ≡
1
γ̇ : γ̇
2
12
(3)
Another two-parameter viscoplastic equation is the Casson model, which is mostly used in hemodynamics. The Herschel–Bulkley model is the generalisation of the Bingham-plastic equation, which involves
a power-law exponent allowing shear-thinning or shear thickening.
Simulating viscoplastic flows poses extra-ordinary difficulties due to the discontinuity of the constitutive equations. In most cases, it is necessary to determine the location and the shape of the yield
surface at which the flow switches from one branch of the constitutive equation to the other, e.g. from
solid to liquid behavior. Mathematically, the yield surface is the locus of points where τ = τy . A
common approach to overcoming this burden is to regularise the constitutive equation, i.e. to describe
the two branches of (1) by one smooth equation using a stress-growth parameter. The most popular
regularization is that proposed by Papanastasiou [3]:
τy
{1 − exp(−mγ̇)} + µ γ̇
(4)
τ =
γ̇
where m denotes the stress growth parameter, which needs to be “sufficiently” large. Frigaard and
Nouar [4] reviewed systematically the convergence of the Papanastasiou and other regularized models
including the bi-viscosity model [5].
Another approach in solving viscoplastic flows is based on the use of variational inequalities, that is
rate-of-strain minimization or stress maximization, which form the basis of the Augmented Lagrangian
Method [6, 7]. Dean et al. [8] reviewed and compared numerical methods based on the variational
inequality approach.
In this paper, we adopt the regularisation approach and investigate the performance of the finite
volume/multigrid method in solving Bingham plastic flows. The finite volume method (FVM) is a
popular method for solving fluid flows, employed by many general-purpose CFD solvers. One of the
attractive features of the method is that it can be applied to a variety of physical problems with relative
ease. However, there are a limited number of published results on the use of the finite-volume method
to solve Bingham flow problems. Neofytou [9] used a FVM in conjunction with the SIMPLE algebraic
solver (Patankar and Spalding [10]) to simulate the lid-driven cavity flow of various non-Newtonian
fluids, including a Papanastasiou-regularised Bingham plastic at quite low Bingham numbers (0.01–1).
Turan and co-workers [11, 12] also used a commercial FVM/SIMPLE code employing the bi-viscosity
model in order to simulate natural convection of a Bingham plastic in a square cavity. The FVM was
also used to solve flows of a Casson fluid through a stenosis (Neofytou and Drikakis [13]) and through a
sudden expansion (Neofytou and Drikakis [14]), at rather low yield-stress values. Also, de Souza Mendes
et al. [15] and Naccache and Barbosa [16] used the FVM in order to simulate viscoplastic flow through
an expansion followed by a contraction.
2
A benchmark problem for testing numerical methods for both Newtonian and non-Newtonian flows
is the lid-driven cavity flow problem. For laminar Newtonian flow, the reader is referred to the works
of Botella and Peyret [17], Syrakos and Goulas [18], Bruneau and Saad [19], and the references therein.
The lid-driven cavity flow has also been used as a test case for Bingham flows by Sanchez [20], Mitsoulis
and Zisis [21], Vola et al. [22], Elias et al. [23], Yu and Wachs [24], Olshanskii [25], Zhang [26], and dos
Santos et al. [27], who used solution methods other than the FVM, mostly the Finite Element method.
To the authors’ knowledge, only Neofytou [9] used the FVM in order to solve the driven cavity flow of a
Bingham plastic or any other viscoplastic fluid, albeit for very small Bingham numbers (≤ 1). Although
some of the aforementioned published works contain results for non-zero Reynolds numbers, in most
cases the results concern creeping flows (Re = 0).
The objective of the present work is to apply the FVM along with a multigrid method in order to
improve the convergence of the SIMPLE solver when solving the lid-driven cavity Bingham flow for a
broad range of Bingham numbers. Multigrid methods use a hierarchy of grids of progressive fineness.
By applying the algebraic solver on each of these grids, all wavelengths of the algebraic error are reduced
with equal efficiency. Multigrid methods were originally proposed by Fedorenko [28] and later developed
by Brandt [29]. In the context of the FVM, they were first used in the late 80’s (Sivaloganathan and Shaw
[30]; Hortman et al. [31]). Since then, they have been employed in numerous studies involving the FVM,
many of which used the lid-driven cavity flow of a Newtonian fluid as a test problem; see Syrakos and
Goulas [18] and references therein. These Newtonian studies have shown that using multigrid algorithms
can result in very significant performance gains. To our knowledge, the finite volume/multigrid method
has not been tested in the case of Bingham flow.
The rest of the paper is organized as follows. In Section 2, the integral forms of the governing
equations are presented and dedimensionalized. In Section 3, the numerical method, that is the discretisation of the governing equations and the solution of the resulting algebraic system, is discussed. The
numerical results are presented in Section 4. These compare well with the results of Mitsoulis and Zisis
[21] and Yu and Wachs [24]. The convergence of the algebraic solver has also been studied and possible
ways for its acceleration have been investigated. Finally, Section 5 contains our concluding remarks.
2. Governing equations
We consider the steady-state, two-dimensional flow in a square cavity of side L, the top boundary
(lid) of which moves towards the right with a uniform horizontal velocity U , while the remaining sides
are fixed. We work in Cartesian coordinates (x,y), centered at the lower-left corner of the cavity and
denote the unit vectors in the x and y directions by i and j, respectively. Let also ρ and η = η(γ̇) denote
the density and the viscosity of any generalised-Newtonian fluid. By means of the Gauss theorem, the
integral forms of the continuity and the x– and y–components of the momentum equations over a control
volume P are as follows:
Z
ρu · dS = 0
(5)
SP
Z
∂u
· dS −
pi · dS
ρ u u · dS =
η ∇u +
∂x
SP
SP
SP
Z
Z
Z
∂u
ρ v u · dS =
η ∇v +
· dS −
pj · dS
∂y
SP
SP
SP
Z
Z
3
(6)
(7)
where SP is the boundary surface of the control volume, dS is an infinitesimal element of this surface
oriented so that the normal vector points out of P , u = ui + vj is the velocity vector, and p is the
pressure. In the case of the Papanastasiou model (4), the viscosity is given by
τy
{1 − exp(−mγ̇)} + µ
γ̇
η =
(8)
The advantage of the Papanastasiou regularization is that expression (8) is used over the entire flow
domain, i.e. over both yielded and unyielded regions. At high strain rates the viscosity tends towards
µ if the growth parameter m is large enough. In the unyielded regions the viscosity obtains high values
which result in very small values of γ̇ and thus solid body motion is approximated. When γ̇ tends to
zero, then the viscosity (8) tends not to infinity but to the finite value mτy + µ. Some authors [32, 33]
suggested that lower values of m can be used at higher yield stress values and vice versa.
To dedimensionalise the governing equations (5)–(7), we scale lengths by the cavity side L, the velocity components by U , and the pressure and stress by µU/L, and use stars to denote the dimensionless
variables. Taking into account that ρ is constant, the dimensionless equations are as follows:
Z
u∗ · dS∗ = 0
(9)
SP
Z
∂u∗
∗
∗ ∗
· dS −
p∗ i · dS∗
Re
u u · dS =
η ∇u +
∗
∗
∗
∗
∂x
SP
SP
SP
Z
Z
Z
∗
∂u
Re
v ∗ u∗ · dS∗ =
η ∗ ∇∗ v ∗ +
· dS∗ −
p∗ j · dS∗
∗
∗
∗
∗
∂y
SP
SP
SP
Z
∗
∗
∗
η∗ =
Z
∗
(10)
(11)
Bn
{1 − exp(−M γ̇ ∗ )} + 1
γ̇ ∗
(12)
Re ≡
ρU L
µ
(13)
Bn ≡
τy L
µU
(14)
M ≡
mU
L
(15)
where
is the Reynolds number,
is the Bingham number, and
is the stress-growth number.
For the sake of simplicity, the stars denoting the dimensionless variables are dropped hereafter.
3. Numerical method
3.1. Discretisation of the equations
The domain is split into a number of control volumes (CVs) using a Cartesian grid of equally spaced
horizontal and vertical grid lines. For each control volume, the continuity and momentum equations are
approximated using algebraic expressions involving the values of the unknowns u, v, p at the centre of
that CV and at the centres of neighbouring CVs [34]. The computer code which was used for the present
study has the ability to use curvilinear grids composed of quadrilateral CVs - see [35, 18] for details but in the present section only the simpler form that these discretisation schemes acquire when the grid
4
∆x = h
N
n
W
P
w
e
E
∆y = h
s
S
Figure 1: A control volume P and its neighbours.
is Cartesian will be presented. Although the present work concerns only creeping flow, the treatment of
the convection terms of the momentum equations is also described for completeness.
For each CV, the surface flux integrals are calculated separately on each face. Figure 1 shows a
control volume P and its neighbours, S, E, N and W . The letters P , S, E, N and W will also denote
the position vectors of the centres of the respective CVs. First, the flow variables and their normal
derivatives are calculated at the centre of each face using central differences, e.g. for face e:
ue =
∂u
∂x
=
e
uE + uP
2
uE − uP
∆x
(16)
(17)
The flux integrals are then approximated using the midpoint rule, e.g. for the x–momentum equation:
Z
u u · dS ≈ Fe ue
(18)
e
Z
∂u
η∇u · dS ≈ ηe
∆y
∂x e
e
Z
p i · dS ≈ pe ∆y
(19)
(20)
e
Equations (18) – (19) do not describe the approximation scheme sufficiently, and some definitions are
still missing: Fe denotes the mass flux through face e and will be defined shortly; the term (19) is only
part of the total viscous flux; and the term ηe requires that the viscosity has already been calculated
somehow at CV centres. To define these terms we first approximate the gradient operator at CV centres
(equation (17) only approximates the normal component at face centres). At the centre of control
5
volume P the following approximation is used:
∂u
∂x
=
P
uE − uW
,
2∆x
∂u
∂y
=
P
uN − uS
2∆y
(21)
If P is a boundary CV, and the domain boundary coincides with face w, then the horizontal component
of the gradient is calculated as follows:
∂u
∂x
=
P
uE + uP − 2uw
2∆x
(22)
where uw is the boundary value of u, specified by the Dirichlet boundary condition.
Now, the viscosity is calculated at CV centres from equation (12), where γ̇ is calculated from (3),
using the velocity gradients at CV centres (21):
γ̇P
"
∂u
= 2
∂x
2
+2
P
∂v
∂y
2
+
P
∂u
∂y
P
∂v
+
∂x
2 # 21
(23)
P
Thus ηe in (19) is calculated from ηP and ηE using linear interpolation (16). Since the viscosity is only
needed at face centres, an alternative would be to calculate it there directly. This would require the
velocity gradients at the face centres, the normal component of which is readily available from (17).
But the tangential component is not, and would require interpolation of velocity components at face
vertices. So, this alternative would not be less complex than the approach presently adopted.
For the same reason, the remaining part of the viscous fluxes is also calculated by interpolating
the velocity gradients at face centres from the CV centres. For example, for faces e and n we have,
respectively:
Z
Z
∂u
∂u
∂u
∂v
η
η
∆y ,
∆x
(24)
· dS ≈ ηe
· dS ≈ ηn
∂x e
∂x n
n ∂x
e ∂x
Here the derivatives (∂u/∂x)e and (∂v/∂n)e are not the same as in Eq. (17), but are interpolated
from the derivatives (21) at face centres according to (16). This approach is convenient in the case of
curvilinear grids, which is why it is adopted by the code we used for the present study.
This leaves only the mass fluxes to be defined. These are discretised using the central difference
scheme (16), but with the addition of an artificial pressure term whose role is to not allow the appearance
of spurious pressure oscillations in the discrete solution. The mass flux through face e, in non-dimensional
form, is discretised as:
Z
(∆y)2
1 ∂p
∂p
Fe =
u · dS ≈ ue ∆y +
(pP − pE ) −
+
(xP − xE )
(25)
ae
2 ∂x P ∂x E
e
where:
ae = Re (|ue |∆y + |ve |∆x) + 2ηe
∆y ∆x
+
∆x ∆y
(26)
Without this treatment, artificial pressure oscillations do appear when both velocity and pressure are
stored at the CV centres, as in the present scheme. This technique is known as momentum interpolation
and was first suggested by Rhie and Chow [36]. The above variant, proposed in [35], has the advantage
that it is decoupled from the SIMPLE solution algorithm. The resulting equations simplify further in
case of a uniform grid, ∆x = ∆y = h, but for completeness the more general forms are included here.
6
The artificial pressure term is very small, of order O(h5 ), and has a very small impact on the accuracy
of the discretisation of the mass flux.
The test cases examined in this work involve only no-slip solid wall boundaries, with Dirichlet
boundary conditions. The discretisation there is as follows: The pressure is linearly extrapolated from
the interior. The product ∇u · dS of the main viscous terms (19) is calculated as a one-sided difference,
i.e. if e is a boundary face then ∇u · dS is approximated as ∆y ·(u(e) − uP )/(∆x/2) where u(e) is the
exact value of u at the centre of face e, as defined by the Dirichlet boundary condition. Finally, in the
secondary viscous terms (24) both the derivative and the viscosity at the boundary face centre are taken
as equal to their values at the adjacent CV centre.
By substituting all terms in the continuity and momentum equations of each CV by their discrete
counterparts, a non-linear algebraic system is obtained, with four equations (x,y–momentum, continuity
and constitutive equation) and four unknowns (u,v,p,η) per CV. The overall discretisation scheme is of
second order accuracy, meaning that the discretisation error should decrease as O(h2 ). Solution of this
algebraic system gives the values of the unknowns, up to the discretisation error. The solver used to
solve this system is described next.
3.2. Solution of the algebraic system
The popular SIMPLE algorithm [10] was chosen as the non-linear solver. It is a widely used algorithm
and so it will not be fully described here – the interested reader is referred to [37] or [34]. In brief, SIMPLE
is an iterative algorithm which constructs and solves a number of linear systems within each iteration.
These linear systems come from breaking up and linearising the set of discretised equations: The systems
of equations for the momentum components are converted into linear systems for the corresponding
velocity components by evaluating some of the terms (including the viscosity (12)) using the velocity
and pressure values obtained from the previous SIMPLE iteration; and the continuity equations system
is used to construct an approximate “pressure correction” linear system, which attempts to improve
the current pressure estimate so as to force the velocity field to be more continuity-conservant. A
SIMPLE iteration (outer iteration) consists of the successive solution of the linear systems of the velocity
components and of pressure correction. Only a few iterations (inner iterations) of a linear solver are
applied to each linear system, since the matrices of coefficients will change in the next outer SIMPLE
iteration. In this work we used GMRES as the linear solver for the velocity systems, and conjugate
gradients (CG) for the pressure correction system, preconditioned by incomplete factorisations with
zero fill-in – see [38] for detailed descriptions. To achieve convergence, underrelaxation factors au , ap < 1
are applied to the velocity and pressure correction systems respectively [34]. SIMPLE iterations are
repeated until the residuals of all the original non-linear algebraic equations drop below a selected
threshold.
The only modification needed so that SIMPLE can be used for Bingham or other generalisedNewtonian flows is that, at the start of every SIMPLE iteration, the viscosity must be updated at
each CV according to (12), using the current estimate of the velocity field to calculate γ̇.
The SIMPLE algorithm converges rather slowly, so it was decided to implement it in a multigrid
context to accelerate its convergence. Many algebraic solvers are able to quickly reduce the short
wavelengths of the error but are quite slow in reducing the longer wavelengths. SIMPLE is such a
solver [39], not only because the linear solvers used for the inner iterations may themselves have this
property, but also because of local assumptions made in the linearisation of the non-linear terms and
in the construction of the pressure correction equation, which relate a CV to its direct neighbours. In
such a case, after a few iterations have been performed and the short wavelengths of the error have been
reduced, it is beneficial to move the solution procedure to a coarser grid, where the direct neighbours of
a CV are farther away and so the long wavelengths of the error appear shorter and can be reduced more
7
efficiently by the same algebraic solver. By using a cascade of progressively coarser grids, all wavelengths
of the error can be reduced with equal efficiency [29].
The solution procedure is transferred between grids as follows. Let the system of all algebraic
equations be written as:
Nh (xh ) = bh
(27)
where now the vector xh stores all the unknowns (velocity components, pressure, and viscosity) at all
CV centres of the grid whose spacing is h. After a few SIMPLE iterations an approximate solution x∗h
is obtained which satisfies the above equation up to a residual rh :
Nh (x∗h ) = bh − rh = Nh (xh ) − rh ⇒ Nh (xh ) = Nh (x∗h ) + rh
(28)
The algebraic system (28) can be approximated on a coarser grid of spacing 2h (grid 2h hereafter),
obtained by removing every second grid line of grid h, as follows:
N2h (x2h ) = N2h (Ih2h x∗h ) + Iˆh2h rh
(29)
The operator N2h is constructed on grid 2h using the same discretisation schemes as Nh on grid h, with
2∆x and 2∆y used instead of ∆x and ∆y. The restriction operator Ih2h transfers the variables from grid
h to grid 2h. The notation comes from the identity matrix I, since the variables are not transformed
but rather transferred from one grid to another. In the present study this restriction operator sets the
value of a variable at the centre of a control volume P of grid 2h equal to the average of the values of
that variable over the 4 CVs of grid h that overlap with P , which will hereafter be called the children of
P . The (possibly different) operator Iˆh2h is used for transferring the residuals to grid 2h. In the present
study the residuals at a control volume P of grid 2h are set equal to the sum of the residuals of all its
children. Since the residuals are essentially flux imbalances, this means that the flux imbalance of P is
set equal to the sum of the imbalances of its children.
The right hand side of (29) is known, so the system can be solved to obtain x2h . Then, the correction
x2h − Ih2h x∗h is transferred back to grid h, to obtain a better estimate xnew
of the exact solution xh , and
h
SIMPLE iterations can resume on grid h:
h
xnew
= x∗h + I2h
x2h − Ih2h x∗h
h
(30)
h
The prolongation operator I2h
transfers the correction from grid 2h to grid h; in the present study linear
interpolation is used. It is straightforward to see that if in (28) the exact solution has already been
obtained (x∗h = xh ) then the residuals will be zero and therefore so will be the last term of (29). Then
the solution of (29) will be x2h = Ih2h x∗h (it is important that Ih2h x∗h be used as the initial guess when
solving (29)), and therefore the correction x2h − Ih2h x∗h will be zero. So, (30) will leave the solution
unaltered: xnew
= x∗h = xh .
h
The coarse grid problem (29) can be solved using an even coarser grid 4h and so on. This whole
process of going down from grid h to some coarsest grid and then back again up to grid h constitutes
a multigrid cycle. In general, a number of multigrid cycles will be required to reduce the residuals by
a given amount. Different kinds of cycles have been suggested and used. The most popular are the
V(ν1 , ν2 ) and W(ν1 , ν2 ) cycles. In V(ν1 , ν2 ) cycles, after performing ν1 iterations on the fine grid h,
the coarse grid system (29) is solved using one multigrid cycle, and then the coarse grid correction is
transferred back to grid h according to (30), where another ν2 iterations are performed. The difference
between the W and V cycles is that in W cycles the coarse grid problems such as (29) are solved using
two rather than one cycle. In the present work we have observed that it is sometimes useful to perform
8
ν3 extra SIMPLE iterations on the finest grid h only, between cycles. Such cycles will be denoted
V(ν1 , ν2 )–ν3 or W(ν1 , ν2 )–ν3 respectively. These cycles are shown schematically in Figure 2. The costs
of a V(ν1 , ν2 )–ν3 and a W(ν1 , ν2 )–ν3 cycle are approximately equal to [ 34 (ν1 +ν2 )+ν3 ] and [2(ν1 +ν2 )+ν3 ]
times the cost of a single iteration on the finest grid h, respectively.
h
2h
4h
8h
16h
ν1
ν2 + ν3
ν1
ν2
ν1
ν2
ν1
ν2
ν0
ν1
ν2 + ν3
ν2 + ν1
ν1
ν2 + ν1
ν1
ν2 + ν1
ν1
ν0
ν2 + ν1
ν2 ν1
ν2 + ν1
ν2 ν1
ν0
ν2
ν0
ν0
ν2
ν2 + ν1
ν1
ν0
ν0
ν2
ν2 + ν1
ν2 ν1
ν0
ν2
ν0
Figure 2: Multigrid cycles: V(ν1 , ν2 )–ν3 (left) and W(ν1 , ν2 )–ν3 (right) cycles, using 5 grid levels.
4. Numerical results
The creeping flow (Re = 0) of a Bingham plastic in a square lid-driven cavity has been chosen as
the test case for the finite volume method described in the previous section. The computational domain
is a square, enclosed by solid boundaries of equal length. The top boundary (lid) moves towards the
right with a uniform horizontal velocity, while the rest of the boundaries are still. The problem was
solved for a range of Bingham numbers, up to 1000. In all cases an exponent M = 400 was used, except
for the two highest Bingham numbers tested, Bn = 500 and 1000, for which M = 100 was used due
to convergence difficulties. In order to check grid convergence, the domain was discretised using three
uniform Cartesian grids with 64 × 64, 128 × 128 and 256 × 256 CVs. Unless otherwise stated, the results
presented below were calculated on the 256 × 256 grid. The pressure was set to zero at the centre of the
domain.
Figure 3 shows the streamlines calculated for selected Bingham numbers. These are the Bingham
numbers chosen also by Mitsoulis and Zisis [21] and Yu and Wachs [24] for their corresponding figures.
The “unyielded” regions are also shown. These are defined here as the regions where τ < Bn, where
τ = η γ̇ is calculated from (12) and (23). Of course, they are not actually unyielded, but they approximate
the unyielded regions of an ideal Bingham flow. The results generally agree with those presented by
other researchers, in [21, 24–26]. The agreement is good with the results of Mitsoulis and Zisis [21]
in the whole range of Bingham numbers, except that these authors predict yielded regions with more
rounded corners than the present results. Also, their yield surfaces are not smooth, but this may be
due to the low resolution of the grid that they employed. The results also agree well with those of Yu
and Wachs [24] up to a Bingham number of 20, but there are notable differences at higher Bingham
numbers - however, Yu and Wachs state that, as far as capturing the yield surface is concerned, their
method performs well at low to moderate Bingham numbers but less so at high Bingham numbers.
Direct comparison can also be made between the present results and those of Olshanskii [25] for Bn = 2
and Bn = 5, who used an augmented Lagrangian approach instead of a regularisation method, and the
agreement is very good. Zhang [26], who also used an augmented Lagrangian approach, also provides
results for Bn = 2, 20, 50 and 200, where the unyielded regions appear somewhat smaller, flatter, and
rounded compared to the present results, although qualitatively similar.
The main characteristic of the flow field is the vortex which develops at the upper central region of
the cavity, similarly to the Newtonian case. Two distinct unyielded regions can be observed: A larger one
9
(a) Bn = 2
(b) Bn = 5
(c) Bn = 20
(d) Bn = 50
(e) Bn = 200
(f) Bn = 500
Figure 3: Streamlines plotted at intervals of 0.004 starting from zero. Unyielded areas (τ < Bn) are shown shaded.
10
at the bottom of the cavity, and a smaller one just below the vortex centre. The flow field is symmetric
with respect to the vertical centreline. As the Bn number increases, the unyielded regions expand and
the flow circulation becomes weaker and limited to the upper part of the cavity, with the centre of the
vortex coming closer to the lid. One can notice in Figure 3 that a couple of secondary vortices appear
at the lower two corners of the cavity, which are completely inside the lower unyielded region. These
vortices, which appear to slightly grow in size as Bn increases, are an artifact of the regularisation of the
Bingham model by the Papanastasiou approximation. In fact, the velocity is extremely low throughout
the lower unyielded region. In reality, since this region is in contact with the rigid walls which are
motionless, and a no-slip boundary condition applies, the unyielded material should also be motionless
throughout.
The situation is different at the upper unyielded region, where the velocity is non-zero as can be
clearly seen from the streamlines’ spacing. So, the upper unyielded regions move as solid bodies. In the
present case it appears that the material at the upper unyielded region moves as a solid body, and in
particular it rotates, with the streamlines forming circular arcs inside it. The fact that the streamlines
cross into this region means that the fluid “solidifies” on entry, and becomes fluid again on exit from
the region.
The behaviour of the vortex as a function of the Bn number is described in more detail in Figure 4,
where the vertical position and the strength of the vortex are plotted as functions of Bn. The weakening
of the circulation as Bn increases is accompanied by an increase in the stress and pressure levels in the
cavity, as greater stresses are needed in order to make the material flow. Figure 4 also serves to validate
the results of the present study against those of other researchers. It can be seen that the results are
quite close to previously published data of [21] and [24]. The results are closer to those of Mitsoulis and
Zisis [21] as far as the vortex strength is concerned, and to the results of Yu and Wachs [24] as far as
the vortex position is concerned (possibly because the grid of [21] is coarse).
Figure 5 shows the pressure and shear stress distributions at the left half of the lid. Increasing the
Bingham number from 2 to 50 causes a significant increase in pressure and stress, as discussed previously.
Figure 5 can also be used to verify grid convergence, as it contains results on the three grids used in this
work. In the interior of the lid the pressure and stress converge with grid refinement, and the rate of
convergence is that expected of a 2nd order method. However, grid refinement causes the pressure and
stress to tend to infinity at the lid corners. This is because of the discontinuity which exists at these
corners: the velocity jumps from zero (side walls) to one (lid). Due to this singularity, it was observed
that the pressure and stress integrals over the lid do not converge with grid refinement, and so the
shear force needed to drive the lid cannot be calculated accurately. The results on grid convergence are
discussed later on. The increase in pressure with the Bingham number can also be seen in the pressure
contours of Figure 6. It is clear that the presence of the unyielded regions causes a distortion in the
pressure field. The streamwise pressure gradients are greater inside the upper unyielded region.
An important issue when using the Papanastasiou regularisation is the choice of the exponent M .
The higher the value of M , the better the approximation, but also the more difficult it is to solve the
equations due to the increased degree of nonlinearity. Therefore, for practical reasons M has to be kept
within certain limits. Here we provide some comparisons showing the effect of M on the quality of the
results, while the issue of the computational effort required as a function of M will be investigated in the
next section. One result of importance in a range of applications is the location of the yield lines, which
are approximated in the present method by τ = Bn, as previously stated. Figure 7 shows that the use of
the different values of 100, 200 and 400 for M does not affect the location of the yield lines significantly.
The difference is especially small between M = 400 and M = 200. Figure 8 shows iso-stress lines which
deviate slightly from the yield stress value. Comparing Figures 7 and 8 shows that the lines τ = τy +
( > 0) are less sensitive to the choice of M than the lines τ = Bn. This verifies the result of Alexandrou
11
(a) Vortex centre
(b) Vortex strength
Figure 4: The y coordinate of the centre of the vortex, and vortex strength as a function of the Bingham number.
12
(a) p at the lid
(b) τyx at the lid
Figure 5: Pressure and shear stress at the left half of the lid for Bn = 2 and 50. Results are shown for grids 64 × 64
(chained lines), 128 × 128 (dashed lines), and 256 × 256 (solid lines).
and co-workers [33, 40] that the location of iso-stress lines τ = τy + ( > 0) is predicted almost equally
well for a range of M values, whereas to accurately predict the location of the yield line τ = τy a high
value of M is required. The iso-stress lines τ = τy − ( > 0) nearly coincide with the yield stress line
except near the side walls and the lower corners of the upper unyielded region. Therefore, there is a
sharp decrease in stress as one moves into the unyielded zone. On the contrary, this is not observed
with the τ = τy + lines which are located at a distance from the yield line, meaning that the stress
increases more gradually as one moves into the yielded zone. It is interesting to note that for Bn = 200
the difference between the actual stress and the yield stress is less than 1% throughout the fluid region
that separates the two solid regions, since the τ = 1.01Bn contours do not cross into this region. So,
the deformation rate is quite small there, and most of the shear takes place close to the lid.
(a) Bn = 0
(b) Bn = 2
(c) Bn = 20
Figure 6: Pressure contours for various Bn numbers. For Bn 6= 0 the thick lines outline the unyielded regions. Note that
the pressure scale of the Bn = 20 case is different.
13
(a) Bn = 2
(b) Bn = 20
(c) Bn = 200
Figure 7: Yield lines (τ = Bn) calculated with M = 100 (dotted), M = 200 (dashed) and M = 400 (solid).
(a) Bn = 2, τ = (1 + 0.01)Bn
(b) Bn = 20, τ = (1 ± 0.01)Bn
(c) Bn = 200, τ = (1 ± 0.01)Bn
Figure 8: Iso-stress lines calculated with M = 100 (dotted), M = 200 (dashed) and M = 400 (solid). Unyielded regions
(τ ≤ Bn) are shown shaded.
14
Figure 9: The x component of velocity along the vertical centreline (x = 0.5), for Bn = 0, 2, 5, 20, 50, 200 and 500.
In closing this section we provide some detailed results at the vertical centreline. The x velocity
component (u) is plotted along the vertical centreline in Figure 9. The limits of the yielded / unyielded
zones can be clearly seen on that graph. On each curve (except for Bn = 0) two straight line segments
can be identified. One is completely vertical with u = 0 and stretches from y = 0 up to a certain
height; this segment corresponds to the lower unyielded zone which is motionless. The other segment
corresponds to the upper unyielded zone; there the velocity is non-zero but decreases linearly with
height, which shows that the upper unyielded zone rotates as a solid body. The centre of rotation can
be estimated by extending the straight line segment until it intersects the u = 0 line.
More detailed results are shown in Tables 1 (Bn = 2) and 2 (Bn = 50), where the velocity values
have been obtained at selected points using linear interpolation from the values at adjacent CVs. To
study grid convergence, results of various grids are also included. The exponent given in the tables is
the order of grid convergence, which should be q = 2 for our present second-order accurate scheme. It
is calculated by the following formula (see [34]):
128 −u64
log uu256
−u128
(31)
q =
log(2)
where the subscripts denote the grid where u has been calculated.
Tables 1 and 2 show that q ≈ 1.5 for Bn = 2, and q ≈ 1 for Bn = 50. So, second-order convergence is
not exhibited, which could be a sign that an even finer grid would be useful. However, the discretisation
error is already relatively small as can be seen from the difference between the solutions at various grids.
Besides, the use of a finer grid would significantly increase the computational cost as will be described
in the next section. For Bn = 50 the value of q is especially low at the upper parts of the centreline,
which is an indication that at higher Bn numbers (when most of the flow occurs at the upper part of
15
the domain) it would be of benefit to use non-uniform grids with increased resolution near the top, or
adaptive grids.
The last three columns of the tables are defined as follows:
128
δ256
400
u400
256 − u128
%,
= 100·
u400
256
200
δ400
200
u400
256 − u256
= 100·
%,
u400
256
100
δ400
100
u400
256 − u256
= 100·
%
u400
256
(32)
where um
n has been calculated on grid n × n with M = m. These values must be interpreted with
128
caution, as they may appear large if u400
256 is close to zero. δ256 is a measure of the discretisation error,
and it can be noticed that for Bn = 2 it is quite small, of the order of 0.1–1%, whereas for Bn = 50
200
100
it is markedly larger, of the order of 5%. For Bn = 2, δ400
is of the order of 0.1%, while δ400
is about
200
3 times higher at each point. δ400 shows greater variability for Bn = 50, and is a little higher than for
100
200
Bn = 2. δ400
, for Bn = 50, has about twice the value of δ400
in the upper part of the centreline, and
200
three times the value of δ400 in the lower part. In general, it appears that, with the present choices of
grid density and M , grid coarseness is a larger source of error than the smallness of M .
5. Convergence of the algebraic solver
When using the SIMPLE algorithm to solve Bingham flows, either as a single-grid solver or in
a multigrid context, one realises that there is a severe deterioration of the convergence rates as the
Bingham number increases. In Figure 10 we plot against the number of iterations the L∞ -norm of the
residual vector of the x−momentum equations,
krk∞ =
max {|rP |} ,
P =1,...,N
(33)
where rP is the residual, expressed per unit volume, of the x−momentum equation of control volume P
and N is the total number of control volumes in the grid. The residual norm is scaled by 1000 because
in the actual numerical experiments we solved the dimensional version of the equations, with L = 1 m,
U = 1 m/s, and µ = 1000 Pa·s. Therefore, the residuals of the dimensional equations are 1000 times
larger than those of the dimensionless equations.
Figure 10 shows that the performance of SIMPLE, as a single-grid solver, deteriorates as either Bn
or M increases. The choices au = 0.7 and ap = 0.3, used to obtain these results, are reasonable for
Newtonian flows [34]. In Figure 11, the effect of varying ap is illustrated. As ap is increased, convergence
becomes faster, especially at the initial stages of iteration; however, it also becomes more oscillatory
with large spikes. Beyond a certain value of ap (ap = 0.3 in Figure 11) convergence stalls. On the other
hand, using a very small ap results in smooth but slow convergence. Due to the very large number of
combinations of Bn, M , au , ap and grid density, it is impossible to investigate fully the effect of all these
parameters. We have observed (results not shown) that the performance of SIMPLE is similar also for
other values of au in the range 0.8 - 0.2, but beyond these values it is difficult to obtain convergence.
The performance of SIMPLE as a smoother in a multigrid context is tested next. In applying
multigrid, the coarsest grid used in the cycles was the 8×8 grid. According to our numerical experiments,
the convergence deterioration of the SIMPLE algorithm with increasing Bn number also reflects on the
multigrid performance. In fact, as Figure 12 shows (solid lines), the multigrid method does not converge
beyond a relatively small value of Bn, around 0.5. As a remedy, we applied the suggestion of Ferziger
and Peric [34], that if a fluid property, such as the viscosity when RANS turbulence models are used,
varies by orders of magnitude within the computational domain, then it may be useful to update
that property only on the finest grid and keep it constant within a multigrid cycle. So, we tried an
implementation where the viscosity at coarse grids is not calculated afresh from (12) according to the
16
Table 1: Values of the x-component of velocity along the vertical centreline for Bn = 2.
grid
y
64 × 64
M = 400
128 × 128
M = 400
256 × 256
M = 400
256 × 256
M = 200
256 × 256
M = 100
q
128
δ256
200
δ400
100
δ400
1.00
0.975
0.950
0.900
0.850
0.800
0.750
0.700
0.650
0.600
0.550
0.500
0.450
0.400
0.350
0.300
0.250
0.200
0.150
0.100
0.050
0.000
1.00000
0.81421
0.63974
0.33677
0.10671
-0.04182
-0.10310
-0.13840
-0.17334
-0.20322
-0.20660
-0.18918
-0.16059
-0.12699
-0.09272
-0.06101
-0.03421
-0.01443
-0.00365
-0.00078
-0.00030
0.00000
1.00000
0.81544
0.64222
0.34016
0.11033
-0.03945
-0.10602
-0.14159
-0.17689
-0.20575
-0.20681
-0.18905
-0.16039
-0.12678
-0.09260
-0.06112
-0.03464
-0.01493
-0.00344
-0.00063
-0.00028
0.00000
1.00000
0.81581
0.64279
0.34121
0.11160
-0.03834
-0.10703
-0.14267
-0.17810
-0.20606
-0.20679
-0.18898
-0.16033
-0.12674
-0.09261
-0.06120
-0.03485
-0.01523
-0.00356
-0.00058
-0.00027
0.00000
1.00000
0.81581
0.64280
0.34126
0.11172
-0.03811
-0.10688
-0.14260
-0.17793
-0.20574
-0.20654
-0.18878
-0.16018
-0.12665
-0.09259
-0.06126
-0.03502
-0.01554
-0.00408
-0.00111
-0.00053
0.00000
1.00000
0.81581
0.64282
0.34136
0.11193
-0.03774
-0.10653
-0.14242
-0.17757
-0.20516
-0.20610
-0.18841
-0.15990
-0.12647
-0.09255
-0.06138
-0.03534
-0.01613
-0.00508
-0.00208
-0.00101
0.00000
1.74
2.13
1.69
1.51
1.09
1.53
1.56
1.56
3.05
0.82
1.89
2.45
0.23
1.07
0.78
1.76
1.64
-
0.00%
0.05%
0.09%
0.31%
1.14%
2.90%
0.95%
0.76%
0.68%
0.15%
0.01%
0.04%
0.03%
0.03%
0.02%
0.14%
0.59%
1.92%
3.21%
7.88%
2.42%
0.00%
0.00%
0.00%
0.00%
0.02%
0.11%
0.59%
0.14%
0.05%
0.09%
0.16%
0.12%
0.11%
0.09%
0.07%
0.03%
0.10%
0.48%
2.05%
14.68%
90.21%
94.50%
0.00%
0.00%
0.00%
0.01%
0.04%
0.29%
1.56%
0.47%
0.18%
0.29%
0.44%
0.34%
0.30%
0.27%
0.21%
0.08%
0.28%
1.40%
5.94%
42.93%
257.00%
272.53%
0.00%
Table 2: Values of the x-component of velocity along the vertical centreline for Bn = 50.
grid
y
64 × 64
M = 400
128 × 128
M = 400
256 × 256
M = 400
256 × 256
M = 200
256 × 256
M = 100
q
128
δ256
200
δ400
100
δ400
1.0000
0.9875
0.9750
0.9500
0.9250
0.9000
0.8500
0.8000
0.7500
0.7000
0.6500
0.6000
0.5750
0.5500
0.5250
0.5000
0.4500
0.4000
0.3000
0.2000
0.1000
0.0000
1.00000
0.71205
0.45957
0.10907
-0.01230
-0.02462
-0.03500
-0.04469
-0.05431
-0.06386
-0.07328
-0.08172
-0.08318
-0.07734
-0.06148
-0.03959
-0.00874
-0.00153
-0.00048
-0.00026
-0.00013
0.00000
1.00000
0.72144
0.47518
0.11660
-0.01713
-0.02952
-0.03941
-0.04914
-0.05880
-0.06840
-0.07788
-0.08678
-0.08778
-0.07753
-0.05345
-0.02641
-0.00255
-0.00076
-0.00040
-0.00022
-0.00011
0.00000
1.00000
0.72838
0.49123
0.13617
-0.01946
-0.03123
-0.04142
-0.05148
-0.06148
-0.07142
-0.08125
-0.09080
-0.09275
-0.08027
-0.05286
-0.02406
-0.00137
-0.00069
-0.00038
-0.00021
-0.00010
0.00000
1.00000
0.72886
0.49225
0.13835
-0.01812
-0.03106
-0.04143
-0.05156
-0.06157
-0.07145
-0.08113
-0.09033
-0.09180
-0.07921
-0.05252
-0.02464
-0.00233
-0.00135
-0.00075
-0.00042
-0.00020
0.00000
1.00000
0.72931
0.49324
0.14067
-0.01641
-0.03073
-0.04142
-0.05165
-0.06162
-0.07135
-0.08069
-0.08917
-0.08990
-0.07706
-0.05126
-0.02484
-0.00408
-0.00263
-0.00146
-0.00083
-0.00040
0.00000
0.43
-0.04
-1.38
1.05
1.52
1.14
0.93
0.75
0.59
0.45
0.33
-0.11
-3.91
3.76
2.49
2.38
3.46
1.95
1.99
2.08
-
0.00%
0.95%
3.27%
14.38%
11.97%
5.49%
4.84%
4.55%
4.36%
4.23%
4.15%
4.43%
5.36%
3.42%
1.12%
9.75%
86.69%
10.19%
5.70%
4.82%
4.40%
0.00%
0.00%
0.07%
0.21%
1.60%
6.90%
0.56%
0.03%
0.16%
0.14%
0.04%
0.15%
0.52%
1.03%
1.32%
0.63%
2.41%
70.45%
96.74%
97.59%
97.74%
97.74%
0.00%
0.00%
0.13%
0.41%
3.30%
15.70%
1.60%
0.01%
0.33%
0.23%
0.09%
0.69%
1.80%
3.08%
3.99%
3.01%
3.23%
198.24%
283.34%
287.10%
287.60%
287.34%
0.00%
17
Figure 10: The L∞ norm of the x-momentum residual plotted against the number nS of SIMPLE iterations, for single
grid solution; au = 0.7, ap = 0.3, 256 × 256 grid.
restricted velocity field, but it is directly restricted (interpolated) from the immediately finer grid. The
convergence of this modified multigrid method is also shown in Figure 12 (dashed lines). The method
converges more slowly than the standard multigrid method, but it is more robust and can converge over
a wider range of Bingham numbers. Both methods converge equally fast up to a point, beyond which
the convergence of the modified method suddenly slows down. It appears that both methods are equally
capable of reducing certain components of the residual, which dominate the residual at the initial stages
of iteration. However, the modified method is less capable of reducing certain other components, which
dominate the error beyond a certain point.
The modified method is not strictly “multigrid” since part of the solution, namely the updating
of the viscosity, only takes place on the finest grid. Therefore the method exhibits also some singlegrid convergence characteristics. In particular, as Figure 13 shows for Bn = 0.05, the convergence of
the modified method slows down as the grid is refined (this is true only for those slowly converging
components of the residual). On the contrary, the standard method converges equally fast on all grids,
which is normal multigrid behaviour.
For higher Bingham numbers it is necessary to use the modified multigrid method, and even in that
case convergence difficulties are encountered. However, the gains compared to the single-grid method
are still quite impressive. Figure 14 shows the convergence rates for various Bingham numbers, using
both single-grid (SG) and multigrid (MG) procedures. In each case the solution on the 128 × 128 grid
was used as the initial guess and the coarsest grid used by the multigrid cycles was the 8 × 8 grid. The
computational effort (x–axis) is measured in terms of equivalent fine-grid SIMPLE iterations. For the
single-grid cases, this is just the number of SIMPLE iterations performed. For the multigrid cases, the
number of cycles is multiplied by the number of fine-grid SIMPLE iterations that cost computationally
the same as a single cycle. For example, nC W(ν1 , ν2 )–ν3 cycles cost approximately the same as nS =
18
Figure 11: The L∞ norm of the x-momentum residual plotted against the number nS of SIMPLE iterations, for single
grid solution; Bn = 200, M = 200, 256 × 256 grid.
nC · [2(ν1 + ν2 ) + ν3 ] SIMPLE iterations on the finest grid, as mentioned in section 3.2. It should be
noted that the cost of restriction and prolongation is omitted in this calculation, since the cost of these
operations is very small compared to the cost of the SIMPLE iterations, especially if one considers that
the numbers of pre- and post- smoothing iterations are large, and fine-grid iterations are also carried
out between cycles. Therefore, MG and SG convergence rates are directly comparable in Fig. 14.
One may also observe that as the Bn number increases, the choice of multigrid parameters becomes
rather unusual compared to the usual multigrid practice: the number of pre- and post-smoothing sweeps
is quite large, and a large number of SIMPLE iterations are required between cycles. These choices of
parameters have been found necessary to obtain convergence. It may also be seen that very small
values of ap were used in the multigrid cases. This was necessary, otherwise the procedure did not
converge. This may be associated with the fact that, as ap increases, SIMPLE converges faster but in
a very oscillatory manner, which may cause problems to the multigrid procedure. Since the single-grid
procedure converges faster when ap is as large as possible, the SG results have been obtained with large
values of ap , compared with the MG cases. It can be observed that multigrid is extremely efficient at
low Bn numbers, but has difficulty when Bn is large. Yet, in every case the multigrid convergence
rates are much faster than the single-grid rates, except for the first few iterations. Therefore multigrid
is preferable in any case. Moreover, we note that multigrid (with the particular choices of parameters)
converges faster for Bn = 200 with M = 200 than for Bn = 20 with M = 400. Indeed we observed
in every case that convergence deteriorates significantly not only with increasing Bn but also with
increasing M (for every Bingham number we solved the problem using M = 100, 200, and 400, except
for Bn = 500 and 1000 where only M = 100 was used due to convergence difficulties).
It is worth mentioning that the number of inner iterations seems to play an important role. In Figure
19
Figure 12: The L∞ norm of the x-momentum residual plotted against the number of multigrid cycles nC , for Bn = 0.05,
0.2, and 0.5 (M = 400, grid 256×256). Solid lines depict convergence of the standard multigrid method, while dashed lines
depict convergence of the modified multigrid method where viscosity is not updated at the coarse grids, but interpolated
from the immediately finer grid.
Figure 13: The L∞ norm of the x-momentum residual plotted against the number of multigrid cycles, nC , for Bn = 0.05,
M = 400, on various grids: 64 × 64 (chained lines), 128 × 128 (dashed lines), and 256 × 256 (solid lines). In the modified
multigrid method, viscosity is not updated at the coarse grids, but simply interpolated from the immediately finer grid.
20
(a)
(b)
Figure 14: (a) The L∞ norm of the x-momentum residual on the 256 × 256 grid plotted against the number of equivalent
fine-grid SIMPLE iterations, nS . In all cases M = 400, except M = 200 for Bn = 200. SG and MG denote respectively
single-grid and multigrid solution procedures, the initial guess being the solution of the immediately finer (128 × 128) grid.
(b) Zoom–in of (a) so that the curves 1m and 2m (which are too close to the vertical axis to be visible in (a)) are visible.
Also, to improve the readability, the results for Bn = 200 have been omitted from graph (b).
14 it can be seen that for Bn = 200, using 4 GMRES iterations and 8 CG iterations for the velocity
and pressure correction systems (labelled “ii(4,8)” in the figure) significantly improves the convergence
rate of the single-grid procedure compared to the case that 15 and 7 iterations are used instead (labelled
“ii(15,7)”). It seems that due to the nonlinearity of the algebraic system it is better not to perform
many iterations on the linearised velocity systems within each SIMPLE outer iteration. This was not
investigated in detail, due to the fact that there are already a large number of parameters involved in
the solution procedure.
6. Conclusions
We have solved the creeping lid-driven cavity flow of a Bingham plastic using the Papanastasiou
regularization and the finite volume method combined with a multigrid algorithm. Results have been
obtained for Bingham numbers in the range 0–1000. These compare favorably with the results of other
methods, such as the finite-element and the finite-difference method, and show that the proposed method
21
provides a useful tool in solving viscoplastic flows for a wide range of Bingham numbers. It should be
noted that the convergence of the method becomes slow at high values of the Bingham number and
the regularization parameter. With the use of a modified multigrid method convergence is accelerated
considerably compared to the single-grid SIMPLE method.
Acknowledgements
This work was co-funded by the European Regional Development fund and the Republic of Cyprus
through the Research Promotion Foundation (research project AEIΦOPIA/ΦYΣH/0609(BIE)/15).
REFERENCES
References
[1] H. A. Barnes, The yield stress – a review or ‘παντ α ρι’ – everything flows?, Journal of NonNewtonian Fluid Mechanics 81 (1999) 133 – 178.
[2] E. C. Bingham, Fluidity and plasticity, McGraw-Hill, New York, 1922.
[3] T. C. Papanastasiou, Flows of materials with yield, Journal of Rheology 31 (1987) 385–404.
[4] I. Frigaard, C. Nouar, On the usage of viscosity regularisation methods for visco-plastic fluid flow
computation, Journal of Non-Newtonian Fluid Mechanics 127 (2005) 1 – 26.
[5] E. O’Donovan, R. Tanner, Numerical study of the Bingham squeeze film problem, Journal of
Non-Newtonian Fluid Mechanics 15 (1984) 75 – 83.
[6] M. Fortin, R. Glowinski, Augmented Lagrangian Methods: applications to the numerical solution
of boundary-value problems, North-Holland, Amsterdam, 1983.
[7] R. Glowinski, Numerical methods for nonlinear variational problems, Springer, New York, 1984.
[8] E. J. Dean, R. Glowinski, G. Guidoboni, On the numerical simulation of Bingham visco-plastic
flow: Old and new results, Journal of Non-Newtonian Fluid Mechanics 142 (2007) 36 – 62.
[9] P. Neofytou, A 3rd order upwind finite volume method for generalised Newtonian fluid flows,
Advances in Engineering Software 36 (2005) 664–680.
[10] S. V. Patankar, D. B. Spalding, A calculation procedure for heat, mass and momentum transfer
in three-dimensional parabolic flows, International Journal of Heat and Mass Transfer 15 (1972)
1787–1806.
[11] O. Turan, N. Chakraborty, R. J. Poole, Laminar natural convection of Bingham fluids in a square
enclosure with differentially heated side walls, Journal of Non-Newtonian Fluid Mechanics 165
(2010) 901 – 913.
[12] O. Turan, N. Chakraborty, R. J. Poole, Laminar Rayleigh-Benard convection of yield stress fluids
in a square enclosure, Journal of Non-Newtonian Fluid Mechanics 171-172 (2012) 83 – 96.
[13] P. Neofytou, D. Drikakis, Effects of blood models on flows through a stenosis, International Journal
for Numerical Methods in Fluids 43 (2003) 597–635.
22
[14] P. Neofytou, D. Drikakis, Non-Newtonian flow instability in a channel with a sudden expansion,
Journal of Non-Newtonian Fluid Mechanics 111 (2003) 127 – 150.
[15] P. R. de Souza Mendes, M. F. Naccache, P. R. Varges, F. H. Marchesini, Flow of viscoplastic
liquids through axisymmetric expansions-contractions, Journal of Non-Newtonian Fluid Mechanics
142 (2007) 207 – 217.
[16] M. F. Naccache, R. S. Barbosa, Creeping flow of viscoplastic materials through a planar expansion
followed by a contraction, Mechanics Research Communications 34 (2007) 423 – 431.
[17] O. Botella, R. Peyret, Benchmark spectral results on the lid-driven cavity flow, Computers &
Fluids 27 (1998) 421 – 433.
[18] A. Syrakos, A. Goulas, Finite volume adaptive solutions using SIMPLE as smoother, International
Journal for Numerical Methods in Fluids 52 (2006) 1215–1245.
[19] C.-H. Bruneau, M. Saad, The 2D lid-driven cavity problem revisited, Computers & Fluids 35
(2006) 326 – 348.
[20] F. Sanchez, Application of a first-order operator splitting method to Bingham fluid flow simulation,
Computers & Mathematics with Applications 36 (1998) 71 – 86.
[21] E. Mitsoulis, T. Zisis, Flow of Bingham plastics in a lid-driven square cavity, Journal of NonNewtonian Fluid Mechanics 101 (2001) 173–180.
[22] D. Vola, L. Boscardin, J. Latch, Laminar unsteady flows of Bingham fluids: a numerical strategy
and some benchmark results, Journal of Computational Physics 187 (2003) 441 – 456.
[23] R. Elias, M. Martins, A. Coutinho, Parallel edge-based solution of viscoplastic flows with the
SUPG/PSPG formulation, Computational Mechanics 38 (2006) 365–381.
[24] Z. Yu, A. Wachs, A fictitious domain method for dynamic simulation of particle sedimentation in
Bingham fluids, Journal of Non-Newtonian Fluid Mechanics 145 (2007) 78–91.
[25] M. A. Olshanskii, Analysis of semi-staggered finite-difference method with application to Bingham
flows, Computer Methods in Applied Mechanics and Engineering 198 (2009) 975 – 985.
[26] J. Zhang, An augmented Lagrangian approach to Bingham fluid flows in a lid-driven square cavity
with piecewise linear equal-order finite elements, Computer Methods in Applied Mechanics and
Engineering 199 (2010) 3051 – 3057.
[27] D. D. dos Santos, S. Frey, M. F. Naccache, P. de Souza Mendes, Numerical approximations for flow
of viscoplastic fluids in a lid-driven cavity, Journal of Non-Newtonian Fluid Mechanics 166 (2011)
667 – 679.
[28] R. P. Fedorenko, A relaxation method for solving elliptic difference equations, USSR Computational
Mathematics and Mathematical Physics 1 (1962) 1092–1096.
[29] A. Brandt, Multi-level adaptive solutions to boundary-value problems, Mathematics of Computation 31 (1977) 333–390.
23
[30] S. Sivaloganathan, G. J. Shaw, A multigrid method for recirculating flows, International Journal
for Numerical Methods in Fluids 8 (1988) 417–440.
[31] M. Hortmann, M. Peric, S. G., Finite volume multigrid prediction of laminar natural convection:
benchmark solutions, International Journal for Numerical Methods in Fluids 11 (1990) 189–207.
[32] J. A. Tsamopoulos, M. E. Chen, A. V. Borkar, On the spin coating of viscoplastic fluids, Rheologica
Acta 35 (1996) 597–615.
[33] G. R. Burgos, A. N. Alexandrou, V. Entov, On the determination of yield surfaces in HerschelBulkley fluids, Journal of Rheology 43 (1999) 463–483.
[34] J. H. Ferziger, M. Peric, Computational methods for fluid dynamics, Springer, 3rd edition, 2002.
[35] A. Syrakos, A. Goulas, Estimate of the truncation error of finite volume discretization of the
Navier-Stokes equations on colocated grids, International Journal for Numerical Methods in Fluids
50 (2006) 103–130.
[36] C. M. Rhie, W. L. Chow, Numerical study of the turbulent flow past an airfoil with trailing edge
separation, AIAA Journal 21 (1983) 1525–1532.
[37] S. V. Patankar, Numerical heat transfer and fluid flow, Hemisphere, 1980.
[38] Y. Saad, Iterative Methods for Sparse Linear Systems, 2nd edition, SIAM, 2003.
[39] G. J. Shaw, S. Sivaloganathan, On the smoothing properties of the SIMPLE pressure-correction
algorithm, International Journal for Numerical Methods in Fluids 8 (1988) 441–461.
[40] G. R. Burgos, A. N. Alexandrou, Flow development of Herschel-Bulkley fluids in a sudden threedimensional square expansion, Journal of Rheology 43 (1999) 485–498.
24
| 5cs.CE
|
WRPN: Wide Reduced-Precision Networks
arXiv:1709.01134v1 [cs.CV] 4 Sep 2017
Asit Mishra Eriko Nurvitadhi Jeffrey J Cook Debbie Marr
Accelerator Architecture Lab, Intel
Abstract
For computer vision applications, prior works have shown the efficacy of reducing
numeric precision of model parameters (network weights) in deep neural networks.
Activation maps, however, occupy a large memory footprint during both the training and inference step when using mini-batches of inputs. One way to reduce this
large memory footprint is to reduce the precision of activations. However, past
works have shown that reducing the precision of activations hurts model accuracy.
We study schemes to train networks from scratch using reduced-precision activations without hurting accuracy. We reduce the precision of activation maps (along
with model parameters) and increase the number of filter maps in a layer, and find
that this scheme matches or surpasses the accuracy of the baseline full-precision
network. As a result, one can significantly improve the execution efficiency (e.g.
reduce dynamic memory footprint, memory bandwidth and computational energy)
and speed up the training and inference process with appropriate hardware support. We call our scheme WRPN - wide reduced-precision networks. We report
results and show that WRPN scheme is better than previously reported accuracies
on ILSVRC-12 dataset while being computationally less expensive compared to
previously reported reduced-precision networks.
1 Introduction
A promising approach to lower the compute and memory requirements of convolutional deeplearning workloads is through the use of low numeric precision algorithms. Operating in lower
precision mode reduces computation as well as data movement and storage requirements. Due to
such efficiency benefits, there are many existing works which propose low-precision deep neural
networks (DNNs) [25, 13, 15, 7, 22], even down to 2-bit ternary mode [27, 12, 23] and 1-bit binary
mode [26, 4, 17, 5, 21]. However, the majority of existing works in low-precision DNNs sacrifice
accuracy over the baseline full-precision networks. Further, most prior works target reducing the
precision of the model parameters (network weights). This primarily benefits the inference step only
when batch sizes are small.
To improve both execution efficiency and accuracy of low-precision networks, we reduce both the
precision of activation maps and model parameters and increase the number of filter maps in a layer.
We call networks using this scheme wide reduced-precision networks (WRPN) and find that this
scheme compensates or surpasses the accuracy of the baseline full-precision network. Although the
number of raw compute operations increases as we increase the number of filter maps in a layer, the
compute bits required per operation is now a fraction of what is required when using full-precision
operations (e.g. going from FP32 AlexNet to 4-bits precision and doubling the number of filters
increases the number of compute operations by 4x, but each operation is 8x more efficient than
FP32).
WRPN offers better accuracies, while being computationally less expensive compared to previously
reported reduced-precision networks. We report results on AlexNet [11], batch-normalized Inception [9], and ResNet-34 [8] on ILSVRC-12 [11] dataset. We find 4-bits to be sufficient for training
deep and wide models while achieving similar or better accuracy than baseline network. With 4-bit
activation and 2-bit weights, we find the accuracy to be at-par with baseline full-precision. Making
the networks wider and operating with 1-bit precision, we close the accuracy gap between previously
report binary networks and show state-of-the art results for ResNet-34 (69.85% top-1 with 2x wide)
and AlexNet (48.04% top-1 with 1.3x wide). To the best of our knowledge, our reported accuracies
with binary networks (and even 4-bit precision) are highest to date.
Our reduced-precision quantization scheme is hardware friendly allowing for efficient hardware
implementations. To this end, we evaluate efficiency benefits of low-precision operations (4-bits
to 1-bits) on Titan X GPU, Arria-10 FPGA and ASIC. We see that FPGA and ASIC can deliver
significant efficiency gain over FP32 operations (6.5x to 100x), while GPU cannot take advantage
of very low-precision operations.
2 Motivation for reduced-precision activation maps
!"#$%$"&$
!"#$%$%&
)'#!%
!"#$%
)!#)%
$+#"%
$,#,%
'$#+%
%8>?
%8-@2?
')#$%
'$#(%
',#,%
!)#*%
&'#"%
&(#&%
%89/9:4;8<::2=4312
%89/9:4;8<::2=4312
%8>?
%8-@2?
&#$%
$#'%
+)#*%
+,#$%
$&#(%
((#,%
,'#"%
'+#!%
'(#$%
'*#"%
!'#(%
!"#$%
*(#*%
&&#+%
&#'%
+
$)
-./01/2
+
$)
345)
!#!%
&#(%
(#&%
+
$)
+
4/61/27("
")#'%
!#&%
$)
"
&+
"
-./01/2
4/61/27+"+
&+
345+
"
&+
4/61/27$)
"
&+
4/61/27")"
Figure 1: Memory footprint of activations (ACTs) and weights (W) during training and inference for minibatch sizes 1 and 32.
While most prior works proposing reduced-precision networks work with low precision weights
(e.g. [4, 27, 26, 23, 12, 5, 21]), we find that activation maps occupy a larger memory footprint when
using mini-batches of inputs. Using mini-batches of inputs is typical in training of DNNs and cloudbased batched inference [10]. Figure 1 shows memory footprint of activation maps and filter maps
as batch size changes for 4 different networks (AlexNet, Inception-Resnet-v2 [19], ResNet-50 and
ResNet-101) during the training and inference steps.
W
W
OFM
/IFM
!
"
ACT
OFM
/IFM
!
"
ACT
Layer !"#
$!
Grad
W
OFM
/IFM
!
"
ACT
Layer !
$!
$"
OFM
/IFM
$!
$"
Grad
ACT
Layer !%#
Grad
$"
Grad
Figure 2: Memory requirements of a feed forward convolutional deep neural network. Orange boxes denote
weights (W), blue boxes are activations (ACT) and green boxes are gradient-maps (Grad).
As batch-size increases, because of filter reuse across batches of inputs, activation maps occupy
significantly larger fraction of memory compared to the filter weights. This aspect is illustrated in
Figure 2 which shows the memory requirements of a canonical feed-forward DNN for a hardware
accelerator based system (e.g. GPU, FPGA, PCIe connected ASIC device, etc.). During training,
the sum of all the activation maps (ACT) and weight tensors (W) are allocated in device memory for
forward pass along with memory for gradient maps during backward propagation. The total memory
requirements for training phase is the sum of memory required for the activation maps, weights and
the maximum of input gradient maps (δZ) and maximum of back-propagated gradients (δX). During
inference, memory is allocated for input (IFM) and output feature maps (OFM) required by a single
layer, and these memory allocations are reused for other layers. The total memory allocation during
inference is then the maximum of IFM and maximum of OFM required across all the layers plus the
sum of all W-tensors. At batch sizes 128 and more, activations start to occupy more than 98% of
total memory footprint during training.
2
Overall, reducing precision of activations and weights reduces memory footprint, bandwidth and
storage while also simplifying the requirements for hardware to efficiently support these operations.
3 WRPN scheme and studies on AlexNet
Based on the observation that activations occupy more memory footprint compared to weights, we
reduce the precision of activations to speed up training and inference steps as well as cut down on
memory requirements. However, a straightforward reduction in precision of activation maps leads
to significant reduction in model accuracy [26, 17].
We conduct a sensitivity study where we reduce precision of activation maps and model weights
for AlexNet running ILSVRC-12 dataset and train the network from scratch. Table 1 reports our
findings. Top-1 single-precision (32-bits weights and activations) accuracy is 57.2%. The accuracy
with binary weights and activations is 44.2%. This is similar to what is reported in [17]. 32bA and
2bW data-point in this table is using TTQ technique [27]. All other data points are collected using
our quantization scheme (described later in Section 5), all the runs have same hyper-parameters and
training is carried out for the same number of epochs as baseline network. To be consistent with
results reported in prior works, we do not quantize weights and activations of the first and last layer.
We find that, in general, reducing the precision of activation maps and weights hurts model accuracy.
Further, reducing precision of activations hurts model accuracy much more than reducing precision
of the filter parameters. We find TTQ to be quite effective on AlexNet in that one can lower the
precision of weights to 2b (while activations are still FP32) and not lose accuracy. However, we did
not find this scheme to be effective for other networks like ResNet or Inception.
Table 1: AlexNet top-1 validation set accuracy
% as precision of activations (A) and weight(W)
changes. All results are with end-to-end training
of the network from scratch. − is a data-point
we did not experiment for.
32b W
8b W
4b W
2b W
1b W
32b A
8b A
4b A
2b A
1b A
57.2
–
–
57.5
56.8
54.3
54.5
54.2
50.2
–
54.4
53.2
54.4
50.5
–
52.7
51.5
52.4
51.3
–
–
–
–
–
44.2
Table 2: AlexNet 2x-wide top-1 validation set
accuracy % as precision of activations (A) and
weights (W) changes.
32b W
8b W
4b W
2b W
1b W
32b A
8b A
4b A
2b A
1b A
60.5
–
–
–
–
58.9
59.0
58.8
57.6
–
58.6
58.8
58.6
57.2
–
57.5
57.1
57.3
55.8
–
52.0
50.8
–
–
48.3
To re-gain the model accuracy while working with reduced-precision operands, we increase the
number of filter maps in a layer. Although the number of raw compute operations increase with
widening the filter maps in a layer, the bits required per compute operation is now a fraction of what
is required when using full-precision operations. As a result, with appropriate hardware support,
one can significantly reduce the dynamic memory requirements, memory bandwidth, computational
energy and speed up the training and inference process.
Our widening of filter maps is inspired from Wide ResNet [24] work where the depth of the network
is reduced and width of each layer is increased (the operand precision is still FP32). Wide ResNet requires a re-design of the network architecture. In our work, we maintain the depth parameter same as
baseline network but widen the filter maps. We call our approach WRPN - wide reduced-precision
networks. In practice, we find this scheme to be very simple and effective - starting with a baseline network architecture, one can change the width of each filter map without changing any other
network design parameter or hyper-parameters. Carefully reducing precision and simultaneously
widening filters keeps the total compute cost of the network under or at-par with baseline cost.1
Table 2 reports the accuracy of AlexNet when we double the number of filter maps in a layer. With
doubling of filter maps, AlexNet with 4-bits weights and 2-bits activations exhibits accuracy at-par
with full-precision networks. Operating with 4-bits weights and 4-bits activations surpasses the
baseline accuracy by 1.44%. With binary weights and activations we better the accuracy of XNORNET [17] by 4%.
1
Compute cost is the product of the number of FMA operations and the sum of width of the activation and
weight operands.
3
When doubling the number of filter maps, AlexNet’s raw compute operations grow by 3.9x compared to the baseline full-precision network, however by using reduced-precision operands the overall compute complexity is a fraction of the baseline. For example, with 4b operands for weights and
activations and 2x the number of filters, reduced-precision AlexNet is just 49% of the total compute
cost of the full-precision baseline (compute cost comparison is shown in Table 3).
Table 3: Compute cost of AlexNet 2x-wide vs. 1x-wide as precision of activations (A) and weights (W) changes.
32b W
8b W
4b W
2b W
1b W
32b A
8b A
4b A
2b A
1b A
3.9x
2.4x
2.2x
2.1x
2.0x
2.4x
1.0x
0.7x
0.6x
0.6x
2.2x
0.7x
0.5x
0.4x
0.3x
2.1x
0.6x
0.4x
0.2x
0.2x
2.0x
0.6x
0.3x
0.2x
0.1x
We also experiment with other widening factors. With 1.3x widening of filters and with 4-bits of
activation precision one can go as low as 8-bits of weight precision while still being at-par with
baseline accuracy. With 1.1x wide filters, at least 8-bits weight and 16-bits activation precision
is required for accuracy to match baseline full-precision 1x wide accuracy. Further, as Table 3
shows, when widening filters by 2x, one needs to lower precision to at least 8-bits so that the total
compute cost is not more than baseline compute cost. Thus, there is a trade-off between widening
and reducing the precision of network parameters.
In our work, we trade-off higher number of raw compute operations with aggressively reducing the
precision of the operands involved in these operations (activation maps and filter weights) while
not sacrificing the model accuracy. Apart from other benefits of reduced precision activations as
mentioned earlier, widening filter maps also improves the efficiency of underlying GEMM calls for
convolution operations since compute accelerators are typically more efficient on a single kernel consisting of parallel computation on large data-structures as opposed to many small sized kernels [24].
4 Studies on deeper networks
We study how our scheme applies to deeper networks. For this, we study ResNet-34 [8] and batchnormalized Inception [9] and find similar trends, particularly that 2-bits weight and 4-bits activations
continue to provide at-par accuracy as baseline. We use TensorFlow [2] and tensorpack [1] for all
our evaluations and use ILSVRC-12 train and val dataset for analysis.2
4.1 ResNet
ResNet-34 has 3x3 filters in each of its modular layers with shortcut connections being 1x1. The
filter bank width changes from 64 to 512 as depth increases. We use the pre-activation variant of
ResNet and the baseline top-1 accuracy of our ResNet-34 implementation using single-precision
32-bits data format is 73.59%. Binarizing weights and activations for all layers except the first and
the last layer in this network gives top-1 accuracy of 60.5%. For binarizing ResNet we did not reorder any layer (as is done in XNOR-NET). We used the same hyper-parameters and learning rate
schedule as the baseline network. As a reference, for ResNet-18, the gap between XNOR-NET (1b
weights and activations) and full-precision network is 18% [17]. It is also interesting to note that
top-1 accuracy of single-precision AlexNet (57.20%) is lower than the top-1 accuracy of binarized
ResNet-34 (60.5%).
We experimented with doubling number of filters in each layer and reduce the precision of activations and weights. Table 4 shows the results of our analysis. Doubling the number of filters and
4-bits precision for both weights and activations beats the baseline accuracy by 0.9%. 4-bits activations and 2-bits (ternary) weights has top-1 accuracy at-par with baseline. Reducing precision to
2-bits for both weights and activations degrades accuracy by only 0.2% compared to baseline.
2
We will open-source our implementation of reduced-precision AlexNet, ResNet and batch-normalized
Inception networks.
4
Table 4: ResNet-34 top-1 validation accuracy % and compute cost as
precision of activations (A) and weights (W) varies.
Width
Precision
Top-1 Acc. %
Compute cost
1x wide
32b A, 32b W
1b A, 1b W
73.59
60.54
1x
0.03x
2x wide
4b A, 8b W
4b A, 4b W
4b A, 2b W
2b A, 4b W
2b A, 2b W
1b A, 1b W
74.48
74.52
73.58
73.50
73.32
69.85
0.74x
0.50x
0.39x
0.39x
0.27x
0.15x
3x wide
1b A, 1b W
72.38
0.30x
Binarizing the weights and activations with 2x wide filters has a top-1 accuracy of 69.85%. This
is just 3.7% worse than baseline full-precision network while being only 15% of the cost of the
baseline network. Widening the filters by 3x and binarizing the weights and activations reduces this
gap to 1.2% while the 3x wide network is 30% the cost of the full-precision baseline network.
Although 4-bits precision seems to be enough for wide networks, we advocate for 4-bits activation precision and 2-bits weight precision. This is because with ternary weights one can get rid
of the multipliers and use adders instead. Additionally, with this configuration there is no loss of
accuracy. Further, if some accuracy degradation is tolerable, one can even go to binary circuits for
efficient hardware implementation while saving 32x in bandwidth for each of weights and activations compared to full-precision networks. All these gains can be realized with simpler hardware
implementation and lower compute cost compared to baseline networks.
To the best of our knowledge, our ResNet binary and ternary (with 2-bits or 4-bits activation) top-1
accuracies are state-of-the-art results in the literature including unpublished technical reports (with
similar data augmentation [14]).
4.2 Batch-normalized Inception
We applied WRPN scheme to batch-normalized Inception network [9]. This network includes batch
normalization of all layers and is a variant of GoogleNet [20] where the 5x5 convolutional filters
are replaced by two 3x3 convolutions with up to 128 wide filters. Table 5 shows the results of
our analysis. Using 4-bits activations and 2-bits weight and doubling the number of filter banks in
the network produces a model that is almost at-par in accuracy with the baseline single-precision
network (0.02% loss in accuracy). Wide network with binary weights and activations is within 6.6%
of the full-precision baseline network.
Table 5: Batch-normalized Inception top-1 validation accuracy % and
compute cost as precision of activations (A) and weights (W) varies.
Width
Precision
Top-1 Acc. %
Compute cost
1x wide
32b A, 32b W
71.64
1x
2x wide
4b A, 4b W
4b A, 2b W
2b A, 2b W
1b A, 1b W
71.63
71.61
70.75
65.02
0.50x
0.38x
0.25x
0.13x
5 Hardware friendly quantization scheme
We adopt the straight-through estimator (STE) approach in our work [3]. When quantizing a real
number to k-bits, the ordinality of the set of quantized numbers is 2k . Mathematically, this small
and finite set would have zero gradients with respect to its inputs. STE method circumvents this
problem by defining an operator that has arbitrary forward and backward operations.
5
Prior works using the STE approach define operators that quantize the weights based on the expectation of the weight tensors. For instance, TWN [12] uses a threshold and a scaling factor for each
layer to quantize weights to ternary domain. In TTQ [27], the scaling factors are learned parameters.
XNOR-NET binarizes the weight tensor by computing the sign of the tensor values and then scaling
by the mean of the absolute value of each output channel of weights. DoReFa uses a single scaling
factor across the entire layer. For quantizing weights to k-bits, where k > 1, DoReFa uses:
wk = 2 ∗ quantizek (
tanh(wi )
1
+ ) − 1)
2 ∗ max(| tanh(wi ) |) 2
(1)
Here wk is the k-bit quantized version of inputs wi and quantizek is a quantization function that
quantizes a floating-point number wi in the range [0, 1] to a k-bit number in the same range. The
transcendental tanh operation constrains the weight value to lie in between −1 and +1. The affine
transformation post quantization brings the range to [−1, 1].
We build on these approaches and propose a much simpler scheme. For quantizing weight tensors we first hard constrain the values to lie within the range [−1, 1] using min-max operation (e.g.
tf.clip_by_val when using Tensorflow [2]). For quantizing activation tensor values, we constrain the
values to lie within the range [0, 1]. This step is followed by a quantization step where a real number
is quantized into a k-bit number. This is given as, for k > 1:
wk =
1
2k−1
−1
round((2k−1 − 1) ∗ wi )
and
ak =
2k
1
round((2k − 1) ∗ ai )
−1
(2)
Here wi and ai are input real-valued weights and activation tensor and wk and ak are their quantized versions. One bit is reserved for sign-bit in case of weight values, hence the use of 2k−1 for
these quantized values. Thus, weights can be stored and interpreted using signed data-types and
activations using un-signed data-types. With appropriate affine transformations, the convolution operations (the bulk of the compute operations in the network during forward pass) can be done using
quantized values (integer operations in hardware) followed by scaling with floating-point constants
(this scaling operation can be done in parallel with the convolution operation in hardware). When
k = 1, for binary weights we use the BWN approach [5] where the binarized weight value is computed based on the sign of input value followed by scaling with the mean of absolute values. For
binarized activations we use the formulation in Eq. 2. We do not quantize the gradients and maintain
the weights in reduced precision format.
For convolution operation when using WRPN, the forward pass during training (and the inference
step) involves matrix multiplication of k-bits signed and k-bits unsigned operands. Since gradient values are in 32-bits floating-point format, the backward pass involves a matrix multiplication
operation using 32-bits and k-bits operand for gradient and weight update.
When k > 1, the hard clipping of tensors to a range maps efficiently to min-max comparator units
in hardware as opposed to using transcendental operations which are long latency operations. TTQ
and DoRefa schemes involve division operation and computing a maximum value in the input tensor.
Floating-point division operation is expensive in hardware and computing the maximum in a tensor
is an O(n) operation. Additionally, our quantization parameters are static and do not require any
learning or involve back-propagation like TTQ approach. We avoid each of these costly operations
and propose a simpler quantization scheme (clipping followed by rounding).
5.1 Efficiency improvements of reduced-precision operations on GPU, FPGA and ASIC
In practice, the effective performance and energy efficiency one could achieve on a low-precision
compute operation highly depends on the hardware that runs these operations. We study the efficiency of low-precision operations on various hardware targets – GPU, FPGA, and ASIC.
For GPU, we evaluate WRPN on Nvidia Titan X Pascal and for FPGA we use Intel Arria-10. We
collect performance numbers from both previously reported analysis [16] as well as our own experiments. For FPGA, we implement a DNN accelerator architecture shown in Figure 3(a). This
is a prototypical accelerator design used in various works (e.g., on FPGA [16] and ASIC such as
TPU [10]). The core of the accelerator consists of a systolic array of processing elements (PEs)
6
to perform matrix and vector operations, along with on-chip buffers, as well as off-chip memory
management unit. The PEs can be configured to support different precision – (FP32, FP32), (INT4,
INT4), (INT4, TER2), and (BIN1, BIN1). The (INT4, TER2) PE operates on ternary (+1,0,-1) values and is optimized to include only an adder since there is no need for a multiplier in this case.
The binary (BIN1, BIN1) PE is implemented using XNOR and bitcount. Our RTL design targets
Arria-10 1150 FPGA. For our ASIC study, we synthesize the PE design using Intel 14 nm process
technology to obtain area and energy estimates.
/
45
45
/
!"#$%&'( )*++,-.
<=>$9'+0$
%7<=$)+$
?@AB
!"##$%"&'()#*'+,-.
45
!"##$%"&'()#*'+,-.
45
%76;$(.--,/.($1'-$)'15H&2I$B() +',-'$
-()&01)-(C$(&25-$&)($91"'&5$&($1$I++,$9&)$
9+'$5/()+0$:+3*.'-5&(&+2$+.-'1)&+2(
!"##$%"&'()#*'+,-.
@2$.'15)&5-C$+2:D$/.$)+$EF>$
(.--,/.($&2$678C$(&25-$2+$
(/..+')$9+'$(/"$G*"&)$.'-5&(&+2(
0--12(3+(45.
/
/
/
6,7(89-:
!"#$%&'()*+',-'$
(.--,/.$-()&01)-($
"1(-,$+2$"&)*3&,)4
!-#$678$12,$%76;$.-'9+'0125-$12,$-2-'ID$-99&5&-25D$9+'$K1'&+/($
.'-5&(&+2(L$%76;$,+-($3-::$+2$K-'D$:+3$.'-5&(&+2(L
=0"*()#0#2>&'()#*'+,-.
BF<F
!,#$%76;$(.--,/.($
9+'$:+3*.'-5&(&+2$
+.-'1)&+2(
=0"*()#0#2>&'()#*'+,-.
!5#$ 678$(.--,/.($
9+'$:+3*.'-5&(&+2$
+.-'1)&+2(
,#*/(*0123#452#*67'
89:,4&4;<
,#*/(*0123#'8?:,4&<
!1#$JAA$ 41',31'-$
/2,-'$()/,D
!9#$;M@N$5+0./)-$
/2&)$.-'9+'0125-
!I#$;M@N$5+0./)-$
/2&)$-2-'ID
Figure 3: Efficiency improvements from low-precision operations on GPU, FPGA and ASIC.
Figure 3(b) - (g) summarize our analysis. Figure 3(b) shows the efficiency improvements using
first-order estimates where the efficiency is computed based on number of bits used in the operation.
With this method we would expect (INT4, INT4) and (BIN1, BIN1) to be 8x and 32x more efficient,
respectively, than (FP32, FP32). However, in practice the efficiency gains from reducing precision
depend on whether the underlying hardware can take advantage of such low-precisions.
Figure 3(c) shows performance improvement on Titan X GPU for various low-precision operations
relative to FP32. In this case, GPU can only achieve up to ∼4x improvements in performance over
FP32 baseline. This is because GPU only provides first-class support for INT8 operations, and is not
able to take advantage of the lower INT4, TER2, and BIN1 precisions. On the contrary, FPGA can
take advantage of such low precisions, since they are amenable for implementations on the FPGA’s
reconfigurable fabric.
Figure 3(d) shows that the performance improvements from (INT4, INT4), (INT4, TER2), and
(BIN1, BIN1) track well with the first-order estimates from Figure 3(b). In fact, for (BIN1, BIN1),
FPGA improvements exceed the first-order estimate. Reducing the precision simplifies the design
of compute units and lower buffering requirements on FPGA board. Compute-precision reduction
leads to significant improvement in throughput due to smaller hardware designs (allowing more parallelism) and shorter circuit delay (allowing higher frequency). Figure 3(e) shows the performance
and performance/Watt of the reduced-precision operations on GPU and FPGA. FPGA performs
quite well on very low precision operations. In terms of performance/watt, FPGA does better than
GPU on (INT4, INT4) and lower precisions.
7
ASIC allows for a truly customized hardware implementation. Our ASIC study provides insights to
the upper bound of the efficiency benefits possible from low-precision operations. Figure 3(f) and
3(g) show improvement in performance and energy efficiency of the various low-precision ASIC
PEs relative to baseline FP32 PE. As the figures show, going to lower precision offers 2 to 3 orders
of magnitude efficiency improvements.
In summary, FPGA and ASIC are well suited for our WRPN approach. At 2x wide, our WRPN
approach requires 4x more total operations than the original network. However, for INT4 or lower
precision, each operation is 6.5x or better in efficiency than FP32 for FPGA and ASIC. Hence,
WRPN delivers an overall efficiency win.
6 Related work
Reduced-precision DNNs is an active research area. Reducing precision of weights for efficient
inference pipeline has been very well studied. Works like Binary connect (BC) [5], Ternary-weight
networks (TWN) [12], fine-grained ternary quantization [14] and INQ [25] target reducing the precision of network weights while still using full-precision activations. Accuracy is almost always
degraded when quantizing the weights. For AlexNet on Imagenet, TWN loses 5% top-1 accuracy.
Schemes like INQ, [18] and [14] do fine-tuning to quantize the network weights and do not sacrifice
accuracy as much but are not applicable for training networks from scratch. INQ shows promising
results with 5-bits of precision.
XNOR-NET [17], BNN [4], DoReFa [26] and TTQ [27] target training as well. While TTQ targets
weight quantization only, most works targeting activation quantization hurt accuracy. XNOR-NET
approach reduces top-1 accuracy by 12% and DoReFa by 8% when quantizing both weights and
activations to 1-bit (for AlexNet on ImageNet). Further, XNOR-NET requires re-ordering of layers
for its scheme to work. Recent work in [6] targets low-precision activations
√ and reports accuracy
within 1% of baseline with 5-bits precision and logarithmic (with base 2) quantization. With
fine-tuning this gap can be narrowed to be within 0.6% but not all layers are quantized.
Non-multiples of two for operand values introduces hardware inefficiency in that memory accesses
are no longer DRAM or cache-boundary aligned and end-to-end run-time performance aspect is unclear when using complicated quantization schemes. We target end-to-end training and inference,
using very simple quantization method and aim for reducing precision without any loss in accuracy.
To the best of our knowledge, our work is the first to study reduced-precision deep and wide networks, and show accuracy at-par with baseline for as low a precision as 4-bits activations and 2-bits
weights. We report state of the art accuracy for wide binarized AlexNet and ResNet while still being
lower in compute cost.
7 Conclusions
We present the Wide Reduced-Precision Networks (WRPN) scheme for DNNs. In this scheme, the
numeric precision of both weights and activations are significantly reduced without loss of network
accuracy. This result is in contrast to many previous works that find reduced-precision activations to
detrimentally impact accuracy; specifically, we find that 2-bit weights and 4-bit activations are sufficient to match baseline accuracy across many networks including AlexNet, ResNet-34 and batchnormalized Inception. We achieve this result with a new quantization scheme and by increasing
the number of filter maps in each reduced-precision layer to compensate for the loss of information
capacity induced by reducing the precision.
We motivate this work with our observation that full-precision activations contribute significantly
more to the memory footprint than full-precision weight parameters when using mini-batch sizes
common during training and cloud-based inference; furthermore, by reducing the precision of both
activations and weights the compute complexity is greatly reduced (40% of baseline for 2-bit weights
and 4-bit activations).
The WRPN quantization scheme and computation on low precision activations and weights is hardware friendly making it viable for deeply-embedded system deployments as well as in cloud-based
training and inference servers with compute fabrics for low-precision. We compare Titan X GPU,
Arria-10 FPGA and ASIC implementations using WRPN and show our scheme increases perfor8
mance and energy-efficiency for iso-accuracy across each. Overall, reducing the precision allows
custom-designed compute units and lower buffering requirements to provide significant improvement in throughput.
9
References
[1] https://github.com/ppwwyyxx/tensorpack.
[2] M. Abadi, A. Agarwal, P. Barham, E. Brevdo, Z. Chen, C. Citro, G. S. Corrado, A. Davis, J. Dean,
M. Devin, S. Ghemawat, I. Goodfellow, A. Harp, G. Irving, M. Isard, Y. Jia, R. Jozefowicz, L. Kaiser,
M. Kudlur, J. Levenberg, D. Mané, R. Monga, S. Moore, D. Murray, C. Olah, M. Schuster, J. Shlens,
B. Steiner, I. Sutskever, K. Talwar, P. Tucker, V. Vanhoucke, V. Vasudevan, F. Viégas, O. Vinyals, P. Warden, M. Wattenberg, M. Wicke, Y. Yu, and X. Zheng. TensorFlow: Large-scale machine learning on
heterogeneous systems, 2015. Software available from tensorflow.org.
[3] Y. Bengio, N. Léonard, and A. C. Courville. Estimating or propagating gradients through stochastic
neurons for conditional computation. CoRR, abs/1308.3432, 2013.
[4] M. Courbariaux and Y. Bengio. Binarynet: Training deep neural networks with weights and activations
constrained to +1 or -1. CoRR, abs/1602.02830, 2016.
[5] M. Courbariaux, Y. Bengio, and J. David. Binaryconnect: Training deep neural networks with binary
weights during propagations. CoRR, abs/1511.00363, 2015.
[6] B. Graham. Low-precision batch-normalized activations. CoRR, abs/1702.08231, 2017.
[7] S. Gupta, A. Agrawal, K. Gopalakrishnan, and P. Narayanan. Deep learning with limited numerical
precision. CoRR, abs/1502.02551, 2015.
[8] K. He, X. Zhang, S. Ren, and J. Sun. Deep residual learning for image recognition. CoRR,
abs/1512.03385, 2015.
[9] S. Ioffe and C. Szegedy. Batch normalization: Accelerating deep network training by reducing internal
covariate shift. CoRR, abs/1502.03167, 2015.
[10] N. P. Jouppi, C. Young, N. Patil, D. Patterson, G. Agrawal, R. Bajwa, S. Bates, S. Bhatia, N. Boden,
A. Borchers, R. Boyle, P.-l. Cantin, C. Chao, C. Clark, J. Coriell, M. Daley, M. Dau, J. Dean, B. Gelb,
T. Vazir Ghaemmaghami, R. Gottipati, W. Gulland, R. Hagmann, C. R. Ho, D. Hogberg, J. Hu, R. Hundt,
D. Hurt, J. Ibarz, A. Jaffey, A. Jaworski, A. Kaplan, H. Khaitan, A. Koch, N. Kumar, S. Lacy, J. Laudon,
J. Law, D. Le, C. Leary, Z. Liu, K. Lucke, A. Lundin, G. MacKean, A. Maggiore, M. Mahony, K. Miller,
R. Nagarajan, R. Narayanaswami, R. Ni, K. Nix, T. Norrie, M. Omernick, N. Penukonda, A. Phelps,
J. Ross, M. Ross, A. Salek, E. Samadiani, C. Severn, G. Sizikov, M. Snelham, J. Souter, D. Steinberg,
A. Swing, M. Tan, G. Thorson, B. Tian, H. Toma, E. Tuttle, V. Vasudevan, R. Walter, W. Wang, E. Wilcox,
and D. H. Yoon. In-Datacenter Performance Analysis of a Tensor Processing Unit. ArXiv e-prints, Apr.
2017.
[11] A. Krizhevsky, I. Sutskever, and G. E. Hinton. Imagenet classification with deep convolutional neural
networks. In F. Pereira, C. J. C. Burges, L. Bottou, and K. Q. Weinberger, editors, Advances in Neural
Information Processing Systems 25, pages 1097–1105. Curran Associates, Inc., 2012.
[12] F. Li and B. Liu. Ternary weight networks. CoRR, abs/1605.04711, 2016.
[13] Z. Lin, M. Courbariaux, R. Memisevic, and Y. Bengio. Neural networks with few multiplications. CoRR,
abs/1510.03009, 2015.
[14] N. Mellempudi, A. Kundu, D. Mudigere, D. Das, B. Kaul, and P. Dubey. Ternary Neural Networks with
Fine-Grained Quantization. ArXiv e-prints, May 2017.
[15] D. Miyashita, E. H. Lee, and B. Murmann. Convolutional neural networks using logarithmic data representation. CoRR, abs/1603.01025, 2016.
[16] E. Nurvitadhi, G. Venkatesh, J. Sim, D. Marr, R. Huang, J. Ong Gee Hock, Y. T. Liew, K. Srivatsan,
D. Moss, S. Subhaschandra, and G. Boudoukh. Can fpgas beat gpus in accelerating next-generation
deep neural networks? In Proceedings of the 2017 ACM/SIGDA International Symposium on FieldProgrammable Gate Arrays, FPGA ’17, pages 5–14, New York, NY, USA, 2017. ACM.
[17] M. Rastegari, V. Ordonez, J. Redmon, and A. Farhadi. Xnor-net: Imagenet classification using binary
convolutional neural networks. CoRR, abs/1603.05279, 2016.
[18] W. Sung, S. Shin, and K. Hwang. Resiliency of deep neural networks under quantization. CoRR,
abs/1511.06488, 2015.
[19] C. Szegedy, S. Ioffe, and V. Vanhoucke. Inception-v4, inception-resnet and the impact of residual connections on learning. CoRR, abs/1602.07261, 2016.
[20] C. Szegedy, W. Liu, Y. Jia, P. Sermanet, S. E. Reed, D. Anguelov, D. Erhan, V. Vanhoucke, and A. Rabinovich. Going deeper with convolutions. CoRR, abs/1409.4842, 2014.
[21] Y. Umuroglu, N. J. Fraser, G. Gambardella, M. Blott, P. H. W. Leong, M. Jahre, and K. A. Vissers. FINN:
A framework for fast, scalable binarized neural network inference. CoRR, abs/1612.07119, 2016.
[22] V. Vanhoucke, A. Senior, and M. Z. Mao. Improving the speed of neural networks on cpus. In Deep
Learning and Unsupervised Feature Learning Workshop, NIPS 2011, 2011.
[23] G. Venkatesh, E. Nurvitadhi, and D. Marr. Accelerating deep convolutional networks using low-precision
and sparsity. CoRR, abs/1610.00324, 2016.
[24] S. Zagoruyko and N. Komodakis. Wide residual networks. CoRR, abs/1605.07146, 2016.
[25] A. Zhou, A. Yao, Y. Guo, L. Xu, and Y. Chen. Incremental network quantization: Towards lossless cnns
with low-precision weights. CoRR, abs/1702.03044, 2017.
10
[26] S. Zhou, Z. Ni, X. Zhou, H. Wen, Y. Wu, and Y. Zou. Dorefa-net: Training low bitwidth convolutional
neural networks with low bitwidth gradients. CoRR, abs/1606.06160, 2016.
[27] C. Zhu, S. Han, H. Mao, and W. J. Dally. Trained ternary quantization. CoRR, abs/1612.01064, 2016.
11
| 9cs.NE
|
THE EVENTUAL SHAPE OF BETTI TABLES OF POWERS OF IDEALS
arXiv:1110.0383v3 [math.AC] 1 Jul 2013
AMIR BAGHERI, MARC CHARDIN, AND HUY TÀI HÀ
Abstract. Let G be an abelian group and S be a G-graded a Noetherian algebra over a
commutative ring A ⊆ S0 . Let I1 , . . . , Is be G-homogeneous ideals in S, and let M be a
finitely generated G-graded S-module. We show that the shape of nonzero G-graded Betti
numbers of M I1t1 . . . Ists exhibit an eventual linear behavior as the ti s get large.
1. Introduction
It is a celebrated result (cf. [5, 12, 15]) that if I ⊆ S is a homogeneous ideal in a
Noetherian standard N-graded algebra and M is a finitely generated Z-graded S-module
then the regularity reg(I t M) is asymptotically a linear function for t ≫ 0. This asymptotic
linear function and the stabilization index have also been studied in [1, 4, 7, 8, 9].
In the case S is a polynomial ring over a field, a more precise result is proved in [5]: the
maximal degree of an i-th syzygy of I t M is eventually a linear function of t. Our interest
here is to understand the eventual behavior of the degrees of all the minimal generators of
the i-th syzygy module. This is of particular interest when the grading is given by some
finitely generated abelian group G that is not Z, as in this case the result for the regularity
of powers do not have an evident analogue. One can in particular consider the Cox ring of
a toric variety, graded by the divisor class group.
We shall actually show, in the G-graded setting, that the collection of nonzero G-graded
Betti numbers of MI1t1 . . . Ists exhibits an asymptotic linear behavior as the ti s get large.
Let us also point out that, even when an explicit minimal free resolution of these powers is
known, for instance when I is a complete intersection ideal, the degrees of i-th syzygies of
I t do not exhibit trivially a linear behavior.
Throughout the paper, let G be a finitely generated abelian group and let S = A[x1 , . . . , xn ]
be a G-graded algebra over a commutative ring A ⊆ S0 . Hence A = S/(x1 , . . . , xn ) is a Ggraded S-module concentrated in degree 0.
Our work hinges on the relationship between multigraded Betti numbers and the graded
pieces of TorSi (MI1t1 . . . Ists , A); we, thus, examine the support of TorSi (MI1t1 · · · Ists , A) as the
ti s get large. Our main result, in the case of a single ideal, gives the following:
2000 Mathematics Subject Classification. 13D45, 13D02.
Key words and phrases. Betti numbers, asymptotic linearity, multigraded.
1
2
AMIR BAGHERI, MARC CHARDIN, AND HUY TÀI HÀ
Theorem 1.1 (see Theorem 4.6). Let I = (f1 , . . . , fr ) be a G-homogeneous S-ideal, and let
M be a finitely generated G-graded S-module. Set Γ := {degG (fi )}ri=1 .
Let ℓ ≥ 0, and assume that ℓ = 0 or A is Noetherian.
There exist a finite collection of elements δi ∈ G, a finite collection of integers ti , and
a finite collection of non-empty tuples Ei = (γi,1 , . . . , γi,si ) of elements in Γ, such that the
elements (γi,j+1 − γi,j , 1 ≤ j ≤ si − 1) are linearly independent, satisfying
SuppG (TorSℓ (MI t , A)) =
m
[
i=1
[
c1 +···+csi =t−ti ,
c1 ,...,csi ∈N
c1 γi,1 + · · · + csi γi,si + δi ,
∀t ≥ max{ti }.
i
The condition of linear independence stated for Ei implies that c1 γi,1 + · · · + csi γi,si 6=
c′1 γi,1 + · · · + c′si γi,si if c1 + · · · + csi = c′1 + · · · + c′si and (c1 , . . . , csi ) 6= (c′1 , . . . , c′si ). Notice
the important fact that the elements in Ei are all in Γ, the set of degrees of generators of I.
Theorem 4.6 is proved in the last section of the paper. Our proof is based on the two
following observations. Firstly,
module (sometimes also referred to as the
L the multi-Rees
t1
ts
Rees modification) MR =
MI
.
.
.
I
is
a G × Zs -graded module over the multi1
s
ti ≥0
L
Rees algebra R = ti ≥0 I1t1 . . . Ists . Secondly, if G′ is a finitely generated abelian group and
R = S[T1 , . . . , Tr ] is a G × G′ -graded polynomial extension of S, such that degG×G′ (a) =
(degG (a), 0) for any a ∈ S, then for a graded complex G• of free R-modules,
Hi (G• )G×{δ} ⊗S A = Hi G• ⊗S A G×{δ} ,
where (•)G×{δ} denotes the degree G × {δ}-strand of the corresponding complex. In particular, if G• is a G × G′ -graded free resolution of an R-module N, as (G• )G×{δ} is a G-graded
free resolution of the S-module NG×{δ} , it follows that
TorSi (NG×{δ} , A) = Hi (G• ⊗S A)G×{δ} ,
in which G• ⊗S A is viewed as a G × G′ -graded complex of free A[T1 , . . . , Tr ]-modules. These
observations allow us to bring the problem to studying the support of graded A[T1 , . . . , Tr ]modules. We proceed by making use of the notion of initial submodules to reduce to the case
when the module is a quotient ring obtained by a monomial ideal. The result then follows
by considering the Stanley decomposition of such a quotient ring.
In its full generality, our proof of Theorem 4.6 is quite technical, so we start in Section
3 by considering first the case when each Ii is equi-generated. In fact, in this case, with
the additional assumption that A is a Noetherian ring, our results are much stronger; the
asymptotic linearity appears clearer and the proofs are simpler. We can also examine the
support of local cohomology modules of MI1t1 . . . Ists . Our results, in the case when each Ii
is finitely generated in a single degree γi , are stated as follows.
Theorem 1.2 (Theorem 3.3). Assume that i = 0 or A is Noetherian. Then there exists a
finite set ∆i ⊆ G such that
THE EVENTUAL SHAPE OF BETTI TABLES OF POWERS OF IDEALS
3
(1) For all t = (t1 , . . . , ts ) ∈ Ns , TorSi (MI1t1 · · · Ists , A)η = 0 if η 6∈ ∆i + t1 γ1 + · · · + ts γs .
(2) There exists a subset ∆′i ⊆ ∆i such that TorSi (MI1t1 · · · Ists , A)η+t1 γ1 +···+ts γs 6= 0 for
t ≫ 0 and η ∈ ∆′i , and TorSi (MI1t1 · · · Ists , A)η+t1 γ1 +···+ts γs = 0 for t ≫ 0 and η 6∈ ∆′i .
(3) Let A → k be a ring homomorphism to a field k. Then for any δ and any j, the
function
t1
S
ts
dimk TorA
j (Tori (MI1 · · · Is , A)δ+t1 γ1 +···+ts γs , k)
is polynomial in the ti s for t ≫ 0, and the function
dimk TorSi (MI1t1 · · · Ists , k)δ+t1 γ1 +···+ts γs
is polynomial in the ti s for t ≫ 0.
Theorem 1.3 (Theorem 3.4). Let b be a G-homogeneous ideal in S such that for any i ≥ 0
and δ ∈ G, Hbi (S)δ is a finitely generated A-module. Then, if A is Noetherian, for i ≥ 0,
there exists a subset Λi ⊆ G such that
(1) Hbi (MI1t1 · · · Ists )η+t1 γ1 +···+ts γs 6= 0 for t = (t1 , . . . , ts ) ≫ 0 if only if η ∈ Λi .
(2) Let A → k be a ring homomorphism to a field k. Then for any δ and any j, the
function
t1
i
ts
dimk TorA
j (Hb (MI1 · · · Is )δ+t1 γ1 +···+ts γs , k)
is a polynomial in the ti s for t ≫ 0.
In the simplest scenario, when S is a standard graded polynomial ring over a field k, m ⊆ S
the homogeneous maximal S-ideal, s = 1, and I ⊆ S is a homogeneous ideal generated in
degree d, Theorems 1.2 and 1.3 give the following result.
Corollary 1.4. For i ≥ 0, there exist ti and finite sets ∆′i ⊆ ∆i ⊆ Z such that
(1)
(2)
(3)
(4)
for
for
for
for
all t ∈ N, TorSi (MI t , k)η+td = 0 if η 6∈ ∆i ;
t ≥ ti , TorSi (MI t , k)η+td 6= 0 if and only if η ∈ ∆′i ;
any η ∈ Z, the function dimk TorSi (MI t , k)η+td is a polynomial in t for t ≥ ti ;
any θ ∈ Z, the function dimk Hmi (MI t )θ+td is a polynomial in t, for t ≫ 0.
While writing this paper, we were informed by Whieldon that in her recent work [16],
a similar result to Corollary 1.4 (1)–(3) is proved. We later learned that Pooja Singla also
proved these results, in the second chapter of her thesis [14], independently. She also shows in
[14] that if I is graded ideal, then for any a, b ∈ Z, dimk TorSi (I t , k)a+bt is a quasi-polynomial
in t for t ≫ 0, and give results describing the regularity of I1t1 · · · Ists for t ≫ 0.
In the general setting, when A is not necessarily a field and S is not necessarily standard
graded, the problem is more subtle, and requires more work and a different approach.
We choose not to restrict to a Noetherian base ring in all our results as it seemed to
us that it makes the presentation more clear to put this hypothesis only in the statements
where it is of use in the proof.
4
AMIR BAGHERI, MARC CHARDIN, AND HUY TÀI HÀ
Acknowledgement. Our work was done when the third named author visited the other
authors at the Institut de Mathématiques de Jussieu, UPMC. The authors would like to
thank the Institut for its support and hospitality. The third named author is partially
supported by NSA grant H98230-11-1-0165.
2. Preliminaries
In this section, we collect necessary notations and terminology used in the paper, and
prove a few auxiliary results. For basic facts in commutative algebra, we refer the reader to
[6, 13].
From now on, G denotes a finitely generated abelian group, S = A[x1 , . . . , xn ] is a Ggraded algebra over a commutative ring with identity, A ⊆ S0 and M is a G-graded Smodule. By abusing notation, we shall use 0 to denote the identity of all abelian groups
considered in the paper; the particular group will be understood from the context of its use.
Definition 2.1. Let E ⊆ G be a collection of elements in G. We say that E is a linearly
independent subset of G if E forms a basis for a free submonoid of G.
Definition 2.2. The support of M in G is defined to be
SuppG (M) = {γ ∈ G Mγ 6= 0}.
Remark 2.3. When A is a field, let F• be a minimal G-graded free resolution of M over S,
where
M
i
Fi =
S(−η)βη (M ) .
η∈G
The numbers βηi (M) are called the multigraded (or G-graded) Betti numbers of M and
βηi (M) = dimA TorSi (M, A)η
as, by definition, the maps in F• ⊗S A are zero maps.
More generally, we shall prove the following lemma relating the multigraded Betti numbers
of M and the support of TorS∗ (M, A).
Lemma 2.4. Let F• be a G-graded free resolution of a G-graded S-module M. Then
(1) Fi has a summand S(−γ) for any γ ∈ SuppG (TorSi (M, A)).
(2) Assume that there exists φ ∈ HomZ (G, R) such that φ(deg(xi )) > 0 for all i and M is
finitely generated. Then there exists a G-graded free resolution F′• of M such that
[
M
S(−γi,ℓ ) with γi,ℓ ∈
SuppG (TorSj (M, A)), ∀ℓ.
F′i =
ℓ∈Ei′
j≤i
Notice that, without further restrictions on A and/or M one cannot in general choose F′i
so that γi,ℓ ∈ SuppG (TorSi (M, A)), ∀ℓ.
THE EVENTUAL SHAPE OF BETTI TABLES OF POWERS OF IDEALS
5
Proof. For (1), let K be defined by the exact sequence
0 → K → F0 → M → 0
and notice that F0 ⊗S A → M ⊗S A is onto. As (F0 ⊗S A)γ 6= 0 if and only if S(−γ) is a direct
summand of F0 , the result holds for i = 0. Furthermore, · · · → F1 → K → 0 is a resolution
of K, TorS1 (M, A) is a graded submodule of K ⊗S A and TorSi (M, A) ≃ TorSi−1 (K, A) for
i ≥ 2, which implies the result by induction on i.
To prove the second statement, we relax the finite generation of M by the following
condition, which will enable us to make induction : ∃q ∈ R, φ(deg a) > q, ∀a ∈ M. Notice
that if a module satisfies this condition, any of its submodules satisfies the same condition.
L
Set Tj := SuppG (TorSj (M, A)). Let ψ : F0 = ℓ∈E0 S(−γ0,ℓ ) → M be the
L augmentation
and E0′ := {ℓ ∈ E0 | γ0,ℓ ∈ T0 }. Denote by ψ ′ the restriction of ψ to F′0 := ℓ∈E ′ S(−γ0,ℓ ).
0
We now prove that ψ ′ is onto. First notice that ψ ′ ⊗S A is surjective. Assume that ψ ′ is
not surjective, let M ′ be the image of ψ ′ and
h := inf{φ(γ), Mγ 6= Mγ′ } ≥ inf{φ(deg(a)), a ∈ M} ∈ R.
Set ǫ := min{φ(deg(xi )), i = 1, . . . , n} > 0 and choose m ∈ Mν \ Mν′ for some ν with
h ≤ φ(ν) <
h + ǫ. As ψ ′ ⊗S A is onto, there exists m′ ∈ M ′ such that m is of the
P
n
form m′ + i=1 mi xi . Now, for some value i we have mi 6∈ M ′ . It then follows that
φ(deg(mi )) = φ(ν) − φ(deg(xi )) < h, contradicting the definition of h.
We will now prove (2) by induction on i. To end this, assume that there exists a graded
complex
0 → F′i → · · · → F′0 → M → 0
L
with at most non zero homology in homological degree i ≥ 0, and such that F′i = ℓ∈E ′ S(−γi,ℓ )
i
S
with γi,ℓ ∈ j≤i Tj .
If the complex is exact, our claim is proved; otherwise, by setting Ki ⊂ F′i for the i-th
homology module of the complex and Qi := ker(F′i ⊗S A → F′i−1 ⊗S A), one has
SuppG (Ki ⊗S A) = SuppG (TorSi+1 (M, A)) ∪ SuppG (ker(Qi → TorSi (M, A))).
Applying the argument above to a graded onto map F → Ki , and using that by induction
SuppG (Ki ⊗S A) ⊆ Ti+1 ∪ SuppG (F′i ⊗S A) ⊆ ∪j≤i+1 Tj
one obtains a graded free S-module F′i+1 as claimed mapping onto Ki .
Let t = (t1 , . . . , ts ) ∈ Zs . We shall write t ≥ 0 (respectively, t > 0, t ≤ 0 and t 6= 0)
if ti ≥ 0 (respectively, ti > 0, ti ≤ 0 and ti 6= 0) for all i = 1, . . . , s. For a property that
depends on a parameter t ∈ Zs , one says that the property holds for t ≫ 0 if there exists
t0 ∈ Zs such that it holds for t ∈ t0 + Ns . The following semi-classical lemma will be of use.
6
AMIR BAGHERI, MARC CHARDIN, AND HUY TÀI HÀ
Lemma 2.5. Let R be a finitely generated Ns -graded algebra over a commutative ring A.
Let M be a finitely generated Zs -graded R-module. Then either Mt = 0 for t ≫ 0 or Mt 6= 0
for t ≫ 0.
L
Proof. Let b = ti ≥1 R(t1 ,...,ts ) be the ideal generated by elements of strictly positive degrees.
If M 6= Hb0 (M) one has Mt 6= 0 for any t ∈ t0 + Ns , where t0 is the degree of a non zero
element in M/Hb0 (M).
If M = Hb0 (M) then any generator a of M spans a submodule of M that is zero in degrees
deg(a) + ba (1, . . . , 1) + Ns , where ba is such that for any product p of ba elements among the
finitely many generators of the R-ideal b, one has pa = 0. As M is finitely generated, the
result follows.
We shall make use of the notion of initial modules with respect to a monomial order. This
is a natural extension of the familiarL
notion of initial ideals in a polynomial ring. Let F be
a free S-module. We can write F = i∈I Sei . A monomial in F is of the form xα ei , where
xα is a monomial in S and i ∈ I. A monomial order in F is a total order, say ≺, on the
monomials of F satisfying the following condition:
if u ≺ v and w 6= 1 is a monomial in S then u ≺ uw ≺ vw.
It can be seen that ≺ is a well ordering, i.e., every non-empty subset of the monomials
in F has a minimal element. We refer the reader to [6, 15.2] for more details on monomial
orders on free modules.
Definition 2.6. Let F be a free S-module, and let K be an S-submodule of F . Let ≺ be
a monomial order in F . The initial module of K, denoted by in≺ (K), is defined to be the
S-submodule of F generated by
{xα ei
∃f = xα ei + smaller terms ∈ K}.
Proposition 2.7. Let F be a free G-graded S-module, and let K be a G-graded S-submodule
of F . Let ≺ be a monomial order in F . Then in≺ (K) is a G-graded S-module of F , and
SuppG (F/K) = SuppG (F/ in≺ (K)).
Proof. It is clear from the definition that in≺ (K) is a G-graded S-module. To prove the
proposition, we need to show that for any µ ∈ G, Kµ = Fµ if and only if in≺ (K)µ = Fµ .
Clearly, if Kµ = Fµ then all monomials of degree µ in F are elements of K, and thus, are
elements of in≺ (K). Therefore, in this case, in≺ (K)µ = Fµ .
Suppose now that in≺ (K)µ = Fµ . Let xα ei be the smallest monomial of degree µ in F
but not in K, if it exists. Then xα ei ∈ in≺ (K)µ . Thus, there exists an element f ∈ K of the
form
f = xα ei + g,
where g consists of monomials that are smaller than xα ei with respect to ≺. Since K is
a G-graded S-module, we can choose f to be G-homogeneous of degree µ. That is, all its
THE EVENTUAL SHAPE OF BETTI TABLES OF POWERS OF IDEALS
7
monomials are of degree µ. This implies that all monomials, and thus, all terms in g are
elements in K. In particular, g ∈ K. Therefore, xα ei = f − g ∈ K, a contradiction. Hence,
Fµ = Kµ . The proposition is proved.
One of our techniques is to take the collection of elements of certain degree from a complex.
This construction gives what we shall call strands of the complex.
Definition 2.8. Let F• be a G-graded complex of S-modules and let Γ ⊆ G. The Γ-strand
of F• , often denoted by (F• )Γ , is obtained by taking elements of degrees belonging to Γ
in F• and the boundary maps between these elements L
(since the complex is graded, the
boundary maps are of degree 0). In particular, if F = γ∈G Fγ is a G-graded S-module,
L
then FΓ := γ∈Γ Fγ . Note that the degree Γ-strand of a complex/module is not necessarily
a complex/module over S.
3. Forms of the same degree
In this section, we consider the case when every ideal Ii is generated in a single degree.
That is, when degG (fi,j ) = γi for all j. We keep the notations of Section 2.
Let G′ denote a finitely generated abelian group. The following result, with G′ = Zs , will
be a key ingredient of our proof.
Theorem 3.1. Let R = S[T1 , . . . , Tr ] be a G × G′ -graded polynomial extension of S with
degG×G′ (a) ∈ G × 0 for all a ∈ S and degG×G′ (Tj ) ∈ 0 × G′ for all j. Let M be a finitely
generated G × G′ -graded R-module and let i be an integer. Assume that i = 0 or A is a
Noetherian ring. Then
(1) There exists a finite subset ∆i ⊆ G such that, for any t, TorSi (M(∗,t) , A)δ = 0 for all
δ 6∈ ∆i .
(2) Assume that G′ = Zs . For δ ∈ ∆i , TorSi (M(∗,t) , A)δ = 0 for t ≫ 0 or TorSi (M(∗,t) , A)δ 6=
0 for t ≫ 0. If, furthermore, A → k is a ring homomorphism to a field k, then for
any j, the function
S
dimk TorA
j (Tori (M(∗,t) , A)δ , k)
is polynomial in the ti s for t ≫ 0, and the function
dimk TorSi (M(∗,t) , k)δ
is polynomial in the ti s for t ≫ 0.
L
i
Proof. Let F• be a graded free resolution of M over R, where Fi = η,j R(−η, −j)βη,j is of
finite rank for i = 0, and for any i when A is Noetherian. For t ∈ G′ , the (∗, t)-strand of F• ,
denoted by Ft• , is a G-graded free resolution of M(∗,t) over S = R(∗,0) , that is not necessarily
minimal. Its i-th term is
M
i
S(−η)βη,j ⊗A Bt−j ,
Fti =
η,j
8
AMIR BAGHERI, MARC CHARDIN, AND HUY TÀI HÀ
where B = A[T1 , . . . , Tr ].
i
Let ∆i = {η ∃j : βη,j
6= 0}. The module TorSi (M(∗,t) , A) = Hi (Ft• ⊗S A) is a subquotient
L
i
of the module η,j A(−η)βη,j ⊗A Bt−j , and (1) is proved.
To prove (2), observe first that TorSi (M(∗,t) , A)δ = Hi (Ft• ⊗S A)δ and A(−η) ⊗A Bt−j δ =
[δ]
[δ]
Aδ−η ⊗A Bt−j is zero if η 6= δ. Thus, Hi (Ft• ⊗S A)δ is equal to Hi (F• ⊗S A)t , where F• is
the subcomplex of F• given by
M
M
i
i
[δ]
[S(−δ) ⊗A B(−j)]βδ,j .
Fi =
R(−δ, −j)βδ,j =
j
j
[δ]
[δ]
As F• ⊗S A is a graded complex of finitely generated B-modules, Hi (F• ⊗S A) is a
finitely generated B-module for any i when A is Noetherian. Similarly, TorSi (M(∗,t) , k)δ =
[δ]
Hi (Ft• ⊗S k)δ = Hi (F• ⊗S k) is a finitely generated k[T1 , . . . , Tr ]-module for any i when A is
Noetherian.
This proves (2) in view of Lemma 2.5 and [11, Theorem 1].
Remark 3.2. In the context of point (2) above, there is a graded spectral sequence of
S
2
graded B-modules with second term Ej,i
(t) = TorA
j (Tori (M(∗,t) , A)δ , k) that converges to
p
TorSi+j (M(∗,t) , k)δ . It follows that all terms Ej,i
(t) for p ≥ 2 are finitely generated k[T1 , . . . , Tr ]modules. In particular, one can write :
X
X
p
∞
dimk TorSℓ (M(∗,t) , k)δ =
dimk Ej,i
(t) ≤
dimk Ej,i
(t), ∀p ≥ 2,
i+j=ℓ
i+j=ℓ
which provides a control on the Hilbert function (and polynomial) of TorSℓ (M, k)δ in terms
S
of the ones of TorA
j (Tori (M, A)δ , k) for i + j = ℓ.
We are now ready to examine the asymptotic linear behavior of nonzero G-graded Betti
numbers and non-vanishing degrees of local cohomology modules of MI1t1 · · · Ists .
In the next two theorems, let R = S[Ti,j | 1 ≤ i ≤ s, 1 ≤ j ≤ ri ]. Then R is equipped with
a G × Zs -graded structure in which degG×Zs (xi ) = (degG (xi ), 0) and degG×Zs (Ti,j ) = (0, ei ),
where ei is the i-th element in the canonical basis of Zs . Recall that γi = degG (fi,j ) and let
γ := (γ1 , . . . , γs ). For t = (t1 , . . . , ts ) ∈ Zs , let I t := I1t1 . . . Ists , T t := T1t1 . . . Tsts , I t T t (γ.t) :=
I1t1 (t1 γ1 )T1t1 . . . Ists (ts γs )Tsts and MI t T t (γ.t) := MI1t1 (t1 γ1 )T1t1 . . . Ists (ts γs )Tsts .
Theorem 3.3. Assume that i = 0 or A is Noetherian. Then there exists a finite set ∆i ⊆ G
such that
(1) For all t ∈ Ns , TorSi (MI1t1 · · · Ists , A)η = 0 if η 6∈ ∆i + t1 γ1 + · · · + ts γs .
(2) There exists a subset ∆′i ⊆ ∆i such that TorSi (MI1t1 · · · Ists , A)η+t1 γ1 +···+ts γs 6= 0 for
t ≫ 0 and η ∈ ∆′i and TorSi (MI1t1 · · · Ists , A)η+t1 γ1 +···+ts γs = 0 for t ≫ 0 and η 6∈ ∆′i .
THE EVENTUAL SHAPE OF BETTI TABLES OF POWERS OF IDEALS
9
(3) If, furthermore, A → k is a ring homomorphism to a field k, then for any δ and any
j, the function
t1
S
ts
dimk TorA
j (Tori (MI1 · · · Is , A)δ+t1 γ1 +···+ts γs , k)
is polynomial in the ti s for t ≫ 0, and the function
dimk TorSi (MI1t1 · · · Ists , k)δ+t1 γ1 +···+ts γs
is polynomial in the ti s for t ≫ 0.
L
L
t t
t t
Proof. Let R :=
t≥0 I T (γ.t) and MR :=
t≥0 MI T (γ.t) denote the (shifted) multiRees algebra and the multi-Rees module with respect to I1 , . . . , Is , and M. The natural
surjective map φ : R ։ R that sends xi to xi and Ti,j to fi,j Ti makes MR a finitely
generated G × Zs -graded module over R.
Observe that, for any δ ∈ G and t ∈ Zs , MR(δ,t) ≃ [MI t (γ.t)]δ = [MI t ]δ+γ.t , where in the
last term δ + γ.t := δ + t1 γ1 + · · · + ts γs . Thus, the assertion follows by applying Theorem
3.1 to the R-module M := MR.
Theorem 3.4. Let b be a G-homogeneous ideal in S such that for any i ≥ 0 and δ ∈ G,
Hbi (S)δ is a finitely generated A-module. Then, if A is Noetherian, for i ≥ 0, there exists a
subset Λi ⊆ G such that
(1) Hbi (MI1t1 · · · Ists )η+t1 γ1 +···+ts γs 6= 0 for t = (t1 , . . . , ts ) ≫ 0 if only if η ∈ Λi .
(2) If, furthermore, A → k is a ring homomorphism to a field k, then for any δ and any
j, the function
t1
i
ts
dimk TorA
j (Hb (MI1 · · · Is )δ+t1 γ1 +···+ts γs , k)
is a polynomial in the ti s for t ≫ 0.
Proof. Since taking local cohomology respects the G-homogeneous degree, we have
i
i
Hbi (MI t )δ+γ.t = HbR
(MR)(δ,t) = HbR
(MR)(δ,∗) t .
i
Let B = A[Ti,j , 1 ≤ i ≤ s, 1 ≤ j ≤ ri ]. Since B is a flat extension of A, HbR
(R)(δ,∗) =
i
s
HbR (B ⊗A S)(δ,∗) is a finitely generated B-module. Let F• be the minimal G × Z -graded free
resolution of MR over R. Since A is Noetherian, each term Fj of F• is of finite rank. This
i
implies that for all δ ∈ G, i ≥ 0 and j ≥ 0, HbR
(Fj )(δ,∗) is a finitely generated B-module. The
i−j
i
i
spectral sequence HbR (Fj ) ⇒ HbR (MR) implies that HbR
(MR)(δ,∗) is a finitely generated
multigraded B-module. This proves (2) in view of [11, Theorem 1].
i
i
Notice that HbR
(MR)(δ,t) =
Ts0 for all t ≫ 0 if and only if Kδ = HbR (MR)(δ,∗) is annihilated
by a power of the ideal a := i=1 (Ti,1 , . . . , Ti,ri ). Hence (1) holds with
Λi := {δ ∈ G Kδ 6= Ha0 (Kδ )}.
10
AMIR BAGHERI, MARC CHARDIN, AND HUY TÀI HÀ
4. Forms of arbitrary degrees
This section is devoted to proving our main result in its full generality, when the ideals
Ii s are generated in arbitrary degrees. We start by recalling the notion of a Stanley decomposition of multigraded modules.
Definition 4.1. Let G be a finitely generated abelian group and let B = A[T1 , . . . , Tr ] be
a G-graded polynomial ring over a commutative ring A. Let M be a finitely generated
G-graded B-module. A Stanley decomposition of M is a finite decomposition of the form
m
M
M=
ui A[Zi ],
i=1
where the direct sum is as A-modules, ui s are G-homogeneous elements in M, Zi s are subsets
(could be empty) of the variables {T1 , . . . , Tr }, and ui A[Zi ] denotes the A-submodule of M
generated by elements of the form ui m where m is a monomial in the polynomial ring A[Zi ].
The following lemma is well known in N-graded or standard Nn -graded situations (cf.
[2, 3, 10]).
Lemma 4.2. Let G be a finitely generated abelian group and let B = A[T1 , . . . , Tr ] be a
G-graded polynomial ring. Let I be a monomial ideal in B. Then a Stanley decomposition
of B/I exists.
Proof. The proof follows along the same lines as in the proof of [10, Corollary 6.4] or [2,
Theorem 2.1], as one notices that any monomial is a homogeneous element.
Theorem 4.3. Let G be a finitely generated abelian group, B = A[T1 , . . . , Tr ] be a G-graded
polynomial ring over a commutative ring A and M be a finitely generated G-graded B-module.
Let Γ denote the set of subsets of {degG (Ti )}ri=1 whose elements are linearly independent over
Z. Then there exist a collection of pairs (δp , Ep ) ∈ G × Γ, for p = 1, . . . , m, such that
m
[
SuppG (M) =
δp + hEp i ,
p=1
where hEp i represents the free submonoid of G generated by elements in Ep .
Proof. Since M is a finitely generated G-graded B-module, there exists a homogeneous
surLm
jective map φ : F ։ M from a free B-module F to M. We can write F = i=1 Bei , where
degG (ei ) represents the degree of the i-th generator of M. Let K = ker φ. Then M ≃ F/K.
In particular, SuppG (M) = SuppG (F/K).
Lm
Extend any monomial order on B to a monomial order on F =
i=1 Bei , by ordering
the ei ’s. Since M is G-graded, so is K. Thus, by Proposition 2.7, we have SuppG (F/K) =
SuppG (F/ in≺ (K)). Therefore,
SuppG (M) = SuppG (F/ in≺ (K)).
THE EVENTUAL SHAPE OF BETTI TABLES OF POWERS OF IDEALS
11
Observe that in≺ (K) is generated by monomials of the form Tα ei (where Tα = T1α1 · · · Trαr
for α = (α1 , . . . , αr ) ∈ Nr ). Let Ii be the monomial ideal in B generated by all monomials
L B
ei .
Tα for which Tα ei ∈ in≺ (K). Clearly, F/ in≺ (K) ≃ m
i=1
Ii
B
By Lemma 4.2, for each i = 1, . . . , m, there exists a Stanley decomposition of ei
Ii
m
i
M
B
ei ≃
uij A[Tij ],
Ii
j=1
where uij are homogeneous elements of
variables {T1 , . . . , Tr } in B. This gives
B
ei and Tij are subsets (could be empty) of the
Ii
F/ in≺ (K) ≃
mi
m M
M
uij A[Tij ].
i=1 j=1
Thus, SuppG (F/ in≺ (K)) can be written as a finite union of the form
[
SuppG (F/ in≺ (K)) =
SuppG uij A[Tij ] .
i,j
Let δij = degG (uij ). Then
SuppG (F/ in≺ (K)) =
[
i,j
δij + SuppG (A[Tij ]) .
To prove the theorem, it now suffices to show that SuppG (A[Tij ]) can be decomposed into a
union of free submonoids of G of the form hEi, where E is a linearly independent subset in
Γ.
Since A[Tij ] is a polynomial ring whose variables are variables of B, without loss of
generality, we may assume that Tij = {T1 , . . . , Tr }, i.e., A[Tij ] = B. Let H be the binomial
ideal in B generated by {Tα − Tβ degG (Tα ) = degG (Tβ )}. Then taking the quotient B/H
is the same as identifying monomials of the same degree in B. Thus, we have
SuppG (B) = SuppG (B/H) = SuppG (B/ in≺ (H))
and
(4.1)
B
H
γ
B
=
=
in≺ (H) γ
A if γ ∈ SuppG (B/H)
0 otherwise.
By Lemma 4.2, a Stanley decomposition of B/ in≺ (H) exists. That is, we can write
B/ in≺ (H) =
s
M
j=1
uj A[Zj ]
12
AMIR BAGHERI, MARC CHARDIN, AND HUY TÀI HÀ
where uj are G-homogeneous elements of B/ in≺ (H) and Zj are subsets (could be empty) of
the variables {T1 , . . . , Tr }. Let Ej be the set of degrees of variables in Zj . It further follows
from (4.1) that B/ in≺ (H) has at most one monomial in each degree. This implies that the
support of uj A[Zj ] are all disjoint and each set Ej is linearly independent. Hence, by letting
σj = degG (uj ), we have
SuppG (B) =
s
a
j=1
The theorem is proved.
σj + hEj i .
Remark 4.4. It would be nice if the union in Theorem 4.3 is a disjoint union. However,
this is not true. Let B = A[x, y] be a Z-graded polynomial ring with deg(x) = 4 and
deg(y) = 7 (hence, Γ = {4, 7}). Let M = B/(x) ⊕ B/(y) ≃ A[y] ⊕ A[x]. Then SuppZ (M) =
{4a + 7b a, b ∈ Z≥0 }. Moreover, linearly independent subsets of Γ are {4} and {7}. It can
be easily seen that SuppZ (M) cannot be written as disjoint union of shifted free submonoids
of Z generated by 4 and/or by 7.
For a vector c = (c1 , . . . , cs ) ∈ Zs and a tuple E = (ν1 , . . . , νs ) of elements in G, we shall
denote ∆E the empty
(s − 1)-tuple (ν2 − ν1 , . . . , νs − νs−1 ) else, and by
Ps tuple if s ≤ 1 and the
′
c.E the G-degree j=1 cj νj . If E and E are tuples, we denote by E|E ′ the concatenation
of E and E ′ .
Remark 4.5. With some simple linear algebra arguments, it can be seen that for tuples
E1 , . . . , Es of elements of G, the tuple of elements of G × Zs , E1 × {e1 }| · · · |Es × {es }, where
ei is the i-th basis element of Zs , is linearly independent if and only if ∆E1 | · · · |∆Es is
linearly independent. These equivalent conditions imply that for (c1 , . . . , cs ) 6= (c′1 , . . . , c′s ),
|E |
with ci , c′i ∈ Z≥0i and |ci | = |c′i | for all i, one has c1 .E1 + · · · + cs .Es 6= c′1 .E1 + · · · + c′s .Es .
This last fact is a direct corollary of the linear independence of E1 × {e1 }| · · · |Es × {es }, as
c1 .(E1 × {e1 }) + · · · + cs .(Es × {es }) = (c1 .E1 + · · · + cs .Es ) × (|c1 |e1 + · · · + |cs |es ).
We shall now prove our main result. Recall that Ii = (fi,1 , . . . , fi,ri ) and let γi,j =
degG (fi,j ).
Theorem 4.6. Let G be a finitely generated abelian group and let S = A[x1 , . . . , xn ] be a
G-graded algebra over a commutative ring A ⊆ S0 . Let Ii = (fi,1 , . . . , fi,ri ) for i = 1, . . . s
be G-homogeneous ideals in S, and let M be a finitely generated G-graded S-module. Set
i
Γi = {degG (fi,j )}rj=1
. Let ℓ ≥ 0 and assume that ℓ = 0 or A is Noetherian.
There exist a finite collection of elements δpℓ ∈ G, a finite collection of integers tℓp,i ,
ℓ
ℓ
ℓ
and a finite collection of non-empty tuples Ep,i
⊆ Γi , such that ∆Ep,1
| · · · |∆Ep,s
is linearly
independent for all p, satisfying :
THE EVENTUAL SHAPE OF BETTI TABLES OF POWERS OF IDEALS
SuppG (TorSℓ (MI1t1
· · · Ists , A))
=
m
[
[
δpℓ +
p=1
|E ℓ |
ci ∈Z≥0p,i ,|ci |=ti −tℓp,i
13
ℓ
ℓ
c1 .Ep,1
+ · · · + cs .Ep,s
,
if ti ≥ maxp {tℓp,i } for all i.
L
Proof. As before, we use t to denote (t1 , . . . , ts ) ∈ Zs , and let R := t≥0 I t T t and MR :=
L
t t
s
t≥0 MI T . Consider R = A[x1 , . . . , xn ][Ti,j , 1 ≤ i ≤ s, 1 ≤ j ≤ ri ], the G × Z -graded
polynomial ring over A[x1 , . . . , xn ] with degG×Zs (xi ) = (degG (xi ), 0) and degG×Zs (Ti,j ) =
(degG (fi,j ), ei ), where ei denotes the i-th canonical generator of Zs . The natural surjective
map φ : R ։ R that sends xi to xi and Ti,j to fi,j Ti makes MR a finitely generated
G × Zs -graded module over R.
Let F• be a G × Zs -graded free resolution of MR over R. If A is Noetherian, then each Fi
can be chosen of finite rank, and we make such a choice. For t ∈ Zs , the degree (∗, t)-strand
Ft• of F• provides a G-graded free resolution of MI t over S = R(∗,0) . Thus,
TorSi (MI t , A) = Hi (Ft• ⊗S A).
Moreover, taking homology respects the graded structure, and therefore,
Hi (Ft• ⊗S A) = Hi (F• ⊗R R/mR)(∗,t) ,
where m = (x1 , . . . , xn ) is the homogeneous irrelevant ideal in S.
`
Let Γ′ = {(γi,j , ei ) ∈ G × Zs } = i Γi × {ei } be the set of degrees of the variables Ti,j .
Observe that Hi (F• ⊗R R/mR) is a finitely generated G ×Zs -graded module over R/mR ≃ B
for any i if A is Noetherian, and for i = 0 in any case. Applying Theorem 4.3 to the G × Zs graded module Hi (F• ⊗R R/mR) we obtain a finite collection of elements θpℓ ∈ G × Zs and a
finite collection of linearly independent subsets Epℓ ⊆ Γ′ (which we view in a fixed order as
tuples), for p = 1, . . . , m, such that
SuppG×Zs Hi (F• ⊗R R/mR) =
m
[
p=1
θpℓ + hEpℓ i .
`
ℓ
× {ei }.
Let θpℓ = (δpℓ , tℓ1,p , . . . , tℓs,p ), where δpℓ ∈ G and tℓi,p ∈ Z. One has Epℓ = si=1 Ep,i
ℓ
The linear independence of the elements in Ep is equivalent to the fact that the elements of
ℓ
ℓ
∆Ep,1
| · · · |∆Ep,s
are linearly independent.
Taking the degree (∗, t)-strand of Hi (F• ⊗R R/mR), we get
SuppG (TorSℓ (MI1t1 · · · Ists , A)) =
m
[
p=1
δpℓ +
[
|E ℓ |
ci ∈Z≥0p,i ,|ci |=ti −tℓp,i
ℓ
ℓ
c1 .Ep,1
+ · · · + cs .Ep,s
.
14
AMIR BAGHERI, MARC CHARDIN, AND HUY TÀI HÀ
Example 4.7. Let I ⊆ S be a complete intersection ideal of three forms f1 , f2 , f3 of degrees
a, b, c. On easily sees, for instance by the Hilbert Burch theorem, that
0
/
T T2
1
R(−a −L
b − c, −2) f1 f2
R(−a − b − c, −1)
f2 T3 − f3 T2
− c, −1)
f T − f T
T3 R(−b L
1 3
3 1
f3
f1 T2 − f2 T1
/R
/ R(−a − c, −1)
L
/
RI
/
0
R(−a − b, −1)
is a graded free R-resolution of RI .
It follows that TorS0 (I t , A)µ = Bµ,t , TorS2 (I t , A)µ = Bµ−a−b−c,t−1 , and
Bµ−b−c,t−1
L
T
T
T
1
2
3
/ Bµ−a−c,t−1
TorS1 (I t , A)µ = coker
Bµ−a−b−c,t−2
L
Bµ−a−b,t−1
.
Now, TorS1 (RI , A) has the following Stanley decomposition:
M
M
A[T1 , T2 , T3 ](−b − c, −1)
A[T1 , T3 ](−a − c, −1)
A[T1 , T2 , T3 ](−a − b, −1).
The ideal H generated by the binomials T α − T β with deg(T α ) = deg(T β ) is the kernel of
the map
/ A[u, v]
A[T1 , T2 , T3 ]
T1
T2
uv a
/
/
uv b
/ uv c
T3
and is therefore generated by a single irreducible and homogeneous binomial.
If, for example, (a, b, c) = (2, 5, 8) then this relation is T22 − T1 T3 , and a Stanley decomposition of, for instance, A[T1 , T2 , T3 ]/(T22 ) is A[T1 , T3 ] ⊕ T2 A[T1 , T3 ](−5, −1). It finally gives
the following decomposition for TorS1 (RI , A) : A[T1 , T3 ](−13, −1) ⊕ T2 A[T1 , T3 ](−18, −2) ⊕
A[T1 , T3 ](−10, −1) ⊕ A[T1 , T3 ](−7, −1) ⊕ T2 A[T1 , T3 ](−12, −2).
Setting Et := {2α + 8β | α, β ∈ Z+ , α + β = t} = 2t + 6{0, · · · , t}, one gets that for t ≥ 2,
– SuppZ (TorS0 (I t , A)) = Et ∪ (5 + Et−1 ),
– SuppZ (TorS1 (I t , A)) = (13 + Et−1 ) ∪ (18 + Et−2 ) ∪ (10 + Et−1 ) ∪ (7 + Et−1 ) ∪ (12 + Et−2 ),
– SuppZ (TorS2 (I t , A)) = (15 + Et−1 ) ∪ (20 + Et−2 ).
Notice that one has the simplified expression :
– SuppZ (TorS1 (I t , A)) = (5 + Et ) ∪ (10 + Et−1 ).
THE EVENTUAL SHAPE OF BETTI TABLES OF POWERS OF IDEALS
15
References
[1] D. Berlekamp. Regularity defect stabilization of powers of an ideal. Preprint. arXiv:1105.2260.
[2] K. Baclawski and A.M. Garsia. Combinatorial decompositions of a class of rings. Adv. in Math. 39
(1981), 155-184.
[3] W. Bruns, C. Krattenthaler, and J. Uliczka. Stanley decompositions and Hilbert depth in the Koszul
complex. J. Commutative Algebra, 2 (2010), no. 3, 327-357.
[4] M. Chardin. Powers of ideals and the cohomology of stalks and fibers of morphisms. Preprint.
arXiv:1009.1271.
[5] S.D. Cutkosky, J. Herzog, and N.V. Trung. Asymptotic behaviour of the Castelnuovo-Mumford regularity. Composito Mathematica, 118 (1999), 243-261.
[6] D. Eisenbud. Commutative Algebra: with a View Toward Algebraic Geometry. Springer-Verlag, New
York, 1995.
[7] D. Eisenbud and J. Harris. Powers of ideals and fibers of morphisms. Math. Res. Lett. 17 (2010), no.
2, 267-273.
[8] D. Eisenbud and B. Ulrich. Stabilization of the regularity of powers of an ideal. Preprint.
arXiv:1012.0951.
[9] H.T. Hà. Asymptotic linearity of regularity and a∗ -invariant of powers of ideals. Math. Res. Lett. 18
(2011), no. 1, 1-9.
[10] J. Herzog and D. Popescu. Finite filtrations of modules and shellable multicomplexes. Manuscripta
Math., 121 (2006), no. 3, 385-410.
[11] V. Kodiyalam. Homological invariants of powers of an ideal. Proceedings of the American Mathematical
Society, 118, no. 3, (1993), 757-764.
[12] V. Kodiyalam. Asymptotic behaviour of Castelnuovo-Mumford regularity. Proceedings of the American
Mathematical Society, 128, no. 2, (1999), 407-411.
[13] H. Matsumura. Commutative Ring Theory. Cambridge, 1986.
[14] P. Singla. Convex-geometric, homological and combinatorial properties of graded ideals. genehmitge
Dissertation. Universität Duisburg-Essen, December 2007.
[15] N.V. Trung and H. Wang. On the asymptotic behavior of Castelnuovo-Mumford regularity. J. Pure
Appl. Algebra, 201 (2005), no. 1-3, 42-48.
[16] G. Whieldon. Stabilization of Betti tables. Preprint. arXiv:1106.2355.
Institut de Mathématiques de Jussieu, UPMC, Boite 247, 4, place Jussieu, F-75252 Paris
Cedex, France
E-mail address: [email protected]
Institut de Mathématiques de Jussieu, UPMC, Boite 247, 4, place Jussieu, F-75252 Paris
Cedex, France, http://people.math.jussieu.fr/~chardin
E-mail address: [email protected]
Department of Mathematics, Tulane University, 6823 St. Charles Ave., New Orleans,
LA 70118, USA, www.math.tulane.edu/~tai/
E-mail address: [email protected]
| 0math.AC
|
arXiv:1304.0063v1 [math.AC] 30 Mar 2013
ON THE GRAPH OF DIVISIBILITY OF AN INTEGRAL
DOMAIN
JASON GREENE BOYNTON AND JIM COYKENDALL
Abstract. It is well-known that the factorization properties of a domain are
reflected in the structure of its group of divisibility. The main theme of this
paper is to introduce a topological/graph-theoretic point of view to the current understanding of factorization in integral domains. We also show that
connectedness properties in the graph and topological space give rise to a
generalization of atomicity.
1. Introduction
Let D be an integral domain with field of fractions K. Then, the group of
divisibility G(D) is defined to be the partially ordered additive group of principal
fractional ideals with aD 6 bD if and only if aD ⊇ bD. If K × is the multiplicative
group of K and if U (D) is the group of units of D, then G(D) is order isomorphic
to the quotient group K × /U (D) with the ordering aU (D) 6 bU (D) if and only if
b
a ∈ D.
It is well-known that the factorization properties of a domain are reflected in the
structure of its group of divisibility. For example, an integral domain is a unique
factorization domain if and only if its group of divisibility is a direct sum of copies
of Z equipped with the usual product order. It is also true that the group of
divisibility reflects more than just factorization properties of a domain. Indeed, it
is not hard to check that a domain is a valuation domain if and only if its group of
divisibility is totally ordered. We refer the interested reader to [6] for an excellent
survey of material regarding the group of divisibility.
In 1968, Cohn introduced the notion of an atomic integral domain in [4]. These
are the domains in which every nonzero nonunit admits a finite factorization into
irreducible elements. For several years, it was believed to be the case that atomicity
in an integral domain was equivalent to the ascending chain condition on principal
ideals (ACCP). However, in 1974, Anne Grams demonstrated that an atomic
domain need not satisfy ACCP in [5]. Grams was able to understand the subtle
difference between atomicity and ACCP using the group of divisibility. Ten years
later, Zaks added two more examples of an atomic domain without ACCP in [7].
However, examples of atomic domains without ACCP are still relatively scarce.
The main theme of this work is to introduce a topological/graph-theoretic point
of view to the current understanding of factorization in integral domains. That
is, we find a graphical representation of the group of divisibility in order to detect
various well-studied factorization properties of an integral domain. The contents
of this paper is organized as follows. In Section 2, we recall a topological structure
2010 Mathematics Subject Classification. Primary 13F15; Secondary 13A05.
Key words and phrases. atomic, factorization, divisibility.
1
2
J. G. BOYNTON AND J. COYKENDALL
that is naturally associated to a partially ordered set. In addition, we make the
relevant graph-theoretic definitions needed in the sequel. In Section 3, we introduce
the graph of divisibility of an integral domain and show that this graph detects
the standard factorization properties studied in [1]. In Section 4, we examine the
connectedness properties of the graph of divisibility using some elementary topology
to do so. In Section 5, we will see that a connected graph of divisibility gives rise
to a generalized atomicity. We also provide some examples in order to illustrate
these notions.
2. Some Definitions and Background
In this section, we make some relevant definitions from graph theory and topology
that will be used throughout. We refer the reader to [2] for a survey of known
results about the Alexandrov topology.
Definition 2.1. Let (X, τ ) be a topological space with neighborhood base U(x) =
{U ∈ τ : x ∈ U }.
(1) (X, τ ) is called an Alexandrov space if arbitrary intersections of open sets
remain open.
(2) For every x in an Alexandrov space X, we set M (x) = ∩U∈U (x) U. The set
M (x) is called the minimal open set containing x.
Theorem 2.2. Let (X, τ ) be an Alexandrov space.
(1) The collection of minimal open sets N = {M (x) : x ∈ X} is a basis for the
space (X, τ ).
(2) (X, τ ) is a T0 space if and only if M (x) = M (y) ⇒ x = y.
(3) (X, τ ) is (path and chain) connected if and only if for any pair of points
a, b ∈ X, there exists a finite set of points {a = x0 , x1 , ..., xn = b} such that
N (xi−1 ) ∩ N (xi ) 6= ∅, i 6 n.
In some sense, a T0 Alexandrov space is the most natural topological structure
induced by a partially ordered set. Indeed, if (X, 6) is any partially ordered set,
then the sets of the form M (a) = {x ∈ X : x 6 a} constitute a basis for a T0
Alexandrov space (X, τ ). Conversely, if (X, τ ) is a T0 Alexandrov space, we can
define a relation 6 on X given by a 6 b if and only if a ∈ M (b). More precisely,
we have the following result found in [2]
Theorem 2.3. There is an isomorphism between the category of T0 Alexandrov
spaces with continuous maps and the category of partially ordered sets with order
preserving set maps.
Now, let (X, τ ) be any T0 Alexandrov space with minimal neighborhood base
M = {M (a) : a ∈ X}. One can construct a directed acyclic graph G(V, E)
determined by the space (X, τ ). The set V of vertices is taken to be the underlying
set X. Define an edge a → b if and only if M (a) ( M (b) and there is no minimal
base element M (c) properly between M (a) and M (b). The resulting construction
is a simple (no parallel edges) directed acyclic graph (SDAG). We now make the
graphical representation of the previous constructions precise.
Definition 2.4. Let (X, 6) be any partially ordered set and define intervals (−∞, b] =
{x ∈ X : x 6 b} and [a, b] = {x ∈ X : a 6 x 6 b}
GRAPH OF DIVISIBILITY
3
(1) We write (X, τ (6)) to denote the the Alexandrov topology generated by the
minimal neighborhood base M = {M (a) : a ∈ X} where M (a) = (−∞, a].
(2) We write G(X, E(6)) to denote the directed acyclic graph whose vertices are
the elements of X and edges a → b if and only if a < b and [a, b] = {a, b}.
Definition 2.5. Let V be a nonempty set.
(1) A finite directed path is a sequence of edges {e1 , e2 , ..., en } ⊂ E where ei =
(vi−1 , vi ) for each i ∈ {1, 2, ..., n}. A finite directed path in G(V, E) is also
denoted by
v0 → v1 → v2 → ... → vn .
A directed graph is said to be acyclic if there does not exist a path {e1 , e2 , ..., en } ⊂
E such that v0 = vn .
(2) A finite weak path is a sequence of ordered pairs {e1 , e2 , ..., en } ⊂ V × V
where ei = (vi−1 , vi ) and either (vi−1 , vi ) or (vi , vi−1 ) ∈ E. A finite directed
path in G(V, E) is also denoted by
v0 ↔ v1 ↔ ... ↔ vn .
(3) A directed graph G(V, E) is said to be weakly connected if for every pair of
vertices v, w ∈ V there exists a finite weak path {e1 , e2 , ..., en } such that
v0 = v and vn = w.
3. The Graph of Divisibility
In this section, we introduce the graph of divisibility of an integral domain. We
will see that this graph gives a picture of the group of divisibility and can be used
to detect certain factorization properties of a domain. Although this graph does
not detect all divisibility relations, it does detect enough of the divisibility relation
to clearly differentiate atomicity and ACCP (for example). For the remainder of
this article, we denote the set of irreducible elements (atoms) of D by Irr(D) and
the set of atomic elements (expressible as a finite product of atoms) is denoted by
F (D).
Definition 3.1. Let D be any integral domain with field of fractions K and let
K × denote its multiplicative group.
(1) We write G(D) to denote the group of divisibility K × /U (D) written additively. We write G(D)+ to denote the positive elements of G(D).
(2) We write P(D) to denote the group of nonzero principal fractional ideals of
D partially ordered by inclusion. We write P(D)+ to denote the nonzero
nonunit principal integral ideals of D.
Recall that the ordering in (G(D), 6) is given by a 6 b if and only if ab ∈ D.
It readily follows that 0 6 a if and only if a ∈ D. It is easy to check that there
exists a reverse order group isomorphism G(D) → P(D) given by a 7→ aD. With
Definition 2.1 in hand, we define a partial ordering on the set P(D) and consider
the structure of the associated topological space and directed acyclic graph. The
following lemma is the basis for the remainder of our investigations.
Lemma 3.2. Define a relation ≺ on P given by a ≺ b if and only if ab ∈ F (D).
(1) (P(D), ) is a partially ordered set.
(2) (P(D), τ ()) is a T0 Alexandrov space with neighborhood base given by
the collection M(a) = {x ∈ P : xa ∈ F (D)}.
4
J. G. BOYNTON AND J. COYKENDALL
(3) G(P(D), E()) is a directed acyclic graph with directed edges a → b if and
only if ab ∈ Irr(D).
Proof. (1) It is never the case that a ≺ a since aa is a unit, and hence is not a
product of atoms. Similarly, it is impossible that both a ≺ b and a b can occur.
Finally, if a ≺ b and b ≺ c, then ab ∈ F (D) and bc ∈ F (D). Since the set F (D) is
multiplicatively closed, we have that ab · cb = ac ∈ F (D) so that a ≺ c.
(2) Follows immediately from (1) and the definition of ≺ .
(3) If a → b, then a ≺ b and [a, b] = {a, b}. It follows that ab = π1 · · · πn where each
πi ∈ Irr(D). In other words, [a, b] = {a, π1 · · · πn−1 a, ..., π1 a, b} and the condition
[a, b] = {a, b} forces b = π1 a. Therefore, ab ∈ Irr(D) as needed. Conversely, if
a
b = π ∈ Irr(D), then it is certainly true that a ≺ b and it suffices to check that
[a, b] = {a, b}. But if a ≺ x ≺ b, then xa = π1 · · · πn and xb = ς1 · · · ςm where each
πi , ςi ∈ A(D). But then π = π1 · · · πn ς1 · · · ςm forcing (without loss of generality)
π = π1 with the remaining factors units. It follows that x = b as needed.
With Lemma 3.2 in hand, we make the definition central to our study.
Definition 3.3. We call G(P(D), E()) the graph of divisibility of D. We might
also refer to the subgraph G(P(D)+ , E()) the graph of divisibility.
We illustrate this definition with a few easy examples.
Example 3.4.
(1) Let D be a one-dimensional Noetherian valuation domain.
It is well-known that D is a PID with a unique nonzero prime ideal. So the
the elements of P(D)+ can be enumerated by the positive integers. We
write P(D)+ = {π, π 2 , π 3 , ...} where π is a chosen generator of the unique
maximal ideal. The graph of divisibility G(P(D), E()) is the (branchless)
tree that looks like
1
1
... → π 2 → π → 1 → → 2 → ....
π
π
Similarly, the subgraph G(P(D)+ , E()) looks like
... → π 3 → π 2 → π
(2) Let D be a one-dimensional nondiscrete valuation domain. For the sake
of concreteness, we will say that the corresponding value group is Q. In
this example, there are no irreducible elements and hence no two elements
of P(D) are adjacent. It follows that G(P(D), E()) is just the collection
of vertices corresponding to P(D) with no edges whatsoever. In fact, the
graph of divisibility of any antimatter domain (no irreducible elements)
consists of vertices only. Topologically speaking, (P(D), τ ()) is totally
disconnected. That is, given any a ∈ P(D) we have that M (a) = {a}. The
same is certainly true for the subspace (P(D)+ , τ ()) and the subgraph
G(P(D)+ , E()).
Recall that a sink in a directed graph is a vertex with arrows in but no arrows
out. We have the following lemma.
Lemma 3.5. Let D be an integral domain and let G(P (D)+ , E()) be the associated graph of divisibility. Then,an element π ∈ D• is irreducible in D if and only
if the node π is a sink in G(P (D)+ , E()).
GRAPH OF DIVISIBILITY
5
It is well known that D is atomic if and only if every element of G(D)+ can be
written as a sum of minimal positive elements. Similarly, D is satisfies ACCP if
and only if every descending sequence of elements in G(D)+ stabilizes. As with
the group of divisibility, the graph of divisibility can be used to characterize the
well-studied factorization domains. We close this section with the following result.
Theorem 3.6. Let D be an integral domain and let G(P(D)+ , E()) be the associated graph of divisibility.
(1) D is atomic if and only if for every non unit element a ∈ P(D)+ , there
exists a (finite) path originating from a that terminates at an atom.
(2) D satisfies ACCP if and only if for every a ∈ P(D)+ , every path originating
from a terminates at an atom.
(3) D is a BFD if and only if for every a ∈ P(D)+ , every path originating from
a terminates at an atom and there is an upper bound on the lengths of all
such paths.
(4) D is an FFD if and only if for every a ∈ P(D)+ , every path originating
from a terminates at an atom and there are finitely many such paths.
(5) D is an HFD if and only if for every a ∈ P(D)+ , every path originating
from a terminates at an atom and all such paths are of the same length.
4. Some Connectedness Properties
In this section, we consider the connectedness of the graph of divisibility. To
do this, we will examine the connectedness of the associated Alexandrov topology.
We conclude the section with a few examples.
Theorem 4.1. Let D be an integral domain. The following statements for a, b ∈
K × are equivalent.
···πn
.
(1) There exist atoms πi , ξi ∈ Irr(D) such that ab = πξ11···ξ
m
(2) The points a, b belong to the same connected component in the Alexandrov
topology (P(D), τ ()).
(3) There is a finite weak path connecting a to b in the graph of divisibility
G(P(D), E()).
···πn
Proof. (1)⇒(2) Suppose there exist atoms πi , ξi ∈ Irr(D) such that ab = πξ11···ξ
.
m
To show that a, b belong to the same connected component, it suffices to show that
M (a) ∩ M (b) is nonempty. To this end, note that
aξ1 · · · ξm = c = bπ1 · · · πn
implies that c ≺ a because ac = ξ1 · · · ξm . Similarly, we have that c ≺ b, from which
it follows that c ∈ M (a) ∩ M (b) as needed.
(2)⇒(3) If a, b belong to the same connected component, then there exists
a finite set of points {a = x0 , x1 , ..., xn = b} such that M (xi−1 ) ∩ M (xi ) 6= ∅
for all i ∈ {1, 2..., n}. Hence, we can choose a ci ∈ M (xi−1 ) ∩ M (xi ) so that
ci
ci
ci
ci
xi−1 , xi ∈ F (D), say xi−1 = ξ1 · · · ξm and , xi = π1 · · · πn where πi , ξi ∈ Irr(D). It
follows that there are directed paths
ci → (xi−1 ξ1 · · · ξm−1 ) → ... → (xi−1 ξ1 ) → xi−1
and
ci → (xi π1 · · · πn ) → ... → (xi π1 ) → xi .
6
J. G. BOYNTON AND J. COYKENDALL
Hence, there is a weak path
xi−1 ← ... ← ci → ... → xi
for all i ∈ {1, 2..., n}, and so there is a weak path connecting a to b.
(3)⇒(1) Suppose that a, b are distinct points in the T0 Alexandrov space (P(D), τ (
)). Then there exists a finite weak path connecting a to b say
a = x0 ↔ x1 ↔ ... ↔ xn = b.
Using induction on n, we suppose that the result is true for all k < n. It follows from
···πn
a
the existence of the weak path a = x0 ↔ x1 ↔ ... ↔ xn−1 that xn−1
= πξ11···ξ
where
m
πi , ξi ∈ Irr(D). Now observe that either xn−1 → b or b → xn−1 . If xn−1 → b then
a
1 ···πn
= π ∈ Irr(D). It follows that ab = xn−1
= ππ
· xn−1
by definition, we have xn−1
b
b
ξ1 ···ξm ,
and a similar argument handles the case.
If F is the subgroup of K × generated by Irr(D), then we can relate the number
of connected components of G(P, E()) with the order of the quotient group K × /F
(a homomorphic image of the group of divisibility K × /U (D)). We immediately
get the following result.
Corollary 4.2. There are 1-1 correspondences between the elements of K × /F,
the connected components of G(P(D), E()), and the connected components of
(P(D), τ ()).
Example 4.3. Consider the classical construction D = Z + xQ[x]. It is wellknown that the irreducible elements of D are the primes p ∈ Z and Q[x]-irreducible
polynomials of the form ±1 + xq(x) where q(x) ∈ Q[x] (see [3]). For each a ∈ D
let us write a(x) = (a0 , a1 , a2 , ...) where a0 ∈ Z and ai ∈ Q for all i > 1. As with
power series representations, we define the order of a to be the natural number
ord(a) = min{i ∈ N : ai 6= 0}. It follows from [3] that a(x) ∈ F (D) if and only if
ord(a) = 0. We will now show that two polynomials a, b ∈ D belong to the same
connected component of (P(D)+ , τ ()) if and only if ord(a) = ord(b). Indeed,
write a(x) = xe0 a(x) and b(x) = xf0 b(x) where ord(a) = 0 = ord(b) (allowing
e0 = 0 = f0 ). If a(x) is connected to b(x), then by there exist atoms πi , ξi ∈ Irr(D)
such that
π1 (x) · · · πn (x)
a(x)
=
.
b(x)
ξ1 (x) · · · ξm (x)
We now have the equation
xe0 a(x)ξ1 (x) · · · ξm (x) = xf0 b(x)π1 (x) · · · πn (x)
and one easily checks that
e0 = ord(xe0 ) = ord(xe0 aξ1 · · · ξm ) = ord(xf0 bπ1 · · · πn ) = ord(xf0 ) = f0
For the converse, suppose that e0 = f0 . Again, using the fact that ord(a) = 0 =
ord(b) is equivalent to a, b ∈ F (D), we have the existence of atoms πi , ξi ∈ Irr(D)
such that
a(x)
π1 (x) · · · πn (x)
.
=
ξ1 (x) · · · ξm (x)
b(x)
On the other hand, e0 = f0 implies
xe0 a(x)
a(x)
a(x)
.
= f
=
b(x)
b(x)
x 0 b(x)
GRAPH OF DIVISIBILITY
7
It follows that the distinct connected components of (P(D)+ , τ ()) are given by the
set {Irr(D) = [2], [x], [x2 ], ...}. In other words, there is no weak path xm ↔ ... ↔ xn
in G(P + , E()) whenever m 6= n.
Example 4.4. Let x, y be indeterminates over the field F2 .
k
(1) Now let X = {xα : α ∈ Q+ } and Z1 = { xyα : α ∈ Q+ , k ∈ Z+ , and
k > 2}. We determine the number of connected components in the graph
of divisibility of the domain D1 = F2 [X, y, Z1 ](X,y,Z1 ) . To do this, we
first observe that the integral closure of D1 is the rank 2 valuation domain
V = F2 [X, Z](X,Z) where Z = { xyα : α ∈ Q+ }. The value group of V is
Z ⊕ Q ordered lexicographically and it is easy to check that every element
of V • is a unit multiple of y k or xr y k where (k, r) ∈ Z+ ⊕ Q. It is not
hard to check that every element of Irr(D1 ) has value (1, 0). It is now an
easy matter to check that the connected components of G(P(D1 )+ , E())
are given in terms of their values by
{[(0, α)]}α∈Q+ ∪ {[(k, 0)]}k∈Z+ ∪ {[(k, α)]} k>2 .
α<0
1
2
For example, consider the elements f = x and g =
y3
1
x3
. Then v(f ) = (0, 21 )
and v(g) = (3, 13 ). Then v( fg ) = v(g) − v(f ) = (3, − 61 ) cannot be written
in the form m(1, 0) where m ∈ Z. In other words, fg cannot be expressed
as the quotient of atomic elements.
k
(2) If Z2 = { yxj : j ∈ Z+ , k ∈ Z+ , and k > 2} and D2 = F2 [x, y, Z2 ](x,y,Z2 ) ,
then (P(D2 )+ , τ ()) is a connected Alexandrov space. Equivalently, the
graph of divisibility G(P(D2 )+ , E()) is weakly connected. One need only
check that the integral closure of D2 has the discrete value group Z ⊕ Z
ordered lexicographically. Again, it is not hard to check that every element
of Irr(D2 ) has value (0, 1) or (1, 0) and given any f, g ∈ D2 , we have that
v( fg ) = m(1, 0) + n(0, 1) where m, n ∈ Z.
5. Some Generalizations of Atomicity
In this section, we show that a connected graph of divisibility gives rise to a
generalization of atomicity.
Definition 5.1. Let D be any integral domain.
(1) D is called almost atomic if for every a ∈ D• , there exist atoms {πi } ⊂
Irr(D) such that aπ1 · · · πn ∈ F (D).
(2) D is called quasi atomic if for every a ∈ D• , there exists an element b ∈ D
such that ab ∈ F (D).
It is easy to see that almost atomic implies quasi atomic. Also, if D is quasi
atomic, it is not hard to show that every nonzero prime ideal of D contains an
irreducible element. We have the following lemma.
Lemma 5.2. Given an integral domain D, each condition below implies the next:
(1) D is atomic
(2) D is almost atomic
(3) D is quasi atomic
(4) Every nonzero prime ideal of D contains an irreducible element.
8
J. G. BOYNTON AND J. COYKENDALL
Proof. It suffices to show that (3) implies (4). Suppose that D is quasi atomic. If a
is a nonzero element of a prime ideal P , then there is b ∈ D such that ab = π1 · · · πn
where each πi ∈ Irr(D). But then π1 · · · πn ∈ P so that πi ∈ P for some i 6 n.
These observations give an example of an integral domain that is not quasi
atomic.
Example 5.3. As in Example 4.3, let D = Z + xQ[x]. Then xQ[x] is a prime
ideal of D that contains no irreducible element. To see this, recall from 4.3 that if
f ∈ Irr(D), then ord(f ) = 0. But f ∈ xQ[x] if and only if ord(f ) > 1. It follows
from Lemma 5.2 that D is not quasi atomic.
We now show the connection between almost atomicity and a connected graph
of divisibility.
Theorem 5.4. The following statements are equivalent for a domain D.
(1) D is almost atomic.
(2) (P, τ ()) is connected.
(3) G(P, E()) is weakly connected.
Proof. (1)⇒(2) Choose any two points a, b ∈ (P, τ ()). If D is almost atomic,
there exist atoms πi , ξi ∈ Irr(D) such that aπ1 · · · πn and bξ1 · · · ξm ∈ F (D). In
π ···πn σi ···σj
other words, there exist atoms σi , ςi ∈ A(D) such that ab = ξ11 ···ξm
ςi ···ςk . It follows
from Theorem 4.1 that any pair of points in (P, τ ()) belong to the same connected
component.
(2)⇒(3) Follows immediately from Theorem 4.1.
(3)⇒(1) Since G(P, E()) is weakly connected, there is a weak path connecting
any a1 ∈ P (where a ∈ D) to an element of the form π1 where π ∈ Irr(D). Theorem
···πn
4.1 implies that there exist atoms πi , ξi ∈ Irr(D) such that πa = πξ11···ξ
. In other
m
words, there exist atoms ξi ∈ Irr(D) such that aξ1 · · · ξm ∈ F (D).
Using the Theorem 5.4 and the results from the previous section, we are led to
an example of an almost atomic domain that is not atomic.
(1) As in Example 4.4(1), let D1 = F2 [X, y, Z1 ](X,y,Z1 ) where
Example 5.5.
k
X = {x : α ∈ Q+ } and Z1 = { xyα : α ∈ Q+ , k ∈ Z+ , and k > 2}. Since the
connected components are in a 1-1 correspondence with Q, it is certainly
not the case that D1 is almost atomic (Theorem 5.4). However, it is the
case that D1 is quasi atomic. Indeed, given any f ∈ D1• , we can write
v(f ) = (k, α). There is a g ∈ D1 such that v(g) = (2, −α) and so
α
v(f g) = v(f ) + v(g) = (k + 2, 0) = (1, 0) + ... + (1, 0).
Translating this information back to D1 , we get that f g = y k+2 u for some
unit u ∈ V. Note that if y n+1 u ∈ D1 for some unit u ∈ V , then y n+1 u =
l
v1 xα + v2 xyβ + v3 y where either vi ∈ U (D1 ) or vi = 0. If n > 0, then
l−1
v1 = 0 = v3 . It follows that l > n + 1 so that y n u = v2 yxβ . Therefore,
l−n
yu = v2 yxβ ∈ D1 as l − n > 1. It follows from all of this that f g ∈ F (D)
as needed.
k
(2) As in Example 4.4(2), let D2 = F2 [x, y, Z2 ](x,y,Z2 ) where Z2 = { yxj : j ∈
Z+ , k ∈ Z+ , and k > 2}. Since G(P(D2 )+ , E()) is weakly connected, it
must be the case that D2 is almost atomic. However, it is not atomic since,
GRAPH OF DIVISIBILITY
9
2
for example, v( y 1 ) = (2, 12 ) cannot be written as an N-linear combination
x2
m(1, 0) + n(0, 1).
Acknowledgements. The authors would like to thank the North Dakota State
University Department of Mathematics for their continued support.
References
[1] D.D. Anderson, D.F. Anderson, and M. Zafrullah. Factorization in integral domains. J. Pure
Appl. Algebra, 69:1–19, 1990.
[2] F. G. Arenas. Alexandroff spaces. Acta Math. Univ. Comenianae, 68:17–25, 1999.
[3] S. T. Chapman and W. W. Smith. Restricted elasticity in rings of integer-valued polynomials
determined by finite subsets. Monatsh. Math., 148:195–203, 2006.
[4] P. M. Cohn. Bézout rings and their subrings. Proc. Cambridge Philos. Soc., 75:321–329, 1974.
[5] Anne Grams. Atomic rings and the ascending chain condition for principal ideals. Proc. Cambridge Philos. Soc., 75:321–329, 1974.
[6] Joe L. Mott. The group of divisibility and its applications. In Conference on Commutative
Algebra (Univ. Kansas, Lawrence, Kan., 1972), pages 194–208. Lecture Notes in Math., Vol.
311. Springer, Berlin, 1973.
[7] Abraham Zaks. Atomic rings without a.c.c. on principal ideals. J. Algebra, 74(1):223–231,
1982.
Department of Mathematics, North Dakota State University, Fargo, ND 58108
E-mail address, J. G. Coykendall: [email protected]
Department of Mathematics, North Dakota State University, Fargo, ND 58108
E-mail address, J. Coykendall: [email protected]
| 0math.AC
|
1
Statistical Mechanics of MAP Estimation:
General Replica Ansatz
arXiv:1612.01980v2 [cs.IT] 23 Oct 2017
Ali Bereyhi, Ralf R. Müller, and Hermann Schulz-Baldes
Abstract
The large-system performance of maximum-a-posterior estimation is studied considering a general distortion
function when the observation vector is received through a linear system with additive white Gaussian noise. The
analysis considers the system matrix to be chosen from the large class of rotationally invariant random matrices. We
take a statistical mechanical approach by introducing a spin glass corresponding to the estimator, and employing the
replica method for the large-system analysis. In contrast to earlier replica based studies, our analysis evaluates the
general replica ansatz of the corresponding spin glass and determines the asymptotic distortion of the estimator for
any structure of the replica correlation matrix. Consequently, the replica symmetric as well as the replica symmetry
breaking ansatz with b steps of breaking is deduced from the given general replica ansatz. The generality of our
distortion function lets us derive a more general form of the maximum-a-posterior decoupling principle. Based on
the general replica ansatz, we show that for any structure of the replica correlation matrix, the vector-valued system
decouples into a bank of equivalent decoupled linear systems followed by maximum-a-posterior estimators. The
structure of the decoupled linear system is further studied under both the replica symmetry and the replica symmetry
breaking assumptions. For b steps of symmetry breaking, the decoupled system is found to be an additive system
with a noise term given as the sum of an independent Gaussian random variable with b correlated impairment
terms. The general decoupling property of the maximum-a-posterior estimator leads to the idea of a replica simulator
which represents the replica ansatz through the state evolution of a transition system described by its corresponding
decoupled system. As an application of our study, we investigate large compressive sensing systems by considering
the ℓp norm minimization recovery schemes. Our numerical investigations show that the replica symmetric ansatz for
ℓ0 norm recovery fails to give an accurate approximation of the mean square error as the compression rate grows,
and therefore, the replica symmetry breaking ansätze are needed in order to assess the performance precisely.
Index Terms
Maximum-a-posterior estimation, linear vector channel, decoupling principle, equivalent single-user system, compressive sensing, zero norm, replica method, statistical physics, replica symmetry breaking, replica simulator
The results of this manuscript were presented in parts at 2016 IEEE Information Theory Workshop (ITW) [78] and 2017 IEEE Information
Theory and Applications Workshop (ITA) [79].
This work was supported by the German Research Foundation, Deutsche Forschungsgemeinschaft (DFG), under Grant No. MU 3735/2-1.
Ali Bereyhi and Ralf R. Müller are with the Institute for Digital Communications (IDC), Friedrich Alexander University of Erlangen-Nürnberg
(FAU), Konrad-Zuse-Straße 5, 91052, Erlangen, Bavaria, Germany (e-mails: [email protected], [email protected]).
Hermann Schulz-Baldes is with the Department of Mathematics, FAU, Cauerstraße 11, 91058, Erlangen, Bavaria, Germany (e-mail: schuba@
mi.uni-erlangen.de).
2
I. I NTRODUCTION
Consider a vector-valued Additive White Gaussian Noise (AWGN) system specified by
y = Ax + z
(1)
where the independent and identically distributed (i.i.d.) source vector xn×1 , with components in a support set
X ⊂ R, is measured by the random system matrix Ak×n ∈ Ak×n , with A ⊂ R, and corrupted by the i.i.d. zero-mean
Gaussian noise vector z k×1 , with variance λ0 , i.e., z ∼ N (0, λ0 I). The source vector can be estimated from the
observation vector y k×1 using a Maximum-A-Posterior (MAP) estimator. For a given system matrix A, the estimator
maps the observation vector to the estimated vector x̂n×1 ∈ Xn via the estimation function g(·|A) defined as
1
2
g(y|A) = arg minn
ky − Avk + u(v)
(2)
v∈X
2λ
for some “utility function” u(·) : Rn 7→
R+ and estimation parameter λ ∈ R+ . In (2), k·k2 denotes the Euclidean
norm, and it is assumed that the minimum is not degenerate so that g(·|A) is well-defined, at least for almost all y
and A. In order to analyze the performance of the system in the large-system limit, i.e., k, n ↑ ∞, one considers a
general distortion function d(·; ·) : X × X 7→ R. For some choices of d(·; ·), the distortion function determines the
distance between the source and estimated vector, e.g. d(x̂; x) = |x̂ − x|2 ; however, in general, it takes different
choices. The asymptotic distortion
n
1X
d(x̂i ; xi ),
n↑∞ n
i=1
D = lim
(3)
then, expresses the large-system performance regarding the distortion function d(·; ·). The performance analysis of
the estimator requires (2) to be explicitly computed, and then, x̂ = g(y|A) substituted in the distortion function.
This task, however, is not trivial for many choices of the utility function and the source support
X, and becomes
unfeasible as n grows large. As basic analytic tools fail, we take a statistical mechanical approach and investigate
the large-system performance by studying the macroscopic parameters of a corresponding spin glass. This approach
enables us to use the replica method which has been developed in the context of statistical mechanics.
A. Corresponding Spin Glass
Consider a thermodynamic system which consists of n particles with each having a microscopic parameter vi ∈ V.
T
The vector v n×1 = [v1 , . . . , vn ] , collecting the microscopic parameters, presents then the microscopic state of the
system and is called the “microstate”. The main goal of statistical mechanics is to excavate the “macroscopic
parameters” of the system, such as energy and entropy through the analysis of the microstate in the thermodynamic
limit, i.e., n ↑ ∞. Due to the large dimension of the system, statistical mechanics proposes a stochastic approach in
which the microstate is supposed to be randomly distributed over the support
this system, the Hamiltonian E(·) :
Vn due to some distribution pv . For
Rn 7→ R+ assigns to each realization of the microstate a non-negative energy
level, and H := −E pv log pv denotes the system’s entropy. The “free energy” of the thermodynamic system at the
3
inverse temperature β is then defined as
F(β) := E pv E(v) − β−1 H.
(4)
The second law of thermodynamics states that the microstate at thermal equilibrium takes its distribution such that
the free energy meets its minimum. Thus, the microstate’s distribution at thermal equilibrium reads
pβ
v (v) = [Z(β)]
−1 −βE(v)
e
(5)
where Z(β) is a normalization factor referred to as the “partition function”, and the superscript β indicates the
distribution’s dependence on the inverse temperature. The distribution in (5) is known as the “Boltzmann-Gibbs
distribution” and covers many distributions on
Vn
by specifying E(·) and β correspondingly. Substituting the
Boltzmann-Gibbs distribution in (4), the free energy at thermal equilibrium and inverse temperature β reads
F(β) = −β−1 log Z(β).
(6)
The average energy and entropy of the system at thermal equilibrium are then determined by taking expectation
over the distribution in (5), i.e.,
E(β) := E pβv E(v)
(7a)
H(β) := −E pβv log pβ
v (v),
(7b)
which can be calculated in terms of the free energy via
d
[βF(β)]
dβ
d
[F(β)] .
H(β) = β2
dβ
E(β) =
(8a)
(8b)
In spin glasses [1], the Hamiltonian assigns the energy levels randomly using some randomizer Q resulting
from random interaction coefficients. In fact, each realization of Q specifies a thermodynamic system represented
by the deterministic Hamiltonian E(·|Q). In statistical mechanics, Q is known to have “quenched” randomness
while the microstate is an “annealed” random variable. The analysis of spin glasses takes similar steps as above
considering a given realization of the randomizer, and therefore, as the system converges to its thermal equilibrium
at the inverse temperature β, the microstate’s conditional distribution given Q, i.e., pβ
v|Q , is a Boltzmann-Gibbs
distribution specified by E(·|Q). Consequently, the free energy reads
F(β|Q) = −β−1 log Z(β|Q).
(9)
where Z(β|Q) is the partition function with respect to the Hamiltonian E(·|Q). Here, the free energy, as well
as other macroscopic parameters of the system, is random; however, the physical intuition behind the analyses
suggests that these random macroscopic parameters converge to deterministic values in the thermodynamic limit.
This property is known as the “self averaging property” and has been rigorously justified for some particular classes
4
of Hamiltonians, e.g., [2]–[5]. Nevertheless, in cases where a mathematical proof is still lacking, the property is
supposed to hold during the analysis. According to the self averaging property, the free energy of spin glasses
converges to its expectation in the thermodynamic limit.
As mentioned earlier, the MAP estimator in (2) can be investigated using a corresponding spin glass. To see that,
consider a spin glass whose microstate is taken from
E(v|y, A) =
Xn , and whose Hamiltonian is defined as
1
ky − Avk2 + u(v).
2λ
(10)
Here, the system matrix A and the observation vector y are considered to be the randomizers of the spin glass. In
this case, given A and y, the conditional distribution of the microstate is given by
pβ
v|y,A (v|y, A) = [Z(β|y, A)]
−1 −βE(v|y,A)
e
.
(11)
Taking the limit when β ↑ ∞ and using Laplace method of integration [6], the zero temperature distribution, under
this assumption that the minimizer is unique, reduces to
p∞
v|y,A (v|y, A) = 1{v = arg minn E(v|y, A)}
v∈X
= 1{v = g(y|A)},
(12a)
(12b)
where 1{·} denotes the indicator function, and g(·|A) is defined as in (2). (12b) indicates that the microstate of the
spin glass converges to the estimated vector of the MAP estimator, i.e., x̂ = g(y|A), in the zero temperature limit.
Invoking this connection, we study the corresponding spin glass instead of the MAP estimator. We represent the
input-output distortion of the system regarding a general distortion function d(·; ·) as a macroscopic parameter of the
spin glass. Consequently, the replica method developed in statistical mechanics is employed to determine the defined
macroscopic parameter of the corresponding spin glass. The replica method is a generally nonrigorous but effective
method developed in the physics literature to study spin glasses. Although the method lacks rigorous mathematical
proof in some particular parts, it has been widely accepted as an analysis tool and utilized to investigate a variety
of problems in applied mathematics, information processing, and coding [7]–[10].
The use of the replica method for studying multiuser estimators goes back to [11] where Tanaka determined
the asymptotic spectral efficiency of Marginal-Posterior-Mode (MPM) estimators by employing the replica method.
The study demonstrated interesting large-system properties of multiuser estimators, and consequently, the statistical
mechanical approach received more attention in the context of multiuser systems. This approach was then employed
in the literature to study multiple estimation problems in large vector-valued linear systems, e.g. [12]–[14]. The
method was also utilized to analyze the asymptotic properties of Multiple-Input Multiple-Output (MIMO) systems in
[15] considering an approach similar to [11]. Regarding multiuser estimators, the earlier studies mainly considered
the cases in which the entries of the source vector are binary or Gaussian random variables. The results were
later extended to a general source distribution in [14]. The statistical mechanical approach was further employed
to address mathematically similar problems in vector precoding, compressive sensing and analysis of superposition
codes [16]–[18], to name just a few examples. Despite the fact that the replica method lacks mathematical rigor, a
5
body of work, such as [5], [19]–[25], has shown the validity of several replica-based results in the literature, e.g.,
Tanaka’s formula in [11], using some alternative rigorous approaches. We later discuss these rigorous results with
more details by invoking the literature of compressive sensing.
B. Decoupling Principle
Considering the MAP estimator defined in (2), the entries of the estimated vector x̂ are correlated in general,
since the system matrix couples the entries of x linearly, and g(·|A) performs several nonlinear operations on y.
In the large-system performance analysis, the marginal joint distribution of two corresponding input-output entries
xj and x̂j , 1 ≤ j ≤ n, is of interest. To clarify our point, consider the case in which a linear estimator is employed
instead of (2), i.e., x̂ = GT y. Denote the matrices A and G as A = [a1 · · · an ] and G = [g1 · · · gn ], respectively,
with ai and gi being k × 1 vectors for 1 ≤ i ≤ n. Therefore, x̂j is written as
x̂j = gjT y
#
" n
X
T
xi ai + z
= gj
(13a)
(13b)
i=1
= (gjT aj ) xj +
n
X
(gjT ai ) xi + gjT z.
(13c)
i=1,i6=j
Here, the right hand side of (13c) can be interpreted as the linear estimation of a single-user system indexed by j in
which the symbol xj is corrupted by an additive impairment given by the last two summands in the right hand side
of (13c). The impairment term is not necessarily independent and Gaussian. For some classes of matrix ensembles,
and under a set of assumptions, it is shown that the dependency of the derived single-user systems on the index
j vanishes, and the distribution of the impairment terms converges to a Gaussian distribution in the large-system
limit [26]. As a result, one can assume the vector-valued system described by (1) followed by the linear estimator
G to be a set on n additive scalar systems with Gaussian noise which have been employed in parallel. In other
words, the vector system can be considered to “decouple” into a set of similar scalar systems. Each of them relates
an input entry xj to its corresponding estimated one x̂j . This asymptotic property of the estimator is referred to as
the “decoupling property” and can be investigated through the large-system performance analysis.
The decoupling property was first studied for linear estimators. Tse and Hanly noticed this property while they
were determining the multiuser efficiency of several linear multiuser estimators in the large-system limit [27]. They
showed that for an i.i.d. system matrix, the effect of impairment is similar to the effect of some modified Gaussian
noise when the dimension tends to infinity. This asymptotic property was then investigated further by studying the
asymptotics of different linear receivers and their large-system distributions [28], [29]. In an independent work,
Verdú and Shamai also studied the linear Minimum Mean Square Error (MMSE) estimator and showed that the
conditional output distribution is asymptotically Gaussian [30]. In [31], the authors studied the asymptotics of the
impairment term when a family of linear estimators is employed and proved that it converges in distribution to a
Gaussian random variable. The latter result was further extended to a larger class of linear estimators in [26].
6
Regarding linear estimators, the main analytical tool is random matrix theory [32], [33]. In fact, invoking properties
of large random matrices and the central limit theorem, the decoupling property is rigorously proved, e.g. [34],
[35]. These tools, however, fail for nonlinear estimators as the source symbol and impairment term do not decouple
linearly due to nonlinear operations at the estimators. In [36], Müller and Gerstacker employed the replica method
and studied the capacity loss due to the separation of detection and decoding. The authors showed that the additive
decoupling of the spectral efficiency, reported in [35] for Gaussian inputs, also holds for binary inputs. As a result, it
was conjectured that regardless of input distribution and linearity, the spectral efficiency always decouples in an
additive form [37]. In [14], Guo and Verdú justified this conjecture for a family of nonlinear MMSE estimators,
and showed that for an i.i.d. system matrix, the estimator decouples into a bank of single-user MMSE estimators
under the Replica Symmetry (RS) ansatz. In [38], Rangan et al. studied the asymptotic performance of a class of
MAP estimators. Using standard large deviation arguments, the authors represented the MAP estimator as the limit
of an indexed MMSE estimators’ sequence. Consequently, they determined the estimator’s asymptotics employing
the results from [14] and justified the decoupling property of MAP estimators under the RS ansatz for an i.i.d. A.
Regarding the decoupling property of MAP estimators, there are still two main issues which need further investigations: 1) cases in which the system matrix A is not i.i.d., and 2) the analysis of the estimator under the Replica
Symmetry Breaking (RSB) ansätze. The first issue was partially addressed in [39] where, under the RS assumption,
the authors studied the asymptotics of a MAP estimator employed to recover the support of a source vector from
observations received through noisy sparse random measurements. They considered a model in which a sparse
Gaussian source vector1 is first randomly measured by a square matrix V, and then, the measurements are sparsely
sampled by a diagonal matrix B whose non-zero entries are i.i.d. Bernoulli random variables. For this setup, the
input-output information rate and support recovery error rate were investigated by considering the measuring matrix
V to belong to a larger set of matrix ensembles. These results, moreover, could address the decoupling property
of the considered setting. Although the class of system matrices is broadened in [39], it cannot be considered as
a complete generalization of the property presented in [14] and [38], since it is restricted to cases with a sparse
Gaussian source and loading factors less than one, i.e., kn−1 < 1 in (1). Vehkaperä et al. also tried to investigate
the first issue for a similar formulation in compressive sensing [40]. In fact, the authors considered a linear sensing
model as in (1) for the class of rotationally invariant random matrices2 and under the RS ansatz determined the
asymptotic Mean Square Error (MSE) for the least-square recovery schemes which can be equivalently represented
by the formulation in (2). The large-system results in [40], however, did not address the asymptotic marginal joint
input-output distribution, and the emphasis was on the MSE. Regarding the second issue, the MAP estimator has
not yet been investigated under RSB ansätze in the literature. Nevertheless, the necessity of such investigations was
mentioned for various similar settings in the literature; see for example [41]–[43]. In [41], the performances of Code
Division Multiple Access (CDMA) detectors were investigated by studying both the RS and one-step RSB ansätze
and the impact of symmetry breaking onto the results for low noise scenarios were discussed. The authors in [43]
1 It
means that the entries of the source vector are of the form xi bi where xi and bi are Gaussian and Bernoulli random variables, respectively.
2 The
class of rotationally invariant random matrices is precisely defined later throughout the problem formulation.
7
further studied the performance of vector precoding under both RS and RSB and showed that the analysis under
RS yields a significantly loose bound on the true performance. The replica ansatz with one-step of RSB, however,
was shown to lead to a tighter bound consistent with rigorous performance bound available in the literature. A
similar observation was recently reported for the problem of least-square-error precoding in [44], [45]. The replica
analyses of compressive sensing in [42], [46], moreover, discussed the necessity of investigating the performance
of ℓp minimization recovery schemes under RSB ansätze for some choices of p.
C. Compressive Sensing
The MAP estimation of a source vector from a set of noisy linear observations arises in several applications,
such as MIMO and sampling systems. To address one, we consider large compressive sensing systems and employ
our asymptotic results to analyze the large-system performance [47]–[49]. In context of compressive sensing, (1)
represents a noisy sampling system in which the source vector x is being sampled linearly via A and corrupted by
additive noise z. In the “noise-free” case, i.e. λ0 = 0, the source vector x is exactly recovered from the observation
vector y, if the number of observations k is as large as the source length n and the sampling matrix A is full rank.
As the number of observations reduces, the possible answers to the exact reconstruction problem may increase
regarding the source support
X,
and therefore, the recovered source vector from the observation vector is not
necessarily unique. In this case, one needs to enforce some extra constraints on the properties of the source vector
in order to recover it uniquely among all possible solutions. In compressive sensing, the source vector is supposed
to be sparse, i.e., a certain fraction of entries are zero. This property of the source imposes an extra constraint on
the solution which allows for exact recovery in cases with k < n. In fact, in this case, one should find a solution
to y = Av over
S = {vn×1 ∈ Xn :
kvk0 < αn}
where k·k0 denotes the “ℓ0 norm” and is defined as kvk0 :=
Pn
i=1
(14)
1{vi 6= 0}, and α ≤ 1 is the source’s sparsity
X, the latter problem can have a unique
solution even for k < n [50]–[52]. Searching for this solution optimally over S, however, is an NP-hard problem and
factor defined as the fraction of non-zero entries. Depending on A and
therefore intractable. The main goal in noise-free compressive sensing is to study feasible reconstruction schemes
and derive tight bounds on the sufficient compression rate, i.e., k/n, for exact source recovery via these schemes.
In noisy sampling systems, exact recovery is only possible for some particular choices of
considering either cases in which exact recovery is not possible or choices of
X.
Nevertheless,
X for which the source vector can
be exactly recovered from noisy observations, the recovery approaches in these sensing systems need to take the
impact of noise into account. The classical strategy in this case is to find a vector in
S such that the recovery
distortion is small. Consequently, a recovery scheme for noisy sensing system based on the ℓ0 norm is given by
1
(15)
ky − Avk2 + kvk0
x̂ = arg minn
v∈X
2λ
which is the MAP estimator defined in (2) with u(v) = kvk0 . It is straightforward to show that for λ0 = 0, i.e.,
zero noise variance, (15) reduces to the optimal noise-free recovery scheme as λ ↓ 0. Similar to the noise-free
8
case, the scheme in (15) results in a non-convex optimization problem, and therefore, is computationally infeasible.
Alternatively, a computationally feasible schemes is introduced by replacing the ℓ0 norm in the cost function with
the ℓ1 norm. The resulting recovery scheme is known as LASSO [53] or basis pursuit denoising [54]. Based on
these formulations, several iterative and greedy algorithms have been introduced for recovery taking into account
the sparsity pattern and properties of the sampling matrix [55]–[57]. The main body of work in noisy compressive
sensing then investigates the trade-off between the compression rate and recovery distortion.
For large compressive sensing systems, it is common to consider a random sensing matrix, since for these matrices,
properties such as restricted isometry property are shown to hold with high probability [58]. In this case, the
performance of a reconstruction schemes is analyzed by determining the considered performance metric, e.g., MSE
and probability of exact recovery in the noisy and noise-free case, respectively, for a given realization of the sensing
matrix. The average performance is then calculated by taking the expectation over the matrix distribution. Comparing
(15) with (2), one can utilize the MAP formulation, illustrated at the beginning of this section, to study the largesystem performance of several reconstruction schemes. This similarity was considered in a series of papers, e.g.,
[17], [38], and therefore, earlier replica results were employed to study compressive sensing systems. The extension
of analyses from the context of multiuser estimation had the disadvantage that the assumed sampling settings were
limited to those setups which are consistent with the estimation problems in the literature. Compressive sensing
systems, however, might require a wider set of assumptions, and thus, a large class of settings could not be addressed
by earlier investigations. As the result, a body of work deviated from this approach and applied the replica method
directly to the compressive sensing problem; see for example [39], [40], [42], [59]–[61].
Although in general the replica method is considered to be mathematically non-rigorous, several recent studies
have justified the validity of the replica results in the context of compressive sensing by using some alternative tools
for analysis. A widely investigated approach is based on the asymptotic analysis of Approximate Message Passing
(AMP) algorithms. In the context of compressive sensing, the AMP algorithms were initially introduced to address
iteratively the convex reconstruction schemes based on ℓ1 norm minimization, such as LASSO and basis pursuit,
with low computational complexity [62], [63]. The proposed approach was later employed to extend the algorithm
to a large variety of estimation problems including MAP and MMSE estimation; see for example [64], [65]. The
primary numerical investigations of AMP demonstrated that for large sensing systems the sparsity-compression rate
tradeoff of these iterative algorithms, as well as the compression rate-distortion tradeoff in noisy cases, is derived
by the fixed-points of “state evolution” and recovers the asymptotics of convex reconstruction schemes [63]. This
observation was then rigorously justified for i.i.d. sub-Gaussian sensing matrices in [66] by using the conditioning
technique developed in [67]. The study was recently extended to cases with rotationally invariant system matrices
in [68], [69]. The investigations in [70], [71] moreover showed that using AMP algorithms for spatially coupled
measurements, the fundamental limits on the required compression rate [72], [73] can be achieved in the asymptotic
regime. The methodology proposed by AMP algorithms and their state evolution also provided a justification for
validity of several earlier studies based on the replica method. In fact, the results given by the replica method were
recovered through state evolution of the corresponding AMP algorithms. Invoking this approach along with other
analytical tools, the recent study in [23] further approved the validity of the replica prediction for the asymptotic
9
MMSE and mutual information of the linear system in (1) with i.i.d. Gaussian measurements. Similar results were
demonstrated in [21] using a different approach.
D. Contribution and Outline
In this paper, we determine the asymptotic distortion for a general distortion function for cases where the MAP
estimator is employed to estimate the source vector from the observation given in (1). We represent the asymptotic
distortion in (3) as a macroscopic parameter of a corresponding spin glass and study the spin glass via the replica
method. The general replica ansatz is then given for an arbitrary replica correlation matrix, and its special cases are
studied considering the RS and RSB assumptions. The asymptotic distortion is determined for rotationally invariant
random system matrices invoking results for asymptotics of spherical integrals [74], [75]. Using our asymptotic
results, we derive a more general form of the decoupling principle by restricting the distortion function to be of a
special form and employing the moment method [76], [77]. We show that the vector-valued system in (1) estimated
by (2) decouples into a bank of similar noisy single-user systems followed by single-user MAP estimators. This
result holds for any replica correlation matrix; however, the structure of the decoupled single-user system depends
on the supposed structure of the correlation matrix. Under the RSB assumption with b steps of breaking (bRSB), the
noisy single-user system is given in the form of an input term added by an impairment term. The impairment term,
moreover, is expressed as the sum of an independent Gaussian random variable and b correlated interference terms.
By reducing the assumption to RS, the result reduces to the formerly studied RS decoupling principle of the MAP
estimators [38] for rotationally invariant system matrix ensembles. In fact, our investigations collect the previous
results regarding the decoupling principle in addition to a new set of setups under a single umbrella given by a
more general form of the decoupling principle. More precisely, we extend the scope of the decoupling principle to
•
the systems whose measuring matrix belongs to the class of rotationally invariant random matrices, and
•
the replica ansatz with general replica correlations which include the RS and RSB ansätze.
To address a particular application, we study the large-system performance of a compressive sensing system under
the ℓp minimization recovery schemes. We address the linear reconstruction, as well as the LASSO and the ℓ0 norm
scheme considering both the sparse Gaussian and finite alphabet sources. Our general setting allows to investigate
the asymptotic performance with respect to different metrics and for multiple sensing matrices such as random i.i.d.
and projector. The numerical investigations show that the RS ansatz becomes unstable for some regimes of system
parameters and predicts the performance of ℓ0 minimization recovery loosely within a large range of compression
rates. This observation agrees with the earlier discussions on the necessity of RSB investigations reported in [42],
[46]. We therefore study the performance under RSB and discuss the impact of the symmetry breaking. Throughout
the numerical investigations, it is demonstrated that the performance enhancement obtained via random orthogonal
measurements, reported in [40], also holds for sparse finite alphabet sources in which sensing via random projector
matrices results in phase transitions at higher rates.
The rest of the manuscript is organized as follows. In Section II, the problem is formulated. We illustrate our
statistical mechanical approach in Section III and explain briefly the replica method. The general replica ansatz, as
well as the general decoupling principle is given in Section IV. The ansatz under the RS and RSB assumptions are
10
expressed in Sections V and VI. Based on bRSB decoupled system, we propose the idea of a replica simulator in VII
and describe the given ansätze in terms of the corresponding decoupled systems. To address an application of our
study, we consider large compressive sensing systems in Section VIII and discuss several examples. The numerical
investigations of the examples are then given in Section IX. Finally, we conclude the manuscript in Section X.
E. Notations
Throughout the manuscript, we represent scalars, vectors and matrices with lower, bold lower, and bold upper
case letters, respectively. The set of real numbers is denoted by
R, and AT and AH indicate the transposed and
Hermitian of A. Im is the m × m identity matrix and 1m is an m × m matrix with all entries equal to 1. For
a random variable x, px represents either the Probability Mass Function (PMF) or Probability Density Function
(PDF), and Fx represents the Cumulative Distribution Function (CDF). We denote the expectation over x by Ex , and
an expectation over all random variables involved in a given expression by E.
and real numbers and the superscript +, e.g.
R
+
Z and R represent the set of integer
, indicates the corresponding subset of all non-negative numbers.
For sake of compactness, the set of integers {1, . . . , n} is abbreviated as [1 : n], the zero-mean and unit-variance
Gaussian PDF is denoted by φ(·), and the Gaussian averaging is expressed as
Z
Z +∞
(·) φ(t)dt.
(·) Dt :=
(16)
−∞
Moreover, in many cases, we drop the set on which a sum, minimization or an integral is taken. Whenever needed,
we consider the entries of x to be discrete random variables, namely the support
X to be discrete. The results of
this paper, however, are in full generality and hold also for continuous distributions.
F. Announcements
Some of the results of this manuscript were presented at the IEEE Information Theory Workshop [78] and the
Information Theory and Applications Workshop [79]. Even though the results have a mathematical flavor, the stress
is not on investigating the rigor of the available tools such as the replica method, but rather to employ them for
deriving formulas which can be used in different problems.
II. P ROBLEM F ORMULATION
Consider the vector-valued linear system described by (1). Let the system satisfy the following properties.
(a) xn×1 is an i.i.d. random vector with each entry being distributed with px over
(b) Ak×n is randomly generated over A
k×n
X ⊆ R.
with A ⊆ R from rotationally invariant random ensembles. The random
matrix A is said to be rotationally invariant when its Gramian, i.e., J = AT A, has the eigendecomposition
J = UDUT
(17)
11
with Un×n being an orthogonal Haar distributed matrix and Dn×n being a diagonal matrix. For a given n,
we denote the empirical CDF of J’s eigenvalues (cumulative density of states) with FnJ and define it as
n
FnJ (λ) =
1X
1{λJi < λ}.
n i=1
(18)
where λJi for i ∈ [1 : n] denotes the ith diagonal entry of D. We assume that FnJ converges, as n ↑ ∞, to a
deterministic CDF FJ .
(c) z k×1 is a real i.i.d. zero-mean Gaussian random vector in which the variance of each entry is λ0 .
(d) The number of observations k is a deterministic function of the transmission length n, such that
lim
n↑∞
k(n)
1
= < ∞.
n
r
(19)
For sake of compactness, we drop the explicit dependence of k on n.
(e) x, A and z are independent.
The source vector x is reconstructed from the observation vector y with a system matrix A that is known at
the estimator. Thus, for a given A, the source vector is recovered by x̂ = g(y|A) where g(·|A) is given in (2).
Here, the non-negative scalar λ is the estimation parameter and the non-negative cost function u(·) is referred to
the “utility function”. The utility function u(·) is supposed to decouple which means that it takes arguments with
different lengths, i.e., u(·) : Rℓ 7→ R+ for any positive integer ℓ, and
u(x) =
n
X
u(xi ).
(20)
i=1
In order to use the estimator in (2), one needs to guarantee the uniqueness of the estimation output. Therefore, we
impose the following constraint on our problem.
(f) For a given observation vector y, the objective function in (2) has a unique minimizer over the support
Xn .
A. MAP Estimator
The MAP estimator in (2) can be considered as the optimal estimator in the sense that it minimizes the
reconstruction’s error probability postulating a source prior distribution proportional to e−u(x) and a noise variance λ.
To clarify the argument, assume X is a finite set1 . In this case, we can define the reconstruction’s error probability as
Pe = Pr{x 6= g̃(y|A)}
1 i.e.,
the entries of x are discrete random variables.
(21)
12
for some estimator g̃(·|A). In order to minimize Pe , g̃(·|A) is chosen such that the posterior distribution over the
input support
Xn is maximized, i.e.,
g̃(y|A) = arg max px|y,A (v|y, A)
(22a)
v
= arg max py|x,A (y|v, A)px|A (v|A)
(22b)
⋆
= arg min − log py|x,A (y|v, A) − log px (v) .
(22c)
v
v
where ⋆ comes from the independency of x and A. Here, py|x,A (y|v, A) = pz (y − Av), and px is the prior
distribution of the source vector. Now, let the estimator postulate the noise variance to be λ and the prior to be
e−u(v)
px (v) = P −u(v)
ve
(23)
for some non-negative function u(·). Substituting into (22c), the estimator g̃(·|A) reduces to g(·|A) defined in (2).
The estimator in (2) models several particular reconstruction schemes in compressive sensing. We address some of
these schemes later on in Section VIII.
Remark II.1. In the case that the entries of x are continuous random variables, the above argument needs some
modifications. In fact, in this case the reconstruction’s error probability as defined in (21) is always one, and
therefore, it cannot be taken as the measure of error. In this case, the MAP estimator is considered to maximize
the posterior PDF postulating px (v) ∝ e−u(v) and noise variance λ.
B. Asymptotic Distortion and Conditional Distribution
In many applications, the distortion is given in terms of the average MSE, while in some others the average
symbol error rate is considered. In fact, the former takes the ℓ2 norm as the distortion function, and the latter
considers the ℓ0 norm. The distortion function, however, can be of some other form in general. Here, we study the
asymptotic performance by considering a general distortion function which determines the imperfection level of the
estimation. Thus, we consider a distortion function d(·; ·) which
d(·; ·) : X × X → R.
(24)
The term “average distortion” usually refers to the case when the averaging weights are uniform. It means that
each tuple of source-estimated entries (xi , x̂i ) is weighted equally when the distortion is averaged over all the
entries’ tuples. It is however possible to average the distortion by a non-uniform set of weights. In the following,
we define the average distortion for a class of binary weights which includes the case of uniform averaging, as well.
Xn , and let d(·; ·) be a
distortion function defined in (24). Define the index set W(n) as a subset of [1 : n], and let |W(n)| grow with n
Definition II.1 (Asymptotic distortion). Consider the vectors xn×1 and x̂n×1 defined over
such that
|W(n)|
=η
n↑∞
n
lim
(25)
13
for some η ∈ [0, 1]. Then, the average distortion of x̂ and x over the index set
W(n) regarding the distortion
function d(·; ·) is defined as
DW(n) (x̂; x) :=
1
|W(n)|
X
w∈W(n)
d(x̂w ; xw ).
(26)
Assuming the limit of (26) when n ↑ ∞ exists, we denote
DW (x̂; x) := lim DW(n) (x̂; x)
(27)
n↑∞
and refer to it as the “asymptotic distortion” over the limit of the index set
W(n).
Definition II.1 is moreover utilized to investigate the asymptotic conditional distribution of the estimator which
plays a key role in studying the decoupling principle. For further convenience, we define the asymptotic conditional
distribution of the MAP estimator as follows.
Definition II.2 (Asymptotic conditional distribution). Consider the source vector xn×1 passed through the linear
system in (1), and let x̂n×1 be its MAP estimation as given in (2). For a given n, we take an index j ∈ [1 : n] and
denote the conditional distribution of x̂j given xj by px̂|x which at the mass point (v̂, v) ∈ X × X reads
j(n)
−1
j(n)
px̂j ,xj (v̂, v)
px̂|x (v̂|v) := pxj (v)
(28)
with px̂j ,xj (v̂, v) being the marginal joint distribution of xj and x̂j at the mass point (v̂, v). Then, in the large-system
limit, we define the asymptotic conditional distribution of x̂j given xj at (v̂, v) as
j(n)
pjx̂|x (v̂|v) := lim px̂|x (v̂|v).
(29)
n↑∞
We study the asymptotic distortion over the limit of a desired index set
W(n) and distortion function d(·; ·)
by defining it as a macroscopic parameter of the corresponding spin glass and employing the replica method to
evaluate it. Using the result for the asymptotic distortion, we determine then the asymptotic conditional distribution
and investigate the decoupling property of the estimator.
III. S TATISTICAL M ECHANICAL A PPROACH
The Hamiltonian in (10) introduces a spin glass which corresponds to the MAP estimator. The spin glass at
zero temperature describes the asymptotics of the MAP estimator. For further convenience, we formally define the
“corresponding spin glass” as follows.
Definition III.1 (Corresponding spin glass). Consider the integer n ∈
Z+ . The corresponding spin glass with the
microstate v n×1 ∈ Xn and the quenched randomizers y n×1 and Ak×n is given by
•
A is a rotationally invariant random matrix, and y is constructed as in (1) from the source vector x.
14
•
For any realization of A and y, the Hamiltonian reads
E(v|y, A) =
1
ky − Avk2 + u(v).
2λ
(30)
for v ∈ Xn .
For the corresponding spin glass, at the inverse temperature β, the following properties are directly concluded.
•
The conditional distribution of the microstate reads
e−βE(v|y,A)
Z(β|y, A)
(31)
e−βE(v|y,A) .
(32)
1
E log Z(β|y, A),
βn
(33)
pβ
v|y,A (v|y, A) =
with Z(β|y, A) being the partition function
Z(β|y, A) =
•
X
v
The normalized free energy is given by
F(β) = −
where the expectation is taken over the quenched randomizers.
•
The entropy of the spin glass is determined as
H(β) = β2
d
[F(β)] .
dβ
(34)
Regarding the MAP estimator, one can represent the asymptotic distortion as a macroscopic parameter of the
corresponding spin glass. More precisely, using Definition II.2, the asymptotic distortion reads
W(n) (v; x)
DW (x̂; x) = lim lim Eβ
vD
(35)
β↑∞ n↑∞
where Eβ
v indicates the expectation over v ∈
Xn
with respect to the conditional Boltzmann-Gibbs distribution
W 1
−1
as
pβ
v|y,A defined in (31). In fact, by introducing the macroscopic parameter D (β) at the temperature β
W(n) (v; x),
DW (β) := lim Eβ
vD
(36)
n↑∞
the asymptotic distortion can be interpreted as the macroscopic parameter at zero temperature. Here, we take a
well-known strategy in statistical mechanics which modifies the partition function to
Z(β, h) =
1 In
W
X
e−βE(v|y,A)+hnD
W(n) (v;x)
.
v
general, D (β) is a function of β, y, x and A. However, we drop the other arguments for sake of compactness.
(37)
15
In this case, the expectation in (36) is taken as
DW (β) = lim lim
n↑∞ h↓0
1 ∂
log Z(β, h).
n ∂h
(38)
The macroscopic parameter defined in (36) is random, namely depending on the quenched randomizer {y, A}. As
discussed in Section I-A, under the self averaging property, the macroscopic parameter is supposed to converge
in the large-system limit to its expected value over the quenched random variables. For the corresponding spin
glass defined in Definition III.1, the self averaging property has not been rigorously justified, and the proof requires
further mathematical investigations as in [4]. However, as it is widely accepted in the literature, we assume that
the property holds at least for the setting specified here. Therefore, we state the following assumption.
Assumption 1 (Self Averaging Property). Consider the corresponding spin glass defined in Definition III.1. For
almost all realizations of the quenched randomizers A and y,
DW (β) = Ey,A DW (β).
(39)
Using the self averaging property of the system, the asymptotic distortion is written as
DW (x̂; x) = lim lim lim
β↑∞ n↑∞ h↓0
1 ∂
E log Z(β, h).
n ∂h
(40)
Evaluation of (40), as well as the normalized free energy defined in (33), confronts the nontrivial problem of
determining the logarithmic expectation. The task can be bypassed by using the Riesz equality [80] which for a
given random variable t states that
E log t = lim
m↓0
1
log E tm .
m
(41)
Using the Riesz equality, the asymptotic distortion can be finally written as
DW (x̂; x) = lim lim lim lim
β↑∞ n↑∞ h↓0 m↓0
1 1 ∂
log E[Z(β, h)]m .
m n ∂h
(42)
(42) expresses the asymptotic distortion in terms of the moments of the modified partition function; however, it
does not yet simplify the problem. In fact, one faces two main difficulties when calculating the right hand side of
(42): 1) the moment needs to be evaluated for real values of m (at least within a right neighborhood of 0), and
2) the limits need to be taken in the order stated. Here is where the replica method plays its role. The replica
method suggests to determine the moment for an arbitrary non-negative integer m as an analytic function in m and
then assume the two following statements:
1) The moment function analytically continues from the set of integer numbers onto the real axis (at least for
some m at a right neighborhood of 0) which means that an analytic expression found for integer moments
directly extends to all (or some) real moments. Under this assumption, the expression determined for integer
moments can be replaced in (42), and the limit with respect to m taken when m ↓ 0. This assumption is the
16
main part where the replica method lacks rigor and is known as the “replica continuity”.
2) In (42), the limits with respect to m and n exchange. We refer to this assumption as the “limit exchange”.
In order to employ the replica method, we need to suppose the validity of the above two statements; therefore,
we state the following assumption.
Assumption 2 (Replica Continuity and Limit Exchange). For the spin glass defined in Definition III.1, the replica
continuity and limit exchange assumptions hold.
By means of Assumption 2, the calculation of asymptotic distortion reduces to the evaluation of integer moments
of the modified partition function which is written as
Z(m) := E[Z(β, h)]m
=E
m
Y
a=1
X
(43a)
e−βE(va |y,A)+hnD
W(n)(va ;x)
(43b)
va
= Ex EA Ez
X
{v a }
e−β
Pm
a=1
E(va |y,A)+hn
Pm
a=1
DW(n) (v a ;x)
.
(43c)
Here, we refer to v a ∈ Xn for a ∈ [1 : m] as the replicas, and define {v a } := {v1 , . . . , v m } ∈ Xn × · · · × Xn as the
set of replicas. After taking the expectation with respect to z and A, it is further observed that, in the large limit,
the expectation with respect to x can be dropped due to the law of large numbers. By inserting the final expression
for Z(m) in (42) and taking the limits, the asymptotic distortion is determined as in Proposition IV.1 given below.
IV. M AIN R ESULTS
Proposition IV.1 states the general replica ansatz. The term “general” is emphasized here, in order to indicate
that no further assumption needs to be considered for derivation. Using Proposition IV.1 along with results in the
classical moment problem [77], a general form of the decoupling principle is justified for the MAP estimator.
Before stating the general replica ansatz, let us define the R-transform of a probability distribution. Considering
a random variable t, the corresponding Stieltjes transform over the upper complex half plane is defined as
Gt (s) = E (t − s)−1 .
(44)
Denoting the inverse with respect to composition by G−1
t (·), the R-transform is given by
Rt (ω) = Gt−1 (−ω) − ω −1
(45)
such that limω↓0 Rt (ω) = E t.
The definition can also be extended to matrix arguments. Assume the matrix Mn×n to have the decomposition
M = UΛU−1 where Λn×n is a diagonal matrix whose nonzero entries represent the eigenvalues of M, i.e.,
Λ = diag[λ1 , . . . , λn ], and Un×n is the matrix of eigenvectors. Rt (M) is then an n × n matrix defined as
Rt (M) = U diag[Rt (λ1 ), . . . , Rt (λn )] U−1 .
(46)
17
A. General Replica Ansatz
Proposition IV.1 expresses the macroscopic parameters of the corresponding spin glass, including the asymptotic
distortion, in terms of the parameters of a new spin glass of finite dimension. It is important to note that the new
spin glass, referred to as “spin glass of replicas”, is different from the corresponding spin glass defined in Definition
III.1. In fact, the spin glass of replicas is the projection of the corresponding spin glass on the reduced support Xm
with m indicating the number of replicas. The macroscopic parameters of the spin glass of replicas can therefore
readily be determined.
Definition IV.1 (Spin glass of replicas). For the finite integer m, the spin glass of replicas with the microstate
vm×1 ∈ Xm and quenched randomizer xm×1 is defined as follows.
•
All the entries of x equal to x where x is a random variable distributed with the source distribution px .
•
For a given realization of x, the Hamiltonian reads
E R (v|x) = (x − v)T TRJ (−2βTQ)(x − v) + u(v)
where RJ (·) is the R-transform corresponding to FJ , Tm×m is defined as
1
βλ0
T :=
Im −
1m ,
2λ
λ + mβλ0
(47)
(48)
and Qm×m is the so-called replica correlation matrix defined as
Q = E (x − v)(x − v)T .
(49)
with the expectation taken over x and v at thermal equilibrium.
•
At thermal equilibrium, the microstate is distributed according to the Boltzmann-Gibbs distribution pβ
v|x
R
e−βE (v|x)
=
Z R (β|x)
pβ
v|x (v|x)
(50)
where Z R (β|x) denotes the partition function of the system defined as
Z R (β|x) :=
•
X
e−βE
R
(v|x)
.
(51)
v
The normalized free energy of the system at the inverse temperature β is given by
FR (β, m) = −
1
E log Z R (β|x),
βm
(52)
where the expectation is taken over x with respect to px . The average energy and entropy at the inverse
temperature β are further found using (8a) and (8b).
•
For the system at thermal equilibrium, the replicas’ average distortion regarding the distortion function d(·, ·)
at the inverse temperature β is
m
1 X
d(va , x),
D (β, m) = E
m a=1
R
(53)
18
with expectation being taken over x and v with respect to px pv|x .
Considering Definition IV.1, the evaluation of the system parameters such as the replicas’ average distortion
DR (β, m) or the normalized free energy FR (β, m) needs the replica correlation matrix Q to be explicitly calculated
first. In fact, (49) describes a fixed point equation in terms of Q when one writes out the expectation using the
conditional distribution in (50). The solution can then be substituted in the distribution and the parameters of the
system can be calculated via (51)-(53). The fixed point equation, however, may have several solutions and thus
result in multiple outputs for the system. Nevertheless, we express the asymptotic distortion of the MAP estimator
in terms of a single output of the spin glass of replicas for which the limits exist and the free energy is minimized.
Proposition IV.1. Let the linear system (1) fulfill the constraints of Section II. Suppose Assumptions 1 and 2 hold,
and consider the spin glass of replicas as defined in Definition IV.1 for a finite integer m. Then, the free energy of
the corresponding spin glass defined in Definition III.1 is given by
Z 1
1
T
Tr{TQRJ (−2βωTQ)}dω − Tr{Q TRJ (−2βTQ)} + FR (β, m)
F(β) = lim
m↓0 m
0
(54)
where Q is the replica correlation matrix satisfying (49). The asymptotic distortion of the MAP estimator regarding
the distortion function d(·; ·) is then determined as
DW (x̂; x) = lim lim DR (β, m).
β↑∞ m↓0
(55)
In case that multiple solutions are available for the replica correlation matrix, the replica’s average distortion in
(55) is evaluated via a correlation matrix which minimizes the free energy at zero temperature, i.e., β ↑ ∞.
Proof. The proof is given in Appendix A. However, we explain briefly the strategy in the following.
Starting from (43c), the expectation with respect to the noise term is straightforwardly taken. Using the results
from [75], the expectation with respect to the system matrix is further taken as discussed in Appendix D. Then, by
considering the following variable exchange,
[Q]ab =
1
(x − v a )T (x − v b ).
n
(56)
Z(m) is determined in terms of the replica correlation matrix Q. Finally, by employing the law of large numbers,
the mth moment of the partition function is given as
Z
Z(m) = Ex e−n[G(TQ)−I(Q)]+ǫn dQ
where dQ :=
Qm
a,b=1
d[Q]ab with the integral being taken over
(57)
Rm×m , ǫn tends to zero as n ↑ ∞ and T is given
by (48). Moreover, enI(Q) denotes the non-normalized probability weight of the vectors of replicas with a same
19
correlation matrix and is explicitly determined in (198b) in Appendix A, and G(·) reads
Z β
G(M) =
Tr{MRJ (−2ωM)}dω
(58)
0
for some diagonal matrix M with Tr{·} denoting the trace operator, and RJ (·) being the R-transform with respect
to FJ . In (57), the term enI(Q) dQ is a probability measure which satisfies the large deviations property. Using
results from large deviations [81], the integral in (57) for large values of n can be written as the integrand at the
saddle point Q̃ multiplied by some bounded coefficient Kn which results in
.
Z(m) = Kn e−n[G(TQ̃)−I(Q̃)]
(59)
.
with = denoting the asymptotic equivalence in exponential scale1 . Consequently, by substituting Z(m) in (42), and
exchanging the limits with respect to n and m, as suggested in Assumption 2, the asymptotic distortion is found
as in Proposition IV.1 where (49) determines the saddle point of the integrand function in (57). The free energy is
further determined as in (54) by substituting (59) in (33). Finally by noting the fact that the free energy is minimized
at the equilibrium, the proof is concluded.
N
Proposition IV.1 introduces a feasible way to determine the asymptotics of the MAP estimator; its validity depends
only on Assumptions 1 and 2 and has no further restriction. To pursue the analysis, one needs to solve the fixed
point equation (49) for the replica correlation matrix Q and calculate the parameters of the spin glass of replicas
explicitly. The direct approach to find Q, however, raises both complexity and analyticity issues. In fact, finding
the saddle point by searching over the set of all possible choices of Q is a hard task to do; moreover, several
solutions may not be of use since they do not lead to analytic FR (β, m) and DR (β, m) in m, and thus, they cannot
be continued analytically to the real axis via Assumption 2.
To overcome both the issues, the approach is to restrict our search into a parameterized set of replica correlation
matrices and find the solution within this set. Clearly, the asymptotics found via this approach may fail as several
other solutions are missed by restriction. The result, in this case, becomes more trustable by extending the restricted
set of replica correlation matrices. Several procedures of restrictions can be considered. The procedures introduced
in the literature are roughly divided into RS and RSB schemes. The former considers the m replicas to interact
symmetrically while the latter recursively breaks this symmetry in a systematic manner. In fact, the RS scheme was
first introduced due to some symmetric properties observed in the analysis of the spin glasses [82]. The properties,
however, do not force the correlation matrix to have a symmetric structure, and later several examples were found
showing that RS leads to wrong conclusions. For these examples, the RSB scheme was further considered as an
extension of the symmetric structure of the correlation matrix to a larger set. We consider both the RS and RSB
schemes in this manuscript; however, before pursuing our study, let us first investigate the general decoupling
property of the estimator which can be concluded from Proposition IV.1.
1 a(·)
and b(·) are asymptotically equivalent in exponential scale over a non-bounded set
a(x)
| = 0.
X, if for x ∈ X we have x↑∞
lim log |
b(x)
20
B. General Decoupling Property of MAP Estimator
Regardless of any restriction on Q, the general ansatz leads to the decoupling property of the MAP estimator.
In fact by using Proposition IV.1, it can be shown that for almost any tuple of input-output entries, the marginal
joint distribution converges to a deterministic joint distribution which does not depend on the entries’ index. The
explicit term for the joint distribution, however, depends on the assumptions imposed on the correlation matrix.
Proposition IV.2 (General Decoupling Principle). Let the linear system (1) fulfill the constraints of Section II.
Suppose Assumptions 1 and 2 hold, and consider the spin glass of replicas as defined in Definition IV.1 with the
replica correlation matrix Q. Then, for j ∈ [1 : n], the asymptotic conditional distribution of the MAP estimator
pjx̂|x defined in Definition II.2 is, almost sure in A, independent of j, namely
pjx̂|x (v̂|v) = px̂|x (v̂|v)
(60)
for some conditional distribution px̂|x at the mass point (v̂, v) ∈ X × X. Consequently, the marginal joint distribution
of the entries xj and x̂j is described by the input-output joint distribution of the single-user system specified by
the conditional distribution px̂|x and the input x ∼ px . The explicit form of px̂|x depends on Q.
Proof. The proof follows from Proposition IV.1 and the moment method [76], [77]. From the classical moment
problem, we know that the joint probability of the tuple of random variables (t1 , t2 ) are uniquely specified with
the sequence of integer joint moments, if the joint moments are uniformly bounded. More precisely, by defining
the (k, ℓ)-joint moment of the tuple (t1 , t2 ) as
Mk,ℓ := E tk1 tℓ2 ,
(61)
the sequence of {Mk,ℓ } for (k, ℓ) ∈ Z+ × Z+ is uniquely mapped to the probability distribution pt1 ,t2 , if Mk,ℓ is
uniformly bounded for all integers k and ℓ. Consequently, one can infer that any two tuples of the random variables
(t1 , t2 ) and (t̂1 , t̂2 ) with the same sequences of the joint moments are identical in distribution.
To determine the joint moment of input and output entries, consider the distortion function
d(x̂, x) = x̂k xℓ
in Proposition IV.1, and evaluate the asymptotic distortion over the limit of
(62)
W(n) = [j : j + ηn] for some η in a
right neighborhood of zero. The (k, ℓ)-joint moment of (x̂j , xj ) is then determined by taking the limit η ↓ 0.
Substituting the distortion function and the index set in Proposition IV.1, it is clear that the asymptotic distortion
is independent of η and j, and therefore, the limit with respect to η exists and is independent of j as well. Noting
that the evaluated moments are uniformly bounded, it is inferred that the asymptotic joint distribution of (x̂j , xj ) is
uniquely specified and does not depend on the index j. Finally, by using the fact that the source vector is i.i.d. and
the distribution of the entry j is independent of the index, we conclude that the asymptotic conditional distribution
j
pjx̂|x is independent of j. The exact expression for px̂|x
is then found by determining the solution Q to the fixed
point equation and determining the joint moments.
N
21
Proposition IV.2 is a generalized form of the RS decoupling principle for the MAP estimators studied in [38].
In fact, Proposition IV.2 indicates that a vector system followed by a MAP estimator always decouples into a bank
of identical single-user systems regardless of any restriction on the replica correlation matrix.
C. Consistency Test
If one restricts the replica correlation matrix Q to be of a special form, the asymptotics determined under the
assumed structure do not necessarily approximate true asymptotics accurately. Several methods were introduced in
the literature to check the consistency of the solution. A primary method is based on calculating the entropy of the
corresponding spin glass at zero temperature. As the temperature tends to zero, the distribution of the microstate
tends to an indicator function at the point of the estimated vector, and consequently, the entropy of the corresponding
spin glass converges to zero1 . One consistency check is therefore the zero temperature entropy of a given solution.
Several works invoked this consistency test and showed that for the settings in which the RS ansatz fails to give a
tight bound on the exact solution, the zero temperature entropy determined from the RS ansatz does not converge to
zero; see for example [43]. This observation illustrates the invalidity of the RS assumption and hints at RSB ansätz
giving better bounds on the true solution. Inspired by the aforementioned results, we evaluate the zero temperature
entropy of the corresponding spin glass as a measure of consistency.
In order to determine the zero temperature entropy, we invoke (34) which determines the entropy in terms of
the free energy at inverse temperature β. Considering the free energy of the corresponding spin glass as given in
Proposition IV.1, the entropy H(β) reads
Z 1
β2 ∂
Tr{TQRJ (−2βωTQ)}dω − Tr{QT TRJ (−2βTQ)} + HR (β, m)
H(β) = lim
m↓0 m ∂β
0
(63)
where HR (β, m) denotes the normalized entropy of the spin glass of replicas. As HR (β, m) determines the entropy
of a thermodynamic system, for any m ∈ R+ we have
lim HR (β, m) = 0,
β↑∞
and therefore, the zero temperature entropy is given by
Z 1
β2 ∂
0
T
H = lim lim
Tr{TQRJ (−2βωTQ)}dω − Tr{Q TRJ (−2βTQ)} ,
β↑∞ m↓0 m ∂β
0
(64)
(65)
which obviously depends on the structure of the replica correlation matrix. In [43], the authors determined the zero
temperature entropy for the spin glass which corresponds to the vector precoding problem considering the RS and
1RSB assumptions, and observed that it takes the same form under both assumptions. They then conjectured that
the zero temperature entropy is of a similar form for the general RSB structure regardless of the number of breaking
steps2 . Using (65), we later show that the conjecture in [43] is true.
1 Note that the entropy of the spin glass at temperature β−1 denotes either the conditional entropy (for discrete supports) or the conditional
differential entropy (for continuous supports) of a random vector v distributed conditioned to y and A with Boltzmann-Gibbs distribution pβ
.
v|y,A
2 In fact in this case, the dependence of the zero temperature entropy on the correlation matrix is completely described via the scalar χ which
corresponds to the diagonal entries of the correlation matrix. See Assumption 3-5 for more illustrations.
22
V. RS A NSATZ AND RS D ECOUPLING P RINCIPLE
The most elementary structure which can be imposed on the replica correlation matrix is RS. Here, one assumes
the correlation matrix to be of a symmetric form which means that the replicas of the spin glass defined in
Definition IV.1 are invariant under any permutation of indices. Using the definition of the replica correlation matrix
as given in (49), it consequently reads that
(x − va )(x − vb ) =
q0
q1
if a 6= b
(66)
if a = b.
Assumption 3 (RS Structure). Considering the spin glass of replicas as defined in Definition IV.1, the replica
correlation matrix is of the form
Q=
χ
Im + q1m
β
where χ and q are some non-negative real scalars.
(67)
Considering (66), Assumption 3 supposes q0 = q and q1 = χβ−1 +q. Substituting (67) in Definition IV.1, the spin
glass of replicas is then specified by the scalars χ and q. The scalars are moreover related via a set of saddle point
equations which are obtained from (49). Finally, using Proposition IV.1, the asymptotics of the system are found.
Proposition V.1 (RS Ansatz). Let the linear system (1) fulfill the constraints of Section II. Suppose Assumptions 1
and 2, as well as Assumption 3 hold. Let x ∼ px , and
p
1
2
s
(x + λ0 z − v) + u(v)
g = arg min
v
2λs
(68)
with v ∈ X, and λs0 and λs being defined as
h
χ o
χ i−2 ∂ n
[λ0 χ − λq] RJ (− )
λs0 := RJ (− )
λ
∂χ
λ
h
i
−1
χ
λs := RJ (− )
λ
λ
(69a)
(69b)
for some non-negative real χ and q and the real variable z. Then, the asymptotic distortion defined in (27) reads
Z
(70)
DW = E d(g, x)Dz,
for χ and q which satisfy
Z
p
λs0 χ = λs E (g − x)z Dz
Z
q = E (g − x)2 Dz
and minimize the zero temperature free energy F0rs which is given by
Z 1
Z
i
p
1 h
1
0
s z − g)2 − λs z 2 + u(g) Dz
λ
(x
+
F(ω)dω − F(1) + E
Frs =
0
0
2λ 0
2λs
(71a)
(71b)
(72)
23
z p
x
λs0
×
y
+
gmap [(·); λs , u]
x̂
Fig. 1: The decoupled single-user system under the RS ansatz.
with F(·) being defined as
χ i
λ0
d h
ωRJ (− ω) .
F(ω) = q − χ
λ
dω
λ
(73)
Proof. See Appendix B.
N
The asymptotic distortion under the RS ansatz is equivalent to the average distortion of a scalar AWGN channel
followed by a single-user estimator as shown in Fig. 1. In this block diagram, the single-user estimator gmap [(·); λs , u]
maximizes the posterior probability over a postulated scalar AWGN channel. We refer to this estimator as the
“decoupled MAP estimator” and define it as follows.
Definition V.1 (Decoupled MAP estimator). The decoupled MAP estimator gmap [(·); λs , u] with the estimation
parameter λs and the utility function u(·) is defined as
gmap [(y); λs , u] := arg max qλs (v|y)
v
1
2
(y
−
v)
+
u(v)
,
= arg min
v
2λs
(74a)
(74b)
where qλs (v|y) denotes the “decoupled posterior distribution” postulated by the estimator which reads
qλs (v|y) = K e
h
i
1
− 2λs (y−v)2 +u(v)
for some real constant K.
(75)
Using the definition of the decoupled MAP estimator, the RS decoupled system is defined next.
Definition V.2 (RS decoupled system). Define the RS decoupled system to be consistent with the block diagram
in Fig. 1 in which
X,
•
the source symbol x is distributed with px over the support
•
z is a zero-mean and unit-variance Gaussian random variable,
•
x and z are independent,
•
x̂ is estimated from the observation y := x +
p s
λ0 z as
x̂ = gmap [(y); λs , u].
(76)
24
•
gmap [(·); λs , u] is the decoupled MAP estimator with the estimation parameter λs and the utility function u(·)
as defined in Definition V.1.
•
λs0 and λs are defined as in Proposition V.1.
Using Proposition IV.2, the equivalency in the asymptotic distortion can be extended to the asymptotic conditional
distribution as well. In fact, by considering the decoupling principle, Definition V.2 describes the structure of the
decoupled single-user system under the RS assumption.
Proposition V.2 (RS Decoupling Principle). Let the linear system (1) fulfill the constraints of Section II and be
estimated via the MAP estimator in (2). Consider the RS decoupled system as defined in Definition V.2, and suppose
Assumptions 1, 2 and 3 hold. Then, for any j ∈ [1 : n], the tuple (x̂j , xj ) converges in distribution to (x̂, x) if
px = px , almost sure in A.
Proof. Using Proposition IV.2, for any two different indices j, q ∈ [1 : n] we have
pjx̂|x (v̂|v) = pqx̂|x (v̂|v)
(77)
at the mass point (v̂, v). Therefore for any index j, we have
n
1 X k ℓ
x̂i xi .
E
n↑∞ n
i=1
E x̂kj xℓj = lim
(78)
Consequently, the asymptotic (k, ℓ)-joint moment of (x̂j , xj ) under the RS assumption is determined by letting
W(n) = [1 : n] and the distortion function in the RS ansatz
d(x̂; x) = x̂k xℓ
(79)
and determining the asymptotic distortion. Substituting in Proposition V.1, the asymptotic joint moment reads
Z
Mjk,ℓ = E gk xℓ Dz
(80)
where g is defined in (68). Considering Definition V.2 and assuming px = px , (80) describes the (k, ℓ)-joint moment
of (x̂, x) as well. Noting that Mjk,ℓ is uniformly bounded for any pair of integers k and ℓ, it is concluded that the
asymptotic joint distribution of (x̂j , xj ) and the joint distribution of (x̂, x) are equivalent.
N
Proposition V.2 gives a more general form of the RS decoupling principles investigated in [38] and [39]. In fact,
by restricting the system matrix and source distribution as in [38] and [39], one can recover the formerly studied
RS decoupling principles.
RS Zero Temperature
To have a basic measure of RS ansatz’s consistency, we evaluate the zero temperature entropy under the RS
assumption following the discussion in Section IV-C. Substituting (67) in (65) and taking the same steps as in
25
Appendix B, the RS zero temperature entropy is determined as
Z 1
β2 ∂
Fβ (ω)dω − Fβ (1)
H0rs = lim
β↑∞ 2λ ∂β
0
(81)
where the function Fβ (·) is defined as
Fβ (ω) =
χ
χ i
χ
λ0
d h
ωRJ (− ω) .
RJ (− ω) + q − χ
β
λ
λ
dω
λ
(82)
Taking the derivative first and then the limit, it finally reads
Z 1
χ
χ
χ
H0rs =
RJ (− ) −
RJ (− ω)dω .
2λ
λ
λ
0
(83)
We later on see that the zero temperature entropy takes the same form under the RSB assumptions.
VI. RSB A NS ÄTZE AND RSB D ECOUPLING P RINCIPLE
In [83], Parisi introduced a breaking scheme to broaden the restricted set of correlation matrices. The scheme
recursively extends the set of matrices to larger sets. The breaking scheme was then employed to broaden the RS
structure of the correlation matrices, and therefore, the obtained structure was identified as the broken RS or, RSB
structure. The key feature of Parisi’s breaking scheme is that, by starting from the RS structure, the new structure
after breaking can be reduced to the structure before breaking. Thus, the set of fixed point solutions found by
assuming the broken structure includes the solutions of the previous structure as well.
Definition VI.1 (Parisi’s breaking scheme). Let m be a multiple of the integer ξ, and Q[ℓ] represent an
correlation matrix. Parisi’s breaking scheme then constructs a new m × m correlation matrix Q[ℓ+1] as
Q[ℓ+1] = Iξ ⊗ Q[ℓ] + κ1m
for some real scalar κ.
m
ξ
×
m
ξ
(84)
By choosing Q[0] to be an RS correlation matrix in Definition VI.1, the matrix Q[1] finds the RSB structure with
one step of breaking (1RSB). The steps of breaking can be further increased recursively by inserting Q[1] in the
next breaking scheme, determining the new correlation matrix Q[2] , and repeating the procedure. We start by the
1RSB correlation matrix, and then, extend the result to higher RSB ansätze with more steps of breaking.
Assumption 4 (1RSB Structure). Considering the spin glass of replicas as defined in Definition IV.1, the replica
correlation matrix is of the form
Q=
χ
Im + pI mβ ⊗ 1 βµ + q1m
µ
β
where χ, p, q and µ are some non-negative real scalars.
(85)
Regarding Parisi’s breaking scheme, Assumption 4 considers Q = Q[1] by letting Q[0] have the RS structure
with parameters χβ−1 and p, ξ = µβ−1 and κ = q. Here, the 1RSB structure reduces to RS by setting p = 0.
26
Therefore, the set of 1RSB correlation matrices contains the one considered in Assumption 3.
Proposition VI.1 (1RSB Ansatz). Let the linear system (1) fulfill the constraints of Section II. Suppose Assumptions
1 and 2, as well as Assumption 4 hold. Let x ∼ px and
p
p
1
2
s
s
(x + λ0 z0 + λ1 z1 − v) + u(v)
g = arg min
v
2λs
(86)
with v ∈ X and λs0 , λs1 and λs being defined by
χ + µp
χ i−2 ∂
[λ0 (χ + µp) − λq] RJ (−
) ,
= RJ (− )
λ
∂χ
λ
h
χ i−2
χ
χ + µp
λs1 = RJ (− )
RJ (− ) − RJ (−
) λµ−1 ,
λ
λ
λ
h
χ i−1
λ
λs = RJ (− )
λ
h
λs0
(87a)
(87b)
(87c)
for some non-negative real χ, p, q and µ and real variables z0 and z1 . Then, the asymptotic distortion in (27) reads
Z
W
(88)
D = E d(g, x) Λ̃(z1 |z0 , x)Dz1 Dz0 ,
with Λ̃(z1 |z0 , x) :=
R
Λ(z1 , z0 , x)Dz1
Λ(z1 , z0 , x) = e
−1
Λ(z1 , z0 , x) and Λ(z1 , z0 , x) being defined by
i
h
√
√
√
1 √
1
−µ 2λs (x+ λs0 z0 + λs1 z1 −g)2 − 2λs ( λs0 z0 + λs1 z1 )2 +u(g)
.
(89)
The scalars χ, p and q satisfy
Z
p
s
s
λ0 [χ + µp] = λ E (g − x)z0 Λ̃(z1 |z0 , x)Dz1 Dz0
Z
p
s
s
λ1 [χ + µq + µp] = λ E (g − x)z1 Λ̃(z1 |z0 , x)Dz1 Dz0
Z
q + p = E (g − x)2 Λ̃(z1 |z0 , x)Dz1 Dz0
and µ is a solution of the fixed point equation
s
Z
Z µp
µ
λ1
1
χ+ω
µ
q
+
p
−
)dω
=
E
log Λ̃(z1 |z0 , x) Λ̃(z1 |z0 , x)Dz1 Dz0 .
R
(−
J
2λs
λs
2λ 0
λ
(90a)
(90b)
(90c)
(91)
0[1]
Moreover, χ, p, q and µ minimize the zero temperature free energy Frsb which reads
Z 1
Z
Z
1
1
0[1]
F(ω)dω − F(1) − E log
Λ(z1 , z0 , x)Dz1 Dz0
Frsb =
2λ 0
µ
(92)
with F(·) being defined as
1 d
F(ω) =
µ dω
"Z
[χ+µp]ω
χω
#
χ + µp
χ + µp d
t
ωRJ (−
ω) .
RJ (− )dt + q − λ0
λ
λ
dω
λ
(93)
Proof. See Appendix C.
N
27
pz1 |z0 ,x
z1 p
x
×
z0
λs1
+
×
+
p
λs0
y
gmap [(·); λs , u]
x̂
Fig. 2: The decoupled scalar system under the 1RSB ansatz.
Similar to our approach under the RS ansatz, we employ Proposition VI.1 to introduce the decoupled 1RSB
single-user system which describes the statistical behavior of the MAP estimator’s input-output entries under 1RSB
assumption. The decoupled system under 1RSB differs from RS within a new impairment term which is correlated
with the source and noise symbols through a joint distribution. The impairment term intuitively plays the role of
a correction factor which compensates the possible inaccuracy of the RS ansatz. The decoupled MAP estimator,
however, follows the same structure as for RS.
Definition VI.2 (1RSB decoupled system). Fig. 2 defines the 1RSB decoupled system in which
X,
•
the source symbol x is distributed with px over the support
•
z0 is a zero-mean and unit-variance Gaussian random variable,
•
z1 is a random variable correlated with x and z0 ,
•
x and z0 are independent, and
pz1 |x,z0 = Λ̃(z1 |z0 , x)φ(z1 )
(94)
with Λ̃ defined in Proposition VI.1.
•
x̂ is estimated from the observation y := x +
p
p s
λ0 z0 + λs1 z1 as
x̂ = gmap [(y); λs , u].
•
(95)
gmap [(·); λs , u] is the decoupled MAP estimator with the estimation parameter λs and the utility function u(·)
as defined in Definition V.1.
•
λs0 , λs1 and λs are defined as in Proposition VI.1.
Proposition VI.2 (1RSB Decoupling Principle). Let the linear system (1) fulfill the constraints of Section II and
be estimated via the MAP estimator in (2). Consider the 1RSB decoupled system as defined in Definition VI.2, and
suppose Assumptions 1, 2 and 4 hold. Then, for all j ∈ [1 : n], the tuple (x̂j , xj ) converges in distribution to (x̂, x)
if px = px .
Proof. The proof takes exactly same steps as for the RS decoupling principle using Proposition VI.1.
N
28
The 1RSB decoupled system, in general, provides a more accurate approximation of the estimator’s asymptotics
by searching over a larger set of solutions which include the RS ansatz. To investigate the latter statement, consider
the case of p = 0. In this case, the 1RSB structure reduces to RS. Setting p = 0 in Proposition VI.1, λ1 becomes
zero, and consequently Λ̃(z1 |z0 , x) = 1. Moreover, the fixed point equations in (91) hold for any choice of µ, and
the scalars χ and q couple through the same set of equations as in the RS ansatz. The zero temperature free energy
of the system, furthermore, reduces to its RS form under the assumption of p = 0. Denoting the parameters of the
RS ansatz by [χrs , qrs ], it is then concluded that [χ, p, q, µ] = [χrs , 0, qrs , 0] is a solution to the 1RSB fixed point
equations, when an stable solution to the RS fixed point exists. The solution, however, does not give necessarily
the 1RSB ansatz, the stable solution to the 1RSB fixed point equations with minimum free energy may occur at
some other point. We investigate the impact of replica breaking for some examples later through numerical results.
Parisi’s breaking scheme can be employed to extend the structure of the correlation matrix to the RSB structure
with more steps of breaking by recursively repeating the scheme. In fact, one can start from an RS structured
Q[0] and employ the breaking scheme for b steps to determine the correlation matrix Q[b] . In this case, the replica
correlation matrix is referred to as the bRSB correlation matrix.
Assumption 5 (bRSB Structure). Considering the spin glass of replicas as defined in Definition IV.1, the replica
correlation matrix is of the form
Q=
b
X
χ
Im +
pκ I mβ ⊗ 1 µβκ + q1m
µκ
β
κ=1
(96)
where scalars χ and q, and the sequences {pκ } and {µκ } with κ ∈ [1 : b] are non-negative reals, and {µκ } satisfies
the following constraint
µκ+1
∈ Z+
µκ
for κ ∈ [1 : b − 1].
(97)
Considering the correlation matrix in Proposition IV.1 to be of the form indicated in Assumption 5 the previous
ansätze are extended to a more general ansatz which can reduce to the 1RSB as well as RS ansatz. Proposition VI.3
expresses the replica ansatz under the bRSB assumption.
Proposition VI.3 (bRSB Ansatz). Let the linear system (1) fulfill the constraints of Section II. Suppose Assumptions
1 and 2, as well as Assumption 5 hold. For κ ∈ [0 : b], define the sequence {χ̃κ }, such that χ̃0 = χ and
χ̃κ := χ +
κ
X
µς p ς
(98)
ς=1
for κ ∈ [1 : b]. Let x ∼ px , and
#
b
X
p
1
g = arg min
λsκ zκ − v)2 + u(v)
(x +
v
2λs
κ=0
"
(99)
29
with v ∈ X and λs0 , {λsκ } and λs being defined as
−2
χ̃b
χ̃0
∂
[λ0 χ̃b − λq] RJ (− ) ,
= RJ (− )
λ
∂ χ̃b
λ
−2
χ̃0
χ̃κ−1
χ̃κ
λsκ = RJ (− )
RJ (−
) − RJ (− ) λµ−1
κ ,
λ
λ
λ
−1
χ̃0
s
λ
λ = RJ (− )
λ
λs0
(100a)
(100b)
(100c)
for some non-negative real scalar q, sequences {χ̃κ } and {µκ } and sequence of real variables {zκ }. Then, the
asymptotic distortion defined in (27) reads
DW = E
Z
d(g; x)
b
Y
κ=1
Λ̃κ (zκ |{zς }bκ+1 , z0 , x) Dzκ Dz0 ,
where {zς }bκ := {zκ , . . . , zb } and Λ̃κ (zκ |{zς }bκ+1 , z0 , x) =
Λ1 ({zς }b1 , z0 , x) := e
R
Λκ ({zς }bκ , z0 , x)Dzκ
−1
(101)
Λκ ({zς }bκ , z0 , x) with
"
#
b √
b √
P
1
1 P
−µ1 2λs (x+
λsκ zκ −g)2 − 2λs (
λsκ zκ )2 +u(g)
κ=0
κ=0
(102)
and {Λκ ({zς }bκ , z0 , x)} for κ ∈ [2 : b] being recursively determined by
Λκ ({zς }bκ , z0 , x)
:=
Z
Λκ−1 ({zς }bκ−1 , z0 , x)
Dzκ−1
µκ
µκ−1
.
(103)
The scalar q and sequences {χ̃κ } and {pκ } satisfy
b
X
pκ + q = E
κ=1
χ̃κ−1 + µκ
b
X
ς=κ
pς + q
!
Z
(g − x)2
λs
= p E
λsκ
λs
χ̃b = p s E
λ0
b
Y
κ=1
Λ̃κ (zκ |{zς }bκ+1 , z0 , x) Dzκ Dz0 ,
Z
(g − x)zκ
Z
(g − x)z0
b
Y
κ=1
b
Y
κ=1
(104a)
Λ̃κ (zκ |{zς }bκ+1 , z0 , x) Dzκ Dz0 ,
(104b)
Λ̃κ (zκ |{zς }bκ+1 , z0 , x) Dzκ Dz0 ,
(104c)
and the sequence {µκ } is given by
Z 1
Z
Z
1
1
1
{µκ } = arg min
F(ω; {µ̃κ })dω − E log
Λb (zb , z0 , x)Dzb Dz0 − s ∆({µ̃κ })
µ̃b
2λ
{µ̃κ } 2λ 0
(105)
where the function F(·; {µκ }) is defined as
Z χ̃κ ω
b
X
1 d
χ̃b
d
λ0
t
F(ω; {µκ }) :=
ωRJ (− ω) ,
RJ (− )dt + q − χ̃b
µ dω χ̃κ−1 ω
λ
λ
dω
λ
κ=1 κ
(106)
∆(·) is defined as
∆({µκ }) =
b
X
1
λs
[ζκ χ̃κ − ζκ−1 χ̃κ−1 ] + ζb q − 0s χ̃b
µ
λ
κ=1 κ
(107)
30
pz1 |z2 ,...,zb ,z0 ,x
z1 p
×
x
λs1
+
pzb |z0 ,x
z0
zb p
×
λsb
×
+
p s
λ0
+
y
gmap [(·); λs , u]
x̂
Fig. 3: The decoupled scalar system under the bRSB ansatz.
with ζ0 = 1, and ζκ for κ ∈ [1 : b] denoting
ζκ := 1 −
and {µ̃κ } ∈ Sµ in which
Sµ :=
{µ1 , . . . , µb } ∋ µκ ∈ R
+
κ
X
ς=1
µς
λsς
,
λs
µκ+1
+
∧
∈ Z ∀κ ∈ [1 : b − 1] .
µκ
(108)
(109)
In the case of multiple solutions for χ, q, {pκ } and {µκ }, the ansatz is chosen such that the free energy at zero
temperature
0[b]
Frsb =
1
2λ
Z
0
1
Z
Z
1
F(ω; {µκ })dω − F(1; {µκ }) − E log
Λb (zb , z0 , x)Dzb Dz0
µb
(110)
is minimized.
Proof. See Appendix D.
N
One can simply observe that Proposition VI.3 reduces to Propositions VI.1 and V.1 by letting b = 1 and pκ = 0
for κ ∈ [1 : b], respectively. The ansatz, moreover, extends the corresponding decoupled single-user system of the
estimator considering the general decoupling principle investigated in Proposition IV.2. By taking the same steps as
in Proposition VI.2, the decoupled bRSB single-user system is found which represents the extended version of the
1RSB system with b additive impairment taps. In fact, considering the impairment terms to intuitively play the role
of correction factors, the bRSB ansatz takes more steps of correction into account. The decoupled MAP estimator,
moreover, remains unchanged.
Definition VI.3 (bRSB decoupled system). Define the bRSB decoupled system as a single-user system illustrated
in Fig. 3 in which
X,
•
the source symbol x is distributed with px over the support
•
z0 is a zero-mean and unit-variance Gaussian random variable,
•
zκ is a random variable correlated with x, z0 and {zκ+1 , . . . , zb }
•
x and z0 are independent, and
pzκ |zκ+1 ,...,zb ,z0 ,x = Λ̃κ (zκ |{zς }bκ+1 , z0 , x)φ(zκ )
(111)
31
with Λ̃ defined in Proposition VI.3.
•
x̂ is estimated from the observation y := x +
b p
P
λsκ zκ as
κ=0
x̂ = gmap [(y); λs , u].
•
(112)
gmap [(·); λs , u] is the decoupled MAP estimator with the estimation parameter λs and the utility function u(·)
as defined in Definition V.1, and
•
λs0 , {λsκ } and λs for κ ∈ [1 : b] are as in Proposition VI.3.
Proposition VI.4 (bRSB Decoupling Principle). Let the linear system (1) fulfill the constraints of Section II and
be estimated via the MAP estimator in (2). Consider the bRSB decoupled system as defined in Definition VI.3, and
suppose Assumptions 1, 2 and 5 hold. Then, for all j ∈ [1 : n], the tuple (x̂j , xj ) converges in distribution to (x̂, x)
if px = px .
Proof. Using Proposition VI.1, it takes same steps as for Proposition V.2.
N
RSB Zero Temperature
In Appendix D, it is shown that under the bRSB assumption on the replica correlation matrix the free energy of
the corresponding spin glass at the inverse temperature β reads
Z 1
1
Fβ (ω)dω − Fβ (1) + FR (β).
F(β) =
2λ 0
(113)
Here, FR (β) denotes the normalized free energy of the spin glass of replicas defined in (52) in the limit m ↓ 0,
and the function Fβ (·) is defined as
Z χ̃κ ω
b
X
d
χ
χ
λ0
t
χ̃b
1 d
RJ (− )dt + RJ (− ω) + q − χ̃b
ωRJ (− ω) .
F (ω) =
µ dω χ̃κ−1 ω
λ
β
λ
λ
dω
λ
κ=1 κ
β
Following the discussion in Section IV-C, the entropy at the zero temperature reads
Z 1
β2 ∂
0[b]
β
β
F (ω)dω − F (1)
Hrsb = lim
β↑∞ 2λ ∂β
0
(114)
(115)
which reduces to
0[b]
Hrsb
Z 1
χ
χ
χ
=
RJ (− ) −
RJ (− ω)dω .
2λ
λ
λ
0
(116)
(116) justifies the conjecture in [43] and states that the zero temperature entropy under any number of breaking
steps, including the RS case, is of the similar form and only depends on the scalar χ. In fact, the Hamiltonian in
(10) reduces to the one considered in vector precoding by considering x to be the deterministic vector of zeros,
λ0 = 0, λ = 1 and u(v) = 0. Substituting in (116), the zero temperature entropy reduces to the one determined in
32
[43] within a factor of 2. The factor comes from the difference in the prior assumption on the support of microstate1 .
VII. R EPLICA S IMULATOR : C HARACTERIZATION
VIA THE
S INGLE -U SER R EPRESENTATION
The general bRSB decoupling principle determines an equivalent single-user system which describes the inputoutput statistics of the MAP estimator under the bRSB ansatz. In order to specify the exact parameters of the
decoupled single-user system, the set of fixed point equations needs to be solved explicitly. In this section, we
propose an alternative approach which describes an ansatz in terms of the corresponding decoupled system’s inputoutput statistics. We define the exact form of the decoupled system as the “steady state” of a transition system
named “replica simulator”. The proposed approach enables us to investigate the properties of the RS and RSB
ansätze by studying the replica simulator. In order to clarify the idea of the replica simulator, let us define a set of
input-output statistics regarding the bRSB decoupled system.
Definition VII.1. Consider the single-user system consistent with the block diagram in Fig. 3. Denote the joint
distribution of the source and impairment terms with px,z0 ,...,zb . For this system,
•
the κth noise-error correlation is defined as
for κ ∈ [0 : b], and
•
1
Cκ = p E (x̂ − x)zκ
λsκ
(117)
the MSE is denoted by
MSE = E (x̂ − x)2 .
(118)
Invoking Definition VII.1, the bRSB ansatz can be completely represented in terms of the input-output statistics
of the decoupled system. In fact, by means of Definition VII.1, the fixed point equations in (104a)-(104c) can be
expressed as
b
X
pκ + q = MSE,
(119a)
κ=1
χ̃κ−1 + µκ
b
X
ς=κ
!
= λs Cκ ,
(119b)
χ̃b = λs C0 ,
(119c)
pς + q
for κ ∈ [1 : b]; moreover, the factor Λ1 is given as
Λ1 ({zς }b1 , z0 , x) = e
1 In
[43], the authors considered v to be a complex vector.
i
h
1
1
−µ1 2λs (y−x̂)2 − 2λs (y−x)2 +u(x̂)
,
(120)
33
which reduces to
Λ1 ({zς }b1 , z0 , x) = px (x)µ1
qλs (x̂|y)
qλs (x|y)
µ1
(121)
with qλs (·|y) indicating the decoupled posterior distribution defined in Definition V.1. The second term on the right
hand side of (121) is an extended form of the likelihood ratio. By defining
µ
qλs (x̂|y) 1
b
Γ1 ({zς }1 , z0 , x) =
,
qλs (x|y)
(122)
(121) reads
Λ1 ({zς }b1 , z0 , x) = px (x)µ1 Γ1 ({zς }b1 , z0 , x),
(123)
and Λκ ({zς }bκ , z0 , x) for κ ∈ [2 : b] are determined by
Λκ ({zς }bκ , z0 , x) = px (x)µκ Γκ ({zς }bκ , z0 , x)
(124)
where Γκ ({zς }bκ , z0 , x) are recursively defined as
Γκ ({zς }bκ , z0 , x)
=
Z
Γκ−1 ({zς }bκ−1 , z0 , x)
Dzκ−1
µκ
µκ−1
.
(125)
The fixed point in (105) is therefore rewritten accordingly.
The above alternative representation of the bRSB ansatz leads us to a new interpretation. In fact, one can define a
transition system in which the vector of replica parameters denotes the state, and the decoupled single-user system
defines the transition rule [84], [85]. We refer to this transition system as the “replica simulator”, and define it
formally as the following.
Definition VII.2 (Replica simulator). Let b be a non-negative integer. Consider the vector s as
s := [χ, µ1 , . . . , µb , p1 , . . . , pb , q]
(126)
with entries satisfying the corresponding constraints in Proposition VI.3, and denote its support by Sb . The transition
rule TRb : Sb 7→ Sb maps the prior state si ∈ Sb to the posterior state sf ∈ Sb in the following way:
TRb realizes the bRSB decoupled system considering the entries of si as the replica parameters. It then constructs
the entries of sf by determining a new set of replica parameters from the statistics of the decoupled system using
the equivalent representation of the fixed point equations given in (119a)-(125).
The replica simulator of breaking order b is then defined as the transition system SimR [b] :=
Sb, TRb
.
The structure of the replica simulator is illustrated in Fig. 4. For the replica simulator of breaking order b, a
sequence of states {st } is considered to be a “process”, if for t ∈ [1 : ∞]
TR
b
st+1 .
st −→
(127)
The state s⋆ is then called the “steady state”, if setting st = s⋆ results in st+1 = s⋆ . Regarding Proposition VI.3,
34
si
TRb
bRSB decoupling
px̂|x
{Cκ }, MSE, · · ·
sf
Fig. 4: Replica Simulator of breaking order b
the bRSB ansatz is in fact the steady state of the replica simulator which minimizes the free energy function. Our
conclusion also extends to the RS case, if we set b = 0.
Considering Definition VII.2, as well as the above discussions, the bRSB ansatz can be numerically investigated
using the methods developed in the literature of transition systems. This approach may reduce the complexity of
numerical analysis; however, it does not impact the computational complexity1. In fact, assuming that one realizes
the bRSB decoupled system for any desired state vector denoted in (126) via some methods of realization, e.g.,
Monte Carlo simulation, the bRSB ansatz can be found by means of an iterative algorithm which has been designed
to find the steady state of a transition system. The latter statement can be clarified as in Scheme 1.
Scheme 1 Analysis via Replica Simulator
begin
set replica simulator of breaking order b, SimR [b]
if b = 0 then
TRb corresponds to the RS decoupled system
else
TRb corresponds to the bRSB decoupled system
end if
initiate initial state s0 ∈ Sb
TR
b
evaluate s0 −→
s
0
if s = s then
s⋆ ← s
break
else
consider mapping rule IM(·, ·) : Sb × Sb 7→ Sb
s0 ← IM(s, s0 )
return A
end if
output s⋆
end
⊲A
⊲B
In Scheme 1, TRb in step A can be realized via different methods. One may determine the input-output distribution of the single-user system analytically or simulate the system by generating impairment and source samples
numerically via Monte Carlo technique. Another degree of freedom is in step B where different mapping rules with
1 We conjecture that in some cases, our bRSB decoupled system represents the asymptotic of a decision-feedback system. The validity of this
conjecture can further reduce the computation complexity.
35
different convergence speeds can be employed. For algorithms designed based on Scheme 1, the computational
complexity depends on the realization method while the convergence speed is mainly restricted with some given
mapping rule IM(·, ·).
The replica simulator introduces a systematic approach for investigating the replica anätze based on the decoupling
principle. Moreover, it gives an intuition about the impact of symmetry breaking. To clarify the latter statement, let
us consider an example.
Example. (RS vs. 1RSB ansatz) Let b = 0; thus, the RS fixed point equations read
q = MSE,
(128a)
χ = λs C0 .
(128b)
The equations under the 1RSB assumption are moreover given by
q + p = MSE,
(129a)
χ + µp = λs C0 ,
(129b)
χ + µp + µq = λs C1 ,
(129c)
and µ is determined through the fixed point equation
s
Z µp
λ1
1
χ+ω
µ
µ
q
+
p
−
)dω = I(z1 ; z0 , x) + DKL (pz1 kφ)
RJ (−
s
s
2λ
λ
2λ 0
λ
(130)
where I(·; ·) and DKL (· k ·) denote the mutual information and Kullback-Leibler divergence, respectively. Assuming
the system matrix to be i.i.d. and setting z1 to be independent of z0 and x, the right hand side of (130) tends to
zero, and therefore, the solutions µ = 0 and p = 0 are concluded. Consequently, (129c) becomes ineffective, and
the fixed point equations reduce to RS. The latter observation can be interpreted in terms of the “state evolution” of
the replica simulator. More precisely, assume that the initial state of the replica simulator with breaking order one
is chosen such that in the corresponding decoupled system, z1 is sufficiently correlated with the source and noise
symbols. In this case, by assuming the mapping rule IM(·, ·) to be converging, the correlation in each iteration of
Scheme 1 reduces, and thus, at the steady state, z1 becomes independent of z0 and x.
The above discussion can be extended to replica simulators with larger breaking orders. Moreover, further
properties of the RSB ansätze could be studied using methods developed in the literature of transition systems1 .
We leave the further investigations as a possible future work.
1 The concept of replica simulator may further clarify the connection between RSB ansätze and message passing based algorithms. Such
investigations however are skipped here and can be considered as a possible future work.
36
VIII. L ARGE C OMPRESSIVE S ENSING S YSTEMS
Considering the setting represented in Section II, a large compressive sensing system can be studied through our
results by restricting the source’s CDF to be of the form
Fx (x) = (1 − α)1 {x ≥ 0} + αF̆x (x).
(131)
In the large limit, the source vector distributed as (131) has (1 − α)n entries equal to zero while the remaining αn
entries are distributed with F̆x . In this case, x is an αn-sparse vector, and thus, (1) is considered to represent a
large compressive sensing system with the sensing matrix A.
Considering the prior as in (131), different recovery schemes are then investigated by restricting the prior setup
of the system, correspondingly. In this section, we study the asymptotics of several recovery schemes using our
bRSB decoupling principle for both the cases of continuous and finite alphabet sources.
A. Continuous Sources
Assuming
X = R, (131) describes a continuous random variable multiplied by an α-Bernoulli random variable.
In this case, by varying the utility function u(·), different reconstruction schemes are considered. Here, we address
the linear, LASSO and ℓ0 norm recovery schemes. The results can however be employed to investigate a general
ℓp norm recovery scheme [61].
Example 1. (linear recovery scheme) The MAP estimation is reduced to the linear recovery scheme when the
utility function is set to be
u(v) =
v2
.
2
(132)
In fact, in this case, the MAP estimator postulates the prior distribution to be a zero-mean and unit-variance Gaussian
and performs considerably inefficient when the source is sparse. Using the bRSB decoupling principle, we conclude
that in the large-system limit the source entry xj and the estimated entry x̂j , for any j ∈ [1 : n], converge in
probability to a sparse random variable x distributed as in (131) and the estimated symbol x̂ := gmap [(y); λs , u]
where the decoupled system reduces to
gmap [(y); λs , u] =
y
,
1 + λs
(133)
with y being given by
y =x+
b
X
p
λsκ zκ ,
(134)
κ=0
and the scalars λs and {λsκ } for κ ∈ [0 : b] are defined as in Proposition VI.3. By letting b = 0, the result reduces
to the RS decoupling principle reported in the literature, see [38], [39]; however, the result here holds for a wider
set of sensing matrices and source distributions.
37
Example 2. (LASSO recovery scheme) To study the LASSO recovery scheme, we set
u(v) = |v|.
(135)
Regarding the bRSB decoupling principle, the prior distribution of the decoupled system’s input x is postulated to
be “Laplacian” or “double exponential” with unit variance. This postulation results in a better performance of the
recovery scheme in many cases, since the Laplacian PDF could be a more realistic approximation of the sparse
source distribution. Consequently, the decoupled system’s output is found as x̂ = gmap [(y); λs , u] with
gmap [(y); λs , u] = [y − λs sgn(y)] wλs (y)
(136)
where y is denoted as in (134), wλs (·) is the null window function with window width λs defined as
1 |y| > λs
,
wλs (y) =
0 |y| ≤ λs
(137)
and sgn(y) is the sign indicator. The decoupled single-user estimator in (137) which is often referred to as the
soft-thresholding operator recovers the earlier RS results by setting b = 0 and the sensing matrix to be i.i.d. [38].
Example 3. (ℓ0 norm recovery scheme) The ℓ0 norm recovery scheme which considers
u(v) = 1 {v 6= 0}
(138)
can perform significantly better that the latter schemes when the sparsity increases. In this case, the prior distribution
of the bRSB decoupled system’s input x is found as the limit of
e
1
pθ,δ
x (x) =
2θ + 2δ (e − 1) 1
|x| ≤ δ
δ < |x| ≤ θ
,
(139)
when θ ↑ ∞ and δ ↓ 0. For finite values of θ and δ, pθ,δ
can be considered as a sparse distribution in which
x
non-zero symbols occur uniformly. The prior explains the better performance of the ℓ0 norm recovery scheme
compared to the linear and LASSO schemes. For this case, the output of the decoupled single-user system reads
x̂ = gmap [(y); λs , u], such that
gmap [(y); λs , u] = ywϑ (y)
where y is denoted as in (134), and wϑ (·) is the null window function with ϑ =
(140)
√
2λs . Here, gmap [(·); λs , u] is the
hard-thresholding operator and recovers the analysis in [38] for a wider class of settings.
The above examples have been also studied in earlier replica based studies, e.g., [38], [40]. The given results
can be analytically derived from the above expressions by considering the RS ansatz and properly substituting the
corresponding R-transforms. We address the results of two important cases reported in the literature in following.
Special Case 1: In [38], the authors addressed the case of which an i.i.d. sparse source is sampled by an i.i.d.
sensing matrix where the matrix entries are zero-mean random variables with the variance vanishing proportional
38
to k −1 . The asymptotic performance of the estimator was then addressed when the linear, LASSO, and ℓ0 norm
recovery schemes are employed using the RS MAP decoupling principle. The results reported in [38] can be derived
directly by setting the R-transform in Proposition V.2 to be
RJ (ω) =
1
.
1 − rω
(141)
Special Case 2: The results of [38] extended in [40] to a larger set of sensing matrices, and the RS prediction of
the asymptotic MSE was determined for sparse Gaussian sources. The given results can be recovered by considering
the distortion function
d(x̂; x) = kx̂ − xk2 ,
(142)
and the source distribution to be (131) with F̆x representing the zero-mean and unit-variance Gaussian CDF.
B. Finite Alphabet Sources
Our result can be further employed to study the sampling problem of finite alphabet sources. Considering
X = {0, t1, . . . , tℓ−1 } ,
(143)
in which the symbol 0 occurs with probability 1−α, and other ℓ−1 outcomes are distributed due to pt . Consequently,
the source distribution reads
px (x) = (1 − α) 1{x = 0} + α1{x 6= 0}pt (x)
(144)
which can be interpreted as the multiplication of the non-zero discrete random variable t distributed with pt and
an α-Bernoulli random variable. For sake of brevity, we denote the sorted version of the symbols in the support
X
by c1 , . . . , cℓ in which
−∞ < c1 < c2 < . . . < cℓ < +∞.
(145)
For notational compactness, we further define c0 and cℓ+1 to be −∞ and +∞, respectively. Similar to the continuous
case, different choices of the utility function u(·) address different types of reconstruction schemes which we
investigate in the sequel.
Example 4. (linear recovery scheme) We consider the case in which the finite alphabet source is reconstructed
via the linear recovery scheme as introduced in Example 1. Using the bRSB decoupling principle, the source and
estimated symbols (xj , x̂j ) converge to the random variables x and x̂ for all j ∈ [1 : n] where x is distributed with
px defined in (144), and x̂ is found as x̂ = gmap [(y); λs , u] with
gmap [(y); λs , u] = ck
if
i
ℓ2
y ∈ vkℓ2 , vk+1
(146)
39
for k ∈ [1 : ℓ]. The scalar y indicates the observation symbol in the equivalent decoupled system, and the boundary
point vkℓ2 is defined as
vkℓ2 :=
1 + λs
(ck−1 + ck ) .
2
(147)
Example 5. (LASSO recovery scheme) Replacing the reconstruction scheme in Example 4 with LASSO, the
single-user estimator of the bRSB decoupled system is of the form
gmap [(y); λs , u] = ck
if
for k ∈ [1 : ℓ] where the boundary point vkℓl reads
vkℓ1 :=
i
ℓ1
y ∈ vkℓ1 , vk+1
1
|ck | − |ck−1 |
.
(ck−1 + ck ) + λs
2
ck − ck−1
(148)
(149)
Example 6. (ℓ0 norm recovery scheme) For finite alphabet sources, the ℓ0 norm recovery scheme is optimal
in terms of symbol error rate, since it realizes the sparse uniform distribution. In fact, for the case of which the
non-zero symbols of the source are uniformly distributed, the ℓ0 norm utility function exactly models the source’s
true prior, and therefore, can be considered as the optimal scheme. Under the ℓ0 norm recovery scheme, the bRSB
decoupled system reduces to
gmap [(y); λs , u] = ck
with the boundary point vkℓ0 being defined as
vkℓ0 :=
if
i
ℓ0
y ∈ vkℓ0 , vk+1
1{ck 6= 0} − 1{ck−1 6= 0}
1
.
(ck−1 + ck ) + λs
2
ck − ck−1
(150)
(151)
for k ∈ [1 : ℓ].
IX. N UMERICAL R ESULTS
FOR
L ARGE C OMPRESSIVE S ENSING S YSTEMS
In this section, we numerically investigate the examples of large compressive sensing systems for some known
setups. For this purpose, we simulate the decoupled systems by setting the source distribution and sensing matrix to
a specific form and determine the expected distortion of the equivalent scalar system. We, then, discuss the validity
of RS and RSB assumptions for these examples.
A. Simulation Setups
The settings and distortion functions being considered in the numerical investigations are as follows.
1) Sensing Matrices: Throughout the numerical investigations, we consider the two important cases of random
“i.i.d.” and “projector” matrices.
40
•
i.i.d. Random Matrix: In this case, the entries of Ak×n are supposed to be generated i.i.d. from an arbitrary
distribution pa . Without loss of generality, we assume that the entries are zero-mean random variables with
variance k −1 . This structure is the most primary and also the most discussed case in random matrix theory. For
this matrix, regardless of pa , it is well known that the asymptotic empirical eigenvalue CDF of the Gramian
J follows the Marcenko-Pastur law which states
FJ (λ) = 1 − r
−1 +
1{λ > 0} +
Z
λ
−∞
p
r − (1 − u)2
du
2πru
(152)
where [x]+ returns x when x is non-negative and is zero otherwise [32], [33], [86]. Using the definition of
R-transform, it is straightforward to show that RJ (·) reads
RJ (ω) =
•
1
.
1 − rω
(153)
Projector Matrix: Here, the only constraint on the sensing matrix is that the row vectors are orthogonal. The
matrices are also referred to as the “row orthogonal” matrices. For the sensing matrix Ak×n , we assume the
case that the row vectors are normalized by the number of rows and k ≤ n; thus, the outer product AAT reads
AAT =
n
Ik .
k
(154)
Consequently, the Gram matrix J takes two different eigenvalues: λ = 0 with multiplicity n−k, and λ = nk −1
with multiplicity k. Considering the definition of the factor r in (19), the asymptotic empirical CDF of the
eigenvalues read
FJ (λ) = 1 − r−1 1{λ > 0} + r−1 1{λ > r}
(155)
which results in the R-transform of the form
RJ (ω) =
ω−1+
p
(rω − 1)2 + 4ω
.
2ω
(156)
2) Source Model: We consider the continuous and finite alphabet sources to be distributed with “sparse Gaussian”
and “sparse uniform” distributions, respectively. More precisely, we assume the entries of the continuous and finite
alphabet sources to be generated from a Gaussian and uniform distribution, respectively, and multiplied by a
Bernoulli random variable with probability α to take 1 and 1 − α to take 0. Moreover, we assume the nonzero
outcomes of the finite alphabet source to be of the symmetric form
{±a, . . . , ±κa}
(157)
for some positive real a and integer κ.
3) Distortion Function: Regarding the continuous sources, we determine the performance of different estimators
by considering the MSE as the distortion function, i.e.,
d(x̂; x) = kx̂ − xk2 .
(158)
−18
−20
41
−2
−4
MSE0 in [dB]
−6
−8
−10
Linear-i.i.d.
Linear-Projector
LASSO-i.i.d.
LASSO-Projector
ℓ0 norm-i.i.d.
ℓ0 norm-Projector
−12
−14
−16
0.05
0.1
0.2
0.15
0.25
0.3
0.35
0.4
0.45
0.5
λ
Fig. 5: RS predicted normalized MSE versus the estimation parameter λ for the linear, LASSO and ℓ0 norm recovery schemes
considering the compression rate r = 2. The sparsity factor is set to be α = 0.1 and the noise variance λ0 is set such that
the source power to noise power ratio becomes 10 dB. The dashed and solid lines respectively indicate the cases with random
projector and i.i.d. measurements. The curves match the numerical results reported in [38], [40].
For the finite alphabet sources, moreover, we determine the probability of the error as the measure for which
d(x̂; x) =
n
X
i=1
1 {x̂i 6= xi }
(159)
as well as the MSE. Considering either case, it is clear that the MSE obtained by any ℓp norm recovery scheme is
bounded from below by the MMSE bound reported in the literature [21], [23].
B. Numerical Results for Continuous Sources
Considering Examples 1-3, we consider the case in which a sparse Gaussian source with sparsity factor α = 0.1
is sampled via a random sensing matrix. Fig. 5 shows the RS prediction of normalized MSE, defined as
MSE0 =
MSE
= α−1 MSE,
E |x|2
(160)
as a function of the estimation parameter λ. The compression rate is set to be r = 2, and both the i.i.d. random
and projector sensing matrices are considered. The curves match the results reported in [40] and [38]. As it is seen,
the ℓ0 -norm recovery scheme with the optimal choice of estimation parameter outperforms the LASSO scheme;
42
r = 11
0
−2
r=5
MSE0 in [dB]
−4
−6
−8
r=3
−10
r=1
−12
−14
i.i.d.
Projector
−16
0.05
0.1
0.15
0.2
0.25
0.3
0.35
0.4
0.45
0.5
λ
Fig. 6: RS predicted normalized MSE versus the estimation parameter λ for different compression rates considering LASSO
recovery. The sparsity factor is set to be α = 0.1 and the true noise variance λ0 is set to be 0.01. The dashed and solid lines
respectively indicate the random projector and i.i.d. matrices. As the compression rate increases, MSE0 converges to 0 dB.
however, the non-optimal choice of the estimation parameter can make the ℓ0 norm’s performance even worse than
the LASSO. Moreover, in contrast to the noiseless case, the projector matrix is always outperforming the i.i.d.
matrix in the noisy case; this fact has been also reported in [40].
In [42], the authors showed that in the noiseless sampling case with an i.i.d. Gaussian matrix, the RS ansatz
for linear and LASSO recovery schemes is locally stable against perturbations that break the symmetry of the
replica correlation matrix. This result in fact agrees with the general belief that convex optimization problems do
not exhibit RSB [87]. The result in [42], however, indicated that for the ℓ0 norm reconstruction, the RS ansatz
becomes unstable, and therefore the RSB ansätze are needed for accurately assessing the performance.
In order to investigate the observation of [42], we have plotted the normalized MSE of the LASSO recovery
scheme predicted by the RS ansatz in terms of the estimation parameter λ in Fig. 6 considering different compression
rates. It is observed that for a given estimation parameter λ, the normalized MSE increases as the compression
rate grows. For large compression rates, the normalized MSE converges to 0 dB which agrees with the fact that
for asymptotically large compression rates, the source and observation vectors are independent, and thus, the MSE
converges to the source power. To investigate the ℓ0 norm recovery scheme, we plot the corresponding curves for
the ℓ0 norm scheme considering Fig. 6 as the reference.
For ℓ0 norm recovery, Fig. 7 shows the normalized MSE predicted by the RS ansatz as a function of the estimation
0.75
43
0.85
0
r=4
0.95
MSE0 in [dB]
−5
r = 11
invalid ansätze
−10
i.i.d.
Projector
r=1
−15
0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
1
λ
Fig. 7: The normalized MSE as a function of the estimation parameter λ determined via the RS ansatz for different compression
rates considering the ℓ0 norm recovery scheme. The sparsity factor and the noise variance are considered to be α = 0.1
and λ0 = 0.01, respectively. The dashed and solid lines respectively indicate the random projector and i.i.d. matrices. As the
compression rate increases, the RS ansatz starts to give invalid solutions at low estimation parameters. In fact, as r grows, the
RS fixed point equations have no valid solutions as λ reduces. The interval in which the solution is invalid becomes larger as
the compression rate increases.
parameter for two different compression rates. The system setup has been set to be similar to the one considered in
Fig. 6, and the curves have been plotted for both the i.i.d. and projector measurements. In contrast to the LASSO
recovery scheme, the RS ansatz starts to give invalid predictions for the ℓ0 norm scheme as the compression
rate increases. As it is observed, the normalized MSE drops unexpectedly down for an interval of the estimation
parameters when the compression rate grows.
In fact, in this interval, the RS fixed point equations have either an unstable solution or no solution. To illustrate
this result further, let us consider Examples 2 and 3 under the RS ansatz when an i.i.d. sensing matrix is employed.
In this case, the equivalent noise power and estimation parameter λs0 and λs read
λs = λ + rχ
(161a)
λs0 = λ0 + rq.
(161b)
By increasing the compression rate, the interference increases, and thus, the MSE takes larger values. Therefore,
44
for small λ and λ0 , one can consider rχ ≫ λ and rq ≫ λ0 as r takes large values and write
λs ≈ rχ
(162a)
λs0 ≈ rq.
(162b)
Considering Example 2, by substituting (162a) and (162b) in the RS ansatz, the fixed point equations, as r grows
large, is written in the following form
√
√
u r φ(u r) ≈
Z
∞
√
u r
q ≈ 2α
Z
z 2 Dz + ǫr
(163a)
√
u r
Dz + ǫr .
(163b)
0
for some ǫr tending to zero as r ↑ ∞, and the bounded real scalar u defined as
χ
u := √ .
q
(164)
Taking the limit r ↑ ∞, (163a) is valid for any bounded real value of u and (163b) reduces to q ≈ α for large
compression rates. Noting that for this setup q = MSE, one concludes that MSE0 ≈ 1 which agrees with the results
given in Fig. 6. A similar approach for the ℓ0 norm recovery scheme in Example 3, however, results in the following
contradicting equations
Z
∞
z 2 Dz ≈ ǫr
u
Z
0
(165a)
u
Dz ≈ ǫr .
(165b)
for the scalar u defined as
r
χ
u := 2 .
q
(166)
Clearly, the set of equations in (165a) and (165b) have no solution as r ↑ ∞. The approximated fixed point equations
explain the invalidity of the RS predicted performance of the ℓ0 norm recovery scheme for large compression rates.
In order to further investigate the RS ansatz, we plot the optimal normalized MSE as a function of the compression
rate in Fig. 8. Here, we consider the case with i.i.d. sensing matrix when the sparsity factor is set to be α = 0.1
and source to noise power ratio to be 10 dB. The normalized MSE is minimized numerically over the estimation
parameter λ. As the figure illustrates, the MSE of the RS ansatz starts to drop in moderate compression rates.
The observation confirms the discussion on the stability of the RS saddle point given in [41], [42]. In fact, the
unexpected drop of the RS ansatz is caused by the limited stability region of the RS fixed point solutions. More
precisely, for a given source to noise power ratio and estimation parameter, the RS fixed point equations have stable
solutions within a certain interval of compression rates. The interval widens as λ grows.
Fig. 9 compares the RS and 1RSB ansätze for ℓ0 norm recovery. In this figure, the optimal normalized MSE
is plotted for the case with i.i.d. sensing matrix. The sparsity factor is considered to be α = 0.1 and the noise
45
0
−2
Linear
LASSO
ℓ0 norm
−4
4.5
5
5.5
6
MSE0 in [dB]
−6
−8
−10
−12
invalid ansätze
−14
−16
−18
1
1.5
2
2.5
3
3.5
4
compression rate r
Fig. 8: Optimal normalized MSE versus the compression rate r under RS assumption considering the linear, LASSO and ℓ0 -norm
recovery schemes. The sparsity factor and the source to noise power ratio are considered to be α = 0.1 and 10 dB, respectively.
MSE0 has been minimized numerically over the estimation parameter λ. The solid and dashed lines show the results for random
i.i.d. and orthogonal measurements, respectively. As r grows, the RS ansatz for ℓ0 norm recovery starts to give invalid solutions.
variance is set to be λ0 = 0.01. For the RS ansatz, we have considered two cases, namely when MSE0 is minimized
over 1) all possible estimation parameters and 2) the interval of λ’s in which the RS ansatz is valid within1 . The
latter case is referred to as the “RS restricted” curve. As the figure depicts, the difference between the RS and the
1RSB ansatz is quite small for low compression rates. The 1RSB prediction however deviates from RS at larger
compression rates. This observation indicates that the performance analysis of the ℓ0 norm recovery exhibits RSB.
For sake of comparison, we also plot the curve for the LASSO recovery scheme as well as the MMSE bound. As
it is observed, the normalized MSE for LASSO bounds the 1RSB curve from above. The 1RSB ansatz is moreover
consistent with the MMSE bound which is rigorously justified in the literature [21], [23]. The RS restricted ℓ0
norm’s curve, moreover, crosses the LASSO curve. This deviation indicates that, at large compression rates, the
optimal estimation parameter λ, which minimizes the MSE, lies somewhere out of the interval in which the RS
ansatz returns a valid approximation for the MSE.
To check the consistency of the solutions under the RS and 1RSB ansätze, we also sketch the zero temperature
entropy of the corresponding spin glass as a function of the compression rate in Fig. 10, considering the same setup
1 In
fact, we consider λ > λd , where λd is the point in Fig. 7 in which the normalized MSE curve is not differentiable.
46
0
LASSO-RS
ℓ0 norm-RS
−2
ℓ0 norm-RS restricted
ℓ0 norm-1RSB
MMSE Bound
−4
4.5
5
5.5
6
MSE0 in [dB]
−6
−8
−10
−12
−14
−16
1
1.5
2
2.5
3
3.5
4
compression rate r
Fig. 9: RS versus 1RSB prediction of the optimal normalized MSE considering the ℓ0 norm recovery scheme and random i.i.d.
measurements. The sparsity factor and the source to noise power ratio are considered to be α = 0.1 and 10 dB respectively. The
green dashed line shows the RS prediction of optimal MSE0 when it is numerically minimized over λ; however, the green solid
line indicates the restricted RS predicted MSE0 minimized numerically over the intervals of λ in which the RS ansatz is stable.
The black solid line denotes the 1RSB ansatz which deviates both the RS and restricted RS curves. For sake of comparison,
the normalized MSE for LASSO recovery as well as the MMSE bound have been plotted.
as in Fig. 8 and Fig. 9. The zero temperature entropy also confirms the better accuracy of the 1RSB ansatz.
Based on the above investigations, the exact performance of the ℓ0 norm recovery scheme needs the RSB ansätze
to be accurately studied. It is however useful to see the validity region of the system parameters in which the
prediction given by the RS ansatz lies approximately on the 1RSB curve. For this purpose, we define for the ℓ0
RS
norm recovery, the “RS break compression rate” rbr
(·) as a function of system parameters to be the compression
rate that the prediction via the RS ansatz starts to deviate from 1RSB. More precisely, for a given setting with
parameters in the set
P, the “RS ǫ-validity region” for a given ǫ ∈ R+ is defined as
VbrRS (P; ǫ) :=
r : |MSE0RS (r, P) − MSE01RSB (r, P)| < ǫ .
(167)
where MSE0RS (r, P) and MSE01RSB (r, P) indicate, respectively, the normalized MSE at the compression rate r for
the setup specified by the set P calculated from the RS and 1RSB ansatz, respectively. Consequently, the “RS break
47
−10 −3
zero temperature entropy H0
−10 −2
5
6
−10−1
−10 0
H0rs
0[1]
Hrsb
1
1.5
2
2.5
3
3.5
4
compression rate r
0[1]
Fig. 10: H0rs and Hrsb as functions of the compression rate considering the spin glass which corresponds to the ℓ0 norm recovery
scheme. The setting is considered to be the same as in the setup investigated in Fig. 8 and Fig. 9. The better approximation of
the performance via the 1RSB ansatz is demonstrated in this figure.
RS
compression rate” rbr
(·) for maximum deviation ǫ is given by
RS
rbr
(P; ǫ) := max r.
VbrRS (P;ǫ)
In general, the set
(168)
P includes several setup parameters such as sensing matrix and source’s distribution, the sparsity
factor, noise variance and estimation parameter. Considering a case with i.i.d. sensing matrix and sparse Gaussian
source with a given sparsity factor α, the RS break compression rate is a function of snr and the estimation parameter
RS
λ, i.e., rbr
(snr, λ; ǫ), where we define the source to noise power ratio snr as
snr :=
In this case, the RS ǫ-validity region
E x2
= αλ−1
0 .
λ0
(169)
VbrRS (snr, λ; ǫ) is the area enclosed by both the λ and snr axes as well as the
RS
rbr
(snr, λ; ǫ) surface.
Fig. 11 illustrates the intersection of the RS ǫ-validity region and the planes snr = 9 and snr = 12 for ǫ = 10−3 .
The break compression rate increases with respect to both λ and snr which agrees with the intuition. Moreover,
48
4
snr = 12 dB
3.5
snr = 9 dB
RS
rbr
(snr, λ; ǫ)
3
2.5
2
VbrRS(9, λ; ǫ)
1.5
1
0.01
0.02
0.03
0.04
0.06
0.05
0.07
0.08
0.09
0.10
λ
RS
Fig. 11: rbr
(snr, λ; ǫ) in terms of the estimation parameter λ for ǫ = 10−3 , and snr = 9 dB and snr = 12 dB considering an
i.i.d. matrix employed for sensing. The break compression rate starts to saturate as the estimation parameter grows. As λ ↓ 0,
the break compression rate takes values close to one. This observation agrees with RS instability reported in [42].
RS
rbr
(snr, λ; ǫ) starts to saturate as λ grows. Another extreme case is when the estimation parameter tends to zero in
which the MAP estimator reduces to
g(y) = arg
min
ky−Avk≤ǫ0
kvk0 .
(170)
with ǫ0 = O( √1snr ). For this case, the break compression rate converges to the minimum compression rate r = 1.
This observation in the large snr limit agrees with the instability of the RS ansatz reported in [42].
0.06
0.07
0.08
0.09
0.10
49
10
0
10 −2
Error Probability
10 −4
10 −8
κ=4
10
−8
Single-user bound
i.i.d. matrix
Projector matrix
10−10
κ=2
5
15
10
20
snr
Fig. 12: The error probability versus snr when the ℓ0 norm recovery scheme is employed and the compression rate is set to
be r = 1. By i.i.d. measurements, the error probability deviates from the single-user bound within an interval of snr while
random orthogonal measurements meets the single-user bound for almost every snr and alphabet size. Here, the sparsity factor
and estimation parameter are considered to be α = 0.1 and λ = 0.1, respectively.
C. Numerical Results for Finite Alphabet Sources
Considering the finite alphabet source given in (157), the boundary points for linear recovery read
2k − 1
ℓ2
s
vk = ±
a(1 + λ )
2
for k ∈ [1 : κ]. In the case of LASSO reconstruction, the boundary points are given by
2k − 1
ℓ1
s
a+λ
vk = ±
2
s
for k ∈ [1 : κ]. Finally, by employing ℓ0 norm recovery, we have v1ℓ0 = ± 21 a + λa and
vkℓ0 = ±
2k − 1
a
2
(171)
(172)
(173)
for k ∈ [2 : κ]. The source to noise power ratio snr is then defined as
snr :=
αa2 (κ + 1)(2κ + 1)
E x2
=
.
λ0
6λ0
(174)
5
10
50
0
−4
MSE0 in [dB]
satz-i.i.d.
projector
projector
d. Matrix
r Matrix
−8
−12
−16
obability
MMSE bound
ℓ0 norm
1
1.5
2
2.5
3
3.5
4
compression rate r
Fig. 13: Normalized MSE versus the compression rate r when snr = 5 dB and a = 1. Here, κ = 1 and the sparsity factor is
considered to be α = 0.1. The ℓ0 norm recovery scheme is employed for estimation and the MSE is numerically minimized
over λ considering an i.i.d. sensing matrix. The red curve indicates the MMSE bound on the normalized MSE. The figure
demonstrates a phase transition in the normalized MSE.
Fig. 12 shows the RS predicted error probability of the finite alphabet system as a function of snr for the unit
compression rate when the ℓ0 norm recovery scheme is employed. The cases with κ = 2 and κ = 4 are considered
for both random i.i.d. and orthogonal measurements. Here, the sparsity factor is set to be α = 0.1 and λ = 0.1.
As the figure illustrates, for either sensing matrix, the error probability meets the single-user bound in a relatively
large snr regime. The deviation from the single-user bound in the i.i.d. case, moreover, occurs in a larger interval
of snr as κ grows. In contrast to i.i.d. measurements, the coincidence occurs for almost every snr and κ when a
projector sensing matrix is employed. This observation is intuitively justified due to the orthogonality of the rows
in the latter case.
For the MSE, similar behavior as in Fig. 12 is observed in Fig. 13. As the compression rate increases, considering
either the MSE or error probability, a deviation from the single-user bound is observed for both random i.i.d. and
orthogonal measurements. Considering a fixed snr, it is observed that the MSE-compression rate has a discontinuity
point in which the MSE suddenly jumps from a lower value to an upper value at a certain compression rate. This
phenomenon is known as the “phase transition” in which the macroscopic parameters suddenly change the phase
within a minor variation of the setting. The compression rate in which the phase transition occurs is referred to as
“transition rate” which increases as snr grows. Phase transitions were reported in the literature of communications
51
and information theory for several problems such as turbo coding and CDMA systems [11], [11], [88].
Fig. 13 illustrates the phase transition under the RS assumption for the case of sparse binary source, i.e., κ = 1,
when the source vector is sensed via a random i.i.d. matrix and snr = 5 dB with a = 1. For estimation, the ℓ0 norm
recovery scheme is employed which is equivalent to LASSO in this particular example. The estimation parameter λ
is moreover optimized numerically such that the MSE is minimized. To determine each point on the curve, the RS
fixed point equations have been solved for the specific compression rate and all possible solutions have been found.
Within this set of solutions, those which are physically stable have been considered and the stable solution with
minimum free energy has been taken for the RS ansatz. The normalized MSE at each compression rate has then
been plotted using the RS ansatz. As the figure shows, the normalized MSE suddenly jumps to an upper bound
which converges to MSE0 = 0 dB. For sake of comparison, we have plotted the MMSE bound as well which
shows a samilar behavior at higher rates. In Fig. 14, we have further plotted the normalized MSE in terms of the
compression rate for the same setting with the sensing matrix replaced by a random projector matrix. The figure
shows a phase transition for orthogonal measurements at a higher transition rate compared to the case with random
i.i.d. sensing. Moreover, the improvement in terms of MSE observed in Fig. 14 by employing a projector matrix
extends the conclusion given for the sparse Gaussian sources in [40] to the sparse finite alphabet sources as well.
In order to obtain a more accurate approximation of the performance and transition point, one may investigate the
free energy and entropy of the corresponding spin glass by also considering RSB ansätze [41]. The RS ansatz,
however, gives an accurate approximation of the MSE for the specific setup considered here.
X. C ONCLUSION
This manuscript considered the performance of the MAP estimator in the large-system limit with respect to a
general distortion function. Taking a statistical mechanical approach, the replica method was employed to determine
the asymptotic distortion of the system. We deviated from the earlier approaches, e.g., [38]–[40], by evaluating the
general replica ansatz of the corresponding spin glass. The general ansatz let us derive the RS, as well as the bRSB
ansatz considering the class of rotationally invariant random matrices. The results recover the previous studies
[38]–[40] in special cases and justifies the uniqueness of the zero temperature entropy’s expression under the bRSB
assumption conjectured in [43]. The replica ansatz evaluated here led us to a more general form of the decoupling
principle. In fact, invoking the general replica ansatz, it was shown that for any tuple of input-output entries, the
marginal joint distribution converges to the input-output empirical distribution asymptotically. This means that in
the large-system limit, the marginal joint distribution of the entries determined by any replica ansatz decouples
into a set of identical joint distributions. The form of the asymptotic decoupled distribution, however, depends on
the structure imposed on the ansatz. For the bRSB ansatz, the vector-valued AWGN system estimated by a MAP
estimator decouples into single-user MAP-estimated AWGN channels followed by some correlated impairment terms
which intuitively model the interference of the system and vanish by reducing the bRSB assumption to RS. The
general decoupling principle justified here confirms the conjecture that decoupling is a generic property of MAP
estimators, since its validity relies only on the replica continuity assumption. Recent results in statistical mechanics
have shown that failures in finding the exact solution via the replica method are mainly caused by the structure
5
10
52
0
−4
MSE0 in [dB]
satz-i.i.d.
projector
projector
d. Matrix
r Matrix
−8
−12
−16
obability
Projector
i.i.d.
1
1.5
2
2.5
3
3.5
4
compression rate r
Fig. 14: MSE0 in terms of the compression rate r for snr = 5 dB and a = 1 considering i.i.d. and projector sensing matrices.
Here, α = 0.1 and κ = 1, and the MSE is numerically minimized with respect to the estimation parameter. As the figure shows,
the projector matrix enhances the reconstruction performance with respect to the both MSE and transition rate.
imposed on the ansatz, and not replica continuity [3], [89]–[91]. The decoupling property enabled us to represent
the equivalent “replica simulator” interpretation of the replica ansatz. In fact, the bRSB fixed point equations
are completely described by the statistics of the corresponding decoupled system. The bRSB ansatz is therefore
represented through the state evolution of a transition system which takes an initial set of ansatz parameters as the
input and determines a new set of parameters via simulating the bRSB decoupled system as the output.
As an example of vector-valued MAP-estimated systems, we considered the noisy compressive sensing system
and studied different sparse recovery schemes, including the linear, LASSO, and ℓ0 norm recovery schemes. The
numerical investigations showed that for sparse Gaussian sources, the performance of the linear and LASSO recovery
schemes are accurately approximated by the RS ansatz within a large interval of compression rates. The RS prediction
of the ℓ0 norm scheme’s performance, however, was observed to lack the stability within the moderate and high
compression rates. As a result, the 1RSB ansatz for this scheme was considered which deviates from the RS ansatz
significantly as the rate grows. For sparse finite alphabet sources, the RS prediction of the error probability and
MSE reported phase transitions with respect to the compression rate. The numerical results, moreover, showed a
better performance of the random orthogonal measurements compared to the random i.i.d. sensing in both the cases
which were previously reported for sparse Gaussian sources in [40].
The current work can be pursued in several directions. As an example, the “replica simulator” introduced in
53
Section VII can be studied further by methods developed in the context of transition systems. The analysis may
result in proposing a new framework which simplifies the evaluation of fixed point equations. Another direction is
the analysis of the conditional distribution of the correlated impairment terms in the bRSB decoupled system. The
study can lead us to understand the necessary or sufficient conditions in which the RS ansatz gives an accurate
approximation of the system’s performance. Inspired by the Sherrington-Kirkpatrick model of spin glasses, for
which the full RSB ansatz, i.e., b ↑ ∞, has been proved to give a stable solution for all system parameters [82],
our conjecture is that the exact solution at all large compression rates, for the cases exhibiting RSB, is given by
a large number of breaking steps. Nevertheless, as our numerical investigations depicted, even in those cases, the
RS ansatz, or the bRSB ansatz with small b, can give good approximations of the solution up to some moderate
compression rates. The latter study, furthermore, gives us some insights about the accuracy of the approximation
given by a finite number of RSB steps. Investigating the connection between the replica simulator and message
passing based algorithms is another interesting topic for future work.
XI. ACKNOWLEDGMENT
The authors would like to acknowledge German Research Foundation, Deutsche Forschungsgemeinschaft (DFG),
for support of this work under the Grant No. MU 3735/2-1.
54
A PPENDIX A
P ROOF
P ROPOSITION IV.1
OF
Starting from (42), we define the moment function Z(m) as in (43a). Therefore,
DW (x̂; x) = lim lim lim lim
1 1 ∂
log Z(m).
β↑∞ n↑∞ h↓0 m↓0 m n ∂h
(175)
Taking the expectation with respect to the noise term first, the moment function is extended as
Z(m) = Ex EA Ez
e−β
Pm
E(va |y,A)+hn
X
e−β
Pm
1
1 T
1
(x−v a )T J(x−va )+ λ
z A(x−v a )+ 2λ
kzk2 +u(v a )}+hn
{ 2λ
{v a }
= Ex EA Ez
{v a }
⋆
=
λ
λ + mβλ0
k2
a=1
a=1
Ex EA
X
e−β{
Pm
DW(n) (v a ;x)
X
a=1
Pm
{v a }
a,b=1
ṽ T
a Jṽ b ζab +
(176a)
Pm
a=1
u(v a )}+hn
Pm
a=1
Pm
a=1
DW(n) (va ;x)
DW(n) (v a ;x)
(176b)
(176c)
where ṽ a = x − v a for a ∈ [1 : m] is the unbiased1 replica vector, J is the Gramian of the system matrix defined
as J := AT A and satisfies the properties stated in Section II, ⋆ comes from taking expectation over z, and the
factor ζab is defined as
ζab
λ0
1
:=
1{a = b} −
β
2λ
λ + mβλ0
(177)
with λ0 being the true noise variance as specified in Section II.
Remark A.1. Considering (176b), one could drop the term
1
2
2λ kzk ,
since it plays no rule in the optimization
problem (2). More precisely, one could redefine the Hamiltonian in (10) to be
Enew (v|y, A) = E(v|y, A) −
1
kzk2
2λ
(178)
without loss of generality. In this case, the coefficient at the right hand side of (176c) reduces to 1, and ζab reads
1
λ0
new
:=
ζab
1{a = b} − β .
(179)
2λ
λ
It is, however, clear from (176c) and (177) that as m tends to zero, both the approaches result in a same result.
Considering the expression in (176c), we define the random variable Z(m; x) as
Z(m; x) := EJ
X
{v a }
e−β{
Pm
a,b=1
ṽ T
a Jṽb ζab +
Pm
a=1
u(v a )}+hn
Pm
a=1
DW(n) (v a ;x)
.
(180)
Consequently, the moment function is given by taking the expectation over Z(m; x) with respect to x and multiplying
k2
λ
. However, we later on show that for almost all realizations of the source vector, Z(m; x)
with the scalar λ+mβλ
0
converges to a deterministic value, and therefore, the expectation with respect to x can be dropped. For Z(m; x),
1 We
call it unbiased since it expresses the deviation of the replicas from the source vector.
55
we have
Z(m; x) = E J
X
{v a }
=
e−β
Pm
a,b=1
ṽ T
a Jṽ b ζab
× e−β
Pm
a=1
u(v a )+hn
Pm
a=1
DW(n) (va ;x)
i
Xh
Pm
Pm
W(n)(va ;x)
E J e−nβTr{JG} × e−β a=1 u(va )+hn a=1 D
(181a)
(181b)
{v a }
where Gn×n is defined as
G :=
m
1 X
ṽ b ṽ T
a ζa,b .
n
(182)
a,b=1
Considering the eigendecomposition of the Gramian J = UDUT , the expectation in (181b) can be expressed in
terms of a spherical integral where the integral measure is the probability measure of U. Regarding the system setup
specified in Section II, the matrix U is distributed over the orthogonal group On with Haar probability distribution.
Therefore, the corresponding spherical integral is the so-called “Harish-Chandra” or “Itzykson & Zuber” integral.
This integral has been extensively studied in the physics and mathematics literature, see for example [92], [93]
and [74]. A brief discussion on the spherical integral and its closed form solution has been given in Appendix F.
√
Invoking Theorem 1.2 and 1.7 in [75], as long as rank(G) = O( n)1 , the expectation in (181b) can be written as
E Je
−nβTr{JG}
=e
Pn R βλG
i R (−2ω)dω +ǫ
−n
J
n
i=1 0
(183)
with G being defined in (182), and ǫn ↓ 0 as n ↑ ∞. In order to employ the above result and substitute it in (181b),
we need to check the rank condition.
Lemma A.1. Considering G to be defined as in (182), the following argument holds.
√
rank(G) = O( n).
(184)
Proof. First, we rewrite G as
#
"m
m
m
X
X
1 X
λ0
T
G=
(
ṽ a )(
ṽ b )T
ṽ a ṽ a − β
2λn a=1
λ + mβλ0 a=1
(185a)
b=1
=
λ0
1
Ṽ(Im − β
1m )ṼT
2λn
λ + mβλ0
(185b)
where Ṽ = [ṽ 1 , . . . , ṽ m ] is an n × m matrix with the columns being the unbiased replicas. Then, by considering
(185b), it is obvious that G could be, at most, of rank m. As Assumption 2 indicates, Z(m) analytically continues
to the real axis, and the limit with respect to m is taken in a right neighborhood of 0. Therefore, for all values of
1 O(·)
indicates the growth order with respect to composition, i.e., lim [f (n)]−1 O(f (n)) = K < ∞.
n↑∞
56
n there exists a constant K ∈ R+ , such that m ≤ K. Consequently, one can write
lim
n↑∞
rank(G)
m
K
√
≤ lim √ ≤ lim √ = 0
n↑∞
n
n n↑∞ n
(186)
√
which concludes that rank(G) = O( n).
N
Lemma A.1 ensures that (183) always holds; therefore, noting the fact that G has only m non-zero eigenvalues,
the expectation in the right hand side of (181b) reduces to
v
E J e−nβTr{JG} = e−nG(TQ
)+ǫn
(187)
where the function G(·) is defined as
G(M) :=
Z
β
Tr{MRJ (−2ωM)}dω
(188)
0
for some square matrix M, T is an m × m deterministic matrix given by
λ0
1
Im −
β1m ,
T :=
2λ
λ + mβλ0
(189)
and Qv is the m × m “correlation matrix” defined as
Qv =
1 T
Ṽ Ṽ.
n
(190)
Remark A.2. Note that although Qv is symmetric, TQv is not a symmetric matrix, in general; however, due to
the symmetry of G, the eigenvalues of TQv are real, and therefore, the sequence of integrals over the real axis in
(183) exists for all indices.
By substituting (187) in (181b), Z(m; x) is given as
Z(m; x) =
X
{v a }
v
e−nG(TQ
)−β
Pm
a=1
u(v a )+hn
Pm
a=1
DW(n) (v a ;x)+ǫn
.
(191)
In order to determine the sum in (191), we follow the technique which has been employed in [16] and [43]. We
split the space of all replicas into subshells defined by the correlation matrices in which all the vectors of replicas
in each subshell have a same correlation matrix. More precisely, for a given source vector x, the subshell of the
matrix Qm×m is defined as
S(Q) = {v1 , . . . , vm |(x − v a )T (x − vb ) = nqab }
(192)
with qab = [Q]ab denoting the entry (a, b) of Q. The sum in (191) is determined first over each subshell, and then,
57
over all the subshells as the following.
X Z
Pm
Pm
W(n) (va ;x)+ǫn
e−nG(TQ) δ(Qv − Q)dQ e−β a=1 u(va )+hn a=1 D
Z(m; x) =
(193a)
{v a }
=
=
where dQ :=
Qm
a,b=1
Z
Z
e−nG(TQ)+ǫn
X
{v a }
δ(Qv − Q)e−β
Pm
a=1
u(v a )+hn
Pm
a=1
e−n[G(TQ)−I(Q)]+ǫn dQ
DW(n) (v a ;x)
dQ
(193b)
(193c)
Rm×m ,
dqab , the integral is taken over
δ(Qv − Q) :=
m
Y
a,b=1
δ(ṽ T
a ṽ b − nqab ),
(194)
and the term enI(Q) which determines the probability weight of the subshell S(Q) is defined as
enI(Q) :=
X
{v a }
δ(Qv − Q)e−β
Pm
a=1
u(v a )+hn
Pm
a=1
DW(n) (v a ;x)
.
(195)
Remark A.3. One may define the subshells over the transferred correlation matrix TQ instead of correlation matrix
Q. In this case the subshells over Q defined in (192) only rotate in the m-dimensional space with respect to T.
The rotation, however, does not have any impact on the analysis.
The last step is to determine enI(Q) . To do so, we represent the Dirac impulse function using its inverse Laplace
transform. By defining sab as the complex frequency corresponding to δ(ṽ T
a ṽ b − nqab ),
Z
T
dsab
δ(ṽ T
esab (ṽa ṽb −nqab )
a ṽ b − nqab ) =
2πj
(196)
where the integral is taken over the imaginary axis J = (t − j∞, t + j∞), for some t ∈ R. Consequently, by defining
the frequency domain correlation matrix S to be
Z
δ(Qv − Q) =
Z
=
with dS being defined as dS :=
an m × m matrix with [S]ab = sab , (194) reads
e
Pm
a,b=1
sab (ṽ T
a ṽ b −nqab )
dS
(197a)
i Pm
h
T
T
e−nTr{S Q} e a,b=1 sab ṽa ṽb dS
(197b)
m
Y
dsab
, and the integral being taken over Jm×m . Substituting in (195), enI(Q)
2πj
a,b=1
reduces to
enI(Q) =
=
Z
e−nTr{S
Z
e−nTr{S
T
Q}
X
{v a }
T
Q}
e
Pm
a,b=1
enM(S) dS
sab ṽ T
a ṽ b −β
Pm
a=1
u(v a )+hn
Pm
a=1
DW(n) (v a ;x)
dS
(198a)
(198b)
58
with M(S) being defined as
M(S) =
Pm
Pm
X Pm
T
W(n) (va ;x)
1
e a,b=1 sab ṽa ṽb −β a=1 u(va )+hn a=1 D
log
.
n
(199)
{v a }
Thus, Z(m; x) finally reads
Z(m; x) =
Z Z
e−n{G(TQ)+Tr{S
T
Q}−M(S)}+ǫn
dSdQ.
(200)
Consequently, one needs to evaluate the expectation of Z(m; x) with respect to x, in order to determine the
moment function Z(m). However, Z(m; x) for almost all realizations of x converges to a deterministic asymptotic
as n ↑ ∞, and thus, the expectation with respect to x can be dropped. To show the latter statement, we note that
the only term in (200) which depends on x is M(S). Therefore, it is sufficient to study the convergence of M(S).
Lemma A.2 justifies this property of M(S) using the law of large numbers, and the decoupling property of the
functions u(·) and d(·; ·).
Lemma A.2. Consider the system specified in Section II, and let Assumption 2 hold. Then, as n ↑ ∞, M(S)
defined in (199) is given by
(
"
M(S) = E
(1 − η) log
X
e(x−v)
T
S(x−v)−βu(v)
v
#
"
+ η log
X
e(x−v)
T
S(x−v)−βu(v)+hη
−1
d(v;x)
v
#)
(201)
where vm×1 ∈ Xm , xm×1 is a vector with all the elements being the random variable x which is distributed with the
Pm
source distribution px , the expectation is taken over px , and d(v; x) is defined as d(v; x) := a=1 d(va ; xa ).
Proof. Consider the decoupling property of the functions u(·) and d(·; ·). Define the vector vm×1 over the
support
Xm , and the coefficients {wi} for i ∈ [1 : n] as
wi =
Then, M(S) reads
M(S) =
0
|W(n)|−1
if i ∈
/ W(n)
if i ∈ W(n).
n P
Pm
Pm
XY
m
1
e a,b=1 sab (xi −vai )(xi −vbi )−β a=1 u(vai )+hn a=1 wi d(vai ;xi )
log
n
i=1
(202)
(203a)
{v a }
n X P
Pm
Pm
Y
m
1
log
e a,b=1 sab (xi −va )(xi −vb )−β a=1 u(va )+hn a=1 wi d(va ;xi )
n
i=1 v
"
#
X
X
1
=
M0 (S; xi ) +
M1 (S; xi )
n
=
i∈
/W
i∈W
(203b)
(203c)
59
where the functions M0 (·; ·) and M1 (·; ·) are defined as
M0 (S; xi ) = log
X
e(x−v)
M1 (S; xi ) = log
X
e
T
S(x−v)−βu(v)
(204a)
v
n
(x−v)T S(x−v)−βu(v)+h |W(n)| d(v;x)
(204b)
v
where xm×1 is a vector with all the elements being xi , and we define d(v; x) :=
Pm
a=1
d(va ; xa ) for compactness.
As Assumption 2 suggests the limits with respect to m and n can be exchanged in (175); thus, one can consider
the asymptotics of M(S) for a given m when n tends to its large limit. Regarding the fact that x is collected from
an i.i.d. distribution, the term in the right hand side of (203c) converges to the expectation over the distribution px
due to the law of large numbers; more precisely, as n ↑ ∞
X
|W(n)|
1 X
1
M0 (S; xi ) = 1 −
M0 (S; xi ) −→ (1 − η) Ex M0 (S; x)
n
n
n − |W(n)|
i∈
/W
i∈
/W
X
1 X
1
|W(n)|
M1 (S; xi ) =
M1 (S; xi ) −→ η Ex M1 (S; x)
n
n
|W(n)|
i∈W
i∈W
with η being defined as in (25). Substituting (205a) and (205b) in (203c), Lemma A.2 is concluded.
(205a)
(205b)
N
Remark A.4. Considering Lemma A.2, it eventually says that the probability weight enI(Q) for a given correlation
matrix Q converges to a deterministic weight as n tends to infinity. This statement equivalently states that for almost
any given realization of the source vector, the correlation matrix converges to its expectation. In fact, considering
the correlation matrix Qv , as defined in (190), the entries are functions of x, and therefore, variate randomly due
to the source distribution for a given n. Lemma A.2, however, indicates that, as n ↑ ∞, the entries converge to
some deterministic asymptotics for almost any realization of x. As an alternative approach, one could study the
convergence property of the correlation matrix Qv by means of the law of large numbers first, and then, conclude
Lemma A.2 by rewriting M(S) in proper way, and replacing it with the expectation using the fact that the probability
weight enI(Q) needs to converge deterministically as n ↑ ∞. Nevertheless, the approach taken here seems to be
more straightforward.
Using Lemma A.2, we drop the expectation with respect to x in (176c). Replacing in (175), the asymptotic
distortion is found by taking the limits. As Assumption 2 suggests, we exchange the order of the limits and take
the limit with respect to n at first. Denoting that the probability measure defined with enI(Q) dQ satisfy the large
deviation properties [81], we can use the saddle point approximation to evaluate the integral in (200) which says
that as n ↑ ∞
Z(m) =
λ
λ + mβλ0
k2 Z Z
e−n{G(TQ)+Tr{S
T
Q}−M(S)}
T
.
dSdQ = Kn e−n{G(TQ̃)+Tr{S̃ Q̃}−M(S̃)} ,
(206)
where we drop ǫn given in (200) regarding the fact that it vanishes in the large limit. Here, (Q̃, S̃) is the saddle
.
point of the integrand function’s exponent, Kn is a bounded coefficient, and = indicates the asymptotic equivalency
in exponential scale defined as the following.
60
Definition A.1. The functions a(·) and b(·) defined over the non-bounded set
X are said to be asymptotically
equivalent in exponential scale, if
lim log |
n↑∞
a(xn )
| = 0.
b(xn )
for an unbounded sequence {xn ∈ X}.
(207)
As n ↑ ∞, the mth moment can be replaced with its asymptotic equivalent in (175). Consequently, by substituting
the equivalent term and exchanging the limits’ order, we have
log Kn
1 ∂
−G(TQ̃) − Tr{S̃T Q̃} + M(S̃) +
DW (x̂; x) = lim lim lim lim
β↑∞ m↓0 h↓0 n↑∞ m ∂h
n
∂
1
⋆
= lim lim lim
M(S̃)
β↑∞ m↓0 h↓0 m ∂h
P
(x−v)T S̃(x−v)−βu(v)
v d(v; x)e
= lim lim E
P
β↑∞ m↓0
m v e(x−v)T S̃(x−v)−βu(v)
(208a)
(208b)
(208c)
where ⋆ comes from the fact that Kn is bounded, and G(TQ̃) and Tr{S̃T Q̃} are not functions of h.
Thesaddle point (Q̃, S̃) is found by letting the derivatives of the exponent zero. Using the standard definition
∂
∂
:=
, the saddle point is given by the following fixed point equations.
∂M ab
∂[M]ab
∂
[G(TQ) + Tr{SQ} − M(S)] |(Q̃,S̃) = 0
∂Q
∂
[G(TQ) + Tr{SQ} − M(S)] |(Q̃,S̃) = 0.
∂S
(209a)
(209b)
(209a) reduces to
S̃ = −βTRJ (−2βTQ̃),
(210)
and (209b) results in
Q̃ = E
P
v (x
T
− v)(x − v)T e(x−v) S̃(x−v)−βu(v)
.
P (x−v)T S̃(x−v)−βu(v)
ve
(211)
By replacing (210) in (208c) and (211), the expression for the asymptotic distortion and the saddle point correlation
matrix can be considered as expectations over a conditional Boltzmann-Gibbs distribution pβ
v|x defined as
e−β[(x−v) TRJ (−2βTQ̃)(x−v)+u(v)]
:= P
pβ
v|x (v|x)
−β[(x−v)T TRJ (−2βTQ̃)(x−v)+u(v)]
ve
T
(212)
which simplifies the expressions in (208c) and (211) to those given in Proposition IV.1.
In general the fixed point equation (211) can be satisfied with several saddle points, and therefore, multiple
asymptotic distortions might be found. In this case, one should note that the valid solution is the one which
minimizes the free energy of the spin glasses at the zero temperature, i.e., β ↑ ∞. Using the mth moment, the free
61
energy of the system reads
F(β) = − lim lim lim
n↑∞ h↓0 m↓0
1 11
log Z(m)
mβn
(213a)
i
1 h
⋆
= lim lim
G(TQ̃) + Tr{S̃T Q̃} − M(S̃)
(213b)
m↓0 h↓0 βm
"
#
X
T
1 1
1
†
= lim
G(TQ̃) − Tr{Q̃T TRJ (−2βTQ̃)} − E log
e−β(x−v) TRJ (−2βTQ̃)(x−v)−βu(v) (213c)
m↓0 m β
β
v
where ⋆ comes from the facts that Kn is bounded and the limits with respect to m and n are supposed to exchange,
and † is deduced from (210) and Lemma A.2. Finally by considering the definition of G(·), Proposition IV.1 is
concluded.
62
A PPENDIX B
P ROOF
OF
P ROPOSITION V.1
Starting from Assumption 3, the replica correlation matrix is
Q=
χ
Im + q1m
β
(214)
for some non-negative real χ and q. Considering Definition IV.1, the Hamiltonian of the spin glass of replicas is
given by
E R (v|x) = (x − v)T TRJ (−2βTQ)(x − v) + u(v)
(215)
with T being defined in (48). Denoting R := TRJ (−2βTQ), it is shown in Appendix E that R has the same
structure as the correlation matrix; thus, one can write
R = eIm − β
f2
1m ,
2
(216)
for some real f and e which are functions of χ and q. Denoting the eigendecomposition of Q as Q = VDQ VT ,
we have1
T = VDT VT
(217a)
R = VDR VT
(217b)
where DQ , DT and DR are the diagonal matrices of eigenvalues. Therefore, we have
DR = DT RJ (−2βDT DQ )
(218)
which equivalently states that for a ∈ [1 : m]
T
T Q
λR
a = λa RJ (−2βλa λa )
(219)
Q
T
with λR
to the ath column of V. The matrices
a , λa and λa being the eigenvalue of R, Q and T corresponding
mβλ0
1
f2 χ
1−
R, Q and T have two different corresponding eigenvalues, namely e − βm , + mq,
2 β
2λ
λ + mβλ0
χ 1
which occur with multiplicity m − 1. Substituting in (219) and
which occur with multiplicity 1 and e, ,
β 2λ
taking the limit when m ↓ 0, e and f are found as
χ
1
RJ (− ),
2λ
λ
1 ∂ n
χ o
2
f = 2
[λ0 χ − λq] RJ (− ) .
λ ∂χ
λ
e=
1 Note
that Q is full-rank and symmetric.
(220a)
(220b)
63
To pursue the analysis, we rewrite the Hamiltonian using (216)
E R (v|x) = ekx − vk2 − β
f2
Tr{(x − v)(x − v)T 1m } + u(v),
2
(221)
and therefore, the partition function Z R (β|x) is given by
Z R (β|x) =
X
e−βekx−vk
2
2
+β2 f2 Tr{(x−v)(x−v)T 1m }−βu(v)
.
(222)
{va }
Using the Gaussian integral, we have
2
eβ
f2
T
2 Tr{(x−v)(x−v) 1m }
=
Z
e−βf [
Pm
a=1 (x−va )
]z Dz,
(223)
and thus, the partition function reduces to
R
Z (β|x) =
Z "X
e
−β[e(x−v)2 +f (x−v)z+u(v)]
v
#m
Dz
(224)
with v ∈ X. The parameters of the spin glass of replicas are then determined using the partition function. Starting
with the normalized free energy, it reads
1
F (β, m) = −
Ex log
βm
R
Noting that
R
Z "X
e
−β[e(x−v)2 +f (x−v)z+u(v)]
v
#m
Dz.
(225)
Dz takes expectation over the Gaussian distribution, one can use the Riesz equality in (41) to show
that when m varies in a vicinity of 0
1
FR (β, m) = − E
β
Z
log
X
e−β[e(x−v)
2
+f (x−v)z+u(v)]
Dz + ǫm
(226)
v
where ǫm tends to 0 as m ↓ 0 and the expectation is taken over x ∼ px . Consequently, as m ↓ 0 the normalized
free energy reads
1
FR (β) = lim FR (β, m) = − E
m↓0
β
Z
log
X
e−β[e(x−v)
2
+f (x−v)z+u(v)]
Dz
(227)
v
The next parameters to be specified are χ and q. By determining the conditional distribution pβ
v|x and substituting
in (49), the following fixed point equations are deduced
X
χ
+ q m = Ex
kx − vk2 pβ
v|x (v|x),
β
v
X
χ
+ mq m = Ex
Tr{(x − v)(x − v)T 1m } pβ
v|x (v|x).
β
v
(228a)
(228b)
where (228a) and (228b) are found by taking the trace and sum over all the entries of the both sides of (49),
respectively. One can directly evaluate the right hand sides of (228a) and (228b); however, considering (222), it is
64
straightforward to show that
Ex
kx − vk2 pβ
v|x (v|x) = m
X
Tr{(x − v)(x − v)T 1m } pβ
v|x (v|x) = −
v
Ex
∂ R
F (β, m),
∂e
X
v
(229a)
m ∂ R
F (β, m).
βf ∂f
(229b)
After substituting and taking the limit m ↓ 0, the fixed point equations finally read
χ
+q =E
β
Z P
1
χ= E
f
2
x)2 e−β[e(x−v) +f (x−v)z+u(v)]
Dz
−β[e(x−v)2 +f (x−v)z+u(v)]
ve
−
v (v
P
Z P
x)ze−β[e(x−v) +f (x−v)z+u(v)]
Dz.
−β[e(x−v)2 +f (x−v)z+u(v)]
ve
2
−
v (v
P
with f and e defined in (220a) and (220b).
(230a)
(230b)
In order to determine the replicas’ average distortion defined in (53) regarding the distortion function d(·; ·), we
replace the Hamiltonian by
EhR (v|x) = E R (v|x) + h
m
X
d(va ; x)
(231)
a=1
with E R (v|x) given in (221), and take the steps as in (222)-(225) to find the modified form of the normalized free
energy, i.e. FR (β, h, m). The replicas’ average distortion is then evaluated as
∂ R
F (β, h, m)|h=0
∂h
2
Z P
d(v; x)e−β[e(x−v) +f (x−v)z+u(v)]
vP
Dz.
=E
−β[e(x−v)2 +f (x−v)z+u(v)]
ve
DR (β, m) =
(232a)
(232b)
which does not depend on m, and thus, taking the limit m ↓ 0 is not needed.
The last step is to take the zero temperature limit. Using the Laplace method of summation, as β ↑ ∞ the fixed
point equations reduce to
Z
(g − x)2 Dz,
Z
1
χ = E (g − x)z Dz.
f
q=E
(233a)
(233b)
with g being defined as
g := arg min e(x − v)2 + f (x − v)z + u(v) .
v
Taking the same approach, the replicas’ average distortion at zero temperature reads
Z
DW = E d(g; x) Dz.
(234)
(235)
In order to avoid multiple solutions, we need to find the normalized free energy of the corresponding spin glass
as given in Proposition IV.1. In fact, the fixed point equations in (233a) and (233b) may have different solutions,
and therefore, the several asymptotics for the distortion can be obtained. In this case, the fixed point solution which
65
minimizes the zero temperature free energy of the system and its corresponding asymptotic distortion are taken.
Substituting in Proposition IV.1, the free energy of the corresponding spin glass at the inverse temperature β is
found as
F(β) =
1
2λ
Z
0
1
Fβ (ω)dω − Fβ (1) + FR (β)
(236)
where the function Fβ (·) is defined as
Fβ (ω) =
χ
χ i
χ
λ0
d h
ωRJ (− ω) .
RJ (− ω) + q − χ
β
λ
λ
dω
λ
By taking the limit as β ↑ ∞, the zero temperature free energy reads
Z 1
Z
1
F0 =
F∞ (ω)dω − F∞ (1) + E e(x − g)2 + f (x − g)z + u(g) Dz
2λ 0
(237)
(238)
−1 2
−1
f
with g being defined as in (234) and F∞ (ω) := limβ↑∞ Fβ (ω). By defining λs := [2e] and λs0 := 4e2
Proposition V.1 is concluded.
66
A PPENDIX C
P ROOF
OF
P ROPOSITION VI.1
We take the same approach as in Appendix B. Considering the replica correlation matrix to be of the form
Q=
χ
Im + pI mβ ⊗ 1 βµ + q1m ,
µ
β
(239)
for some non-negative real χ, p, q, and µ, we need to evaluate the parameters of the spin glass of replicas defined
in Definition IV.1. Starting with the Hamiltonian,
E R (v|x) = (x − v)T TRJ (−2βTQ)(x − v) + u(v)
(240)
where T is given in (48). As discussed in Appendix E, for a given µ the matrix R := TRJ (−2βTQ) is of the
following form
R = eIm − β
g2
f2
I mβ ⊗ 1 βµ − β 1m
2 µ
2
(241)
where e, g and f can be found in terms of χ, p and q. Using the eigendecomposition of R, Q and T, it is then
straightforward to show that for a ∈ [1 : m]
T
T Q
λR
a = λa RJ (−2βλa λa )
(242)
Q
T
where λR
a , λa and λa denote the ath eigenvalues of R, Q and T, respectively. Regarding the structure of Q and
R, there are three different sets of corresponding eigenvalues for R, Q and T, namely
•
•
•
mβλ0
f 2 χ + µp
1
g2
1−
with multiplicity 1,
+ mq,
e − µ − mβ ,
2
2
2λ
λ + mβλ0
β
2
g χ + µp 1
with multiplicity mβµ−1 − 1, and
,
e−µ ,
2 β
2λ
χ 1
with multiplicity m − mβµ−1 .
e, ,
β 2λ
Thus, by substituting in (242) and taking the limit when m ↓ 0, e, g and f are given as
1
χ
RJ (− ),
2λ
λ
χ
1
χ + µp
2
RJ (− ) − RJ (−
g =
) ,
λµ
λ
λ
χ + µp
1 ∂
2
[λ0 (χ + µp) − λq] RJ (−
) .
f = 2
λ ∂χ
λ
e=
(243a)
(243b)
(243c)
The next step is to evaluate the partition function. Substituting (241) in (240), the Hamiltonian reads
E R (v|x) = ekx − vk2 − β
g2
f2
Tr{(x − v)(x − v)T I mβ ⊗ 1 βµ } − β Tr{(x − v)(x − v)T 1m } + u(v).
µ
2
2
(244)
67
The partition function is then determined as in (51). Substituting in (51) and using the equalities
1 2 2
T
e 2 β f Tr{(x−v)(x−v) 1m }
e
1 2 2
T
⊗1 µ }
2 β g Tr{(x−v)(x−v) I mβ
β
µ
=
=
Z
e
−βf
Ξ Z
Y
e
m
P
(x−va ) z0
a=1
−βg
"
̺
˘k
P
Dz0 ,
#
(245a)
(x−va ) z1
a=̺k
Dz1 ,
(245b)
k=0
where ̺k = kµβ−1 + 1, ̺˘k = (k + 1)µβ−1 , and Ξ = mβµ−1 − 1, the partition function is found as
Z R (β; µ|x) =
Z
Z "X
e−β[e(x−v)
2
+(f z0 +gz1 )(x−v)+u(v)]
v
# βµ
mβ
µ
Dz1
Dz0
(246)
with v ∈ X where we denoted µ in the argument of the partition function to indicate that the expression is determined
for a given µ. The normalized free energy of the spin glass of replicas then reads
FR (β, m; µ) = −
1
Ex log
βm
Z
"
mβ
# βµ
µ
Z X
−β[e(x−v)2 +(f z0 +gz1 )(x−v)+u(v)]
e
Dz1
Dz0 .
(247)
v
Using the Riesz equality and taking the limit m ↓ 0, the normalized free energy reduces to
"
# βµ
Z
Z X
2
1
Dz1 Dz0 .
FR (β; µ) = lim FR (β, m; µ) = − E log
e−β[e(x−v) +(f z0 +gz1 )(x−v)+u(v)]
m↓0
µ
(248)
v
In order to find the fixed point equations, we use (49); therefore,
X
χ
+ q + p m = Ex
kx − vk2 pβ
v|x (v|x),
β
v
X
χ µp µq
m = Ex
Tr{(x − v)(x − v)T I mβ ⊗ 1 βµ } pβ
+
+
v|x (v|x),
µ
β
β
β
v
X
χ µp
+
+ mq m = Ex
Tr{(x − v)(x − v)T 1m } pβ
v|x (v|x)
β
β
v
(249a)
(249b)
(249c)
where (249a), (249b) and (249c) are concluded by taking the trace, sum over the diagonal blocks and sum over all
the entries of the both sides of (49), respectively. To evaluate the right hand sides of (249a)-(249c), we take the
alternative approach and express the expectations as
Ex
kx − vk2 pβ
v|x (v|x) = m
X
Tr{(x − v)(x − v)T I mβ ⊗ 1 βµ } pβ
v|x (v|x) = −
X
Tr{(x − v)(x − v)T 1m } pβ
v|x (v|x) = −
v
Ex
v
Ex
∂ R
F (β, m; µ),
∂e
X
v
µ
(250a)
m ∂ R
F (β, m; µ),
βg ∂g
m ∂ R
F (β, m; µ).
βf ∂f
(250b)
(250c)
68
Taking the derivatives and limit m ↓ 0, the fixed point equations finally reduce to
2
x)2 e−β[e(x−v) +(f z0 +gz1 )(x−v)+u(v)] β
Λ̃ Dz1 Dz0
−β[e(x−v)2 +(f z0 +gz1 )(x−v)+u(v)]
ve
2
Z P
− x)z1 e−β[e(x−v) +(f z0 +gz1 )(x−v)+u(v)] β
1
v (v
P −β[e(x−v)2 +(f z +gz )(x−v)+u(v)]
χ + µp + µq = E
Λ̃ Dz1 Dz0 ,
0
1
g
ve
2
Z P
− x)z0 e−β[e(x−v) +(f z0 +gz1 )(x−v)+u(v)] β
1
v (v
P −β[e(x−v)
χ + µp = E
Λ̃ Dz1 Dz0
2 +(f z +gz )(x−v)+u(v)]
0
1
f
ve
χ
+q+p=E
β
with Λ̃β :=
R
Λβ Dz1
−1
Z P
−
v (v
P
(251a)
(251b)
(251c)
Λβ and Λβ being defined as
β
Λ :=
"
X
e
−β[e(x−v)2 +(f z0 +gz1 )(x−v)+u(v)]
v
# βµ
.
(252)
The replicas’ average distortion regarding the distortion function d(·; ·) is further determined by modifying the
Hamiltonian as
EhR (v|x) = E R (v|x) + h
m
X
d(va ; x)
(253)
a=1
with E R (v|x) given in (244), and taking the steps as in (244)-(247) to find the modified form of the normalized
free energy, i.e. FR (β, h, m; µ). The replicas’ average distortion then reads
∂ R
F (β, h, m; µ)|h=0
∂h
2
Z P
d(v; x)e−β[e(x−v) +(f z0 +gz1 )(x−v)+u(v)] β
vP
Λ̃ Dz1 Dz0 .
=E
−β[e(x−v)2 +(f z0 +gz1 )(x−v)+u(v)]
ve
DR (β; µ) = lim
m↓0
The analysis is concluded by taking the zero temperature limit. As β ↑ ∞, (251a)-(251c) read
Z
q + p = E (g − x)2 Λ̃Dz1 Dz0 ,
Z
1
χ + µq + µp = E (g − x)z1 Λ̃Dz1 Dz0 ,
g
Z
1
χ + µp = E (g − x)z0 Λ̃Dz1 Dz0 ,
f
(254a)
(254b)
(255a)
(255b)
(255c)
where g is defined as
g := arg min e(x − v)2 + (f z0 + gz1 )(x − v) + u(v)
v
and Λ̃ :=
R
ΛDz1
−1
(256)
Λ with Λ denoting
Λ := lim Λβ
(257a)
β↑∞
= e−µ[e(x−g)
2
+(f z0 +gz1 )(x−g)+u(g)]
.
(257b)
69
Moreover, the asymptotic distortion for a given µ reads
Z
DW = E d(g; x)Λ̃Dz1 Dz0 .
(258)
(248) as well as (255a)-(258) are determined in terms of µ. Moreover, for a given µ, multiple solution to the fixed
point equations can be found. Proposition IV.1 suggests us to choose the solution which minimizes the free energy.
Therefore, one needs to find the optimal µ, and its corresponding χ, p and q, such that the free energy meets its
minimum value. As the second law of thermodynamics is satisfied at any inverse temperature, we should initially
search for the optimal µ considering a given β. We, then, find the corresponding χ, p, and q which minimize the
zero temperature free energy. Using Proposition IV.1, the free energy at the inverse temperature β for a given µ is
written as
F(β; µ) =
1
2λ
Z
0
1
Fβ (ω; µ)dω − Fβ (1; µ) + FR (β; µ)
(259)
where the function Fβ (·; µ) is defined as
1 d
F (ω; µ) =
µ dω
β
Z
[χ+µp]ω
χω
χ + µp
χ + µp d
χ
χ
t
ωRJ (−
ω) .
RJ (− )dt + RJ (− ω) + q − λ0
λ
β
λ
λ
dω
λ
(260)
To find µ at the thermal equilibrium, we let
∂
F(β; µ) = 0
∂µ
Using the equalities (243a)-(243c), (261) concludes that µ satisfies
Z χ+µp
1
χ
χ
χ + µp
1
t
R
pRJ (− ) + qRJ (− ) − qRJ (−
) = F (β; µ) +
RJ (− )dt
2λ
λ
λ
λ
2λµ χ
λ
"
#
Z
X
2
1
+E
log
e−β[e(x−v) +(f z0 +gz1 )(x−v)+u(v)] Λ̃β Dz1 Dz0
β
v
which as β ↑ ∞ reduces to
Z χ+µp
Z
χ
χ
χ + µp
1
1
t
1
pRJ (− ) + qRJ (− ) − qRJ (−
) =
RJ (− )dt + E
log Λ̃ Λ̃Dz1 Dz0 .
2λ
λ
λ
λ
2λµ χ
λ
µ
(261)
(262)
(263)
Denoting the solution to (263) by µ⋆ , the free energy of the corresponding spin glass is then given as F(β) =
F(β; µ⋆ ) which at the zero temperature reads
Z 1
Z
Z
1
1
0
∞
∞
F =
F (ω)dω − F (1) − E log
ΛDz1 Dz0
2λ 0
µ
(264)
with
F∞ (ω) := lim Fβ (ω; µ⋆ ).
β↑∞
−1 2
−1 2
−1
g1 , Proposition VI.1 is concluded.
f and λs1 := 4e2
Finally by defining λs := [2e] , λs0 := 4e2
(265)
70
A PPENDIX D
P ROOF
OF
P ROPOSITION VI.3
The strategy here is to extend the approach in Appendix C to a general number of breaking steps. Following
Appendix E and considering Q as
Q=
b
X
χ
Im +
pκ I mβ ⊗ 1 µβκ + q1m ,
µκ
β
κ=1
(266)
the frequency domain correlation matrix R := TRJ (−2βTQ) is written as
b
X
gκ2
f2
I mβ ⊗ 1 µβκ − β 1m
R = eIm − β
2 µκ
2
κ=1
(267)
T
considering a given vector µ = [µ1 , . . . , µb ] , such that
µκ+1 = ϑκ+1 µκ ,
(268)
with {ϑκ } being non-negative integers, e, f and {gκ } are then found in terms of χ, q and {pκ } by letting
T
T Q
λR
a = λa RJ (−2βλa λa )
(269)
Q
T
for a ∈ [1 : m] where λR
a , λa and λa denote the ath corresponding eigenvalues of R, Q and T. As long as the
constraint in (268) holds, Q, T and R have b + 2 different sets of corresponding eigenvalues specified by
•
•
•
•
b
X
)
b
gκ2
mβλ0
f 2 χ X µκ
1
e−
µκ
1−
with multiplicity Θb+1 (m) = 1,
− mβ , +
pκ
+ mq,
2
2 β κ=1
β
2λ
λ + mβλ0
κ=1
)
(
b
b
X
g 2 χ X µκ 1
with multiplicity Θb (m) = mβµ−1
pκ ,
e−
µκ κ , +
b − 1,
2
β
β
2λ
κ=1
κ=1
)
(
κ
κ
X
gς2 χ X µς 1
−1
with multiplicity Θκ (m) = mβ µ−1
pς ,
e−
µς , +
κ − µκ+1 for κ ∈ [1 : b − 1], and
2 β ς=1 β 2λ
ς=1
χ 1
e, ,
with multiplicity Θ0 (m) = m − mβµ−1
1 .
β 2λ
(
Substituting in (269) e, f and {gκ } for κ ∈ [1 : b] are determined in terms of χ, q and {pκ } as
χ
1
RJ (− ),
2λ λ
χ̃κ−1
1
χ̃κ
2
RJ (−
gκ =
) − RJ (− ) ,
λµκ
λ
λ
1
χ̃
∂
b
2
f = 2
[λ0 χ̃b − λq] RJ (− ) .
λ ∂ χ̃b
λ
e=
(270a)
(270b)
(270c)
where we define χ̃0 := χ and
χ̃κ := χ +
κ
X
ς=1
µς p ς
(271)
71
for κ ∈ [1 : b]. The Hamiltonian of the spin glass of replicas is then determined as in (47). Substituting the
Hamiltonian in (51) and using the equalities
1 2 2
T
e 2 β f Tr{(x−v)(x−v) 1m }
e
1 2 2
T
⊗1 µκ }
2 β gκ Tr{(x−v)(x−v) I mβ
β
µ
κ
=
=
Z
e−βf [
Ξκ Z
Y
e
k=0
Pm
a=1 (x−va )
−βgκ
]z0 Dz ,
0
κ
̺
˘k
P
a=̺κ
k
(272a)
(x−va )zκ
Dzκ ,
(272b)
with ̺κk = kµκ β−1 + 1, ̺˘κk = (k + 1)µκ β−1 , and Ξκ = mβµ−1
κ − 1, the partition function finally reads
Z R (β; µ|x) =
Z
b
^
ς=2
Z
Z
X
e
"
−β e(x−v)2 +(f z0 +
b
P
β
gκ zκ )(x−v)+u(v)
κ=1
v
with v ∈ X where for the sequences {ξς } and {zς } we define
b Z
^
ς=1
# µ1
µς
µς−1
Dz1
mβ
µb
Dzς
#ξb
ξ2
Z " Z Z
ξ1
Dz2 · · ·
Dzb .
···
F Dz1
F Dzς :=
ξς
Dz0
(273)
(274)
Consequently, one evaluates the free energy as in (52) which by using the Riesz equality when m ↓ 0 reduces to
"
# µς
µς−1
b
P
Z
2
b Z
−β
e(x−v)
+(f
z
+
g
z
)(x−v)+u(v)
^
X
0
κ κ
1
R
κ=1
(275)
F (β; µ) = − E log
Dzς Dz0
e
µb
ς=1
v
where we have defined µ0 = β for sake of compactness. The fixed point equations are, moreover, found via (49)
where we have
#
b
X
χ X
+
pκ + q m = E x
kx − vk2 pβ
v|x (v|x),
β κ=1
v
!#
b
X
µκ X
+
pς + q
m = Ex
Tr{(x − v)(x − v)T I mβ ⊗ 1 µβκ } pβ
v|x (v|x),
µκ
β ς=κ
v
X
χ̃b
+ mq m = Ex
Tr{(x − v)(x − v)T 1m } pβ
v|x (v|x)
β
v
"
"
χ̃κ−1
β
(276a)
(276b)
(276c)
for κ ∈ [1 : b]. (276a) and (276c) are found by taking trace and sum over all the entries from both sides of (49),
respectively. (276b) is moreover concluded by adding up the entries over the diagonal blocks of size µκ β−1 . The
right hand sides of (276a)-(276c) can then be evaluated using the equalities
Ex
kx − vk2 pβ
v|x (v|x) = m
X
Tr{(x − v)(x − v)T I mβ ⊗ 1 µβκ } pβ
v|x (v|x) = −
X
Tr{(x − v)(x − v)T 1m } pβ
v|x (v|x) = −
v
Ex
v
Ex
∂ R
F (β, m; µ),
∂e
X
v
µκ
(277a)
m ∂ R
F (β, m; µ),
βgκ ∂gκ
m ∂ R
F (β, m; µ).
βf ∂f
(277b)
(277c)
72
Thus, the fixed point equations are finally concluded as
b
χ X
+
pκ + q = E
β κ=1
χ̃κ−1 + µκ
b
X
Z P
1
E
f
Z P
− x)2 e
P
v
pς + q
ς=κ
χ̃b =
v (v
!
=
e
1
E
gκ
"
−β e(x−v)2 +(f z0 +
"
−β e(x−v)2 +(f z0 +
Z P
v (v
P
− x)z0 e
P
v
e
"
e(x−v)2 +(f z
b
P
0+
κ=1
gκ zκ )(x−v)+u(v)
"
−β e(x−v)2 +(f z0 +
"
−β e(x−v)2 +(f z0 +
b
P
gκ zκ )(x−v)+u(v)
κ=1
gκ zκ )(x−v)+u(v)
κ=1
for κ ∈ [1 : b] in which we denote Λ̃β
κ :=
Λβ
κ Dzκ
R
b
P
#
#
b
P
b
Y
Λ̃β
κ Dzκ Dz0
gκ zκ )(x−v)+u(v)
κ=1
gκ zκ )(x−v)+u(v)
κ=1
−1
#
#
b
Y
(278a)
κ=1
#
×
#
×
−β e(x−v)2 +(f z0 +
"
−β
e
gκ zκ )(x−v)+u(v)
κ=1
− x)zκ e
v
v (v
b
P
b
P
b
Y
Λ̃β
κ Dzκ Dz0
(278b)
κ=1
Λ̃β
κ Dzκ Dz0 .
(278c)
κ=1
β
Λβ
κ with Λ1
"
# µ1
b
β
P
2
X −β e(x−v) +(f z0 + gκ zκ )(x−v)+u(v)
β
κ=1
Λ1 :=
e
(279)
v
and {Λβ
κ } for κ ∈ [2 : b] being recursively defined as
Λβ
κ
:=
Z
Λβ
κ−1
Dzκ−1
µκ
µκ−1
.
(280)
The replicas’ average distortion regarding the distortion function d(·; ·) is further determined using the Hamiltonian
modification technique employed in Appendix B and C. After modifying the Hamiltonian and taking the derivatives,
the average distortion at the inverse temperature β is given by
DR (β; µ) = E
Z P
v
d(v; x)e
P
v
e
"
−β e(x−v)2 +(f z0 +
"
−β
e(x−v)2 +(f z
0+
b
P
b
P
gκ zκ )(x−v)+u(v)
κ=1
gκ zκ )(x−v)+u(v)
κ=1
#
#
b
Y
Λ̃β
κ Dzκ Dz0 .
(281)
κ=1
Finally, by taking the limit β ↑ ∞, we find the asymptotic distortion as
DW = E
Z
d(g; x)
b
Y
Λ̃κ Dzκ Dz0
(282)
κ=1
where g is defined as
"
2
g := arg min e(x − v) + (f z0 +
v
b
X
κ=1
#
gκ zκ )(x − v) + u(v) ,
(283)
73
β
and Λ̃κ denotes the limiting factor Λ̃∞
κ . Considering the definition of Λ̃κ , Λ̃κ reads Λ̃κ =
Λ1 := e
"
−µ1 e(x−g)2 +(f z0 +
b
P
gκ zκ )(x−g)+u(g)
κ=1
#
R
Λκ Dzκ
−1
Λκ with
(284)
and {Λκ } for κ ∈ [2 : b]
Λκ :=
Z
Λκ−1 Dzκ−1
µκ
µκ−1
.
(285)
Moreover, the fixed point equations reduce to
b
X
pκ + q = E
κ=1
b
X
χ̃κ−1 + µκ
pς + q
ς=κ
!
Z
(g − x)2
1
=
E
gκ
1
χ̃b = E
f
Z
Z
b
Y
Λ̃κ Dzκ Dz0 ,
(286a)
κ=1
(g − x)zκ
(g − x)z0
b
Y
Λ̃κ Dzκ Dz0 ,
(286b)
κ=1
b
Y
Λ̃κ Dzκ Dz0 ,
(286c)
κ=1
for κ ∈ [1 : b].
As in the 1RSB ansatz, we set µ to be the extreme point of the free energy at a given inverse temperature β,
in order to satisfy the second law of thermodynamics. The solution needs to be found over the set of non-negative
real vectors which satisfy the constraint in (268). The parameters of the ansatz, however, are finally taken such that
the zero temperature free energy is minimized.
Using Proposition IV.1, the free energy of the corresponding spin glass for a given vector µ is written as
Z 1
1
β
β
F(β; µ) =
F (ω; µ)dω − F (1; µ) + FR (β; µ)
(287)
2λ 0
where the function Fβ (·; µ) is defined as
Fβ (ω; µ) =
Z χ̃κ ω
b
X
1 d
χ̃b
d
χ
χ
λ0
t
ωRJ (− ω) .
RJ (− )dt + RJ (− ω) + q − χ̃b
µ dω χ̃κ−1 ω
λ
β
λ
λ
dω
λ
κ=1 κ
(288)
Therefore, the vector µ⋆ , for a given β, is set as
µ⋆ = arg min F(β; µ)
(289)
µ
with µ ∈
Sµ where Sµ is the set of non-negative real vectors satisfying the constraint in (268). By substituting
(270a)-(270c) in (288), µ⋆ reduces to
µ⋆ = arg min
µ
with µ ∈ Sµ where ∆(·) reads
1
∆(µ) :=
e
(
1
2λ
Z
0
1
Fβ (ω; µ)dω + FR (β; µ) − e∆(µ)
b
X
1
f2
ẽ0 χ̃0
[ẽκ χ̃κ − ẽκ−1 χ̃κ−1 ] +
+ ẽb q −
χ̃b
µ
β
2
κ=1 κ
(290)
)
(291)
74
with ẽ0 := e and
ẽκ := e −
κ
X
µς
ς=1
gς2
2
(292)
for κ ∈ [1 : b]. The vector µ⋆ is then determined such that it minimizes the free energy . Finally by taking the
limit β ↑ ∞, the zero temperature free energy is evaluated as
Z 1
Z
Z
1
1
0
∞
∞
F =
F (ω)dω − F (1) − E log
Λb Dzb Dz0
2λ 0
µb
(293)
where we define
F∞ (ω) := lim Fβ (ω; µ⋆ ).
β↑∞
(294)
−1 2
−1 2
−1
gκ for κ ∈ [1 : b], and defining the sequence {ζκ }
f and λsκ := 4e2
Denoting λs := [2e] , λs0 := 4e2
such that ζ0 = 1 and
ζκ := 1 −
for κ ∈ [1 : b], Proposition VI.3 is concluded.
κ
X
ς=1
µς
λsς
λs
(295)
75
A PPENDIX E
G ENERAL RSB F REQUENCY D OMAIN C ORRELATION M ATRIX
Consider the spin glass of replicas defined in Definition IV.1, the Hamiltonian reads
E R (v|x) = (x − v)T R(x − v) + u(v).
(296)
where R := TRJ (−2βTQ) is referred to as the “frequency domain correlation matrix”. In this appendix, we show
that under the general RSB assumption on Q, including the RS case, the frequency domain correlation matrix has
the same structure with different scalar coefficients. To show that, let the correlation matrix be of the form
Q = q0 Im +
b
X
i=1
qi I ξm ⊗ 1ξi + qb+1 1m
i
(297)
for some integer b where q0 , qb+1 6= 0. (297) represents the bRSB as well as RS structures by setting the coefficients
correspondingly. Considering T as defined in (48), TQ is then written as
βλ0
1
Q−
1m Q .
TQ =
2λ
λ + mβλ0
(298)
Defining the vector um×1 as a vector with all entries equal to 1, it is clear that u is an eigenvector of Q, and
therefore, by denoting the eigendecomposition of Q as VDQ VT , 1m reads
1m = uuT = VD1 VT
(299)
where D1 is a diagonal matrix in which all the diagonal entries expect the entry corresponding to the eigenvector
u are zero. Consequently, (298) reduces to
TQ =
βλ0
1
D1 DQ VT
V DQ −
2λ
λ + mβλ0
(300)
which states that TQ and Q span the same eigenspace. The eigenvalues of TQ and Q are also distributed with
the same frequencies. In fact, as the eigenvalue corresponding to u occurs with multiplicity 1, the second term
on the right hand side of (298) does not change the distribution of eigenvalues and only modifies the eigenvalue
corresponding to u. Therefore, TQ can be also represented as in (297) with different scalar coefficient.
To extend the scope of the analysis to R, we note that the function RJ (·) is strictly increasing for any FJ different
from the single mass point CDF1 [43]. Consequently, the eigenvalues’ distribution remains unchanged, and thus,
R = r0 Im +
b
X
i=1
ri I ξm ⊗ 1ξi + rb+1 1m .
i
(301)
for some real {ri }. In the case that FJ is the single mass point CDF, the R-transform becomes a constant function
which results in RJ (−2βTQ) = KIm for some constant K. Therefore, R = KT which is again represented as in
(301) by setting ri = 0 for i ∈ [1 : b]. This concludes that R has the same structure as Q for any FJ .
1 In
the single mass point CDF, we have FJ (λ) = 1{λ ≥ K} for some real constant K.
76
A PPENDIX F
A SYMPTOTICS
OF
S PHERICAL I NTEGRAL
Consider µζn to be the Haar measure on the orthogonal group
On for ζ = 1, and on the unitary group Un for
ζ = 2. Let Gn and Dn be n × n matrices; then, the integral of the form
Z
†
ζ
In (Gn , Dn ) := enTr{UGn U Dn } dµζn (U),
(302)
is known as the “spherical integral”. This integral has been extensively studied in the mathematics literature, as well
as physics where it is often called “Harish-Chandra” or “Itzykson & Zuber” integral. In a variety of problems, such
as ours, the evaluation of spherical integrals in asymptotic regime is interesting, and therefore, several investigations
have been done on this asymptotics. In [74], the asymptotics of the integral has been investigated when the matrices
Gn and Dn have n distinct eigenvalues with converging spectrums, and under some assumptions, a closed form
formula has been given; however, the final formula in [74] is too complicated and hard to employ. In [75], the
authors showed that, for a low-rank Gn , the asymptotics of the integral can be written directly in terms of the
R-transform corresponding to the asymptotic eigenvalue distribution of Dn . As long as the replica analysis is being
considered, we can utilize the result from [75], since the number of replicas can be considered to be small enough.
In [75], Theorem 1.2, it is shown that when Gn is a rank-one matrix, under the assumption that the spectrum of
Dn asymptotically converges to a deterministic CDF FD with compact and finite length support, the asymptotics
of the integral can be written in terms of the R-transform RD (·) as
1
log Iζn (Gn , Dn ) =
n↑∞ n
lim
Z
θ
RD (
0
2ω
)dω,
ζ
(303)
in which θ denotes the single nonzero eigenvalue of Gn . The authors further showed in Theorem 1.7 that in the
√
case of rank(Gn ) = O( n), under the same assumption as in Theorem 1.2, the spherical integral asymptotically
factorizes into product of rank-one integrals, and therefore,
m
X
1
log Iζn (Gn , Dn ) =
n↑∞ n
i=1
lim
Z
0
θi
RD (
2ω
)dω,
ζ
(304)
with {θi } denoting the nonzero eigenvalues of Gn for i ∈ [1 : m], and m = rank(Gn ).
In Appendix A, one can employ (304) in order to evaluate the asymptotics over the system matrix consistent to
the system setup illustrated in Section II. Moreover, by using the above discussion, the investigations in Appendix
A can be extended to the case of complex variables. More about the spherical integral and its asymptotics can be
found in [75], and the references therein.
77
R EFERENCES
[1] S. F. Edwards and P. W. Anderson, “Theory of spin glasses,” Journal of Physics F: Metal Physics, vol. 5, no. 5, p. 965, 1975.
[2] L. Pastur and M. Shcherbina, “Absence of self-averaging of the order parameter in the Sherrington-Kirkpatrick model,” Journal of Statistical
Physics, vol. 62, no. 1-2, pp. 1–19, 1991.
[3] F. Guerra and F. L. Toninelli, “The thermodynamic limit in mean field spin glass models,” Communications in Mathematical Physics, vol.
230, no. 1, pp. 71–79, 2002.
[4] ——, “The infinite volume limit in generalized mean field disordered models,” arXiv preprint cond-mat/0208579, 2002.
[5] S. B. Korada and N. Macris, “Tight bounds on the capacity of binary input random CDMA systems,” Information Theory, IEEE Transactions
on, vol. 56, no. 11, pp. 5590–5613, 2010.
[6] N. Merhav, Statistical physics and information theory.
Now Publishers Inc, 2010.
[7] M. Mézard and G. Parisi, “A replica analysis of the travelling salesman problem,” Journal de Physique, vol. 47, no. 8, pp. 1285–1296,
1986.
[8] Y. Fu and P. W. Anderson, “Application of statistical mechanics to NP-complete problems in combinatorial optimisation,” Journal of
Physics A: Mathematical and General, vol. 19, no. 9, p. 1605, 1986.
[9] H. Nishimori, Statistical physics of spin glasses and information processing: an introduction.
Clarendon Press, 2001, no. 111.
[10] A. Montanari, “Turbo codes: The phase transition,” The European Physical Journal B-Condensed Matter and Complex Systems, vol. 18,
no. 1, pp. 121–136, 2000.
[11] T. Tanaka, “A statistical-mechanics approach to large-system analysis of CDMA multiuser detectors,” Information Theory, IEEE
Transactions on, vol. 48, no. 11, pp. 2888–2910, 2002.
[12] ——, “Average-case analysis of multiuser detectors,” 2001.
[13] D. Guo and S. Verdú, “Multiuser detection and statistical mechanics,” in Communications, Information and Network Security.
Springer,
2003, pp. 229–277.
[14] ——, “Randomly spread CDMA: Asymptotics via statistical physics,” Information Theory, IEEE Transactions on, vol. 51, no. 6, pp.
1983–2010, 2005.
[15] R. R. Müller, “Channel capacity and minimum probability of error in large dual antenna array systems with binary modulation,” IEEE
Transactions on Signal Processing, vol. 51, no. 11, pp. 2821–2828, 2003.
[16] R. Müller, D. Guo, and A. L. Moustakas, “Vector precoding for wireless MIMO systems and its replica analysis,” Selected Areas in
Communications, IEEE Journal on, vol. 26, no. 3, pp. 530–540, 2008.
[17] D. Guo, D. Baron, and S. Shamai, “A single-letter characterization of optimal noisy compressed sensing,” in Communication, Control, and
Computing, 2009. Allerton 2009. 47th Annual Allerton Conference on.
IEEE, 2009, pp. 52–59.
[18] J. Barbier and F. Krzakala, “Replica analysis and approximate message passing decoder for superposition codes,” in Information Theory
(ISIT), 2014 IEEE International Symposium on.
IEEE, 2014, pp. 1494–1498.
[19] A. Montanari and D. Tse, “Analysis of belief propagation for non-linear problems: The example of CDMA (or: How to prove Tanaka’s
formula),” in Information Theory Workshop, 2006. ITW’06 Punta del Este. IEEE. IEEE, 2006, pp. 160–164.
[20] W. Huleihel and N. Merhav, “Asymptotic MMSE analysis under sparse representation modeling,” Signal Processing, vol. 131, pp. 320–332,
2017.
[21] G. Reeves and H. D. Pfister, “The replica-symmetric prediction for compressed sensing with Gaussian matrices is exact,” in Information
Theory (ISIT), 2016 IEEE International Symposium on. IEEE, 2016, pp. 665–669.
[22] J. Barbier, M. Dia, N. Macris, and F. Krzakala, “The mutual information in random linear estimation,” in Communication, Control, and
Computing (Allerton), 2016 54th Annual Allerton Conference on. IEEE, 2016, pp. 625–632.
[23] J. Barbier, N. Macris, M. Dia, and F. Krzakala, “Mutual information and optimality of approximate message-passing in random linear
estimation,” arXiv preprint arXiv:1701.05823, 2017.
[24] J. Barbier, M. Dia, and N. Macris, “Threshold saturation of spatially coupled sparse superposition codes for all memoryless channels,” in
Information Theory Workshop (ITW), 2016 IEEE. IEEE, 2016, pp. 76–80.
[25] ——, “Universal sparse superposition codes with spatial coupling and GAMP decoding,” arXiv preprint arXiv:1707.04203, 2017.
[26] D. Guo, S. Verdú, and L. K. Rasmussen, “Asymptotic normality of linear multiuser receiver outputs,” Information Theory, IEEE Transactions
on, vol. 48, no. 12, pp. 3080–3095, 2002.
78
[27] D. N. Tse and S. V. Hanly, “Linear multiuser receivers: Effective interference, effective bandwidth and user capacity,” Information Theory,
IEEE Transactions on, vol. 45, no. 2, pp. 641–657, 1999.
[28] D. N. Tse and O. Zeitouni, “Linear multiuser receivers in random environments,” Information Theory, IEEE Transactions on, vol. 46,
no. 1, pp. 171–188, 2000.
[29] Y. C. Eldar and A. M. Chan, “On the asymptotic performance of the decorrelator,” Information Theory, IEEE Transactions on, vol. 49,
no. 9, pp. 2309–2313, 2003.
[30] S. Verdú and S. Shamai, “Spectral efficiency of CDMA with random spreading,” Information Theory, IEEE Transactions on, vol. 45, no. 2,
pp. 622–640, 1999.
[31] J. Zhang, E. K. Chong, and D. N. Tse, “Output MAI distributions of linear MMSE multiuser receivers in DC-CDMA systems,” Information
Theory, IEEE Transactions on, vol. 47, no. 3, pp. 1128–1144, 2001.
[32] A. M. Tulino and S. Verdú, “Random matrix theory and wireless communications,” Communications and Information theory, vol. 1, no. 1,
pp. 1–182, 2004.
[33] R. R. Müller, G. Alfano, B. M. Zaidel, and R. de Miguel, “Applications of large random matrices in communications engineering,” arXiv
preprint arXiv:1310.5479, 2013.
[34] D. Guo, L. K. Rasmussen, and T. J. Lim, “Linear parallel interference cancellation in long-code CDMA multiuser detection,” Selected
Areas in Communications, IEEE Journal on, vol. 17, no. 12, pp. 2074–2081, 1999.
[35] S. Shamai and S. Verdú, “The impact of frequency-flat fading on the spectral efficiency of CDMA,” Information Theory, IEEE Transactions
on, vol. 47, no. 4, pp. 1302–1327, 2001.
[36] R. R. Müller and W. H. Gerstacker, “On the capacity loss due to separation of detection and decoding,” Information Theory, IEEE
Transactions on, vol. 50, no. 8, pp. 1769–1778, 2004.
[37] R. R. Müller, “On channel capacity, uncoded error probability, ML-detection, and spin glasses,” in Proc. Workshop on Concepts in
Information Theory.
Citeseer, 2002, pp. 79–81.
[38] S. Rangan, A. K. Fletcher, and V. Goyal, “Asymptotic analysis of MAP estimation via the replica method and applications to compressed
sensing,” in IEEE Transactions on Information Theory, 2012, pp. 1902–1923.
[39] A. M. Tulino, G. Caire, S. Verdu, and S. Shamai, “Support recovery with sparsely sampled free random matrices,” Information Theory,
IEEE Transactions on, vol. 59, no. 7, pp. 4243–4271, 2013.
[40] M. Vehkaperä, Y. Kabashima, and S. Chatterjee, “Analysis of regularized LS reconstruction and random matrix ensembles in compressed
sensing,” IEEE Transactions on Information Theory, vol. 62, no. 4, pp. 2100–2124, 2016.
[41] M. Yoshida, T. Uezu, T. Tanaka, and M. Okada, “Statistical mechanical study of code-division multiple-access multiuser detectorsanalysis
of replica symmetric and one-step replica symmetry breaking solutions,” Journal of the Physical Society of Japan, vol. 76, no. 5, pp.
054 003–054 003, 2007.
[42] Y. Kabashima, T. Wadayama, and T. Tanaka, “A typical reconstruction limit for compressed sensing based on ℓp -norm minimization,”
Journal of Statistical Mechanics: Theory and Experiment, vol. 2009, no. 09, p. L09003, 2009.
[43] B. M. Zaidel, R. Müller, A. L. Moustakas, and R. de Miguel, “Vector precoding for Gaussian MIMO broadcast channels: Impact of replica
symmetry breaking,” Information Theory, IEEE Transactions on, vol. 58, no. 3, pp. 1413–1440, 2012.
[44] M. A. Sedaghat, A. Bereyhi, and R. Mueller, “LSE precoders for massive MIMO with hardware constraints: Fundamental limits,” arXiv
preprint arXiv:1612.07902, 2016.
[45] A. Bereyhi, M. A. Sedaghat, and R. R. Müller, “Asymptotics of nonlinear LSE precoders with applications to transmit antenna selection,”
in Information Theory (ISIT), 2017 IEEE International Symposium on.
IEEE, 2017, pp. 81–85.
[46] K. Takeda and Y. Kabashima, “Statistical mechanical assessment of a reconstruction limit of compressed sensing: Toward theoretical
analysis of correlated signals,” EPL (Europhysics Letters), vol. 95, no. 1, p. 18006, 2011.
[47] D. L. Donoho, “Compressed sensing,” Information Theory, IEEE Transactions on, vol. 52, no. 4, pp. 1289–1306, 2006.
[48] E. J. Candès, J. Romberg, and T. Tao, “Robust uncertainty principles: Exact signal reconstruction from highly incomplete frequency
information,” Information Theory, IEEE Transactions on, vol. 52, no. 2, pp. 489–509, 2006.
[49] E. J. Candes and T. Tao, “Near-optimal signal recovery from random projections: Universal encoding strategies?” Information Theory,
IEEE Transactions on, vol. 52, no. 12, pp. 5406–5425, 2006.
[50] D. L. Donoho and X. Huo, “Uncertainty principles and ideal atomic decomposition,” Information Theory, IEEE Transactions on, vol. 47,
no. 7, pp. 2845–2862, 2001.
79
[51] M. Elad and A. M. Bruckstein, “A generalized uncertainty principle and sparse representation in pairs of bases,” Information Theory, IEEE
Transactions on, vol. 48, no. 9, pp. 2558–2567, 2002.
[52] D. L. Donoho and M. Elad, “Optimally sparse representation in general (nonorthogonal) dictionaries via ℓ1 minimization,” Proceedings
of the National Academy of Sciences, vol. 100, no. 5, pp. 2197–2202, 2003.
[53] R. Tibshirani, “Regression shrinkage and selection via the LASSO,” Journal of the Royal Statistical Society. Series B (Methodological),
pp. 267–288, 1996.
[54] S. S. Chen, D. L. Donoho, and M. A. Saunders, “Atomic decomposition by basis pursuit,” SIAM review, vol. 43, no. 1, pp. 129–159, 2001.
[55] D. Needell and J. A. Tropp, “Cosamp: Iterative signal recovery from incomplete and inaccurate samples,” Applied and Computational
Harmonic Analysis, vol. 26, no. 3, pp. 301–321, 2009.
[56] W. Dai and O. Milenkovic, “Subspace pursuit for compressive sensing signal reconstruction,” Information Theory, IEEE Transactions on,
vol. 55, no. 5, pp. 2230–2249, 2009.
[57] T. T. Cai and L. Wang, “Orthogonal matching pursuit for sparse signal recovery with noise,” Information Theory, IEEE Transactions on,
vol. 57, no. 7, pp. 4680–4688, 2011.
[58] S. Foucart and H. Rauhut, A mathematical introduction to compressive sensing.
Birkhäuser Basel, 2013, vol. 1, no. 3.
[59] Y. Kabashima, T. Wadayama, and T. Tanaka, “Statistical mechanical analysis of a typical reconstruction limit of compressed sensing,”
arXiv preprint arXiv:1001.4298, 2010.
[60] C.-K. Wen, J. Zhang, K.-K. Wong, J.-C. Chen, and C. Yuen, “On sparse vector recovery performance in structurally orthogonal matrices
via LASSO,” IEEE Transactions on Signal Processing, vol. 64, no. 17, pp. 4519–4533, 2016.
[61] L. Zheng, A. Maleki, H. Weng, X. Wang, and T. Long, “Does ℓp -minimization outperform ℓ1 -minimization?” IEEE Transactions on
Information Theory, 2017.
[62] D. L. Donoho, A. Maleki, and A. Montanari, “Message-passing algorithms for compressed sensing,” Proceedings of the National Academy
of Sciences, vol. 106, no. 45, pp. 18 914–18 919, 2009.
[63] ——, “Message passing algorithms for compressed sensing: I. Motivation and construction,” in Information Theory (ITW 2010, Cairo),
2010 IEEE Information Theory Workshop on. IEEE, 2010, pp. 1–5.
[64] S. Rangan, “Generalized approximate message passing for estimation with random linear mixing,” in Information Theory Proceedings
(ISIT), 2011 IEEE International Symposium on.
IEEE, 2011, pp. 2168–2172.
[65] A. Montanari, “Graphical models concepts in compressed sensing,” Compressed Sensing: Theory and Applications, pp. 394–438, 2012.
[66] M. Bayati and A. Montanari, “The dynamics of message passing on dense graphs, with applications to compressed sensing,” IEEE
Transactions on Information Theory, vol. 57, no. 2, pp. 764–785, 2011.
[67] E. Bolthausen, “On the high-temperature phase of the Sherrington-Kirkpatrick model,” in Seminar at EURANDOM, Eindhoven, 2009.
[68] S. Rangan, P. Schniter, and A. K. Fletcher, “Vector approximate message passing,” in Information Theory (ISIT), 2017 IEEE International
Symposium on. IEEE, 2017, pp. 1588–1592.
[69] K. Takeuchi, “Rigorous dynamics of expectation-propagation-based signal recovery from unitarily invariant measurements,” arXiv preprint
arXiv:1701.05284, 2017.
[70] D. L. Donoho, A. Javanmard, and A. Montanari, “Information-theoretically optimal compressed sensing via spatial coupling and
approximate message passing,” IEEE transactions on information theory, vol. 59, no. 11, pp. 7434–7464, 2013.
[71] F. Krzakala, M. Mézard, F. Sausset, Y. Sun, and L. Zdeborová, “Statistical-physics-based reconstruction in compressed sensing,” Physical
Review X, vol. 2, no. 2, p. 021005, 2012.
[72] Y. Wu and S. Verdú, “Optimal phase transitions in compressed sensing,” Information Theory, IEEE Transactions on, vol. 58, no. 10, pp.
6241–6263, 2012.
[73] ——, “MMSE dimension,” IEEE Transactions on Information Theory, vol. 57, no. 8, pp. 4857–4879, 2011.
[74] A. Guionnet and O. Zeitouni, “Large deviations asymptotics for spherical integrals,” Journal of functional analysis, vol. 188, no. 2, pp.
461–515, 2002.
[75] A. Guionnet, M. Maı et al., “A fourier view on the R-transform and related asymptotics of spherical integrals,” Journal of functional
analysis, vol. 222, no. 2, pp. 435–490, 2005.
[76] N. I. Akhiezer, The classical moment problem: and some related questions in analysis.
Oliver & Boyd, 1965, vol. 5.
[77] B. Simon, “The classical moment problem as a self-adjoint finite difference operator,” Advances in Mathematics, vol. 137, no. 1, pp.
82–203, 1998.
80
[78] A. Bereyhi, R. Müller, and H. Schulz-Baldes, “RSB decoupling property of MAP estimators,” in Information Theory Workshop (ITW),
2016 IEEE. IEEE, 2016, pp. 379–383.
[79] A. Bereyhi, R. R. Müller, and H. Schulz-Baldes, “Replica symmetry breaking in compressive sensing,” in Information Theory and
Applications Workshop (ITA), 2017 IEEE. IEEE, 2017, pp. 1–7.
[80] F. Riesz, “Sur les valeurs moyennes des fonctions,” Journal of the London Mathematical Society, vol. 1, no. 2, pp. 120–121, 1930.
[81] A. Dembo and O. Zeitouni, Large deviations techniques and applications.
[82] M. Mezard and A. Montanari, Information, physics, and computation.
Springer Science & Business Media, 2009, vol. 38.
Oxford University Press, 2009.
[83] G. Parisi, “A sequence of approximated solutions to the SK model for spin glasses,” Journal of Physics A: Mathematical and General,
vol. 13, no. 4, p. L115, 1980.
[84] M. N. Hansen and E. M. Schmidt, Algorithms and data structures: transition systems.
Datalogisk Institut, Aarhus Universitet, 2003.
[85] A. Finkel and P. Schnoebelen, “Fundamental structures in well-structured infinite transition systems,” in LATIN’98: Theoretical Informatics.
Springer, 1998, pp. 102–118.
[86] R. Couillet and M. Debbah, Random matrix methods for wireless communications.
Cambridge University Press, 2011.
[87] A. L. Moustakas and S. H. Simon, “On the outage capacity of correlated multiple-path MIMO channels,” Information Theory, IEEE
Transactions on, vol. 53, no. 11, pp. 3887–3903, 2007.
[88] D. Agrawal and A. Vardy, “The turbo decoding algorithm and its phase trajectories,” IEEE Transactions on Information Theory, vol. 47,
no. 2, pp. 699–722, 2001.
[89] M. Talagrand, Spin glasses: a challenge for mathematicians: cavity and mean field models.
Springer Science & Business Media, 2003,
vol. 46.
[90] F. Guerra and F. L. Toninelli, “Quadratic replica coupling in the Sherrington–Kirkpatrick mean field spin glass model,” Journal of
Mathematical Physics, vol. 43, no. 7, pp. 3704–3716, 2002.
[91] F. Guerra, “Broken replica symmetry bounds in the mean field spin glass model,” Communications in mathematical physics, vol. 233,
no. 1, pp. 1–12, 2003.
[92] H. Chandra, “Differential operators on a semisimple Lie algebra,” American Journal of Mathematics, pp. 87–120, 1957.
[93] C. Itzykson and J.-B. Zuber, “The planar approximation. II,” Journal of Mathematical Physics, vol. 21, no. 3, pp. 411–421, 1980.
| 7cs.IT
|
Asking the Difficult Questions: Goal-Oriented Visual Question Generation
via Intermediate Rewards
Junjie Zhang∗2,3 , Qi Wu†1 , Chunhua Shen1 , Jian Zhang2 , Jianfeng Lu3 , and Anton van den Hengel1
arXiv:1711.07614v1 [cs.CV] 21 Nov 2017
1
Australian Centre for Robotic Vision, The University of Adelaide, Australia
Faculty of Engineering and Information Technology, University of Technology Sydney, Australia
3
School of Computer Science and Technology, Nanjing University of Science and Technology, China
2
Abstract
Hi, Robby, can you get my cup
from the cupboard?
Despite significant progress in a variety of vision-andlanguage problems, developing a method capable of asking
intelligent, goal-oriented questions about images is proven
to be an inscrutable challenge. Towards this end, we
propose a Deep Reinforcement Learning framework based
on three new intermediate rewards, namely goal-achieved,
progressive and informativeness that encourage the generation of succinct questions, which in turn uncover valuable
information towards the overall goal. By directly optimizing
for questions that work quickly towards fulfilling the overall
goal, we avoid the tendency of existing methods to generate
long series of insane queries that add little value. We evaluate our model on the GuessWhat?! dataset and show that
the resulting questions can help a standard Guesser identify
a specific object in an image at a much higher success rate.
Is it tall one?
No
Is it from IKEA?
Does it have Elsa on it?
Is it pink?
No
Yes
Does it have a handle?
I’ll get it myself ….
Brilliant!
Figure 1: Two illustrative examples of potential conversations between a
human and a robot. The left conversation clearly makes people frustrated
while the right one makes people happy because the robot achieves the
goal in a quicker way via less but informative questions.
A well-posed question extracts the most informative answer towards achieving a particular goal, and thus reflects
the knowledge of the asker, and their estimate of the capabilities of the answerer. Although the information would
be beneficial in identifying a particular object in an image,
there is little value in an agent asking a human about the
exact values of particular pixels, the statistics of their gradients, or the aspect ratio of the corresponding bounding box.
The fact that the answerer is incapable of providing the requested information makes such questions pointless. Selecting a question that has a significant probability of generating an answer that helps achieve a particular goal is a
complex problem.
Asking questions is an essential part of the way humans
communicate and learn. Any intelligent agent that seeks to
interact flexibly and effectively with humans thus needs to
be able to ask questions. The ability to ask intelligent questions is even more important than receiving intelligent, ac-
Judge a man by his questions rather than by his answers.
-Voltaire
Although Visual Question Answering (VQA) [3, 30, 31]
has attracted more attention, Visual Question Generation
(VQG) is a much more difficult task. Obviously, generating facile, repetitive questions represents no challenge at all,
but generating a series of questions that draw out useful information towards an overarching goal, however, demands
consideration of the image content, the goal, and the conversation thus far. It could, generally, also be seen as requiring consideration of the abilities and motivation of the
other participant in the conversation.
† The
Yes
No
1. Introduction
∗ The
Is it pink?
Is it short one?
Yes
work was done while visiting The University of Adelaide.
first two authors contributed to this work equally.
1
tionable answers. A robot, for example in Fig. 1, has been
given a task and realized that it is missing critical information required to carry it out, needs to ask a question. It will
have a limited number of attempts before the human gets
frustrated and carries out the task themselves. This scenario
applies equally to any intelligent agent that seeks to interact with humans, as we have surprisingly little tolerance for
agents that are unable to learn by asking questions, and for
those that ask too many.
As a result of the above, Visual Question Generation
(VQG) has started to receive research attention, but primarily as a vision-to-language problem [16, 20, 32]. Methods
that approach the problem in this manner tend to generate
arbitrary sequences of questions that are somewhat related
to the image [21], but which bare no relationship to the goal.
This reflects the fact that these methods have no means of
measuring whether the answers generated to assist in making progress towards the goal. Instead, in this paper, we
ground the VQG problem as a goal-oriented version of the
game - GuessWhat?!, introduced in [10]. The method presented in [10] to play the GuessWhat game is made up of
three components: the Questioner asks questions to the
Oracle, and the Guesser tries to identify the object that
the Oracle is referring to, based on its answers. The quality of the generated questions is thus directly related to the
success rate of the final task.
Goal-oriented training that uses a game setting has been
used in visual dialog generation previously [7, 8]. However, these work focus on generating more human-like dialogs, not on helping the agent achieve the goal through
better question generation. Moreover, previous work [26]
only uses the final goal as the reward to train the dialog generator, which might be suitable for dialog generation but is
a rather weak and undirected signal by which to control the
quality, effectiveness, and informativeness of the generated
question in a goal-oriented task. In other words, in some
cases, we want to talk to a robot because we want it to finish a specific task but not to hold the meaningless boring
chat. Therefore, in this paper, we use intermediate rewards
to encourage the agent to ask short but informative questions to achieve the goal. Moreover, in contrast to previous
works that only consider the overall goal as the reward, we
assign different intermediate rewards for each posed question to control the quality.
This is achieved through fitting the goal-oriented VQG
into a reinforcement learning (RL) paradigm and devising
three different intermediate rewards, which are our main
contributions in this paper, to explicitly optimize the question generation. The first goal-achieved reward is designed
to encourage the agent to achieve the final goal (pick out
the object that the Oracle is ‘thinking’) via asking multiple questions. However, different from only considering whether the goal is achieved, additional rewards are
awarded if the agent can use fewer questions to achieve it.
This is a reasonable setting because you do not need a robot
that can finish a task but has to ask you hundreds of questions. The second reward we proposed is the progressive
reward, which is established to encourage questions that
generated by the agent can progressively increase the probability of the right answer. This is an intermediate reward
for the individual question, and the reward is decided by
the change of the ground-truth answer probability. A negative reward will be given if the probability decreases. The
last reward is the informativeness reward, which is used to
restrict the agent not to ask ‘useless’ questions, for example, a question that leads to the identical answer for all the
candidate objects (this question cannot eliminate any ambiguous). We show the whole framework in Fig. 2.
We evaluate our model on the GuessWhat?! dataset [10],
with the pre-trained standard Oracle and Guesser, we
show that our novel Questioner model outperforms the
baseline and state-of-the-art model by a large margin. We
also evaluate each reward respectively, to measure the individual contribution. Qualitative results show that we can
produce more informative questions.
2. Related Works
Visual Question Generation Recently, the visual question generation problem has been brought to the computer
vision community, aims at generating visual-related questions. Most of the works treat the VQG as a standalone
problem and follow an image captioning style framework,
i.e., translate an image to a sentence, in this case, a question. For example, in [20], Mora et al. use a CNN-LSTM
model to generate questions and answers directly from the
image visual content. Zhang et al. [32] focus on generating questions of grounded images. They use Densecap [13]
as region captioning generator to guide the question generation. In [21], Mostafazadeh et al. propose a dataset to
generate natural questions about images, which are beyond
the literal description of image content. Li et al. [16] view
the VQA and VQG as a dual learning process by jointly
training them in an end-to-end framework. Although these
works can generate meaningful questions that are related
to the image, the motivation of asking these questions are
rather weak because they are not related to any goals. Moreover, it is hard to conduct the quality measurement on this
type of questions. Instead, in our work, we aim to develop
an agent that can learn to ask realistic questions, which can
contribute to achieving a specific goal.
Goal-oriented Visual Dialogue generation has attracted
many attentions at most recently. In [8], Das et al. introduce
a reinforcement learning mechanism for visual dialogue
generation. They establish two RL agents corresponding to
question and answer generation respectively, to finally locate an unseen image from a set of images. The question
Rounds of Dialogue
…
…
Oracle
𝑞1 : Is it a person?
< S𝑡𝑎𝑟𝑡 >
CNN
Image Feature
Guesser
𝑎1𝑜1 : Yes
…
𝑎1𝑜∗ : No
……
…
Oracle
𝑎𝑗𝑜1 : No
…
𝑎𝑗𝑜∗ : No
𝑞𝑗 : Is it a funiture?
(𝑞1:𝑗−1 , 𝑎1:𝑗−1,𝑜∗ )
[𝑝1𝑜1 , … , 𝑝1𝑜∗ ]
Guesser
Oracle
……
𝑞𝐽 : Is it a drink?
(𝑞1:𝐽 , 𝑎1:𝐽,𝑜∗ )
[𝑝𝑗𝑜1 , … , 𝑝𝑗𝑜∗ ]
Image Feature
VQG
𝒓𝒊𝒏𝒇𝒐𝒓𝒎𝒂𝒕𝒊𝒗𝒆𝒏𝒆𝒔𝒔
Question
Generator
Guesser
[𝑝𝐽𝑜1 , … , 𝑝𝐽𝑜∗ ]
𝒓𝒑𝒓𝒐𝒈𝒓𝒆𝒔𝒔𝒊𝒗𝒆
VQG
𝑎2𝑜1 : No
…
𝑎𝐽𝑜∗ : Yes
Intermediate
Rewards
𝒓𝒈𝒐𝒂𝒍−𝒂𝒄𝒉𝒊𝒆𝒗𝒆𝒅
Success
VQG
Figure 2: The framework of proposed VQG agent plays in the whole game environment. A target object o∗ is assigned to the Oracle, but it is unknown to
VQG and Guesser. Then VQG generates a series of questions, which are answered by Oracle. During training, we let Oracle answer the question based on
all the objects at each round, and measure the informativeness reward, and we also let Guesser generate probability distribution to measure the progressive
reward. Finally, we consider the number of rounds J and set the goal-achieved reward based on the status of success. These intermediate rewards are
adopted for optimizing the VQG agent by the REINFORCE.
agent predicts the feature representation of the image and
the reward function is given by measuring how close the
representation is compared to the true feature. However,
we focus on encouraging the agent to generate questions
that directed towards the final goal, and we adopt different
kinds of intermediate rewards to achieve that in the question generation process. Moreover, the question generation
agent in their model only asks questions based on the dialogue history, which does not involve visual information. In
[26], Florian et al. propose to employ reinforcement learning to solve question generation of the GuessWhat game
by introducing the final status of success as the sole reward. We share the similar backbone idea, but there are
several technical differences. One of the most significant
differences is that the previous work only considers using
whether achieving the final goal as the reward but we assign different intermediate rewards for each posed question
to push VQG agent to ask short but informative questions
to achieve the goal. The experimental results and analysis
in Section 4 show that our model not only outperforms the
state-of-art but also achieves higher intelligence, i.e., using
as few questions as possible to finish the task.
Question Generation in NLP There is a long history
of works on grammar question generation from text domain in natural language processing (NLP) [6, 9, 25, 29].
In [1, 4], authors focus on automatically generating gapfill questions, while crowdsourcing templates and manually
built templates are used for question generation in [15] and
[19] respectively. These works focus on constructing formatted questions from the text corpus.
Reinforcement Learning for V2L Reinforcement learning [14, 27] has been adopted in several vision to language
(V2L) problems, including image captioning [17, 23, 24],
VQA [2, 12, 33], and aforementioned visual dialogue system [8, 18] etc. In [23], Ren et al. use a policy network
and a value network to collaboratively generate image captions, while different optimization methods for RL in image captioning are explored in [17] and [24], called SPIDEr
and self-critical sequence training. Zhu et al. [33] introduce
knowledge source into the iterative VQA and employ RL
to learn the query policy. In [2], authors use RL to learn
the parameters of QA model for both images and structured
knowledge bases. These works solve V2L related problems
by employing RL as an optimization method, while we focus on using RL with carefully designed intermediate rewards to train the VQG agent for goal-oriented tasks.
3. Goal-Oriented VQG
We ground our goal-oriented VQG problem on a Guess
What game, specifically, on the GuessWhat?! dataset [10].
GuessWhat?! is a three-role interactive game, where all
roles observe the same image of a rich visual scene that contains multiple objects. We view this game as three parts:
Oracle, Questioner and Guesser. In each game,
a random object in the scene is assigned to the Oracle,
where this process is hidden to the Questioner. Then
the Questioner can ask a series of yes/no questions to
locate this object. The list of objects is also hidden to the
Questioner during the question-answer rounds. Once
the Questioner has gathered enough information, the
Guesser can start to guess. The game is considered as
successful if the Guesser selects the right object.
The Questioner part of the game is a goal-oriented
VQG problem; each question is generated based on the visual information of the image and the previous rounds of
question-answer pairs. The goal of VQG is to successfully
finish the game, in this case, to locate the right object. In this
paper, we fit the goal-oriented VQG into a reinforcement
learning paradigm and propose three different intermediate rewards, namely the goal-achieved reward, progressive
reward, and informativeness reward, to explicitly optimize
the question generation. The goal-achieved reward is established to lead the dialogue to achieve the final goal, the
progressive reward is used to push the intermediate generation process towards the optimal direction, while the informativeness reward is used to ensure the quality of generated questions. To better express the generation process, we
first introduce the notations of GuessWhat?! game, which
is used throughout the rest of sections.
Each game is defined as a tuple (I, D, O, o∗ ), where I
is the observed image, D is the dialogue with J rounds
of question-answer pairs (qj ,aj )Jj=1 , O = (on )N
n=1 is the
list of N objects in the image I, where o∗ is the target obj Mj
ject. Each question qj = (wm
)m=1 is a sequence of Mj
tokens, which are sampled from the pre-defined vocabulary
V . The V is composed of word tokens, a question stop
token <?> and a dialogue stop token <End>. The answer aj ∈ {<Yes>,<No>,<NA>} is set to be yes, no or
not applicable. For each object o, it has an object category
co ∈ {1 . . . C} and a segment mask.
3.1. Learning Environment
We build the learning environment to generate visual dialogues based on the GuessWhat?! dataset. Since we focus on the goal-oriented VQG, for a fair comparison, the
Oracle and Guesser are produced by referring to the
original baseline models in GuessWhat?! [10]. We also
introduce the VQG supervised learning model, which is referred as the baseline for the rest of the paper.
Oracle The Oracle requires generating answers for all
kinds of questions about any objects within the image scene.
We build the neural network architecture for Oracle by
referring to [10]. The bounding box (obtained from the
segment mask) of the object o are encoded into an eight
dimensional vector to represent the spatial feature, where
ospa = [xmin ,ymin ,xmax ,ymax ,xcenter ,ycenter ,wbox ,hbox ]
indicates the box coordinates, width and height. The category co is embedded using a learned look-up table, while
the current question is encoded by an LSTM [11]. All three
features are concatenated into a single vector and fed into a
one hidden layer MLP followed by a softmax layer to produce the answer probability p(a|ospa ,co ,q).
Guesser Given an image I and a series of questionanswer pairs, the Guesser requires predicting right object
o∗ from a list of objects. By referring to [10], we consider
the generated dialogue as one flat sequence of tokens and
encode it with an LSTM. The last hidden state is extracted
as the feature to represent the dialogue. We also embed all
the objects’ spatial features and categories by an MLP. We
perform a dot-product between dialogue and object features
with a softmax operation to produce the final prediction.
VQG Baseline Given an image I and a history of the
question-answer pairs (q,a)1:j−1 , the VQG requires generating a new question qj . We build the VQG baseline
based on an RNN generator. The RNN recurrently produces a series of state vectors sj1:m by transitioning from
j
.
the previous state sjm−1 and the current input token wm
We use an LSTM as the transition function f , that is,
j
sjm = f (sjm−1 ,wm
). In our case, the state vector s is conditioned on the whole image and all the previous questionanswer tokens. We add a softmax operation to produce
the probability distribution over the vocabulary V , where
j
j
). This baseline is conducted by
p(wm
|I,(q,a)1:j−1 ,w1:m−1
employing the supervised training. We train the VQG by
minimizing the following negative log loss function:
L = − log p(q1:J |I,a1:J )
=−
M
J X
X
j
j
log p(wm
|I,w1:m−1
,(q,a)1:j−1 )
(1)
j=1 m=1
During the test stage, the question can be sampled from the
j
model by starting from state sj1 ; a new token wm
is sampled
from the probability distribution, then embedded and fed
back to the LSTM. We repeat this operation until the end of
question token is encountered.
3.2. Reinforcement Learning of VQG
We use our established Oracle, Guesser and VQG
baseline model to simulate a complete GuessWhat?! game.
Given an image I, an initial question q1 is generated by
sampling from the VQG baseline until the stop question token is encountered. Then the Oracle receives the question
q1 along with the assigned object category o∗ and its spatial
information o∗spa , and output the answer a1 , the questionanswer pair (q1 ,a1 ) is appended to the dialogue history. We
repeat this loop until the end of dialogue token is sampled,
or the number of questions reaches the maximum. Finally,
the Guesser takes the whole dialogue D and the object
list O as inputs to predict the object. We consider the goal
reached if o∗ is selected. Otherwise, it failed.
To more efficiently optimize the VQG towards the final
goal and generate informative questions, we adopt three intermediate rewards (which will be introduced in the following sections) into the RL framework.
3.2.1
State, Action & Policy
We view the VQG as a Markov Decision Process (MDP),
the VQG is noted as the agent. For the dialogue generated based on the image I at time step t, the state of agent
is defined as the image visual content with the history of
question-answer pairs and the tokens of current question
j
generated so far: St = (I,(q,a)1:j−1 ,(w1j , . . . ,wm
)), where
Pk=j−1
t = k=1 Mk + m. The action At of agent is to select
j
the next output token wm+1
from the vocabulary V . Depends on the actions that agent takes, the transition between
two states falls into one of the following cases:
j
=<?>: The current question is finished, the
1) wm+1
Oracle from the environment will answer aj , which is
appended to the dialogue history. The next state St+1 =
(I,(q,a)1:j ).
j
2) wm+1
=<End>: The dialogue is finished, the
Guesser from the environment will select the object from
the list O.
j
keeps ap3) Otherwise, the new generated token wm+1
pending to the current question qj , the next state St+1 =
j
j
)).
,wm+1
(I,(q,a)1:j−1 ,(w1j , . . . ,wm
The maximum length of question qj is Mmax , and the
maximum rounds of the dialogue is Jmax . Therefore, the
number of time steps T of any dialogue are T ≤ Mmax ∗
Jmax . We model the VQG under the stochastic policy
πθ (A|S), where θ represents the parameters of the deep
neural network we used in the VQG baseline that produces
the probability distributions for each state. The goal of the
policy learning is to estimate the parameter θ.
After we set up the components of MDP, the most significant aspect of the RL is to define the appropriate reward
function for each state-action pair (St ,At ). As we emphasized before, the goal-oriented VQG aims to generate the
questions that lead to achieving the final goal. Therefore,
we build three kinds of intermediate rewards to push the
VQG agent to be optimized towards the optimal direction.
The whole framework is shown in Fig. 2.
3.2.2 Goal-Achieved Reward
One basic rule of the appropriate reward function is that
it cannot conflict with the final optimal policy [22]. The
primary purpose of the VQG agent is to gather enough information as soon as possible to help Guesser to locate
the object. Therefore, we define the first reward to reflect
whether the final goal is achieved. Moreover, we take the
number of rounds into consideration to accelerate the questioning part and let the reward nonzero when the game is
successful.
Given the state St , where the <End> token is sampled
or the maximum number of rounds Jmax is reached, the
reward of the state-action pair is defined as:
(
1+λ · Jmax /J, If Guesser(St ) = o∗
rg (St ,At ) =
0,
Otherwise
(2)
We set the reward as one plus the weighted maximum number of rounds Jmax against the actual rounds J of the current dialogue if the dialogue is successful, and zero other-
wise. This is based on that we want the final goal to motivate the VQG to generate useful questions. Moreover, the
intermediate process is considered into the reward function
as the rounds of the question-answer pairs J, which guarantees the efficiency of the generation process; the fewer
questions are generated, the more reward VQG agent can
get at the end of the game (if and only if the game succeed).
This is a quite useful setting in the realistic because we do
want to use fewer orders to guide the robot to finish more
tasks. λ is a weight to balance between the contribution of
the successful reward and the dialogue round reward.
3.2.3 Progressive Reward
Based on the intuition and the observation of the human interactive dialogues, we find that the questions of a successful game, are ones that progressively achieve the final goal,
i.e., as long as the questions being asked and answered, the
confidence of referring to the target object becomes higher
and higher. Therefore, at each round, we define an intermediate reward for state-action pair as the improvement of
target probability that Guesser outputs. More specific,
we interact with the Guesser at each round to obtain the
probability of predicting target object. If the probability increases, it means that the generated question qj is a positive
question that leads the dialogue towards the right direction.
We set an intermediate reward called progressive reward
to encourage VQG agent progressively generate these positive questions. At each round j, we record the probability
pj (o∗ |I,(q,a)1:j ) returned by Guesser, and compare it with
the last round j − 1. The difference between two probabilities is used as the intermediate reward. That is:
rp (St ,At ) = pj (o∗ |I,(q,a)1:j ) − pj−1 (o∗ |I,(q,a)1:j−1 )
(3)
In this way, the question is considered high-quality and has
a positive reward, if it leads to a higher probability to guess
the right object. Otherwise, the reward is negative.
3.2.4
Informativeness Reward
When we human ask questions (especially in a guess what
game), we expect an answer that can help us to eliminate
the confusion and distinguish the candidate objects. Hence,
imagine that if a posed question that leads to the same answer for all the candidate object, this question will be useless. For example, all the candidate objects are ‘red’ and if
we posed a question that ‘Is it red?’, we will get the answer
‘Yes.’ However, this question-answer pair cannot help us
to identify the target. We want to avoid this kind of questions because they are non-informative. In this case, we
need to evaluate the question based on the answer from the
Oracle.
Given generated question qj , we interact with the
Oracle to answer the question. Since the Oracle takes
Algorithm 1 Training procedure of the VQG agent.
the image I, the current question qj , and the target object
o∗ as inputs, and outputs the answer aj , we let the Oracle
answer question qj for all objects in the image. If the answers are different from each other, we consider qj is useful
for locating the right object. Otherwise, it does not contribute to the final goal. Therefore, we set the reward positive, which we called informativeness reward, for these useful questions.
Formally, during each round, the Oracle receives the
image I, the current question qj and the list of objects O,
and then outputs the answer set ajO = {ajo1 , . . . ,ajoN },
where each element corresponds to each object. Then the
informativeness reward is defined as:
(
η, If all ajon are not identical
ri (St ,At ) =
(4)
0,
Otherwise
Input: Oracle(Ora), Guesser(Gus), V QG, batch size H
1: for Each update do
2:
# Generate episodes τ
3:
for h = 1 to H do
4:
select image Ih and one target object o∗h ∈ Oh
5:
# Generate question-answer pairs (q,a)h1:j
6:
for j = 1 to Jmax do
7:
qjh = V QG(Ih ,(q,a)h1:j−1 )
8:
# N is the number of total objects
9:
for n = 1 to N do
10:
ahjohn = Ora(Ih ,qjh ,ohn )
11:
12:
13:
By giving a positive reward to the state-action pair, we improve the quality of the dialogue by encouraging agent to
generate more informative questions.
14:
15:
16:
17:
18:
3.2.5
19:
20:
Training with Policy Gradient
21:
22:
23:
24:
Now we have three different kinds of rewards that take
the intermediate process into consideration, for each stateaction pair (St ,At ), we add three rewards together as the
final reward function:
r(St ,At ) = rg (St ,At ) + rp (St ,At ) + ri (St ,At )
(5)
25:
26:
27:
28:
Considering the large action space in the game setting,
we adopt the policy gradient method [28] to train the VQG
agent with proposed intermediate rewards. The goal of policy gradient is to update policy parameters with respect to
the expected return by gradient descent. Since we are in the
episodic environment, given policy πθ , which is the generative network of the VQG agent, in this case, the policy
objective function takes the form:
T
X
J(θ) = Eπθ [
r(St ,At )]
(6)
if all ahjohn are not identical then
ri (St ,At ) = η
else ri (St ,At ) = 0
r(St ,At ) = ri (St ,At )
pj (o∗h |·) = Gus(Ih ,(q,a)h1:j ,Oh )
if j > 1 then
rp (St ,At ) = pj (o∗h |·) − pj−1 (o∗h |·)
r(St ,At ) = r(St ,At ) + rp (St ,At )
if <End>∈ qjh then
break;
p(oh |·) = Gus(Ih ,(q,a)h1:j ,Oh )
if argmaxoh p(oh |·) = o∗h then
rg (St ,At ) = 1 + λ · Jmax /j
else rg (St ,At ) = 0
r(St ,At ) = r(St ,At ) + rg (St ,At )
Define τ = (Ih ,(q,a)h1:jh ,rh )1:H
Evaluate OJ(θ) as Eq. 9 and update VQG agent
Evaluate OL(ϕ) as Eq. 10 and update bϕ baseline
OJ(θ) ≈
X
Mj
J X
j=1 m=1
πθ
(Q
ϕ
τ
t0 =t
by substituting the notations with VQG agent, we have the
following policy gradient:
− bϕ )
τ
The parameters θ then can be optimized by following the
gradient update rule. In REINFORCE algorithm [14], the
gradient of J(θ) can be estimated from a batch of episodes
τ that are sampled from the policy πθ :
X
T X
πθ
OJ(θ) ≈
Oθ log πθ (St ,At )(Q (St ,At )−bϕ )
(7)
where Qπθ (St ,At ) is the state-action value function that returns the expectation of cumulative reward at (St ,At ):
T
X
Qπθ (St ,At ) = Eπθ [
r(St0 ,At0 )]
(8)
j
j
(I,(q,a)1:j−1 ,w1:m−1
,wm
)
(9)
bϕ is a baseline function to help reduce the gradient variance, which can be chosen arbitrarily. We use a one-layer
MLP that takes state St as input in VQG agent and outputs
the expected reward. The baseline bϕ is trained with mean
squared error as:
T
X
min L(ϕ) = [bϕ (St ) −
r(St0 ,At0 )]2
(10)
t=1
t=1 At ∈V
j
j
Oθ log πθ (wm
|I,(q,a)1:j−1 ,w1:m−1
)
t0 =t
τ
The whole training procedure is shown in Alg.1.
4. Experiment
In this section, we present our VQG results and conduct
comprehensive ablation analysis about each intermediate
reward. As mentioned above, the proposed method is evaluated on the GuessWhat?! game dataset [10] with pre-trained
standard Oracle and Guesser. By comparing with the
baseline and the state-of-the-art model, we show that proposed model can efficiently generate informative questions,
which serve the final goal.
4.1. Dataset & Evaluation Metric
The GuessWhat?! Dataset [10] is composed of 155,281
dialogues grounded on the 66,537 images with 134,074
unique objects. There are 821,955 question-answer pairs
in the dialogues with vocabulary size 4,900. We use the
standard split of training, validation and test in [10, 26].
Following [26], we report the accuracies of the games as
the evaluation metric. Given a J-round dialogue, if the target object o∗ is located by Guesser, the game is noted as
successful, which indicates that the VQG agent has generated the qualified questions to serve the final goal. There are
two kinds of test runs on the training set and test set respectively, named as NewObject and NewImage. NewObject
is randomly sampling target objects from the training images (but we restrict only to use new objects that are not
seen before), while NewImage is sampling objects from the
test images (unseen). We report three inference methods
namely sampling, greedy and beam-search (beam size is 5)
for these two test runs.
4.2. Implementation Details
The standard Oracle, Guesser and VQG baseline
are reproduced by referring to [26]. The error of trained
Oracle, Guesser on test set are 21.1% and 35.8% respectively. The VQG baseline is referred as Baseline in
Tab.1 and 2 1 .
We initialize the training environment with the standard
Oracle, Guesser and VQG baseline, then start to train
the VQG agent with proposed reward functions. We train
our models for 100 epochs with stochastic gradient descent
(SGD) [5]. The learning rate and batch size are 0.001 and
64, respectively. The baseline function bϕ is trained with
SGD at the same time. During each epoch, each training
image is sampled once, and one of the objects inside it is
randomly assigned as the target. We set the maximum round
Jmax = 5 and maximum length of question Mmax = 12.
The weight of the dialog round reward is set to λ = 0.1.
The progressive reward is set as η = 0.12 .
4.3. Results & Ablation Analysis
In this section, we give the overall analysis on proposed
intermediate reward functions. To better show the effectiveness of each reward, we conduct comprehensive ablation studies. Moreover, we also carry out a human interpretability study to evaluate whether human subjects can
1 These results are reported on https://github.com/GuessWhatGame by
original authors.
2 We use a grid search to select the hyper-parameters λ and η, we find
0.1 produces the best results.
Table 1: Results on training images (NewObject).
NewObject
Baseline [10]
Sole-r [26]
VQG-rg
VQG-rg +rp
VQG-rg +ri
VQG-rg +rp +ri
Sampling
41.6
58.5
60.6
62.1
61.3
63.2
Greedy
43.5
60.3
61.7
62.9
62.4
63.6
Beam-Search
47.1
60.2
61.4
63.1
62.7
63.9
Table 2: Results on test images (NewImage).
NewImage
Baseline [10]
Sole-r [26]
VQG-rg
VQG-rg +rp
VQG-rg +ri
VQG-rg +rp +ri
Sampling
39.2
56.5
58.2
59.3
58.5
59.8
Greedy
40.8
58.4
59.3
60.6
59.7
60.7
Beam-Search
44.6
58.4
59.4
60.5
60.1
60.8
understand the generated questions and how well the human can use these question-answer pairs to achieve the final goal. We note VQG agent trained with goal-achieved
reward as VQG-rg , trained with goal-achieved and progressive rewards as VQG-rg +rp , trained with goal-achieved and
informativeness rewards as VQG-rg +ri . The final agent
trained with all three rewards is noted as VQG-rg +rp +ri .
Overall Analysis Tab. 1 and 2 show the comparisons between VQG agent optimized by proposed intermediate rewards and the state-of-the-art model proposed in [26] noted
as Sole-r, which uses indicator of whether reaching the final goal as the sole reward function. As we can see, with
proposed intermediate rewards and their combinations, our
VQG agents outperform both compared models on all evaluation metrics. More specifically, our final VQG-rg +rp +ri
agent surpasses the Sole-r 4.7%, 3.3% and 3.7% accuracy
on NewObject sampling, greedy and beam-search respectively, while obtains 3.3%, 2.3% and 2.4% higher accuracy
on NewImage sampling, greedy and beam-search respectively. Moreover, all of our agents outperform the supervised baseline by a significant margin.
To fully show the effectiveness of our proposed intermediate rewards, we train three VQG agents using rg , rg +rp ,
and rg +ri rewards respectively, and conduct ablation analysis. As we can see, the VQG-rg already outperforms both
baseline and the state-of-the-art model, which means that
controlling dialogue round can push the agent to ask more
wise questions. With the combination of rp and ri reward
respectively, the performance of VQG agent further improved. We find that the improvement gained from rp reward is higher than ri reward, which suggests that the intermediate progressive reward contributes more in our experiment. Our final agent combines all rewards and achieves the
best results. Fig. 3 shows some qualitative results. More results can be found in the supplementary material, including
Baseline
Sole-r
Our VQG
Baseline
Sole-r
Our VQG
Is it a person ? No
Is it food ? No
Is it a plate ? No
Is it a table ? No
Is it a person ? No
Is it a food ? No
Is it in left ? No
Is it in front ? Yes
Is it a person ? No
Is it a food ? No
Is it a drink ? Yes
Is it in front ? Yes
Is it a donut ? Yes
Is it on the left ? No
Is it a person ? No
Is it a food ? Yes
Is it in left ? No
Is it in middle? No
Is it a food ? Yes
Is it on right ? Yes
Is it front one? Yes
[0.08, 0.12, 0.12, 0.03]
[0.09, 0.11, 0.02, 0.02]
[0.09, 0.11, 0.92, 0.98]
[0.19, 0.28]
[0.13, 0.13, 0.26, 0.22]
[0.11, 0.56, 0.72]
Failure (Table)
Failure (Table)
Success (Cup)
Failure (Donut)
Failure (Donut)
Success (Donut)
Is it a phone? No
Is it a book ? No
Is it a remote ? No
Is it in left ? No
Is it in middle? No
Is it a remote? No
Is it a laptop? Yes
Is it on right? Yes
Is it in front? Yes
Is it food ? No
Is it a plate ? No
Is it a spoon ? No
Is it food ? No
Is it a bow ? Yes
Is it in left ? Yes
Is it in front? Yes
Is it food ? No
Is it a drink ? Yes
Is it on right ? No
Is it in front ? Yes
[0.10, 0.20]
[0.07, 0.03, 0.02]
[0.20, 0.48, 0.99, 1.00]
[0.09, 0.08, 0.08, 0.07]
[0.09, 0.09, 0.05, 0.01]
[0.08, 0.67, 0.86, 0.89]
Failure (Table)
Failure (Keyboard)
Success (Laptop)
Failure (Knife)
Failure (Fork)
Success (Glass)
Figure 3: Some qualitative results of our VQG agent (green), and the comparisons with baseline (blue) and Sole-r model (brown). The elements in the
middle array indicate the probabilities of successfully locating the target object after each round. Better viewed in color.
15500
15500
13500
13500
0.7
0.7
Ours
Sole-r
Baseline
Ours
Sole-r
Baseline
0.6
0.6
11500
11500
0.5
0.5
9500
9500
0.4
0.4
7500
7500
0.3
0.3
5500
5500
11
22
33
44
55
Question Informativeness We also investigate the informativeness of the questions generated by different models.
We let Oracle answer questions for all the objects at each
round, and count the percentage of high-quality questions in
the successful game. We define that a high-quality question
is a one does not lead to the same answer for all the candidate objects. The experimental results show that our VQG
agent has 87.7% high-quality questions, which is higher
than the baseline (84.7%) and Sole-r (86.3%). This confirms the contribution of the ri reward.
0.2
0.2
4.4. Human Study
Figure 4: The comparisons of success ratio between our agent and Soler, as well the baseline model, at different dialogue round. The left y-axis
indicates the number of successful dialogues, which corresponds to the bar
chart. The right y-axis indicates the success ratio, which corresponds to
the line chart. Better viewed in color.
some fail cases.
Dialogue Round We conduct an experiment to investigate the relationship between the dialogue round and the
game success ratio. More specifically, we let Guesser to
select the object at each round and calculate the success ratio at the given round, the comparisons of different models
are shown in Fig. 4. As we can see, our agent can achieve
the goal at fewer rounds compared to the other models, especially at the round three.
Progressive Trend To prove our VQG agent can learn a
progressive trend on generated questions, we count the percentage of the successful game that has a progressive (ascending) trend on the target object, by observing the probability distributions generated by Guesser at each round.
Our agent achieves 60.7%, while baseline and Sole-r are
50.8% and 57.3% respectively, which indicates that our
agent is better at generating questions in a progressive trend
considering we introduce the progressive reward, rp . Some
qualitative results of the ‘progressive trend’ are shown in
the Fig. 3, i.e., the probability of the right answer is progressively increasing.
We conduct a human study to see how well human can
guess the target object based on the questions generated by
these models. We show human subjects 50 images with
generated question-answer pairs from baseline, Sole-r, and
our final VQG agent, and let them guess the objects, i.e.,
replacing the AI guesser to a real human. We ask three
human subjects to play on the same split, and the game is
recognized as successful if at least two of them give the
right answer. Based on our experiments, averagely, subjects achieve the highest accuracy 76% based on our agent,
which achieves 52% and 70% accuracies on the baseline
and Sole-r questions respectively. These results indicate
that our agent can generate higher qualitative questions that
can benefit the human to achieve the final goal.
5. Conclusion
The ability to devise concise questions that lead to two
parties to a dialog satisfying a shared goal as effectively as
possible has important practical applications and theoretical
implications. By introducing suitably crafted intermediate
rewards into a deep reinforcement learning framework, we
have shown that it is possible to achieve this result, at least
for a particular class of goal.
The method we have devised not only achieves the Guess
What goal reliably and succinctly but also outperforms the
state-of-art. However, since the Oracle and Guesser are
fixed, they are inaccurate to a certain extent. Consider the
main objective of this paper is to show the effectiveness of
our proposed intermediate rewards on the VQG problem,
we leave it as the further work to train the three components
jointly, with a reinforcement learning framework.
References
[1] M. Agarwal and P. Mannem. Automatic gap-fill question
generation from text books. In Proc. Workshop Inno. Use
of NLP for Buil. Educ. Appl., pages 56–64. Association for
Computational Linguistics, 2011. 3
[2] J. Andreas, M. Rohrbach, T. Darrell, and D. Klein. Learning
to compose neural networks for question answering. arXiv
preprint arXiv:1601.01705, 2016. 3
[3] S. Antol, A. Agrawal, J. Lu, M. Mitchell, D. Batra,
C. Lawrence Zitnick, and D. Parikh. Vqa: Visual question
answering. In Proceedings of the IEEE International Conference on Computer Vision, pages 2425–2433, 2015. 1
[4] L. Becker, S. Basu, and L. Vanderwende. Mind the gap:
learning to choose gaps for question generation. In Proc.
Conf. North Amer. Chap. Asso. for Comp. Ling.: Human
Lang. Tech., pages 742–751. Association for Computational
Linguistics, 2012. 3
[5] L. Bottou. Large-scale machine learning with stochastic gradient descent. In Proceedings of COMPSTAT’2010, pages
177–186. Springer, 2010. 7
[6] J. C. Brown, G. A. Frishkoff, and M. Eskenazi. Automatic
question generation for vocabulary assessment. In Proc.
Conf. Human Lang. Tech. and Empi. Meth. in Natu. Lang.
Process, pages 819–826. Association for Computational Linguistics, 2005. 3
[7] A. Das, S. Kottur, K. Gupta, A. Singh, D. Yadav, J. M.
Moura, D. Parikh, and D. Batra. Visual dialog. arXiv
preprint arXiv:1611.08669, 2016. 2
[8] A. Das, S. Kottur, J. M. Moura, S. Lee, and D. Batra. Learning cooperative visual dialog agents with deep reinforcement
learning. arXiv preprint arXiv:1703.06585, 2017. 2, 3
[9] B. Davey and S. McBride. Effects of question-generation
training on reading comprehension. Journal of Educational
Psychology, 78(4):256, 1986. 3
[10] H. de Vries, F. Strub, S. Chandar, O. Pietquin, H. Larochelle,
and A. C. Courville. Guesswhat?! visual object discovery
through multi-modal dialogue. In Proc. IEEE Conf. Comp.
Vis. Patt. Recogn., 2017. 2, 3, 4, 6, 7
[11] F. A. Gers, J. Schmidhuber, and F. Cummins. Learning to
forget: Continual prediction with lstm. 1999. 4
[12] R. Hu, J. Andreas, M. Rohrbach, T. Darrell, and K. Saenko.
Learning to reason: End-to-end module networks for visual
question answering. arXiv preprint arXiv:1704.05526, 2017.
3
[13] J. Johnson, A. Karpathy, and L. Fei-Fei. Densecap: Fully
convolutional localization networks for dense captioning. In
Proc. IEEE Conf. Comp. Vis. Patt. Recogn., pages 4565–
4574, 2016. 2
[14] L. P. Kaelbling, M. L. Littman, and A. W. Moore. Reinforcement learning: A survey. J. Arti. Intell. Research, 4:237–285,
1996. 3, 6
[15] I. Labutov, S. Basu, and L. Vanderwende. Deep questions
without deep understanding. In ACL (1), pages 889–898,
2015. 3
[16] Y. Li, N. Duan, B. Zhou, X. Chu, W. Ouyang, and X. Wang.
Visual question generation as dual task of visual question
answering. arXiv preprint arXiv:1709.07192, 2017. 2
[17] S. Liu, Z. Zhu, N. Ye, S. Guadarrama, and K. Murphy. Optimization of image description metrics using policy gradient
methods. arXiv preprint arXiv:1612.00370, 2016. 3
[18] J. Lu, A. Kannan, J. Yang, D. Parikh, and D. Batra. Best
of both worlds: Transferring knowledge from discriminative
learning to a generative visual dialog model. arXiv preprint
arXiv:1706.01554, 2017. 3
[19] K. Mazidi and R. D. Nielsen. Linguistic considerations in
automatic question generation. In ACL (2), pages 321–326,
2014. 3
[20] I. M. Mora, S. P. de la Puente, and X. Giro-i Nieto. Towards
automatic generation of question answer pairs from images,
2016. 2
[21] N. Mostafazadeh, I. Misra, J. Devlin, M. Mitchell, X. He,
and L. Vanderwende. Generating natural questions about an
image. arXiv preprint arXiv:1603.06059, 2016. 2
[22] A. Y. Ng, D. Harada, and S. Russell. Policy invariance under
reward transformations: Theory and application to reward
shaping. In Proc. Int. Conf. Mach. Learn., volume 99, pages
278–287, 1999. 5
[23] Z. Ren, X. Wang, N. Zhang, X. Lv, and L.-J. Li. Deep reinforcement learning-based image captioning with embedding
reward. Proc. IEEE Conf. Comp. Vis. Patt. Recogn., 2017. 3
[24] S. J. Rennie, E. Marcheret, Y. Mroueh, J. Ross, and V. Goel.
Self-critical sequence training for image captioning. arXiv
preprint arXiv:1612.00563, 2016. 3
[25] H. Singer and D. Donlan. Active comprehension: Problemsolving schema with question generation for comprehension
of complex short stories. Reading Research Quarterly, pages
166–186, 1982. 3
[26] F. Strub, H. de Vries, J. Mary, B. Piot, A. C. Courville, and
O. Pietquin. End-to-end optimization of goal-driven and visually grounded dialogue systems. In Proc. Int. Joint Conf.
Artificial Intell., 2017. 2, 3, 7
[27] R. S. Sutton and A. G. Barto. Reinforcement learning: An
introduction, volume 1. MIT press Cambridge, 1998. 3
[28] R. S. Sutton, D. A. McAllester, S. P. Singh, and Y. Mansour. Policy gradient methods for reinforcement learning
with function approximation. In Advances in neural information processing systems, pages 1057–1063, 2000. 6
[29] W. J. Therrien, K. Wickstrom, and K. Jones. Effect of a combined repeated reading and question generation intervention
on reading achievement. Learning Disabilities Research &
Practice, 21(2):89–97, 2006. 3
[30] Q. Wu, P. Wang, C. Shen, A. Dick, and A. van den Hengel. Ask me anything: Free-form visual question answering
based on knowledge from external sources. In Proc. IEEE
Conf. Comp. Vis. Patt. Recogn., June 2016. 1
[31] H. Xu and K. Saenko. Ask, attend and answer: Exploring
question-guided spatial attention for visual question answering. In Proc. Eur. Conf. Comp. Vis., pages 451–466. Springer,
2016. 1
[32] S. Zhang, L. Qu, S. You, Z. Yang, and J. Zhang. Automatic
generation of grounded visual questions. arXiv preprint
arXiv:1612.06530, 2016. 2
[33] Y. Zhu, J. J. Lim, and L. Fei-Fei. Knowledge acquisition for
visual question answering via iterative querying. Proc. IEEE
Conf. Comp. Vis. Patt. Recogn., 2017. 3
| 1cs.CV
|
arXiv:1603.02864v1 [math.SG] 9 Mar 2016
THE AUTONOMOUS NORM ON Ham (R2n ) IS
BOUNDED
MICHAEL BRANDENBURSKY AND JAREK KĘDRA
Abstract. We prove that the autonomous norm on the group of
compactly supported Hamiltonian diffeomorphisms of the standard
R2n is bounded.
Let (M, ω) be a symplectic manifold and let Ham(M, ω) be the group
of compactly supported Hamiltonian diffeomorphisms of (M, ω). Recall that a Hamiltonian diffeomorphism f is a time-one map of the
flow generated by the vector field XFt defined by ω(XFt , −) = dFt .
Here F : M × S 1 → R is a smooth compactly supported function and
F (x, t) = Ft (x) (see [9, Section 5.1] for details). The function F is
called a Hamiltonian of f . If F does not depend on the second variable
or is time independent then f is called autonomous. It is known that
every Hamiltonian diffeomorphism is a product of autonomous ones [3].
The autonomous norm on Ham(M, ω) is defined by:
kf k = min{k ∈ N | f = a1 · · · an , where ai is autonomous}.
It is a conjugation invariant norm and is known to be unbounded on
the group of compactly supported Hamiltonian diffeomorphisms of an
oriented surface of finite area [2, 4, 3, 6].
This paper is concerned with the group Ham(R2n ) of compactly supported Hamiltonian diffeomorphism of the Euclidean space equipped
with the standard symplectic form. We prove the following result.
Theorem. The diameter of the autonomous norm on Ham (R2n ) is
bounded above by 3.
Proof. Let f ∈ Ham (R2n ). Let f = a1 · · · am , where ai ∈ Ham (R2n )
are autonomous diffeomorphisms with compactly supported Hamiltonian functions Fi : R2n → R. Let B(r) ⊂ R2n be an Euclidean ball of
radius r > 0, centered at the origin and containing the union of the
supports of the functions Fi .
Lemma. There exists an autonomous diffeomorphism h ∈ Ham (R2n )
displacing the ball B(r) m times.
1
2
MICHAEL BRANDENBURSKY AND JAREK KĘDRA
The statement of the lemma means that hi (B(r)) ∩ hj (B(r)) = ∅ for
0 ≤ i 6= j ≤ m. It follows from [5, Lemma 2.6] that there exists
g ∈ Ham (R2n ) such that the following equality holds:
2
m
f = a1 · · · am = [h, g]ah1 ah2 · · · ahm ,
i
where ahi = hi ai h−i and h is the diffeomorphism from the Lemma.
Observe that, since the supports of Fi ◦ hi are pairwise disjoint for i ∈
2
m
{1, . . . , m}, we obtain that the composition ah1 ah2 · · · ahm is autonomous
with the Hamiltonian function equal to
F1 ◦ h + F2 ◦ h2 + · · · + Fm ◦ hm .
Since the commutator [h, g] = h · hg is a product of two autonomous
diffeomorphisms we obtain that f is a product of three autonomous
diffeomorphisms.
Proof of the Lemma. Let H1 : R → R be a smooth function satisfying
the following conditions:
(1) H1 (y) = 0 for |y| > r + 1
(2) H1′ (y) = r for |y| ≤ r.
Let H(x1 , y1 , . . . , xn , yn ) = H1 (y1 ). We have that dH = rdx1 and that
the induced Hamiltonian vector field X is equal to r ∂x∂ 1 . Thus the
induced Hamiltonian diffeomorphism displaces the ball B(r) as many
times as we like. Taking an appropriate cut off function we obtain the
required compactly supported diffeomorphism h.
Remarks. If f in Theorem is contained in the kernel of the Calabi
homomorphism (see Section 8.B of [1] for a definition) then the same
argument shows that it is a product of up to three autonomous diffeomorphisms with trivial Calabi invariant.
It is known that the Hofer norm on Ham (R2n ) is unbounded and stably
bounded [7]. The kernel of the Calabi homomorphism does not admit
nontrivial quasimorphisms, however, it is stably unbounded [8].
It is not difficult to see that the diameter of the autonomous norm on
Ham (R2n ) is at least 2. To the best of our knowledge it is an open
question whether there exists a Hamiltonian diffeomorphism of R2n of
autonomous norm equal to 3.
THE AUTONOMOUS NORM ON Ham R2n IS BOUNDED
3
Acknowledgements. We thank the Center for Advanced Studies in
Mathematics at Ben Gurion University for supporting the visit of the
second author at BGU.
References
[1] Vladimir I. Arnold and Boris A. Khesin. Topological methods in hydrodynamics,
volume 125 of Applied Mathematical Sciences. Springer-Verlag, New York, 1998.
[2] Michael Brandenbursky and Jarek Kedra. On the autonomous metric on the
‘
group of area-preserving diffeomorphisms of the 2-disc. Algebr. Geom. Topol.,
13(2):795–816, 2013.
[3] Michael Brandenbursky, Jarek Kedra, and Egor Shelukhin. On the autonomous
norm on the group of Hamiltonian diffeomorphisms of the torus. Available at
http://arxiv.org/abs/1602.03287v2.
[4] Michael Brandenbursky and Egor Shelukhin. On the Lp -geometry of autonomous Hamiltonian diffeomorphisms of surfaces. Math. Res. Lett. to appear.
Available at http://arxiv.org/abs/1405.7931v2.
[5] Dmitri Burago, Sergei Ivanov, and Leonid Polterovich. Conjugation-invariant
norms on groups of geometric origin. In Groups of diffeomorphisms, volume 52
of Adv. Stud. Pure Math., pages 221–250. Math. Soc. Japan, Tokyo, 2008.
[6] Jean-Marc Gambaudo and Étienne Ghys. Commutators and diffeomorphisms
of surfaces. Ergodic Theory Dynam. Systems, 24(5):1591–1617, 2004.
[7] Helmut Hofer and Eduard Zehnder. Symplectic invariants and Hamiltonian dynamics. Birkhäuser Advanced Texts: Basler Lehrbücher. [Birkhäuser Advanced
Texts: Basel Textbooks]. Birkhäuser Verlag, Basel, 1994.
[8] Morimichi Kawasaki. Relative quasimorphisms and stably unbounded norms on
the group of symplectomorphisms of the euclidean spaces. Journal of Symplectic
Geometry, 2014.
[9] Leonid Polterovich. The geometry of the group of symplectic diffeomorphisms.
Lectures in Mathematics ETH Zürich. Birkhäuser Verlag, Basel, 2001.
Ben Gurion University, Israel
E-mail address: [email protected]
University of Aberdeen and University of Szczecin
E-mail address: [email protected]
| 4math.GR
|
Macroscopic Modeling and Simulation of Managed Lane-Freeway
Networks
arXiv:1609.09470v2 [cs.SY] 21 Dec 2016
Matthew A. Wright, Roberto Horowitz, Alex A. Kurzhanskiy
March 29, 2018
Abstract
To help mitigate road congestion caused by the unrelenting growth of traffic demand, many transit
authorities have implemented managed lane policies. Managed lanes typically run parallel to a freeway’s
standard, general-purpose (GP) lanes, but are restricted to certain types of vehicles. It was originally
thought that managed lanes would improve the use of existing infrastructure through incentivization of
demand-management behaviors like carpooling, but implementations have often been characterized by
unpredicted phenomena that is often to detrimental system performance. Development of traffic models
that can capture these sorts of behaviors is a key step for helping managed lanes deliver on their promised
gains.
Towards this goal, this paper presents several macroscopic traffic modeling tools we have used for
study of freeways equipped with managed lanes, or “managed lane-freeway networks.” The proposed
framework is based on the widely-used first-order kinematic wave theory. In this model, the GP and
the managed lanes are modeled as parallel links connected by nodes, where certain type of traffic may
switch between GP and managed lane links. Two types of managed lane configuration are considered:
full-access, where vehicles can switch between the GP and the managed lanes anywhere; and separated,
where such switching is allowed only at certain locations called gates.
We incorporate two phenomena into our model that are particular to managed lane-freeway networks:
the inertia effect and the friction effect. The inertia effect reflects drivers’ inclination to stay in their lane
as long as possible and switch only if this would obviously improve their travel condition. The friction
effect reflects the empirically-observed driver fear of moving fast in a managed lane while traffic in the
adjacent GP links moves slowly due to congestion.
Calibration of models of large road networks is difficult, as the dynamics depend on many parameters whose numbers grow with the network’s size. We present an iterative learning-based approach
to calibrating our model’s physical and driver-behavioral parameters. Finally, we validate our model
and calibration methodology with case studies of simulations of two managed lane-equipped California
freeways.
Keywords: macroscopic first order traffic model, first order node model, multi-commodity traffic, managed
lanes, HOV lanes, dynamic traffic assignment, dynamic network loading, inertia effect, friction effect
1
Introduction
Traffic demand in the developed and developing worlds shows no sign of decreasing, and the resulting
congestion remains a costly source of inefficiency in the built environment. One study (Lomax et al., 2015)
estimated that, in 2014, delays due to congestion cost drivers 7 billion hours and $160B in the United States
alone, leading to the burning of 3 billion extra gallons of fuel. The historical strategy for accommodating more
demand has been construction of additional infrastructure, but in recent years planners have also developed
1
strategies to improve the performance of existing infrastructure, both through improved road operations and
demand management, which seeks to lower the number of vehicles on the road (Kurzhanskiy and Varaiya,
2015). One such strategy that has been widely adopted in the United States and other developed countries
is the creation of so-called managed lanes (Obenberger, 2004). Managed lanes are implemented on freeways
by restricting the use of one or more lanes to certain vehicles. As an example, high-occupancy-vehicle (HOV)
lanes are intended to incentivize carpooling, which reduces the total number of cars on the road as a demand
management outcome (Chang et al., 2008).
In addition to demand management, managed lanes provide an opportunity for improved road operations
through real-time, responsive traffic control. For example, tolled express lanes give drivers the opportunity
to pay a toll to drive parallel to the general-purpose lanes on a (presumably less-congested) express lane.
Traffic management authorities here have an opportunity to adjust the toll amount in response to the
real-time state of traffic on the network. The potential for managed lanes as instruments for reactive, realtime traffic operations–in addition to their demand-management purpose–has made them popular among
transportation authorities (Kurzhanskiy and Varaiya, 2015).
However, the traffic-operational effects of managed lanes are not always straightforward or as rehabilitative
as expected, as their presence can create complex traffic dynamics (Jang and Cassidy, 2012). Even in a
freeway with simple geometry, the dynamics of traffic flow are complex and not fully understood, and adding
managed lanes alongside the non-managed, general-purpose (GP) lanes only exacerbates this. In effect,
adding a managed lane creates two parallel and distinct, but coupled, traffic flows on the same physical
structure. When used as intended, managed lanes carry flows with different density-velocity characteristics
and vehicle-type (e.g., strictly HOVs) compositions than the freeway. When vehicles move between the two
lane flows, complex phenomena that are unobserved in GP-only freeways can emerge (see e.g. Daganzo and
Cassidy (2008); Liu et al. (2011); Jang and Cassidy (2012); Cassidy et al. (2015), and others).
Making better use of managed lanes requires an understanding of the macroscopic behavior they induce. One
widely-used tool for understanding macroscopic traffic flow behavior is the macroscopic traffic flow model.
A rich literature exists on macroscopic models for flows on long roads, and at junctions where those roads
meet, but an extension to the parallel-flows situation created by placing a managed lane in parallel with
a freeway (a “managed lane-freeway network”) is not straightforward. Such a model should describe when
vehicles enter or exit the managed lane, and capture the interactions caused by the two flows’ proximity,
such as the so-called “friction effect” (Liu et al., 2011; Jang and Cassidy, 2012).
This paper presents macroscopic flow modeling tools we have used for simulation of managed lane-freeway
networks. We begin in Section 2 with a discussion of relevant modeling tools from the literature, and
how we make use of them. Section 3 describes network structures for the two common managed lane
configurations: gated-access and full-access (ungated) lanes. This Section also describes models for friction
and inertia effects that arise in managed lanes. Section 4 outlines how one may take these managed lanespecific constructions and add them to a general macroscopic traffic simulation process. Section 5 describes
calibration methodologies for both network configurations. Section 6 presents case studies of two networks,
one each of gated- and full-access, and typical macroscopic simulation results.
2
Managed Lane Modeling
The modeling techniques presented in this paper are based on the first-order “kinematic wave” macroscopic
traffic flow model. These models describe aggregate traffic flows as fluids following a one-dimensional conservation law. We briefly introduce our notation here, but do not discuss the basics of this class of models.
Detailed reviews are available in many references.
2
2.1
Modeling basics
In this simulation framework, a road is divided into discrete cells, which we refer to as links. Links are drawn
between nodes: a link begins at one node and ends at another. Many links may begin and end at each node.
Each link l is characterized by density nl , the number of cars in the link. In a first-order model, the traffic
flows are fully prescribed by the density. From timestep t to t + 1, link l’s density updates according to the
equation
M
N
X
1 X
fil (t) −
flj (t) ,
(2.1)
nl (t + 1) = nl (t) +
Ll i=1
j=1
where Ll is the length of link l, fil (t) is the flow (number of vehicles) leaving link i and entering link l at
time t, flj (t) is the flow leaving link l and entering link j at time t, M is the number of links that end at
link l’s beginning node, and N is the number of links that begin at link l’s ending node.
Computing the inter-link flows requires the use of two intermediate quantities for each link. These are the
link demand Sl (t), which is the number of vehicles that wish to exit link l at timestep t; and the link supply,
Rl (t), which is the number of vehicles link l can accept at time t. Both Sl and Rl are functions of the density
nl . The model that computes Sl and Rl from nl is often called the “fundamental diagram” or “link model,”
and the model that computes the flows from all links’ supplies and demands is often called the “node model.”
A brief outline of how first-order macroscopic simulation of a road network (sometimes called a dynamic
network loading simulation) is performed could be:
1. At time t, use the link model for each link l to compute the link’s demand Sl (t) and supply Rl (t) as a
function of its density nl (t).
2. Use the node model for each node to compute the inter-link flows fij for all incoming links i and
outgoing links j as functions of Si (t), Rj (t), and information about vehicles’ desired movements ij.
3. Update the state of each link using (2.1).
4. Increment t and repeat until the desired simulation end time is reached.
The tools described in this paper are compatible with any such link model. We make use of a particular
node model that we have studied in Wright et al. (2016a,b). An aspect of this node model of particular
relevance to managed lane modeling is our “relaxed first-in-first-out (FIFO) rule” construction (Wright et al.,
2016a). This is necessary for modeling the flows between the GP and managed lanes. Without a FIFO
relaxation, congestion in one of the two lane groups could block traffic in the other when that may be
unrealistic (see Wright et al. (2016a, Section 2.2) for a detailed discussion).
So far, we have presented ingredients for a model that, while able to express many simple network topologies
by joining links and nodes, does not capture several important behaviors in managed lane-freeway networks.
The next three Sections briefly overview the additions to the standard model that will be explained in greater
detail in the remainder of the paper.
2.2
Multiple classes of vehicles and drivers
In (2.1), we describe the number of vehicles in a link as a single number, nl . In this formulation, all vehicles
are treated the same. However, for simulation in a managed lane-freeway network, it makes sense to break
3
nl into different classes of vehicles and/or drivers. For example, for a freeway with an HOV lane facility, we
might consider two classes: HOVs and non-HOVs. To this end, (2.1) can be rewritten as
M
N
X
X
1
ncl (t + 1) = ncl (t) +
f c (t) −
fljc (t) ,
(2.2)
Ll i=1 il
j=1
where c ∈ {1, . . . , C} indexes vehicle classes (often called “commodities” in the traffic literature).
Extending the density update equation to multiple classes means that the link and node models must also
c
be extended to produce per-class flows fij
. In this paper, we will not specify a particular link model, but
assume use of one that produces per-class demands Sic and overall supplies Rj (the node model, in computing
c
the fij
, is responsible for splitting the available supply Rj among the different demanding vehicle classes).
Examples of this type of link model include those considered in Wong and Wong (2002), Daganzo (2002),
and van Lint et al. (2008) (examples of multi-class link models of second- or higher-order include those of
Hoogendoorn and Bovy (2000). These types of models have higher-order analogs of supply and demand).
2.3
Toplogical expression of managed lane-freeway networks
In both (2.1) and (2.2), we describe a link in terms of its total density nl and its breakdown into percommodity portions, ncl . By discretizing the road into these one-dimensional links, we lose information
about differences between vehicle proportions across lanes, as well as inter-lane and lane-changing behavior.
This becomes a problem if such unmodeled behavior is of interest. In our setting, this means that modeling
a freeway with a managed lane should not be done with a single link following (2.1) or (2.2), as it would be
impossible to study the managed lane-freeway network behavior of interest.
Modeling differences in vehicle density across lanes is natural in microscopic and mesoscopic (see, for example,
Treiber et al. (1999); Hoogendoorn and Bovy (1999); Ngoduy (2006), and others) models, but macroscopic
models, in their simplicity, have less readily-accessible avenues for including these differences. One straightforward method is to model each lane as a separate link, as in, e.g. Bliemer (2007) or Shiomi et al. (2015).
However, this method has a few drawbacks. First, it requires the addition of some lane-assignment method
to prescribe the proportions of each vehicle class c for each lane (such as a logit model as used in Farhi
et al. (2013) and Shiomi et al. (2015)), which requires not-always-accessible data for calibration. Second,
drastically increasing the number of links in a macroscopic model will necessarily increase the size of the
state space and model complexity, which is, in a sense, incompatible with the overall goal of selecting a
macroscopic model over a micro- or mesoscopic model: some of the “macro” in the macroscopic model is lost.
Instead, in this paper we choose to model the GP lanes (or “GP lane group”) as one link and the parallel
managed lanes (or “managed lane group”) as another link. Applied to an entire length of road, this creates
a network topology of two “parallel chains” of links - one GP and one managed. The two chains will share
nodes, but cross-flows between the chains are permitted only in locations where there is physical access
(i.e., no physical barriers) and policy access (i.e., no double solid lines under U.S. traffic markings). Where
cross-flows are possible, we do not use a logit model, but instead a driver behavior model first introduced in
Wright et al. (2016a). This two-chain model is similar to the one described in Liu et al. (2012), though in
this reference, GP-managed lane crossflows were not considered.
2.4
Observed phenomena in managed lane-freeway networks
The above-mentioned friction effect is one of several emergent behaviors thought to be important for understanding traffic phenomena present on managed lane-freeway networks, and has been blamed for some of the
“underperformance” of managed lanes (Jang and Cassidy, 2012). To expand on our earlier description, the
4
friction effect describes a tendency of vehicles in an HOV lane to reduce their speed when the vehicles in the
adjacent GP lane(s) congest to the point where they slow down. It has been hypothesized that this occurs
due to HOV drivers being uncomfortable when traveling at a drastically higher speed than the adjacent GP
vehicles, and slowing down to reduce the speed differential (Jang and Cassidy, 2012).
In Liu et al. (2012), the authors created a macroscopic model of a managed lane-freeway network with a
friction effect model. They proposed modeling the friction effect by having two separate link models for the
managed lanes – one each for when the GP lane is above or below some threshold density value.
In Section 3, we propose a more expressive model of the friction effect by replacing this piecewise implementation with a linear implementation, where the reduction in demands Slc in the managed lane is proportional
to the speed differential between the managed lane(s) and GP lane(s). A linear relationship like this was
hypothesized in, e.g., Jang and Cassidy (2012, Fig. 5(c)). The magnitude of the friction effect (i.e., the degree to which managed lane drivers slow down to match GP lane drivers’ speed) is thought to be dependent
on the type of separation between the managed lane(s) and GP lane(s) (e.g., painted lines vs. a concrete
barrier) (Jang et al., 2012), which can be encoded by selecting the linear coefficient to this speed differential.
3
Full- and Gated-Access Managed Lane-Freeway Network Topologies
We will consider two types of managed lane-freeway network configurations: full access and separated with
gated access. In a a full-access configuration, the managed lane(s) are not physically separated from the
GP lane(s), and eligible vehicles may switch between the two lane groups at any location. Often, full-access
managed lane(s) are special-use only during certain periods of the day, and at other times they serve as GP
lane(s) (e.g., HOV lanes are often accessible to non-HOVs outside of rush hour). On the other hand, in a
gated-access configuration, traffic may switch between the managed lane(s) and GP lane(s) only at certain
locations, called gates; at non-gate locations, the two lane groups are separated by road markings (i.e., a
double solid line in the U.S.) or a physical barrier. Usually, gated-access managed lanes are special-use at
all times. The implemented managed lane access scheme depends on jurisdiction. For example, full-access
lanes are common in Northern California, and separated lanes are common in Southern California.
The differences in physical geometry and access points between the two access types requires two different
types of topology in constructing a network for a macroscopic model.
3.1
Note on link and node models used in this section
As discussed in the previous Section, we attempt to be agnostic with regards to the particular link model
(e.g., first-order fundamental diagram) used in our implementations. However, in our model of the friction
effect in Section 3.3, we parameterize friction being in effect on a particular link l for a particular vehicle
class c at time t by adjusting that link’s demand Sl (t). For that discussion only, we specify a particular link
model. This link model is reviewed in Appendix A. This link model is also the one used in the simulations
presented in Section 6.
The node model used here and for the remainder of this paper, when a particular form is necessary, is the
one discussed in Wright et al. (2016a) and Wright et al. (2016b). We use this node model because it handles
multi-commodity traffic, optimizes the utilization of downstream supply, makes use of input link priorities
and has a relaxation of the “conservation of turning ratios” or “first-in-first-out” (FIFO) constraint of most
node models. This last feature allows us to describe a set of GP lanes just upstream of an offramp with
one link, and handle a condition of a congested offramp by having the congestion spill back onto only the
offramp-serving lanes of the GP link (as opposed to the entirety of the GP link). See Wright et al. (2016a,
Section 3) and Wright et al. (2016b) for more discussion.
5
Figure 1: Freeway with full-access managed lane. ML = Managed Lane.
3.2
Full-access managed lanes
A full-access managed lane configuration is presented in Figure 1: GP and managed links (recall, as discussed
above, all GP lanes and all managed lanes are collapsed into one link each) are parallel with the same geometry
and share the same beginning and ending node pairs; traffic flow exchange between GP and managed lanes
can happen at every node. Note that in Figure 1, we use a slightly irregular numbering scheme so that it is
clear whether a link is a GP link, managed lane link, or ramp link. Parallel links in the graph have numbers
made up of the digit of their terminating node, with GP links having one digit (i.e., link 1), managed lane
links having two (i.e., link 11), and ramp links having three (i.e., link 111). Note also that we use a U.S.-style,
driving-on-the-right convention here, with the managed lane(s) on the left of the GP lanes and the ramps
on the far right.
Links that are too long for modeling purposes (i.e., that create too low-resolution a model) may be broken
up into smaller ones by creating more nodes, such as nodes 2 and 3 in Figure 1. Fundamental diagrams for
parallel GP and managed lane links may be different (Liu et al., 2011).
We introduce two vehicle classes (C = 2): c = 1 corresponds to the GP-only traffic and c = 2 corresponds
to the special traffic. When the managed lane(s) is (are) active, c = 1-traffic is confined to the GP link,
whereas c = 2-traffic can use both the GP and managed lane links. We denote the portion of vehicles of
c
class c in link i that will attempt to enter link j as βi,j
. This quantity is called the split ratio.
3.2.1
Split ratios for full-access managed lanes
We make an assumption that both vehicle classes take offramps at the same rate. For example, for node 1 in
1
2
Figure 1, we might say that βi,222
= βi,222
, βi,222 . Strictly speaking, it is not necessary to assume that the
c
βi,222 are equal for all c. However, in practice the offramp split ratios are typically estimated from flow count
data taken from detectors on the offramp and freeway. Generally speaking, these detectors cannot identify
c
vehicle type, so the only quantity estimable is a flow-weighted average of the quantities βi,222
. To estimate
the class-specific split ratios, one needs some extra knowledge of the tendency of each class to take each
offramp (for example, that GP-only vehicles are half as likely as special traffic to take a certain offramp).
Assuming that each class exits the freeway at the same rate is a simple and reasonable-seeming assumption.
This same problem of unidentifiability from typical data appears in several other split ratios. First, it may
not be possible to tell how many vehicles taking an offramp link come from the upstream GP, managed lane,
or (if present) onramp link. In this case, some assumptions must then be made. For example, three different
assumptions that may be reasonable are (1) that vehicles in each link i take the offramp at the same rate;
or (2) that no vehicles from the managed lane(s) are able to cross the GP lanes to take the offramp at this
node, and that no vehicles entering via the onramp, if one is present, exit via the offramp at the same node;
or (3) that vehicles in GP and managed links take the offramp at the same rate, while no vehicles coming
from the onramp are directed to the offramp. Looking back at node 1 in Figure 1 again, assumption (1)
would say the βi,222 are equal for all i; assumption (2) would say β11,222 = β111,222 = 0; and assumption (3)
6
would say β1,222 = β11,222 and β111,222 = 0. The best assumption for each node will depend on the road
geometry for that particular part of the road (how near the offramp is to any onramps, how many GP lanes
a vehicle in a managed lane would have to cross, etc.).
Second, the crossflows between the GP and managed lane links are not observable. Even if there exist
detectors immediately upstream and downstream of the node where traffic can switch between GP and
managed lanes, it is impossible to uniquely identify the crossflows. In a simulation, these crossflows must be
governed by some driver choice model.
Putting together these assumptions and the special-traffic-only policy for the managed lane, we can summarize most of the necessary split ratios needed for computing flows in a node model. For example, for node 1
in Figure 1,
j=2
j = 22 j = 222
i
=
1
1
−
β
0
β1,222
1,222
1
βi,j
=
i
=
11
n/a
n/a
n/a
i = 111 1 − β111,222 0
β111,222
j = 2 j = 22 j = 222
i
=
1
−
−
β1,222
2
βi,j
=
i
=
11
−
−
β
11,222
i = 111 −
−
β111,222
1
where “n/a” means that the split ratios β11,j
are not applicable, as there should be no vehicles of class
c = 1 in the managed lane. The split ratios marked with a dash are those above-mentioned flows that are
unobservable and come from some driver choice model. Of course, whatever method is chosen to compute
2
2
2
these unknown split ratios, we must have βi,2
+ βi,22
= 1 − βi,222
.
Similarly, for node 2, which does not have an onramp or an offramp,
j = 3 j = 33
1
i=2
1
0
βi,j
=
i = 22 n/a
n/a
j = 3 j = 33
2
i=2
−
−
βi,j
=
i = 22 −
−
with “n/a” and the dash meaning the same as above.
As previously mentioned, full-access managed lanes often have certain time periods during which nonspecial
(c = 1) traffic is allowed into the managed lane. We can model this change in policy simply by changing the
split ratios at the nodes. For node 1, for example, the nonrestrictive policy is encoded as
j = 2 j = 22 j = 222
i=1
−
−
β1,222
c
for c = {1, 2},
βi,j =
i
=
11
−
−
β
11,222
i = 111 −
−
β111,222
and for node 2 as
j = 3 j = 33
c
i=2
−
−
βi,j
=
i = 22 −
−
for c = {1, 2}.
In other words, the managed lane link is treated as additional GP lane(s), and the split ratios governing the
crossflows between the two links should be found from the driver choice model for both vehicle classes.
7
3.2.2
Node model for full-access managed lane-freeway networks
As mentioned in Section 3.1, we make use of the node model discussed in Wright et al. (2016a) and Wright
et al. (2016b) to describe a freeway network with managed lanes. This node model differentiates itself from
others in that it deals with multi-commodity traffic flow, optimally utilizes the available supply, makes use
of input link priorities, and has a relaxation of the common FIFO constraint. By default, link priorities can
be taken proportional to link capacities. To explain relaxed FIFO, say that some link i has vehicles that
wish to enter both links j and j 0 . If link j 0 is jammed and cannot accept any more vehicles, a strict FIFO
constraint would say that the vehicles in i that wish to enter j 0 will queue at i’s exit, and block the vehicles
that wish to enter j. In a multi-lane road, however, only certain lanes may queue, and traffic to j may still
pass through other lanes. The relaxation is encoded in so-called “mutual restriction intervals” ηji 0 j ⊆ [0, 1].
This interval partly describes the overlapping regions of link i’s exit that serve both links j and j 0 . For
ηji 0 j = [y, z], a z − y portion of i’s lanes that serve j also serve j 0 , and will be blocked by the cars queueing
to enter j 0 when j 0 is congested. For example, if j is served by three lanes of i, and of those three lanes, the
leftmost also serves j 0 , we would have ηji 0 j = [0, 1/3].
As an example, we consider again node 1 in Figure 1. Say that the GP links (1 and 2) have four lanes, that
the managed lane links (11 and 22) have two lanes, and that the onramp merges into and the offramp diverges
from the rightmost GP lane. Further, we say that when the managed lane link is congested, vehicles in the
GP lanes that wish to enter the managed lanes will queue only in the leftmost GP lane. On the other hand,
when the GP link is congested, vehicles in the managed lanes that wish to enter the GP lanes will queue only
in the rightmost managed lane. Finally, we suppose that jammed offramp (222) will cause vehicles to queue
only in the rightmost GP lane. Taking together all of these statements, our mutual restriction intervals for
this example are:
j = 2 j = 22 j = 222
0
j =2
[0, 1]
[0, 1]
[0, 1]
1
ηj 0 j =
0
1/4]
j
=
22
[0,
[0,
1]
∅
0
j = 222 [3/4, 1]
∅
[0, 1]
j = 2 j = 22 j = 222
0
j =2
[0, 1]
[0, 1/2]
∅
11
ηj 0 j =
0
j = 22
[0, 1]
[0, 1]
∅
0
j = 222
∅
∅
[0, 1]
j = 2 j = 22 j = 222
0
j =2
[0, 1]
[0, 1]
[0, 1]
111
ηj 0 j =
.
∅
[0, 1]
∅
j 0 = 22
0
j = 222 [0, 1]
[0, 1]
[0, 1]
To read the above tables, recall that as written, j 0 is the congested, restricting link, and j is the restricted
link. These chosen restriction intervals allow for expected behavior in this network, such as a congested GP
link causing possible queueing in the right managed lane (if some drivers are trying to enter the GP link),
but no spillback into the left managed lane.1
For a detailed discussion on how mutual restriction intervals are included in the node model’s flow calculations
and solution algorithms, see Wright et al. (2016a) and Wright et al. (2016b).
3.3
Friction effect
The friction effect is an empirically-observed phenomenon in situations where managed lanes are relatively
uncongested, but the managed-lane traffic will still slow down when the adjacent GP lanes congest and
1 In
this example, we assume that the managed lane has two sublanes — left and right.
8
slow down (see Daganzo and Cassidy (2008), Liu et al. (2011), Cassidy et al. (2015), etc.). It has been
hypothesized (Jang and Cassidy, 2012) that this phenomenon arises from the managed-lane drivers’ fear
that slower-moving vehicles will suddenly and dangerously change into the managed lane ahead of them.
We suggest modeling the friction effect based on a feedback mechanism that uses the difference of speeds in
the parallel GP and managed lane links to scale down the flow (and therefore the speed) out of the managed
lane link if necessary.
To explain the concept, we again refer to Figure 1 and consider parallel links 1 (GP) and 11 (managed lane).
Recall that, under a first-order model (2.2), the speed of traffic in link l at time t is
( PC PN f c (t)
Pc
lj
c=1
c
P j=1
if
c
c=1 nl (t) > 0,
c=1 nl (t)
(3.1)
vl (t) =
f
vl (t)
otherwise,
where vlf (t) is the theoretical free flow speed of link l at time t.
We say that the friction effect is present in managed lane link 11 (following the notation of Figure 1) at time
t if
n
o
v1 (t − 1) < min v1f , v11 (t − 1) ,
(3.2)
which means that (1) the GP link is in congestion (its speed is below its current free flow speed), and (2)
the speed in the GP link is less than the speed in the managed lane link. We denote this speed differential
as:
f
∆11 (t) = v11
− v1 (t − 1).
(3.3)
It has been observed (Jang et al., 2012) that the magnitude of the friction effect — the degree to which
managed-lane drivers slow down towards the GP lane’s traffic speed — depends on the physical configuration
of the road. For example, less of a friction effect will be present on managed lanes that are separated from
the GP lanes by a buffer zone than those that are contiguous with the GP lanes (Jang et al., 2012), and the
presence of a concrete barrier would practically eliminate the friction effect. Other factors that may affect
this magnitude include, for example, whether there is more than one managed lane, or whether there is a
shoulder lane to the left of the managed lane that drivers could swerve into if necessary.
To encode this variability in the magnitude of the friction effect in managed lane link 11, we introduce
σ11 ∈ [0, 1] the friction coefficient of this link. The friction coefficient reflects the strength of the friction.
Its value depends on the particular managed configuration and is chosen by the modeler. A value of σ11 = 0
means there is no friction (which may be appropriate if, perhaps, the managed lane(s) are separated from
the GP lanes by a concrete barrier), and σ11 = 1 means that the managed lane link speed tracks the GP
link speed exactly.
When the friction effect is active (i.e., when (3.2) is true), we adjust the fundamental diagram of the managed
lane link by scaling down its theoretical free flow speed vlf (t), and propagate that change through the rest of
the fundamental diagram parameters. The exact mathematical changes will of course be different for every
different form of fundamental diagram. For the particular fundamental diagram discussed in Appendix A,
this means adjusting the free flow speed and capacity as follows:
f
v̂11 (t) = v11
(t) − σ11 ∆11 (t);
(3.4)
F̂11 (t) = v̂11 (t)n+
11 ,
(3.5)
where n+
11 is the high critical density (see Appendix A for its definition), and using these adjusted values in
the calculation of the sending function (A.1),
(
)
F̂
(t)
11
c
S11
(t) = v̂11 (t)nc11 (t) min 1,
.
(3.6)
PC
v̂11 (t) c=1 nc11 (t)
9
For the fundamental diagram of Appendix A, we must also check whether
C
X
c=1
nc11 (t) <
F̂11 (t)
f
v11
− ∆11 (t)
=
v̂11 (t)n+
11 (t)
f
v11
− ∆11 (t)
(3.7)
.
If not, then applying friction will lead to the managed lane link speed falling below the GP link speed, and
the unadjusted sending function should be used (the possibility of this happening is due to the use of two
critical densities to create Appendix A’s fundamental diagram’s “backwards lambda” shape).
Evidence of a friction effect reducing managed lane speed in a form that is linear in the speed differential
(i.e., (3.4)) can be found in e.g., Jang et al. (2012). Again, the exact form of (3.6), the sending function with
friction, will depend on the link’s original fundamental diagram model.
3.4
Separated managed lanes with gated access
A separated, gated-access managed lane configuration is presented in Figure 2. Unlike the full-access configuration, the GP and managed lane link chains do not necessarily meet at every node. Instead, they need
only meet at a few locations, where vehicles can move into and out of the managed lane(s). Note that, unlike
the full-access configuration, there is no need for GP and managed lane links to be aligned.
As labeled in Figure 2, the nodes where the two link chains meet are called gates (as an aside, one way
to describe the full-access managed lane configuration would be that every node is a gate). Similar to our
construction of excluding GP-only traffic from the managed lane in the full-access case, we can disable flow
exchange at a given gate by fixing split ratios so that they keep traffic in their lanes. For example, to disable
c
c
the gate (the flow exchange between the two lanes) at node 2 in Figure 1, we set β2,3
= 1 and β22,33
= 1 (this
c
c
means that β2,33 = 0 and β22,3 = 0), c = 1, 2. Thus, the full-access managed lane can be easily converted into
the separated managed lane by setting non-exchanging split ratios everywhere but designated gate-nodes.
In practice, a gate is stretch of freeway that may be a few hundreds of meters long (Cassidy et al., 2015),
and, potentially, we can designate two or three sequential nodes as gates. In this paper, however, we model
a gate as a single node.
Figure 2: Freeway with separated managed lane and gates. ML = Managed Lane.
For the gated-access configuration, we suggest setting mutual restriction coefficients in the same manner as
full-access managed lanes, in Section 3.2.2.
Compared to the full-access managed lane configuration, the gated-access configuration has a smaller friction
effect (Jang et al., 2012): drivers in the separated managed lane feel somewhat protected by the buffer,
whether it is virtual (double solid line) or real (concrete), from vehicles changing abruptly from the slow
moving GP lane and, therefore, do not drop speed as dramatically. The degree to which the friction effect is
mitigated is disputed (e.g., see Footnote 3 in Cassidy et al. (2015)), but overall the bottlenecks created by
the gates are much greater instigators of congestion (Cassidy et al., 2015). Inclusion of the friction effect in
modeling separated managed lane configurations is thus not as essential as in modeling the full-access case.
10
3.4.1
Modeling a flow of vehicles from the managed lanes to the offramps
Recall from Section 3.2.1 that, in the full-access managed lane model, we can model vehicles moving from
the managed lane link to offramps in a straightforward manner, by setting corresponding split ratios (for
c
example, β11,222
, c = 1, 2, for node 1 in configuration from Figure 1). For the gated-access configuration,
however, modeling traffic as moving from the managed lanes to offramps is more complicated: generally,
gates do not coincide with offramp locations. In fact, there are typically between two and five offramps
between two gate locations. We denote the offramps in the GP road segment connecting two gates as exits
e1 , e2 , . . . , eK (see Figure 2). These offramps cannot be accessed directly from the managed lane. Instead,
vehicles traveling in the managed lane link that intend to take one of the exits e1 , . . . , eK , must switch from
the managed lane link to the GP link at gate-node 1 and then be directed to the correct offramp. Creating
a simulation where vehicles behave like this requires a more involved modeling construction.
To resolve this challenge, our gated-access model introduces new vehicle classes in addition to the c = 1
(GP-only) and c = 2 (special) traffic used in the full-access model of Section 3.2. These additional classes
will be used to distinguish subsets of the special traffic population by its destination offramp. If K is the
largest number of offramps between two adjacent gates, then altogether we have C = K + 2 vehicle classes:
c = 1, 2, e1 , . . . , eK , where ek indicates the class of vehicles that will exit through the k-th offramp after
leaving the managed lane through the gate. By definition, traffic of type c = ek may exist in the GP lane
segment between gate 1 and offramp ek , but there is no traffic of this type either in the GP link segment
between offramp ek and gate 2 or in the managed lane link. This movement pattern is ensured by setting
constant split ratios:
βiekx1 = 1, i = 1, 11, 111,
βxekk ek = 1,
βxekk0 ek0 = 0, k 0 6= k,
direct all ek -type traffic to the GP link at gate 1;
direct all ek -type traffic to offramp ek ;
do not send any ek -type traffic to other offramps,
(3.8)
where k = 1, . . . , K, and xk denotes the input GP link for the node that has the output link ek (see Figure 2).
Vehicles of class c = ek do not enter the network via onramps or the upstream boundary, but instead
are converted from special c = 2 traffic as it leaves the managed lane(s) through the gate. We perform
this
as part of the link model computation, such that the total demand for the switching link,
PC conversion
c
c=1 Sl remains the same before and after the switch. The exact link on which this switching takes place
is the managed lane link immediately upstream of the gate (e.g., link 11 in Figure 2).
We say that the amount of traffic that should change from vehicle class c = 2 to vehicle class ek at time
t is (using link 11 as an example) n211 (t)βx2k ek (t)v11 (t), where v11 (t) is in units of vehicles per simulation
timestep (this factor is included so that the switching done is proportional to link 11’s outflow, rather than
its density), and βx2k ek (t) is the split ratio from xk , the GP link immediately upstream of exit ek , and the
exit ek at time t. This statement is based on the assumption that the vehicles in the managed lane link will
exit the freeway through exit ek at the same rate as special (c = 2) vehicles that happened to stay in the GP
lanes. That is, if a βx2k ek portion of c = 2 vehicles intend to leave the GP lanes through exit ek , then a βx2k ek
portion of the c = 2 vehicles in the managed lane(s) will leave the managed lane link at the closest upstream
gate and leave the network at exit ek when they reach it. Note that if there are K 0 < K offramps between
two particular gates, then no vehicles should switch to type c = ek , k ∈ {K 0 + 1, . . . , K} at the upstream
gate, as they would have no ramp to exit through.
Using Figure 2 as a reference, we can now formally describe the procedure for destination assignment to
traffic in the managed lane link.
1. Given are vehicle counts per commodity nc11 , c = 1, 2, e1 , . . . , eK ; free flow speed v11 ; and offramp split
ratios βx1k ,ek and βx2k ,ek , k = 1, . . . , K.2
a given GP segment connecting two adjacent gates has K 0 offramps, where K 0 < K, then assume βx1k ,ek = βx2k ,ek = 0
for k ∈ (K 0 , K].
2 If
11
2. Initialize:
ñc11 (0)
k
:= nc11 ,
:=
c = 1, 2, e1 , . . . eK ;
1.
3. Assign ek -type traffic:
ñe11k (k)
ñ111 (k)
ñ211 (k)
=
ñe11k (k − 1) + βx1k ,ek v11 ñ111 (k − 1) + βx2k ,ek v11 ñ211 (k − 1);
=
ñ111 (k
ñ211 (k
=
− 1) −
− 1) −
βx1k ,ek v11 ñ111 (k
βx2k ,ek v11 ñ211 (k
(3.9)
− 1);
(3.10)
− 1).
(3.11)
4. If k < K, then set k := k + 1 and return to step 3.
5. Update the state:
nc11 = ñc11 (K),
c = 1, 2, e1 , . . . , eK .
After the switches to c = ek class traffic have been done, there may be unresolved split ratios for both
classes c = 1 and c = 2 at the gates (similar to the dashed split ratios in the tables in Section 3.2.1). These
split ratios should be filled in with the same tools as those in Section 3.2.1: some sort of driver lane choice
behavior.
3.5
Inertia effect
At the end of the previous Section, we mentioned that some split ratios will likely be undefined after switching
c = 2 class vehicles to the c = ek classes. In particular, c = 2 class vehicles in the upstream GP link (e.g.,
link 1 in Figure 2) and remaining c = 2 vehicles in the upstream managed lane link (e.g. link 11 in Figure
2) will need to decide whether to pass through the gate or remain in their current lane. Sample tools for
modeling driver lane choices such as these include the class of “logit” logistic regression models (McFadden,
1973), or dynamic split ratio solvers such as the one presented in Wright et al. (2016a) and reviewed in
Appendix B.
When applied to lane choice (e.g., as in Farhi et al. (2013)), logit models produce a set of portions in [0, 1],
one for each lane, that sum to one. Each lane’s value is the equilibrium portion of vehicles that will travel
on that lane. In a dynamic simulation context, such as considered in this article, the differences between
the logit model’s equilibrium portions and the current distribution of vehicles across lanes at time t are used
to select split ratios at time t such that the actual distribution approaches the logit equilibrium. We argue,
though, that use of this sort of split ratio solver in unmodified form might be inappropriate for computing
gate split ratios in the gated-access managed lane configuration. Unmodified, a value function for either the
GP or managed lane link might include terms such as the link’s speed of traffic, density, etc. However, often
a gated-access managed lane is separated from the GP lanes by some buffer zone or visibility-obstructing
barrier that makes switching between the two links more hazardous than switching between two contiguous
lanes. Therefore, if one uses a logit-based model, it would be appropriate to modify the logit model’s value
function such that staying in the current link (e.g., the movements (1,2) and (11,22) in Figure 3) has some
positive value for drivers, and the gains (e.g., in travel time) for ingress/egress movements (e.g., (1,22) and
(11,2) in Figure 3) must be of more value than the staying-in-the-lane value. We refer to this model as the
inertia effect.
We may also incorporate the inertia effect into dynamic split ratio solvers, such as the one introduced in
Wright et al. (2016a) and reviewed in Appendix B. At time t, this particular
algorithm selects split ratios
P
in an attempt to balance the density ratio at the next timestep t + 1, c ncl (t + 1)/nJl , where nJl is the jam
12
Figure 3: A node where some of the input links form travel facilities with some of the output links. ML =
Managed Lane.
density, or the maximum number of vehicles that
P link l can hold. For
P example, if applied to the node in
Figure 3, this algorithm would attempt to make c nc2 (t + 1)/nJl and c nc22 (t + 1)/nJl as equal as possible.
For more details, see Appendix B and Wright et al. (2016a).
Here, we modify several steps of the solver so that this equality-seeking goal is balanced with a goal towards
enforcing the inertia effect. We illustrate these changes with the particular example of the node in Figure 3.
Ensuring that the split ratio assignment algorithm gives preferences to movements (1,2) and (11,22) over
(1,22) and (11,2) can be done in step 5 of the original algorithm, setting of oriented priorities. Specifically,
we modify (B.4). For this particular example, the original formula gives us:
c
c
γ1,2
(k) = β̃1,2
(k) +
c
c
γ11,2
(k) = β̃11,2
(k)
c
β 1 (k)
2c
β (k)
+ 112
c
c
γ1,22
(k) = β̃1,22
(k) +
c
c
γ11,22
(k) = β̃11,22
(k)
c
β 1 (k)
2c ;
β (k)
+ 112 .
c
c
c
Since for k = 0 β̃i,j
(0) = 0, i = 1, 11, j = 2, 22, we get γ1,2
(0) = γ1,22
(0) =
c
β 11 (k)
2
c
β 1 (k)
2
c
c
and γ11,2
(0) = γ11,22
(0) =
, which, according to (B.3), yields p̃1,2 (0) = p̃1,22 (0) and p̃11,2 (0) = p̃11,22 (0).
For each input link i that forms one lane with an output link ̂ and class c, such that ̂ ∈ Vic , (B.4) can be
modified as follows:
c
if split ratio is defined a priori: {i, j, c} ∈ B,
βij ,
c
c
c
c
β̃
(k)
+
β
(k)λ
,
i
and j form one lane: j = ̂,
γij (k) =
(3.12)
i
ij
i
c
β̃ c (k) + β c (k) 1−λ
i
,
i
and
j
are
in
different
lanes:
j
=
6
̂,
c
i
ij
|V |−1
i
where the parameter λci ∈
h
1
|Vic | , 1
i
is called the inertia coefficient, and indicates how strong the inertia effect
is. With
=
(3.12) reduces to the original formula, (B.4). With λci = 1, all the a priori unassigned
traffic from link i must stay in its lane — be directed to output link ̂. The choice of λi lies with the modeler.
In the case of example from Figure 3, the modified formula (3.12) yields:
λci
1
|Vic | ,
c
c
c
c
c
c
(k) = β̃1,22
(k) + β 1 (k)(1 − λc1 );
γ1,2
(k) = β̃1,2
(k) + β 1 (k)λc1
γ1,22
(3.13)
c
c
c
c
c
c
γ11,2
(k) = β̃11,2
(k) + β 11 (k)(1 − λc11 ) γ11,22
(k) = β̃11,22
(k) + β 11 (k)λc11 ,
where λ1 , λ11 ∈ 12 , 1 , and picking λ1 > 12 (λ11 > 12 ) would give preference to movement (1,2) over (1,22)
(and (11,22) over (11,2)).
The way of choosing λci for multiple input links is not obvious and an arbitrary choice may result in an
unbalanced flow distribution among output links. Therefore, we suggest picking just one input-output pair
(ı̂, ̂), and for that input link setting λcı̂ = 1, while for other input links i setting λci = |V1c | , c = 1, . . . , C.
i
The input link ı̂ must be from the lane that is expected to have a positive net inflow of vehicles as a result
of the split ratio assignment and flows computed by the node model. So,
P
P
P
c
c c
i:{i,j,c}∈B
c:{i,j,c}∈B βij Si
c:{i,j,c}∈B S i +
ı̂ = arg min =
,
(3.14)
Rj
i∈Û
13
where
Û = {input links i : ∃j, c, s.t. j ∈ Vic and the pair of links (i, j) belongs to the same lane} ,
(3.15)
and j denotes the output link that is in the same lane as input link i.
For the particular example node in Figure 3, we need to determine whether flow from link 1 will proceed to
link 2 or flow from link 11 to link 22, while other a priori undefined split ratios will be computed according
to the split ratio assignment algorithm. If ı̂ = 11, then λ11 = 1, λ1 = 21 , and a priori unassigned traffic in
c
the managed lane will stay in the managed lane (β11,2
= 0), while the a priori unassigned traffic coming
from links 1 and 111, will be distributed between links 2 and 22 according to the dynamic split ratio solver.
On the other hand, if ı̂ = 1, then λ1 = 1, λ11 = 12 , and the a priori unassigned traffic in the GP lane will
c
stay in the GP lane (β1,22
= 0), while the a priori unassigned traffic coming from links 11 and 111, will be
distributed between links 2 and 22 according to the dynamic split ratio solver.
4
Putting Together the Pieces: A Managed Lane-Freeway Network
Simulation Model
In this Section, we present a unified simulation algorithm for managed lane-freeway networks. This can
be considered a fleshing-out of the simplified first-order macroscopic simulation method briefly outlined in
Section 2.1, with extensions made by incorporating the additional items we have described in the sections
between that and this one.
4.1
Definitions
• We have a network consisting of set of links L and a set of nodes N .
– A node always has at least one incoming link and one outgoing link.
– A link may have an upstream node, a downstream node, or both.
• We have C different vehicle classes traveling in the network, with classes indexed by c ∈ {1, . . . , C}.
• Let t ∈ {0, . . . , T } denote the simulation timestep.
• In addition, while we have not covered them here, the modeler may optionally choose to define control
inputs to the simulated system that modify system parameters or the system state. Such control
inputs may represent operational traffic control schemes such as ramp metering, changeable message
signs, a variable managed-lane policy, etc. In the context of this paper, we suggest including the
parameterization of the friction effect and the class-switching construction of the separated-access
managed lane model as control actions.
4.1.1
Link model definitions
• For each link l ∈ L, let there be a time-varying C-dimensional state vector ~nl (t), which denotes the
density of the link of each of the C vehicle classes at time t.
– Each element of this vector, ncl (t), updates between timestep t and timestep t + 1 according to
(4.1).
– Also define ncl,0 for all l, c, the initial condition of the system.
14
• We define three types of links:
– Ordinary links are those links that have both beginning and ending nodes.
– Origin links are those links that have only an ending node. These links represent the roads that
vehicles use to enter the network.
– Destination links are those links that have only a beginning node. These links represent the roads
that vehicles use to exit the network.
• For each link l ∈ L, define a “link model” that computes the per-class demands Slc (t) and link supply
Rl (t) as a function of t and ~nl (t). Appendix A describes a particular example link model that will be
used in the example simulations in Section 6.
• For each origin link l ∈ L, define a C-dimensional time-varying vector, d~l (t), where the c-th element
dcl (t) denotes the exogenous demand of class c into the network at link l.
4.1.2
Node model definitions
• For each node ν ∈ N , let i ∈ {1, . . . , Mν } denote the incoming links and j ∈ {1, . . . , Nν } denote the
outgoing links.
P c
c
• For each node, define {βij
(t) :
j βij (t) = 1 ∀i, c} the time-varying split ratios for each triplet {i, j, c}.
Each split ratio may be fully defined, partially defined, or fully undefined. If some split ratios for a
node are undefined, then also define for that node some split ratio solver (e.g., logit or dynamic) to fill
c
in undefined split ratios βij
(t) at time t.
• For each node, define a “node model” that, at each time t, takes its incoming links’ demands Sic (t)
c
and split ratios βij
(t), its outgoing links’ supplies Rj (t), and other nodal parameters, and computes
c
the flows fij (t). In this paper, we refer to a specific node model with a relaxed first-in-first-out
(FIFO) construction that has additional parameters ηji 0 j (t) (mutual restriction intervals) and pi (t)
(the incoming links’ priorities).
4.1.3
State update equation definitions
• All links l ∈ L update their states according to the equation
ncl (t + 1) = ncl (t) +
1
c
f c (t) − fl,out
(t)
Ll l,in
∀c ∈ {1, . . . , C},
(4.1)
which is a slightly generalized form of (2.2).
• For all ordinary and destination links,
c
fl,in
(t) =
Mν
X
(4.2)
filc (t),
i=1
where ν is the beginning node of link l.
• For all origin links,
(4.3)
c
fl,in
(t) = dcl (t).
• For all ordinary and origin links,
c
fl,out
(t) =
Nν
X
(4.4)
fljc (t),
j=1
where ν is the ending node of link l.
15
• For all destination links,
(4.5)
c
fl,out
(t) = Slc (t).
4.2
Simulation algorithm
1. Initialization:
ncl (0)
:=
ncl,0
t
:=
0,
for all l ∈ L, c ∈ {1, . . . , C}.
2. Perform all control inputs that have been (optionally) specified by the modeler. For our purposes, this
includes:
(a) For each managed lane link, modify the sending function of the link model in accordance with the
friction effect model. Recall that, for the particular link model of Appendix A, we use (3.1)-(3.7).
(b) For each managed lane link whose downstream node is a gate node in a gated-access managed
lane configuration, perform class switching as detailed in Section 3.4.1
3. For each link l ∈ L and commodity c ∈ {1, . . . , C}, compute the demand, Slc (t) using the link’s link
model.
4. For each ordinary and destination link l ∈ L, compute the supply Rl (t) using the link model. For
origin links, the supply is not used.
c
5. For each node ν ∈ N that has one or more undefined split ratios βij
(t), use the node’s split ratio solver
to complete a fully-defined set of split ratios. Note that if an inertia effect model is being used, the
modified split ratio solver, e.g. the one described in Section 3.5, should be used where appropriate.
c
6. For each node ν ∈ N , use the node model to compute throughflows fij
(t) for all i, j, c.
7. For every link l ∈ L, compute the updated state ~nl (t + 1):
• If l is an ordinary link, use (4.1), (4.2), and (4.4).
• If l is an origin link, use (4.1), (4.3), and (4.4)
• If l is a destination link, use (4.1), (4.2), and (4.5).
8. If t = T , then stop. Otherwise, increment t := t + 1 and return to step 2.
5
Calibrating the Managed Lane-Freeway Network Model
Typically, a traffic modeler will have some set of data collected from traffic detectors (e.g., velocity and flow
readings), and will create a network topology with parameter values that allow the model to reproduce these
values in simulation. Then, the parameters can be tweaked to perform prediction and analysis. For our
managed lane-freeway networks, the parameters of interest are:
1. Fundamental diagram parameters for each link. Calibration of a fundamental diagram is typically
agnostic to the node model and network topology, and there exists an abundant literature on this
topic. For the purposes of the simulations in the following Section, we used the method of Dervisoglu
et al. (2009), but any other method is appropriate.
16
2. Percentage of special (that is, able to access the managed lane) vehicles in the traffic flow entering
the system. This parameter depends on, e.g., the time of day and location as well as on the type of
managed lane. It could be roughly estimated as a ratio of the managed lane vehicle count to the total
freeway vehicle count during periods of congestion at any given location.
3. Inertia coefficients. These parameters affect only how traffic of different classes mixes in different links,
but they have no effect on the total vehicle counts produced by the simulation.
4. Friction coefficients. How to tune these parameters is an open question. In Jang and Cassidy (2012)
the dependency of a managed lane’s speed on the GP lane speed was investigated under different
densities of the managed lane, and the presented data suggests that although the correlation between
the two speeds exists, it is not overwhelmingly strong, below 0.4. Therefore, we suggest setting friction
coefficients to values not exceeding 0.4.
5. Mutual restriction intervals. It is also an open question how to estimate mutual restriction intervals
from the measurement data. See the discussion in Section 3.2.2 for some guidelines.
6. Offramp split ratios.
Calibrating a traffic model, or identifying the best values of its parameters to match real-world data, is
typically an involved process for all but the simplest network topologies. In particular, once we consider
more than a single, unbroken stretch of freeway, the nonlinear nature and network effects of these systems
means that estimating each parameter in isolation might lead to unpredictable behavior. Instead, nonlinear
and/or non-convex optimization techniques such as genetic algorithms (Poole and Kotsialos, 2012), particle
swarm methods (Poole and Kotsialos, 2016), and others (Ngoduy and Maher (2012); Fransson and Sandin
(2012), etc.) are employed.
In the managed lane-freeway networks we have discussed, the key unknown parameters we have introduced
are the offramp split ratios, item 6, which may be particularly hard to estimate as they are typically timevarying and explicitly represent driver behavior, rather than physical parameters of the road. The other
items are not too difficult to identify using methods from the literature. The remainder of this section
describes iterative methods for identification of the offramp split ratios for both the full-access and gatedaccess configurations.
5.1
Split ratios for a full-access managed lane
Consider a node, one of whose output links is an offramp, as depicted in Figure 3. We shall make the
following assumptions.
in
1. The total flow entering the offramp, fˆ222
, at any given time is known (from measurements) and is not
in
ˆ
restricted by the offramp supply: f222 < R222 .
2. The portions of traffic sent to the offramp from the managed lane and from the GP lane at any given
c
c
= β11,222
, β, c = 1, . . . , C.
time are equal: β1,222
3. None of the flow coming from the onramp (link 111), if such flow exists, is directed toward the offramp.
c
In other words, β111,222
= 0, c = 1, . . . , C.
4. The distribution of flow portions not directed to the offramp between the managed lane and the GP
c
c
c
output links is known. This can be written as: βij
= (1 − β)δij
, where δij
∈ [0, 1], as well as β111,j ,
i = 1, 11, j = 2, 22, c = 1, . . . , C, are known.
5. The demand Sic , i = 1, 11, 111, c = 1, . . . , C, and supply Rj , j = 2, 22, are given.
17
At any given time, β is unknown and is to be found.
If β were known, the node model would compute the input-output flows, in particular, fi,222 =
i = 1, 11. Define
PC
c=1
in
ψ(β) = f1,222 + f11,222 − fˆ222
.
c
fi,222
,
(5.1)
Our goal is to find β from the equation
ψ(β) = 0,
(5.2)
i
h ˆin
PC
f222
in
,
1
, where Si = c=1 Sic . Obviously, if S1 + S11 < fˆ222
, the solution does not exist,
such that β ∈ S1 +S
11
in
ˆ
and the best we can do in this case to match f222 is to set β = 1, directing all traffic from links 1 and 11 to
the offramp.
in
in
Suppose now that S1 + S11 ≥ fˆ222
. For any given fˆ222
, we assume ψ(β) is a monotonically increasing
function
of
β
(this
assumption
is
true
for
the
particular
node
model of Wright et al. (2016a)). Moreover,
ˆin
f222
ψ S1 +S11 ≤ 0, while ψ(1) ≥ 0. Thus, the solution of (5.2) within the given interval exists and can be
obtained using the bisection method.
The algorithm for finding β follows.
1. Initialize:
b(0)
:=
b(0)
:=
in
fˆ222
;
S1 + S11
1;
k
:=
0.
in
2. If S1 + S11 ≤ fˆ222
, then are not enough vehicles to satisfy the offramp demand. Set β = 1 and stop.
3. Use the node model with β = b(0) and evaluate ψ(β). If ψ(b(0)) ≥ 0, then set β = b(0) and stop.
4. Use the node model with β = b(k)+b(k)
and evaluate ψ(β). If ψ b(k)+b(k)
= 0, then set β = b(k)+b(k)
2
2
2
and stop.
5. If ψ b(k)+b(k)
< 0, then update:
2
b(k + 1)
=
b(k + 1)
=
b(k) + b(k)
;
2
b(k).
b(k + 1)
=
b(k);
b(k + 1)
=
b(k) + b(k)
.
2
Else, update:
6. Set k := k + 1 and return to step 4.
Here, b(k) represents the lower bound of the search interval at iteration k and b(k) the upper bound.
18
5.2
Split ratios for the separated managed lane with gated access
The configuration of a node with an offramp as one of the output links is simpler in the case of a separated
HOV lane, as shown in Figure 4. Here, traffic cannot directly go from the HOV lane to link 222, and, thus, we
have to deal only with the 2-input-2-output node. There is a caveat, however. Recall from Section 3.4 that
in the separated managed lane case we have destination-based traffic classes, and split ratios for destinationbased traffic are fixed.
Figure 4: A node with a GP link and an onramp as inputs, and a GP link and an offramp as outputs.
We shall make the following assumptions:
in
1. The total flow entering the offramp, fˆ222
, at any given time is known (from measurements) and is not
in
restricted by the offramp supply: fˆ222 < R222 .
2. All the flow coming from the onramp (link 111), if such flow exists, is directed toward the GP link 2.
c
c
In other words, β111,2
= 1 and β111,222
= 0, c = 1, . . . , C.
3. The demand Sic , i = 1, 111, c = 1, . . . , C, and supply R2 are given.
c
4. We denote the set of destination-based classes as D. The split ratios β1j
for c ∈ D are known. Let
c
the split ratios β1j = β for c ∈ {1, . . . , C} \ D, where β is to be determined (i.e., we assume all
non-destination-based classes exit at the same rate).
The first three assumptions here reproduce assumptions 1, 3 and 5 made for the full-access managed lane
case. Assumption 4 is a reminder that there is a portion of traffic flow that we cannot direct to or away from
the offramp, but we have to account for it.
Similarly to the full-access managed lane case, we define the function ψ(β):
X
X
c
c
in
f1,222
+
f1,222
− fˆ222
,
ψ(β) =
c∈D
(5.3)
c∈D
c
where f1,222
, c = 1, . . . , C are determined by the node model. The first term of the right-hand side of (5.3)
depends on β. As before, we assume ψ(β) is a monotonically increasing function. We look for the solution
of equation (5.2) on the interval [0, 1]. This solution exists iff ψ(0) ≤ 0 and ψ(1) ≥ 0. The algorithm for
finding β is the same as the one presented in the previous section, except that b(0) should be initialized to
0, and S11 is to be assumed 0.
5.3
An iterative full calibration process
For the purposes of the simulations presented in the following Section, we placed the iterative split ratio
identification methods of Sections 5.1 and 5.2 within a larger iterative loop for the remaining parameters.
The model calibration follows the flowchart shown in Figure 5.
19
Initial Input Data
•
•
•
•
•
Link fundamental
diagrams
Input demand
Special traffic
portion of input
demand
Initially guessed
off-ramp split ratios
Off-ramp demand 1
Run simulation in
normal mode to
compute split ratios
that distribute traffic
between GP and
managed lane links 2
Start
Run simulation in
off-ramp split ratio
solver mode to
compute off-ramp
split ratios
3
No
Off-ramp
flows match
off-ramp
demand?
Run simulation in
normal mode to
produce density,
flow, speed, VMT
and VHT
4
5
Yes
Tune Input Data
• Off-ramp demand
• Input demand
• Special traffic
portion of input
demand
7
No
Simulation
results
match
benchmark?
6
Yes
Stop
Figure 5: Calibration workflow.
1. We start by assembling the available measurement data. Fundamental diagrams are assumed to be
given. Mainline and onramp demand are specified per 5-minute periods together with the special vehicle
portion parameter indicating the fraction of the input demand that is able to access the managed lane.
Initially, we do not know offramp split ratios as they cannot be measured directly. Instead, we use
some arbitrary values to represent them and call these values “initially guessed offramp split ratios”.
Instead of the offramp split ratios, we have the flows directed to offramps, to which we refer to as
offramp demand.
2. We run our network simulation outlined in Section 4.2 for the entire simulation period. At this point,
in step 5 of the simulation, the a priori undefined split ratios between traffic in the GP and in the
managed lanes are assigned using a split ratio solver.
3. Using these newly-assigned split ratios, we run our network simulation again, only this time, instead
of using the initially guessed offramp split ratios, we compute them from the given offramp demand as
described in Sections 5.1 and 5.2. As a result of this step, we obtain new offramp split ratios.
4. Now we run the network simulation as we did originally, in step 2, only this time with new offramp
split ratios, and record the simulation results — density, flow, speed, as well as performance measures
such as vehicle miles traveled (VMT) and vehicle hours traveled (VHT).
5. Check if the resulting offramp flows match the offramp demand. If yes, proceed to step 6, otherwise,
repeat steps 2-5. In our experience (i.e., the case studies in the following Section), it takes the process
described in steps 2-5 no more than two iterations to converge.
20
6. Evaluate the simulation results:
• correctness of bottleneck locations and activation times;
• correctness of congestion extension at each bottleneck;
• correctness of VMT and VHT.
If the simulation results are satisfactory, stop. Otherwise, proceed to step 7.
7. Tune/correct input data in the order shown in block 7 of Figure 5.
6
6.1
Simulation Results
Full-access managed lane case study: Interstate 680 North
We consider a 26.8-mile stretch of I-680 North freeway in Contra Costa County, California, from postmile
30 to postmile 56.8, shown in Figure 6, as a test case for the full-access managed lane configuration. This
freeway’s managed lane is a high-occupancy-vehicle (HOV) lane, which allows entry to vehicles with two
or more passengers. This stretch contains two HOV lane segments whose beginning and ending points are
marked on the map. The first HOV segment is 12.3 miles long and will be converted to HOT in spring
2017 (Metropolitan Transportation Commission), and the second HOV segment is 4.5 miles long. There are
26 onramps and 24 offramps. The HOV lane is active from 5 to 9 AM and from 3 to 7 PM. The rest of the
time, the HOV lane is open to all traffic, and behaves as a GP lane.
Figure 6: Map of I-680 North in Contra Costa County.
To build the model, we used data collected for the I-680 Corridor System Management Plan (CSMP)
study (System Metrics Group, Inc., 2015). The bottleneck locations as well as their activation times and
congestion extension were identified in that study using video monitoring and tachometer vehicle runs. Onand offramp flows were given in 5-minute increments. Here, we assume that the HOV portion of the input
demand is 15%. The model was calibrated to a typical weekday, as suggested in the I-680 CSMP study.
For this simulation, we used the fundamental diagram described in Appendix A, with parameters as follows:
21
• The capacity of the ordinary GP lane is 1,900 vehicles per hour per lane (vphl);
• The capacity of the auxiliary GP lane is 1,900 vphl;
• The capacity of HOV lane is 1,800 vphl while active and 1,900 vphl when it behaves as a GP lane;
• The free flow speed varies between 63 and 70 mph — these measurements came partially from the
California Performance Measurement System (PeMS) (California Department of Tranportation, 2016)
and partially from tachometer vehicle runs.
• The congestion wave speed for each link was taken as 1/5 of the free flow speed.
The modeling results are presented in Figures 7, 8 and 9 showing density, flow and speed contours, respectively, in the GP and the HOV lanes. In each plot, the top contour corresponds to the HOV lanes, and the
bottom to the GP lanes. In all the plots traffic moves from left to right along the “Absolute Postmile” axis,
while the vertical axis represents time. Bottleneck locations and congestion areas identified by the I-680
CSMP study are marked by blue boxes in GP lane contours. The HOV lane does not get congested, but
there is a speed drop due to the friction effect. The friction effect, when vehicles in the HOV lane slow down
because of the slow moving GP lane traffic, can be seen in the HOV lane speed contour in Figure 9.
Figure 10 shows an example of how well the offramp flow computed by the simulation matches the target,
referred to as offramp demand, as recorded by the detector on the offramp at Crow Canyon Road. We can
see that in the beginning and in the end of the day, the computed flow falls below the target (corresponding
areas are marked with red circles). This is due to the shortage of the mainline traffic in the simulation —
the offramp demand cannot be satisfied.
Finally, Table 1 summarizes the performance measurements — vehicle miles traveled (VMT), vehicle hours
traveled (VHT) and delay in vehicle-hours — computed by simulation versus those collected in the course
of the I-680 CSMP study. Delay is computed for vehicles with speed below 45 mph.
GP Lane VMT
HOV Lane VMT
Total VMT
GP Lane VHT
HOV Lane VHT
Total VHT
GP Lane Delay
HOV Lane Delay
Total Delay
Simulation result
1,687,618
206,532
1,894,150
27,732
3,051
30,783
2,785
6
2,791
Collected data
1,888,885
31,008
2,904
Table 1: Performance measures for I-680 North.
22
Figure 7: I-680 North density contours for GP and HOV lanes produced by simulation. Density values are
given in vehicles per mile per lane. Blue boxes on the GP lane speed contour indicate congested areas as
identified by the I-680 CSMP study.
23
Figure 8: I-680 North flow contours for GP and HOV lanes produced by simulation. Flow values are give in
vehicles per hour per lane. Blue boxes on the GP lane speed contour indicate congested areas as identified
by the I-680 CSMP study.
24
Figure 9: I-680 North speed contours for GP and HOV lanes produced by simulation. Speed values are given
in miles per hour. Blue boxes on the GP lane speed contour indicate congested areas as identified by the
I-680 CSMP study.
25
Figure 10: Flow at the Crow Canyon Road offramp over 24 hours — collected (offramp demand) vs. computed
by simulation (offramp flow).
26
6.2
Gated-access managed lane case study: Interstate 210 East
We consider a 20.6-mile stretch of SR-134 East/ I-210 East in Los Angeles County, California, shown in
Figure 11, as a test case for the separated managed lane configuration. This freeway’s managed lane is also
an HOV lane. This freeway stretch consists of 3.9 miles of SR-134 East from postmile 9.46 to postmile 13.36,
which merges into 16.7 miles of I-210 East from postmile 25 to postmile 41.7. Gate locations where traffic
can switch between the GP and the HOV lanes are marked on the map. At this site, the HOV lane is always
active. There are 28 onramps and 25 offramps. The largest number of offramps between two gates is 5.
Thus, our freeway model has 7 vehicle classes - LOV, HOV and 5 destination-based.
Figure 11: Map of SR-134 East/ I-210 East freeway in Los Angeles County.
To build the model, we used PeMS data for the corresponding segments of the SR-134 East and I-210 East
for Monday, October 13, 2014 (California Department of Tranportation, 2016). Fundamental diagrams were
calibrated using PeMS data following the methodology of Dervisoglu et al. (2009). As in the I-680 North
example, we assume that HOV portion of the input demand is 15%.
The modeling results are presented in Figures 12, 13 and 14 showing density, flow and speed contours,
respectively, in the GP and the HOV lanes. In each plot, the top contour corresponds to the HOV lanes, and
the bottom to the GP lanes. As before, in all the plots traffic moves from left to right along the “Absolute
Postmile” axis, while the vertical axis represents time. The HOV lane does not get congested. Dashed blue
lines on the contour plots indicate HOV gate locations.
Figure 15 shows the PeMS speed contours for the SR-134 East/ I-210 East GP and HOV lanes that were
used as a target for our simulation model. In these plots, traffic also travels from left to right, with the
horizontal axis representing postmiles, while the vertical axis represents time.
Figure 16 shows an example of how well the offramp flow computed by the simulation matches the target,
referred to as offramp demand, as recorded by the detector on the offramp at North Hill Avenue. The
simulated offramp flow matches the offramp demand fairly closely. Similar results were found for the other
offramps.
Finally, Table 2 summarizes the performance measurements — VMT, VHT and delay — computed by
simulation versus those values obtained from PeMS. The PeMS data come from both SR-134 East and I-210
East, and VMT, VHT and delay values are computed as sums of the corresponding values from these two
freeway sections. Delay values are computed in vehicle-hours for those vehicles traveling slower than 45 mph.
27
GP Lane VMT
HOV Lane VMT
Total VMT
GP Lane VHT
HOV Lane VHT
Total VHT
GP Lane Delay
HOV Lane Delay
Total Delay
Simulation result
2,017,322
378,485
2,395,807
33,533
6,064
39,597
3,078
584
3,662
PeMS data
414,941 + 2,006,457 = 2,421,398
6,416 + 36,773 = 43,189
1 + 3,802 = 3,803
Table 2: Performance measures for SR-134 East/ I-210 East.
28
Figure 12: SR-134 East/ I-210 East density contours for GP and HOV lanes produced by simulation. Density
values are given in vehicles per mile per lane.
29
Figure 13: SR-134 East/ I-210 East flow contours for GP and HOV lanes produced by simulation. Flow
values are give in vehicles per hour per lane.
30
Figure 14: SR-134 East/ I-210 East speed contours for GP and HOV lanes produced by simulation. Speed
values are given in miles per hour.
31
Figure 15: SR-134 East/ I-210 East speed contours for GP and HOV lanes obtained from (California Department of Tranportation, 2016) for Monday, October 13, 2014. The horizontal axis represents absolute
postmile, and the vertical axis represents time in hours. The four contours share the same color scale.
32
Figure 16: Flow at the North Hill Avenue offramp over 24 hours — PeMS data (offramp demand) vs.
computed by simulation (offramp flow).
33
7
Conclusion
In this paper we discussed modeling procedures for two managed lane configurations: (1) full access, where
special traffic can switch between the GP and the managed lanes at any node; and (2) separated, where
special traffic can switch between the two lanes only at specific nodes, called gates. We have introduced the
friction effect (Section 3.3) and the inertia effect (Section 3.5). The friction effect reflects the empiricallyobserved drivers’ fear of moving fast in the managed lane while traffic in the adjacent GP links moves slowly
due to congestion. The inertia effect reflects drivers’ inclination to stay in their lane as long as possible and
switch only if this would obviously improve their travel condition.
The presence of interchanges in freeways with managed lanes produces modeling complications beyond what
one would see in a generic freeway model. For example, the separated managed lane requires a special
modeling trick for sending vehicles traveling in the managed lane to offramps, as locations of gates and
offramps generally do not coincide. In this paper we proposed such a mechanism. It employs destinationbased traffic classes, which are populated at modeling links approaching gates using offramp split ratios
(Section 3.4). This approach is a natural use of the multi-class feature of many traffic models.
Freeways with managed lanes feature many parameters, and calibrating them can be difficult. We presented
an iterative learning approach for estimating some of the harder-to-estimate parameters. Our simulation
results comparing our model and calibration results showed good agreement in two case studies, validating
both our full-access and gated-access modeling techniques. In the sequel to this paper, we will further extend
these results to include traffic control, with simulation of a reactive tolling controller on the managed lane.
Acknowledgement
We would like to express great appreciation to several of our colleagues. To Elena Dorogush and Ajith
Muralidharan for sharing ideas, and to Gabriel Gomes and Pravin Varaiya for their critical reading and their
help in clarifying some theoretical issues.
This research was funded by the California Department of Transportation.
A
Link Model
Figure 17: The “backwards lambda” fundamental diagram.
For the majority of this paper, we remain agnostic as to the particular functional relationship between density
nl , demand Sl (per commodity, ncl , Slc ) and supply Rl , and flow fl (also called the fundamental diagram)
34
used in our first-order macroscopic model (2.1). Where a particular fundamental diagram is required, i.e.
for the simulation results presented in Section 6 and the example implementation of the friction effect in
(3.4), (3.5), (3.6), and (3.7), we use a fundamental diagram from Horowitz et al. (2016), shown in Figure 17.
This fundamental diagram captures the traffic hysteresis behavior with the “backwards lambda” shape often
observed in detector data (Koshi et al., 1983):
)
(
C
X
Fl (t)
f
c
c
, Sl (t) =
Slc (t),
(A.1)
Sl (t) =vl (t)nl (t) min 1, f PC
c
vl (t) c=1 nl (t)
c=1
!
C
X
J
c
Rl (t) = (1 − θl (t)) Fl (t) + θl (t)wl (t) nl (t) −
nl (t) ,
(A.2)
c=1
where, for link l, Fl is the capacity, vlf is the free flow speed, wl is the congestion wave speed, nJl is the
jam density, and n−
l =
w l nJ
l
vlf +wl
and n+
l =
Fl
vlf
Fl , vlf ,
are called the low and high critical densities, respectively. As
written here and used in this paper,
and wl are in units per simulation timestep. The variable θl (t)
is a congestion metastate of l, which encodes the hysteresis:
nl (t) ≤ n−
0
l ,
(A.3)
θl (t) = 1
nl (t) > n+
l ,
−
+
θl (t − 1) nl < nl (t) ≤ nl ,
where nl (t) =
PC
c=1
ncl (t).
Examining (A.3) and (A.2), we see that when a link’s density goes above n+
l (i.e., when it becomes congested),
its ability to receive flow is reduced until the density falls below n−
.
l
An image of (A.1) and (A.2) overlaid on each other, giving a schematic image of the fundamental diagram,
+
is shown in Figure 17. Unless n−
shape, the fundamental diagram is not
l = nl , when it assumes triangular
+
a function of density alone (i.e., without θl (t)): nl (t) ∈ n−
,
n
admits
two possible flow values.
l
l
B
Dynamic Split Ratio Solver
Throughout this article, we have made reference to a dynamic-system-based method for solving for partiallyor fully-undefined split ratios from Wright et al. (2016a). This split ratio solver is designed to implicitly
solve the logit-based split ratio problem
P P
M
C
c
i=1
c=1 Sij
exp
Rj
c
PM PC c ,
βij
=
(B.1)
PN
i=1
c=1 Sij 0
exp
0
j =1
R 0
j
c
c
which cannot be solved explicitly, as the Sij
’s are also functions of the βij
’s. The problem (B.1) is chosen
to be a node-local problem that does not rely on information from the link model (beyond supplies and
demands), and is thus independent of the choice of link model (Wright et al., 2016a).
The solution algorithm is as follows, reproduced from Wright et al. (2016a). More discussion is available in
the reference.
c
• Define the set of commodity movements for which split ratios are known as B = {i, j, c} : βij
∈ [0, 1] ,
and the set of commodity movements for which split ratios are to be computed as B =
c
{i, j, c} : βij
are unknown .
35
• For a given input link i and commodity c such that Sic = 0, assume that all split ratios are known:
{i, j, c} ∈ B.3
• Define the set of output links for which there exist unknown split ratios as V = j : ∃ {i, j, c} ∈ B .
• Assuming that for a given input link i and
c, the split ratios must sum up to 1, define the
P commodity
c
c
unassigned portion of flow by β i = 1 − j:{i,j,c}∈B βij
.
• For a given input link i and commodity c such that there exists at least one commodity movement
c
{i, j, c} ∈ B, assume β i > 0, otherwise the undefined split ratios can be trivially set to 0.
• For every output link j ∈ V , define the set
of input links that have an unassigned demand portion
directed toward this output link by Uj = i : ∃ {i, j, c} ∈ B .
• For a given input link i and commodity c, define the set of output links for which split ratios for which
are to be computed as Vic = {j : ∃i ∈ Uj }, and assume that if nonempty, this set contains at least two
c
elements, otherwise a single split ratio can be trivially set equal to β i .
PM
• Assume that input link priorities are nonnegative, pi ≥ 0, i = 1, . . . , M , and i=1 pi = 1.
• Define the set of input links with zero priority: Uzp = {i : pi = 0}. To enable split ratio assignment
for inputs with zero priorities, perform regularization:
|Uzp |
1 |Uzp |
M − |Uzp | |Uzp |
p̃i = pi 1 −
+
= pi
+
,
(B.2)
M
M M
M
M2
where |Uzp | denotes the number of elements in set Uzp . Expression (B.2) implies that the regularized
input priority p̃i consists of two parts: (1) the original input priority pi normalized to the portion of
1
input links with positive priorities; and (2) uniform distribution among M input links, M
, normalized
to the portion of input links with zero priorities.
PM
Note that the regularized priorities p̃i > 0, i = 1, . . . , M , and i=1 p̃i = 1.
c
The algorithm for distributing β i among the commodity movements in B (that is, assigning values to the a
priori unknown split ratios) aims at maintaining output links as uniform in their demand-supply ratios as
possible. At each iteration k, two quantities are identified: µ+ (k), which is the largest oriented demandsupply ratio produced by the split ratios that have been assigned so far, and µ− (k), which is the smallest
oriented demand-supply ratio whose input link, denoted i− , still has some unclaimed split ratio. Once these
two quantities are found, the commodity c− in i− with the smallest unallocated demand has some of its
demand directed to the j corresponding to µ− (k) to bring µ− (k) up to µ+ (k) (or, if this is not possible due
to insufficient demand, all such demand is directed).
To summarize, in each iteration k, the algorithm attempts to bring the smallest oriented demand-supply
thatPall such oriented
ratio µ+ (k) up to the largest oriented demand-supply ratio µ− (k). If it turns out P
c
demand-supply ratios become perfectly balanced, then the demand-supply ratios ( i c Sij
)/Rj are as
well.
The algorithm is:
1. Initialize:
c
β̃ij
(0)
c
3 If
:=
c
βij
,
0,
if {i, j, c} ∈ B,
otherwise;
c
β i (0)
:= β i ;
Ũj (0)
=
Uj ;
Ṽ (0)
=
V;
k
:=
0,
split ratios were undefined in this case, they could be assigned arbitrarily.
36
Here Ũj (k) is the remaining set of input links with some unassigned demand, which may be directed
to output link j; and Ṽ (k) is the remaining set of output links, to which the still-unassigned demand
may be directed.
n
o
c
2. If Ṽ (k) = ∅, stop. The sought-for split ratios are β̃ij
(k) , i = 1, . . . , M , j = 1, . . . , N , c = 1, . . . , C.
3. Calculate the remaining unallocated demand:
c
c
S i (k) = β i (k)Sic ,
i = 1, . . . , M, c = 1, . . . , C.
4. For all input-output link pairs, calculate oriented demand:
c
c
S̃ij
(k) = β̃ij
(k)Sic .
5. For all input-output link pairs, calculate oriented priorities:
PC
p̃ij (k)
c=1
p̃i P
C
=
c c
Si
γij
c=1
(B.3)
Sic
with
(
c
γij
(k)
c
βij
,
=
c
β̃ij
(k)
if split ratio is defined a priori: {i, j, c} ∈ B,
c
β i (k)
|Vic | ,
+
otherwise,
(B.4)
where |Vic | denotes the number of elements in the set Vic . Examining the expression (B.3)-(B.4), one
c
can see that the split ratios β̃ij
(k), which are not fully defined yet, are complemented with a fraction
c
of β i (k) inversely proportional to the number of output links among which the flow of commodity c
from input link i can be distributed.
Note that in this step we are using regularized priorities p̃i as opposed to the original pi , i = 1, . . . , M .
This is done to ensure that inputs with pi = 0 are not ignored in the split ratio assignment.
6. Find the largest oriented demand-supply ratio:
PC
c
S̃ij
(k) X
p̃ij (k).
p̃ij (k)Rj
c=1
+
µ (k) = max max
j
i
i∈Uj
7. Define the set of all output links in Ṽ (k), where the minimum of the oriented demand-supply ratio is
achieved:
PC
c
X
c=1 S̃ij (k)
Y (k) = arg min min
p̃ij (k),
p̃ij (k)Rj
j∈Ṽ (k) i∈Ũj (k)
i∈Uj
and from this set pick the output link j − with the smallest output demand-supply ratio (when there
are multiple minimizing output links, any of the minimizing output links may be chosen as j − ):
j
−
PM PC
= arg min
i=1
c=1
c
S̃ij
(k)
Rj
j∈Y (k)
.
8. Define the set of all input links, where the minimum of the oriented demand-supply ratio for the output
link j − is achieved:
PC
W
j−
(k) = arg
min
i∈Ũj − (k)
c=1
c
S̃ij
− (k) X
p̃ij − (k)Rj −
p̃ij − (k),
i∈Uj −
37
and from this set pick the input link i− and commodity c− with the smallest remaining unallocated
demand:
{i− , c− } = arg
c
S i (k).
min
i ∈ Wj − (k),
c
c : β i− (k) > 0
9. Define the smallest oriented demand-supply ratio:
PC
c
X
c=1 S̃i− j − (k)
−
p̃ij− (k).
µ (k) =
p̃i− j − (k)Rj −
i∈Uj −
• If µ− (k) = µ+ (k), the oriented demands created by the split ratios that have been assigned as
c
of iteration k, β̃ij
(k), are perfectly balanced among the output links, and to maintain this, all
remaining unassigned split ratios should be distributed proportionally to the allocated supply:
c
β̃ij
(k + 1)
c
β i (k
=
Rj
c
β̃ij
(k) + P
j 0 ∈Vic (k)
c
β i (k) > 0,
+ 1)
=
0,
c:
Ũj (k + 1)
=
∅,
j ∈ Ṽ (k);
Ṽ (k + 1)
=
∅.
c
Rj 0
β i (k),
c
c : β i (k) > 0, i ∈ Ũj (k), j ∈ Ṽ (k); (B.5)
i ∈ Ũj (k), j ∈ Ṽ (k);
If the algorithm ends up at this point, we have emptied Ṽ (k + 1) and are done.
• Else, assign:
−
∆β̃ic− j − (k)
=
PC
c
−
S̃
(k)
−
−
µ+ (k)p̃i− j − (k)Rj −
c
c=1 i j
min β i− (k), c−
−
;
P
c−
S i− (k) i∈U − p̃ij − (k)
S i− (k)
(B.6)
=
−
β̃ic− j − (k)
(B.7)
=
c−
β i− (k)
j
−
β̃ic− j − (k
c−
β i− (k
+ 1)
+ 1)
+
−
∆β̃ic− j − (k);
−
+ 1)
=
+ 1)
=
Ũj (k + 1)
=
c
β̃ij
(k) for {i, j, c} =
6 {i− , j − , c− };
c
β i (k) for {i, c} =
6 {i− , c− };
n
o
c
Ũj (k) \ i : β i (k + 1) = 0, c = 1, . . . , C ,
=
n
o
Ṽ (k) \ j : Ũj (k + 1) = ∅ .
c
β̃ij
(k
c
β i (k
Ṽ (k + 1)
(B.8)
− ∆β̃ic− j − (k);
j ∈ Ṽ (k);
c−
In (B.6), we take the minimum of the remaining unassigned split ratio portion β i− (k) and the
split ratio portion needed to equalize µ− (k) and µ+ (k). To better understand the latter, the
second term in min{·, ·} can be rewritten as:
!
PC
+
X
C
c
µ+ (k)p̃i− j − (k)Rj −
µ (k)
1
c=1 S̃i− j − (k)
c
−
=
.
−1
S̃i− j − (k)
−
P
−
c−
c−
c
µ (k)
S i− (k) i∈U − p̃ij − (k)
S i− (k)
S i− (k)
c=1
j
The right hand side of the last equality can be interpreted as: flow that must be assigned for
input i− , output j − and commodity c− to equalize µ− (k) and µ+ (k) minus flow that is already
assigned for {i− , j − , c− }, divided by the remaining unassigned portion of demand of commodity
c− coming from input link i− .
In (B.7) and (B.8), the assigned split ratio portion is incremented and the unassigned split ratio
−
portion is decremented by the computed ∆β̃ic− j − (k).
10. Set k := k + 1 and return to step 2.
38
References
M. Bliemer. Dynamic queueing and spillback in an analytical multiclass dynamic network loading model.
Transportation Research Record, 2029:14–21, 2007.
California Department of Tranportation. PeMS homepage, 2016. http://pems.dot.ca.gov.
M. J. Cassidy, K. Kim, W. Ni, and W. Gu. A problem of limited-access special lanes. Part I: Spatiotemporal
studies of real freeway traffic. Transportation Research Part A: Policy and Practice, 80:307 – 319, 2015.
ISSN 0965-8564. doi: http://dx.doi.org/10.1016/j.tra.2015.07.001. URL http://www.sciencedirect.
com/science/article/pii/S0965856415001834.
M. Chang, J. Wiegmann, A. Smith, and C. Bilotto. A Review of HOV Lane Performance and Policy Options
in the United States. Report FHWA-HOP-09-029, Federal Highway Administration, 2008.
C. F. Daganzo. A behavioral theory of multi-lane traffic flow. Part i: Long homogeneous freeway sections.
Transportation Research Part B: Methodological, 36(2):131 – 158, 2002. ISSN 0191-2615. doi: http://
dx.doi.org/10.1016/S0191-2615(00)00042-4. URL http://www.sciencedirect.com/science/article/
pii/S0191261500000424.
C. F. Daganzo and M. J. Cassidy. Effects of high occupancy vehicle lanes on freeway congestion. Transportation Research Part B: Methodological, 42(10):861–872, Dec. 2008. ISSN 01912615. doi: 10.1016/j.
trb.2008.03.002. URL http://linkinghub.elsevier.com/retrieve/pii/S0191261508000325.
G. Dervisoglu, G. Gomes, J. Kwon, A. Muralidharan, P. Varaiya, and R. Horowitz. Automatic calibration of
the fundamental diagram and empirical obsevations on capacity. 88th Annual Meeting of the Transportation
Research Board, Washington, D.C., USA, 2009.
N. Farhi, H. Haj-Salem, M. Khoshyaran, J.-P. Lebacque, F. Salvarani, B. Schnetzler, and F. De Vuyst. The
Logit lane assignment model: first results. In TRB 92nd Annual Meeting Compendium of Papers, 2013.
M. Fransson and M. Sandin. Framework for Calibration of a Traffic State Space Model. Masters Thesis,
Linkoping University, Norrkoping, Sweden, Oct. 2012.
S. Hoogendoorn and P. Bovy. Gas-Kinetic Model for Multilane Heterogeneous Traffic Flow. Transportation Research Record: Journal of the Transportation Research Board, 1678:150–159, 1999. doi:
10.3141/1678-19.
S. P. Hoogendoorn and P. H. Bovy. Continuum modeling of multiclass traffic flow. Transportation Research Part B: Methodological, 34(2):123 – 146, 2000. ISSN 0191-2615. doi: http://dx.doi.
org/10.1016/S0191-2615(99)00017-X. URL http://www.sciencedirect.com/science/article/pii/
S019126159900017X.
R. Horowitz, A. A. Kurzhanskiy, A. Siddiqui, and M. A. Wright. Modeling and control of hot lanes. Technical
report, Partners for Advanced Transportation Technologies, University of California, Berkeley, Berkeley,
CA, 2016.
K. Jang and M. Cassidy. Dual influences on vehicle speed in special-use lanes and critique of US regulation.
Transportation Research, Part A, 46(2012):1108–1123, 2012.
K. Jang, S. Oum, and C.-Y. Chan. Traffic Characteristics of High-Occupancy Vehicle Facilities: Comparison
of Contiguous and Buffer-Separated Lanes. Transportation Research Record: Journal of the Transportation
Research Board, 2278:180–193, 2012. doi: 10.3141/2278-20.
M. Koshi, M. Iwasaki, and I. Ohkura. Some findings and an overview on vehicular flow characteristics. In
Proceedings of the 8th International Symposium on Transportation and Traffic Flow Theory, volume 198,
pages 403–426, 1983.
A. A. Kurzhanskiy and P. Varaiya. Traffic management: An outlook. Economics of Transportation, 4(3):
135–146, Sept. 2015. ISSN 22120122. doi: 10.1016/j.ecotra.2015.03.002.
39
X. Liu, B. Schroeder, T. Thomson, Y. Wang, N. Rouphail, and Y. Yin. Analysis of operational interactions
between freeway managed lanes and parallel, general purpose lanes. Transportation Research Record:
Journal of the Transportation Research Board, 2262:62–73, 2011.
X. Liu, G. Zhang, Y. Lao, and Y. Wang. Modeling Traffic Flow Dynamics on Managed Lane Facility: Approach Based on Cell Transmission Model. Transportation Research Record: Journal of the Transportation
Research Board, 2278:163–170, 2012.
T. Lomax, D. Schrank, and B. Eisele. The 2015 Annual Urban Mobility Scorecard. Technical report, Texas
Transportation Institute, 2015. http://mobility.tamu.edu.
D. McFadden. Conditional Logit Analysis of Qualitative Choice Behavior. In P. Zarembka, editor, Frontiers
in Econometrics, pages 105–142. Academic Press, New York, 1973.
Metropolitan Transportation Commission. Bay Area Express Lanes. http://bayareaexpresslanes.org.
D. Ngoduy. Macroscopic discontinuity modeling for multiclass multilane traffic flow operations. PhD thesis,
Delft University of Technology, 2006.
D. Ngoduy and M. Maher. Calibration of second order traffic models using continuous cross entropy method.
Transportation Research Part C: Emerging Technologies, 24:102–121, Oct. 2012. ISSN 0968090X. doi:
10.1016/j.trc.2012.02.007.
J. Obenberger. Managed lanes: Combining access control, vehicle eligibility, and pricing strategies can help
mitigate congestion and improve mobility on the nationâĂŹs busiest roadways. Public Roads, 68(3):48–55,
2004. URL http://www.fhwa.dot.gov/publications/publicroads/04nov/08.cfm.
A. Poole and A. Kotsialos. METANET Model Validation using a Genetic Algorithm. IFAC Proceedings
Volumes, 45(24):7–12, Sept. 2012. ISSN 14746670. doi: 10.3182/20120912-3-BG-2031.00002.
A. Poole and A. Kotsialos. Second order macroscopic traffic flow model validation using automatic differentiation with resilient backpropagation and particle swarm optimisation algorithms. Transportation Research
Part C: Emerging Technologies, 71:356–381, Oct. 2016. ISSN 0968090X. doi: 10.1016/j.trc.2016.07.008.
URL http://linkinghub.elsevier.com/retrieve/pii/S0968090X16301152.
Y. Shiomi, T. Taniguchi, N. Uno, H. Shimamoto, and T. Nakamura. Multilane first-order traffic flow
model with endogenous representation of lane-flow equilibrium. Transportation Research Part C: Emerging
Technologies, 59:198–215, Oct. 2015. ISSN 0968090X. doi: 10.1016/j.trc.2015.07.002. URL http://
linkinghub.elsevier.com/retrieve/pii/S0968090X15002429.
System Metrics Group, Inc. Contra Costa County I-680 Corridor System Management Plan Final Report.
Technical report, Caltrans District 4, 2015. http://dot.ca.gov/hq/tpp/corridor-mobility/CSMPs/
d4_CSMPs/D04_I680_CSMP_Final_Revised_Report_2015-05-29.pdf.
M. Treiber, A. Hennecke, and D. Helbing. Derivation, properties, and simulation of a gas-kinetic-based,
nonlocal traffic model. Phys. Rev. E, 59:239–253, Jan 1999. doi: 10.1103/PhysRevE.59.239. URL http:
//link.aps.org/doi/10.1103/PhysRevE.59.239.
J. van Lint, S. Hoogendoorn, and M. Schreuder. Fastlane: New Multiclass First-Order Traffic Flow Model.
Transportation Research Record: Journal of the Transportation Research Board, 2088:177–187, 2008. doi:
10.3141/2088-19.
G. Wong and S. Wong. A multi-class traffic flow model - an extension of LWR model with heterogeneous drivers. Transportation Research Part A: Policy and Practice, 36(9):827 – 841, 2002. ISSN 09658564. doi: http://dx.doi.org/10.1016/S0965-8564(01)00042-8. URL http://www.sciencedirect.com/
science/article/pii/S0965856401000428.
M. A. Wright, G. Gomes, R. Horowitz, and A. A. Kurzhanskiy. On node and route choice models for
high-dimensional road networks. Submitted to Transportation Research Part B, 2016a. Online: http:
//arxiv.org/abs/1601.01054.
40
M. A. Wright, R. Horowitz, and A. A. Kurzhanskiy. A dynamic system characterization of road network
node models. In Proceedings of the 10th IFAC Symposium on Nonlinear Control Systems, pages 1053–1058,
August 2016b. Online: http://arxiv.org/abs/1608.07623.
41
| 3cs.SY
|
A Verified Compiler for
Probability Density Functions
Manuel Eberl, Johannes Hölzl, and Tobias Nipkow
arXiv:1707.06901v1 [cs.PL] 21 Jul 2017
Technische Universität München
Bhat et al. [4] developed an inductive compiler that computes density
functions for probability spaces described by programs in a simple probabilistic functional language. In this work, we implement such a compiler for
a modified version of this language within the theorem prover Isabelle and
give a formal proof of its soundness w. r. t. the semantics of the source and
target language. Together with Isabelle’s code generation for inductive predicates, this yields a fully verified, executable density compiler. The proof is
done in two steps, using a standard refinement approach: first, an abstract
compiler working with abstract functions modelled directly in the theorem
prover’s logic is defined and proven sound. Then, this compiler is refined
to a concrete version that returns a target-language expression.
1. Introduction
1.1. Motivation
Random distributions of practical significance can often be expressed as probabilistic
functional programs. When studying a random distribution, it is often desirable to
determine its probability density function (PDF). This can be used to e. g. determine the
expectation or sample the distribution with a sampling method such as Markov-chain
Monte Carlo (MCMC).
In 2013, Bhat et al. presented a compiler that computes the probability distribution
function of a program in the probabilistic functional language Fun [4]. They evaluated the compiler on a number of practical problems and concluded that it reduces the
amount of time and effort required to model them in an MCMC system significantly
compared to hand-written models.
Bhat et al. also stated that their eventual goal is the formal verification of such a compiler in a theorem prover [3]. This has the advantage of providing guaranteed correctness,
i. e. the result of the compilation is provably a PDF for the source expression, according
to the formal semantics. This greatly increases the confidence one has in the compiler.
The final publication is available at Springer via https://doi.org/10.1007/978-3-662-46669-8_4.
1
In this Master’s thesis, we implemented such a compiler for a similar probabilistic
functional language in the interactive theorem prover Isabelle/HOL and formally proved
its correctness.
1.2. Related work
This work is based on the work of Bhat et al. [3, 4], which presents a density compiler for probability spaces described by expressions in the language Fun. This is a
small functional language with basic arithmetic, Boolean logic, product and sum types,
conditionals, and a number of built-in discrete and continuous distributions. It does,
however, not support lists or recursion. The correctness proof is purely pen and paper,
but the authors stated that formalisation in a proof assistant such as Coq is the ultimate
goal [3].
Park et al. [13] developed a probabilistic extension to Objective CAML called λ .
Their work focuses on sample generation, i. e. using an infinite stream of random numbers to compute samples of the distribution described by a probabilistic program. While
Bhat et al. generate density functions of functional programs, Park et al. generate
sampling functions. These are functions that map [0; 1]∞ to the sample space. The
sampling function effectively uses a finite number of uniformly-distributed random
numbers from the interval [0; 1] and returns a sample of the desired random variable. This approach allows them to handle much more general distributions, even
recursively-defined ones and distributions that do not have a density function, but it
does not allow precise reasoning about these distributions (such as determining the
exact expectation). No attempt at formal verification is made.
pGCL, another probabilistic language, has been formalised with a proof assistant –
first in HOL by Hurd et al. [12], then in Isabelle/HOL by David Cock [6, 7]. pGCL
contains a rich set of language features, such as recursion and probabilistic and nondeterministic choice. David Cock formally proved a large number of results about
the semantics of the language and developed mechanisms for refinement and program
verification; however, the focus of this work was verification of probabilistic programs,
not compiling them to density functions. Bhat et al. mention that reconciling recursion with probability density functions is difficult [3], so a feature-rich language such
as pGCL is probably not suited for developing a density compiler.
Audebaud and Paulin-Mohring [1] also implemented a probabilistic functional language with recursion in a theorem prover (namely Coq). Their focus was also on verification of probabilistic programs. While Cock uses a shallow embedding in which any
type and operation of the surrounding logic of the theorem prover (i. e. HOL) can be
used, Audebaud and Paulin-Mohring use a deep embedding with a restricted type system, expression structure, predefined primitive functions, etc. Like Bhat et al. and we,
they also explicitly reference the Giry monad as a basis for their work.
2
1.3. Utilised tools
As stated before, we work with the interactive theorem prover Isabelle/HOL. Isabelle
is a generic proof assistant that supports a number of logical frameworks (object logics)
such as Higher-Order Logic (HOL) or Zermelo-Fraenkel set theory (ZF). The most widely
used instance is Isabelle/HOL, which is what we use.
We heavily rely on Johannes Hölzl’s Isabelle formalisation of measure theory [11],
which is already part of the Isabelle/HOL library. We also use Sudeep Kanav’s formalisation of the Gaussian distribution and a number of libraries from the proof of the
Central Limit Theorem by Avigad et al. [2], namely the notion of interval and set integrals and the Fundamental Theorem of Calculus for these integrals. All of these have
been or will be moved to the Isabelle/HOL library by their respective maintainers as
well.
1.4. Outline
In Sect. 2.1, we will explain the notation we will use and then give a brief overview of
the mathematical basics we require – in particular the Giry Monad – in Sect. 2.2.
Section 3 contains the definition and the semantics of the source and target language.
Section 4 defines the abstract compiler and gives a high-level outline of the soundness
proof. Section 5 then explains the refinement of the abstract compiler to the concrete
compiler and the final correctness result and evaluates the compiler on a simple example.
Section 6 gives an overview of how much work went into the different parts of the
project and what difficulties we encountered. It also lists possible improvements and
other future work before summarising the results we obtained. Finally, the appendix
contains a table of all notation and auxiliary functions used in this work for reference.
2. Preliminaries
2.1. Notation
In the following, we will explain the conventions and the notation we will use.
2.1.1. Typographical notes
We will use the following typographical conventions in mathematical formulæ:
• Constants, functions, datatype constructors, and types will be typeset in regular
roman font: max, return, π, etc.0
• Free and bound variables (including type variables) are in italics: x, A, dst, etc.
0 In formulæ embedded in regular text, we will often set types and constants
from the surrounding text.
3
in italics to distinguish them
• Isabelle keywords are in bold print: lemma, datatype, primrec, etc.
• σ-algebras are typeset in calligraphic font: A, B , M, etc.
• Categories are typeset in Fraktur: Set, Grp, Meas, etc.
• File names of Isabelle theories are set in a monospaced font: PDF_Compiler.thy.
2.1.2. Deviations from standard mathematical notation
In order to maintain coherence with the notation used in Isabelle, we will deviate from
standard mathematical notation in the following ways:
• As is convention in functional programming (and therefore in Isabelle), function
application is often written as f x instead of f ( x) . It is the operation that binds
strongest and it associates to the left, i. e. f x + 1 is ( f x) + 1 and f x y is ( f x) y .
• The integral over integrand f with variable x and the measure µ is written as
Z
x. f x ∂µ
instead of
Z
f ( x) dµ( x) .
• The special notion of a non-negative integral1 is used; this integral ‘ignores’ the
negative part of the integrand. Effectively:
Z +
x. f x ∂µ
Z
=
ˆ
max(0, f ( x)) dµ( x)
• Lambda abstractions, such as λx. x2 , are used to specify functions. The corresponding standard mathematical notation would be x 7→ x2 .
Table 1 lists a number of additional mathematical notation.
2.1.3. Semantics notation
We will always use Γ to denote a type environment, i. e. a function from variable names
to types, and σ to denote a state, i. e. a function from variable names to values. Note
that variable names are always natural numbers as we use de Bruijn indices. The letter
Υ (Upsilon) will be used to denote density contexts, which are used in the compiler.
We also employ the notation t • Γ resp. v • σ to denote the insertion of a new variable
with the type t (resp. value v) into a typing environment (resp. state). This is used when
entering the scope of a bound variable; the newly inserted variable then has index 0 and
all other variables are incremented by 1. The precise definition is as follows:2
1 This
is a very central concept in the measure theory library in Isabelle. We will mostly use it with
non-negative functions anyway, so the distinction is purely formal.
2 Note the analogy to the notation x # xs for prepending an element to a list. This is because contexts with
de Bruijn indices are generally represented by lists, where the n-th element is the entry for the variable
n, and inserting a value for a newly bound variable is then simply the prepending operation.
4
Table 1: General notation
N OTATION
D ESCRIPTION
D EFINITION
f ‘X
undefined
image set
arbitrary value
f ( X ) or { f ( x ) | x ∈ X }
merge V V ′ (ρ, σ )
merging disjoint states
h Pi
density M f
distr M N f
indicator function
measure with density
push-forward/image measure
(y • f )( x) =
(
y
f ( x − 1)
if x ∈ V
ρ x
σy
if x ∈ V ′
undefined otherwise
1 if P is true, 0 otherwise
result of applying density f to M
( B, B , λX. µ( f −1 ( X ))) for
M = ( A, A, µ), N = ( B, B , µ′ )
if x = 0
otherwise
We use the same notation for inserting a new variable into a set of variables, shifting
all other variables, i. e.:
x • V = { x } ∪ {y + 1 | y ∈ V }
In general, the notation Γ ⊢ e : t and variations thereof will always mean ‘The
expression e has type t in the type environment Γ’, whereas Υ ⊢ e ⇒ f and variations
thereof mean ‘The expression e compiles to f under the context Υ’.
2.2. Mathematical basics
The category theory part of this section is based mainly on a presentation by ErnstErich Doberkat [9]. For a more detailed introduction, see his textbook [8] or the original
paper by Michèle Giry [10].
2.2.1. Basic measure spaces.
There are two basic measure spaces we will use:
Counting space. The counting space on a countable set A has
• the carrier set A
• the measurable sets P ( A), i. e. all subsets of A
• the measure λX. | X |, i. e. the measure of a subset of A is simply the number of
elements in that set (which can be ∞)
5
Borel space. The Borel space is a measure space on the real numbers which has
• the carrier set R
• the Borel σ-algebra as measurable sets, i. e. the smallest σ-algebra containing all
open subsets of R
• the Borel measure as a measure, i. e. the uniquely defined measure that maps real
intervals to their lengths
2.2.2. Sub-probability spaces.
A sub-probability space is a measurable space ( A, A) with a measure µ such that every
set X ∈ A has a measure ≤ 1, or, equivalently, µ( A) ≤ 1 .
For technical reasons, we also assume A 6= ∅ . This is required later in order to define
the bind operation in the Giry monad in a convenient way within Isabelle. This nonemptiness condition will always be trivially satisfied by all the measure spaces used in
this work.
2.2.3. The category Meas.
Note that:
• For any measurable space ( A, A), the identity function is A-A-measurable.
• For any measurable spaces ( A, A), ( B, B), (C, C), an A-B -measurable function f ,
and a B -C -measurable function g, the function g ◦ f is A-C -measurable.
Therefore, measurable spaces form a category Meas where:
• the objects of the category are measurable spaces
• the morphisms of the category are measurable functions
• the identity morphism 1( A,A) is the identity function id A : A → A, λx. x
• morphism composition is function composition
2.2.4. Kernel space.
The kernel space S ( A, A) of a measurable space ( A, A) is the natural measurable space
over the measures over ( A, A) with a certain property. For our purposes, this property
will be that they are sub-probability measures (as defined in Sect. 2.2.2).
Additionally, a natural property the kernel space should satisfy is that measuring a
fixed set X ∈ A while varying the measure within S ( A, A) should be a Borel-measurable
function; formally:
For all X ∈ A, f : S ( A, A) → R, λµ. µ( X ) is S ( A, A)-Borel-measurable
6
We can now simply define S ( A, A) as the smallest measurable space with the carrier
set
M := {µ | µ is a measure on ( A, A), µ( A) ≤ 1}
that fulfils this property, i. e. we let S ( A, A) := ( M, M) with
M := {(λµ. µ( X ))−1 (Y ) | X ∈ A, Y ∈ B}
where B is the Borel σ-algebra on R.
Additionally, for a measurable function f , we define S ( f ) = λµ. f (µ), where f (µ)
denotes the push-forward measure (or image measure)3 . Then S maps objects of Meas
to objects of Meas and morphisms of Meas to morphisms of Meas. We can thus see that
S is an endofunctor in the category Meas, as (id( A,A) )(µ) = µ and ( f ◦ g)(µ) = f ( g(µ)) .
2.2.5. Giry monad.
The Giry monad naturally captures the notion of choosing a value according to a (sub-)probability distribution, using it as a parameter for another distribution, and observing the
result.
Consequently, return (or η) yields a Dirac measure, i. e. a probability measure in
which all the ‘probability’ lies in a single element, and bind (or ≫=) integrates over
all the input values to compute one single output measure. Formally, for measurable
spaces ( A, A) and ( B, B), a measure µ on ( A, A), a value x ∈ A, and an A-S ( B, B)measurable function f :
(
Z
1 if x ∈ X
µ ≫= f := λX. x. f ( x)( X ) ∂µ
return( A,A) x := λX.
0 otherwise
Unfortunately, restrictions due to Isabelle’s type system require us to determine the
σ-algebra of the resulting measurable space for bind M f since this information cannot
be provided by the type of f . This can be done with an additional parameter, but it is
more convenient to define bind in such a way that it chooses an arbitrary value x ∈ M
and takes the σ-algebra of f x (or the count space on ∅ if M is empty)4 .
This choice is somewhat non-standard, but the difference is of no practical significance as we will not use bind on empty measure spaces.
To simplify the proofs for bind, we instead define the join operation (also known as
µ in category theory) first and use it to then define bind. The join operation ‘flattens’
objects, i. e. it maps an element of S (S ( A, A)) to one of S ( A, A) . Such an operation can
be naturally defined as:
Z
µ′ . µ′ ( X ) ∂µ
join µ = λX.
Note that in Isabelle, join has an additional explicit parameter for the measurable
space of the result to avoid the problem we had with bind. This makes expressions
= λX. µ ( f −1 ( X )) .
that for any A − S ( B, B)-measurable function f , the σ-algebra thus obtained is independent of
which value is chosen, by definition of the kernel space.
3 f (µ )
4 Note
7
containing join more complicated; this is, however, justified by the easier proofs and
will be unproblematic later since we will never use join directly, only bind.
Now, bind can be defined using join in the following way, modulo handling of empty
measure spaces5 :
µ ≫= f = join (S ( f )(µ)) = join ( f (µ))
2.2.6. The ‘do’ syntax.
For better readability, we employ a Haskell-style ‘do notation’ for operations in the
Giry monad. The syntax of this notation is defined recursively, where M stands for a
monadic expression and h patterni stands for arbitrary ‘raw’ text:
do { M } =
ˆ M
do { x ← M; h patterni} =
ˆ M ≫= (λx. do {h pattern i})
3. Language Syntax and Semantics
The source language used in the formalisation was modelled after the language Fun
described by Bhat et al. [4]; similarly, the target language is almost identical to the
target language used by Bhat et al. However, we have made the following changes in
our languages:
• Variables are represented by de Bruijn indices to avoid handling freshness, captureavoiding substitution, and related problems.
• No sum types are supported. Consequently, the match . . . with . . . command is
replaced with an IF . . . THEN . . . ELSE . . . command. Furthermore, booleans are
a primitive type rather than represented as unit + unit .
• The type double is called real and it represents a real number with absolute precision as opposed to an IEEE 754 floating point number.
• Beta and Gamma distributions are not included.
In the following sections, we give the precise syntax, typing rules, and semantics of
both our source language and our target language.
3.1. Types, values, and operators
The source language and the target language share the same type system and the same
operators. Figure 1 shows the types and values that exist in our languages.6 Additionally, standard arithmetical and logical operators exist.
5 S ( f ),
i. e. lifting a function on values to a function on measure spaces, is done by the distr function
in Isabelle. This function was already defined in the library, which simplifies our proofs about bind,
seeing as some of the proof work has already been done in the proofs about distr.
6 Note that bool, int, and real stand for the respective Isabelle types, whereas B, Z, and R stand for the
source-/target-language types.
8
datatype pdf_type =
UNIT | B | Z | R | pdf_type × pdf_type
datatype val =
UnitVal | BoolVal bool | IntVal int | RealVal real | <| val, val |>
datatype pdf_operator =
Fst | Snd | Add | Mult | Minus | Less | Equals | And | Or | Not | Pow |
Fact | Sqrt | Exp | Ln | Inverse | Pi | Cast pdf_type
Figure 1: The types and values in the source and target language
All operators are total, meaning that for every input value of their parameter type,
they return a single value of their result type. This requires some non-standard definitions for non-total operations such as division, the logarithm, and the square root – we
define them to be 0 outside their domain. Non-totality could also be handled by implementing operators in the Giry monad by letting them return either a Dirac distribution
with a single result or, when evaluated for a parameter on which they are not defined,
the null measure. This, however, would probably complicate many proofs significantly.
To increase readability, we will use the following abbreviations:
• TRUE and FALSE stand for BoolVal True and BoolVal False, respectively.
• RealVal, IntVal, etc. will be omitted in expressions when their presence is implicitly
clear from the context.
• a − b stands for a + (−b) and a/b for a · b−1 .
3.2. Auxiliary definitions
A number of auxiliary definitions are used in the definition of the semantics; for a full
list of auxiliary functions see table 2. The following two notions require a detailed
explanation:
3.2.1. Measure embeddings.
A measure embedding is the measure space obtained by ‘tagging’ values in a measure
space M with some injective function f (in fact, f will always be a datatype constructor).
For instance, a set of values of type R can naturally be measured by stripping away the
RealVal constructor and using a measure on real numbers (e. g. the Lebesgue-Borel
measure) on the resulting set of reals. Formally:
embed_measure ( A, A, µ) f = f ( A), { f ( X ) | X ∈ A}, λX. µ( f −1 ( X ) ∩ A)
9
3.2.2. Stock measures.
The stock measure for a type t is the ‘natural’ measure on values of that type. This is
defined as follows:
• For the countable types UNIT, B, Z: the count measure over the corresponding
type universes
• For type R: the embedding of the Lebesgue-Borel measure on R with RealVal
• For t1 × t2 : the embedding of the product measure
stock_measure t1
with λ(v, w). <| v, w |>
N
stock_measure t2
R
Note that
R in order to save space and increase readability, we will often write x. f x ∂t
instead of x. f x ∂stock_measure t in integrals.
3.2.3. The state measure.
Using the stock measure, we can also construct a measure on states in the context of a
typing environment Γ. A state on the variables V is a function that maps a variable in
V to a value. A state σ is well-formed w. r. t. to V and Γ if it maps every variable x ∈ V to
a value of type Γ x and every variable ∈
/ V to undefined.
We now fix Γ and a finite V and consider the set of well-formed states w. r. t. V and
Γ. Another representation of these states are tuples in which the i-th component is the
value of the i-th variable in V. The natural measure that can be given to such tuples is
then the finite product measure of the stock measures of the types of the variables:
state_measure Γ V :=
3.3. Source language
O
stock_measure (Γ x)
x ∈V
Figures 2 and 3 show the syntax resp. the typing rules of the source language. Figure
4 defines the source language semantics as a primitively recursive function. Similarly
to the abbreviations mentioned in Sect. 3.1, we will omit Val when its presence is implicitly obvious from the context; e. g. if in some context, e is an expression and c is a
constant real number, we will write e + Val (RealVal c) as e + c .
3.3.1. Built-in distributions.
Our language contains the built-in distributions Bernoulli, UniformInt, UniformReal, Gaussian, and Poisson. The uniform distributions are parametrised with a pair of numbers
( a, b) and return a uniform distribution over the interval [a; b]. The Gaussian distribution is parametrised with a pair of real numbers (µ, σ), i. e. mean and standard deviation.
10
Table 2: Auxiliary functions
F UNCTION
D ESCRIPTION
op_sem op v
op_type op t
dist_param_type dst
dist_result_type dst
dist_measure dst x
dist_dens dst x y
type_of Γ e
val_type v
type_universe t
countable_type t
free_vars e
e det
extract_real x
return_val v
null_measure M
semantics of operator op applied to v
result type of operator op for input type t
parameter type of the built-in distribution dst
result type of the built-in distribution dst
built-in distribution dst with parameter x
density of the built-in distribution dst w. parameter x at value y
the unique t such that Γ ⊢ e : t
the type of value v, e. g. val_type (IntVal 42) = INTEG
the set of values of type t
true iff type_universe t is a countable set
the free (i. e. non-bound) variables in the expression e
true iff e does not contain Random or Fail
returns y for x = RealVal y (analogous for int, pair, etc.)
return (stock_measure (val_type v)) v
measure with same measurable space as M, but 0 for all sets
datatype expr =
Var nat | Val val | LET expr IN expr | pdf_operator $ expr | < expr, expr> |
Random pdf_dist | IF expr THEN expr ELSE expr | Fail pdf_type
Figure 2: The source language syntax
For invalid parameters, i. e. Bernoulli 2 or UniformReal 3 2, the built-in distributions
return the null measure.
3.4. Deterministic expressions
We call an expression e deterministic (written as ‘e det’) if it contains no occurrence of
Random or Fail. Such expressions are of particular interest: if all their free variables have
a fixed value, they return precisely one value, so we can define a function expr_sem_rf 7
that, when given a state σ and a deterministic expression e, returns this single value.
7 In Isabelle, the expression randomfree is used instead of deterministic, hence
the ‘rf’ suffix. This is in order
to emphasise the syntactical nature of the property. Additionally, it is worth noting that a syntactically
deterministic expression is not truly deterministic if the variables it contains are randomised over,
which is the case sometimes.
11
ET _ VAL
ET _ VAR
ET _ FAIL
Γ ⊢ Val v : val_type v
Γ ⊢ Var x : Γ x
Γ ⊢ Fail t : t
ET _ OP
Γ⊢e : t
ET _ IF
Γ⊢b : B
op_type op t = Some t
Γ ⊢ op $ e : t
Γ ⊢ e1 : t
′
′
ET _ PAIR
Γ ⊢ e1 : t1
Γ ⊢ e2 : t2
Γ ⊢ < e1 , e2> : t1 × t2
ET _ LET
Γ ⊢ e2 : t
Γ ⊢ IF b THEN e1 ELSE e2 : t
ET _ RAND
Γ ⊢ e1 : t1
t1 • Γ ⊢ e2 : t2
Γ ⊢ LET e1 IN e2 : t2
Γ ⊢ e : dist_param_type dst
Γ ⊢ Random dst e : dist_result_type dst
Figure 3: The typing rules for source-language expressions
The definition is obvious and leads to the following equality (assuming that e is deterministic and well-typed and σ is a valid state):
expr_sem σ e = return (expr_sem_rf σ e)
This property will later enable us to also convert deterministic source-language expressions into ‘equivalent’ target-language expressions.
3.5. Target language
The target language is again modelled very closely after the one used by Bhat et al. [4].
The type system and the operators are the same as in the source language, but the key
difference is that while expressions in the source language return a measure space on
their result type, the expressions in the target language always return a single value.
Since our source language lacks sum types, so does our target language. Additionally, our target language differs from that used by Bhat et al. in the following respects:
• Our language has no function types; since functions only occur as integrands and
as final results (as the compilation result is a density function), we can simply
define integration to introduce the integration variable as a bound variable and
let the final result contain a single free variable with de Bruijn index 0, i. e. there
is an implicit λ abstraction around the compilation result.
• Evaluation of expressions in our target language can never fail. In the language
by Bhat et al., failure is used to handle undefined integrals; we, on the other hand,
use the convention of Isabelle’s measure theory library, which returns 0 for integrals of non-integrable functions. This has the advantage of keeping the semantics
simple, which makes proofs considerably easier.
12
• Our target language does not have Let bindings, since, in contrast to the source
language, they would be semantically superfluous here. However, they are still
useful in practice since they yield shorter expressions and can avoid multiple
evaluation of the same term; they could be added with little effort.
Figures 5, 6, and 7 show the syntax, typing rules, and semantics of the target language.
primrec expr_sem :: state ⇒ expr ⇒ val measure where
expr_sem σ (Val v) = return_val v
| expr_sem σ (Var x) = return_val (σ x)
| expr_sem σ (LET e1 IN e2 ) =
do {
v ← expr_sem σ e1 ;
expr_sem (v • σ) e2
}
| expr_sem σ (op $ e) =
do {
v ← expr_sem σ e;
return_val (op_sem op v)
}
| expr_sem σ < e1 , e2> =
do {
v ← expr_sem σ e1 ;
w ← expr_sem σ e2 ;
return_val <| v, w |>
}
| expr_sem σ (IF b THEN e1 ELSE e2 ) =
do {
b′ ← expr_sem σ b;
if b′ = TRUE then expr_sem σ e1 else expr_sem σ e2
}
| expr_sem σ (Random dst e) =
do {
p ← expr_sem σ e;
dist_measure dst p
}
| expr_sem σ (Fail t) = null_measure (stock_measure t)
Figure 4: The semantics of source-language expressions
13
datatype cexpr =
CVar nat | CVal val | pdf_operator $c cexpr | < cexpr, cexpr>c |
R
IFc cexpr THEN cexpr ELSE cexpr | c cexpr ∂pdf_type
Figure 5: The expressions of the target language
CET _ VAL
CET _ VAR
Γ ⊢c CVal v : val_type v
Γ ⊢c CVar x : Γ x
CET _ OP
Γ ⊢c e : t
op_type op t = Some t
Γ ⊢c op $c e : t
′
′
CET _ PAIR
Γ ⊢ c e1 : t1
Γ ⊢c < e1 , e2>c : t1 × t2
CET _ IF
Γ ⊢c b : B
Γ ⊢ c e2 : t2
CET _ INT
Γ ⊢ c e1 : t
Γ ⊢ c e2 : t
Γ ⊢c IFc b THEN e1 ELSE e2 : t
t • Γ ⊢c e : R
R
Γ ⊢c c e ∂t : R
Figure 6: The typing rules for the target language
primrec cexpr_sem :: state ⇒ cexpr ⇒ val where
cexpr_sem σ (CVal v) = v
| cexpr_sem σ (CVar x) = σ x
| cexpr_sem σ < e1 , e2>c = <| cexpr_sem σ e1 , cexpr_sem σ e2 |>
| cexpr_sem σ op $c e = op_sem op (cexpr_sem σ e)
| cexpr_sem σ (IFc b THEN e1 ELSE e2 ) =
(if cexpr_sem σ b = TRUE then cexpr_sem σ e1 else cexpr_sem σ e2 )
R
| cexpr_sem σ ( c e ∂t) =
R
RealVal ( x. extract_real (cexpr_sem ( x • σ) e) ∂stock_measure t)
Figure 7: The semantics of the target language
14
Converting deterministic expressions. The auxiliary function expr_rf _to_cexpr, which
will be used in some rules of the compiler that handle deterministic expressions, is of
particular interest. We mentioned earlier that deterministic source-language expressions can be converted to equivalent target-language expressions.8 This function does
precisely that. Its definition is obvious.
expr_rf _to_cexpr satisfies the following equality for any deterministic source-language
expression e:
cexpr_sem σ (expr_rf_to_cexpr e) = expr_sem_rf σ e
4. Abstract compiler
4.1. Density contexts
First, we define the notion of a density context, which holds the acquired context data
the compiler will require to compute the density of an expression. A density context is
a tuple Υ = (V, V ′ , Γ, δ) that contains the following information:
• The set V of random variables in the current context. These are the variables that
are randomised over.
• The set V ′ of parameter variables in the current context. These are variables that
may occur in the expression, but are not randomised over but treated as constants.
• The type environment Γ
• A density function δ that returns the common density of the variables V under
the parameters V ′ . Here, δ is a function from space (state_measure (V ∪ V ′ ) Γ) to
the extended real numbers.
A density context (V, V ′ , Γ, δ) describes a parametrised measure on the states on
V ∪ V ′ . Let ρ ∈ space (state_measure V ′ Γ) be a set of parameters. Then we write
dens_ctxt_measure (V, V ′ , Γ, δ) ρ for the measure that we obtain by taking state_measure V Γ,
transforming it by merging a given state σ with the parameter state ρ and finally applying the density δ on the resulting image measure. The Isabelle definition of this
is:
definition dens_ctxt_measure :: dens_ctxt ⇒ state ⇒ state_measure where
dens_ctxt_measure (V, V ′ , Γ, δ) ρ = density (distr (state_measure V Γ)
(state_measure (V ∪ V ′ ) Γ) (λσ. merge V V ′ (σ, ρ))) δ
8 Bhat
et al. say that a deterministic expression ‘is also an expression in the target language syntax, and
we silently treat it as such’ [4]
15
Informally, dens_ctxt_measure describes the measure obtained by integrating over the
variables v1 , . . . , vm ∈ V while treating the variables v1′ , . . . , v′n ∈ V ′ as parameters. The
evaluation of an expression e with variables from V ∪ V ′ in this context is effectively a
function
λv1′ . . . v′n .
Z +
v1 . . . .
Z +
vm . expr_sem (v1 , . . . , vm , v1′ , . . . , v′n ) e ·
δ (v1 , . . . , vm , v1′ , . . . , v′n ) ∂Γ v1 . . . ∂Γ vm .
A density context is well-formed (implemented in Isabelle as the locale density_context)
if:
• V and V ′ are finite and disjoint
• δ σ ≥ 0 for any σ ∈ space (state_measure (V ∪ V ′ ) Γ)
• δ is Borel-measurable w. r. t. state_measure (V ∪ V ′ ) Γ
• for any ρ ∈ space (state_measure V ′ Γ), the dens_ctxt_measure (V, V ′ , Γ, δ) ρ is a
sub-probability measure
4.2. Definition
As a first step, we have implemented an abstract density compiler as an inductive predicate Υ ⊢d e ⇒ f , where Υ is a density context, e is a source-language expression and
f is a function of type val state ⇒ val ⇒ ereal . Its first parameter is a state that assigns
values to the free variables in e and its second parameter is the value for which the
density is to be computed. The compiler therefore computes a density function that is
parametrised with the values of the non-random free variables in the source expression.
The compilation rules are virtually identical to those by Bhat et al. [4], except for the
following adaptations:
• Bhat et al. handle the IF . . . THEN . . . ELSE case with the ‘match’ rule for sum
types. As we do not support sum types, we have a dedicated rule for IF . . . THEN
. . . ELSE . . . .
• The use of de Bruijn indices requires shifting of variable sets and states whenever
the scope of a new bound variable is entered; unfortunately, this makes some
rules somewhat technical.
• We do not provide any compiler support for deterministic ‘let’ bindings. They are
semantically redundant, as they can always be expanded without changing the
semantics of the expression. In fact, they have to be unfolded for compilation, so
they can be regarded as a feature that adds convenience, but no expressivity.
16
The following list shows the standard compilation rules adapted from Bhat et al.,
plus a rule for multiplication with a constant.9 Note that the functions marg_dens and
marg_dens2 compute the marginal density of one (resp. two) variables by ‘integrating
away’ all the other variables from the common density δ. The function branch_prob
computes the probability of being in the current branch of execution by integrating
over all the variables in the common density δ.
HD _ VAL
countable_type (val_type v)
(V, V , Γ, δ) ⊢d Val v ⇒ λρ x. branch_prob (V, V ′ Γ, δ) ρ · h x = vi
′
HD _ VAR
x∈V
(V, V , Γ, δ) ⊢d Var x ⇒ marg_dens (V, V ′ , Γ, δ) x
′
HD _ PAIR
x∈V
y∈V
x 6= y
(V, V ′ , Γ, δ) ⊢d <Var x, Var y> ⇒ marg_dens2 (V, V ′ , Γ, δ) x y
HD _ FAIL
(V, V ′ , Γ, δ) ⊢d Fail t ⇒ λρ x. 0
HD _ LET
(∅, V ∪ V ′ , Γ, λx. 1) ⊢d e1 ⇒ f
(0 • V, { x + 1 | x ∈ V }, type_of Γ e1 • Γ, ( f · δ)(ρ ◦ (λx. x + 1)) ⊢d e2 ⇒ g
(V, V ′ , Γ, δ) ⊢d LET e1 IN e2 ⇒ λρ. g (undefined • ρ)
′
HD _ RAND
(V, V ′ , Γ, δ) ⊢d e ⇒ f
(V, V ′ , Γ, δ) ⊢d Random dst e ⇒
R+
λρ y. x. f ρ x · dist_dens dst x y ∂dist_param_type dst
HD _ RAND _ DET
e det
free_vars e ⊆ V ′
(V, V ′ , Γ, δ) ⊢d Random dst e ⇒
λρ x. branch_prob (V, V ′ , Γ, δ) ρ · dist_dens dst (expr_sem_rf ρ e) x
HD _ IF
(∅, V ∪ V ′ , Γ, λρ. 1) ⊢d b ⇒ f
(V, V , Γ, λρ. δ ρ · f TRUE) ⊢d e1 ⇒ g1
(V, V ′ , Γ, λρ. δ ρ · f FALSE) ⊢d e2 ⇒ g2
(V, V ′ , Γ, δ) ⊢d IF b THEN e1 ELSE e2 ⇒ λρ x. g1 ρ x + g2 ρ x
′
HD _ IF _ DET
9 Additionally, three congruence rules are required for technical reasons.
These rules are required because
the abstract and the concrete result may differ on a null set and outside their domain.
17
b det
(V, V ′ , Γ, λρ. δ ρ · hexpr_sem_rf ρ b = TRUEi) ⊢d e1 ⇒ g1
(V, V ′ , Γ, λρ. δ ρ · hexpr_sem_rf ρ b = FALSEi) ⊢d e2 ⇒ g2
(V, V ′ , Γ, δ) ⊢d IF b THEN e1 ELSE e2 ⇒ λρ x. g1 ρ x + g2 ρ x
HD _ FST
(V, V ′ , Γ, δ) ⊢d e ⇒ f
R+
(V, V ′ , Γ, δ) ⊢d fst e ⇒ λρ x. y. f ρ <| x, y |> ∂type_of Γ (snd e)
HD _ SND
(V, V ′ , Γ, δ) ⊢d e ⇒ f
R+
(V, V ′ , Γ, δ) ⊢d snd e ⇒ λρ y. x. f ρ <| x, y |> ∂type_of Γ (fst e)
HD _ OP _ DISCR
HD _ NEG
countable_type (type_of (op $ e))
(V, V ′ , Γ, δ) ⊢d e ⇒ f
R
+
(V, V ′ , Γ, δ) ⊢d op $ e ⇒ λρ y. x. hop_sem op x = yi · f ρ x ∂type_of Γ e
HD _ ADDC
HD _ MULTC
(V, V ′ , Γ, δ) ⊢d e ⇒ f
(V, V ′ , Γ, δ) ⊢d −e ⇒ λρ x. f ρ (− x )
e′ det
free_vars e′ ⊆ V ′
(V, V ′ , Γ, δ) ⊢d e ⇒ f
(V, V ′ , Γ, δ) ⊢d e + e′ ⇒ λρ x. f ρ ( x − expr_sem_rf ρ e′ )
c 6= 0
(V, V ′ , Γ, δ) ⊢d e ⇒ f
(V, V , Γ, δ) ⊢d e · Val (RealVal c) ⇒ λρ x. f ρ ( x/c) / |c|
′
HD _ ADD
HD _ INV
(V, V ′ , Γ, δ) ⊢d <e1 , e2> ⇒ f
R+
(V, V ′ , Γ, δ) ⊢d e1 + e2 ⇒ λρ z. x. f ρ <| x, z − x |> ∂type_of Γ e1
(V, V ′ , Γ, δ) ⊢d e ⇒ f
(V, V ′ , Γ, δ) ⊢d e−1 ⇒ λρ x. f ρ ( x −1 ) / x2
HD _ EXP
(V, V ′ , Γ, δ) ⊢d e ⇒ f
(V, V ′ , Γ, δ) ⊢d exp e ⇒ λρ x. if x > 0 then f ρ (ln x ) / x else 0
Figure 8: The abstract compilation rules
18
4.3. Soundness proof
We show the following soundness result for the abstract compiler:
10
lemma expr_has_density_sound :
assumes (∅, ∅, Γ, λρ. 1) ⊢d e ⇒ f and Γ ⊢ e : t and free_vars e = ∅
shows has_subprob_density (expr_sem σ e) (stock_measure t) ( f (λx. undefined))
Here, has_subprob_density M N f is an abbreviation for the following four facts:
• applying the density f to N yields M
• M is a sub-probability measure
• f is N-Borel-measurable
• f is non-negative on its domain
We prove this with the following generalised auxiliary lemma:
lemma expr_has_density_sound_aux :
assumes (V, V ′ , Γ, δ) ⊢d e ⇒ f and Γ ⊢ e : t and free_vars e = ∅
and density_context V V ′ Γ δ and free_vars e ⊆ V ∪ V ′
shows
has_parametrized_subprob_density (state_measure V ′ Γ)
(λρ. do {σ ← dens_ctxt_measure (V, V ′ , Γ, δ) ρ; expr_sem σ e})
(stock_measure t) f
The predicate has_parametrized_subprob_density R M N f simply means that f is Borelmeasurable w. r. t. R ⊗ N and that for any parameter state ρ from R, the predicate
has_subprob_density M N f holds.
The proof is by straightforward induction following the inductive definition of the
abstract compiler. In many cases, the monad laws for the Giry monad allow restructuring the induction goal in such a way that the induction hypothesis can be applied
directly; in the other cases, the definitions of the monadic operations need to be unfolded and the goal is essentially to show that two integrals are equal and that the output
produced is well-formed.
The proof given by Bhat et al. [5] is analogous to ours, but much more concise due to
the fact that side conditions such as measurability, integrability, non-negativity, and so
on are not proven explicitly and many important (but uninteresting) steps are skipped
or only hinted at.
10 Note
that since the abstract compiler returns parametrised density functions, we need to parametrise
the result with the state λx. undefined, even if the expression contains no free variables.
19
5. Concrete compiler
5.1. Approach
The concrete compiler is another inductive predicate, modelled directly after the abstract compiler, but returning a target-language expression as the compilation result
instead of a HOL function. We will use a standard refinement approach to relate the
concrete compiler to the abstract compiler. We thus lift the soundness result on the abstract compiler to an analogous one on the concrete one. This effectively shows that the
concrete compiler always returns a well-formed target-language expression that represents a density for the sub-probability space described by the source language.
The concrete compilation predicate is written as
(vs, vs′ , Γ, δ) ⊢c e ⇒ f
Here, vs and vs′ are lists of variables, Γ is a typing environment, and δ is a targetlanguage expression describing the common density of the random variables vs in the
context. It may be parametrised with the variables from vs′ .
5.2. Definition
The concrete compilation rules are, of course, a direct copy of the abstract ones, but
with all the abstract HOL operations replaced with operations on target-language expressions. Due to the de Bruijn indices and the lack of functions as explicit objects in
the target language, some of the rules are somewhat complicated – inserting an expression into the scope of one or more bound variables (such as under an integral) requires
shifting the variable indices of the inserted expression correctly. For this reason, we will
not print the rules here; they can be found in the Isabelle theory file PDF_Compiler.thy.
5.3. Refinement
The refinement relates the concrete compilation (vs, vs′ , Γ, δ) ⊢c e ⇒ f to the abstract
compilation
(set vs, set vs′ , Γ, λσ. cexpr_sem σ δ) ⊢c e ⇒ λρ x. cexpr_sem ( x • ρ) f
In words: we take the abstract compilation predicate and
• variable sets are refined to variable lists
• the typing context and the source-language expression remain unchanged
• the common density in the context and the compilation result are refined from
HOL functions to target-language expressions (by applying the target language
semantics)
20
The main refinement lemma states that the concrete compiler yields a result that is
equivalent to that of the abstract compiler, modulo refinement. Informally, the statement is the following: if e is ground and well-typed under some well-formed concrete
density context Υ and Υ ⊢c e ⇒ f , then Υ′ ⊢d e ⇒ f ′ , where Υ′ and f ′ are the abstract
versions of Υ and f .
The proof for this is conceptually simple – induction over the definition of the concrete compiler; in practice, however, it is quite involved. In every single induction step,
the well-formedness of the intermediary expressions needs to be shown, congruence
lemmas for the abstract compiler need to be applied, and, when integration is involved,
non-negativity and integrability have to be shown in order to convert non-negative integrals to Lebesgue integrals and integrals on product spaces to iterated integrals.
Combining this main refinement lemma and the abstract soundness lemma, we can
now easily show the concrete soundness lemma:
lemma expr_has_density_cexpr_sound :
assumes ([], [], Γ, 1) ⊢c e ⇒ f and Γ ⊢ e : t and free_vars e = ∅
shows
has_subprob_density (expr_sem σ e) (stock_measure t)
(λx. cexpr_sem ( x • σ) f )
′
′
Γ 0 = t =⇒ Γ ⊢c f : REAL
free_vars f ⊆ {0}
Informally, the lemma states that if e is a well-typed, ground source-language expression, compiling it with an empty context will yield a well-typed, well-formed targetlanguage expression representing a density function on the measure space described
by e.
5.4. Final result
We will now summarise the soundness lemma we have just proven in a more concise
manner. For convenience, we define the symbol e : t ⇒c f (read ‘e with type t compiles
to f ’), which includes the well-typedness and groundness requirements on e as well as
the compilation result:11
e : t ⇒c f ←→ (λx. UNIT ⊢ e : t ∧ free_vars e = ∅ ∧ ([], [], λx. UNIT, 1) ⊢c e ⇒ f )
The final soundness theorem for the compiler, stated in Isabelle syntax, is then:12
11 In
this definition, the choice of the typing environment is, of course, completely arbitrary since the
expression contains no free variables.
12 To be precise, the lemma statement in Isabelle is slightly different; for better readability, we unfolded
one auxiliary definition here and omitted the type cast from real to ereal.
21
lemma expr_compiles_to_sound :
assumes e : t ⇒c f
fixes Γ Γ′ σ σ′
shows expr_sem σ e = density (stock_measure t) (λx. cexpr_sem ( x • σ′ ) f )
∀ x ∈ type_universe t. cexpr_sem ( x • σ′ ) f ≥ 0
Γ⊢e : t
t • Γ′ ⊢c f : REAL
free_vars f ⊆ {0}
In words, this result means the following:
Theorem
Let e be a source-language expression. If the compiler determines that e is
well-formed and well-typed with type t and returns the target-language
expression f , then:
• the measure obtained by taking the stock measure of t and using the
evaluation of f as a density is the measure obtained by evaluating e
• f is non-negative on all input values of type t
• e has no free variables and indeed has type t (in any type context Γ)
• f has no free variable except the parameter (i. e. the variable 0) and is a
function from t to REAL13
Isabelle’s code generator now allows us to execute our inductively-defined verified
compiler using the values command14 or generate code in one of the target languages
such as Standard ML or Haskell.
5.5. Evaluation
As an example on which to test the compiler, we choose the same expression that was
chosen by Bhat et al. [4]:15
LET x = Random UniformReal < 0, 1> IN
LET y = Random Bernoulli x IN
IF y THEN x + 1 ELSE x
14 Our
compiler is inherently non-deterministic since it may return zero, one, or many density functions,
seeing as an expression may have no matching compilation rules or more than one. Therefore, we must
use the values command instead of the value command and receive a set of compilation results.
15 Val and RealVal were omitted for better readability and symbolic variable names were used instead of
de Bruijn indices.
22
We abbreviate this expression with e. We can then display the result of the compilation using the Isabelle command values ”{(t, f ) | t f . e : t ⇒c f }” .
The result is a singleton set which contains the pair (REAL, f ), where f is a very long
and complicated expression. Simplifying constant subexpressions and expressions of
the form fst < e1 , e2 > and again replacing de Bruijn indices with symbolic identifiers,
we obtain:
R
b. (IF 0 ≤ x − 1 ∧ x − 1 ≤ 1 THEN 1 ELSE 0) · (IF 0 ≤ x − 1 ∧ x − 1 ≤ 1 THEN
IF b THEN x − 1 ELSE 1 − ( x − 1) ELSE 0) · hbi +
R
b. (IF 0 ≤ x ∧ x ≤ 1 THEN 1 ELSE 0) · (IF 0 ≤ x ∧ x ≤ 1 THEN
IF b THEN x ELSE 1 − x ELSE 0) · h¬ bi
Further simplification yields the following result:
h 1 ≤ x ≤ 2i · ( x − 1) + h 0 ≤ x ≤ 1i · ( 1 − x )
While this result is the same as that which Bhat et al. have reached, our compiler generates a much larger expression than the one they printed. The reason for this is that
they printed a β-reduced version of the compiler output; in particular, constant subexpressions were evaluated. While such simplification is, of course, very useful when
using the compiler in practice, we have not implemented it since it is not conceptually
interesting and outside the scope of this work.
6. Conclusion
6.1. Breakdown
All in all, the formalisation of the compiler took about three months. It contains a
total of roughly 10000 lines of Isabelle code (definitions, lemma statements, proofs, and
examples). Table 3 shows a detailed breakdown of this.
Table 3: Breakdown of the Isabelle code
Type system and semantics
Abstract compiler
Concrete compiler
General Measure Theory / auxiliary lemmas
2900 lines
2600 lines
1400 lines
3400 lines
As can be seen from this table, a sizeable portion of the work was the formalisation
of results from general measure theory, such as the substitution lemmas for Lebesgue
integrals, and auxiliary notions and lemmas, such as measure embeddings. Since the
utility of these formalisations is not restricted to this particular project, they will be
moved to Isabelle’s Measure Theory library.
23
6.2. Difficulties
The main problems we encountered during the formalisation were:
Missing background theory. As mentioned in the previous section, a sizeable amount
of Measure Theory and auxiliary notions had to be formalised. Most notably, the existing Measure Theory library did not contain integration by substitution. As a side
product of their formalisation of the Central Limit Theorem [2], Avigad et al. proved
the Fundamental Theorem of Calculus and, building thereupon, integration by substitution. However, their integration-by-substitution lemma only supported continuous
functions, whereas we required the theorem for general Borel-measurable functions.
Using their proof of the Fundamental Theorem of Calculus, we proved such a lemma,
which initially comprised almost 1000 lines of proof, but has been shortened significantly thereafter.
Proving side conditions. Many lemmas from the Measure Theory library require
measurability, integrability, non-negativity, etc. In hand-written proofs, this is often
‘hand-waved’ or implicitly dismissed as trivial; in a formal proof, proving these can
blow up proofs and render them very complicated and technical. The measurability
proofs in particular are ubiquitous in our formalisation. The Measure Theory library
provides some tools for proving measurability automatically, but while they were quite
helpful in many cases, they are still work in progress and require more tuning.
Lambda calculus. Bhat et al. use a simply-typed Lambda-calculus-like language
with symbolic identifiers as a target language. For a paper proof, this is the obvious
choice, since it leads to concise and familiar definitions, but in a formal setting, it always comes with the typical problems of having to deal with variable capture and
related issues. For that reason, we chose to use de Bruijn indices instead; however, this
makes handling target-language terms less intuitive, since variable indices need to be
shifted whenever several target-language terms are combined.
Another issue was the lack of a function type in our target language; allowing firstorder functions, as Bhat et al. did, would have made many definitions easier and more
natural, but would have complicated others significantly due to measurability issues.
Furthermore, we effectively needed to formalise a number of properties of Lambda
calculus that can be used implicitly in a paper proof.
6.3. Future work
The following improvements to the Isabelle formalisation could probably be realised
with little effort:
• sum types and a match . . . with . . . statement
• a compiler rule for deterministic ‘let’ bindings
24
• ‘let’ bindings for the target language
• a preprocessing stage to allow ‘normal’ variables instead of de Bruijn indices
• a postprocessing stage to automatically simplify the density expression as far as
possible
The first of these is interesting because it is the only substantial weakness of our
formalisation as opposed to that by Bhat et al.; the remaining four merely make using
the compiler more convenient or more efficient.
Additionally, in the long term, a Markov-chain Monte Carlo (MCMC) sampling method
such as the Metropolis–Hastings algorithm could also be formalised in Isabelle and integrated with the density compiler. However, this would be a more involved project as it
will require the formalisation of additional probability theory in Isabelle.
6.4. Summary
Using Isabelle/HOL, we formalised the semantics of a simple probabilistic functional
programming language with predefined probability distributions and a compiler that
returns the probability distribution that a program in this language describes. These
are modelled very closely after those given by Bhat et al. [4]. We then used the existing
formalisations of measure theory in Isabelle/HOL to formally prove the correctness of
this compiler w. r. t. the semantics of the source and target languages.
This shows not only that the compiler given by Bhat et al. is correct, but also that
a formal correctness proof for such a compiler can be done with reasonable effort and
that Isabelle/HOL in general and its Measure Theory library in particular are suitable
for it.
References
[1] Audebaud, P., Paulin-Mohring, C.: Proofs of randomized algorithms in Coq. In:
Mathematics of Program Construction, Lecture Notes in Computer Science, vol.
4014, pp. 49–68. Springer Berlin Heidelberg (2006),
http://dx.doi.org/10.1007/11783596_6
[2] Avigad, J., Hölzl, J., Serafin, L.: A formally verified proof of the Central Limit
Theorem. CoRR abs/1405.7012 (2014)
[3] Bhat, S., Agarwal, A., Vuduc, R., Gray, A.: A type theory for probability density
functions. In: Proceedings of the 39th Annual ACM SIGPLAN-SIGACT
Symposium on Principles of Programming Languages. pp. 545–556. POPL ’12,
ACM, New York, NY, USA (2012),
http://doi.acm.org/10.1145/2103656.2103721
[4] Bhat, S., Borgström, J., Gordon, A.D., Russo, C.: Deriving probability density
functions from probabilistic functional programs. In: Tools and Algorithms for
25
the Construction and Analysis of Systems, Lecture Notes in Computer Science,
vol. 7795, pp. 508–522. Springer Berlin Heidelberg (2013),
http://dx.doi.org/10.1007/978-3-642-36742-7_35, (best paper award)
[5] Bhat, S., Borgström, J., Gordon, A.D., Russo, C.: Deriving probability density
functions from probabilistic functional programs. In: TODO. Springer Berlin
Heidelberg (2014)
[6] Cock, D.: Verifying probabilistic correctness in Isabelle with pGCL. In:
Proceedings of the 7th Systems Software Verification. pp. 1–10 (November 2012)
[7] Cock, D.: pGCL for Isabelle. Archive of Formal Proofs (Jul 2014),
http://afp.sf.net/entries/pGCL.shtml, Formal proof development
[8] Doberkat, E.E.: Stochastic relations: foundations for Markov transition systems.
Studies in Informatics, Chapman & Hall/CRC (2007)
[9] Doberkat, E.E.: Basing Markov transition systems on the Giry monad.
http://www.informatics.sussex.ac.uk/events/domains9/Slides/Doberkat_GiryMonad.pdf
(2008)
[10] Giry, M.: A categorical approach to probability theory. In: Categorical Aspects of
Topology and Analysis, Lecture Notes in Mathematics, vol. 915, pp. 68–85.
Springer Berlin Heidelberg (1982), http://dx.doi.org/10.1007/BFb0092872
[11] Hölzl, J.: Construction and stochastic applications of measure spaces in
Higher-Order Logic. PhD thesis, Technische Universität München, Institut für
Informatik (2012)
[12] Hurd, J., McIver, A., Morgan, C.: Probabilistic guarded commands mechanized in
HOL. Electron. Notes Theor. Comput. Sci. 112, 95–111 (Jan 2005),
http://dx.doi.org/10.1016/j.entcs.2004.01.021
[13] Park, S., Pfenning, F., Thrun, S.: A probabilistic language based upon sampling
functions. In: Proceedings of the 32Nd ACM SIGPLAN-SIGACT Symposium on
Principles of Programming Languages. pp. 171–182. POPL ’05, ACM, New York,
NY, USA (2005), http://doi.acm.org/10.1145/1040305.1040320
26
Appendix
A. Operator semantics
O PERATOR
Add
Minus
Mult
Inverse
I NPUT
TYPE
O UTPUT
Z×Z
Z
R×R
R
Z
Z
R
R
Z×Z
Z
R×R
S EMANTICS
TYPE
a+b
−a
a·b
R
R
(
R
Sqrt
R
R
Exp
R
R
Ln
R
R
Pi
UNIT
R
1
a
for a 6= 0
otherwise
0
(√
a
0
for a ≥ 0
otherwise
ea
(
ln a
0
for a > 0
otherwise
π
(
ab
0
Z×Z
Z
R×Z
R
Fact
Z
Z
a!
And
B×B
B
a∧b
Or
B×B
B
a∨b
Not
B
B
¬a
Equals
t×t
B
a=b
Z×Z
B
Less
Pow
R×R
B
27
for a 6= 0, b ≥ 0
otherwise
a<b
Fst
t1 × t2
t1
a for z = ( a, b)
Snd
t1 × t2
t2
b for z = ( a, b)
B
R
Z
R
h a = TRUEi
a as a real number
B
Z
R
Z
h a = TRUEi
Cast REAL
Cast INT
⌊a⌋
B. Built-in distributions
D ISTRIBUTION
PARAM .
D OMAIN
T YPE
D ENSITY FUNCTION
(
p
1− p
for x = TRUE
for x = FALSE
Bernoulli
R
p ∈ [0; 1]
B
UniformInt
Z×Z
p1 ≤ p2
Z
h x ∈ [ p1 ; p2 ]i
p2 − p1 + 1
UniformReal
R×R
p1 < p2
R
h x ∈ [ p1 ; p2 ]i
p2 − p1
Gaussian
R×R
p2 > 0
R
Poisson
R
p≥0
Z
( x − p1 )2
q
exp −
2p22
2πp22
1
(
exp(− p) · p x /x!
0
!
for x ≥ 0
otherwise
Figure 9: The built-in distributions of the source language.
The density functions are given in terms of the parameter p, which is of the type given in the
column ‘parameter type’. If p is of a product type, p1 and p2 stand for the two components of
p. x is the function variable, e. g. the point at which the density function is to be evaluated.
28
C. Target language auxiliary functions
For the definitions of the following functions, see the corresponding Isabelle theory file
PDF_Target_Semantics.thy. We will not print them here since they are rather technical.
In this section, x will always be a variable, v will always be a value, and e and e′
will always be target language expressions. f and g will always be target language
expressions interpreted as functions, i. e. with an implicit λ abstraction around them.
All functions in the table take de Bruijn indices into account, i. e. they will shift variables when they enter the scope of an integral.
F UNCTION
D ESCRIPTION
case_nat y f x
if x = 0 then y else f ( x − 1)
map_vars h e
transforms all variables in e with the function h
ins_var0 e
map_vars Suc e
(prepares expression e for insertion into scope of new
variable)
ins_var1 f
map_vars (case_nat 0 (λx. x + 2)) f
(prepares function f for insertion into scope of new variable)
del_var0 e
map_vars (λx. x − 1) e
(shifts all variables down, deleting the variable 0)
cexpr_subst x e e′
substitutes e for x in e′
cexpr_subst_val e v
substitutes v for CVar 0 in e, effectively applying e interpreted as a function to the argument v.
cexpr_comp f g
function composition, i. e. λx. f ( g x )
f ◦c g
alias for cexpr_comp
expr_rf_to_cexpr e
converts a deterministic source language expression
into a target language expression, see Sect. 3.5
integrate_var Γ x e
integrates e over the free variable x
integrate_vars Γ xs e
integrates e over the free variables in the list xs
branch_prob_cexpr (vs, vs′ , Γ, δ)
returns an expression that computes branch_prob
marg_dens_cexpr Γ vs x δ
returns an expression that computes the marg_dens
marg_dens2_cexpr Γ vs x y δ
returns an expression that computes the marg_dens2
dist_dens_cexpr dst e
e′
returns an expression that computes the density of the
built-in distribution dst, parametrised with e and evaluated at e′
29
D. Concrete compiler rules
EDC _ VAL
countable_type (val_type v)
(vs, vs′ , Γ, δ) ⊢c Val v ⇒
ins_var0 (branch_prob_cexpr (vs, vs′ , Γ, δ)) · hCVar 0 = CVal vi
EDC _ VAR
x ∈ set vs
(vs, vs′ , Γ, δ) ⊢c Var x ⇒ marg_dens_cexpr Γ vs x δ
EDC _ PAIR
′
x ∈ set vs
y ∈ set vs
(vs, vs , Γ, δ) ⊢c <x, y> ⇒ marg_dens2_cexpr Γ vs x y δ
EDC _ FAIL
(vs, vs′ , Γ, δ) ⊢c Fail t ⇒ CVal (RealVal 0)
EDC _ LET
([], vs @ vs′ , Γ, 1) ⊢c e ⇒ f
(0 # map Suc vs, map Suc vs′ , type_of Γ e • Γ, ins_var0 δ · f ) ⊢c e′ ⇒ g
(vs, vs′ , Γ, δ) ⊢c LET e IN e′ ⇒ del_var0 g
EDC _ RAND
(vs, vs′ , Γ, δ) ⊢c e ⇒ f
(vs, vs ,RΓ, δ) ⊢c Random dst e ⇒
c ins_var1 f · dist_dens_cexpr dst (CVar 0) (CVar 1)
∂ dist_param_type dst
′
EDC _ RAND _ DET
free_vars e ⊆ set vs′
e det
′
(vs, vs , Γ, δ) ⊢c Random dst e ⇒
ins_var0 (branch_prob_cexpr (vs, vs′ , Γ, δ)) ·
dist_dens_cexpr dst (ins_var0 (expr_rf_to_cexpr e)) (CVar 0)
EDC _ IF
([], vs @ vs′ , Γ, 1) ⊢c b ⇒ f
(vs, vs′ , Γ, δ · hcexpr_subst_val f TRUEi) ⊢c e1 ⇒ f 1
(vs, vs′ , Γ, δ · hcexpr_subst_val f FALSEi) ⊢c e2 ⇒ f 2
(vs, vs′ , Γ, δ) ⊢c IF b THEN e1 ELSE e2 ⇒ f 1 + f 2
EDC _ IF _ DET
b det
(vs, vs′ , Γ, δ · hexpr_rf_to_cexpr b i) ⊢c e1 ⇒ f 1
(vs, vs′ , Γ, δ · h¬ expr_rf_to_cexpr b i) ⊢c e2 ⇒ f 2
(vs, vs′ , Γ, δ) ⊢c IF b THEN e1 ELSE e2 ⇒ f 1 + f 2
EDC _ OP _ DISCR
Γ⊢e : t
op_type oper t = Some t′
(vs, vs′ , Γ, δ) ⊢c e ⇒ f
(vs, vs , Γ, δ) ⊢c oper $ e ⇒ c hoper $c CVar 0 = CVar 1i · ins_var1 f ∂t
′
EDC _ FST
countable_type t′
R
Γ ⊢ e : t × t′
(vs, vs′ , Γ, δ) ⊢c e ⇒ f
R
(vs, vs′ , Γ, δ) ⊢c fst e ⇒ c ins_var1 f ◦c <CVar 1, CVar 0>c ∂t′
30
EDC _ SND
Γ ⊢ e : t × t′
(vs, vs′ , Γ, δ) ⊢c e ⇒ f
(vs, vs , Γ, δ) ⊢c snd e ⇒ c ins_var1 f ◦c <CVar 0, CVar 1>c ∂t
′
EDC _ NEG
EDC _ ADDC
R
(vs, vs′ , Γ, δ) ⊢c e ⇒ f
(vs, vs′ , Γ, δ) ⊢c − e ⇒ f ◦c (−CVar 0)
(vs, vs′ , Γ, δ) ⊢c e ⇒ f
e′ det
free_vars e′ ⊆ set vs′
′
(vs, vs , Γ, δ) ⊢c e + e ⇒ f ◦c (CVar 0 − ins_var0 (expr_rf_to_cexpr e′ ))
′
EDC _ MULTC
(vs, vs′ , Γ, δ) ⊢c e ⇒ f
c 6= 0
(vs, vs , Γ, δ) ⊢c e · Val (RealVal c) ⇒ ( f ◦c (CVar 0 / c))/| c|
′
EDC _ ADD
Γ ⊢<e, e′> : t × t
(vs, vs′ , Γ, δ) ⊢c <e, e′> ⇒ f
R
′
(vs, vs , Γ, δ) ⊢c e + e ⇒ c ins_var1 f ◦c <CVar 0, CVar 1 − CVar 0 > ∂t
′
EDC _ INV
(vs, vs′ , Γ, δ) ⊢c e ⇒ f
(vs, vs′ , Γ, δ) ⊢c e−1 ⇒ ( f ◦c (1/CVar 0)) / (CVar 0)2
EDC _ EXP
(vs, vs′ , Γ, δ) ⊢c e ⇒ f
(vs, vs , Γ, δ) ⊢c exp e ⇒
IF CVar 0 > 0 THEN ( f ◦c ln (CVar 0)) / CVar 0 ELSE 0
′
31
| 6cs.PL
|
Flare: Native Compilation for Heterogeneous Workloads
in Apache Spark
Grégory M. Essertel1 , Ruby Y. Tahboub1 , James M. Decker1 ,
Kevin J. Brown2 , Kunle Olukotun2 , Tiark Rompf1
1
Purdue University, 2 Stanford University
arXiv:1703.08219v1 [cs.DB] 23 Mar 2017
{gesserte,rtahboub,decker31,tiark}@purdue.edu, {kjbrown,kunle}@stanford.edu
ABSTRACT
SQL still performs at least an order of magnitude worse
than best-of-breed relational query engines like HyPer [35].
This is, in part, due to the fact that these relational query
engines not only optimize query plans aggressively on the
relational operator level, but also compile queries into native
code, thus operating closer to the metal than Spark’s current
Java-based techniques.
While one might argue that this is not a fair comparison
due to the added expressiveness of Spark and the different
nature of these systems, we show that it is actually possible
to bring the performance of Spark much closer to such
highly-optimized relational engines without sacrificing this
added expressiveness. We present Flare, a new back-end for
Spark that yields significant speedups by compiling entire
query plans obtained from Spark’s query optimizer to native
code, bypassing inefficient abstraction layers of the Spark
runtime system.
The need for modern data analytics to combine relational,
procedural, and map-reduce-style functional processing is
widely recognized. State-of-the-art systems like Spark have
added SQL front-ends and relational query optimization,
which promise an increase in expressiveness and performance. But how good are these extensions at extracting high
performance from modern hardware platforms?
While Spark has made impressive progress, we show that
for relational workloads, there is still a significant gap compared with best-of-breed query engines. And when stepping
outside of the relational world, query optimization techniques are ineffective if large parts of a computation have
to be treated as user-defined functions (UDFs).
We present Flare: a new back-end for Spark that brings
performance closer to the best SQL engines, without giving
up the added expressiveness of Spark. We demonstrate order
of magnitude speedups both for relational workloads such as
TPC-H, as well as for a range of machine learning kernels
that combine relational and iterative functional processing.
Flare achieves these results through (1) compilation to
native code, (2) replacing parts of the Spark runtime system, and (3) extending the scope of optimization and code
generation to large classes of UDFs.
1.
Heterogeneous Workloads. While Flare’s native code
translation alone already provides excellent performance
for SQL queries and DataFrame operations, performance
still suffers for heterogeneous workloads, e.g., when executing many small queries interleaved with user code, or when
combining relational with iterative functional processing,
as is common in machine learning pipelines. While the extract, transfer, and load part (ETL) of such pipelines can
often be implemented as DataFrames, the compute kernels
often have to be supplied as user-defined functions (UDFs),
which appear as black boxes to the query optimizer, and
thus remain as unoptimized library code.
For such workloads, Flare takes advantage of Delite [13,
14, 44], an existing compiler framework for high-performance
domain-specific languages (DSLs). As an alternative to generating target code in a single step, Flare can map Spark’s
query plans to Delite’s intermediate language, DMLL. This
enables Spark to interface with UDFs written in any of the
existing Delite DSLs, which cover domains such as machine
learning (OptiML [45]), graph processing (OptiGraph [46]),
or mesh-based PDE solvers (OptiMesh [46]). All of these
DSLs are embedded in Scala and provide APIs comparable
to those built on top of Spark. But unlike Scala code that
uses normal Spark APIs, Delite DSL code is amenable to optimization and native code generation, including GPU code
for computational kernels [28, 29]. We show that combining
Spark SQL queries and DataFrames with machine learning
kernels written in OptiML, in particular, results in order of
magnitude speedups compared to a standard Spark implementation.
INTRODUCTION
Modern data analytics applications require a combination of different programming paradigms, spanning relational, procedural, and map-reduce-style functional processing. Shortcomings in both expressiveness and performance
are key reasons why the excitement around early MapReduce tools, heralded as the solution to all big data problems, has tapered off. Instead, state-of-the-art systems like
Spark have added SQL front-ends and APIs which enable
relational query optimization [7].
But how good are these extensions? We demonstrate that
on standard relational benchmarks such as TPC-H, Spark
[Preprint, March 2017. Copyright by the authors.]
1
Front-end
User Programs
Front-end
SQL
OptiQL
(Java, Scala, Python, R) (JDBC, Console)
OptiML
OptiGraph
Spark SQL DataFrame API Export query plan
Delite’s Back-end
Catalyst Optimizer
DMLL
Code
Generation
JNI
Flare’s
Code Generation
Flare’s
Code Generation
LMS Code Generation
Optimized Scala, C
Flare’s Runtime
Spark
native code
Delite’s Runtime
Resilient Distributed Dataset
(a) Spark SQL
(b) Flare Level 1
(c) Flare Level 2
(d) Flare Level 3
Figure 1: Flare system overview: (a) architecture of Spark and Spark SQL, adapted from [7]; (b) Flare Level 1 replaces Spark’s
JVM code generation with native code generation, at the granularity of Spark’s query stages; (c) Flare Level 2 generates code
for entire queries, eliminating the RDD layer, and orchestrating parallel execution optimized for shared memory architectures;
(d) Flare Level 3 translates Spark query plans to Delite, an established framework for high-performance domain-specific
languages, which enables optimization of relational queries together with UDFs implemented in one of the front-end DSLs.
Scale-Up over Scale-Out. With the implementation of
Flare, we revisit some design decisions of Spark SQL rooted
in the legacy of Spark and earlier systems. We show that
alternative implementations, which start from different assumptions, can provide better performance with the same
expressive user-facing API.
One key assumption of MapReduce, Hadoop, and Spark is
that of a Google-scale, distributed, shared-nothing architecture, focusing primarily on scale-out vs scale-up. This makes
it easy to scale computation by adding more machines, but
it also means that each individual machine may not be used
efficiently [31]. An immediate consequence is an increase in
datacenter bills, but on a global scale, the effects of scaleout-first approaches (a.k.a “nobody ever got fired for running
a Hadoop cluster”) and their inefficient use of energy may
be as far-reaching as accelerating global warming.
Today, machines with dozens of cores and memory in
the TB range are readily available, both for rent and for
purchase. At the time of writing, Amazon EC2 instances
offer up to 2 TB main memory, with 64 cores and 128
hardware threads. Built-to-order machines at Dell can be
configured with up to 12 TB, 96 cores and 192 hardware
threads. NVIDIA advertises their latest 8-GPU system as
a “supercomputer in a box,” with compute power equal to
hundreds of conventional servers [36]. With such powerful
machines becoming increasingly commonplace, large clusters
are less and less frequently needed. Many times, “big data” is
not that big, and often computation is the bottleneck [37].
As such, a small cluster or even a single large machine is
sufficient, if it is used to its full potential.
With this scenario as the primary target – heterogeneous workloads and small clusters of powerful machines,
potentially with accelerators such as GPUs – Flare prioritizes bare-metal performance on all levels of scaling, with
the option of bypassing mechanisms such as fault-tolerance
for shared-memory-only execution. Thus, Flare strengthens
Spark’s role as a unified, efficient, big data platform, as opposed to a mere cluster computing fabric.
The overall architecture of Flare, with three possible levels
of integration, is shown in Figure 1. The paper first reviews
the design of Spark and Spark SQL (Section 2) and continues
with our main contributions:
• We present Flare: a new accelerator back-end for
Spark. We identify key impediments to performance
in Spark SQL, particularly related to Spark’s Java execution environment. As an immediate remedy, Flare
Level 1 generates native code instead of Java for improved performance (Section 3).
• We identify further performance issues in Spark SQL,
specifically related to joins in shared-memory environments. Flare Level 2 drastically reduces these overheads by making different starting assumptions than
Spark, focusing on scale-up instead of scale-out, and
reducing constant factors throughout. In particular,
Flare Level 2 compiles whole queries at once, as opposed to individual query stages, which results in an
end-to-end optimized data path (Section 4).
• We extend Flare further to accelerate heterogeneous
workloads, consisting of relational queries combined
with iterative machine learning kernels written as userdefined functions. Flare Level 3 uses the Delite compiler framework to optimize relational queries together
with such UDFs, by means of an intermediate language
(Section 5).
• We evaluate Flare in comparison to Spark both on
TPC-H, reducing the gap to best-of-breed relational
query engines, and on benchmarks involving heterogeneous machine learning workloads. In both settings,
Flare exhibits order-of-magnitude speedups. Our evaluation spans single-core, multi-core, NUMA, cluster,
and GPU targets (Section 6).
2.
BACKGROUND ON SPARK
Apache Spark [51, 52] is today’s most popular and most
widely-used big data framework. The core programming
abstraction is an immutable, implicitly distributed, collection data type called RDD (Resilient Distributed Dataset).
RDDs serve as high-level programming interfaces and also
transparently manage fault-tolerance.
2
Here is a quick example (from [7]) that counts the number
of errors in a (potentially distributed) log file:
== Physical Plan ==
*Filter StartsWith(value#894, ERROR)
+- *Scan text [value#894]
Format: ...TextFileFormat@18edbdbb,
InputPaths: ...,
ReadSchema: struct<value:string>
val lines = spark.sparkContext.textFile("...")
val errors = lines.filter(s => s.startsWith("ERROR"))
println("Total errors: " + errors.count())
Spark’s RDDs provide a deferred API: in the above example, the calls to textFile and filter merely construct a
computation graph. Actual computation only takes place at
the point where errors.count is invoked. Sometimes, RDDs
are described as lazily evaluated. This is somewhat misleading, as a second call to errors.count will re-execute the entire computation.1 However, RDDs support memoization via
explicit calls to errors.persist(), which will mark the data
set to be kept in memory for future operations.
2.1
From the high-level DataFrame operations, Spark SQL
internally computes a query plan, much like a relational
DBMS. Spark SQL optimizes query plans using its relational query optimizer called Catalyst, and may even generate Java code at runtime to accelerate parts of the query
plan using a component named Tungsten (see Section 2.3).
It is hard to overstate the benefits of this kind of API,
which generates a complete program (i.e. query) representation at runtime. First, it enables various kinds of optimizations, including classic relational query optimizations. Second, one can use this API from multiple front-ends, which
exposes Spark to non-JVM languages such as Python and
R, and the API can also serve as a translation target from
literal SQL:
Spark SQL and the DataFrame API
The directed acyclic computation graph represented by
an RDD describes the distributed operations in a coarsegrained way, at the granularity of map, filter, and so on.
This level of detail is enough to enable demand-driven computation, scheduling, and fault-tolerance via selective recomputation along the “lineage” of a result [51], but it does
not provide a full view of the computation applied to each element of a data set. For example, in the code snippet above,
the argument to lines.filter is a normal Scala closure. This
makes it easy to integrate RDD code with arbitrary external
libraries, but it also means that the given closure needs to
be invoked as-is for every element in the data set.
Thus, the performance of RDDs suffers from two limitations: first, limited visibility for analysis and optimizations,
especially standard optimizations like join re-ordering for
relational workloads expressed as RDDs; and second, interpretive overhead, i.e. function calls for each processed tuple.
Recent Spark versions aim to ameliorate both of these issues
with the introduction of the Spark SQL subsystem [7].
2.2
lines.createOrReplaceTempView("lines")
val errors = spark.sql("select * from lines
where value like ’ERROR%’")
println("Total errors: " + errors.count())
Third, one can use the full host language to structure code,
and use small functions that pass DataFrames between them
to build up a logical plan that is then optimized as a whole.
Of course, this is only true as long as one stays in the relational world, and does not use UDFs. As part of our contribution, we will show in Section 5 how the DataFrame model
extends to UDFs in Flare. Flare uses existing generative programming frameworks [42] to make larger classes of expressions available for DataFrame-like multi-stage programming
patterns, where a general piece of program code builds up an
intermediate representation (IR) of a more specific computation at runtime, like DataFrames with query plans. We argue
that such multi-stage APIs, including Spark’s DataFrames,
are the real key to success for unified and efficient big data
platforms, completely independent of RDDs. We will show
in Section 3.1 that the underlying RDD layer can actually be
an impediment to performance in Spark. Flare demonstrates
that Spark’s DataFrame API can successfully be supported
by a generic, high-performance DSL compiler framework.
This removes the confinement to relational query optimizations, and enables optimization of entire data processing
pipelines that combine relational processing with iterative
machine learning kernels that would require unoptimized
UDFs in plain Spark.
The Power of Multi-Stage APIs and DSLs
The core addition of Spark SQL is an alternative API
based on DataFrames.2 A DataFrame is conceptually equivalent to a table in a relational database, i.e., a collection
of rows with named columns. However, like RDDs, the
DataFrame API only records operations, but does not compute the result right away. We can write the same example
as before:
val lines = spark.read.textFile("...")
val errors = lines.filter($"value".startsWith("ERROR"))
println("Total errors: " + errors.count())
This is just like the RDD API, in that the call to errors.count
will trigger execution. Unlike RDDs, however, DataFrames
capture the full computation/query to be executed. We can
obtain the internal representation using
2.3
Catalyst and Tungsten
Spark SQL queries are optimized using the Catalyst optimizer that supports both rule-based and cost-based optimizations [7]. Starting with a logical query plan tree, Catalyst performs optimizations as tree transformations in order
to realize an optimized plan. It is important to note that at
the time of writing, Catalyst does not yet perform any kind
of join reordering. Instead, it simply joins tables in the order
they appear in the where clause. Hence, users need to manually encode a good join in order to avoid big intermediate
tables or other bad surprises.
Tungsten is the execution back-end that aims to improve
Spark performance by reducing the allocation of objects on
the JVM (Java Virtual Machine) heap, controlling off-heap
memory management, employing cache-aware data struc-
errors.explain()
which produces the following output:
1
In its original definition, the term “lazy evaluation” means
that each term is evaluated only when needed, and not more
than once [22].
2
Internally, Spark distinguishes the type Dataset[T], which
provides a typesafe collection API for elements of type T,
from the type DataFrame = Dataset[Row], which provides an
untyped API for rows with arbitrary arity and column
names. For the purpose of this paper, this difference is
immaterial; hence, we use the terms Dataset and DataFrame
interchangeably.
3
val tpchq6 = spark.sql("""
select
sum(l_extendedprice*l_discount) as revenue
from
lineitem
where
l_shipdate >= to_date(’1994-01-01’)
and l_shipdate < to_date(’1995-01-01’)
and l_discount between 0.05 and 0.07
and l_quantity < 24
""")
// data loading elided ...
for (i = 0; i < size; i++) {
double l_quantity = l_quantity_col[i];
double l_extendedprice = l_extendedprice_col[i];
double l_discount = l_discount_col[i];
long l_shipdate = l_shipdate_col[i];
if (l_shipdate >= 19940101L && l_shipdate < 19950101L &&
l_discount >= 0.05 && l_discount <= 0.07 &&
l_quantity < 24.0) {
revenue += l_extendedprice * l_discount;
}
} ...
Figure 2: Query 6 from the TPC-H benchmark in Spark.
Figure 3: Q6 hand-written C code
tures, and generating Java code that is compiled to JVM
bytecode at runtime [26]. These optimizations simultaneously improve the performance of all Spark SQL libraries
and DataFrame operations [52].
Spark SQL
Preload ms
Direct CSV
Preload CSV
118,062
Hand-Written C / Flare
Preload CSV
2,847
3.
Figure 4: Running times for Q6 in Spark, with and without
pre-loading, and compared to hand-written code and Flare.
FLARE: ADDING FUEL TO THE FIRE
We present Flare: a new back-end for Spark SQL that
achieves better performance without changing Spark’s userfacing front-end. Flare efficiently compiles Spark SQL
queries into native code, reduces internal overhead, and
brings performance closer to both hand-written C and the
best-of-breed SQL engines.
Flare can integrate with Spark at three levels, illustrated
in Figure 1. Level 1 enables native compilation within Tungsten, at the same level of granularity. Level 2 goes further
by compiling whole queries instead of only query stages, effectively bypassing Spark’s RDD layer and runtime for operations like hash joins in shared-memory environments. Finally, Level 3 goes beyond purely relational workloads by
adding another intermediate layer between query plans and
generated code. Flare Level 3 employs the Delite compiler
framework [28] to efficiently compile query pipelines with
heterogeneous workloads consisting of user-defined code in
multiple DSLs (e.g., SQL and ML). We will motivate and
describe each of these three levels in the following sections.
3.1
Query ms
24,400
1,418
45
As our benchmark, we pick the simplest query from the
industry-standard TPC-H benchmark: Query 6 (shown in
Figure 2). We define the schema of table lineitem, provide
the source file, and finally register it as a temporary table
for Spark SQL (steps not shown). For our experiments, we
use scale factor 2 (SF2) of the TPC-H data set, which means
that table lineitem is stored in a CSV file of about 1.4 GB.
Following the setup by McSherry et al., we run our tests on
a fairly standard laptop.3 We run our query, Q6, straight
from the CSV file as input, and record the running time:
scala> val q = spark.sql(tpchq6)
q: org.apache.spark.sql.DataFrame = [revenue: double]
scala> time(q.show)
Completed in 24,400 ms
Clearly, this result of 24 seconds is not the best we can do
(see Figure 4). We could convert our data to the columnar
Apache Parquet format [6] for increased performance, or
we can just preload the data so that subsequent runs are
purely in-memory. Since we are mainly interested in the
computational part, we opt to preload. We note in passing
that preloading is quite slow (almost 2 min), which may be
due to a variety of factors. Now we can execute our query
in-memory, and we get a much better result. Running the
query a few more times yields further speedups, but timings
stagnate at around 1s. The key question now is: how good
is this result?
How Fast is Spark?
Spark performance studies primarily focus on the scaleout performance, e.g., running big data benchmarks [51]
on high-end clusters, performing Terabyte sorting [52], etc.
However, as McSherry, Isard, and Murray have eloquently
argued in their 2015 HotOS paper [31] and accompanying
blog post [30], big data systems such as Spark tend to
scale well, but often just because there is a lot of internal
overhead. In particular, McSherry et al. demonstrate that
a straightforward native implementation of the PageRank
algorithm [38] running on a single laptop can outperform a
Spark cluster with 128 cores, using the then-current version.
Hand-Written C. Since Query 6 is very simple, it is perfectly feasible to write a program in C that performs exactly
the same computation: map the input file into memory using the mmap system call, load the data into an in-memory
columnar representation, and then execute the main query
loop, shown in Figure 3. If we compile this C program with
gcc -O3 Q6.c and run it, it will take 2.8s in total (including
data loading), and just 45ms for the actual query computation. So compared to Spark 2.0, the C program runs 20×
faster!
Is there any good reason why the C code needs to be faster
than Spark? We believe not, and in fact, running the same
Laptop vs Cluster. Inspired by this setup, we are interested
in gauging the inherent overheads of Spark and Spark SQL
in absolute terms.
“You can have a second computer once you’ve
shown you know how to use the first one.”
– Paul Barham, via [31]
We focus on a simple relational query, which will profit
from query optimization and code generation, and should
thus be a best-case scenario for Spark. We launch a Spark
shell configured to use a single worker thread:
3
MacBook Pro Retina 2012, 2.6 GHz Intel Core i7, 16 GB
1600 MHz DDR3, 500 GB SSD, Spark 2.0, Java HotSpot
VM 1.8.0_112-b16
./bin/spark-shell --master local[1]
4
java.lang.Double.isNaN(double) (148 samples, 14.52%)
2/7/17 7:28 PM
2.0.0
(/)
Jobs (/jobs/)
Stages (/stages/)
Storage (/storage/)
Details
for
Query
10
Submitted
Time:
2017/02/22 01:25:19
Duration:
5 s
Succeeded
Jobs:
13 (/jobs/job?id=13) 14 (/jobs/job?id=14)
Reset Zoom
java.lang.Double.isNaN..
java.n.. ja..
org.apache.spark.sql.c..
org.apache.spark.sql.catalyst.expressions.codegen.UnsafeRowWri..
org.apache.spark.sql.execution.colum..
org.apache.spark.sql.executio..
org.apache.spark.sql.execution.columnar.MutableUnsafeRow.setIn..
org.apache.spark.sql.execution.columnar.DOUBLE$.extract(ByteBuffer,..
org.apache.spark.sql.execution.columnar.compression.IntDelta$D..
org.apache.spark.sql.execution.columnar.compression.PassThrough$Dec..
org.apache.spark.sql.execution.columnar.compression.CompressibleColumnAccessor$class.extractSingle(NativeColumnAccessor,MutableRow,..
org.apache.spark.sql.execution.columnar.NativeColumnAccessor.extractSingle(MutableRow,int)
org.apache.spark.sql.execution.columnar.BasicColumnAccessor.extractTo(MutableRow,int)
org.apache.spark.sql.execution.columnar.NativeColumnAccessor.org$apache$spark$sql$execution$columnar$NullableColumnAccessor$$super$..
org.apache.spark.sql.execution.columnar.NullableColumnAccessor$class.extractTo(NullableColumnAccessor,MutableRow,int)
org.apache.spark.sql.execution.columnar.NativeColumnAccessor.extractTo(MutableRow,int)
org.apache.spark.sql.catalyst.expressions.GeneratedClass$SpecificColumnarIterator.next()
org.apache.spark.sql.catalyst.expressions.GeneratedClass$SpecificColumnarIterator.next()
org.apache.spark.sql.catalyst.expressions.GeneratedClass$GeneratedIterator.agg_doAggregateWithoutKey$(GeneratedClass$GeneratedIterator)
org.apache.spark.sql.catalyst.expressions.GeneratedClass$GeneratedIterator.processNext()
org.apache.spark.sql.execution.BufferedRowIterator.hasNext()
org.apache.spark.sql.execution.WholeStageCodegenExec$$anonfun$8$$anon$1.hasNext()
InMemoryTableScan
Spark
Sort-merge join
14,937 ms
Broadcast-hash join
4,775 ms
of which in exchange 2,232 ms
Flare
In-memory hash join
136 ms
number of output rows: 3000000
WholeStageCodegen
1.2 s (110 ms, 405 ms, 654 ms)
InMemoryTableScan
Filter
number of output rows: 11997996
number of output rows: 3000000
WholeStageCodegen
381 ms (10 ms, 32 ms, 43 ms)
BroadcastExchange
Filter
number of output rows: 11997996
data size (bytes): 48000000
time to collect (ms): 2885
time to build (ms): 1587
time to broadcast (ms): 239
BroadcastHashJoin
number of output rows: 11997996
Project
Figure 6: Join lineitem ./ orders: cost
of different operators (left); Spark’s hash
join plan (right) shows three separate code
generation regions, which communicate
through Spark’s runtime system.
Figure 5: CPU profile of TPC-H Q6 in Spark SQL, after
pre-loading the lineitem table. 80% of time (the large hump
middle to right) is spent accessing and decoding the inmemory data representation.
HashAggregate
number of output rows: 12
aggregate time total (min, med, max):
352 ms (10 ms, 30 ms, 35 ms)
Exchange
data size total (min, med, max):
180.0 B (15.0 B, 15.0 B, 15.0 B)
WholeStageCodegen
0 ms (0 ms, 0 ms, 0 ms)
HashAggregate
number of output rows: 1
aggregate time total (min, med, max):
0 ms (0 ms, 0 ms, 0 ms)
CollectLimit
Details
emitted in the generated code. LMS can be compared to
other systems like LLVM [27], but operates on a higher level
of abstraction. In the context of query compilation, LMS has
been used as part of the LegoBase engine [23].
query, Q6, accelerated with Flare yields exactly the same
performance as the hand-written C code. The same holds
for other best-of-breed, main-memory RDBMSs like HyPer
[35], which takes 46.58ms on a comparable machine.
Identifying the Bottlenecks. To understand these perfor-
4.
mance gains, we need to investigate where Spark SQL spends
its time. There are two reasons for the performance difference, both of which are particularly visible in the case of Q6,
which has low computational content and uses only trivial
query operators. First, we observe that Spark’s Tungsten
layer generates Java code for this query. In fact, two pieces
of code are generated: one piece for the main query loop, the
other an iterator to traverse the in-memory data structure.
When looking at the CPU profile (Figure 5), we can see
that 80% of the execution time is spent in accessing and decoding the in-memory data representation, and going back
and forth between the two pieces of generated code through
code paths that are part of the pre-compiled Spark runtime.
Second, even if we remove this indirection and replace it
with a unified piece of Java code, the performance remains
about 30% lower than C, a difference that gets more pronounced for other queries where tight low-level control over
data structures and memory management is required.
While Flare Level 1 can already provide significant
speedups for a number of queries, it is also constrained by
the architecture of the Spark runtime. Flare Level 2 takes
more liberties and revisits key design assumptions of Spark
SQL. Spark SQL operates on the legacy of Spark where
workloads are assumed to be Google-scale big, and the only
reasonable way of processing these humongous amounts of
data is to scale-out on a large number of distributed, unreliable, shared-nothing machines. Furthermore, Spark SQL
inherits Spark’s RDD abstraction which encodes lineage to
support fault tolerance, which is a necessity in distributed
environments of 100s or 1000s of machines. However, many
realistic workloads are not Google-scale, and, in fact, can
be scaled-up on modern big-memory hardware, where faults
are statistically highly improbable and fault tolerance can
often be supported in hardware, e.g., through RAID arrays,
redundant power supplies, or similar facilities. Accelerating
Spark SQL’s performance all the way to the level of bestof-breed modern query engines such as HyPer [35] requires
a fully compiled back-end that removes the remaining inefficiencies of legacy Spark.
file:///Users/me/Desktop/tryout/javaFlameGraph/FlameGraph/test.html
3.2
Page 1 of 1
Flare Level 1 Architecture
Flare Level 1 (Figure 1b) adds native compilation within
Tungsten. At startup, Flare installs a number of rules into
the Catalyst optimization logic. After a query is optimized
by Catalyst, these rules will trigger Tungsten to invoke Flare
to generate C code for supported operators or combinations
of operators. Flare then invokes a C compiler, loads the generated native code into the running JVM through the Java
Native Interface (JNI), and Spark’s runtime will then execute the generated code as part of the query evaluation.
Flare Level 1 does not modify any of Spark’s other internals, e.g., memory management, data format, RDD layer,
etc. Hence, Flare Level 1 can serve as a lightweight accelerator that increases performance for certain queries, but
not necessarily all, and that does not interfere with Sparks
execution model, i.e., transparently supports the standard
cluster execution and faul tolerance behavior. Like Tungsten itself, Flare’s query compiler implements Neumann’s
[35] data-centric model, which fuses pipelines of operators
that do not need to materialize intermediate results. Code
generation in Flare is realized using Lightweight Modular
Staging (LMS) [42], a generative programming and compiler
framework that uses the Scala type system to distinguish
normal code from expressions that generate code. In LMS,
a special type constructor Rep[T] is used to denote a staged
expression, which will cause an expression of type T to be
FLARE LEVEL 2
More Bottlenecks. For complex queries, concerns about
granularity of code generation and interfacing with the runtime system become more pronounced than in our previous
example, TPC-H Query 6. In fact, queries with joins exhibit some unfortunate consequences for main-memory execution due to Spark’s design as primarily a cluster computing framework. Figure 6 shows timings for a simple join
query that joins the lineitem and orders tables of the TPCH benchmark. Spark’s query optimizer picks an expensive
sort-merge join by default, which may be the right choice
for distributed or out-of-core execution, but is suboptimal
for main-memory. With the right flags, it is possible to tune
Spark’s Catalyst query planner to prefer a hash join instead,
which is more efficient. But even the hash join operator follows a broadcast model, with high overhead for the internal exchange operator (2.2s of 4.7s), which is present in the
physical plan even when running on a single core. Looking
at Spark’s query plan for this hash join query reveals that
Tungsten has split the query into three code generation regions (dark blue areas, right side in Figure 6), which need to
communicate through Spark’s runtime system. It is therefore no surprise that Flare can achieve much faster query
execution by generating tight code for the entire query.
5
Environment (/environment/)
Executors (/executors/)
SQL (/SQ
1
2
3
4
val tol = 0.001
val k = 4
def findNearestCluster(x_i: Rep[DenseVector[Double]],
mu: Rep[DenseMatrix[Double]]): Rep[Int] = {
(mu mapRowsToVector {
row => dist(x_i, row, SQUARE) }).minIndex}
5 /* Relational ETL */
6 val data = spark.read.csv[Data]("input")
7 val q = data.select(...)
8 val mat = flare(q).toMatrix
9 /* ML DSL with user code (k-Means training loop) */
10 val x = DenseTrainingSet(mat, DenseVector[Double]())
11 val m = x.numSamples
12 val n = x.numFeatures
13 val mu = (0::k, *) { i => x(randomInt(m)) }
14 val newMu = untilconverged_withdiff(mu, tol){ (mu, iter) =>
15
val c = (0::m) { e => findNearestCluster(x(e), mu) }
16
val allWP = (0::m).groupByReduce(...)
17
...
18 }((x, y) => dist(x, y, SQUARE))
19 /* Relational and ML DSLs */
20 val r = spark.sql("""select ... from data
where class = findNearestCluster(...)
group by class""")
21 flare(r).show
Memory
Columnar data
C
O
R
E
0
C
O
R
E
1
C
O
R
E
2
C
O
R
E
3
C
O
R
E
4
C
O
R
E
5
C
O
R
E
6
C
O
R
E
7
C
O
R
E
8
C C C
O O O
R R R
E E E
9 10 11
C
O
R
E
12
C C C
O OO
R R R
E E E
13 1415
Figure 7: Thread pinning to specific cores in NUMA.
4.1
Flare Level 2 Architecture
The architecture of Flare Level 2 is illustrated in Figure
1c. Spark SQL’s front-end, DataFrame API, and Catalyst
optimizer remain the same. The optimized query plan is
exported wholesale from Catalyst to Flare. Flare’s compiler
iterates recursively over each operator node in the query
plan tree and maps it to Flare’s internal optimized data
structures and query execution logic, represented in LMS.
After that, LMS performs some light optimizations like
common subexpression and dead code elimination, generates
C, invokes a C compiler, and Flare launches the resulting
binary either inside the JVM, like in Level 1, or as a separate
process. This cuts out the rest of Spark and relies solely on
Flare’s runtime to trigger execution of the generated code.
In contrast to Flare Level 1, which operates transparently to
the user, Flare Level 2 exposes a dedicated API to let users
pick which DataFrames to evaluate through Flare:
Figure 8: k-means clustering in Flare, combining SQL, Spark
DataFrames, and OptiML
4.3
val df = spark.sql("...") // create DataFrame (SQL or direct)
val fd = flare(df)
// turn it into a FlareDataFrame
fd.show()
// execute query plan with Flare
Flare Level 2 does not currently support cluster execution,
but it would be possible to extend the Spark runtime with a
set of hooks that would delegate to Flare within a single
machine, and coordinate multiple cluster nodes and the
necessary data exchanges through the existing Spark fabric.
Such a setup would be similar to HadoopDB [3], which
combines MapReduce processing between cluster nodes with
fast DBMS instances on each node that are able to handle
the per-node computation more efficiently than Hadoop.
4.2
Parallel and NUMA Execution
Query engines realize parallelism either explicitly by implementing special split and merge operators, or internally
by modifying the operator’s internal logic to orchestrate
parallel execution. Flare Level 2 uses the latter, and realizes parallelism using OpenMP [1]. On the architectural
level, Flare’s operators’ implementations take care of splitting their work internally across multiple threads, accumulating final results, etc. For instance, the parallel scan operator starts a parallel section in its produce method, which sets
the number of threads, and invokes consume on the downstream operators in parallel. Join and aggregate operators,
in turn, which implement materialization points, implement
their consume method in such a way that parallel invocations are possible without conflict, either through per-thread
data structures that are merged after the parallel section or
through lock-free data structures.
Flare also contains specific optimizations for environments
with non-uniform memory access (NUMA), including pinning threads to specific cores and optimizing the memory
layout of various data structures to reduce the need for
accessing non-local memory. For instance, memory-bound
workloads (e.g., TPC-H Q6) perform small amounts of computation, and do not scale-up given a large number of
threads on a single CPU socket. Flare’s code generation
supports such workloads through various data partitioning
strategies, in order to maximize local processing and to reduce the need for threads to access non-local memory as
illustrated in Figure 7 and Section 6.1.
Optimizing Data Loading
Data loading is an often overlooked factor data processing,
and is seldom reported in benchmarks. However, we recognize that data loading from CSV can often be the dominant
performance factor for Spark SQL queries. The Apache Parquet [6] format is an attractive alternative, modeled after
Dremel [32]. As a binary columnar format, it offers opportunities for compression, and queries can load only required
columns instead of all data.
In keeping with our running theme, can we do better?
While Parquet allows for irrelevant data to be ignored almost entirely, Spark’s code to read Parquet files is very
generic, resulting in undue overhead. This generality is primarily due to supporting multiple compression and encoding techniques, but there also exists overhead in determining
which column iterators are needed. While these sources of
overhead seem somewhat unavoidable, in reality they can
be resolved by generating specialized code. In Flare, we implement compiled CSV and Parquet readers that generate
native code specialized to a given schema. As a result, Flare
Level 2 can compile data paths end-to-end.
5.
FLARE LEVEL 3
Many data analytics applications require a combination
of different programming paradigms, i.e., relational, procedural, and map-reduce-style functional processing. For
example, a machine learning (ML) application might use
relational APIs for ETL, and dedicated ML libraries for
computations. Spark provides specialized libraries, e.g., ML
6
pipelines, and supports user-defined functions to support
domain-specific applications. Unfortunately, Spark’s performance falls off a proverbial cliff once DataFrame operations
are interleaved with user code. Currently, Spark SQL optimization and code generation treats user code as a black
box. Hence, Flare Level 3 focuses on generating efficient code
for heterogeneous workloads.
5.1
where each data point is assigned to the partition with the
nearest mean [10]. The k-means application mixes multiple
DSLs, i.e., SQL and OptiML [45], with user code. In lines
6-8, Spark SQL reads data from a file and preprocesses input. Lines 5-16 show the k-means processing code using OptiML. Delite provides optimized data types (e.g., Vectors)
and expressive libraries to assist ML computations. For instance, mapRowsToVector, dist, and untilconverged_withdiff
are ML-specific optimized methods. The final result can be
post-processed using SQL as illustrated in lines 20-21.
User Defined Functions (UDF)
Spark SQL uses Scala functions which appear as a black
box to the optimizer. As mentioned in Section 3, Flare’s internal code generation logic is based on a technique called
Lightweight Modular Staging (LMS) [42], which uses a special type constructor Rep[T] to denote staged expressions
of type T, that should become part of the generated code.
Extending UDF support to Flare is achieved by injecting
Rep[A] => Rep[B] functions into DataFrames in the same
way as normal A => B functions in plain Spark. As an example, here is a UDF sqr that squares a given number:
6.
// define and register UDF
def sqr(fc: FlareUDFContext) = { import fc._;
(y: Rep[Int]) => y * y }
flare.udf.register("sqr", sqr)
// use UDF in query
val df = spark.sql("select ps_availqty from partsupp where
sqr(ps_availqty) > 100")
flare(df).show()
6.1
Bare-Metal Relational Query Execution
The first set of experiments focuses on a standard relational workload, and demonstrates that the inherent overheads of Spark SQL cause a slowdown of at least 10× compared to the best available query engines for in-memory execution on a single core. Our experiments show that Flare
Level 2 is able to bridge this gap, accelerating Spark SQL
to the same level of performance as state-of-the-art query
compiler systems, while retaining the flexibility of Spark’s
DataFrame API. We also compare parallel speedups, the
effect of NUMA-aware optimization, and evaluate the performance benefits of optimized data loading.
Environment. We conducted our experiments on a single
NUMA machine with 4 sockets, 12 Xeon E5-4657L cores
per socket, and 256GB RAM per socket (1 TB total). The
operating system is Ubuntu 14.04.1 LTS. We use Spark
2.0, Scala 2.11, Postgres 9.4, HyPer v0.5-222-g04766a1, and
GCC 6.3 with optimization flags -O3.
Dataset. We use the standard TPC-H [47] benchmark
with scale factor SF10 for sequential execution, and SF20
and SF100 for parallel execution.
Notice that the definition of sqr uses an additional argument
of type FlareUDFContext, from which we import overloaded
operators such as +, -, *, etc., to work on Rep[Int] and other
Rep[T] types. The staged function will become part of the
code as well, and will be optimized along with the relational
operations. This provides benefits for UDFs (general purpose code embedded in queries) and enables queries to be
be optimized with respect to their surrounding code (e.g.,
queries run within a loop).
5.2
EXPERIMENTAL EVALUATION
To assess the performance and acceleration potential of
Flare in comparison to Spark, we present two sets of experiments. The first set focuses on a standard relational benchmark; the second set evaluates heterogeneous workloads,
consisting of relational processing combined with machine
learning kernels as UDFs. Our experiments span single-core,
multi-core, NUMA, cluster, and GPU targets.
Heterogeneous Workloads and DSLs
Flare Level 3 (Figure 1d) goes even further and leverages
an existing compiler framework called Delite [14], which is
built on top of LMS, to efficiently compile applications that
mix multiple DSLs or “query languages.” Delite has a particular focus on parallel and distributed applications, as well as
heterogeneous hardware such as GPUs [28]. Delite’s frontend provides several Scala DSLs, covering various domains,
e.g., databases, machine learning, linear algebra, etc. Since
these DSLs are embedded in Scala, they provide user-facing
APIs similar to libraries built on top of Spark.
On the front-end, Flare Level 3 exports a Spark SQL optimized query plan and maps it to Delite’s OptiQL (similar to
the process explained in Section 4). The core of Delite is built
around a first-order functional intermediate language called
Distributed Multi-Loop Language (DMLL) that models parallel patterns. As the name suggests, DMLL represents parallel patterns as a highly flexible loop abstraction that fuses
multiple loops to maximize pipeline and horizontal fusion
[13]. Furthermore, DMLL provides implicitly parallel and
nested collection operations, e.g., Collect, Group by-Reduce,
as well as optimizations like loop fusion, loop nesting optimizations, and so on. Finally, DMLL performs analysis and
optimizations for multiple hardware targets (e.g. NUMA,
clusters, GPU, etc.).
Figure 8 shows code highlights from the popular k-means
application, which partitions n data points into k clusters
Single-Core Running Time. In this experiment, we compare the single-core, absolute running time of Flare Level 2
with Postgres, HyPer, and Spark using the TPC-H benchmark with scale factor SF10. In the case of Spark, we use a
single executor thread, though the JVM may spawn auxiliary threads to handle GC or the just-in-time compilation.
Postgres and HyPer implement cost-based optimizers that
can avoid inefficient query plans, in particular by reordering
joins. While Spark’s Catalyst optimizer [7] is also cost-based,
it does not perform any kind of join re-ordering. Hence, we
match the join ordering of the query plan in Spark SQL and
Flare with HyPer’s, with a small number of exceptions: in
Spark SQL, the original join ordering given in the TPC-H
reference outperformed the HyPer plans for Q5, Q9, Q10,
and Q11 in Spark SQL, and for Q10 in Flare. For these
queries, we kept the original join ordering as is. For Spark
SQL, this difference is mainly due to Catalyst picking sortmerge joins over hash joins. It is worth pointing out that
HyPer and Postgres plans can use indexes on primary keys,
which may give an additional advantage.
Figure 9 gives the absolute execution time of Postgres,
HyPer, Spark SQL, and Flare for all TPC-H queries. For all
7
PostgreSQL
Running Time (ms)
1x106
Spark
HyPer
Flare
100000
10000
1000
100
10
1
Q1
SF10
Postgres
Spark SQL
HyPer
Flare
Q2
Q3
Q4
Q5
Q6
Q7
Q8
Q9
Q10
Q11
Q12
Q13
Q14
Q15
Q16
Q17
Q18
Q19
Q20
Q21
Q22
Q1
Q2
Q3
Q4
Q5
Q6
Q7
Q8
Q9
Q10 Q11
Q12
Q13
Q14
Q15
Q16
Q17
Q18
Q19
Q20
Q21
Q22
241404 6649 33721 7936 30043 23358 32501 29759 64224 33145 7093 37880 31242 22058 23133 13232 155449 90949 29452 65541 299178 11703
18219 23741 47816 22630 51731 3383 31770 63823 88861 42216 3857 17233 28489 7403 14542 23371 70944 53932 13085 31226 128910 10030
603
59 1126
842
941
232
943
616 1984
967 131
501 3625
330
253 1399
563 3703 1980
434
1626
180
530
139
532
521
748
198
830 1525 3124 1436
56
656 3727
278
302
620
2343
823
909
870
1962
177
Figure 9: Performance comparison of Postgres, HyPer, Spark SQL, Flare Level 2 in SF10
systems, data loading time is excluded, i.e., only execution
time is reported. In Spark and Flare, we use persist to
ensure that the data is loaded from memory. At first glance,
the performance of Flare and HyPer lie within the same
range, and notably outperform Postgres and Spark in all
queries. Similarly, Spark’s performance is comparable to
Postgres’s in most of the queries. Unlike the other systems,
Postgres does not compile queries at runtime, and relies on
the Volcano model [20] for query evaluation, which incurs
significant overhead. Hence, we can see that Spark’s query
compilation does not provide a significant advantage over a
standard interpreted query engines on most queries.
At a closer look, Flare outperforms Spark SQL in aggregate queries Q1 and Q6 by 34× and 17× respectively. We
observe that Spark is an order of magnitude slower than
Flare in nested queries like those found in Q2. After examining the execution plans of Q2, we found that Catalyst’s plan does not detect all patterns that help with avoiding re-computations, e.g., a table which has been previously
scanned or sorted. In join queries, e.g., Q5, Q10, Q14, etc.,
Flare is faster than Spark SQL by 20×-60×. Likewise, in join
variants outer join Q13, semi-join Q21, and anti-join Q22,
Flare is faster by 8×, 89× and 57× respectively.
The single-core performance gap between Spark SQL and
Flare is attributed to the bottlenecks identified in Sections 3
and 4. First, overhead associated with low-level data access
on the JVM. Second, Spark SQL’s distributed-first strategy that employs costly distributed operators, e.g., sortmerge join and broadcast hash join, even when running on
a single core. Third, internal bottlenecks in in-memory processing, the overhead of RDD operations, and communication through Spark’s runtime system. By compiling entire
queries, instead of isolated query stages, Flare effectively
avoids these bottlenecks.
HyPer [35] is a state-of-the-art compiled relational query
engine. A precursory look shows that Flare is faster than
HyPer by 10%-70% in Q1, Q4-Q6, Q7, and Q14. Moreover,
Flare is faster by 2×-4.5× in Q3, Q11, Q16, Q18, and Q19.
On the other hand, HyPer is faster than Flare by 20%-60%
in Q9, Q10, Q12, Q15, and Q21. Moreover, HyPer is faster
by 2×-4.1× in Q2, Q8, Q17, and Q20. This performance gap
is, in part, attributed to (1) HyPer’s use of specialized operators like GroupJoin [2], and (2) employing indexes on primary keys as seen in Q2, Q8, etc., whereas Flare (and Spark
SQL) currently does not support indexes. In order to understand the performance gained by using indexes, we disabled
indexing in HyPer and re-ran the benchmarks (detailed re-
sults omitted). We observed that Flare outperformed HyPer
by 80% in Q13 and matched the performance of Q12 and
Q21. For Q2, Q8-Q10, and Q17, where HyPer outperformed
Flare, the performance gap was shrunk by 20% to 1.2×. Finally, in Q4, Q7, Q11, and Q18 where Flare outperformed
HyPer, the performance gap had increased by 10% to 1.2×.
In summary, while both Flare and HyPer generate native query code at runtime, subtle implementation differences in query evaluation and code generation can result in
faster code. For instance. HyPer uses proper decimal precision numbers, whereas Flare follows Spark in using double
precision floating point values. which are native to the architecture. Furthermore, HyPer generates LLVM code, whereas
Flare generates C code which is then compiled with GCC.
Compilation Time. We compared the compilation time for
each TPCH-H query on Spark and Flare (detailed results
not shown). For Spark, we measured the time to generate
the physical plan, which includes Java code generation and
compilation. We do not quantify JVM-internal JIT compilation, as this is hard to measure, and code may be recompiled
multiple times. For Flare, we measured C code generation
and compilation with GCC. Both systems spend a similar
amount of time on code generation and compilation, on average 20% more in Flare. Compilation time depends on the
complexity of the query but is less than 1.5s for all queries,
i.e., well in line with interactive, exploratory, usage.
Parallel Scaling. In this experiment, we compare the scalability of Spark SQL and Flare Level 2. The experiment
focuses on the absolute performance and the Configuration
that Outperforms a Single Thread (COST) metric proposed
by McSherry et al. [31]. We pick four queries that represent
aggregate and join variants.
Figure 11 presents speedup numbers for Q6, Q13, Q14,
and Q22 when scaled up to 32 cores. At first glance, Spark
appears to have good speedups in Q6 and Q13 whereas
Flare’s Q6 speedup drops for high core counts. However,
examining the absolute running times, Flare is faster than
Spark SQL by 9×. Furthermore, it takes Spark SQL estimated 12 cores in Q6 to match the performance of Flare’s
single core. In scaling-up Q13, Flare is consistently faster
by 8× on all cores. Similarly, Flare continues to outperform
Spark by a steady 25× in Q14 and by 20×-80× in Q22 as
the number of cores reaches 32. Notice the COST metric in
the last two queries is infinity, i.e., there is no Spark configuration that matches Flare’s single-core performance.
8
Spark CSV
Speedup
1000
Spark Parquet
Flare CSV
Flare Parquet
100
10
1
Q1
Q2
Q3
Q4
Q5
Q6
Q7
Q8
Q9
Q10
Q11
Q12
Q13
Q14
Q15
Q16
Q17
Q18
Q19
Q20
Q21
Figure 10: Speedup for TPC-H SF1 when streaming data from SSD on a single thread.
#Tuples Postgres HyPer
CSV
CSV
Spark
Spark Flare
Flare
CSV Parquet
CSV Parquet
CUSTOMER
1500000
11664
LINEITEM
NATION
59986052
25
ORDERS
PART
15000000
2000000
7067
1102
377765 49408 471207
1
8
106
60214 33195
8807 1393
85985
11154
329
266
257898 11167
110
< 1
9730
10668
< 1
54124
7601
2028
351
16
8000000
37408
5265
28748
17731
1164
1010
REGION
SUPPLIER
5
100000
1
478
8
66
102
616
90
522
< 1
28
< 1
16
14
Spark
20
Q6 aggregate
Q13 outer-join
Q14 join
Q22 semi/anti-join
18
1786
340
PARTSUPP
Flare
20
Speedup
Table
18
16
14
12
12
10
10
8
8
6
6
4
4
2
2
1
2
4
8
16
32
1
2
4
8
Q6
7000
Spark SQL
Flare Level 2
6000
Running Time (ms)
Table 1: Loading time in ms for TPC-H SF10 in Postgres,
HyPer, and SparkSQL.
The seemingly good scaling for Spark reveals that the
runtime incurs significant overhead. In particular, we would
expect Q6 to become memory-bound as we increase the level
of parallelism. In Flare we can directly observe this effect as
a sharp drop from 16 to 32 cores. Since our machine has 18
cores per socket, for 32 cores, we start accessing non-local
memory (NUMA). The reason Spark scales better is because
the internal overhead, which does not contribute anything
to query evaluation, is trivially parallelizable and hides the
memory bandwidth effects. In summary, Flare scales as
expected for both of memory and CPU-bound workloads,
and reflects the hardware characteristics of the workload,
which means that query execution takes good advantage of
the available resources – with the exception of multiple CPU
sockets, a problem we address next.
As a next step, we evaluate NUMA optimizations in Flare
and show that these enable us to scale queries like Q6 to
higher core numbers. In particular, we pin threads to individual cores and lay out memory such that most accesses
are to the local memory region attached to each socket (Figure 12). Q6 performs better when the threads are dispatched
on different sockets. This is due to the computation being
bounded by the memory bandwidth. As such, when dividing
the threads on multiple sockets, we multiply the available
bandwidth proportionally. However, as Q1 is more computation bound, dispatching the threads on different sockets
has little effect. For both Q1 and Q6, we see scaling up to
the capacity of the machine (in our tests, up to 72 cores).
This is seen in a maximum speedup of 46× and 58× for Q1
and Q6, respectively.
32
60000
50000
5000
40000
4000
30000
3000
20000
2000
10000
1000
0
1
2
4
8
16
0
32
1
2
4
Q14
Running Time (ms)
16
Q13
8
16
32
8
16
32
Q22
14000
16000
12000
14000
12000
10000
10000
8000
8000
6000
6000
4000
4000
2000
0
2000
1
2
4
8
16
0
32
1
2
4
# Cores
# Cores
Figure 11: Scaling-up Flare and Spark SQL in SF20, without
NUMA optimizations: Spark has good nominal speedups
(top), but Flare has better absolute running time in all
configurations (bottom). For both systems, NUMA effects
for 32 cores are clearly visible.
Q1
Running Time (ms)
5500
Q6
3700
1
5400
3600
5300
600
500
400
300
200
100
0
3500
300
one socket
two sockets
four sockets
1
14
12 12 12
200
23 24
46
1
18
36
# Cores
72
22 23
29
100
0
44
1
18
36
# Cores
58
72
Figure 12: Scaling-up Flare for SF100 with NUMA optimizations on different configurations: threads pinned to one,
two, or four sockets. The speedups relative to a single thread
are shown on top of the bars.
faster in only one case: reading nation, a table with only 25
rows. In all other cases, Spark’s Parquet reader was 1.33×1.81× faster. However, Flare’s highly optimized CSV reader
operates at a closer level of performance to the Parquet
reader, with all tables except supplier having a benefit of
less than a 1.25× speedup by using Parquet.
Performing queries. Figure 10 shows speedups gained
from executing queries without preloading data for both systems. Whereas reading an entire table gives Spark and Flare
marginal speedups, reading just the required data gives
speedups in the range of 1.22×-22.96× for Spark (excluding Q2 and Q16, which performed 9% and 41% slower, re-
Optimized Data Loading. An often overlooked part of
data processing is data loading. Flare contains an optimized
implementation for both CSV files and the columnar Apache
Parquet format.4 We show loading times for each of the
TPC-H tables in Table 1.
Full table read. From the data in Table 1, we see that in
both Spark and Flare, the Parquet file readers outperform
the CSV file readers in most scenarios, despite this being
a worst-case scenario for Parquet. Spark’s CSV reader was
4
All Parquet files tested were uncompressed and encoded
using Parquet’s PLAIN encoding.
9
LogReg
50
6.2
Speedup
45
45
40
40
35
30
30
25
25
20
20
15
15
10
10
5
5
1
12
Speedup
24
48
1
Gene
50
12
45
40
40
35
35
30
30
25
25
20
20
15
15
24
48
24
48
GDA
50
45
10
10
5
5
1
12
# Cores
24
48
1
12
# Cores
Figure 13: Level 3: Machine learning kernels, scaling on
shared memory NUMA with thread pinning and data partitioning
Heterogeneous Workloads and UDFs
LogReg
Speedup over Spark
We now turn our attention to heterogeneous workloads
that combine relational processing with iterative functional
computation. We study a range of machine learning kernels,
where the input data is loaded via DataFrames. The kernel
computation is expressed as a UDF written in the OptiML
DSL, which is used as part of a SQL query.
k-means
GPU Cluster
8
8
7
7
6
6
5
5
5
4
4
4
3
3
3
2
2
2
1
1
0
Shared-Memory NUMA. In a recent study, Brown et
3.4 GB
17 GB
0
Spark
Delite-CPU
Delite-GPU
8
7
6
1
1.7 GB
17 GB
0
k-means
LogReg
Figure 14: Level 3: Machine learning kernels running on a
20 node Amazon cluster (left, center) and on a 4 node GPU
cluster connected within a single rack.
al. [13] compared Spark with Delite with regards to NUMA
scalability on a single shared-memory machine with 4 CPU
sockets and a total of 48 cores. Specifically, Gaussian Discriminant Analysis (GDA), logistic regression (LogReg), kmeans clustering, and a gene barcoding application (Gene)
were chosen as machine learning benchmarks. As shown in
Figure 13, Delite gains significant speedups over Spark in
every application studied, with thread pinning and NUMAaware data partitioning contributing in different ways for
each application. As stated previously, Flare Level 3 generates code in Delite’s intermediate language DMLL, and we
have verified that the generated code matches perfectly the
code in [13], thus guaranteeing the same performance.
data between iterations. Likewise, Spark [51, 52] tackles
the issue of data reuse among MapReduce jobs or applications by explicitly persisting intermediate results in memory. Along the same lines, the need for an expressive programming model to perform analytics on structured and
semistructured data motivated Hive [48], Dremel [32], Impala [24], Shark [49] and Spark SQL [7] and many others.
SnappyData [41] integrates Spark with a transactional mainmemory database to realize a unified engine that supports
streaming, analytics and transactions. Moreover, Asterix [9],
Stratosphere [4], and Tupleware [16] provide data flow-based
programming models and support user defined functions
(UDFs). Flare distinguishes itself by maintainng expressiveness while achieving performance close to highly optimized
systems like HyPer [35]. Flare keeps Spark SQL’s front-end
intact, and integrates its own code generation and runtime
(in Level 2) replacing Spark’s RDDs and distributed-bydefault runtime system. Furthermore, Flare Level 3 integrates with Delite as its execution back-end in order to support multi-DSL applications. By design, Flare’s main target
is scaled-up in-memory clusters where faults are improbable.
Query Compilation Recently, code generation for SQL
queries has regained momentum. Historic efforts go back all
the way to System R [8]. Query compilation can be realized
using code templates. e.g., Daytona [21] or HIQUE [25], general purpose compilers, e.g., HyPer [35] and Hekaton [18], or
DSL compiler frameworks, e.g., Legobase [23], DryadLINQ
[50], and DBLAB [43].
Embedded DSL frameworks and intermediate languages address the compromise between productivity and
performance in writing programs that can run under diverse programming models. Voodoo [40] addresses compiling portable query plans that can run on CPUs and GPUs.
Voodoo’s intermediate algebra is expressive and captures
hardware optimizations, e.g., multicores, SIMD, etc. Fur-
Clusters and GPU. Brown et al. [13] also showcase
speedups over Spark when running on a small cluster of
machines, and when employing GPU accelerators on the kmeans and LogReg applications. Despite running on Spark’s
intended architecture, when run on Amazon EC2 using 20
nodes, the code generated by Delite demonstrated a 2×-3.5×
speedup over Spark for k-means, and approximately a 2.5×3× speedup for LogReg (see Figure 14, adapted from [13]).
When these applications were moved to a cluster of higherend machines with more CPU cores as well as GPUs, this
performance gap widened even further. The study shows
that when running on a GPU cluster of 4 nodes, each with
12 cores, performance jumped to above a 7× speedup for
each application. Again, for completeness, Flare Level 3 generates DMLL, and we verified that all generated code was
identical to that used in [13].
7.
k-means
50
Spark
C++(nopin)
C++(pin)
35 C++(numa)
spectively) and 2×-14× for Flare. Across systems, however,
Flare’s Parquet reader demonstrated between a 5×-795×
speedup over Spark’s, and between 35×-720× over Spark’s
CSV reader. While the speedup over Spark lessens slightly
in higher scale factors, we found that Flare’s Parquet reader
consistently performed on average at least one order of magnitude faster across each query, regardless of scale factor.
In nearly every case, reading from a Parquet file in Flare is
approximately 2×-4× slower than in-memory processing (as
expected). However, reading from a Parquet file in Spark is
rarely significantly slower than in-memory processing. These
results show that while reading from Parquet certainly provides performance gains for Spark when compared to reading from CSV, the overall performance bottleneck of Spark
surely does not lie in the cost of reading from SSD compared
to in-memory processing.
RELATED WORK
Numerous cluster computing frameworks implement
a combination of parallel, distributed, relational, procedural, and MapReduce computations. The MapReduce model
[17] realized in Hadoop [5] performs big data analysis under shared-nothing, unreliable machines. Twister [19] and
Haloop [15] support iterative MapReduce workloads by
avoiding reading unnecessary data and keeping invariant
10
thermore, Voodoo is used as an alternative back-end for
MonetDB [12]. Delite [44], a general purpose compiler framework, implements high-performance DSLs (e.g., SQL, Machine Learning, graphs and matrices), provides parallel patterns and generates code for heterogeneous targets. The
Distributed Multiloop Language (DMLL) [13] provide rich
collections and parallel patterns and supports big-memory
NUMA machines. Weld [39] is another recent system that
aims to provide a common runtime for diverse libraries e.g.,
SQL and machine learning. Steno [33] performs optimizations similar to DMLL to compile LINQ queries. Furthermore, Steno uses DryadLINQ [50] runtime for distributed
execution. Nagel et. al. [34] generates efficient code for LINQ
queries. Weld is similar to DMLL in supporting nested parallel structures. However, Weld focuses on a larger set of
data science frameworks (Spark, TensorFlow, NumPy and
Pandas) and does not, e.g, support a large enough set of
joins and other operators to run all TPC-H queries. Query
compilation inside Flare Level 3 is based on Delite’s OptiQL
DSL and DMLL intermediate language.
Performance evaluation in data analytics frameworks
aims to identify performance bottlenecks and study the parameters that impact performance the most, e.g., workload,
scaling-up/scaling-out resources, probability of faults, etc.
A recent study [37] on a single Spark cluster revealed that
CPU, not I/O, is the source of bottlenecks. McSherry et
al. [31] proposed the COST (Configuration that Outperforms a Single Thread) metric, and showed that in many
cases, single-threaded programs can outperform big data
processing frameworks running on large clusters. TPC-H [47]
is a decision support benchmark that consists of 22 analytical queries that address several “choke points,” e.g., aggregates, large joins, arithmetic computations, etc. [11]. Flare
Level 2’s performance evaluation is done using TPC-H.
8.
[5]
[6]
[7]
[8]
[9]
[10]
[11]
[12]
[13]
CONCLUSION
Modern data analytics need to combine multiple programming models and make efficient use of modern hardware with
large memory, many cores, and accelerators such as GPUs.
We introduce Flare: a new back-end for Spark that brings
relational performance on par with the best SQL engines,
and also enables highly optimized of heterogeneous workloads. Most importantly, all of this comes without giving up
the expressiveness of Spark’s high-level APIs.
We believe that multi-stage APIs, in the spirit of
DataFrames, and compiler systems like Flare and Delite,
will play an increasingly important role in the future to satisfy the increasing demand for flexible and unified analytics
with high efficiency.
[14]
9.
[18]
[15]
[16]
[17]
REFERENCES
[1] OpenMP. http://openmp.org/.
[2] Accelerating queries with group-by and join by
groupjoin. PVLDB, 4(11), 2011.
[3] A. Abouzeid, K. Bajda-Pawlikowski, D. Abadi,
A. Silberschatz, and A. Rasin. Hadoopdb: an
architectural hybrid of mapreduce and dbms
technologies for analytical workloads. PVLDB,
2(1):922–933, 2009.
[4] A. Alexandrov, R. Bergmann, S. Ewen, J.-C. Freytag,
F. Hueske, A. Heise, O. Kao, M. Leich, U. Leser,
[19]
[20]
11
V. Markl, et al. The stratosphere platform for big data
analytics. The VLDB Journal, 23(6):939–964, 2014.
Apache. Hadoop. http://hadoop.apache.org/.
Apache. Parquet. https://parquet.apache.org/.
M. Armbrust, R. S. Xin, C. Lian, Y. Huai, D. Liu,
J. K. Bradley, X. Meng, T. Kaftan, M. J. Franklin,
A. Ghodsi, and M. Zaharia. Spark SQL: relational
data processing in spark. In SIGMOD, pages
1383–1394. ACM, 2015.
M. M. Astrahan, M. W. Blasgen, D. D. Chamberlin,
K. P. Eswaran, J. N. Gray, P. P. Griffiths, W. F. King,
R. A. Lorie, P. R. McJones, J. W. Mehl, et al. System
r: relational approach to database management.
TODS, 1(2):97–137, 1976.
A. Behm, V. R. Borkar, M. J. Carey, R. Grover, C. Li,
N. Onose, R. Vernica, A. Deutsch,
Y. Papakonstantinou, and V. J. Tsotras. Asterix:
towards a scalable, semistructured data platform for
evolving-world models. Distributed and Parallel
Databases, 29(3):185–216, 2011.
P. Berkhin. A survey of clustering data mining
techniques. In Grouping multidimensional data, pages
25–71. Springer, 2006.
P. Boncz, T. Neumann, and O. Erling. TPC-H
analyzed: Hidden messages and lessons learned from
an influential benchmark. In Technology Conference
on Performance Evaluation and Benchmarking, pages
61–76. Springer, 2013.
P. A. Boncz, M. Zukowski, and N. Nes.
Monetdb/x100: Hyper-pipelining query execution. In
CIDR, 2005.
K. J. Brown, H. Lee, T. Rompf, A. K. Sujeeth,
C. De Sa, C. Aberger, and K. Olukotun. Have
abstraction and eat performance, too: Optimized
heterogeneous computing with parallel patterns. CGO
2016, pages 194–205. ACM, 2016.
K. J. Brown, A. K. Sujeeth, H. Lee, T. Rompf,
H. Chafi, M. Odersky, and K. Olukotun. A
heterogeneous parallel framework for domain-specific
languages. In PACT, 2011.
Y. Bu, B. Howe, M. Balazinska, and M. D. Ernst.
Haloop: efficient iterative data processing on large
clusters. PVLDB, 3(1-2):285–296, 2010.
A. Crotty, A. Galakatos, K. Dursun, T. Kraska,
C. Binnig, U. Çetintemel, and S. Zdonik. An
architecture for compiling udf-centric workflows.
PVLDB, 8(12):1466–1477, 2015.
J. Dean and S. Ghemawat. MapReduce: Simplified
Data Processing on Large Clusters. In OSDI, pages
137–150, 2004.
C. Diaconu, C. Freedman, E. Ismert, P.-A. Larson,
P. Mittal, R. Stonecipher, N. Verma, and M. Zwilling.
Hekaton: Sql server’s memory-optimized oltp engine.
In SIGMOD, pages 1243–1254. ACM, 2013.
J. Ekanayake, H. Li, B. Zhang, T. Gunarathne, S.-H.
Bae, J. Qiu, and G. Fox. Twister: a runtime for
iterative mapreduce. In ACM HPCD, pages 810–818.
ACM, 2010.
G. Graefe. Volcano-an extensible and parallel query
evaluation system. TKDE, 6(1):120–135, 1994.
[21] R. Greer. Daytona and the fourth-generation language
cymbal. In ACM SIGMOD Record, volume 28, pages
525–526. ACM, 1999.
[22] P. Henderson and J. H. M. Jr. A lazy evaluator. In
POPL, pages 95–103. ACM Press, 1976.
[23] Y. Klonatos, C. Koch, T. Rompf, and H. Chafi.
Building efficient query engines in a high-level
language. PVLDB, 7(10):853–864, 2014.
[24] M. Kornacker, A. Behm, V. Bittorf, T. Bobrovytsky,
C. Ching, A. Choi, J. Erickson, M. Grund, D. Hecht,
M. Jacobs, I. Joshi, L. Kuff, D. Kumar, A. Leblang,
N. Li, I. Pandis, H. Robinson, D. Rorke, S. Rus,
J. Russell, D. Tsirogiannis, S. Wanderman-Milne, and
M. a. Yoder. Impala: A modern, open-source sql
engine for hadoop. In CIDR, 2015.
[25] K. Krikellas, S. D. Viglas, and M. Cintra. Generating
code for holistic query evaluation. In ICDE, pages
613–624. IEEE, 2010.
[26] J. Laskowski. Mastering Apache Spark 2.
Technical Report 1999-66, Stanford InfoLab,
November 1999.
[39] S. Palkar, J. J. Thomas, A. Shanbhag, D. Narayanan,
H. Pirk, M. Schwarzkopf, S. Amarasinghe, M. Zaharia,
and S. InfoLab. Weld: A common runtime for high
performance data analytics. In CIDR, 2017.
[40] H. Pirk, O. Moll, M. Zaharia, and S. Madden. Voodoo
- a vector algebra for portable database performance
on modern hardware. VLDB, 9(14):1707–1718, 2016.
[41] J. Ramnarayan, B. Mozafari, S. Wale, S. Menon,
N. Kumar, H. Bhanawat, S. Chakraborty,
Y. Mahajan, R. Mishra, and K. Bachhav. Snappydata:
A hybrid transactional analytical store built on spark.
In SIGMOD, pages 2153–2156, 2016.
[42] T. Rompf and M. Odersky. Lightweight modular
staging: a pragmatic approach to runtime code
generation and compiled dsls. Commun. ACM,
55(6):121–130, 2012.
[43] A. Shaikhha, Y. Klonatos, L. Parreaux, L. Brown,
M. Dashti, and C. Koch. How to architect a query
compiler. In SIGMOD, pages 1907–1922. ACM, 2016.
[44] A. K. Sujeeth, K. J. Brown, H. Lee, T. Rompf,
H. Chafi, M. Odersky, and K. Olukotun. Delite: A
compiler architecture for performance-oriented
embedded domain-specific languages. TECS,
13(4s):134, 2014.
[45] A. K. Sujeeth, H. Lee, K. J. Brown, T. Rompf,
M. Wu, A. R. Atreya, M. Odersky, and K. Olukotun.
OptiML: an implicitly parallel domain-specific
language for machine learning. In ICML, 2011.
[46] A. K. Sujeeth, T. Rompf, K. J. Brown, H. Lee,
H. Chafi, V. Popic, M. Wu, A. Prokopec,
V. Jovanovic, M. Odersky, and K. Olukotun.
Composition and reuse with compiled domain-specific
languages. In ECOOP, 2013.
[47] The Transaction Processing Council. TPC-H Version
2.15.0.
[48] A. Thusoo, J. S. Sarma, N. Jain, Z. Shao, P. Chakka,
S. Anthony, H. Liu, P. Wyckoff, and R. Murthy. Hive:
a warehousing solution over a map-reduce framework.
PVLDB, 2(2):1626–1629, 2009.
[49] R. S. Xin, J. Rosen, M. Zaharia, M. J. Franklin,
S. Shenker, and I. Stoica. Shark: Sql and rich analytics
at scale. In ACM SIGMOD, pages 13–24, 2013.
[50] Y. Yu, M. Isard, D. Fetterly, M. Budiu, Ú. Erlingsson,
P. K. Gunda, and J. Currey. Dryadlinq: A system for
general-purpose distributed data-parallel computing
using a high-level language. OSDI, 8, 2008.
[51] M. Zaharia, M. Chowdhury, M. J. Franklin,
S. Shenker, and I. Stoica. Spark: cluster computing
with working sets. In USENIX, HotCloud’10, 2010.
[52] M. Zaharia, R. S. Xin, P. Wendell, T. Das,
M. Armbrust, A. Dave, X. Meng, J. Rosen,
S. Venkataraman, M. J. Franklin, A. Ghodsi,
J. Gonzalez, S. Shenker, and I. Stoica. Apache spark:
a unified engine for big data processing. Commun.
ACM, 59(11):56–65, 2016.
https://www.gitbook.com/book/jaceklaskowski/
mastering-apache-spark/details, 2016.
[27] C. Lattner and V. Adve. LLVM: A compilation
framework for lifelong program analysis and
transformation. In CGO, 2004.
[28] H. Lee, K. J. Brown, A. K. Sujeeth, H. Chafi,
T. Rompf, M. Odersky, and K. Olukotun.
Implementing domain-specific languages for
heterogeneous parallel computing. IEEE Micro,
31(5):42–53, 2011.
[29] H. Lee, K. J. Brown, A. K. Sujeeth, T. Rompf, and
K. Olukotun. Locality-aware mapping of nested
parallel patterns on gpus. In MICRO, 2014.
[30] F. McSherry. Scalability! but at what cost.
http://www.frankmcsherry.org/graph/scalability/
cost/2015/01/15/COST.html, 2015.
[31] F. McSherry, M. Isard, and D. G. Murray. Scalability!
but at what cost? In HotOS, 2015.
[32] S. Melnik, A. Gubarev, J. J. Long, G. Romer,
S. Shivakumar, M. Tolton, and T. Vassilakis. Dremel:
interactive analysis of web-scale datasets. Proceedings
of the VLDB Endowment, 3(1-2):330–339, 2010.
[33] D. G. Murray, M. Isard, and Y. Yu. Steno: automatic
optimization of declarative queries. In ACM
SIGPLAN Notices, volume 46, pages 121–131, 2011.
[34] F. Nagel, G. Bierman, and S. D. Viglas. Code
generation for efficient query processing in managed
runtimes. VLDB, 7(12):1095–1106, 2014.
[35] T. Neumann. Efficiently compiling efficient query
plans for modern hardware. PVLDB, 4(9):539–550,
2011.
[36] NVIDIA. The world’s first AI supercomputer in a box.
http://www.nvidia.com/object/
deep-learning-system.html, 2016.
[37] K. Ousterhout, R. Rasti, S. Ratnasamy, S. Shenker,
and B. Chun. Making sense of performance in data
analytics frameworks. In NSDI, pages 293–307, 2015.
[38] L. Page, S. Brin, R. Motwani, and T. Winograd. The
pagerank citation ranking: Bringing order to the web.
12
| 6cs.PL
|
1
A High-Level Rule-based Language for Software
Defined Network Programming based on OpenFlow
arXiv:1712.04706v2 [cs.NI] 14 Dec 2017
Mehdi Mohammadi, Ala Al-Fuqaha, Zijiang James Yang
Computer Science Department
Western Michigan University
{mehdi.mohammadi, ala.al-fuqaha, zijiang.yang}@wmich.edu
Abstract—This paper proposes XML-Defined Network policies
(XDNP), a new high-level language based on XML notation,
to describe network control rules in Software Defined Network
environments. We rely on existing OpenFlow controllers specifically Floodlight but the novelty of this project is to separate
complicated language- and framework-specific APIs from policy
descriptions. This separation makes it possible to extend the
current work as a northbound higher level abstraction that
can support a wide range of controllers who are based on
different programming languages. By this approach, we believe
that network administrators can develop and deploy network
control policies easier and faster.
Index Terms—Software Defined Networks; OpenFlow; Floodlight; SDN compiler; SDN programming languages; SDN abstraction.
I. I NTRODUCTION
By Software-Defined Network (SDN) technology, network
engineers and administrators can control and manage network
services through abstraction of lower level functionality. This
end can be achieved by splitting the system that makes
decisions where to send the packets (control plane) from the
underlying systems known as data plane that is in charge of
forwarding packets to the selected destinations. In order to be
practical, SDN needs some mechanism for the control plane to
interact with the data plane. OpenFlow [1] is such a protocol
that has been used in network research community in recent
years. There are couples of OpenFlow controller frameworks
such as POX1 , NOX2 , Beacon3 , Floodlight4 , Trema5 , NodeFlow6 and Ryu7 .
The emerge of OpenFlow simplified network management
by providing high-level abstractions to control a set of switches
remotely. An OpenFlow framework requires network engineers or administrators to write programs to control and
manage data traffic in their network and control network
devices. However, one potential problem that network engineers face is that the programming languages that support
OpenFlow are complex and administrators are required to
1 http://www.noxrepo.org/pox/about-pox/
2 http://www.noxrepo.org/
3 https://openflow.stanford.edu/display/Beacon/Home
4 http://www.projectfloodlight.org/
5 http://trema.github.io/trema/
6 http://garyberger.net/?p=537
7 http://osrg.github.io/ryu/
know much irrelevant information to develop and deploy a
control policy. The problem will be more prohibitive for
a beginner network engineer who does not have a good
background in the programming language of the controller.
There are other challenges for programmers including [2]: the
interaction between concurrent modules, low-level interface
to switch hardware, and multi-tiered programming model. A
simple and unified language in a higher abstraction level that
does not depend on a specific language can fill this gap.
We have developed a text-based format based on XML by
which a human-friendly semantic of operations is developed
for policy description. XML has been used widely in network
management and configuration protocols like NETCONF. This
work can be annexed to the current OpenFlow controller as a
top layer service.
Two contributions of this paper are:
• proposing an XML-based script language for describing
network control policies; and
• implementing a translator that converts the XML file to
a Java source code containing the controller program for
Floodlight.
The emulation experiments affirms the applicability of this
XML notation as a policy describer in SDNs.
II. R ELATED W ORK
XML documents are widely used to describe systems, to
configure or control them. One attractive usage is code generation. There are several works that try to use XML notation as
a representation of source codes. JavaML [3] is such a work
that represents Java source code in XML notation. The source
code representation in JavaML is in a way that constructs like
superclasses, methods, message sends, and literal numbers are
all directly represented in the elements and attributes of the
document content. XML notation is also used in another work
named srcML [4] by which structural information is added to
unstructured source code files. srcML aims to enhance source
code representation by adding syntactic information obtained
from parse tree.
In [5], Liang et al. proposed a proof of concept to use
XML as a description scheme for OpenFlow Networking
experiments. They defined networking experiments in a hierarchical model in which the experiment is at the highest
level, and each experiment contains information, topology,
deployment, control, and output components. However, they
This paper has been presented in the poster section of GENI Engineering Conference 22 (GEC 22), Washington D.C., March 23-26, 2015.
2
just provided the format description, but not mentioned by
which way they generated their XML parser. Furthermore,
they have not provided a full evaluation of their system in
the real environments. Our work differs from their work in
the way we design a platform by which network engineers
can describe their network control rules by XML notations
regardless of the underlying topology or network elements. In
other words, Liang’s work focuses on description of network
environment not the control of network traffic and behavior.
There are several works and projects aim to propose a
higher level abstraction above the OpenFlow APIs in their
development frameworks [2], [6]–[8]. Frenetic [2] which has
been implemented in python emerged with some simple
rules including predicate-action pairs, in which actions support filtering, forwarding, duplicating, and modifying packets. Later it included other more complicated operators like
packet processing functionalities [6]. Pyretic [7] as a modern
SDN programming language based on Frenetic and beyond
the current parallel composition operator, presents two more
complex abstractions: sequential composition operator and
applying control policies over abstract topologies. By these
abstractions the development of modular control programs
becomes simpler. NetKat [8] is another language for SDN with
a solid mathematical foundation that supports primitives like
filtering, modifying, and transmitting packets as well as union
and sequential composition operators.
There are other works that have more concentration to
provide a high-level language for policy description. Procera
[9] and FML [10] are designed based on this aim. The
declarative policy language in Procera is based on functional
reactive programming. This domain-specific language relies
on Haskell and is able to use the constructs and data types
defined in Haskell. FML has been built on Datalog language
to support a declarative logic-based programming interface.
In [11], as part of their resilient network management system,
authors have proposed a policy description language based on
management patterns. The management patterns describe the
handling and controlling of individual OpenFlow resilience
services.
III. S OFTWARE D EFINED N ETWORKING
Software-Define Networking defines two separate layers
as the new architecture of network environments. A data
plane that is supposed to do operations like buffering packets,
forwarding, dropping, tagging and collecting packet statistics.
Control plane, on the other hand, may have the algorithms
to track the dynamic topology of the network and has route
processing capability. The control plane usually consists of
a separate powerful machine called controller. The control
plane uses its computed data and the data plane’s statistics to
manage and govern a set of dependent switches by installing
or removing packet forwarding rules over them.
OpenFlow as a realization of SDN follows this two-layer
architecture. In a typical OpenFlow network, if a switch can
find a rule match to the received packet in its flow table, then
it proceeds with that rule. Otherwise, the packet is sent to the
controller for more processing. The controller examines packet
Fig. 1. The overall architecture of the system.
header and establishes a rule based on that packet. The next
similar packets arrived in the switches then are not required to
go to the controller since the switches have appropriate rule to
process them. Although forwarding packets to the controller
increases their latency but it occurs not very much.
IV. T HE P ROPOSED S YSTEM
Our XML translator consists of a lexical analyzer (lexer)
and a syntax analyzer (parser). The lexer tokenizes the input
file and matches the tags, attributes, identifiers, constant values
and so on based on regular grammars. Syntax analyzer, on the
other hand, performs syntactic analysis of the input file and if it
does not find any problem in this step, generates appropriate
Java source code. The overall architecture of the system is
depicted in Figure 1.
A. Language Specification
The overall design of an XML file to be used as control
program should follow the format of figure 2. Rule description
is defined in a hierarchical structure in which at the top level,
we define the class name in the SDN element. Then a list of
rules contacting zero or more rule elements should be declared.
Inside each rule, one or more conditions are expressed. To have
compositional conditions, condition elements support logical
operators which are stated by attribute “connector”. For example, if a condition element has an “or” connector, it will be
joined to the previous condition by logical “or” operator in java
source code. The conditions themselves comply with a simple
pattern “variable op value”. Variables can be chosen from
src_ip (source IP), dest_ip (destination IP), src_prt
(source port) and dest_prt (destination port). The current
supported operator is equal sign. Value can be a port number
or IP address. An example of XML file is shown in Figure 3.
3
construct the lexer. For example, to define the pattern for IP
address, we define these regular expressions in the lexer file:
oct10 [1-9][0-9]|1[0-9][0-9]|2[0-4][0-9]|25[0-5]
start_oct [1-9]|oct10
octet [0-9]|oct10
ipv4 start_octȯctetȯctetȯctet
Tag names are considered as keywords in the script and are
required to comply exactly with the XML specification.
C. Syntax Analyzer
Fig. 2. The format of a policy description file.
The grammar section of translator is defined in syntax
analyzer. We used YACC tool to generate the translator. In
its input file, we define all the tokens that are introduced in
the lexer, grammar production rules that examine the syntax of
the XML file and the output associated with each production
rule leading to code generation for the XML file. The syntax
analyzer description follows the Backus–Naur Form (BNF)
notation for context-free grammars. For example, the following production rules show the nesting and iteration nature of
XML script for control policies:
ruleslist :
r u l e s l i s t rule
| rule
| /∗ nothing ∗/
;
rule :
RULE_S { f p r i n t f ( fp , " \ t \ t i f
conditionCounter = 0;}
c o n d i t i o n L i s t a c t i o n RULE_E
;
Fig. 3. A sample XML policy description.
The XML sample contains two rules. The first one says
that all the packets who are going to IP address 10.0.0.2
or who they are coming from IP address 192.168.0.1 should
be forwarded to port 1 of the switches (which assigned to
IP address 10.0.0.1). The second rule indicates that each
packet originated from Telnet service (port 23) should be
dropped (specified by port 0). This example shows that it is
possible to implement all network policy management schemes
and services like firewalls or load balancing with this XML
notation.
");
conditionList :
conditionList condition
| condition
;
condition :
COND_S GT { c o n d i t i o n C o u n t e r ++; }
c o n d T e x t COND_E
| ...
condText :
SD_IP EQ I P v 4 {
f p r i n t f ( fp , "(% s . e q u a l s ( \ " % s \ " ) ) " ,
$1 , $3 ) ; }
| SD_PRT EQ NUMBER{
f p r i n t f ( fp , "(% s == %d ) " , $1 , $3 ) ; } ;
Each string in the input XML file that is not matched to the
designed tokens and rules will cause an error to the program
and termination of the program with an error message.
B. Lexical Analyzer
We designed our lexer with LEX format by which we
defined the matching patterns for XML elements, attributes
and literals that are needed in a typical network controller. The
source of lexer file has to be fed to Flex to generate a source
code in C language. We used regular expression notations to
V. E XPERIMENTS
When our LEX and YACC input files are ready, we can
generate the final translator by calling the following commands. We have used equivalent versions of LEX and YACC
called FLEX and bison respectively. In the following command
4
xmlparser.l and xmlparser.y are the lexer and parser grammar
respectively:
flex xmlparser.l
bison -dy xmlparser.y
gcc lex.yy.c y.tab.c -o xmlparser.exe
The translator file is named xmlparser.exe. In order to
generate the Java source code for a desired XML file we
can use the following command in a console command line:
xmlparser.exe Demo.xml
The output of this command is a file named Demo.java
containing a Java class for the control policies of the network.
This file can be pushed to the controller machine and compiled
to be used for network management.
We used Mininet which is a network emulator to examine
the behavior of the proposed translator. Mininet supports
OpenFlow, so we can register the generated Java source
code with a Floodlight controller and attach the new
controller to a mininet topology. Here we explain how
to setup the experiments: After generating Java source
code form XML file, it should be added to the Floodlight
controller. We use Floodlight source code and add the
generated class to its main package. It is needed to set
Floodlight to load this class at startup. To do that, we
first tell the loader that this module exists by adding the
fully qualified module name on a configuration file named
net.floodlight.core.module.IFloodlightModule in the path
src/main/resources/META-INF/services. Then
we tell the module to be loaded. So we need to modify the
Floodlight module configuration file to append the Demo
file. The full name of this class should be added to the
file src/main/resources/floodlightdefault.properties. Now we
can compile the Floodlight project and have an OpenFlow
controller with our designed services running over it.
The next step is running Mininet. We setup a simple network
topology containing one switch (s0) and three hosts (h1, h2,
h3) connected to the switch. We also set the controller to be
remote. This configuration can be obtained by this command:
sudo mn -controller remote -topo
single,3.
We can test our module by calling ping command. Since
our first rule was forwarding those packets going to 10.0.0.2
to port 1, so a ping from h3 to h2 or h1 to h2 will install that
rule on switches and we will not receive reply from h2 while
other hosts can ping each other like p1 and p3. By this simple
rule we can block the traffic to h2. Running command pingall
we get 33% packet loss that means h2 is not responding to
none of ping requests.
Table I below shows the comparison of the XML file versus
generated Java file in terms of number of code lines and size of
code (Kilo Byte). Using this system, we reduce coding efforts
to about 1/8 of the original Java source code while the size of
rule description is about 1/5 of the generated source code.
VI. C ONCLUSION
This paper describes a new approach of software-defined
networks and presents a higher level of abstraction compared
to current software-defined network programming languages.
TABLE I
C OMPARISON OF XML AND JAVA SOURCE CODES .
XML file
Java file
# Line of Code
18
147
Size (Kilo Byte)
1
5
We defined a scripting language based on XML notation
by which network administrators can define control policies
without concerning the complexities of underlying controller
framework. Indeed, this will make software-defined networking easier and more attractive for network administrators.
This work opens up the opportunity for using service
oriented architecture and web services as a model of collaboration between SDN controllers (e.g., a controller offers load
balancing or firewalling).
Extending this work to support more APIs and more
complicated scenarios is intended to be done in the future
works. Technically, examining and manipulating other OpenFlow headers is possible by this approach. Supporting more
controller frameworks with different languages is also another
direction for future works.
R EFERENCES
[1] N. McKeown, T. Anderson, H. Balakrishnan, G. Parulkar, L. Peterson,
J. Rexford, S. Shenker, and J. Turner, “Openflow: enabling innovation in
campus networks,” ACM SIGCOMM Computer Communication Review,
vol. 38, no. 2, pp. 69–74, 2008.
[2] N. Foster, R. Harrison, M. J. Freedman, C. Monsanto, J. Rexford,
A. Story, and D. Walker, “Frenetic: A network programming language,”
in ACM Sigplan Notices, vol. 46, no. 9. ACM, 2011, pp. 279–291.
[3] G. J. Badros, “Javaml: a markup language for java source code,”
Computer Networks, vol. 33, no. 1, pp. 159–177, 2000.
[4] J. I. Maletic, M. L. Collard, and A. Marcus, “Source code files as
structured documents,” in Program comprehension, 2002. proceedings.
10th international workshop on. IEEE, 2002, pp. 289–292.
[5] J. Liang, Z. Lin, and Y. Ma, “Of-nedl: an openflow networking experiment description language based on xml,” Web Information Systems and
Mining, pp. 686–697, 2012.
[6] C. Monsanto, N. Foster, R. Harrison, and D. Walker, “A compiler
and run-time system for network programming languages,” in ACM
SIGPLAN Notices, vol. 47, no. 1. ACM, 2012, pp. 217–230.
[7] C. Monsanto, J. Reich, N. Foster, J. Rexford, D. Walker et al., “Composing software defined networks.” in NSDI, vol. 13, 2013, pp. 1–13.
[8] C. J. Anderson, N. Foster, A. Guha, J.-B. Jeannin, D. Kozen,
C. Schlesinger, and D. Walker, “Netkat: Semantic foundations for
networks,” in ACM SIGPLAN Notices, vol. 49, no. 1. ACM, 2014,
pp. 113–126.
[9] A. Voellmy, H. Kim, and N. Feamster, “Procera: a language for highlevel reactive network control,” in Proceedings of the first workshop on
Hot topics in software defined networks. ACM, 2012, pp. 43–48.
[10] T. L. Hinrichs, N. S. Gude, M. Casado, J. C. Mitchell, and S. Shenker,
“Practical declarative network management,” in Proceedings of the 1st
ACM workshop on Research on enterprise networking. ACM, 2009,
pp. 1–10.
[11] P. Smith, A. Schaeffer-Filho, D. Hutchison, and A. Mauthe, “Management patterns: Sdn-enabled network resilience management,” in Network
Operations and Management Symposium (NOMS), 2014 IEEE. IEEE,
2014, pp. 1–9.
| 6cs.PL
|
Approximate Generalized Matching:
f -Factors and f -Edge Covers
arXiv:1706.05761v2 [cs.DS] 27 Jul 2017
Dawei Huang
University of Michigan
Seth Pettie∗
University of Michigan
Abstract
In this paper we present linear time approximation schemes for several generalized matching problems on nonbipartite graphs. Our results include O (m)-time algorithms for (1 − )maximum weight f -factor and (1 + )-approximate minimum weight f -edge cover. As a byproduct, we p
also obtain direct algorithms for the exact cardinality versions of these problems running
in O(m f (V )) time.
The technical contributions of this work include an efficient method for maintaining relaxed complementary slackness in generalized matching problems and approximation-preserving
reductions between the f -factor and f -edge cover problems.
∗
Supported by NSF grants CCF-1217338, CNS-1318294, CCF-1514383, and CCF-1637546.
1
Introduction
Many combinatorial optimization problems are known to be reducible to computing optimal matchings in non-bipartite graphs [8, 7]. These problems include computing b-matchings, f -factors, f edge covers, T -joins, undirected shortest paths (with no negative cycles), and bidirected flows,
see [17, 12, 21, 9]. These problems have been investigated heavily since Tutte’s work in the 1950s
[22, 20]. However, the existing reductions to graph matching are often inadequate: they blow up the
size of the input [17], use auxiliary space [10], or piggyback on specific matching algorithms [10] like
the Micali-Vazirani algorithm [18]. Moreover, most existing reductions destroy the dual structure
of optimal solutions and are therefore not approximation preserving.
In this paper we design algorithms for computing f -factors and f -edge covers in a direct fashion,
or through efficient, approximation-preserving reductions. Because our algorithms are based on the
LP formulations of these problems (in contrast to approaches using shortest augmenting walks [10,
18]), they easily adapt to weighted and approximate variants of the problems. Let us define these
problems formally. We assume graphs may have multiple parallel edges and loops.
f -factor An f -factor is a subset F ⊂ E such that degF (v) ≤ f (v). F is perfect if the degree
constraints hold with equality.1
f -edge cover An f -edge cover is a subset F ⊂ E such that degF (v) ≥ f (v). It is perfect if all
degree constraints hold with equality.
On unweighted graphs the f -factor objective is to maximize |F |, and for f -edge cover it is
to minimize |F |. On weighted graphs it is to maximize/minimize w(F ), possibly subject to the
additional constraint that F is perfect.
Classic Reductions. The classical reduction from f -factor to standard graph matching uses the
b-matching problem as a stepping stone. A b-matching is a function
x : E → Z≥0 (where x(e)
P
indicates how many copies of e are in the matching) such that e3v x(e) ≤ b(v), i.e., the number
of matches edges incident to v, counting multiplicity, is at most b(v). The f -factor problem on
G = (V, E, w) is reduced to b-matching by subdividing each edge e = (u, v) ∈ E into a path
(u, ue , ve , v). Here ue , ve are new vertices, w(u, ue ) = w(ve , v) = w(u, v) while w(ue , ve ) = 0, and
b(ue ) = b(ve ) = 1. The original vertices have b(u) = f (u). This reduction blows up the number of
vertices to O(m) and is not approximation preserving. The b-matching problem is easily reduced to
standard matching by replicating each vertex u b(u) times, and replacing each edge (u, v) with a
bipartite b(u) × b(v) clique on its endpoints’ replicas. This step of the reduction is approximation
preserving, but blows up the number of vertices and edges. Both reductions together reduce f factor to a graph matching problempon O(m) vertices and O(fmax m) edges. Gabow [10] gave a
method for solving f -factor in O(m f (V )) time using black-box calls to single iterations of the
Micali-Vazirani [18] algorithm.
Observe that f -factor and f -edge cover are complementary problems: if C is an fC -edge cover,
the complementary edge set F = E\C is necessarily an fF -factor, where fF (v) = deg(v) − fC (v).
Complementarity implies that any polynomial-time algorithm for one problem solves the other
in polynomial time, but it says nothing about the precise complexity of solving them exactly or
approximately. Indeed, this phenomenon is very well known in the realm of NP-complete problems.
1
In the literature the term f -factor often refers to our definition of a perfect f -factor.
1
For example, Maximum Independent Set and Minimum Vertex Coverpare complementary problems,
but have completely different approximation profiles. Gabow’s O(m
p fF (V )) cardinality fF -factor
algorithm [10] implies that fC -edge cover is computed in O(m m − fC (V )) = O(m3/2 ) time,
and says nothing about the approximability of fC -edge cover. As far as we are aware, the fastest
approximation algorithms for fC -edge cover (see [2]) treat it as a general weighted Set Cover problem
on 2-element sets. Chvátal’s analysis [3] shows the greedy algorithm is an H(2)-approximation,
where H(2) = 3/2 is the 2nd harmonic number.
Our interest in the approximate f -edge cover problem is inspired by a new application to
anonymizing data in environments where users have different privacy demands; see [2, 1, 16]. Here
the data records correspond to edges and the privacy demand of v is measured by f (v); the goal is
to anonymize as few records to satisfy everyone’s privacy demands.
New Results. We give new algorithms for computing f -factors and f -edge covers approximately
and exactly.
• We show that a folklore reduction from minimum weight 1-edge cover to maximum weight
1-factor (matching) is approximation-preserving, in the sense that any (1 − )-approximation
for matching gives a (1 + )-approximation for edge cover. This implies that 1-edge cover can
be (1 + )-approximated in O (m) time [5], and that one can apply any number of simple and
practical algorithms [4, 19, 5] to approximate 1-edge cover. This simple reduction does not
extend to f -factors/f -edge covers.
• We give an O (m)-time (1 + )-approximation algorithm for weighted fC -edge cover, for any
fC . Our algorithm follows from two results, both of which are somewhat surprising. First,
any approximate weighted fF -factor algorithm that reports a (1 ± )-optimal dual solution
can be transformed into a (1 + O())-approximate weighted fC -edge cover algorithm. Second,
such an fF -factor algorithm exists, and its running time is O (m). The first claim is clearly
false if we drop the approximate dual solution requirement (for the same reason that an O(1)approximate vertex cover does not translate into an O(1)-approximate maximum independent
set), and the second is surprising because the running time is independent of the demand
function fF and the magnitude of the edge weights.
• As corollaries of these reductions,
p we obtain a new exact algorithm for minimum cardinality
fC -edge cover running in O(m fC (V )) time, ratherpthan O(m3/2 ) time ([10]), and a direct
algorithm for cardinality fF -factor that runs in O(m fF (V )) time, without reduction [10] to
the Micali-Vazirani algorithm [18].
The blossom structure and LP characterization of b-matching is considerably simpler than the
corresponding blossoms/LPs for f -factor and f -edge cover. In the interest of simplicity, one might
want efficient code that solves (approximate) b-matching directly, without viewing it as an f -factor
problem on a multigraph in which there is implicitly an infinite supply of each edge. We do not
know of such a direct algorithm. Indeed, the structure of b-matching blossoms seems to rely on strict
complementary slackness, and is incompatible with our main technical tool, relaxed complementary
slackness.2
2
Using relaxed complementary slackness, matched and unmatched edges have different eligibility criteria (to be
included in augmenting paths and blossoms) whereas b-matching blossoms require that all copies of an edge—matched
and unmatched alike—are all eligible or all ineligible.
2
Structure of the Paper. In Section 2 we give an introduction to the LP-formulation of generalized matching problems and the structure of their blossoms and augmenting walks. In Section 3.1
we show that a folklore reduction from 1-edge cover to 1-factor is approximation-preserving and in
Section 3.2 we reduce approximate f -edge cover to approximate f -factor. In Section 4 we give an
O(W m−1 )-time algorithm for (1 − )-approximate f -factor in graphs with weights in [1, W ] and
then speed it up to O(m−1 log −1 ), independent of weight. Section 5 gives a linear time algorithm
to compute a maximal set of augmenting walks; cf. [15, §8].
2
Basis of f -factor and f -edge cover
This section reviews basic algorithmic concepts from matching theory and their generalizations to
the f -factor and f -edge cover problems, e.g., LPs, blossoms, and augmenting walks. These tools
will let us generalize the Duan-Pettie algorithm [5] for Approximate Maximum Weight Matching to
Approximate Maximum Weight f -factor and Approximate Minimum Weight f -edge cover.
Notation. The input is a multigraph G = (V, E). For S ⊆ V , let δ(S) and γ(S) be the sets of
edges with exactly one endpoint and both endpoints in S, respectively. For T ⊆ E, δT (S) denotes
the intersection of δ(S) and T . By definition, degT (S) = |δT (S)|.
2.1
LP formulation
The maximum weight f -factor problem can be expressed as maximizing
the following constraints:
X
x(e) ≤ f (v), for all v ∈ V
P
e∈E
w(e)x(e), subject to
e∈δ(v)
X
e∈γ(B)∪I
f (B) + |I|
x(e) ≤
, for all B ⊆ V, I ⊆ δ(B)
2
(1)
0 ≤ x(e) ≤ 1, for all e ∈ E
j
P
f (B)+|I|
2
k
Here, the blossom constraint e∈γ(B)∪I x(e) ≤
is a generalization of blossom conj k
P
|B|
straint e∈γB x(e) ≤ 2 in ordinary matching. The reason that we have a subset I of incident
edges in the sum is that the subset allow us to distinguish between matched edges that have both
endpoints inside B with those with exactly one endpoints. Any basic feasible solution x of this LP
is integral [21, §33], and can therefore be interpreted as membership vector of an f -factor F . To
ensure optimality of the solution, the algorithm works with the dual LP, which is:
X
X
X
f (B) + |I|
minimize
f (v)y(v) +
z(B, I) +
u(e)
2
e
v∈V
B⊆V,I⊆δ(B)
(2)
subject to yzF (e) + u(e) ≥ w(e), for all e ∈ E
y(v) ≥ 0, z(B, I) ≥ 0, u(e) ≥ 0
3
Here the aggregated dual yzF : E 7→ R≥0 is defined as:
X
yzF (u, v) = y(u) + y(v) +
z(B, I).
B,I:(u,v)∈γ(B)∪I,
I⊆δ(B))
Unlike matching, each z-value here is associated with the combination of a vertex set B and a
subset I of its incident edges.
P
The minimum weight f -edge cover problem can be expressed as minimizing
e∈E w(e)x(e),
subject to:
X
x(e) ≥ f (v), for all v ∈ V
e∈δ(v)
X
e∈γ(B)∪(δ(B)\I)
f (B) − |I|
x(e) ≥
, for all B ⊆ V and I ⊆ δ(B)
2
(3)
0 ≤ x(e) ≤ 1, for all e ∈ E
With the dual program being:
maximize
X
v∈V
X
f (v)y(v) +
B⊆V,I⊆δ(B)
X
f (B) − |I|
z(B, I) −
u(e)
2
e∈E
subject to yzC (e) − u(e) ≤ w(e), for all e ∈ E
(4)
y(v) ≥ 0, z(B, I) ≥ 0, u(e) ≥ 0
where
X
yzC (u, v) = y(u) + y(v) +
z(B, I).3
B,I:(u,v)∈γ(B)∪(δ(B)\I),I⊆δ(B))
Both of our f -factor and f -edge cover algorithms maintain a dynamic feasible solution F ⊆ E
that satisfies the primal constraints. We call edges in F matched and all other edges unmatched,
which is referred to as the type of an edge. A vertex v is saturated if degF (v) = f (v). It is unsaturated/oversaturated if degF (v) is smaller/greater than f (v). Given an f -factor F , the deficiency
def(v) of a vertex v is defined as def(v) = f (v) − degF (v). Similarly, for an f -edge cover F , the
surplus of a vertex is defined as surp(v) = degF (v) − f (v).
2.2
Blossoms
We follow Gabow’s [11] definitions and terminology for f -factor blossoms, augmenting walks, etc.
A blossom is a tuple (B, EB , β(B), η(B)) where B is the vertex set, EB the edge set, β(B) ∈ B the
base vertex and η(B) ⊂ δ(β(B)) ∩ δ(B), |η(B)| ≤ 1 the base edge set, which may be empty. We
often refer to the blossom by referring to its vertex set B. Blossoms can be defined inductively as
follows.
Definition 1. A single vertex v forms a trivial blossom, or a singleton. Here B = {v}, EB = ∅,
β(B) = v, and η(B) = ∅.
3
We use yzF and yzC to denote the aggregated dual yz for f -factor and f -edge cover respectively. We will omit
the subscript if it is clear from the context.
4
Inductively, let B0 , B1 , ... , Bl−1 be a sequence of disjoint singletons or nontrivial blossoms. Suppose there exists a closed walk CB = {e0 , e1 , ...S, el−1 } ⊆ E starting and ending with B0 such that
ei ∈ Bi × Bi+1 (mod l). The vertex set B = l−1
i=0 Bi is identified with a blossom if the following
are satisfied:
1. Base Requirement: If B0 is a singleton, the two edges incident to B0 on CB , i.e., e0 and el−1 ,
must both be matched or both be unmatched.
2. Alternation Requirement: Fix a Bi , i 6= 0. If Bi is a singleton, exactly one of ei−1 and ei is
matched. If Bi is a nontrivial blossom, η(Bi ) 6= ∅ and must be either {ei−1 } or {ei }.
S
The edge set of the blososm B is EB = CB ∪ ( l−1
i=1 EBi ) and its base is β(B) = β(B0 ). If B0
is not a singleton, η(B) = η(B0 ). Otherwise, η(B) may either be empty or contain one edge in
δ(B) ∩ δ(B0 ) that is the opposite type of e0 and el−1 .
Blossoms are classified as light/heavy. If B0 is a singleton, B is light/heavy if e0 and el−1 are
both unmatched/matched. Otherwise, B is light/heavy if B0 is light/heavy. Note that blossoms in
the usual matching problem (1-factor) are always light.
The purpose of blossoms is to identify the part of graph that behaves as a unit when searching
for an augmenting walk. This property can be formally stated as follows:
Lemma 2. Let v be an arbitrary vertex in B. There exists an even length alternating walk P0 (v)
and an odd length alternating walk P1 (v) from β(B) to v using edges in EB . Moreover, The terminal
edge incident to β(B) must have a different type than η(B), if η(B) exists.
Proof. We prove this by induction. The base case is a blossom B consisting of singletons hv1 , v2 , ... , v2k+1 i,
where v = vi . Then one of the two walks hv1 , v2 , ... , vi i and hv1 , v2k+1 , v2k ... , vi i must be odd and
the other must be even.
Consider the cycle CB = hB0 , e0 , B1 , ... , el−2 , Bl−1 el−1 , B0 i. Suppose the claim holds inductively
for all nontrivial blossoms in B0 , B1 , ... , Bl−1 . Let v be an arbitrary vertex in B. We use PBi ,j
(0 ≤ i < l, j ∈ {0, 1}) to denote the walk P0 and P1 guaranteed in blossom Bi . There are two cases:
Case 1: When v is contained in a singleton Bk . We examine the two walks Pb = he0 , e1 , ... ek−1 i
and Pb0 = hel−1 , el−2 , ... ek i. Notice that Pb and Pb0 are walks in the graph obtained by contracting all
subblossoms B0 , B1 , ... Bl−1 of B. By the inductive hypothesis, we can extend Pb and Pb0 to P and
P 0 in G by replacing each Bi on the walk with PBi ,j for suitable j connecting the endpoints of ei−1
and ei . In particular, if ei−1 and ei are of different types, we replace Bi with the even length walk
PBi ,0 . Otherwise, we replace with PBi ,1 . By the alternation requirement, one of P and P 0 must be
odd and the other must be even.
Case 2: When v is contained in a non-trivial blossom Bk . Without loss of generality, ek−1
η(Bk ). Consider the contracted walk Pb = he0 , e1 , ... , ek−1 i. We extend Pb to an alternating walk
in EB terminating at ek−1 similar to Case 1. Then P0 (v) or P1 (v) is obtained by concatenating
with the alternating walk PBk ,0 (v) or PBk ,1 (v) of the right parity.
Notice that in both cases, the base requirement in Definition 1 guarantees the starting edge
the alternating walk P1 (v) and P0 (v) alternates with the base edge η(B).
=
P
P
of
The main difference between blossoms in generalized matching problems and blossoms in ordinary matching problem is that P0 and P1 are both meaningful for finding augmenting walks or
blossoms. In ordinary matching, since each vertex has at most 1 matched edge incident to it, an
5
alternating walk enters the blossom at the base vertex via a matched edge and must leave with
an unmatched edge. As a result the subwalk inside the blossom is always even. In generalized
matching problems, this subwalk can be either even or odd, and may contain a cycle. In general,
an alternating walk enters the blossom at the base edge and can leave the blossom at any nonbase
edge.
We also define the notion of maturity for blossoms, for both f -factor and f -edge cover. For
simplicity let us focus on f -factor first. By complementary slackness, we can only assign a positive
z(B, I) for the pair (B, I) if it satisfies the constraint |F ∩ (γ(B) ∪ I)| ≤ b(f (B) + |I|)/2c with
equality. This requirement can be captured in the following definition:
Definition 3 (Mature Blossom). A blossom is mature w.r.t an f -factor F if it satisfies the following:
1. Every vertex v ∈ B \ {β(B)} is saturated: degF (v) = f (v).
2. def(β(B)) = 0 or 1. If def(β(B)) = 1, B must be a light blossom and η(B) = ∅; If def(β(B)) =
0, η(B) 6= ∅.
This is motivated by two observations:
1. Complementary slackness: dual variables can be positive only if its primal constraint is satisfied
with kequality, i.e. a blossom can have a positive z-value only if |F ∩ (γ(B) ∪ I(B))| =
j
f (B)+I(B)
.
2
2. Topology of Augmenting Walks: Unsaturated heavy blossoms cannot start or end an augmenting walk since we cannot extend it to an augmenting walk in G that starts with an unmatched
edge.
Several remarks can be made here:
• A blossom that is not mature may contain an augmenting walk. Specifically, suppose B is light
and unsaturated. If any nonbase vertex v 6= β(B) in B is also unsaturated, the odd length
alternating walk from β(B) to v satisfies the definition of an augmenting walk. Moreover,
if β(B) has deficiency of 2 or more, the odd length alternating walk from β(B) to β(B) is
augmenting. For these reasons we never contract immature blossoms.
j
k
• The maturity of an f -factor blossom B will imply |F ∩ (γ(B) ∪ I)| = f (B)+I
, for some set
2
Il ⊆ δ(B).
m Similarly, a mature f -edge cover blossom B satisfies |F ∩ (γ(B) ∪ (δ(B) \ I))| =
f (B)−I
for some I ⊆ δ(B).
2
• Augmentation never destroys maturity. In particular, it never creates an unsaturated heavy
blossom.
Maturity for f -edge cover blossoms can be defined similarly and has similar properties.
Definition 4 (Mature Blossom for f -edge cover). A blossom is mature w.r.t an f -edge cover F if
it satisfies the following:
1. Every vertex v ∈ B \ {β(B)} is saturated: degF (v) = f (v).
2. surp(β(B)) = 0 or 1. If surp(β(B)) = 1, B must be a heavy blossom and η(B) = ∅; If
surp(β(B)) = 0, η(B) 6= ∅.
6
v0
v0
v1
v1
v2
v2
v3
v4
v12
v3
v4
v12
v5
v6
v13
v5
v6
v13
v7
v8
v10
v9
v7
v14
v8
v9
v10
v11
v15
v16
v11
v14
v15
v16
Figure 1: Two example of contractable blossoms: Bold edges are matched and thin ones are unmatched. Blossoms are circled with border. Base edges are represented with arrow pointing away
from the blossom.
The algorithm keeps track of a laminar set Ω ⊂ 2V of mature blossoms and maintains a nonnegative z value for each B ∈ Ω with one subset of I(B) ⊆ δ(B) of its neighborhood, defined as
follows.
I(B) = δF (B) ⊕ η(B),
where ⊕ is the symmetric difference operator (XOR). All other subsets
I of δ(B)
j
k will have z(B, I) =
f (B)+|I(B)|
0. If B is a mature blossom, then we have |F ∩ (γ(B) ∪ I(B))| =
(or in f -edge cover:
2
l
m
|F ∩ (γ(B) ∪ (δ(B) \ I(B)))| = f (B)−|I(B)|
).
2
2.3
Augmenting/Reducing Walk
Augmenting walks are the analogy for augmenting paths from ordinary matching. Complications
arise from the fact that an f -factor blossom cannot be treated identically to a single vertex after
it is contracted. For example, in Figure 2, the two edges (v0 , v1 ) and (v4 , v6 ) incident to blossom
{v1 , v2 , v4 , v5 , v3 } are of the same type, both before and after augmenting along hv0 , v1 , v3 , v5 , v4 , v6 i.
This can never happen in ordinary matching! Moreover, augmenting walks can begin and end at the
same vertex and can visit the same vertex multiple times. Hence a naive contraction of a blossom
into a single vertex loses key information about the internal structure of blossoms. Definition 5
characterizes when a walk in the contracted graph can be extended to an augmenting walk.
b be the graph obtained from G by contracting a laminar set Ω of blossoms. Let
Definition 5. Let G
b
b Here {ei } are edges and {Bi } are blossoms or
P = hB0 , e0 , B1 , e1 , ... Bl−1 , el−1 , Bl i be a walk in G.
singletons, with ei ∈ Bi × Bi+1 for all 0 ≤ i < l. We say Pb is an augmenting walk with respect to
the f -factor F if the following requirements are satisfied:
7
v0
v0
v1
v1
v2
v3
v2
v3
v4
v5
v4
v5
v6
v6
Figure 2: An example for how a blossom changes with an augmentation: here the augmenting walk
is hv0 , v1 , v3 , v5 , v4 , v6 i. Notice after rematching, the base edge of the blossom changes from (v0 , v1 )
to (v4 , u6 ), and the blossom turns from a heavy blossom to a light one.
1. Terminal Vertices Requirement: The terminals B0 and Bl must be unsaturated singletons
or unsaturated light blossoms. If P is a closed walk (B0 = Bl ), B0 must be a singleton and
def(β(B0 )) ≥ 2. Otherwise B0 and Bl can be either singletons or blossoms and their deficiency
must be positive.
2. Terminal Edges Requirement: If the terminal vertex B0 (Bl ) is a singleton, the incident terminal edges e0 (el−1 ) must be unmatched. Otherwise they can be either matched or unmatched.
3. Alternation Requirement: Let Bi , 0 < i < l, be an internal singleton or blossom. If Bi is a
singleton, the exactly one of ei−1 and ei is matched. If Bi is a nontrivial blossom, η(Bi ) 6= ∅
and must be one of {ei−1 } or {ei }.
b can be extended
A natural consequence of the above definition is that an augmenting walk in G
to an augmenting walk in G. This is proved exactly as in Lemma 2. Rematching F along an
augmenting walk P means updating F to the symmetric difference F ⊕ P . This will decrease the
total deficiency by 2. Moreover, after rematching along P , every blossom B ∈ Ω still satisfies the
definition for blossoms (with different base vertices and edges).
In f -edge cover, the corresponding notion is called reducing walk. The definition of reducing
walk can be naturally obtained from Definition 5 while replacing “unsaturated”, “deficiency” and
“light” with “oversaturated”, “surplus” and “heavy”. It is also worth pointing out that if an f -factor
F and an f 0 -edge cover F 0 are complement to each other, i.e. F 0 = E \ F and f (v) + f 0 (v) = deg(v),
and they have the same blossom set Ω, then an augmenting walk Pb for F is also a reducing walk
for F 0 .
8
2.4
Complementary Slackness
To characterize an (approximately) optimal solution, we maintain dual functions: y : V 7→ R≥0 and
z : 2V 7→ R≥0 . Here z(B) is short for z(B, I(B)). We do not explicitly maintain the edge dual
u : E 7→ R≥0 since its maximizing value can be explicitly given by u(e) = max{w(e) − yz(e), 0}.
For f -factor F , the following property characterizes an approximate maximum weight f -factor:
Property 1 (Approximate Complementary Slackness for f -factor). Let δ1 , δ2 ≥ 0 be nonnegative
parameters. We say an f -factor F , duals y, z, and the set of blossoms Ω satisfies (δ1 , δ2 )-approximate
complementary slackness if the following hold:
1. Approximate Domination. For each unmatched edge e ∈ E \ F , yz(e) ≥ w(e) − δ1 .
2. Approximate Tightness. For each matched edge e ∈ F , yz(e) ≤ w(e) + δ2 .
j
k
3. Blossom Maturity. For each blossom B ∈ Ω, |F ∩ (γ(B) ∪ I(B))| = f (B)+|I(B)|
.
2
4. Unsaturated Vertices’ Duals. For each unsaturated vertex v, y(v) = 0.
The following lemma states Property 1 characterizes an approximately optimal solution.
Lemma 6. Let F be an f -factor in G along with duals y, z and let F ∗ be the maximum weight
f -factor. If F, Ω, y, z satisfy Property 1 with parameters δ1 and δ2 , we have
w(F ) ≥ w(F ∗ ) − δ1 |F ∗ | − δ2 |F |.
Proof. We first define u : E 7→ R as
(
w(e) − yz(e) + δ2 ,
u(e) =
0,
if e ∈ F .
otherwise.
From Approximate Tightness, we have u(e) ≥ 0 for all e ∈ E. Moreover, yz(e) + u(e) ≥ w(e) − δ1
for all e ∈ E and yz(e) + u(e) = w(e) + δ2 for all e ∈ F . This gives the following:
X
X
w(F ) =
w(e) =
(yz(e) + u(e) − δ2 )
e∈F
=
X
e∈F
degF (v)y(v) +
v∈V
X
|F ∩ (γ(B) ∪ I(B))|z(B) +
B∈Ω
X
u(e) − |F |δ2
e∈F
By Property 1 (Unsaturated Vertices’ Duals, Blossom Maturity, and the definition of u), this is
equal to
X
X
X f (B) + |I(B)|
z(B) +
u(e) − |F |δ2
=
f (v)y(v) +
2
e∈E
v∈V
B∈Ω
X
X
X
∗
≥
degF ∗ (v)y(v) +
|F ∩ (γ(B) ∪ I(B))|z(B) +
u(e) − |F |δ2
v∈V
=
X
e∈F ∗
B∈Ω
(yz(e) + u(e)) − |F |δ2
e∈F ∗
≥
X
(w(e) − δ1 ) − |F |δ2 = w(F ∗ ) − |F ∗ |δ1 − |F |δ2 .
e∈F ∗
9
Connection Between f -Factors and f -Edge Covers
3
The classical approach for solving fC -edge cover problem is reducing it to fF -factor. Specifically,
looking for a minimum weight fC -edge cover C can be seen as choosing edges that are not in C,
which is a maximum weight fF -factor where fF (u) = deg(u) − fC (u).
The main drawback of this reduction is that it yields inefficient algorithms. For example,
Gabow’s algorithms [11] for solving maximum weight fF -factor scales linearly with fF (V ), which
makes it undesirable when fC is small. Even for fC (V ) = O(n), Gabow’s algorithm runs in
O(m2 + mn log n) time. Moreover, this reduction is not approximation-preserving. In other words,
the complement of an arbitrary (1 − )-approximate maximum weight fF -factor is not guaranteed
to be a (1 + )-approximate fC -edge cover.
In this section we establish two results: First we prove that a folklore reduction from 1-edge cover
to matching is approximation preserving. This allows us to use an efficient approximate matching
algorithm to solve the weighted 1-edge cover problem. Then we establish the connection between
approximate fF -factor and approximate fC -edge cover using approximate complementary slackness
from the previous section. This will give a (1 + )-approximate minimum weight fC -edge cover
algorithm from our (1 − ) approximate maximum weight fF -factor algorithm.
3.1
Approximate Preserving Reduction from 1-Edge Cover to 1-Factors
The edge cover problem is a special case of f -edge cover where f is 1 everywhere. The minimum
weight edge cover problem is reducible to maximum weight matching, simply reweighting edges [21].
Let e(v) be any edge with minimum weight in δ(v) and let µ(v) = w(e(v)). Define a new weight
function w0 as follows
w0 (u, v) = µ(u) + µ(v) − w(u, v).
Schrijver [21, §27] showed the following theorem:
Theorem 7. Let M ∗ be a maximum weight matching with respect to weight function w0 , and C =
M ∗ ∪{e(v) : v ∈ V \V (M )}. Then C is a minimum weight edge cover with respect to weight function
w.
We show this reduction is also approximation preserving.
Theorem 8. Let M 0 be a (1 − )-maximum weight matching with respect to weight function w0 , and
C 0 = M 0 ∪ {e(v) : v ∈ V \ V (M 0 )}. Then C 0 is a (1 + )-minimum weight edge cover with respect to
weight function w.
Proof. Let C ∗ and M ∗ be the optimal edge cover and matching defined previously. By construction,
we have
w(C 0 ) = w(M 0 ) + µ(V \ V (M 0 )) = µ(V (M 0 )) − w0 (M 0 ) + µ(V \ V (M 0 )) = µ(V ) − w0 (M 0 )
Similarly, we have w(C ∗ ) = µ(V ) − w0 (M ∗ ). Then
w(C 0 ) = µ(V ) − w0 (M ) ≤ µ(V ) − (1 − )w0 (M ∗ ) = w(C ∗ ) + w0 (M ∗ ) ≤ (1 + )w(C ∗ ).
The last inequality holds because we have w0 (u, v) = µ(u) + µ(v) − w(u, v) ≤ w(u, v).
The reduction does not naturally extend to f -edge cover. In the next section we will show how
to obtain a (1 + )-approximate f -edge cover algorithm from a (1 − )-approximate f -factor within
the primal-dual framework.
10
3.2
From f -factor to f -edge cover
We show that a primal-dual algorithm computing a (1−)-approximate f -factor can used to compute
an (1 + )-approximate f -edge cover. We start by giving the approximate complementary slackness
for f -edge cover. Similar to f -factor, the property characterizes a good approximate f -edge cover.
Property 2 (Approximate Complementary Slackness for f -edge cover). Let δ1 , δ2 ≥ 0 be positive
parameters. We say an f -edge cover C, with duals y, z and blossom family Ω satisfies the (δ1 , δ2 )approximate complementary slackness if the following requirements holds:
1. Approximate Domination. For each unmatched edge e ∈ E \ C, yzC (e) ≤ w(e) + δ1 .
2. Approximate Tightness. For each matched edge e ∈ C, yzC (e) ≥ w(e) − δ2 .
3. Blossom Maturity. For each blossom B ∈ Ω, |C ∩ (γ(B) ∪ (δ(B) \ IC (B)))| =
l
f (B)−|IC (B)|
2
m
.
4. Oversaturated Vertices’ Duals. For each oversaturated vertex v, y(v) = 0.
Recall that we are using the aggregated duals yzC for f -edge cover:
X
yzC (u, v) = y(u) + y(v) +
z(B)
B:(u,v)∈γ(B)∪(δ(B)\IC (B))
The proof for the following Lemma 9 is identical to Lemma 6.
Lemma 9. Let C be an f -edge cover with duals y, z, Ω satisfying Property 2 with parameters δ1
and δ2 , and let C ∗ be the minimum weight f -edge cover. We have w(C) ≤ w(C ∗ ) + δ1 |C ∗ | + δ2 |C|.
Suppose we have an f 0 -factor F with blossoms Ω and duals y, z satisfying Property 1 and let C
be an f -edge cover that is F ’s complement. The key observation is that the same blossom set Ω
and duals y, z can be used to show Property 2 for C. Specifically, we have the following lemma,
Lemma 10. If the duals y, z, Ω and an f 0 -factor F satisfy Property 1 with parameters δ10 , δ20 , then
the same duals y, z, Ω and the complementary f -edge cover C = E \ F satisfies Property 2 with
parameter δ1 = δ20 and δ2 = δ10 .
Proof. It is trivial to see Property 2(4) (Oversaturated Vertices’ Duals) and Property 1(4) (Unsaturated Vertices’ Duals) are equivalent to each other.
To show Property 2(1,2) is equivalent to Property 1(2,1), it suffices to show that the function
yzF for f 0 -factor F agrees with the function yzC for its f -edge cover complement C. Recall that
X
yzC (u, v) = y(u) + y(v) +
z(B)
B:(u,v)∈γ(B)∪(δ(B)\IC (B))
X
yzF (u, v) = y(u) + y(v) +
z(B)
B:(u,v)∈γ(B)∪IF (B)
Here IC and IF refer to the I-sets of a blossom with respect to the f -edge cover C and the
f 0 -factor F . It suffices to show that IF (B) = δ(B) \ IC (B):
IF (B) = η(B) ⊕ δF (B) = η(B) ⊕ (δ(B) ⊕ δC (B))
= δ(B) ⊕ (η(B) ⊕ δC (B)) = δ(B) \ IC (B)
11
v0
v0
v1
v6
v1
v7
v2
v3
v8
v4
v5
v9
v6
v7
v2
v3
v8
v4
v5
v9
Figure 3: Illustration on relation between I-set of an f -factor and the I-set of its complementary
f 0 -edge cover. Here figures on the left demonstrate an f -factor and its complementary f 0 -cover is
on the right. Their I-sets are circled (dashed).
Therefore, in yzF (e) and yzC (e), z-values are summed up over the same set of blossoms in Ω. In
other words, yzF (e) = yzC (e) for each e ∈ E and the claim follows
For Blossom Maturity in Property 1 and Property 2, we argue that one equality implies the other.
Suppose η(B) ⊂ F . We have IC (B) = δC (B) \ η(B). Therefore the edge set C ∩ (γ(B) ∪ (δ(B) \
IC (B))) contains edges with both endpoints inside B plus the edge η(B), while f (B) − |IC (B)| =
f (B) − degC (B) + 1 is a lower bound on how
this set of edges can have inside B.
l many endpoints
m
f (B)−|IC (B)|
Equality |C ∩ (γ(B) ∪ (δ(B) \ IC (B)))| =
implies f (v) = degC (v) for every v ∈ B.
2
This further implies that f 0 -factor F also saturates every vertex v ∈ B w.r.t. f 0 . Therefore, equality
in Property 1(3) holds. If η(B) 6⊂ C, then IC (B) = δC (B), the edge set C ∩ (γ(B) ∪ (δ(B) \ IC (B)))
contains edges with both endpoints inside B only. Similarly, f (B) − |IC (B)| = f (B) − degC (B) is
al lower bound
m on the total number of their endpoints. Equality |C ∩ (γ(B) ∪ (δ(B) \ IC (B)))| =
f (B)−|IC (B)|
again implies every vertex in B is saturated so Property 1(3) is satisfied. Notice
2
that the same arugment also applies to the case when η(B) = ∅. This completes the proof that
Property 2(3) implies Property 1(3). The other direction is symmetric.
4
Approximation Algorithms for f -Factor and f -Edge Cover
In this section, we prove the main result by giving an approximation algorithm for computing
(1 − )-approximate maximum weight f -factor. The crux of the result is an implementation of
Edmonds’ search with relaxed complementary slackness as the eligibility criterion. The notion
of approximate complementary slackness was introduced by Gabow and Tarjan for both bipartite
matching [14] and general matching [15]. Gabow gave an implementation of Edmonds’ search with
exact complementary slackness for f -factor problem [11], which finds augmenting walks one at a
time. The main contribution of this section is to adapt [11] to approximate complementary slackness
to facilitate finding augmenting walks in batches.
To illustrate how this works, we will first give an approximation algorithm for f -factor in graph
with small edge weights. Let w(·) be a positive weight function w : E 7→ {1, ... , W }. The algorithm
12
computes a (1 − )-approximate maximum weight f -factor in O(mW −1 ) time, independent of f .
We also show how to use scaling techniques to transform this algorithm to run in O(m−1 log −1 )
time, independent of W .
4.1
Approximation for small weights
The main procedure in our O(mW −1 ) time algorithm is called Edmonds’ search. In one iteration,
Edmonds’ search finds a set of augmenting walks using eligible edges, creates and dissolve blossoms,
and performs Dual Adjustment on y and z while maintaining the following Invariant:
Invariant 1 (Approximate Complementary Slackness). Let δ > 0 be some parameter such that w(e)
is a multiple of δ, for all e ∈ E:
1. Granularity. y-values are multiples of δ/2 and z-values are multiples of δ.
2. Approximate
Domination. For each unmatched edge and each blossom edge e ∈ (E \ F ) ∪
S
( B∈Ω EB ), yz(e) ≥ w(e) − δ.
S
3. Tightness. For each matched and each blossom edge e ∈ F ∪ ( B∈Ω EB ), yz(e) ≤ w(e).
j
k
4. Blossom Maturity. For each blossom B ∈ Ω, |F ∩ (γ(B) ∪ I(B))| = f (B)+|I(B)|
.
2
5. Unsaturated Vertices. All unsaturated vertices have the same y-value; their y-values are strictly
less than the y-values of other vertices.
Notice that here we relax Property 1(4) to allow unsaturated vertices to have positive y-values.
The purpose of Edmonds’ search is to decrease the y-values for all unsaturated vertices while maintaining Invariant 1. Following [15, 5, 6], we define the following eligibility criterion:
Criterion 1. An edge (u, v) is eligible if it satisfies one of the following:
1. e ∈ EB for some B ∈ Ω.
2. e 6∈ F and yz(e) = w(e) − δ.
3. e ∈ F and yz(e) = w(e).
A key property of this definition is that it is asymmetric for matched and unmatched edges. As
a result, if we augment along an eligible augmenting walk P , all edges in P , except for those in
contracted blossoms, will become ineligible.
[
Let Gelig be the graph obtained from G by discarding all ineligible edges, and let G
elig = Gelig /Ω
be obtained from Gelig by contracting all blossoms in Ω. For initialization, we set F = ∅, y = W/2,
z = 0, Ω = ∅. Edmonds’ search repeatedly executes the following Augmentation and Blossom
Formation, Dual Adjustment, and Blossom Dissolution steps until all unsaturated vertices have 0
y-values. See Figure 4.
Let us define what we mean by reachable vertices in Steps 1–3 of the algorithm, as well as
inner/outer labelling of blossoms and singletons. We start by defining the notion of alternation
that follows from Definition 5 of an augmenting walk. We say two edges e, e0 incident to a blossom/singleton B alternate if either B is a singleton and e and e0 are of different types, or B is a
13
b of augmenting path with
1. Augmentation and Blossom Formation. Find a maximal set Ψ
0
b in Gelig .
[
a maximal set of reachable mature blossom Ω in Gelig ; Ψ be the preimage of Ψ
S
[
Update F ← F ⊕ P ∈Ψ P and Ω ← Ω ∪ Ω0 . After this step, the new G
elig contains no
4
augmenting walk as well as no reachable blossoms .
[
2. Dual Adjustment. Let Sb be the set of vertices from G
elig reachable from an unsaturated
vertex via an eligible alternating walk. We classify vertices in Sb into Vc
in , the set of inner
5
vertices and Vd
out , the set of outer vertices . Let Vin and Vout be the set of original vertices
d
in V represented by Vc
in and Vout . Adjust the y and z values as follows:
y(v) ← y(v) − δ/2, if v ∈ Vout
y(v) ← y(v) + δ/2, if v ∈ Vin
z(B) ← z(B) + δ, if B is a root blossom in Vbout
z(B) ← z(B) − δ, if B is a root blossom in Vbin
3. Blossom Dissolution. After Dual Adjustment some root blossoms in Ω might have 0 zvalue. Remove them from Ω until none exists. Update Ω and Gelig . Notice that root
blossoms with 0 z-value can be generated because some root blossoms have their z-value
decremented in Dual Adjustment step, or root blossoms formed in Augmentation step
become unreachable after augmentation and thus do not get their z-value incremented.
Figure 4: A (1 − )-approximate f -factor algorithm for small integer weights.
nontrivial blossom, and |η(B) ∩ {e, e0 }| = 1. An alternating path P in the contracted graph is one
where every two consecutive edges alternates.
The search forest Sb consists of blossoms and singletons that are reachable from some unsaturated
[
blossom/singleton, via an eligible alternating path in G
elig . Furthermore, we require that the
preimage of those paths in Gelig start with an unsaturated vertex and an unmatched edge. It
follows that the roots of Sb can only be unsaturated singletons or unsaturated light blossoms. We
label the root vertices outer. If v is a nonroot vertex in the search forest let τ (v) be the edge in Sb
pointing to the parent of v. The inner/outer status of vertices is defined as follows:
Definition 11. A vertex v is outer if one of the following is satisfied:
1. v is the root of a search tree.
5
We observe that since augmenting along an eligible augmenting walk does not make any ineligible edges eligible,
any blossoms that is reachable after augmentation must also be reachable before it. Hence we do not have to further
contract any blossom after the augmentation.
5
In actual implementation we can reuse the inner/outer label from the previous Augmentation and Blossom
Formation step for dual adjustment. The reason is that augmenting along an eligible augmenting path will not create
any eligible edges, and will not make any unreachable vertex reachable. Moreover, the depth first nature of the search
tree guarantees that vertices on the augmenting path, if remains reachable, will still inherit its inner/outer label
before augmentation.
14
2. v is a singleton and τ (v) ∈ F .
3. v is a nontrivial blossom and τ (v) = η(v).
Otherwise, one the the following holds and v is classified as inner:
1. v is a singleton and τ (v) ∈ E\F .
2. v is a nontrivial blossom and τ (v) 6= η(v).
b call it Tb, can be grown by repeatedly attaching a child v to
An individual search tree in S,
b Let Bu denote the root blossom in Ω
its parent u using an edge (u, v) that is eligible for u in S.
containing u. We say an edge (u, v) ∈ E is eligible for u if it is eligible and one of the following is
satisfied:
1. u is an outer singleton and e 6∈ F .
2. Bu is an outer blossom and e 6= η(Bu ).
3. u is an inner singleton and e ∈ F .
4. Bu is an inner blossom and e = η(Bu ).
Hence, Sb consists of singletons and blossoms that are reachable from an unsaturated singleton
or light blossom, via an eligible alternating path whose preimage starts with an unsaturated vertex
and an unmatched edge. For simplicity, we call such blossoms and singletons reachable, and all
other singletons and blossoms unreachable. A vertex v from the original graph Gelig is reachable
[
(unreachable) if Bv is reachable (unreachable) in G
elig .
In our version of Edmonds’ Search, primal and dual variables are initialized in a way that Property 1(1) (Approximate Domination) is always satisfied, and Property 1 (Approximate Tightness) is
vacuous (as the f -factor is initially empty) but Property 1(4) (Unsaturated Vertices) is not. For this
reason, there is a large gap between primal and dual objective at the beginning of the algorithm,
which can be evaluated by the following:
yz(V ) − w(F ) =
X
f (v)y(v) +
v∈V
=
X
X f (B) + |I(B)|
2
B∈Ω
z(B) +
X
u(e) −
e∈E
X
w(e)
e∈F
def(v)y(v).
v∈V
The goal of the algorithm can be seen as bridging the gap between the primal objective and dual
objective while preserving all other complementary slackness properties. It can be achieve in two
ways. Augmentations enlarge the f -factor by augmenting F along some augmenting walk P . This
will reduce the total deficiency on the vertex set V . Dual Adjustments change the dual variables in
a way that decreases the y-value on unsaturated vertices while maintaining other complementary
slackness condition. In this algorithm, the progress of Edmonds’ Search is measured by the latter,
i.e., the overall reduction in y-values of unsaturated vertices.
The correctness of our algorithm reduces to showing that Augmentation and Blossom Formation,
Blossom Dissolution, and Dual Adjustment all preserve Invariant 1.
Lemma 12. Augmentation and Blossom Formation preserves Invariant 1.
15
v0
v1
v2
v18
v3
v4
v5
v6
v7
v19
v20
v8
v9
v10
v21
v22
v11
v12
v13
v23
v24
v14
v15
v16
v27
v25
v17
v26
Figure 5: An example of an eligible alternating search tree. Outer blossoms and singletons are
labeled using solid boundaries while inner blossoms and singletons have dashed boundaries.
16
Proof. We first show that the identity of I(B) is invariant under augmentation; in particular,
augmenting walks that intersect B do not affect I(B). As a result, the function yz(·) is invariant
under augmentation. We use I(B), η(B) and I 0 (B), η 0 (B) to denote the I-set and base edge of B
before and after the augmentation. By Definition 5 (augmenting walks), if P intersects B, then
δP (B) = η(B) ∪ η 0 (B) = η(B) ⊕ η 0 (B)
Let F and F 0 be the f -factor before and after augmentation. We have
δF 0 (B) = δF (B) ⊕ δP (B)
Combining both equations, we have
δF 0 (B) = δF (B) ⊕ (η(B) ⊕ η 0 (B))
Hence
I 0 (B) = δF 0 (B) ⊕ η 0 (B) = δF (B) ⊕ η(B) = I(B).
S
By Invariant 1, any blossom edge e ∈ B∈Ω EB satisfies both Approximate Domination as well
as Tightness, so it continues to satisfy these Invariants after augmentation. For any eligible edge not
in EB for any B ∈ Ω, by Criterion 1, if e ∈ F , yz(e) = w(e) − δ, thus after augmentation its duals
satisfy Approximate Domination. If e is unmatched, yz(e) = w(e), so its duals satisfy Tightness
after augmentation.
Augmentation preserves maturity of blossoms. For any vertex v in a nonterminal blossom B,
degF (v) = deg0F (v) = f (v), so maturity is naturally preserved. If B is a terminal blossom, we have
degF (v) = f (v)−1 for v = β(B) and degF (v) = f (v) for all v 6= β(B). Moreover, after augmentation
B always has a base edge η(B) = δP (B). Therefore, B is also mature after augmentation.
All the newly formed blossom in this step must be mature and have 0 z-values, so the value of
the yz function is unchanged and all the invariants are preserved.
Lemma 13. Blossom Dissolution both preserve Invariant 1.
Proof. Discarding blossoms of 0 z-values preserves the value of the yz function and Hence preserves
the invariants.
The crux of the proof is to show that Dual Adjustment also preserves Invariant 1, in particular
Approximate Domination and Tightness. Before proving the correctness of Dual Adjustment, we
first prove the following parity lemma, which was first used in [15]; we generalize it to f -factor:
Lemma 14 (Parity). Let Sb be the search forest defined as above. Let S be the preimage of Sb in G.
The y-value of every vertex in S has the same parity, as a multiple of δ/2.
Proof. The claim clearly holds after initialization as all vertices have the same y-values. Because
every eligible P
edge (u, v) ∈ Sb that straddles two singletons or blossoms must have yz(u, v) =
y(u) + y(v) + B:(u,v)∈γ(B)∪I(B) z(B) being a multiple of δ, and z-values are always multiples of δ,
y(u) and y(v) will always share the same parity, as multiple of δ/2. Therefore it suffices to show
that every vertex in a blossom B ∈ Ω has the same parity.
To prove this, we only need to show that Blossom Formation step only groups vertices with same
parity together. This is because new blossoms B are formed with when edges in CB are eligible
17
because of Criterion 1(2,3), this means their endpoints share the same parity. Hence by induction,
all vertices in B also share the same parity. The Dual Adjustment step also preserves this property
as vertices in a blossom will have the same inner/outer classification and thus have their y-values
all incremented or decremented by δ/2.
Lemma 15. Dual Adjustment preserves Invariant 1.
Proof. We focus on (2)(Approximate Domination) and (3)(Tightness). Other invariants are not
affected by Dual Adjustment.
There are much more cases to consider in f -factor compared to ordinary matching. Different
cases can be generated for an edge (u, v) by considering (1) the inner/outer classification of both
endpoints, (2) whether (u, v) is matched, (3) whether (u, v) is the base edge for its respective
endpoints, if they are in blossoms, and (5) whether (u, v) is eligible. In the following analysis, we
narrow down the number of meaningful cases to just 8.
We consider an edge e = (u, v). If u and v are both unreachable from any unsaturated vertex,
or both are in the same root blossom, yz(u, v) clearly remains unchanged after the dual adjustment.
Therefore we can assume Bu 6= Bv and at least one of them, say Bu , is reachable. Every reachable
endpoint will contribute a change of ±δ/2 to yz(u, v). This is the adjustment ofPy(u), plus the
adjustment of z(Bu ) if e ∈ I(Bu ). Let ∆e (u) be the net change of the quantity y(u)+ e∈I(Bu ) z(Bu ).
We omit e and use ∆(u) when the edge e we are considering is clear. By how we perform Dual
Adjustment, we have the following scenarios:
1. ∆(u) = +δ/2: This occurs if u is an inner singleton, or Bu is an outer blossom with e ∈ I(Bu ),
or an inner blossom with e 6∈ I(Bu ).
2. ∆(u) = −δ/2: This occurs if u is an outer singleton, or Bu is an inner blossom with e ∈ I(Bu ),
or an outer blossom with e 6∈ I(Bu ).
Then we consider the effect of a Dual Adjustment on edge e = (u, v). First we consider the case
b In this case only u will introduce a change on
when exactly one of Bu and Bv , say Bu , is in S.
yz(u, v):
Case 1: u is an inner singleton: Here ∆(u) = +δ/2. In this case Approximate Domination is
preserved, so we only need to worry about approximate tightness and hence assume e ∈ F . Since
b e cannot be eligible, or Bv would have been included in Sb as a child of Bu . Hence
Bv is not in S,
yz(e) < w(e). By Granularity, yz(e) ≤ w(e) − δ/2. Therefore we have yz(e) ≤ w(e) after the Dual
Adjustment.
Case 2: u is an outer singleton: Here ∆(u) = −δ/2. In this case Tightness is preserved and
we only need to worry about approximate dominination when e 6∈ F . Similar to Case 1, e must be
ineligible and yz(e) ≥ w(e) − δ/2. After Dual Adjustment we have yz(e) ≥ w(e) − δ.
Case 3: Bu is an inner blossom: We divide the cases according to whether e is matched or not.
Subcase 3.1: e ∈ F . If e 6∈ η(Bu ), then e ∈ I(Bu ) and ∆(u) = −δ/2. In this case Tightness
is preserved. If e ∈ η(Bu ), then e 6∈ I(Bu ) and ∆(u) = +δ/2. But e cannot be eligible since
otherwise Bv must be in the search tree, so we have yz(e) ≤ w(e) − δ/2 and yz(e) ≤ w(e) after
Dual Adjustment.
Subcase 3.2: e 6∈ F . This is basically symmetric to Subcase 3.1. If e ∈ η(Bu ), then e ∈ I(Bu )
and ∆(u) = −δ/2. But e cannot be eligible therefore yz(e) ≥ w(e) − δ/2, and yz(e) ≥ w(e) − δ after
18
Dual Adjustment. If e 6∈ η(Bu ), then e 6∈ I(Bu ) and ∆(u) = +δ/2, so approximate Domination is
preserved.
Case 4: Bu is an outer blossom:
Subcase 4.1: e ∈ F . If e ∈ η(Bu ), then Bv must be the parent of Bu in the search tree,
b Thus e 6∈ η(Bu ), so e ∈ I(Bu ) and ∆(u) = +δ/2. Since
contradicting the fact that Bv 6∈ S.
Bv is not reachable, e cannot be eligible, so yz(u, v) ≤ w(e) − δ/2 before Dual Adjustment and
yz(u, v) ≤ w(e) afterward.
Subcase 4.2: e 6∈ F . Similarly, e 6∈ η(Bu ), so e 6∈ I(Bu ) and ∆(u) = −δ/2. Similarly Bv is not
reachable so e cannot be eligible. Therefore we have yz(u, v) ≥ w(e) − δ/2 and yz(u, v) ≥ w(e) − δ
after Dual Adjustment.
This completes the case when exactly one of e’s endpoints is reachable. The following part will
complete the argument for when both endpoints are reachable. We argue that three scenarios can
happen: either ∆(u) and ∆(v) are of opposite signs and cancel each other out, or ∆(u) and ∆(v)
are of the same sign and the sign aligns with the property we wish to keep, or if both cases does
not hold, we use Lemma 14(Parity) to argue that there is enough room for Dual Adjustment not to
violate Approximate Domination or Tightness.
b In this case we assume Bu is the parent of Bv and e is the
We first examine tree edges in S.
parent edge of Bv . Hence e must be eligible for Bu . We argue by the sign of ∆(u).
Case 5: ∆(u) = +δ/2:
There are three cases here: u is an inner singleton, Bu is an outer blossom with e ∈ I(Bu ), or
Bu is an inner blossom with e 6∈ I(Bu ). We first observe that in all three cases, e ∈ F . This is
straightforward when u is an inner singleton. If Bu is an outer blossom with e ∈ I(Bu ), we know
that since Bu is outer, e 6∈ η(Bu ), so therefore e ∈ F . If Bu is an inner singleton with e 6∈ I(Bu ),
since Bu is inner, e ∈ η(Bu ), so combined with the fact that e 6∈ I(Bu ) we have e ∈ F .
Notice that since Bu is the parent of Bv , and e ∈ F , v can be an outer singleton, or Bv is
an outer blossom with e ∈ η(Bv ), or Bv is an inner blossom with e 6∈ η(Bv ). In the second case
e 6∈ I(Bv ) and in the third case e ∈ I(Bv ). In all three cases we have ∆(v) = −δ/2, and yz(e)
remains unchanged.
Case 6: ∆(u) = −δ/2: Case 6 is symmetric to Case 5. Bu can either be an outer singleton,
an inner blossom with e ∈ I(Bu ) or an outer blossom with e 6∈ I(Bu ). In all cases, the fact that e
must be eligible for Bu implies e 6∈ F , and Bv can only be an inner singleton, an outer blossom with
e ∈ I(Bu ) or an inner blossom with e 6∈ I(Bv ). Hence we have ∆(v) = +δ/2 so yz(e) still remains
constant.
Now suppose Bu and Bv are both in Sb but (u, v) is not a tree edge. We still break the cases
according to the sign of ∆(u) and ∆(v). Here we only need to consider when ∆(u) = ∆(v), since
otherwise they cancel each other and yz(e) remains constant.
Case 7: ∆(u) = ∆(v) = δ/2. In this case yz(e) is incremented by δ. Therefore we only need to
worry about Tightness when e ∈ F . Notice that Bu can only be an inner singleton, an outer blossom
with e ∈ I(Bu ) or an inner blossom with e 6∈ I(Bu ). When Bu is an outer blossom, e 6∈ η(Bu ).
When Bu is an inner blossom, since e ∈ F and e 6∈ I(Bu ), e ∈ η(Bu ). The same holds for the other
endpoint Bv .
It is easy to verify that in all cases, e is eligible for Bu (or Bv ) if and only if e is eligible.
But notice that after Augmentation and Blossom Formation steps, there is no augmenting walk or
19
[
reachable blossom in G
elig , i.e., there cannot be an edge (u, v) that is eligible for both endpoints
Bu and Bv since otherwise you can find an augmenting walk or a new reachable blossom. Thus
e is ineligible and yz(e) < w(e). But by Invariant 1 (1) (Granularity) and Lemma 14 (Parity),
both w(e) and yz(e) must be multiples of δ. Therefore we have yz(e) ≤ w(e) − δ. This implies
yz(e) ≤ w(e) after Dual Adjustment.
Case 8: ∆(u) = ∆(v) = −δ/2. Here yz(e) is decremented by δ. Similar to the case above, we
can assume e 6∈ F and only focus on Approximate Domination. Bu can be an outer singleton, inner
blossom with e ∈ I(Bu ), or outer blossom with e 6∈ I(Bu ). Since e 6∈ F , e ∈ I(Bu ) if and only if
e ∈ η(Bu ). Therefore if e is eligible, e must be eligible for both Bu and Bv . But similar to Case 7, e
being eligible for both endpoints will lead to the discovery of additional blossom or augmenting walk
in Gelig , which is impossible after Augmentation and Blossom Formation. Therefore we conclude in
this case e is ineligible and yz(e) > w(e) − δ. By Lemma 14(Parity), we have yz(e) ≥ w(e) before
Dual Adjustment, so Approximate Domination still holds after Dual Adjustment.
Theorem 16. A (1 − )-approximate f -factor can be computed in O(W m−1 ) time.
Proof. By setting δ = 1/ −1 ≤ , y-values of unsaturated vertices takes (W/2)/(δ/2) = O(W −1 )
iterations to reach 0 and thus satisfying Property 1 with δ1 = δ, δ2 = 0. By invoking Lemma 6, with
F ∗ being the optimum f -factor, we have
w(F ) ≥ w(F ∗ ) − |F ∗ |δ ≥ w(F ∗ ) − w(F ∗ )δ ≥ (1 − )w(F ∗ ).
For the running time, each iteration of Augmentation, Blossom Formation, Dual Adjustment
and Blossom Dissolution can be implemented in linear time. We defer the detail implementation to
Section 5. There are a total of W/δ = O(W −1 ) iterations, so the running time is O(W m−1 ).
As a result of Lemma 10 and Lemma 9, we also obtain the following result:
Corollary 17. A (1 + )-approximate f -edge cover can be computed in O(W m−1 ) time.
4.2
A Scaling Algorithm for General Weight
In this section, we can modify the O(W m−1 ) weighted f -factor algorithm to work on graphs with
general weights. The modification is based on the scaling framework in [5]. The idea is to divide
the algorithm into into L = log W + 1 scales that execute Edmonds’ search with exponentially
diminishing δ. The goal of each scale is to use O(−1 ) Edmonds’ searches to halve the y-value of
all unsaturated vertices while maintaining a more relaxed version of approximate complementary
slackness. By manipulating the weight function, Approximate Domination, which is weak at the
beginning, is strengthened over scales, while Approximate Tightness is weakened in exchange. Assume without loss of generality that W and are powers of two. We define δi , 0 ≤ i < L be the
error parameter for each scale, where δ0 = W and δi = δi−1 /2 for 0 < i < L. Each scale works with
a new weight function wi which is the old weight function rounded down to the nearest multiple of
δi , i.e, wi (e) = δi bw(e)/δi c. In the last scale WL = w. We maintain a scaled version of Invariant 1
at each scale:
Invariant 2 (Scaled approximate complementary slackness with positive unsaturated vertices). At
scale i = 0, 1, ... , L = log W , we maintain the f -factor F , blossoms Ω and duals y, z to satisfy the
following invariant:
20
1. Granularity. All y-values are multiple of δi /2, and z-values are multiple of δi .
2. Approximate Domination. For each e 6∈ F or e ∈ EB for any B ∈ Ω, yz(e) ≥ wi (e) − δi .
3. Approximate Tightness. For each e ∈ F or e ∈ EB for any B ∈ Ω, let j be the index of the
earliest scale that e became matched (e can be unmatched and rematched afterwards). We have
yz(e) ≤ wi (e) + 2δj − 2δi .
j
k
4. Mature Blossoms. For each blossom B ∈ Ω, |F ∩ (γ(B) ∪ I(B))| = f (B)+I(B)
.
2
5. Unsaturated Vertices’ Duals. The y-values of all free vertices are the same and will always be
smaller or equal to the y-values of other vertices.
Based on Invariant 2, Edmonds’ search will use the following Eligibility criterion:
Criterion 2. At scale i, an edge e ∈ E is eligible if one of the following holds
1. e ∈ EB for some B ∈ Ω.
2. e 6∈ F and yz(e) = wi (e) − δi .
3. e ∈ F and yz(e) − wi (e) is a nonnegative multiple of δi .
Before the start of scale 0, the algorithm initializes F, Ω, y, z similar to the algorithm for small
edge weights: y(u) ← W/2, Ω ← ∅, F ← ∅. At scale i, the duals of unsaturated vertices start at
W/2i+1 . We execute (W/2i+2 )/(δi /2) = O(−1 ) iterations of Edmonds’ search with parameter δi
and Criterion 2. The scale terminates when the y-values of unsaturated vertices are decreased to
W/2i+2 , or in the last iteration, terminates as they reach 0.
Notice that although the invariant and the eligibility criterion are changed, the fact that Edmonds’ search preserves the complementary slackness invariant, in particular that Dual Adjustment
preserves approximate domination and approximate tightness still holds here. By the proof of
Lemma 15, as long as the definition of eligibility guarantees the following parity property:
Lemma 18. At any point of scale i, let S be the set of vertices in Gelig reachable from an unsaturated
vertices using eligible edges. The y-value of any vertex v ∈ V with Bv ∈ S has the same parity as
multiple of δi /2.
Dual adjustment never changes the yz-values of edges inside any blossom B ∈ Ω, while it will
have the following effect on edge e if its endpoints are lying in different blossoms:
1. If e 6∈ F and is ineligible, yz(e) might decrease but will never exceed the threshold for eligibility,
i.e, it will not drop below wi (e) − δi .
2. If e 6∈ F and is eligible, yz(e) will never decrease.
3. If e ∈ F and is ineligible, yz(e) might increase but will never exceed the threshold for eligibility,
i.e, it will not raise above wi (e) + 2δj − 2δi .
4. If e ∈ F and is eligible, yz(e) will never increase.
In order words, Dual Adjustment will not destroy Approximate Domination and Approximate Tightness. Therefore Edmonds’ search within scale i will preserve Invariant 2.
We also need to manipulate the duals between different scales to ensure Invariant 2. Formally,
after completion of scale i, we increment all the y-values by δi+1 . This will ensure both approximate
21
domination and approximate tightness holds at scale i + 1: wi+1 (e) ≤ wi (e) + δi+1 , so increasing
y-values by δi+1 ensures Approximate Domination. For Approximate Tightness, we have yz(e) −
wi+1 (e) ≤ 2δj − 2δi + 2δi+1 = 2δj − 2δi+1 , since δi+1 = δi /2.
The algorithm terminates when the y-values of all unsaturated vertices reach 0. It terminates
with an f -factor F and its corresponding duals y, z and Ω satisfying the following property:
Property 3 (Final Complementary Slackness).
1. Approximate Domination. For all e 6∈ F or e ∈ EB for any B ∈ Ω, yz(e) ≥ w(e) − δL .
2. Approximate Tightness. For all e ∈ F or e ∈ EB for any B ∈ Ω, let j be the index of the
earliest scale that e becomes matched (e can be unmatched and rematched afterwards). We
have yz(e) ≤ w(e) + 2δj .
j
k
3. Blossoms Maturity. For all blossoms B ∈ Ω, |F ∩ (γ(B) ∪ I(B))| = f (V )+I(B)
).
2
4. Unsaturated Vertices Duals. The y-values of all free vertices are 0.
This implies domination and tightness are satisfied within some factor 1±O(). For Approximate
Domination this is easy to see since w(e) ≥ 1 and δL = , thus yz(e) ≥ (1 − )w(e) if e 6∈ F . For
Approximate Tightness, we can lower bound the weight of e if e first enters F at scale j. Throughout
scale j, the y-values are at least W/2j+2 , so w(e) ≥ wj (e) ≥ 2(W/2j+2 ) − δj . Since δj = W/2j ,
yz(e) ≤ w(e) + 2δj ≤ (1 + O())w(e) when e ∈ F .
With a similar proof to Lemma 6, we can show this characterizes an (1 − O())-approximate
minimum weight f -factor.
The running time of the algorithm is O(m−1 log W ) because there is a total of log W + 1 scales,
and each scale consists of O(−1 ) iterations of Edmonds’ search that can be implemented in linear
time.
Theorem 19. A (1 − )-approximate maximum weight f -factor can be computed in O(m−1 log W )
time.
Corollary 20. A (1+)-approximate minimum weight f -edge cover can be computed in O(m−1 log W )
time.
4.3
A Linear Time Algorithm
We also point out that by applying techniques in [5, §3.2], the algorithm can be modified to run
in time independent of W . The main idea is to force the algorithm to ignore an edge e for all
but O(log −1 ) scales. Let µi = W/2i be the maximum possible weight of an edge, matched or
unmatched, being eligible during scale i. Let scale(e) be the i such that w(e) ∈ [µi , µi−1 ). We
notice that we can ignore e in any scale j < scale(e). Moreover, we will also forcibly ignore e at
scale j > scale(e) + λ where λ = log −1 + O(1). Ignoring an otherwise eligible edge might cause
violation of approximate tightness and approximate domination. However, since the y-values of free
vertices are O(w(e)) at this point, this violation will only amount to O(w(e)).
To see this, notice that µi is also an upper bound to the amount of change to yz(e) caused
by all Dual Adjustment after scale i. Hence, after scale scale(e) + λ, the total amount of violation to approximate tightness and approximate domination on e can be bounded by µscale(e)+λ =
O()µscale(e) = O()w(e), which guarantees we still get an (1 − O()) approximate solution.
Therefore, every edge takes part in at most log −1 + O(1) scales, with O(−1 ) cost per scale.
The total running time is O(m−1 log −1 ).
22
Theorem 21. A (1−)-approximate maximum weight f -factor and a (1+)-approximate minimum
weight f -edge cover can be computed in O(m−1 log −1 ) time, independent of the weight function.
5
A Linear Time Augmenting Walk Algorithm
This section presents the linear time augmenting walk algorithm used in the Augmentation step of
Edmonds’ search. The goal here is to find a maximal set Ψ of edge disjoint augmenting walks in
the graph Gelig .
The main idea is to apply the modified depth first search of [15, §8], which finds a maximal set of
vertex disjoint augmenting walks for ordinary matching in linear time. There are several difficulties
in adapting algorithm in [15] for f -factor:
1. Instead of vertex disjoint, we are looking for a maximal set of edge disjoint augmenting walks,
which might lead to intersecting search trees.
2. The algorithm needs to be able to identify augmenting cycles, which are not present in the
standard matching problem.
3. The depth first search tree branches on both outer and inner singletons, as well as outer
blossoms.
b = G/Ω with f -factor F . At any moment, we maintain a
The input is a contracted graph G
D
E
c1 , T
c2 , ... , T
ck on a minor of
search forest, which is an ordered set of depth first search trees Sb = T
b where vertices on Tbi are
G . The minor is defined by a dynamic laminar set Ω0 of blossoms on G,
0
6
singletons or maximal blossoms of Ω ∪ Ω. For a vertex v in G, we use Bv to denote the largest
blossom in Ω ∪ Ω0 that contains v. We use Ti to refers to the set of vertices contained in blossoms
of Tbi .
Since search trees are rooted at unsaturated vertices, we can use Definition 11 to define inner/outer vertices in the search forest. This also gives us the set of edges that are eligible for a
vertex in the search forest.
c1 , T
c2 , ... , T
[
Similar to depth first search, T
k−1 are search trees from previously terminated searches.
b
Each Ti can either be a successful search tree, which contains an augmenting walk Pi , or an exhausted
search tree, which is a maximal search tree that does
S not lead to any augmenting walk, in which case
Pi = ∅. The final output of the algorithm is Ψ = i Pi , along with the newly discovered blossoms
in Ω0 , such that Ψ is a maximal set of augmenting walks in Ω0 ∪ Ω.
Invariant 3. For each vertex u in Ti \ Pi , i < k, there is no edge (u, v) that is eligible for u, but v
is in some later search tree Tj with j > i.7
In other words, each vertex u in Ti \ Pi , i < k, is exhausted, that is, every edge that is eligible
for u is scanned by the depth first search procedure. Notice that a vertex can go from being already
exhausted to unexhausted because its inner/outer classification changes and the change introduces
a new set of eligible edges. Therefore, we say a vertex v is outer(inner)-exhausted if it is currently
outer(inner) and all edges eligible for v are scanned.
As in Edmonds Algorithm, Ω0 is the set of blossoms we discovered during the search for augmenting walk
In standard matching, this corresponds to the property that there does not exist an unmatched edge connecting
two outer vertices from different search trees.
6
7
23
Walks in Ψ must be edge disjoint but are not necessarily vertex disjoint. Suppose we scan the
neighborhood of v in the order given by its adjacency list e1 , e2 , ... , edeg(v) . During the algorithm,
each edge in δ(v) is in one of the following three states.
• Explored, indicating the edge is explored from v when building some search tree Tbi , and is
either included in Tbi , or is not included in Tbi because it cannot be used to enlarge Tbi .
• Augmented, indicating the edge is explored from v in some Tbi and is part of Pi .
• Unexplored, indicating it has not been explored in any Tbi yet.
When the algorithm terminates all the unexplored edges in δ(v) will be declared unreachable,
i.e., after removing all the edges in Ψ, the edge cannot be reached via an alternating path starting
from any unsaturated vertex. This will be stated in Lemma 22.
The main complication of our algorithm is the alternating structure of the search tree, which
necessitates the maintenance of a maximal set of blossoms. Maintaining a maximal set of blossoms
allows us to completely decide for each vertex v, whether there is an even and odd length alternating
path from an unsaturated vertex to v. We can reduce it to maintaining the following Invariant:
Invariant 4. If (u, v) is an edge such that (u, v) is scanned from both u and v, u and v must be in
the same blossom.
ck , referred to as the active search tree, is the search
The last search tree in the search forest T
tree that the algorithm is currently growing. The algorithm grows the search tree in a depth first
ck consisting of all vertices that we have
manner by extending the active walk, Pactive , i.e., a walk in T
not finish exploring yet. The rest of the Invariant states the topology of the search tree as well as
its depth first property.
Invariant 5.
E
D
c1 , T
c2 , ... , T
ck consists of edge disjoint alternating trees in the graph
1. The sequence Sb = T
b = G/(Ω ∪ Ω0 ) that are rooted at unsaturated singletons/light blossoms.
G
2. Edges in Tbi are all explored. If an edge (u, v) is explored but not augmented, then all edges
(v, w) eligible for v must also be explored. If (u, v) is augmented, then its parent edge τ (Bu )
must also be augmented.
3. Vertices in Tbi are singletons or blossoms of Ω ∪ Ω0 . For each tree edge (u, v) = τ (v), (u, v) is
eligible for u.
ck descendant.
4. The active walk Pactive consists of a walk from root to a T
A search tree initially starts with a single unsaturated vertex as its root and the only vertex in
the active path. Suppose a blossom B is the last blossom on the active path. In each iteration, we
scan the next unexplored edge (u, v), u ∈ B that is eligible for B. Depending on where the other
endpoint v is located, we do one of the following:
1. Augmentations: We form an augmenting walk in two situations: (1) when (u, v) is unmatched,
Bv is not saturated and not yet in the active search tree, and (2) when (u, v) is unmatched,
Bv is the singleton root of the active search tree, and def(v) ≥ 2. In both cases we extend
the active walk with (u, v), terminate the search, and make the state of every Pactive edge
augmenting.
24
2. DFS extensions: If the exploration of (u, v) does not lead to an Augmentation step, and either
Bv is a nontrivial blossom not in any search tree, or Bv is a singleton, extend the active walk
ck as Bu ’s child. This corresponds to a grow step in Edmonds’ Search.
to v and put v into T
3. Blossom Formation: If Bv is a descendant of B, and edge (u, v) is also eligible for Bv , we form
ck path from B to Bv . Contract this
a blossom consisting of blossoms and singletons from the T
blossom, call it B 0 , and place the blossom in the active path. Label all these vertices now as
outer and ready to be scanned. Note that all previously inner vertices will go from exhausted
to not exhausted as an outer vertex now.
4. DFS retractions: If every edge (u, v) eligible for B is already scanned, mark B as exhausted
and remove it from the active walk. If B is the only vertex in the active walk, terminate the
search.
After the search terminates and returns with a search tree, we update the deficiency function
(in case we discover an augmenting walk/cycle, and some unsaturated vertices become saturated)
and start a new search at an unsaturated vertex that has not been an exhausted search tree root,
as long as it exists. Notice that the root of a successful search tree can be reused as long as it is
unsaturated after all augmentations so far. The algorithm terminates when all unsaturated vertices
are exhausted, where each vertex is either saturated, or is unsaturated but we cannot discover any
augmenting path starting from it.
We first show after the search terminates, all unexplored edges are not reachable from any
alternating path starting from an unsaturated vertex. This guarantees we did not ignore any edges
that might lead to a new augmenting path.
Lemma 22. When the search terminates, each unexplored edges (u, v) is unreachable from any
alternating path that does not contain any edge in Ψ.
Proof. Suppose an unexplored edge (u, v) is reachable from an alternating path consisting of explored but not augmented edges. Without loss of generality let (u, v) be the first reachable unexplored edge in δ(u) and let (w, u) be the edge that precedes (u, v) in the alternating path. Since
(w, u) is explored but not augmented, all unexplored edges must be explored from u, leading to a
contradiction.
Lemma 23. All operations preserve Invariants 3, 4, and 5.
Proof. Invariant 3 is a natural consequence of depth first search. Invariant 5 (1) is immediate from
how we maintain blossoms and the active walk. Invariant 5 (2) holds because if (u, v) is explored
but, for each (v, w) eligible for v, the search along (v, w) finds no augmenting walk, then every edge
eligible for v must also be explored before the search retracts from u.
To show Invariant 4, notice by the depth first nature of the search tree, Bu and Bv must be
ancestor of one another (otherwise, if Bu is first explored while Bv is not in the search tree, DFS
extension will put Bv into the search tree as Bu child). Suppose Bu is the ancestor of Bv , when we
scan the edge (u, v) from Bu , if (u, v) is not eligible for v and no blossom step can be performed,
the edge (u, v) must be made eligible for Bv in a later blossom step from Bu or an ancestor of Bu ,
to Bv or a descendant of Bv . This puts u and v into the same blossom.
With these invariants, we can show the following lemma:
25
Lemma 24. Any augmenting walk P 0 w.r.t. F ⊕
P ∈ Ψ at an edge.
S
P ∈Ψ P
must intersect some augmenting walk
Proof. Suppose we have an augmenting walk P 0 = hu0 , u1 , ... u2k−2 , u2k−1 , u2k i that does not intersect Ψ at any edge. By Lemma 22, we can assume all edges (ui , ui+1 ) in P 0 must be explored or
ineligible for ui . By termination of the algorithm, u0 must be an exhausted unsaturated vertex so
must be a search tree root. And by definition of augmenting walk (u1 , u2 ) must be eligible for u0
so must be explored. We show by induction that for 0 ≤ i ≤ 2k − 1, (ui , ui+1 ) must be explored
b
and ui+1 must be in the search forest S.
For i = 0, by definition u0 is a search tree root and (u0 , u1 ) is explored from u0 and is eligible for
u0 . This means u1 must also be in the search tree. If ui is in the search tree and (ui , ui+1 ) is explored
from ui . Consider vertex (ui+1 ), since it is in the augmenting walk, by the alternation requirement
for augmenting walk one of (ui , ui+1 ) and (ui+1 , ui+2 ) must be eligible for ui+1 , if (ui+1 , ui+2 ) is
eligible for ui+1 , (ui+1 , ui+2 ) must be explored and we are done. Otherwise, (ui , ui+1 ) is eligible
for both of its endpoints and thus (ui , ui+1 ) must be in the same blossom. In this case, Blossom
Formation step makes (ui+1 , ui+2 ) eligible for ui+1 .
Since all ui ’s is in the search tree and the preceding edge is eligible for its predecessor, ui ’s,
in particular u2k can not be a unsaturated inner singleton or an unsaturated light inner blossom
because otherwise an Augmentation step will be performed. This contradicts the fact that P 0 is an
augmenting walk.
Theorem 25. A maximal set of augmenting walks can be found in linear time.
Proof. The algorithm runs in linear time because each edge is explored at most twice, i.e. once
in each direction. Moreover, blossom structures here can also be maintained in linear time using
incremental-tree disjoint set data structure in [13].
We say a set of nested blossoms Ω0 is maximal if after contracting blossoms in Ω0 , no more contractible blossom that is reachable from an unsaturated vertex can be found. Since all augmenting
walk has been eliminated after Augmentation, this is equivalent to saying that no edge (u, v) is
eligible for both Bu and Bv . Thus, to implement the blossom step for Edmonds’ search, it suffices
to run the linear time find-aw once again (which will not find any augmenting walk) and return
the blossom set Ω0 discovered at the Blossom Formation step. By Invariant 5 (4), Ω0 is maximal.
Corollary 26. A maximal set of nested blossom can be found in linear time.
Corollary 27. One iteration of Edmonds’ Search with Criterion 1 (Approximate Complementary
Slackness) can be implemented in linear time.
6
Algorithms for unweighted f -factor and f -edge cover
p
In this section we will give an O( f (V )m) algorithm for both maximum cardinality f -factor and
minimum cardinality f -edge cover. This is a direct consequence of the O(W m−1 ) algorithm for the
weighted problem. This algorithm matches the running time of [10] but does not rely on reduction
to Micali-Vazirani Algorithm. Moreover, it is much simpler to state and to analyze.
For illustration purposes we focus on maximum cardinality f -factor. The algorithm consists of
two phases. The first phase, referred to as batch augmentation, finds an f -factor F that is close to
26
optimal using an instance of the O(W m−1 ) algorithm. After F is close to optimum, we discard
all dual variables y and z, dissolve all the blossoms in Ω and uses our linear time augmenting walk
algorithm to increase the cardinality of F until F becomes optimum.
This is stated formally in the following theorem:
p
Theorem 28. A maximum cardinality f -factor can be computed in O( f (V )m) time.
Proof. We can view an maximum cardinality fp
-factor problem as an maximum weight problem
with weight function w(e) = 1. Choose = 1/ f (V ), by p
Theorem 16, we can compute a (1 −
√ 1 )-approximate maximum cardinality matching F in O( f (V )m) time. If F ∗ is the maximum
f (V )
cardinality f -factor, we have
p
1
f (V )
)|F ∗ | > |F ∗ | − p
> |F ∗ | − f (V )/2
f (V )
f (V ) 2
p
This means F is only O( f (V )) augmentations away from optimal. Hence we can then discard
blossom structure Ω with duals y and z from the approximate f -factor and run the linear time
augmenting
pwalk algorithm of Lemma 25 in G with respect to F until F is optimal. By
p the discussion
above, O( f (V )) iterations suffice. The total running time of the algorithm is O( f (V )m).
p
Theorem 29. A minimum cardinality f -edge cover can be computed in O( f (V )m) time.
|F | ≥ (1 − p
1
Proof. This is similar to Theorem 28. We first use the O(W m−1 ) algorithm for f -edge cover given
p
−1
in Section 3 to find an (1 + f (V ) )-approximate minimum cardinality f -edge
p cover F by viewing
the graph as a weighted graph with weight 1 everywhere. Choosing = 1/ f (V ) will give us an
p
−1
f -edge cover F with |F | ≤ |F ∗ | + f (V ) |F ∗ |. Notice that we always have |F ∗ | ≤ f (V ) because
taking f (v) arbitrary incident edges for each v and taking their union will
p always give a trivial
∗| +
f -edge cover
with
cardinality
at
most
f
(V
).
Hence
we
have
|F
|
≤
|F
f (V ), which means at
p
most O( f (V )) reductions are needed to make F optimal. Therefore we can run the augmenting
path algorithm from Lemma 25 to find reducing paths (P is a reducing path w.r.t. F if and
p only
if it is an augmenting path w.r.t. E \ F ) until no reducing path can be
p found. There are f (V )
iterations in this phase. The total running time of the algorithm is O( f (V )m).
27
References
[1] K. Choromanski, T. Jebara, and K. Tang. Adaptive anonymity via b-matching. In Proceedings
of the 26th International Conference on Neural Information Processing Systems, pages 3192–
3200, 2013.
[2] K. M. Choromanski, A. Khan, A. Pothen, and T. Jebara. Large scale robust adaptive anonymity
via b-edge cover and parallel computation. 2015.
[3] V. Chvatal. A greedy heuristic for the set-covering problem. Mathematics of Operations Research, 4(3):233–235, 1979.
[4] D. E. Drake and S. Hougardy. A simple approximation algorithm for the weighted matching
problem. Information Processing Letters, 85(4):211–213, 2003.
[5] R. Duan and S. Pettie. Linear time approximation for maximum weight matching. J. ACM,
61(1), 2014.
[6] R. Duan, S. Pettie, and H.-H. Su. Scaling algorithms for weighted matching in general graphs.
In Proceedings 28th Annual ACM-SIAM Symposium on Discrete Algorithms (SODA), pages
781–800, 2017.
[7] J. Edmonds. Maximum matching and a polyhedron with 0, 1 vertices. J. of Res. the Nat.
Bureau of Standards, 69 B:125–130, 1965.
[8] J. Edmonds. Paths, trees and flowers. Canadian Journal of Mathematics, pages 449–467, 1965.
[9] J. Edmonds and E. Johnson. Matching: A Well-Solved Class of Integer Linear Programs. In
Combinatorial Optimization - Eureka, You Shrink!, pages 27–30. 2003.
[10] H. N. Gabow. An efficient reduction technique for degree-constrained subgraph and bidirected
network flow problems. In Proceedings of the Fifteenth Annual ACM Symposium on Theory of
Computing, pages 448–456, 1983.
[11] H. N. Gabow. Data structures for weighted matching and extensions to b-matching and f factors. CoRR, abs/1611.07541, 2016.
[12] H. N. Gabow and P. Sankowski. Algebraic algorithms for b-matching, shortest undirected
paths, and f-factors. In 2013 IEEE 54th Annual Symposium on Foundations of Computer
Science, pages 137–146, 2013.
[13] H. N. Gabow and R. E. Tarjan. A linear-time algorithm for a special case of disjoint set union.
In Proceedings of the Fifteenth Annual ACM Symposium on Theory of Computing, STOC ’83,
pages 246–251, 1983.
[14] H. N. Gabow and R. E. Tarjan. Faster scaling algorithms for network problems. SIAM Journal
on Computing, 18(5):1013–1036, 1989.
[15] H. N. Gabow and R. E. Tarjan. Faster scaling algorithms for general graph matching problems.
J. ACM, 38(4):815–853, 1991.
28
[16] A. Khan and A. Pothen. A new 3/2-approximation algorithm for the b-edge cover problem.
In 2016 Proceedings of the Seventh SIAM Workshop on Combinatorial Scientific Computing,
pages 52–61.
[17] E. Lawler. Combinatorial Optimization: Networks and Matroids. Dover Books on Mathematics
Series. 2001.
p
[18] S. Micali and V. V. Vazirani. An O( |V |E) algoithm for finding maximum matching in general
graphs. In 21st Annual Symposium on Foundations of Computer Science, pages 17–27, 1980.
[19] S. Pettie and P. Sanders. A simpler linear time 2/3 − approximation for maximum weight
matching. Inf. Process. Lett., 91(6):271–276, Sept. 2004.
[20] W. Pulleyblank. Faces of matching polyhedra. PhD thesis, University of Waterloo, Ontario,
Canada, 1973.
[21] A. Schrijver. Combinatorial Optimization: Polyhedra and Efficiency. Algorithms and Combinatorics. Springer, 2002.
[22] W. T. Tutte. On the problem of decomposing a graph into n connected factors. Journal of the
London Mathematical Society, s1-36(1):221–230, 1961.
29
| 8cs.DS
|
ROBUST TORIC IDEALS
arXiv:1304.0603v2 [math.AC] 19 Jun 2013
ADAM BOOCHER AND ELINA ROBEVA
Abstract. We call an ideal in a polynomial ring robust if it can be minimally generated
by a universal Gröbner basis. In this paper we show that robust toric ideals generated
by quadrics are essentially determinantal. We then discuss two possible generalizations to
higher degree, providing a tight classification for determinantal ideals, and a counterexample
to a natural extension for Lawrence ideals. We close with a discussion of robustness of higher
Betti numbers.
1. Introduction
Let S = k[x1 , . . . , xn ] be a polynomial ring over a field k. We call an ideal robust if it can
be minimally generated by a universal Gröbner basis, that is, a collection of polynomials
which form a Gröbner basis with respect to all possible monomial term orders. Robustness
is a very strong condition. For instance, if I is robust then the number of minimal generators
of each initial ideal is the same:
(1.1)
µ(I) = µ(in< I) for all term orders <.
In general, we can only expect an inequality (≤).
For trivial reasons, all monomial and principal ideals are robust. Simple considerations
show that robustness is preserved upon taking coordinate projections and joins (see Section
2). However, nontrivial examples of robust ideals are rare. A difficult result of [BZ93, SZ93]
(recently extended by [Boo12] and [CDNG13]) shows that the ideal of maximal minors of
a generic matrix of indeterminates is robust. In the toric case, the class of Lawrence ideals
(which naturally includes determinantal ideals) is the largest known class of robust ideals.
This paper discusses robustness for toric ideals. Our main result is the following:
Theorem 1.2. Let F be a set of irreducible binomials that minimally generate an ideal.
(Assume that F cannot be partitioned into disjoint sets of polynomials in distinct variables.)
If F consists of polynomials of degree 2 then the following are equivalent:
• The ideal generated by F is robust
• |F | = 1 or F consists of the 2 × 2 minors of a generic 2 × n matrix
x1 · · · xn
y1 · · · yn
up to a rescaling of the variables.
In higher degree, it is not true that the robustness of (F ) implies that it is determinantal (or
even Lawrence).
1
2
ADAM BOOCHER AND ELINA ROBEVA
Our motivation for studying robust toric ideals is threefold. First, the study of ideals
minimally generated by a Gröbner basis (for some term order) is ubiquitous in the literature.
In [CHT06], Conca et al studied certain classical ideals and determined when they are
minimally generated by some Gröbner basis. In the study of Koszul algebras, one of the
most fruitful approaches has been via G-quadratic ideals - those generated by a quadratic
Gröbner basis. We are not aware, however, of any systematic study of ideals minimally
generated by a universal Gröbner basis; robust ideals.
Second, we focus on toric ideals, because they are a natural testing ground for algebraic
questions. Toric ideals possess binomial universal Gröbner bases, and hence and there is a
rich theory of Gröbner basis theory, see e.g. [Stu96].
Finally, passing from an ideal to an initial ideal is a particular type of flat deformation.
In this phrasing, robustness is almost equivalent to the property that the minimal number
of generators is preserved by these deformations.1 This interpretation suggests there might
be a geometric interpretation of robustness in terms of the Hilbert scheme.
The paper is organized as follows: In section 2, we prove our main result, Theorem 2.2
characterizing robust toric ideals generated in degree two. The methods are mainly combinatorial. In Sections 3 and 4 we pose two questions concerning extensions of Theorem 2.2
using Lawrence ideals. We provide negative and positive answers respectively. Section 5
closes with a discussion of “robustness of higher Betti numbers,” our original motivation for
this project.
2. Quadratic Robust Toric Ideals are Determinantal
In the sequel, by toric ideal we will always mean a prime ideal generated minimally by
homogeneous binomials with nonzero coefficients in k. By the support of a polynomial we
mean the set of variables appearing in its terms.
Definition 2.1. A set F of polynomials in S is called robust if F is a universal Gröbner
basis and the elements of F minimally generate their ideal.
If a set of polynomials F can be written as a union F = G ∪ H of polynomials in disjoint
sets of variables, then we say that G is a robust component of F . If F admits no such
decomposition, then we say the set F is irreducibly robust. Notice that robustness is preserved
under these disjoint unions, so to classify robust ideals, it suffices to study the irreducible
ones. We remark that the ideal of F corresponds to the join of the varieties corresponding
to the G and H.
The goal of this section is to prove the following theorem:
Theorem 2.2. Let F be an irreducibly robust set consisting of irreducible quadratic binomials. Then F is robust if and only if |F | = 1 or F consists of the 2 × 2 minors of a generic
2 × n matrix
x1 · · · xn
y1 · · · yn
up to a rescaling of the variables.
1Note
y, y − z).
that the equality 1.1 is in general, weaker than robustness. For example, consider the ideal (x −
ROBUST TORIC IDEALS
3
Remark 2.3. In the statement of the Theorem, we only assume that the generators are
irreducible. It turns out that this is sufficient to show that the ideal they generate is prime.
Notice that one direction follows immediately from the results of [SZ93] which show that
the 2 × 2 minors are a universal Gröbner basis. To prove the converse, our technique is
essentially to eliminate certain combinations of monomials from appearing in F . To simplify
notation, we will omit writing coefficients in the proofs when it is clear that they do not
affect the argument. In particular, we treat the issue of coefficients only in tackling the proof
of Theorem 2.2 itself and not in earlier lemmas.
Lemma 2.4. Let F be a robust set of prime quadratic binomials. Then no monomial appears
as a term in two different polynomials in F .
Proof. Suppose that the monomial m appears in the polynomials f, g ∈ F . Let < be a Lex
term order taking the support of m to be first. Since f and g are prime, < will select m
as the lead term of both f and g, and applying Buchberger’s algorithm, we would obtain a
degree zero syzygy of the elements in F , contradicting the minimality of F .
Proposition 2.5. If F is robust, and 0 ≤ k ≤ n, then so is F ∩ k[x1 , . . . , xk ].
Proof. Write Fk = F ∩ k[x1 , . . . , xk ]. It is clear that Fk minimally generates the ideal (Fk ).
Let < be any term order on k[x1 , . . . , xk ]. Extend < to a term order <S on S, taking
x1 , . . . , xk last. Then since F is a Gröbner basis with respect to <S , by basic properties of
Gröbner bases, we know that Fk will be a Gröbner basis with respect to <.
The above proposition is extremely useful, because in our analysis it will be helpful to
assume we are working in a ring with few variables. We will use this reduction extensively
in the following main technical lemma. We use the letters a, . . . , z when convenient for ease
of reading.
Lemma 2.6. Let F be a robust set of prime quadratic binomials:
(a) F cannot contain two polynomials of the form
f = x2 + yz, g = xy + m
or
f = x2 + y 2 , g = xy + m
or
f = x2 + y 2 , g = xz + m.
where m is any monomial.
(b) F cannot contain two polynomials of the form
f = xi x j + xk x l , g = xi xk + xp xq
(here we do not assume i, j, k, l, p, q are distinct.)
(c) F cannot contain two polynomials of the form
f = x2 + yz, g = xw + m
4
ADAM BOOCHER AND ELINA ROBEVA
or
f = x2 + yz, g = yw + m
where m is any monomial.
(d) If f, g ∈ F are two polynomials whose supports share a variable, then all terms of f and
g are squarefree.
(e) If F contains two polynomials whose supports share a variable, then (up to coefficients)
F must contain the 2 × 2 minors of a generic matrix.
a b c
d e f
Proof. a) We prove the first statement. The proofs of the others are similar. Suppose that
f, g ∈ F . Notice that by primality, m cannot contain a factor of y. Let < be the lex term
order with (y > x > z > all other variables). Then the S-pair of f and g is x3 − mz, whose
lead term is x3 . Since F is a Gröbner basis with respect to < we must have some polynomial
h whose lead term divides x3 . Since x2 is not the lead term of f , h 6= f and we must have
two distinct polynomials in F with x2 appearing. This contradicts Lemma 2.4.
b) Suppose that f, g ∈ F . By restricting to the subring k[xi , xj , xk , xp , xq ], Proposition 2.5
tells us we can assume F involves only these variables. Notice by primality (and part (a))
we know that k and i are distinct from j, l, p, q. Let < be the lex term order with (xk > xi >
all other variables.) Then the S-pair of f and g is x2i xj − xl xp xq , whose lead term is x2i xj .
As in part a) we must have some polynomial h ∈ F whose lead term divides x2i xj . The only
possible monomials are x2i and xi xj . And since xi xj appears in f (and is not a lead term)
we must have a polynomial h = x2i + xa xb ∈ F for some a, b ∈ {i, j, k, l, p, q}. So F contains
f = xi xj + xk xl , g = xi xk + xp xq , h = x2i + xa xb .
Applying part a), and primality, we know that a, b ∈ {l, p, q}. By part a), we know that
xa xb must be squarefree, and since xp xq already appears, we can say (renaming p and q if
necessary,) that xa xb = xl xp . But now choosing < to be the lex term order with (xl > xp >
xi > xk > all other variables,) we see that the S-pair of f and h is x2i xk − xi xj xk whose lead
term is x2i xk which is only divisible by the monomials x2i and xi xk , neither of which can be
a lead term of a polynomial in F by Lemma 2.4.
c) We will prove the first statement. The second proof is similar. Suppose that f, g ∈ F .
First restrict, using Proposition 2.5 to assume we are working only with the variables x, y, z, w
and the factors of m. Let < be the lex term order with (w > x > all other variables. Taking
the S-pair of f and g, we obtain wyz − mx, whose lead term is wyz. Since this must be
divisible by the lead term of some polynomial h ∈ F , without loss of generality, we assume
h = wy + n. Consider now the possibilities for n. By primeness n cannot contain a factor
of w or y. By part b) it cannot contain a factor of x or z. Hence, the only possible options
left are that the factors of n are contained in the factors of m. But this means that m, n are
ab, a2 for some (new, distinct) variables a, b. As in (b), we can conclude this is impossible.
d) This follows immediately from parts a) - c).
e) We assume that F contains two polynomials whose supports intersect. By d) we can
assume that these polynomials are squarefree, and we write them as p = ae − bd, q =
af − m1 m2 where mi represents some variable. Notice that by primality and part b), neither
ROBUST TORIC IDEALS
5
m1 nor m2 can be a, e or f . Nor can m1 m2 = bd (since it would be a repetition). Hence we
may as well assume m1 is different from the other variables, and call it c. There are now two
cases: Either m2 is also a new variable g, or it isn’t, in which case we can see that without
loss of generality, m2 = d. Rewriting: If p, q are two polynomials whose supports intersect,
then they must contain either 6 or 7 distinct variables.
In the case of 6 variables, restrict F to the subring k[a, b, c, d, e, f ]. Now
p = ae − bd, q = af − cd.
Computing an S-pair with the lex order (a > b > d > all other variables) we obtain f bd−ecd
with lead term f bd. This must be divisible by the lead term of some polynomial r ∈ F . But
this lead term cannot be bd (by its presence in p), hence it must be either bf or df . In case
it is bf then F contains a polynomial of the form r = bf − n1 n2 . Now n1 , n2 ∈ {a, c, d, e}
by primality. And part b) of this lemma allows us to further say n1 , n2 ∈ {c, e} which along
with squarefreeness implies that r = bf − ce as required. In case the term is df , similar
considerations show that parts a) - d) will not allow any n1 n2 .
In the case of 7 variables, restrict F to the subring k[a, b, c, d, e, f, g]. Now
p = ae − bd, q = af − cg.
Computing an S-pair with the lex term order (a > b > d > all other variables): we obtain
f bd − ebg with lead term f bd. This must be divisible by the lead term of some polynomial
r ∈ F . But this lead term cannot be bd (by its presence in p), hence it must be either bf
or df . By symmetry we can assume that it is df and that F contains a polynomial of the
form df − n1 n2 . Now n1 , n2 ∈ {a, b, c, e, g} by primality, and part b) restricts us further
to n1 , n2 ∈ {c, e, g}. And since cg already appears, we can conclude that n1 n2 = eg or ec
(and again by symmetry, we may assume n1 n2 = ec. But now notice that q and r are two
polynomials whose supports intersect, and involve only 6 variables. Hence, by the previous
part of this proof, we can conclude that F contains a polynomial s = gd − ae. But this is a
contradiction by Lemma 2.4.
Proof of Theorem 2.2. Suppose that |F | > 1. Since F is irreducible, it must contain two
polynomials whose supports intersect. By Lemma 2.6 we can conclude that F contains
polynomials of the form:
p1 = ae − bd, p2 = af − cd, p3 = bf − ce
up to coefficients. However, computing an S-pair with the lex order on (a > b > · · · > f )
we obtain:
S(p1 , p2 ) = bdf − cde (with some nonzero coefficients)
which after reducing by p3 we obtain either zero, or a constant multiple of cde. In the latter
case, in order to continue the algorithm, we would have to have another polynomial in F
whose lead term divided cde. By the presence of p1 , p2 , p3 , the terms cd and ce are prohibited.
And by Lemma 2.6 b), de is also prohibited. Hence, this S-pair must reduce to zero after
only two subtractions.
This means in fact, that the polynomials are precisely determinants of some matrix
λ1 a λ2 b λ3 c
µ1 d µ 2 e µ 3 f
6
ADAM BOOCHER AND ELINA ROBEVA
for some nonzero constants λi , µi .
To complete the proof, suppose that F 6= {p1 , p2 , p3 }. Since F is irreducible, one polynomial p4 ∈ F must share a variable with say, p1 . Renaming variables if necessary, say that
variable is a. Then (ignoring constants for the moment) by applying the proof of Lemma 2.6
e), to the polynomials p1 = ae − bd and p4 = ah − m1 m2 we can conclude that F contains
the minors of the matrix
λ1 a λ 2 b λ 4 g
.
µ1 d µ 2 e µ 4 h
Applying this technique to p2 and p4 as well, shows that we in fact get all 2 × 2 minors of
the full matrix
λ1 a λ2 b λ3 c λ4 g
.
µ1 d µ 2 e µ 3 f µ 4 h
Inductively we continue this process until we obtain all of F .
Notice that in our proof, every term order we used was a Lex term order, and we ended
up a with a prime ideal. Hence, we have the following:
Corollary 2.7. If F is a set of prime quadratic binomials that minimally generate an ideal.
Then the following are equivalent:
(i) F is a Gröbner basis with respect to every Lex term order.
(ii) F is a Gröbner basis with respect to every term order.
(iii) F generates a prime ideal and the irreducible robust components of F are generic
determinantal ideals and hypersurfaces.
Corollary 2.8. If X is a generic k × n matrix, and F is the set of 2 × 2 minors, then F is
a universal Gröbner basis if and only if k = 2.
Remark 2.9. It is almost the case that every irreducibly robust component is determinantal.
Indeed, every prime binomial is up to rescaling either xy − zw or x2 − yz. The former is
determinantal. Thus the only possible non-determinantal robust component is {x2 − yz}.
3. From Determinants to Lawrence Ideals
Encouraged by the result of the previous section, it is natural to ask to what extent
robustness classifies generic determinantal ideals. Indeed, it is easy to see that the ideal of
minors of any 2 × n matrix whose entries are relatively prime monomials will be robust.
There are two questions we consider:
Question 3.1.
1. If I is a robust toric ideal, is I generated by the 2×2 minors of some matrix of monomials?
2. Precisely which matrices of monomials provide robust ideals of 2 × 2 minors?
The answer to the first question is negative. Examples are provided by Lawrence ideals,
studied in [Stu96]. If I is any toric ideal, with corresponding variety X ⊂ Pn−1 , then the
ideal J corresponding to the re-embedding of X in (P1 )n−1 is called the Lawrence lifting of
I. Its ideal is generated by polynomials of the following form:
JL = xa yb − xb ya | a − b ∈ L ⊂ S = k[x1 , . . . , xn , y1 , . . . , yn ],
ROBUST TORIC IDEALS
7
where L is a sublattice of Zn and k is a field. Here a = xa11 xa22 · · · xann for a = (a1 , . . . , an ) ∈ Nn .
Binomial ideals of the form JL are called Lawrence ideals. The following result is Theorem
7.1 in [Stu96].
Proposition 3.2. The following sets of binomials in a Lawrence ideal JL coincide:
a) Any minimal set of binomial generators of JL .
b) Any reduced Gröbner basis for JL .
c) The universal Gröbner basis for JL .
d) The Graver basis for JL .
Hence Lawrence ideals provide a large source of robust toric ideals, and naturally include
the class of generic determinantal ideals. Given this, it is natural to rephrase the first part
of Question 3.1 as:
Question 3.3. Does robustness characterize Lawrence ideals?
Again the answer is negative.
Example 3.4. The ideal
I = (b2 e − a2 f, bc2 − adf, ac2 − bde, c4 − d2 ef)
in the polynomial ring Q[a, b, c, d, e, f ] is robust but not Lawrence. This example was found
using the software Macaulay2 and Gfan [Gra, Jen]. It is the toric ideal IL corresponding to
the lattice defined by the kernel of
1 1 1 1 1 1
1 1 0 0 0 0
L=
0 0 1 2 0 0
1 0 1 0 3 1
Given this counterexample we ask
Question 3.5. Is there a nice combinatorial description of robust toric ideals?
Remark 3.6. Heuristically, it is very easy to find robust ideals that are not Lawrence by
starting with a Lawrence ideal given by JL . This lattice L gives rise to a lattice L̃ ⊂ N2n such
that JL = IL̃ . By modifying L̃ slightly, it is very often the case that the resulting toric ideal
is robust (though often non-homogeneous). The ubiquity of these examples computationally
suggests that a nice combinatorial description of robustness may require imposing further
hypotheses.
4. Matrices of Monomials
In this section we answer Question 3.1.2.
Theorem 4.1. Suppose that Xi , Yj are monomials of degree at least 1 in some given set of
variables U = {u1 , u2 , . . . , ud }. Let
X1 X2 · · · Xn
A=
,
Y1 Y2 · · · Yn
8
ADAM BOOCHER AND ELINA ROBEVA
where n > 3 and suppose that the set F of 2 × 2-minors Xi Yj − Xj Yi , i 6= j consists of
irreducible binomials. Then F is robust if and only if all the monomials Xi , Yj are relatively
prime.
The proof is technical, so we begin by fixing notation. Since we will assume that each
Xi Yj − Xj Yi is prime for all i 6= j, then, gcd(Xi , Xj ) = gcd(Yi , Yj ) = gcd(Xi , Yi ) = 1 for all
i 6= j. Therefore, if we define
zij = gcd(Xi , Yj ),
then, we can write
X i = xi
Y
zij
and
j6=i
Yj = yj
Y
zij .
i6=j
Thus,
(i)
gcd(xi , xj ) = gcd(yi , yj ) = 1 for all i 6= j
(ii)
gcd(zij , zkl ) = 1 whenever i 6= k or j 6= l,
(iii)
gcd(xi , zkl ) = 1 if i 6= k and gcd(yj , zkl ) = 1 if j 6= l.
Our goal is to show that zij = 1 for all i 6= j.
Lemma 4.2. If z12 6= 1 then n > 4 and for each m > 3, there exist permutations im , lm 6=
1, 2, m and jm , km 6= 1, 2 and term orders >1 and >2 such that
X1
and Xim Yjm >1 Xjm Yim
(iv)
Xim Yjm | X2 Ym2
z12
2
Y1
Xkm Ylm | Xm
(iv’)
Y2
and Xkm Ylm >2 Xlm Ykm .
z12
Moreover,
Xim = xim zim m and xim |zim m ,
Ylm = ylm zmlm and ylm |zmlm .
Proof. We will build <1 and <2 in several steps. To begin, take a lex term order > where
the variables in z12 are first. Consider the S-pair:
S X1 Ym − Xm Y1 , Xm Y2 − X2 Ym =
=
lcm(X1 Ym , Xm Y2 )
lcm(X1 Ym , Xm Y2 )
(X1 Ym − Xm Y1 ) −
(Xm Y2 − X2 Ym ) =
X 1 Ym
Xm Y2
= Xm
Y2
X1
(X1 Ym − Xm Y1 ) − Ym (Xm Y2 − X2 Ym ) =
z12
z12
X1
Y2
2
− Xm
Y1 .
z12
z12
Y2
2 X1
2
Since all of the variables in X2 Ym z12 − Xm Y1 z12 are different from the variables in z12 , then,
there exist term orders >1 and >2 refining > for which X2 Ym2 zX121 is the leading term for >1
= X2 Ym2
ROBUST TORIC IDEALS
9
2
and Xm
Y1 zY122 is the leading term for >2 .
2
Consider first >1 : X2 Ym2 zX121 >1 Xm
Y1 zY122 . Since the Xi Yj − Xj Yi form a Gröbner basis with
respect to >1 , there exist im 6= jm such that
(iv)
Xim Yjm | X2 Ym2
X1
and Xim Yjm >1 Xjm Yim
z12
in this ordering. If jm = 2, then, z12 | Y2 | X2 Ym2 zX121 , which is not true. If jm = 1, then, since
z12 | X1 | X1 Yim and z12 - Xim Y1 and z12 was chosen to have its variables first in >1 , then,
Xjm Yim = X1 Yim >1 Xim Y1 = Xim Yjm , which is not true by (iv). Thus, jm > 3. Similarly,
we can deduce that im > 3. Thus, im , jm > 3.
Since im > 3, gcd(X1 , Xim ) = gcd(X2 , Xim ) = 1, and we are assuming that Xi , Yi 6= 1
for all i and Xim | X2 Ym2 zX121 then, Xim | Ym2 . Since gcd(Xm , Ym ) = 1, im 6= m either.
Thus, we have that im 6= 1, 2, m and , jm > 3. So, in particular, ifQn = 3, we already
have a contradiction. We assume now that n > 4. Since Xim = xim j6=im zim j and Ym =
Q
ym k6=m zkm , properties (i) and (iii) allow us to conclude that Xim = xim zim m and xim | zim m .
Going back to when we chose the ordering >1 , consider now the ordering >2 for which
2
Xm
Y1 zY122 >2 X2 Ym2 zX121 . By a symmetric argument we find that there exist km > 3, lm 6= 1, 2, m
2
such that Xkm Ylm | Xm
Y1 zY122 and, thus, Ylm = ylm zmlm with ylm | zmlm .
Hence, there exist im , lm 6= 1, 2, m and jm , km 6= 1, 2 such that Xim = xim zim m , xim | zim m
and Ylm = ylm zmlm , ylm | zmlm .
Lemma 4.3. If z12 6= 1, then, n is even and we can rearrange the numbers 1, .., n so that
for each i 6 n2 ,
X2i = x2i z2i,2i+1 and Y2i = y2i z2i+1,2i
X2i+1 = x2i+1 z2i+1,2i and Y2i+1 = y2i+1 z2i,2i+1
and x2i , y2i+1 | z2i,2i+1 and x2i+1 , y2i | z2i+1,2i .
Proof. By property (ii) and Lemma 4.2 we have that m 7→ im and m 7→ lm are permutations
on {3, .., n} with no fixed points. Thus, for each i > 3, there exists m > 3 such that m 6= i
and i = im , Xi = xi zim with xi | zim and for each l > 3, there exists m > 3, m 6= l such that
l = lm and Yl = yl zml .
Fix m0 6= 1, 2. Then, Xim0 = xim0 zim0 m0 and xim0 | zim0 m0 . Since we assumed that Xim0 6= 1,
then, zim0 m0 6= 1. So now, repeating the whole argument with m0 and im0 instead of with 1
and 2 (recall that im0 6= 1, 2, m0 ), we would get similar permutations on {1, . . . , n}\{m0 , im0 }.
But, by property (ii), these permutations have to agree with the permutations m 7→ im and
m 7→ lm from above on the set {1, . . . , n} \ {1, 2, m0 , im0 }. Thus, (1, 2) and (m0 , im0 ) will be
transpositions in all of these permutations (and, in particular, im0 = lm0 ).
Since we can run the above argument with any m0 6= 1, 2, we have that the permutations
m 7→ im and m 7→ lm agree and are composed of transpositions (m, im ). In particular, n
10
ADAM BOOCHER AND ELINA ROBEVA
is even and, after rearranging the numbers from {1, .., n} so that i2k = 2k + 1 and, thus,
i2k+1 = 2k for all k = 1, .., n/2, our matrix A looks as follows:
x1 z12 x2 z21 x3 z34 x4 z43 · · ·
A=
.
y1 z21 y2 z12 y3 z43 y4 z34 · · ·
The rest of the statement of the Lemma follows from Lemma 4.2.
Proof of Theorem: (⇒): It suffices to show that zij = 1 for all i, j. Without loss of generality,
assume that z12 6= 1.
By Lemma 4.2 we have that
X1
Y2
2
and Xkm Ylm | Xm
Y1
Xim Yjm | X2 Ym2
z12
z12
for all m > 3. But by (the proof of) Lemma 4.3 we know that the above hold when we
substitute 1 and 2 with any m0 and im0 such that m 6= im0 , m0 , i.e.
Yi 0
Xm 0
2
Xim Yjm | Xim0 Ym2
and Xkm Ylm | Xm
Ym0 m
zm0 im0
zm0 im0
Rewriting out the expressions in the form Xim = xim zim m and Ylm = ylm zmlm and then
canceling repeating factors, we get that
2
zim m xm0
xim yjm zljm jm | xim0 zim0 m0 ym
and xkm zkm lkm ylm | x2m zmlm ym0 zim0 m0 yim0
for all m0 6= m, im . Therefore, we have that for every m0 6= m, im
2
zim m xm0
zljm jm | xim0 zim0 m0 ym
and zkm lkm | x2m zmlm ym0 zim0 m0 yim0
Noting that xim0 | zim0 m0 and xm0 | zm0 lm0 and, similarly for ym0 , yim0 , switching m0 with im0 ,
and using (ii), shows us that
2
zim m
zljm jm | ym
and zkm lkm | x2m zmlm
Again, by property (ii), the only way for this to happen is if jm = km = m. In that case, we
have that
2
xim ym zlm m | xim0 zim0 m0 ym
zim m xm0
and xm zmlm ylm | x2m zmlm ym0 zim0 m0 yim0
Again, by (i),(ii), and (iii), and by switching m0 and im0 , we have that
2
xim ym zim m | ym
zim m
and xm zmlm ylm | x2m zmlm
After cancelations,
xim | ym and ylm | xm .
Thus, xim = ylm = 1. Since i and l are permutations, we have that xm = ym for all m. Thus,
our matrix looks like this
z12 z21 z34 z43 · · ·
A=
.
z21 z12 z43 z34 · · ·
2
2
But then, we have that, for example, X1 Y2 − X2 Y1 = z12
− z21
, which is not prime! Contradiction! Thus, zij = 1 for all i 6= j.
Thus, z12 = 1 and by symmetry, zij = 1 for all i 6= j and gcd(Xi , Yj ) = 1 for all
i 6= j. Combined with the assumptions of the theorem statement, we get that gcd(Xi , Xj ) =
ROBUST TORIC IDEALS
11
gcd(Xi , Yi ) = gcd(Yi , Yj ) = 1 for all i 6= j, that is, all the entries of A are pairwise relatively
prime.
(⇐): Assume that the entries of A are pairwise relatively prime. Let < be any monomial
term order. To show that the 2×2 minors of A are a Gröbner basis with respect to <, we just
need to show that all S-pairs reduce to zero. By the result of Bernstin-Sturmfels-Zelevinsky,
such a reduction is guaranteed to exist for a generic matrix X = (xij ) . Since all entries of
A are relatively prime, it is clear that such a reduction will extend simply by the ring map:
xij 7→ Xij .
5. Robustness of Higher Betti Numbers
Our interest in robust ideals originated with the following classical inequality:
(5.1)
βi (S/ in< I) ≤ βi (S/I) for all i.
It is natural to ask for which ideals and term orders equality holds (for all i). In the setting
of determinantal ideals, Conca et al proved in [CHT06] that the ideal of maximal minors
of a generic matrix has some initial ideal with this property. They also gave examples
of determinantal ideals for which no initial ideal has this property. In a different vein,
Conca, Herzog and Hibi showed in [CHH04] that if the generic initial ideal Gin(I) has
βi (I) = βi (Gin(I)) for some i > 0, then βk (I) and βk (Gin(I)) also agree for k > i.
Our interest was to instead approach the inequality 5.3 in a universal setting, i.e. to
consider when equality holds for all term orders <. In this case we say that I has robust
Betti numbers. The following result is due to the first author [Boo12]
Theorem 5.2. If I := Ik (X) is the ideal of maximal minors of a generic k × n matrix X
and < is any term order, then
(5.3)
βij (S/ in< I) = βij (S/I)
for all i, j.
In particular, every initial ideal is a Cohen-Macaulay, squarefree monomial ideal with a linear
free resolution. Further, the resolution can be obtained from the Eagon-Northcott complex by
taking appropriate lead terms of each syzygy.
A combination of Theorems 5.2 and 2.2 yields
Corollary 5.4. Let I be a toric ideal generated in degree two. If I is robust, then I has
robust Betti numbers.
Our original hope with this project was that all robust toric ideals had robust Betti
numbers. Unfortunately, the situation seems much more delicate.
Example 5.5. Using Gfan [Jen], we were able to check that the Lawrence ideal JL corresponding to the lattice L given by the matrix
1 1 1 1 1
0 1 2 7 8
has initial ideals with different Betti numbers.
12
ADAM BOOCHER AND ELINA ROBEVA
6. Acknowledgments
Many of the results in this paper were discovered via computations using Macaulay 2 [Gra]
and Gfan [Jen]. We thank David Eisenbud and Bernd Sturmfels for many useful discussions.
Finally we thank Seth Sullivant for introducing us to Lawrence ideals and starting us on the
path toward Example 3.4.
References
[Boo12] A. Boocher, Free resolutions and sparse determinantal ideals, Math. Res. Letters 19 (2012), no. 4,
805–821. ↑1, 11
[BZ93] D. Bernstein and A. Zelevinsky, Combinatorics of maximal minors, J. Algebraic Combin. 2
(1993), no. 2, 111–121. MR1229427 (94j:52021) ↑1
[CDNG13] A. Conca, E. De Negri, and E. Gorla, Universal groebner bases for maximal minors, ArXiv (2013).
↑1
[CHH04] A. Conca, J. Herzog, and T. Hibi, Rigid resolutions and big Betti numbers, Comment. Math.
Helv. 79 (2004), no. 4, 826–839. MR2099124 (2005k:13025) ↑11
[CHT06] A. Conca, S. Hoşten, and R. R. Thomas, Nice initial complexes of some classical ideals, Algebraic
and geometric combinatorics, 2006, pp. 11–42. MR2298753 (2008h:13035) ↑2, 11
[Gra] D. R. Grayson, Macaulay 2, a software system for research in algebraic geometry. Available at
http://www.math.uiuc.edu/Macaulay2/. ↑7, 12
[Jen] A. N. Jensen, Gfan, a software system for Gröbner fans and tropical varieties. ↑7, 11, 12
[Stu96] B. Sturmfels, Gröbner bases and convex polytopes, University Lecture Series, vol. 8, American
Mathematical Society, Providence, RI, 1996. MR1363949 (97b:13034) ↑2, 6, 7
[SZ93] B. Sturmfels and A. Zelevinsky, Maximal minors and their leading terms, Adv. Math. 98 (1993),
no. 1, 65–112. MR1212627 (94h:52020) ↑1, 3
| 0math.AC
|
Segmentation of nearly isotropic overlapped tracks in
photomicrographs using successive erosions as
watershed markers
arXiv:1706.03282v2 [cs.CV] 18 Jan 2018
Alexandre Fioravante de Siqueiraa,b,∗, Wagner Massayuki Nakasugaa , Sandro
Guedesa , Lothar Ratschbacherb
a Departamento
de Raios Cósmicos e Cronologia, IFGW, University of Campinas
b Institut für Geologie, TU Bergakademie Freiberg
Abstract
The major challenges of automatic track counting are distinguishing tracks and
material defects, identifying small tracks and defects of similar size, and detecting overlapping tracks. Here we address the latter issue using WUSEM,
an algorithm which combines the watershed transform, morphological erosions
and labeling to separate regions in photomicrographs. WUSEM shows reliable
results when used in photomicrographs presenting almost isotropic objects. We
tested this method in two datasets of diallyl phthalate (DAP) photomicrographs
and compared the results when counting manually and using the classic watershed. The mean automatic/manual efficiency ratio when using WUSEM in the
test datasets is 0.97 ± 0.11.
Keywords: Automatic counting, Diallyl phthalate, Digital image processing,
Fission track dating
∗ Corresponding
author. Phone: +55(19)3521-5362.
Email addresses: [email protected] (Alexandre Fioravante de Siqueira),
[email protected] (Wagner Massayuki Nakasuga), [email protected] (Sandro
Guedes), [email protected] (Lothar Ratschbacher)
Authorship statement: A. F. de S. designed the algorithms, wrote the code, analyzed the
data and wrote the paper. W. M. S. counted the tracks in the images, analyzed the data and
wrote the paper. S. G. obtained the images from the samples, analyzed the data, wrote the
paper and supervised the research project. L. R. supervised the research project.
Preprint submitted to Computers & Geosciences
January 19, 2018
1. Introduction
Solid state nuclear track detectors (SSNTD) are materials such as inorganic
crystals, plastics and glasses, known to record the path of charged particles.
There are several applications for SSNTD in nuclear science; for instance, measurements of radon gas [1], boron neutron capture therapy [2], and age determination by fission track dating [3, 4, 5, 6, 7, 8].
When a charged particle collide in a SSNTD, the ionization and/or the
collision with atoms modify the path by which the particle go through. This
damage in the SSNTD structure is called latent track. This track becomes
visible under an optical microscope after a convenient etching process [9], and
these tracks can be counted. Photomicrographs can also be used, which could
improve the counting accuracy [10].
Procedures for measuring and counting tracks are time-consuming and involve practical problems, e.g. variation in observer efficiency [11]. An automatic
method based on image processing techniques could increase the track counting rate and improve counting reproducibility. However, separating elements in
nontrivial images is one of the hardest tasks in image processing [12].
Automatic systems for separating, counting or measuring tracks have been
studied for a while, and several solutions were presented (e.g. [13, 14, 15, 16,
17, 18, 19, 20, 21, 22, 23, 24, 25, 26, 27, 11, 28, 29]. Still, the precision of
automatic methods is not satisfactory yet; an automatic analysis could need to
be manually adjusted by the operator, being more time consuming than the
usual measure [26, 30]. The major challenges to automatic track counting are
detecting overlapping tracks, distinguishing tracks and material defects (e.g.
surface scratches due to polishing), and identifying small tracks and defects of
comparable size in the background of photomicrographs [11].
To address the problem of identifying overlapping ion tracks in photomicrographs, we propose an algorithm based on the watershed transform [31] using
morphological erosions [32] as markers. A similar method was used to separate packings of ellipsoidal particles represented using X-Ray tomography [33].
2
We tested this method in two datasets of diallyl phthalate (DAP) photomicrographs, and used the results to relate the incident track energy with the mean
gray levels and mean diameter products for each sample from the first dataset.
2. Material and methods
We employed the ISODATA threshold to obtain the binary images used in
our tests. The WUSEM algorithm is based on the following techniques: 1.
morphological erosion; 2. watershed transform, and 3. labeling. We describe
these algorithms in this Section.
2.1. The ISODATA threshold
The Iterative Self-Organizing Data Analysis Technique (A) (ISODATA) threshold [34, 35] is an histogram-based method. It returns a threshold that separates
the image into two pixel classes, where the threshold intensity is halfway between
their mean intensities.
When applied for two classes, ISODATA is always convergent [36]; in our
case, these classes are the tracks (our regions of interest, ROI) and the background. We used the algorithm filters.threshold isodata, implemented in
scikit-image [37].
2.2. Structuring elements and morphological erosion
In morphological image processing, a structuring element is a matrix representing a mask, or a shape, that is used to retrieve information about the shapes
in an input image [32]. Erosion, a basic operation in image processing, uses a
chosen structuring element to shrink the border of all ROI in a binary image;
the shrinkage factor is correspondent to the structuring element size [12].
In our tests (Section 3) we used disks as structuring elements, since tracks in
DAP are mostly round in shape. The algorithms used were morphology.disk
and morphology.erosion, contained in scikit-image.
3
2.3. Watershed transform
The watershed algorithm is a non-parametric method which defines the contours as the watershed of the gradient modulus of the gray levels of the input
image, considered as a relief surface detection method [31].
In this algorithm, the input image is seen in a three-dimensional perspective.
Two dimensions correspond to spatial coordinates, and the third represents the
gray levels. In this interpretation, we consider three kinds of points [12]:
1. Points in regional minima.
2. Points where a drop of water would flow to a common minimum.
3. Points where a drop of water would flow to different minima. The set of
these points is named watershed line.
The aim of watershed algorithms is to find the watershed lines. The method
used in this paper is implemented in the function morphology.watershed, from
scikit-image.
2.4. Labeling algorithm
In image processing, the labeling algorithm labels connected regions of a
binary input image, according to the 2-connectivity sense: all eight pixels surrounding the reference pixel. Pixels receive the same label when they are connected and have the same value. We used the algorithm measure.label from
scikit-image, implemented as described in [38].
2.5. Watershed Using Successive Erosions as Markers (WUSEM) algorithm
Here we present the WUSEM (Watershed Using Successive Erosions as
Markers) algorithm, which combines morphological erosions, the watershed transform and labeling algorithms to separate regions of interest (ROI) in binary
images. The WUSEM algorithm and its steps follow (Figure 1):
1. The user define an initial radius (r0 ) and an iterative radius (∆r) to a
structuring element (Figure 2).
4
Figure 1: Representing WUSEM and the functions employed in its implementation.
2. The input binary image is eroded using the structuring element with radius
equal to r0 . This erosion is used as a marker to the watershed algorithm.
Then, the resulting image is labeled. This is the first segmentation.
3. A new structuring element is defined. Its radius is r0 + ∆r. The input
binary image is eroded with this new structuring element, the erosion is
used as a marker to watershed, and the result is labeled. This is the second
segmentation.
4. The process continues until the eroded image does not have objects. Then,
all segmentations are summed.
5. The result is labeled again, to reorder the ROI, and labels with area smaller
than 64 pixels are excluded, to ensure that noise will not affect the results.
Tracks are counted according to the “lower right corner” method, where
objects that touch the bottom and right edges are not counted. This leads
to an accurate, unbiased measure [39].
WUSEM is implemented in the function segmentation wusem(), available
in the Supplementary Material. It receives the arguments str el, initial radius
and delta radius, representing the structuring elements, r0 and ∆r, respectively. To exemplify WUSEM’s capabilities, we used it to separate overlapping
tracks within the test photomicrographs (Figure 3).
2.6. DAP photomicrographs
We used two sets of diallyl phtalate (DAP, C14 H14 O4 ) photomicrographs to
test the WUSEM algorithm. One of them was obtained from detectors irradiated with
78
Kr tracks, and the other has induced fission tracks.
5
Figure 2: Representation of disks as structuring elements defined for each segmentation when
employing WUSEM in an interest region. The radius of the first, second and third structuring
elements are r0 , r0 + ∆r and r0 + 2∆r, respectively.
2.6.1.
78
Kr tracks
The first dataset contains 362 photomicrographs of tracks from nine different
DAP plaques irradiated with
78
Kr ions at a nominal fluence of 1.4 × 105 cm−2 ,
from a beam perpendicular to the detector surfaces at GSI, Darmstadt, Germany2 .
During the irradiation, detectors were covered with aluminum foils forming
a moderation layers of thicknesses varying from zero (no cover) to 90 µm. Detectors were named after their cover thicknesses (K0, K20, K30, ..., K90). Ions
arrived at the setup with initial energy of 865 MeV and were slowed down in
the aluminum cover before hitting the detector surface. Incidence energies were
2 These
photomicrographs are contained in the folder orig figures/dataset 01, available
in the Supplementary Material.
6
0
100
200
300
400
500
0
100
200
300
400
500
600
700
Figure 3: WUSEM algorithm application in an input image. First, the image is binarized
using the ISODATA threshold. Then, the image is eroded using initial radius (r0 ) and iterative
radius (∆r) equal to 1 and 1 respectively, to ease visualization. The process continues until
the eroded image has regions in it. All erosions are summed, and the result is labeled; then,
regions with area smaller than 64 px are excluded. Each labeled track receives a color from the
nipy spectral colormap. Finally, the function enumerate objects() is used to number the
found tracks. Final results are shown to r0 and ∆r equal to 10 and 4, respectively. Animation
also available at https://youtu.be/CtIOxhNISW8.
calculated using the software SRIM [40], and varied from 18 MeV (90 µm cover)
to 865 MeV (no cover). Detectors were etched in a PEW solution (7.5 g KOH,
32.5 g ethanol, 10 g water) solution for 4.5 ± 0.2 min at 65 ± 3 ◦ C.
Images were captured with a CCD camera coupled with a Zeiss microscope,
in reflected light mode, under 1250 × nominal magnification. Then, the detectors were further etched for 4 minutes, total of 8.5 ± 0.3 min, and new images
were captured. This way, we obtained 18 subsets of images. Tracks within
these photomicrographs are almost isotropic and have the same orientation,
resembling circles (Figure 4).
7
Figure 4: Photomicrograph from the first dataset, presenting tracks in a DAP sample.
2.6.2. Induced fission tracks
The second dataset contains 19 photomicrographs with two different magnifications from a DAP plaque used as external detector, coupled to an apatite
sample and irradiated with thermal neutrons to induce fission in the 235 U atoms
inside the mineral3 . During the fission process, two fragments are released and,
eventually, are detected by the DAP plaque. Fragments arrive at the detector surface with different energies and incidence angles, resulting in a variety
of track formats. Hence, counting tracks in these photomicrographs is more
complex than in the previous case (Figure 5).
2.7. License and reusability
The WUSEM algorithm and several functions for its implementation lie
within the packages Numpy [41], Scipy [42], Matplotlib [43], scikit-image [37],
among others. All code published with this paper is written in Python 3 [44]
3 These
photomicrographs are contained in the folder orig figures/dataset 02, available
in the Supplementary Material.
8
Figure 5: Input test photomicrograph from the second dataset, presenting tracks in a DAP
sample.
and is available under the GNU GPL v3.0 license, and all photomicrographs and
figures distributed with this paper are available under the CC-BY 2.0 license.
3. Experimental
3.1. Processing photomicrographs of
78
Kr tracks
3.1.1. Exemplifying the methodology
Here we use a photomicrograph from the first dataset4 to exemplify WUSEM
(Figure 4). We binarized this photomicrograph using the ISODATA threshold
[34, 35] (Figure 6). Different gray levels in some tracks may not be properly separated, complicating the extraction of track features. To address this issue, we
filled the regions in the binary image using the function ndimage.morphology.binary fill holes()
from scipy.
4 Image
K90 incid4,5min 3.bmp from the folder orig figures/dataset 01/Kr-78 4,5min/K90 incid,
available in the Supplementary Material.
9
0
100
200
300
400
500
0
100
200
300
400
500
600
700
Figure 6: Input photomicrograph (Figure 4) binarized using the ISODATA threshold (threshold = 128) and region filling. Colormap: gray.
After binarizing the input image, the WUSEM algorithm separates the overlapping tracks (Figure 7). For this example, we chose initial radius = 10
and delta radius = 4 as parameters.
Tracks were counted using the “lower right corner” method, i.e., border
tracks are only counted if they lie on the right or bottom edges of the image.
We used the function clear rd border(), given in the Supplementary Material,
to remove the tracks in these edges. WUSEM returns a labeled region, which
can be used as parameter to the function enumerate objects() (Figure 8).
3.1.2. Comparison between manual and automatic counting
An experienced observer can easily distinguish tracks in the photomicrograph, even when several tracks are superimposed. For this reason, manual
counting is considered the control in the comparison with the automatic count-
10
ing. In the following, the WUSEM algorithm is applied to the photomicrograph
set and the processing parameters are studied.
We established an arbitrary value of up to two tracks less than the mean
of the manual counting as a tolerance. WUSEM’s parameters become candidates if the automatic counting lies within the tolerance interval, i.e. 0 <
µnmanual –nauto < 2, where n is the track number obtained by each approach
(Figure 9).
WUSEM’s best parameters for this study are the ones within the tolerance
interval for most samples. According to the stated comparison, the best parameters are initial radius = 10, delta radius = 20 for 4.5 min samples, and
initial radius = 10, delta radius = 14 for 8.5 min samples.
Using the best parameters defined, we compared manual, classic (flooding)
watershed and WUSEM counting for each sample (Figure 10, Table 1). The
classic watershed algorithm used is implemented in ImageJ [45], and the full
method consisted in binarizing the input image, applying the watershed and
removing small objects. The source code for these operations and instructions
on how to use it are given in the Supplementary Material.
Since tracks in this dataset have the same shapes, we can attribute an efficiency of 100 % for manual counting. WUSEM counting was initially set to
obtain a smaller number of tracks when compared to the manual counting.
However, WUSEM counting returns false positives, i.e., incorrectly labels background regions as ROI (points above the 1:1 line in Figure 10 (b) and (d)). To
avoid false positives, one could use a more restrictive criterion, such as eccentricity (Section 3.1.3).
Counting reproducibility is important for considering the reliability of track
counting.
Despite the variations in track characteristics, the efficiencies of
WUSEM’s automatic counting remained constant within uncertainties, when
compared to manual counting.
11
Manual counting (µ ± 1σ)
WUSEM counting (µ ± 1σ)
Sample
4.5 min
8.5 min
4.5 min
8.5 min
4.5 min
Efficiency ± 1σ
8.5 min
K0
22 ± 3
20 ± 3
21 ± 2
19 ± 3
0.99 ± 0.12
0.95 ± 0.11
K20
25 ± 5
22 ± 4
23 ± 3
22 ± 3
0.90 ± 0.11
0.99 ± 0.14
K30
23 ± 4
24 ± 4
22 ± 4
22 ± 3
0.96 ± 0.12
0.94 ± 0.10
K40
22 ± 4
22 ± 4
22 ± 3
21 ± 3
1.01 ± 0.10
0.96 ± 0.08
K50
24 ± 2
22 ± 4
23 ± 3
20 ± 3
0.94 ± 0.11
0.95 ± 0.13
K60
25 ± 3
23 ± 4
25 ± 3
22 ± 3
0.99 ± 0.08
0.97 ± 0.13
K70
21 ± 4
17 ± 4
21 ± 4
17 ± 4
1.00 ± 0.07
1.00 ± 0.13
K80
21 ± 4
20 ± 5
20 ± 4
20 ± 4
0.96 ± 0.13
1.02 ± 0.14
K90
24 ± 4
21 ± 4
23 ± 4
21 ± 4
0.99 ± 0.11
0.94 ± 0.09
Table 1: Mean, standard deviation, and automatic/manual efficiency ratio for each sample of
the first dataset.
3.1.3. Relating ion energies with diameter product and mean gray levels
In this application, we use WUSEM to relate the track energies to the product of major and minor diameters (D> and D< , respectively) and the mean
gray level of each track. We considered only approximately circular tracks in
our analysis, based on an eccentricity ()5 criterion: should be equal to or less
than 0.3 (Figure 11). This additional criterion can be used also for counting
tracks, to ensure that false positives (spurious objects counted as tracks) are
avoided.
After separating each track, we can obtain its features such as gray levels
and diameters. Once the mean gray level and diameters of each track in a
photomicrograph are obtained, we can relate them with the incident energy for
each sample. The mean gray level of each sample is obtained getting the mean
of all gray levels of the tracks in the images of that sample. We adopted a
similar process to obtain the mean diameters. Then, the results are related to
the incident energy (Figure 12).
The diameter products roughly reflect the electronic energy loss (dE/dx)
curve for
5 In
78
Kr in DAP (Figure 12 (a) and (c)), calculated with the software
image processing, the eccentricity of an object is a number in the interval [0, 1). The
lower the value, the region becomes closer to a circle.
12
SRIM [40]. The Bragg peak appears around 100 MeV. Further scatter of points
can be attributed to poor control of etching conditions. The uncertainty of 3 ◦ C
in temperature may cause a large variation in etching results [46]. Variations in
gray level means are impaired probably because this set was acquired in reflected
illumination mode, which privileges surface details over depth effects.
3.2. Processing photomicrographs of fission tracks in DAP
In photomicrographs from the first dataset, our main concern was track superposition. However, all tracks were similar, created by a collimated beam of
78
Kr tracks. Here we go one step ahead, applying WUSEM to images where
tracks present a variety of shapes. This image dataset was obtained from DAP
plaques irradiated with thermal neutrons, coupled with apatite mounts. Fission fragments are born in the interior of the mineral, then emitted at different
directions towards the detector. For instance, round tracks were created by perpendicular incident fragments, while the elliptical ones were created by particles
hitting DAP surface at shallower angles (Figure 5).
As in the previous analysis, the manual counting performed by an experienced observer is taken as reference because we expect the observer to be able to
recognize tracks efficiently. However, in this case, we do not expect the observer
efficiency to 100 %. Fragments hitting the detector at lower energies originate
tracks that are very difficult to distinguish from detector surfaces. An experienced observer would avoid counting those tracks to keep hers/his counting
efficiency constant.
Repeating the previous processes for photomicrographs in the second dataset,
we first binarize a test photomicrograph6 (Figure 5) using the ISODATA threshold. The binarized image is generated for two scenarios: considering and ignoring border tracks. Here, regions in the binary image are also filled using the function ndimage.morphology.binary fill holes() from scipy. Then we apply
6 Image
“FT-Lab 19.07.390.MAG1.jpg”, from the folder orig figures/dataset 02. Avail-
able in the Supplementary Material.
13
the WUSEM algorithm. We chose initial radius = 5 and delta radius = 2
as parameters, and the WUSEM result as parameter in the function enumerate objects()
(Figure 13).
For this dataset, we established an arbitrary tolerance of five tracks less
than the mean of the manual counting. In this case, WUSEM’s parameters
become candidates if 0 < µnmanual –nauto < 5, where n is the track number obtained by each approach. According to the stated comparison, the best parameters are initial radius = 5, delta radius = 12 for the first magnification
and initial radius = 10, delta radius = 14 for the second one. Using the
best parameters defined, we compared manual, classic watershed and WUSEM
counting for each sample (Figure 14, Table 2).
Magnification
Manual counting (µ ± 1σ)
WUSEM counting (µ ± 1σ)
Efficiency ± 1σ
1
85 ± 8
83 ± 6
0.98 ± 0.07
2
37 ± 3
37 ± 5
0.99 ± 0.05
Table 2: Mean, standard deviation, and automatic/manual efficiency ratio for each magnification of the second dataset.
WUSEM succeeded in avoiding false positives in this application, when compared to the classic watershed (black line above the 1:1 line in Figure 14 (b)).
Still, the user could apply more restrictive criteria. Also, it is worth noting the
efficiency variation between the two image sets. Bigger objects are easier to be
treated, thus automatic track counting in greater magnification images resulted
in a higher counting efficiency (Table 2).
4. Discussion
4.1. Structuring elements
In this study, we used disks as structuring elements for processing images
in both datasets. Since tracks in photomicrographs from dataset 1 are almost
isotropic (as seen in Figure 4), disks are suitable structuring elements to be
used in their segmentation. However, tracks within images in dataset 2 do not
have a defined format (Figure 5). Employing different structuring elements, e.g.
14
rotated cones or ellipses, could improve their segmentation and the automatic
counting result.
4.2. False positives, false negatives and counting efficiency
In track counting, reproducibility is not about counting every track in the
image, but counting the same types of tracks every time. It is the primary
concern in FTD: for instance, the efficiencies for counting tracks in the standard
sample should not vary when using zeta age calibrations [47, 48]. For absolute
methods of determining neutron fluence [49], the efficiency should be constant
when counting tracks in unknown age samples. The fast-growing areas of FTD
using the Laser Ablation Inductively Couple Plasma Mass Spectrometer (LAICP-MS, [50, 51, 52, 53]) or the Electron Microprobe ([54, 55]) have the same
efficiency issues, and could also benefit from automatic counting.
The major challenge for reproducibility is avoiding false positives, spurious
objects such as scratches on the detector or mineral surface and other etching
figures which automatic counting algorithms could misrepresent as tracks. In
most situations, it is preferable to restrict the criteria, thus increasing the number of false negatives (not counted tracks), even implying in efficiency reduction.
Even more experienced observers expect some decreasing in efficiency due
to superposition when counting tracks in high track density samples. This loss
tends to be more severe in automatic counting. When applying algorithms as
WUSEM, the separation of tracks in objects formed of several tracks is not
always possible.
Counting less tracks than the actual number in a cluster is acceptable; however, when processing a large number of clusters in higher track density samples,
we expect a lower efficiency when compared with lower track density samples.
This effect can be assessed by calibrating the efficiency as a function of track
density. Therefore, efficiencies presented for WUSEM (Tables 1 and 2) only
hold for the track densities of the used samples.
15
4.3. Perspective of future development
The WUSEM algorithm represents an advance in automated track counting,
opening several possibilities. Differently from the direct application of classic
watershed, WUSEM allows adaptation (Figure 15): the user can determine optimal structuring elements and efficiencies using a training image subset and apply
these parameters to hers/his sample photomicrographs. This is very convenient
in fission track dating, where the tracks present the same shapes regardless of
the track source, and efficiency may vary mainly due to superposition of tracks
in higher track density samples.
Another possibility is applying WUSEM in separate parts of the input image,
allowing the determination of specific parameters for each track configuration.
This could bring even better results when dealing with the superposition of two,
three or more tracks.
5. Conclusion
In this paper we present a watershed algorithm using successive erosions as
markers, which we call WUSEM. We employ WUSEM to separate overlapping
tracks in photomicrographs. WUSEM performs well in images containing overlapping circular tracks (from a 78 Kr collimated beam) and in photomicrographs
with fission tracks at various orientations, both in DAP. The results are encouraging: the mean automatic/manual counting efficiency ratio is 0.97 ± 0.11 when
using WUSEM in the test datasets. We show also that diameter and eccentricity
criteria may be used to increase the reliability of this method.
Since WUSEM using circles as structuring elements is aimed to isotropically
shaped regions, this technique is suitable for separating etched tracks in DAP.
Etching velocity in mineral surfaces depends on the crystal orientation, yielding
more complex etching figures. Also, natural minerals are richer in scratches and
other etching figures that can be mistaken with fission tracks, especially when
using image processing techniques. WUSEM could be studied to separate tracks
16
in mineral surfaces; for that, it would need to use different structuring elements,
which have to consider the orientation and shape of each track.
Acknowledgements
The authors would like to thank Raymond Jonckheere for the second photomicrograph dataset and Matthias Schröter for his insights, and also Christina
Trautmann for the sample irradiation at GSI in Darmstadt. This work is supported by the São Paulo Research Foundation (FAPESP), grants # 2014/229220 and 2015/24582-4.
Supplementary Material
The supplementary material contains the complementary results and the
code published with this study, and is available at https://github.com/alexandrejaguar/
publications/tree/master/2017/dap_segmentation.
References
References
[1] J. M. C. Brown, S. Solomon, R. A. Tinker, Development of an energy
discriminate CR-39 R nuclear track etch dosimeter for Radon-220 gas measurements, Journal of Environmental Radioactivity 102 (10) (2011) 901–
905. doi:10.1016/j.jenvrad.2010.09.008.
[2] B. Smilgys, S. Guedes, M. Morales, F. Alvarez, J. C. Hadler, P. R. P.
Coelho, P. T. D. Siqueira, I. Alencar, C. J. Soares, E. A. C. Curvo,
Boron thin films and CR-39 detectors in BNCT: A method to measure
the 10B(n,α)7Li reaction rate, Radiation Measurements 50 (2013) 181–
186. doi:10.1016/j.radmeas.2012.07.001.
URL http://dx.doi.org/10.1016/j.radmeas.2012.07.001
17
[3] G. A. Wagner, P. Van den Haute, Fission-Track Dating, Springer Netherlands, Dordrecht, 1992. doi:10.1007/978-94-011-2478-2.
URL http://link.springer.com/10.1007/978-94-011-2478-2
[4] Y. Goto, T. Danhara, Zircon Fission-track Dating of the Hiyoriyama
Cryptodome at Kuttara Volcano, Southwestern Hokkaido, Japan, Bulletin of the volcanological society of Japan 56 (1) (2011) 19–23. doi:
10.18940/kazan.56.1_19.
[5] T. Imayama, T. Takeshita, K. Yi, D.-L. Cho, K. Kitajima, Y. Tsutsumi,
M. Kayama, H. Nishido, T. Okumura, K. Yagi, T. Itaya, Y. Sano,
Two-stage partial melting and contrasting cooling history within the
Higher Himalayan Crystalline Sequence in the far-eastern Nepal Himalaya,
Lithos 134-135 (2012) 1–22. doi:10.1016/j.lithos.2011.12.004.
URL
http://linkinghub.elsevier.com/retrieve/pii/
S002449371100377X
[6] H. Takagi, K. Tsutsui, H. Arai, H. Iwano, T. Danhara, Geological framework and fission track dating of pseudotachylyte of the Atotsugawa Fault,
Magawa area, central Japan, Island Arc 22 (3) (2013) 318–337. doi:
10.1111/iar.12036.
URL http://doi.wiley.com/10.1111/iar.12036
[7] T. Yuguchi, S. Sueoka, H. Iwano, T. Danhara, M. Ishibashi, E. Sasao,
T. Nishiyama, Spatial distribution of the apatite fission-track ages in the
Toki granite, central Japan: Exhumation rate of a Cretaceous pluton emplaced in the East Asian continental margin, Island Arc 26 (6) (2017) 1–15.
doi:10.1111/iar.12219.
URL http://doi.wiley.com/10.1111/iar.12219
[8] T. Ninomiya, S. Taniguchi, S. Shimoyama, K. Watanabe, T. Danhara,
H. Iwano, D. J. Dunkley, K. Shiraishi, C. Gouzu, U-Pb and fission-track
dating of a submarine pyroclastic rock from southwest Japan, Island Arc
18
26 (6) (2017) 1–11. doi:10.1111/iar.12215.
URL http://doi.wiley.com/10.1111/iar.12215
[9] R. L. Fleischer, P. B. Price, R. M. Walker, E. L. Hubbard, Track Registration in Various Solid-State Nuclear Track Detectors, Physical Review
133 (5A) (1964) A1443–A1449. doi:10.1103/PhysRev.133.A1443.
URL https://link.aps.org/doi/10.1103/PhysRev.133.A1443
[10] K. Moriwaki, Y. Shimpuku, M. Makimura, Improvement of Track Counting Accuracy and Efficiency in the Fission Track Method., Radioisotopes
47 (8) (1998) 611–616. doi:10.3769/radioisotopes.47.611.
URL
https://www.jstage.jst.go.jp/article/radioisotopes1952/
47/8/47_8_611/_article
[11] A. J. W. Gleadow, S. J. Gleadow, D. X. Belton, B. P. Kohn, M. S.
Krochmal, R. W. Brown, Coincidence mapping - a key strategy for the automatic counting of fission tracks in natural minerals, Geological Society,
London, Special Publications 324 (1) (2009) 25–36. doi:10.1144/SP324.2.
URL http://sp.lyellcollection.org/lookup/doi/10.1144/SP324.2
[12] R. C. Gonzalez, R. E. Woods, Digital Image Processing, 3rd Edition, Pearson, Upper Saddle River, NJ, USA, 2007.
[13] D. Azimi-Garakani, J. G. Williams, Automatic fission track counting using
the Quantimet 720, Nuclear Instruments and Methods 147 (1977) 69–73.
[14] J. Beer, J. Pipper, W. Heinrich, Automatic measurements of nuclear tracks
in plastic foils, Journal of Physics E: Scientific Instruments 15 (4) (1982)
439–443.
URL http://stacks.iop.org/0022-3735/15/i=4/a=013
[15] P. B. Price, W. Krischer, Semi-automated, three-dimensional measurement
of etched tracks in solid-state nuclear track detectors, Nuclear Instruments
and Methods in Physics Research Section A: Accelerators, Spectrometers,
Detectors and Associated Equipment 234 (1) (1985) 158–167.
19
[16] W. Trakowski, B. Schöfer, J. Dreute, S. Sonntag, C. Brechtmann, J. Beer,
H. Drechsel, W. Heinrich, An automatic measuring system for particle
tracks in plastic detectors, Nuclear Instruments and Methods in Physics
Research 225 (1) (1984) 92–100. doi:10.1016/0167-5087(84)91342-5.
[17] S. Masumoto, K. Wadatsumi, Computerized Image-Processing System for
Fission-Track Dating, Geoinformatics 1 (1) (1990) 93–102.
[18] K. Wadatsumi, S. Masumoto, Three-dimensional measurement of fissiontracks: Principles and an example in zircon from the Fish Canyon tuff,
International Journal of Radiation Applications and Instrumentation. Part
D. Nuclear Tracks and Radiation Measurements 17 (3) (1990) 399–406.
doi:10.1016/1359-0189(90)90063-4.
URL
http://linkinghub.elsevier.com/retrieve/pii/
1359018990900634
[19] A. P. Fews, Fully automated image analysis of etched tracks in CR39, Nuclear Instruments and Methods in Physics Research Section B:
Beam Interactions with Materials and Atoms 71 (4) (1992) 465–478.
doi:10.1016/0168-583X(92)95367-Z.
URL
http://linkinghub.elsevier.com/retrieve/pii/
0168583X9295367Z
[20] N. Petford, J. A. Miller, Three-dimensional imaging of fission tracks using
confocal scanning laser microscopy, American Mineralogist 77 (1992) 529–
533.
[21] N. Petford, J. A. Miller, J. Briggs, The automated counting of fission
tracks in an external detector by image analysis, Computers & Geosciences
19 (4) (1993) 585–591. doi:10.1016/0098-3004(93)90084-I.
URL
http://linkinghub.elsevier.com/retrieve/pii/
009830049390084I
[22] G. Espinosa, R. B. Gammage, K. E. Meyer, C. S. Dudney, Nuclear track
20
analysis by digital imaging, Radiation Protection Dosimetry 66 (1-4) (1996)
363–366.
[23] A. Boukhair, A. Haessler, J. C. Adlo, A. Nourreddine, New code for digital
imaging system for track measurements, Nuclear Instruments and Methods in Physics Research. Section B, Beam Interactions with Materials and
Atoms 160 (2000) 550–555.
[24] M. Dolleiser, S. R. Hashemi-Nezhad, A fully automated optical microscope
for analysis of particle tracks in solids, Nuclear Instruments and Methods in
Physics Research, Section B: Beam Interactions with Materials and Atoms
198 (1-2) (2002) 98–107. doi:10.1016/S0168-583X(02)01484-2.
[25] R. Bedogni, Development of a reader for the routine analysis of CR-39 fast
neutron dosemeters: improvement of the dosimetric performance using
automatic vision and motion tools, Radiation Measurements 36 (1-6)
(2003) 239–243. doi:10.1016/S1350-4487(03)00131-8.
URL
http://linkinghub.elsevier.com/retrieve/pii/
S1350448703001318
[26] N. Yasuda, K. Namiki, Y. Honma, Y. Umeshima, Y. Marumo, H. Ishii,
E. R. Benton, Development of a high speed imaging microscope and new
software for nuclear track detector analysis, Radiation Measurements 40 (26) (2005) 311–315. doi:10.1016/j.radmeas.2005.02.013.
[27] H. Tawara, K. Eda, K. Takahashi, T. Doke, N. Hasebe, S. Kodaira,
S. Ota, M. Kurano, N. Yasuda, Development of an automated multisample scanning system for nuclear track etched detectors, Nuclear Instruments and Methods in Physics Research, Section A: Accelerators, Spectrometers, Detectors and Associated Equipment 593 (3) (2008) 475–480.
doi:10.1016/j.nima.2008.05.035.
[28] A. F. de Siqueira, W. M. Nakasuga, A. Pagamisse, C. A. Tello Saenz, A. E.
Job, An automatic method for segmentation of fission tracks in epidote
21
crystal photomicrographs, Computers and Geosciences 69 (2014) 55–61.
doi:10.1016/j.cageo.2014.04.008.
URL http://dx.doi.org/10.1016/j.cageo.2014.04.008
[29] J. Ruan, H. Li, J. Zhang, Z. Zhang, L. Chen, J. Song, J. Liu, J. Liu,
Image mosaic method for recognition of proton track in nuclear emulsions,
Nuclear Instruments and Methods in Physics Research Section A: Accelerators, Spectrometers, Detectors and Associated Equipment 762 (2014)
16–21. doi:10.1016/j.nima.2014.05.040.
URL
http://dx.doi.org/10.1016/j.nima.2014.05.040http:
//linkinghub.elsevier.com/retrieve/pii/S0168900214005671
[30] E. Enkelmann, T. A. Ehlers, G. Buck, A. K. Schatz, Advantages and challenges of automated apatite fission track counting, Chemical Geology 322323 (2012) 278–289. doi:10.1016/j.chemgeo.2012.07.013.
URL http://dx.doi.org/10.1016/j.chemgeo.2012.07.013
[31] S. Beucher, C. Lantuejoul, Use of Watersheds in Contour Detection (1979).
doi:citeulike-article-id:4083187.
URL http://www.citeulike.org/group/7252/article/4083187
[32] J. Serra, Image analysis and mathematical morphology, Vol. 1, Academic
press, 1982.
[33] F. M. Schaller, M. Neudecker, M. Saadatfar, G. Delaney, K. Mecke, G. E.
Schröder-Turk, M. Schröter, Tomographic analysis of jammed ellipsoid
packings, in: AIP Conference Proceedings, Vol. 1542, 2013, pp. 377–380.
doi:10.1063/1.4811946.
URL http://aip.scitation.org/doi/abs/10.1063/1.4811946
[34] G. H. Ball, D. J. Hall, A clustering technique for summarizing multivariate data, Behavioral Science 12 (2) (1967) 153–155. doi:10.1002/bs.
3830120210.
22
[35] T. W. Ridler, S. Calvard, Picture Thresholding Using an Iterative Selection
Method, IEEE Transactions on Systems, Man and Cybernetics 8 (8) (1978)
630–632. doi:10.1109/TSMC.1978.4310039.
[36] F. R. Dias Velasco, Thresholding Using the ISODATA Clustering Algorithm, IEEE Transactions on Systems, Man, and Cybernetics 10 (11) (1980)
771–774. doi:10.1109/TSMC.1980.4308400.
URL http://ieeexplore.ieee.org/document/4308400/
[37] S. van der Walt, J. L. Schönberger, J. Nunez-Iglesias, F. Boulogne, J. D.
Warner, N. Yager, E. Gouillart, T. Yu, scikit-image: image processing in
Python, PeerJ 2 (2014) e453. arXiv:1407.6245, doi:10.7717/peerj.453.
URL https://peerj.com/articles/453
[38] C. Fiorio, J. Gustedt, Two linear time Union-Find strategies for image
processing, Theoretical Computer Science 154 (2) (1996) 165–181. doi:
10.1016/0304-3975(94)00262-2.
[39] J. C. Russ, The Image Processing Handbook, 6th Edition, CRC Press, 2011.
URL https://www.crcpress.com/The-Image-Processing-Handbook-Sixth-Edition/
Russ/p/book/9781439840450
[40] J. F. Ziegler, M. D. Ziegler, J. P. Biersack, SRIM - The stopping and range
of ions in matter (2010), Nuclear Instruments and Methods in Physics
Research, Section B: Beam Interactions with Materials and Atoms 268 (1112) (2010) 1818–1823. doi:10.1016/j.nimb.2010.02.091.
URL http://dx.doi.org/10.1016/j.nimb.2010.02.091
[41] S. van der Walt, S. C. Colbert, G. Varoquaux, The NumPy Array: A
Structure for Efficient Numerical Computation, Computing in Science &
Engineering 13 (2) (2011) 22–30. doi:10.1109/MCSE.2011.37.
URL http://ieeexplore.ieee.org/document/5725236/
[42] E. Jones, T. Oliphant, P. Peterson, SciPy: Open source scientific tools for
23
Python (2001–).
URL http://www.scipy.org/
[43] J. D. Hunter, Matplotlib: A 2D Graphics Environment, Computing in
Science & Engineering 9 (3) (2007) 90–95. doi:10.1109/MCSE.2007.55.
URL http://ieeexplore.ieee.org/document/4160265/
[44] G. van Rossum, Python Reference Manual, Tech. rep., Centrum voor
Wiskunde en Informatica (CWI), Amsterdam (1995).
URL https://ir.cwi.nl/pub/5008/05008D.pdf
[45] C. T. Rueden, J. Schindelin, M. C. Hiner, B. E. DeZonia, A. E. Walter, E. T. Arena, K. W. Eliceiri, ImageJ2: ImageJ for the next generation of scientific image data, BMC Bioinformatics 18 (1) (2017) 529.
doi:10.1186/s12859-017-1934-z.
URL https://bmcbioinformatics.biomedcentral.com/articles/10.
1186/s12859-017-1934-z
[46] S. Guedes, J. C. Hadler, P. J. Iunes, S. R. Paulo, C. A. Tello, On the
reproducibility of SSNTD track counting efficiency, Nuclear Instruments
and Methods in Physics Research Section A: Accelerators, Spectrometers, Detectors and Associated Equipment 418 (2-3) (1998) 429–433.
doi:10.1016/S0168-9002(98)00918-8.
URL
http://linkinghub.elsevier.com/retrieve/pii/
S0168900298009188
[47] A. J. Hurford, P. F. Green, The zeta age calibration of fission-track dating,
Chemical Geology 41 (1983) 285–317.
doi:10.1016/S0009-2541(83)
80026-6.
URL
http://linkinghub.elsevier.com/retrieve/pii/
S0009254183800266
[48] A. J. Hurford, Standardization of fission track dating calibration: Recommendation by the Fission Track Working Group of the I.U.G.S. Subcom-
24
mission on Geochronology, Chemical Geology: Isotope Geoscience Section
80 (2) (1990) 171–178. doi:10.1016/0168-9622(90)90025-8.
[49] C. J. Soares, I. Alencar, S. Guedes, R. H. Takizawa, B. Smilgys, J. C.
Hadler, Alpha spectrometry study on LR 115 and Makrofol through measurements of track diameter, Radiation Measurements 50 (2013) 246–248.
doi:10.1016/j.radmeas.2012.06.010.
[50] J. C. Hadler, P. J. Iunes, C. A. Tello, F. Chemale, K. Kawashita, E. A. C.
Curvo, F. G. S. Santos, T. E. Gasparini, P. A. F. P. Moreira, S. Guedes,
Experimental study of a methodology for Fission-track Dating without
neutron irradiation, Radiation Measurements 44 (9-10) (2009) 955–957.
doi:10.1016/j.radmeas.2009.10.081.
URL
http://dx.doi.org/10.1016/j.radmeas.2009.10.081http:
//linkinghub.elsevier.com/retrieve/pii/S1350448709002571
[51] N. Hasebe, J. Barbarand, K. Jarvis, A. Carter, A. J. Hurford, Apatite
fission-track chronometry using laser ablation ICP-MS, Chemical Geology
207 (3-4) (2004) 135–145. doi:10.1016/j.chemgeo.2004.01.007.
[52] C. J. Soares, S. Guedes, J. C. Hadler, R. Mertz-Kraus, T. Zack, P. J.
Iunes, Novel calibration for LA-ICP-MS-based fission-track thermochronology, Physics and Chemistry of Minerals 41 (1) (2014) 65–73.
doi:
10.1007/s00269-013-0624-2.
[53] C. J. Soares, R. Mertz-Kraus, S. Guedes, D. F. Stockli, T. Zack, Characterisation of Apatites as Potential Uranium Reference Materials for Fissiontrack Dating by LA-ICP-MS, Geostandards and Geoanalytical Research
39 (3) (2015) 305–313. doi:10.1111/j.1751-908X.2014.00301.x.
URL http://doi.wiley.com/10.1111/j.1751-908X.2014.00301.x
[54] D. J. Gombosi, J. I. Garver, S. L. Baldwin, On the development of electron microprobe zircon fission-track geochronology, Chemical Geology 363
(2014) 312–321. doi:10.1016/j.chemgeo.2013.11.005.
URL http://dx.doi.org/10.1016/j.chemgeo.2013.11.005
25
[55] A. N. C. Dias, F. Chemale, C. J. Soares, S. Guedes, A new approach
for electron microprobe zircon fission track thermochronology, Chemical
Geology 459 (January) (2017) 129–136. doi:10.1016/j.chemgeo.2017.
04.014.
URL http://dx.doi.org/10.1016/j.chemgeo.2017.04.014
26
Figure 7: Processing an input image using WUSEM. Each line presents a segmentation step.
Left column: original binary image (red) and erosion obtained according to the variation
of delta radius (white). Center column: erosion results labeled, used as markers. Right
column: watershed results when using generated markers. Parameters for WUSEM algorithm:
initial radius = 10, delta radius = 4.
27
Figure 8: (a) Labels generated from tracks in Figure 4 using the WUSEM algorithm. (b)
Tracks in (a) enumerated using enumerate objects(). Tracks in the lower or right corners
are not counted, according to the “lower right corner” method. Parameters for WUSEM
algorithm: initial radius = 10, delta radius = 4. Colormaps: (a) nipy spectral, (b)
gray.
Figure 9: Manual counting mean (top of the blue bar; values on the right) for each sample
and automatic counting results with mean within (orange points) and outside (gray points)
the tolerance interval (blue bar) for the first dataset.
28
Figure 10: Comparison between manual and automatic counting for (a, b) 4.5 min etching
samples and (c, d) 8.5 min etching samples. (a, c) white: manual counting. Gray: flooding
watershed counting. Red line: distribution median. White signal: distribution mean. (b, d)
dashed: 1:1 line. Red line: regression for the WUSEM counting data. Black line: regression
for the flooding watershed counting data.
Figure 11: Regions from Figure 4 complying with ≤ 0.3. (a) labeled regions. (b) tracks
correspondent to (a) in their original gray levels. Colormaps: (a) nipy spectral. (b) magma.
29
Figure 12: Relation between incident energy versus mean diameter product ((a) 4.5 min; (c)
8.5 min samples, left Y axis) and incident energy versus mean gray levels ((b) 4.5 min; (d)
8.5 min samples, left Y axis). Cyan dashed line: electronic energy loss calculated with SRIM
(right Y axis).
Figure 13: (a) Input photomicrograph (Figure 5) binarized using the ISODATA threshold
(threshold = 0.59) and region filling. (b) Tracks separated in Figure 5 using the WUSEM
algorithm, and then enumerated using the function enumerate objects(). Parameters for
WUSEM algorithm: initial radius = 5, delta radius = 12.
30
Figure 14: Comparison between manual and automatic counting for photomicrographs in
dataset 2.
(a) white: manual counting.
Gray: flooding watershed counting.
Red line:
distribution median. White signal: distribution mean. Blue dots: outliers. (b) dashed: 1:1
line. Red line: regression for the WUSEM counting data. Black line: regression for the
flooding watershed counting data.
Figure 15: When using suitable input parameters, WUSEM may perform better in certain
regions where the classic watershed does not return reliable results. For instance, the highlighted region in (a) presents three tracks. WUSEM separates them correctly, but the region
is oversegmented by the classic watershed. The highlighted region in (b), by its turn, is
undersegmented by the classic watershed, which returns two tracks. WUSEM returns three
tracks, being closer to the real number (four tracks). Left: input photomicrographs with
highlighted regions. Center: tracks separated using WUSEM. Right: tracks separated using
classic watershed. Parameters for WUSEM algorithm: initial radius = 15, delta radius
= 4.
31
| 1cs.CV
|
1
Sparse Activity Detection for Massive Connectivity
arXiv:1801.05873v1 [cs.IT] 17 Jan 2018
Zhilin Chen, Student Member, IEEE, Foad Sohrabi, Student Member, IEEE, and Wei Yu, Fellow, IEEE
Abstract—This paper considers the massive connectivity application in which a large number of potential devices communicate
with a base-station (BS) in a sporadic fashion. The detection of
device activity pattern together with the estimation of the channel
are central problems in such a scenario. Due to the large number
of potential devices in the network, the devices need to be assigned
non-orthogonal signature sequences. The main objective of this
paper is to show that by using random signature sequences and
by exploiting sparsity in the user activity pattern, the joint user
detection and channel estimation problem can be formulated as
a compressed sensing single measurement vector (SMV) problem
or multiple measurement vector (MMV) problem, depending on
whether the BS has a single antenna or multiple antennas, and be
efficiently solved using an approximate message passing (AMP)
algorithm. This paper proposes an AMP algorithm design that
exploits the statistics of the wireless channel and provides an
analytical characterization of the probabilities of false alarm and
missed detection by using the state evolution. We consider two
cases depending on whether the large-scale component of the
channel fading is known at the BS and design the minimum
mean squared error (MMSE) denoiser for AMP according to the
channel statistics. Simulation results demonstrate the substantial
advantage of exploiting the statistical channel information in
AMP design; however, knowing the large-scale fading component
does not offer tangible benefits. For the multiple-antenna case,
we employ two different AMP algorithms, namely the AMP with
vector denoiser and the parallel AMP-MMV, and quantify the
benefit of deploying multiple antennas at the BS.
Index Terms—Device activity detection, channel estimation,
approximate message passing, compressed sensing, Internet of
Things (IoT), machine-type communications (MTC)
I. I NTRODUCTION
One of the key requirements for the next-generation wireless cellular networks is to provide massive connectivity for
machine-type communications (MTC), envisioned to support
diverse applications such as environment sensing, event detection, surveillance and control [1], [2]. Machine-centric
communications have two distinctive features as compared to
conventional human-centric communications: (i) the overall
system needs to support massive connectivity—the number
of devices connected to each cellular base-station (BS) may
be in the order of 104 to 106 ; and (ii) the traffic pattern of
each device may be sporadic—at any given time only a small
fraction of potential devices are active. For such a network,
accurate user activity detection and channel estimation are
Manuscript accepted and to appear in IEEE Transactions on Signal Processing. This work has been presented in part at IEEE International Conference on
Acoustics, Speech, and Signal Processing (ICASSP), March 2017. This work
is supported by Natural Sciences and Engineering Research Council (NSERC)
of Canada through a Discovery Grant and through a Steacie Memorial
Fellowship.
The authors are with The Edward S. Rogers Sr. Department of Electrical
and Computer Engineering, University of Toronto, Toronto, ON M5S 3G4,
Canada (e-mails:{zchen, fsohrabi, weiyu}@comm.utoronto.ca).
crucial for establishing successful communications between
the devices and the BS.
To identify active users and to estimate their channels,
each user must be assigned a unique signature sequence.
However, due to the large number of potential devices but
the limited coherence time and frequency dimensions in the
wireless fading channel, the signature sequences for all users
cannot be mutually orthogonal. Non-orthogonal signature sequences superimposed in the pilot stage causes significant
multi-user interference, e.g., when a simple matched filtering or correlation operation is applied at the BS for user
activity detection and channel estimation. A key observation
of this paper is that the sporadic nature of the traffic leads
to sparse user transmission patterns. By exploiting sparsity
and by formulating the detection and estimation problem
with independent identically distributed (i.i.d.) random nonorthogonal pilots as a compressed sensing problem, this multiuser interference problem can be overcome, and highly reliable
activity detection and accurate channel estimation can be made
possible. In the compressed sensing terminology, when the
BS is equipped with a single antenna, activity detection and
channel estimation can be formulated as a single measurement
vector (SMV) problem; when the BS has multiple antennas,
the problem can be formulated as a multiple measurement
vector (MMV) problem.
This paper proposes the use of compressed sensing techniques for the joint user activity detection and channel estimation problem. Due to the large-scale nature of massive
device communications, this paper adopts the computationally
efficient approximate message passing (AMP) algorithm [3]
as the main technique. AMP is an iterative thresholding
method with a key feature that allows analytic performance
characterization via the so-called state evolution. The main
contributions of this paper are: (i) a novel AMP algorithm
design for user activity detection that exploits the statistical
information of the wireless channel; and (ii) a characterization
of the probabilities of false alarm and missed detection for both
SMV and MMV scenarios.
A. Related Work
The user activity detection problem for massive connectivity
has been studied from information theoretical perspectives
in [2], [4]. From an algorithmic point of view, the problem
is closely related to sparse recovery in compressed sensing
and has been studied in a variety of wireless communication
settings. For example, assuming no prior knowledge of the
channel state information (CSI), joint user activity detection
and channel estimation is considered in [5]–[8]. Specifically,
[5] proposes an efficient greedy algorithm based on orthogonal
matching pursuit for sporadic multi-user communication. By
2
exploiting the statistics of channel path-loss and the joint
sparsity structures, [6] proposes a modified Bayesian compressed sensing algorithm in a cloud radio-access network.
In the context of orthogonal frequency division multiplexing
(OFDM) systems, [7] introduces a one-shot random access
protocol and employs the basis pursuit denoising detection
method with a detection error bound based on the restricted
isometry property. The performance of such schemes in a
practical setting is illustrated in [7], [8]. When perfect CSI
is assumed, joint user activity and data detection for code
division multiple access systems (CDMA) is investigated in
[9], [10], where [9] designs the sparsity-exploiting maximum
a posteriori detector by accounting for both sparsity and finitealphabet constraints of the signal, and [10] proposes a greedy
block-wise orthogonal least square algorithm by exploiting
the block sparsity among several symbol durations. Differing
from most of the above works that consider cellular systems,
[11], [12] study the user activity detection in wireless ad hoc
networks, where each node in the system identifies its neighbor
nodes simultaneously. The authors of [11] propose a scalable
compressed neighbor discovery scheme that employs random
binary signatures and group testing based detection algorithms.
In [12], the authors propose a more dedicated scheme that uses
signatures based on Reed-Muller code and a chirp decoding
algorithm to achieve a better performance.
In contrast to the aforementioned works, this paper adopts
the more computationally efficient AMP algorithm for user activity detection and channel estimation, which is more suitable
for large-scale networks with a large number of devices. The
AMP algorithm is first proposed in [3] as a low-complexity
iterative algorithm for conventional compressed sensing with
real-valued signals and real-valued measurements. A framework of state evolution that tracks the performance of AMP
at each iteration is introduced in [3]. The AMP algorithm is
then extended along different directions. For example, [13]
generalizes the AMP algorithm to a broad family of iterative
thresholding algorithms, and provides a rigorous proof of the
framework of the state evolution. To deal with complex-valued
signals and measurements, [14] proposes a complex AMP
algorithm (CAMP). By exploiting the input and output distributions, a generalized AMP (GAMP) algorithm is designed in
[15]. Similarly, a Bayesian approach is used to design the AMP
algorithm in [16], [17] by accounting for the input distribution.
For the compressed sensing problem with multiple signals
sharing joint sparsity, i.e., the MMV problem, [18] designs an
AMP algorithm via a vector form of message passing; and [19]
designs the AMP-MMV algorithm by directly using message
passing over a multi-frame factor graph.
Although the use of the AMP algorithm for user activity
detection has been previously proposed in [20], the statistical
information of the channel is not exploited in the prior work;
also performance analysis is not yet available. This paper
makes progress by showing that exploiting channel statistics
can significantly enhance the Bayesian AMP algorithm. Moreover, analytical performance characterization can be obtained
by using state evolution. Finally, the AMP algorithm can be
extended to the multiple-antenna case.
B. Main Contributions
This paper considers the user activity detection and channel
estimation problem in the uplink of a single-cell network with
a large number of potential users, but at any given time slot
only a small fraction of them are active. To exploit the sparsity
in user activity pattern, this paper formulates the problem as a
compressed sensing problem and proposes the use of random
signature sequences and the computationally efficient AMP
algorithm for device activity detection. This paper provides the
design and analysis of AMP for both cases in which the BS
is equipped with a single antenna and with multiple antennas.
This paper considers two different scenarios: (i) when the
large-scale fading coefficients of all user are known and the
detector is designed based on the statistics of fast fading component only; and (ii) when the large-scale fading coefficients
are not known and the detector is designed based on the
statistics of both fast fading and large-scale fading components
as a function of the distribution of device locations in the cell.
The proposed AMP-based detector exploits the statistics of the
wireless channel by specifically designing the minimum mean
squared error (MMSE) denoiser. This paper provides analytical
characterization of the probabilities of false alarm and missed
detection via the state evolution for both scenarios.
For the case where the BS is equipped with a single
antenna, numerical results indicate that: (i) the analytic performance characterization via state evolution is very close to
the simulation; (ii) exploiting the statistical information of
the channel and user activity can significantly improve the
detector performance; and (iii) knowing the large-scale fading
coefficient actually does not bring substantial performance
improvement as compared to the case that only the statistical
information about the large-scale fading is available.
For the case where the BS is equipped with multiple antennas, this paper considers both the AMP with vector denoiser
[18] and the parallel AMP-MMV [17]. For the AMP with
vector denoiser, this paper exploits wireless channel statistics
in denoiser design and further analytically characterizes the
probabilities of false alarm and missed detection based on
the state evolution. For the parallel AMP-MMV algorithm,
which is more suitable for distributed computation, performance characterization is more difficult to obtain. Simulation
results show that: (i) having multiple antennas at the BS
can significantly improve the detector performance; (ii) the
predicted performance of AMP with vector denoiser is very
close to its simulated performance; and (iii) AMP with vector
denoiser and parallel AMP-MMV achieve approximately the
same performance.
C. Paper Organization and Notations
The remainder of this paper is organized as follows. Section II introduces the system model. Section III introduces
the AMP algorithms for both SMV and MMV problems.
Section IV considers user activity detection and channel estimation when the BS has a single-antenna, while Section V
considers the multiple-antenna case. Simulation results are
provided in Section VI. Conclusions are drawn in Section VII.
3
Throughout this paper, upper-case and lower-case letters
denote random variables and their realizations, respectively.
Boldface lower-case letters denote vectors. Boldface uppercase letters denote matrices or random vectors, where context
should make the distinction clear. Superscripts (·)T , (·)∗
and (·)−1 denote transpose, conjugate transpose, and inverse
operators, respectively. Further, I denotes identity matrix with
appropriate dimensions, E[·] denotes expectation operation,
, denotes definition, | · | denotes either the magnitude of a
complex variable or the determinant of a matrix, depending
on the context, and k · k2 denotes the `2 norm.
II. S YSTEM M ODEL
Consider the uplink of a wireless cellular system with one
BS located at the center and N single-antenna devices located
uniformly in a circular area with radius R, but in each coherence block only a subset of users are active. Let an ∈ {1, 0}
indicate whether or not user n is active. For the purpose of
channel probing and user identification, user n is assigned
a unique signature sequence sn = [s1n , s2n , · · · , sLn ]T ∈
CL×1 , where L is the length of the sequence. This paper
assumes that the signature sequence sn is generated according
to i.i.d. complex Gaussian distribution with zero mean and
variance 1/L such that each sequence is normalized to have
unit power, and the normalization factor 1/L is incorporated
into the transmit power.
We consider a block-fading channel model where the channel is static in each block. In this paper, we consider two
cases where the BS is equipped with either a single antenna
or multiple antennas. When the BS has only one antenna, the
received signal at the BS can be modeled as
y=
N
X
(1)
where hn ∈ C is the channel coefficient between user n
and the BS, w ∈ CL×1 is the effective complex Gaus2
depends on the background
sian noise whose variance σw
noise power normalized by the user transmit power. Here,
x , [x1 , x2 , · · · , xN ]T ∈ CN ×1 where xn , hn an , and
S , [s1 , s2 , · · · , sN ] ∈ CL×N .
We aim to recover the non-zero entries of x based on the
received signals y. We are interested in the regime where
the number of potential users is much larger than the pilot
sequence length, i.e., N L, so that the user pilot sequences
cannot be mutually orthogonal; but due to the sporadic traffic,
only a small number of devices transmit in each block,
resulting in a sparse x. The recovering of x for the single
antenna case is in the form of the SMV problem in compressed
sensing.
This paper also considers the case where the BS is equipped
with M antennas. In this case, the received signal Y ∈ CL×M
at the BS can be expressed in matrix form as
N
X
pR|G (rn |gn ) = (1 − λ)δ0 + λpH|G (rn |gn ),
an sn hn + W , SX + W,
(2)
n=1
where hn ∈ C1×M is the channel vector between user n and
the BS, W ∈ CL×M is the effective complex Gaussian noise,
(3)
where δ0 denotes the point mass measure at 0, pH|G denotes
the probability density function (pdf) of the channel vector
H given prior information G, which has a pdf pG , and gn
denotes the prior information for user n. Note that we use
H to denote the random channel vector and hn to denote its
realization. Based on (3), the pdf of the entries of x is
pX|G (xn |gn ) = (1 − λ)δ0 + λpH|G (xn |gn ).
an sn hn + w , Sx + w,
n=1
Y=
and X , [rT1 , · · · , rTN ]T ∈ CN ×M where rn , an hn ∈
C1×M is the nth row vector of X. We also use cm ∈ CN ×1 to
represent the mth column vector of X, i.e., X = [c1 , · · · , cM ].
Note that an indicates whether the entire row vector rn is zero
or not. In other words, columns of X (i.e., cm ) share the same
sparsity pattern.
We are interested in detecting the user activity an as well
as in estimating the channel gains of the active users, which
correspond to the non-zero rows of the matrix X, based on
the observation Y in the regime where N L. The problem
of recovering X from Y is in the form of the MMV problem
in compressed sensing.
A key observation of this paper is that the design of recovery
algorithm can be significantly enhanced by taking advantage
of the knowledge about the statistical information of x or X.
Toward this end, we provide a model for the distribution of
the entries of x, and the distribution of the rows of X. Since
x is a special case of X when M = 1, we focus on the model
for X.
We assume that each user accesses the channel with a small
probability λ in an i.i.d. fashion, i.e., Pr(an = 1) = λ, ∀n,
and there is no correlation between different users’ channels,
so that the row vectors of X follow a mixture distribution
(4)
To model the distribution of H, we assume that all users are
randomly and uniformly located in a circular coverage area of
radius R with the BS at the center, and the channels between
the users and the BS follow an independent distribution that
depends on the distance. More specifically, H includes pathloss, shadowing, and Rayleigh fading. The path-loss between
a user and the BS is modeled (in dB) as α + β log10 (d),
where d is the distance measured in meter, α is the fading
coefficient at d = 1, and β is the path-loss exponent. The
shadowing (in dB) follows a Gaussian distribution with zero
2
mean and variance σSF
. The Rayleigh fading is assumed to
be i.i.d. complex Gaussian with zero mean and unit variance
across all antennas.
The large-scale fading, which includes path-loss and shadowing, is denoted as G, whose pdf pG can be modeled by the
2
distribution of BS-user distance and shadowing parameter σSF
.
This paper considers both the case where only the statistics
of the the large-scale fading, i.e., pG , is known as well as
the case where the exact large-scale fading coefficient gn is
known at the BS. The latter case is motivated by the scenario
in which the devices are stationary, so that the path-loss and
shadowing can be estimated and stored at the BS as prior
information. When gn is known, pH|G captures the distribution
of the Rayleigh fading component. When only p(G) is known,
we drop G and gn from pH|G (hn |gn ), and write it as pH (hn ),
4
which captures the distribution of both large-scale fading and
Rayleigh fading.
III. AMP A LGORITHM
AMP is an iterative algorithm that recovers sparse signal
for compressed sensing. We introduce the AMP framework
for both the SMV and the MMV problems in this section.
A. AMP for SMV problem
AMP is first proposed for the SMV problem in [3]. Starting
with x0 = 0 and z0 = y, AMP proceeds at each iteration as
xt+1 = η(S∗ zt + xt , g, t),
N
zt+1 = y − Sxt+1 + zt hη 0 (S∗ zt + xt , g, t)i,
L
(5)
(6)
where g , [g1 , · · · , gN ]T , and t = 0, 1, · · · is the index of
iteration, xt is the estimate of x at iteration t, zt is the residual,
η(·, g, t) , [ηt (·, g1 ), · · · , ηt (·, gN )]T where ηt (·, gn ) : C →
C is an appropriately designed non-linear function known as
denoiser that operates on the nth entry of the input vector,
η 0 (·) , [ηt0 (·, g1 ), · · · , ηt0 (·, gN )]T where ηt0 (·, gn ) is the first
order derivative of ηt (·, gn ) with respect to the first argument,
and h·i is averaging operation over all entries of a vector. Note
that the third term in the right hand side of (6) is the correction
term known as the “Onsager term” from statistical physics.
In the AMP algorithm, the matched filtered output x̃t ,
S∗ zt + xt can be modeled as signal x plus noise (including
multiuser interference), i.e., x̃t = x + vt , where vt is
Gaussian due to the correction term. The denoiser is typically
designed to reduce the estimation error at each iteration. In
the compressed sensing literature, the prior distribution of x
is usually assumed to be unknown. In this case, a minimax
framework over the worst case x leads to a soft thresholding
denoiser [21]. When the prior distribution of x is known, the
Bayesian framework then can be used to account for the prior
information on x [16]. In this paper, we adopt the Bayesian
approach and design the MMSE denoiser for the massive
connectivity setup as shown in the next section.
The AMP algorithm can be analyzed in the asymptotic
regime where L, N → ∞ with fixed N/L via the state
evolution, which predicts the per-coordinate performance of
the AMP algorithm at each iteration as follows
2
2
τt+1
= σw
+
N
2
E |ηt (X + τt V, G) − X| ,
L
(7)
where τt is referred to as the state, X, V , and G are random
variables with X following pX|G , V following the complex
Gaussian distribution with zero mean and unite variance, and
G following pG , and the expectation is taken over all X, V ,
and G. We denote X̃ t , X + τt V . The random variables X,
V , G X̃ t capture the distributions of the entries of x, entries
of vt (up to a factor τt ), the
prior informationgn , and entries
of x̃t , respectively, with E |ηt (X̃ t , G) − X|2 characterizing
the per-coordinate MSE of the estimate of x at iteration t.
B. AMP for MMV problem
1) AMP with vector denoiser: One extension of the AMP
algorithm to solve the MMV problem in (2) is proposed in
[18], which employs a vector denoiser that operates on each
row vector of the matched filtered output:
Xt+1 = η(S∗ Zt + Xt , g, t),
N
Zt+1 = Y − SXt+1 + Zt hη 0 (S∗ Zt + Xt , g, t)i,
L
(8)
(9)
where η(·, g, t) , [ηt (·, g1 ), · · · , ηt (·, gN )]T with ηt (·, gn ) :
C1×M → C1×M is a vector denoiser that operates on the nth
row vector of S∗ Zt + Xt , and the other notations are similar
to those used in (5) and (6). The state evolution of the AMP
algorithm for the MMV problem also has a similar form as
N
2
Σt+1 = σw
I + E Dt (Dt )∗ ,
(10)
L
T
where Dt , (ηt (R + Ut , G) − R) ∈ CM ×1 , with random
vector R following pR|G and random vector Ut following
CN (0, Σt ). The expectation is taken over R, Ut , and G. To
minimize the estimation error at each iteration, we can also
design the vector denoiser ηt (·, ·) via the Bayesian approach.
2) Parallel AMP-MMV: A different extension of the AMP
algorithm for dealing with the MMV problem is the parallel
AMP-MMV algorithm proposed in [19]. The basic idea is to
solve the MMV problem iteratively by using multiple parallel
AMP-SMVs then exchanging soft information between them.
Parallelization allows distributed implementation of the algorithm, which can be computationally advantageous, especially
when the number of antennas is large.
The outline of the parallel AMP-MMV algorithm is illustrated in Algorithm 1 which operates on a per-antenna
basis, i.e., on the columns of X and Z, denoted as
cm and zm respectively, and where η(·, g, t, i, m) ,
[ηt,i,m (·, g1 ), · · · , ηt,i,m (·, gN )]T is the denoiser used for the
mth antenna in the iteration (t, i). Note that here we add
index i and m in the notation of denoiser, ηt,i,m (·, gn ), to
indicate the index of outer iteration and the index of SMV
stage, respectively. In the first phase which is called the (into)(
phase, the messages π nm , are calculated and passed to the
mth AMP-SMV stage. These messages convey the current
belief about the probability of being active for each user. In
(
the first iteration, we have π nm = λ, ∀n, m, since no further
information is available. In the next phase, which is called the
(within)-phase, the conventional AMP algorithm is applied to
the received signal of each antenna. Note that the denoiser in
AMP algorithm is a function of the current belief about the
activity of the users which is obtained based on the information
sharing between all M AMP-SMV stages. Finally, in the (out)phase, the estimate of channel gains is used to refine the belief
about the activity of the users.
IV. U SER ACTIVITY D ETECTION : S INGLE -A NTENNA C ASE
A main point of this paper is that exploiting the statistics
of the wireless channel can significantly enhance detector
performance. This section proposes an MMSE denoiser design
for the AMP algorithm for the wireless massive connectivity
5
√
−10 2
−α − β log10 (R) 20
√
b=
−
, c=
.
(ln 10)σSF
βb
2σSF
Proof. See Appendix A.
Algorithm 1 Parallel AMP-MMV Method [19]
1:
2:
3:
4:
5:
6:
7:
8:
9:
10:
11:
12:
*
Initialize π nm = 0.5, ∀n, m.
for i = 1 to I do
Execute the (into)-phase: Q
π nm =
*
m0 6=m π nm0
Q
*
*
m0 6=m (1− π nm0 )+λ
m0 6=m π nm0
λ
(
(1−λ)
Q
, ∀n, m
Execute the (within)-phase:
for m = 1 to M do
Initialize c0m = 0, z0m = ym .
for t = 0 to T do
∗ t
t
ct+1
m = η(S zm + cm , g, t, i, m),
t
0
∗ t
t+1
zm
= ym − Sct+1
+ N
m
L zm hη (S zm +
t
cm , g, t, i, m)i,
end for
end for
Execute the (out)-phase:
*
Calculate π nm , ∀n, m, the probability of user n being
active based on the decision at the mth AMP-SMV stage.
end for
problem that specifically takes wireless channel characteristics
into consideration in the single-antenna case. Two scenarios
are considered: the large-scale fading gn of each user is either
directly available or only its statistics is available at the BS.
This section further studies the optimal detection strategy, and
analyzes the probabilities of false alarm and missed detection
by using the state evolution of the AMP algorithm.
A. MMSE Denoiser for AMP Algorithm
In the scenario where only the statistics about the largescale fading is known at the BS, the distributions of the
channel coefficients pH (hn ) are independent and identical for
all devices. In the scenario where the devices are stationary and
their path-loss and shadowing coefficients can be estimated
and thus the exact large-scale fading is known at the BS,
the distributions of the channel coefficients are of the form
pH|G (hn |gn ), which are complex Gaussian with variance
parameterized by gn , and are independent but not identical
across the devices. To derive the MMSE denoisers via the
Baysian approach for both cases, we first characterize the
distributions pG (gn ), pH (hn ) and pH|G (hn |gn ) as follows.
Proposition 1. Consider a circular wireless cellular coverage area of radius R with BS at the center and uniformly
distributed devices where the channels between the BS and
the devices are modeled with large-scale fading gn with
parameters α, β and shadowing fading parameter σSF as
defined in the system model. Then, gn follows a distribution
as
pG (gn ) = agn−γ Q(gn ),
(11)
R∞
where Q(gn ) , (b ln gn +c) exp(−s2 )ds, γ , 40/β + 1, and
a, b, and c are constants depending on parameters α, β, σSF
and R as
2
40
2(ln 10)2 σSF
2 ln(10)α
a = 2 √ exp
−
,
β2
β
R β π
Proposition 2. Denote hn as the channel coefficient which
contains both the large-scale fading gn and Rayleigh fading.
If only pG (gn ) is known at the BS, the pdf of hn is given by
Z ∞
−|hn |2
a −γ−2
dgn . (12)
g
Q(gn ) exp
pH (hn ) =
π n
gn2
0
If gn is known at the BS, the pdf of hn given gn is
1
−|hn |2
pH|G (hn |gn ) =
exp
.
(13)
πgn2
gn2
Proof. See Appendix B.
Note that for the first scenario, the channel distribution
(12) only depends on a few parameters such as the path-loss
exponent in the path-loss model and the standard deviation in
the shadowing model, which are assumed to be known and can
be estimated in practice. For the second scenario, the channel
distribution (13) is just a Rayleigh fading model parameterized
by the large-scale fading. The large-scale fading information
can be obtained by tracking the estimated channel over a
reasonable period. This second scenario is applicable to the
case where the users are mostly stationary, so the large-scale
fading changes only slowly over time. It is worth noting that
although this paper restricts attention to the Rayleigh fading
model, the approach developed here is equally applicable for
Rician or any other statistical channel model.
In the following, we design the MMSE denoisers for the
AMP algorithm by exploiting pH (hn ) and pH|G (hn |gn ).
1) With Statistical Knowledge of Large-Scale Fading Only:
Since gn is unknown and only the distribution pG (gn ) is
available at the BS, the denoiser ηt (·, gn ) reduces to ηt (·),
which indicates that the denoiser for each entry of the matched
filtered output is the same. By using (12), the pdf of the entries
of x can be expressed as
Z ∞
exp(−|xn |2 gn−2 )
pX (xn ) = (1 − λ)δ0 +
dgn . (14)
πgnγ+2 /(aλQ(gn ))
0
The MMSE denoiser is given by the conditional expectation, i.e., ηt (x̃tn ) = E[X|X̃ t = x̃tn ] where random variable
X̃ t = X + τt V , and x̃tn is a realization of X̃ t . Note that the
denoiser ηt (x̃tn ) depends on t through τt . The expression of the
conditional expectation is given in the following proposition.
Proposition 3. Based on the pdf of xn in (14) and the signalplus-noise model X̃ t = X + τt V at each iteration in AMP,
the conditional expectation of X given X̃ t = x̃tn is given by
E[X|X̃ t = x̃tn ] = x̃tn
ν1 (|x̃tn |2 )
,
ξ1 (|x̃tn |2 )
where functions νi (s) and ξi (s) are defined as
Z ∞ 2−γ
−s
gn Q(gn )
exp
dgn ,
νi (s) ,
(gn2 + τt2 )i+1
gn2 + τt2
0
1−λ
−s
ξi (s) ,
exp
τ2
λaτt2i
Z ∞ −γ t
gn Q(gn )
−s
+
exp
dgn .
(gn2 + τt2 )i
gn2 + τt2
0
(15)
(16)
(17)
6
1.5
hypothesis testing problem
(
H0 : X = 0, inactive user,
10 -5
H1 : X 6= 0, active user;
1
the optimal decision rule is given by
0.5
0
LLR = log
-0.5
MMSE denoiser
t
=10-6
soft thresholding =10-6
=2 10
-6
soft thresholding =2 10
-6
MMSE denoiser
-1
-1.5
-1.5
-1
-0.5
0
0.5
t
1
1.5
10 -5
Fig. 1. MMSE denoiser vs. soft thresholding denoiser [14] ηtsoft (x̃tn ) ,
(x̃tn −
(21)
θ x̃tn
)I(|x̃tn |
|x̃tn |
> θ), where I(·) is the indicator function.
Proof. See Appendix C.
Note that to implement ηt (x̃tn ) at each iteration, the value of
τt is needed. In practice, an empirical estimate τ̂t = √1L kzt k2 ,
where k · k2 denotes the `2 norm, can be used [22]. Although
ηt (x̃tn ) is in a complicated form, we note that it can be precomputed and stored as table lookup, so it does not add to runtime complexity. To gain some intuition, we illustrate the shape
of the MMSE denoiser as compared to the widely used soft
thresholding denoiser in Fig. 1. We observe that the MMSE
denoiser plays a role similar to the soft thresholding denoiser,
shrinking the input towards the origin, especially when the
input is small, thereby promoting sparsity.
2) With Exact Knowledge of Large-Scale Fading: When gn
is available at the BS, we substitute (13) into (4), and the pdf
of the entries of x is simplified to Bernoulli-Gaussian as
λ
−|xn |2
pX|G (xn |gn ) = (1 − λ)δ0 + 2 exp
. (18)
πgn
gn2
The MMSE denoiser is given by ηt (x̃tn , gn ) = E[X|X̃ t =
x̃tn , G = gn ], where the conditional expectation is [17]
E[X|X̃ t = x̃tn , G = gn ] =
1+
gn2 (gn2 + τt2 )−1 x̃tn
,
2 +τ 2
1−λ gn
t
exp (−∆|x̃tn |2 )
λ
τt2
(19)
where
∆ , τt−2 − (gn2 + τt2 )−1 .
(20)
Compared with the MMSE denoiser in (15), we add gn to the
left hand side of (19) to emphasize the dependency on prior
information gn .
B. User Activity Detection
After the AMP algorithm has converged, we employ the
likelihood ratio test to perform user activity detection. For the
pX̃ t |X (x̃tn |X 6= 0)
pX̃ t |X (x̃tn |X
= 0)
!
H0
≶ ln ,
(22)
H1
where LLR denotes the log-likelihood ratio, and ln denotes
the decision threshold typically determined by a cost function.
The performance metrics of interest are the probability of
missed detection PM , defined as the probability that a device
is active but the detector declare the null hypothesis H0 , and
the probability of false alarm, PF , defined as the probability
that a device is inactive, but the detector declare it to be active.
We consider the threshold for two cases depending on whether
the large-scale fading coefficient gn is available at the BS or
not.
1) With Statistical Knowledge of Large-Scale Fading Only:
We first derive the likelihood probabilities in the following.
Proposition 4. Suppose that X follows (14), and V follows
complex Gaussian distribution with zero mean and unit variance, the likelihood of X̃ t = X + τt V given X = 0 or X 6= 0
is given by
−|x̃tn |2
1
exp
,
(23)
pX̃ t |X (x̃tn |X = 0) =
πτt2
τt2
Z ∞ −γ
agn Q(gn )
−|x̃tn |2
pX̃ t |X (x̃tn |X 6= 0) =
exp
dgn .
π(gn2 + τt2 )
gn2 + τt2
0
(24)
Proof. See Appendix D.
Based on (23) and (24), the log-likelihood ratio is given as
Z ∞ 2 −γ
aτt gn
LLR = log
Q(gn ) exp(|x̃tn |2 ∆)dgn ,
(25)
gn2 + τt2
0
where ∆ is defined in (20). By observing that LLR is
monotonic in |x̃tn |, we can simplify the decision rule in (22)
H0
as |x̃tn | ≶ ln , indicating that user activity detection can be
H1
performed based on the magnitude of x̃tn only.
Based on the likelihood probabilities and the threshold ln ,
the probabilities of false alarm and missed detection can be
characterized as follows
2
Z
−ln
t
t
PF =
pX̃ t |X (x̃n |X = 0)dx̃n = exp
, (26)
τt2
|x̃tn |>ln
Z
PM =
pX̃ t |X (x̃tn |X 6= 0)dx̃tn ,
(27)
|x̃tn |<ln
where (26) is simplified by using (23). Note that since only
statistical information of the large-scale fading is known at
the BS, PF and PM are the averaged false alarm and missed
detection probabilities which do not depend on gn .
7
2) With Exact Knowledge of Large-Scale Fading: When
gn is known at the BS, the distribution of X is simplified to
Bernoulli-Gaussian. The likelihood probabilities become
exp −|x̃n |2 τt−2
t
,
(28)
pX̃ t |X,G (x̃n |X = 0, G = gn ) =
πτt2
exp −|x̃n |2 (gn2 + τt2 )−1
t
.
pX̃ t |X,G (x̃n |X 6= 0, G = gn ) =
π(τt2 + gn2 )
(29)
The log-likelihood ratio is then given as
τt2
t 2
LLR(gn ) = log
exp(|x̃
|
∆)
,
n
gn2 + τt2
(30)
where the notation LLR(gn ) emphasizes the dependency on
the prior information gn . Similar to the case where only the
statistics of gn is known, LLR here is also monotonic in |x̃tn |,
which means that the user activity detection can be performed
based on |x̃tn | only.
We also use ln to denote the threshold in the detection.
Based on (28) and (29), the probabilities of false alarm and
missed detection probability are given as follows
Z
PF (gn ) =
pX̃ t |X,G (x̃tn |X = 0, G = gn )dx̃tn
|x̃tn |>ln
= exp −ln2 τt−2 ,
(31)
Z
pX̃ t |X,G (x̃tn |X 6= 0, G = gn )dx̃tn
PM (gn ) =
|x̃tn |<ln
= 1 − exp −ln2 (gn2 + τt2 )−1 ,
(32)
where we use the notation PF (gn ) and PM (gn ) to indicate
the prior known gn . Note that the false alarm probability in
(31) has the form as that in (26) even through the value of τt
may be different due to different denoisers.
A natural question then arises: how to design the threshold
ln as a function of the known large-scale fading gn ? In theory,
we can treat each user separately, i.e., set the thresholding
value of each user separately according to its own cost function. For example, if a specific target false alarm probability is
needed for user n, we can design its thresholding parameter,
ln , using the expression in (31). In order to bring fairness,
this paper considers a common target false alarm probability
for all users. Under this condition, all users share the same
thresholding parameter, i.e., ln = l, ∀n, since the expression of
PF (gn ) in (31) does not depend on gn . In such a case, different
users may have different probabilities of missed detection
depending on their large-scale fading gn . To measure the
performance of the detector for the entire system, we employ
the average probability of missed detection as
N
1 X
−l2
PM =
1 − exp
N n=1
τt2 + gn2
Z
−l2
→ pG (g) 1 − exp
dg, as N → ∞,
τt2 + g 2
(33)
where the distribution PG (g) is given in (11). When N is large,
once τt is given, the averaged performance only depends on
the statistics of the large-scale fading gn .
C. State Evolution Analysis
We have characterized the probabilities of false alarm
PF and missed detection PM for user activity detection
in (26), (27) and (31), (32), but the parameter τt that
represents the standard deviation of the residual noise still
needs to be determined. As AMP proceeds, τt converges to
τ∞
. To compute τt, we use the state evolution (7), where
E |ηt (X̃ t , G) − X|2 in (7) can be interpreted as the MSE of
the denoiser. Note that for the MMSEdenoiser,
MSE can also
be expressed as E |ηt (X̃ t , G) − X|2 = E Var(X|X̃ t , G) ,
where Var(X|X̃ t , G) is the conditional variance of X given
X̃ t and G, and the expectation is taken over both X̃ t and G.
(Note that we drop G if the large-scale fading coefficient is
unknown.) By using conditional variance, we characterize the
MSE of the designed denoisers in the following propositions.
Proposition 5. The MSE of the denoiser for the case where
only the statistics of gn is known to the BS is given by
Z ∞
aQ(gn ) λgn2 τt2
dgn
· 2
MSE(τt ) =
gnγ
gn + τt2
0
Z ∞
+
aλs µ1 (s) − ν12 (s)ξ1−1 (s) ds, (34)
0
where functions νi (s) and ξi (s) are defined in (16) and (17),
respectively, and function µi (s) is defined as
Z ∞ 4−γ
gn Q(gn )
−s
µi (s) ,
exp
dgn .
(35)
(gn2 + τt2 )i+2
gn2 + τt2
0
Proof. See Appendix E.
It is worth noting that λgn2 τt2 (gn2 + τt2 )−1 in the first term
of the right hand side of (34) corresponds to the MSE of the
estimate of xn if the large-scale fading coefficient gn as well
as the user activity is assumed to be a priori known, and the
integral of gn corresponds to the averaging over all possible
gn . The second term then represents the cost of unknown gn
and unknown user activity in reality. Similarly, for the case
where gn is exactly known, the MSE can be characterized as
follows.
Proposition 6. The MSE of the denoiser for the case where
gn is known exactly at the BS is
Z ∞
aQ(gn ) λgn2 τt2
MSE(τt ) =
· 2
dgn
gnγ
gn + τt2
0
Z ∞
aλQ(gn )gn4
+
1 − ϕ1 (gn2 τt−2 ) dgn ,
γ 2
2
gn (gn + τt )
0
(36)
where function ϕi (s) of s is defined as
Z ∞
ti exp(−t)
dt.
ϕi (s) ,
1 + (1 − λ)(1 + s)i exp(−st)/λ
0
(37)
Proof. See Appendix F.
We also observe from (36) that the first term in the right
hand side corresponds to the averaged MSE if the user activity
is assumed to be known, and the second term corresponds the
extra error brought by unknown user activity.
8
Based on the expressions of MSE in (34) and (36), the state
evolution in (7) can be expressed as
N
2
2
τt+1
= σw
+ MSE(τt ),
(38)
L
based on which PF and PM can be evaluated according to
(26), (27), and (31), (32), as functions of the iteration number.
As the AMP algorithm converges, τt converges to the fixed
point τ∞ of the above equation.
Now we compare the resulting MSEs in these two cases.
According to the decomposition of variance, we have
E Var X|X̃ t = E Var X|X̃ t , G
+ E Var E[X|X̃ t , G] X̃ t
≥ E Var X|X̃ t , G ,
(39)
which indicates that knowing the large-scale fading can help
to improve the estimation on X given X̃ t . However, the
simulation results in Section VI show that surprisingly for the
model of the large-scale fading considered in this paper, the
performance improvement is actually minor, indicating that
knowing the large-scale fading does not help to get a much
better estimation. Knowing the exact value of gn is not crucial
in user activity detection and the statistical information of gn
is sufficient for device detection.
V. U SER ACTIVITY D ETECTION : M ULTIPLE -A NTENNA
C ASE
This section designs the AMP algorithms that account for
wireless channel propagation for the massive connectivity
problem in the multiple-antenna case. As mentioned earlier,
two different AMP algorithms can be used for the MMV
problem: the AMP with a vector denoiser operating on each
row of the input matrix, or the parallel AMP-MMV that
divides the MMV problem into parallel SMV problems and
iteratively solves the SMV problem on each antenna separately
with soft information exchange between the antennas. The
AMP with vector denoiser admits a state evolution, which
allows an easier characterization of its performance, whereas
AMP-MMV can be implemented in a distributed way which
is helpful for reducing the running time of the algorithm,
especially when the BS is equipped with large antenna arrays.
A. User Activity Detection by AMP with Vector Denoiser
As in the scenario with single antenna, we consider both
the cases where only the statistical knowledge or the exactly
knowledge of the large-scale fading is known at the BS. To
design the denoisers, we first characterize the pdfs of the row
vectors of X in the following.
Proposition 7. Denote rn as the row vector of X. If only
pG (gn ) is known at the BS, the pdf of rn is given by
Z ∞
exp −krn k22 gn−2
dgn .
pR (rn ) = (1 − λ)δ0 +
π M gnγ+2M /(aλQ(gn ))
0
(40)
If gn is known, the pdf of rn is Bernoulli-Gaussian as
pR|G (rn |gn ) = (1 − λ)δ0 +
λ exp(−krn k22 gn−2 )
.
(πgn2 )M
(41)
Proof. The results are extensions of (14) and (18) by considering multivariate random variables.
Given R̃t = R + Utn with Ut following complex Gaussian
distribution with zero mean and covariance Σt , the MMSE
denoisers ηt (r̃tn ) and ηt (r̃tn , gn ) for both cases are given by
the conditional expectation in the following.
Proposition 8. If only pG (gn ) is known at the BS, the
conditional expectation of R given R̃t = r̃tn is
R∞
Q(gn )ψa (gn )(gn−2 Σt + I)−1 r̃tn dgn
t
t
R∞
,
E[R|R̃ = r̃n ] = 0
ψc (gn ) + 0 Q(gn )ψb (gn )dgn
(42)
where ψa (gn ), ψb (gn ) and ψc (gn ) are defined as follows
t ∗
−2 2 −1
exp −r̃tn Σ−1
(r̃n )
t − (Σt + gn Σt )
ψa (gn ) ,
,
γ
2
gn |Σt + gn I|
(43)
t ∗
t
−2
2
4 −1 −1
exp −r̃n gn I − (gn I + gn Σt )
(r̃n )
,
ψb (gn ) ,
γ
2
gn |Σt + gn I|
(44)
−1 t ∗
−1
t
−1
ψc (gn ) ,(1 − λ)(aλ) exp −r̃n Σt (r̃n ) |Σt | . (45)
If gn is known at the BS, the conditional expectation is
E[R|R̃t = r̃tn , G = gn ] =
(gn−2 Σt + I)−1 r̃tn
, (46)
2
1 + 1−λ
λ |gn I + Σt |ψd (gn )
where ψd (gn ) is defined as follows
t ∗
2 −1
ψd (gn ) , exp −r̃tn Σ−1
(r̃n ) |Σt |−1 .
t − (Σt + gn I)
(47)
Proof. See Appendix G.
The covariance matrix Σt in both (42) and (46) is tracked
via the state evolution (10), and Σt can be further simplified
by the following proposition.
Proposition 9. Based on the pdfs in (40) and (41) and the state
evolution (10), if the initial covariance matrix Σ0 is a diagonal
matrix with identical diagonal entries, i.e., Σ0 = τ02 I, then Σt
stays as a diagonal matrix with identical diagonal entries, i.e.,
Σt = τt2 I, for t ≥ 1, where τt is determined by
N
2
2
τt+1
= σw
+ MSE(τt ).
(48)
L
If only pG (gn ) is known at the BS, MSE(τt ) is given by
Z ∞
aλgn2 τt2 Q(gn )
MSE(τt ) =
dgn
gnγ (gn2 + τt2 )
0
Z ∞
−1
2
µM (s) − νM
(s)ξM
(s)
ds,
(49)
+
M
Γ(M + 1)/(λas )
0
where functions µi (s) , νi (s) and ξi (s) are defined in (35),
(16), and (17), respectively, and Γ(·) is the Gamma function.
If the exact large-scale fading gn is known at the BS, MSE(τt )
is given by
Z ∞
aλgn2 τt2 Q(gn )
MSE(τt ) =
dgn
gnγ (gn2 + τt2 )
0
Z ∞
aλQ(gn )gn4
ϕM (gn2 τt−2 )
+
1−
dgn ,
gnγ (gn2 + τt2 )
Γ(M + 1)
0
(50)
9
where function ϕi (s) is defined in (37).
Proof. See Appendix H.
Note that Σ0 is the noise covariance matrix after the first
matched filtering, which is indeed a diagonal matrix with
identical diagonal entries. Based on Proposition 9, the MMSE
denoiser in (42) can be further simplified as
t
E[R|R̃ =
r̃tn ]
=
νM (kr̃tn k22 )
,
r̃tn
ξM (kr̃tn k22 )
and
(51)
where νi (s) and ξi (s) are defined in (16) and (17), respectively, and the MMSE denoiser in (46) can be simplified as
E[R|R̃t = r̃tn , G = gn ]
gn2 (gn2 + τt2 )−1 r̃tn
=
1+
2
2 +τ
1−λ gn
t M
λ ( τt2 )
exp(−∆kr̃tn k22 )
large-scale fading is unknown at the BS, the probabilities of
false alarm or missed detection are then given as, respectively
Z
exp −kr̃tn k22 τt−2
dr̃tn
PF =
π M τt2M
kr̃tn k2 >ln
1
(a)
= 1−
γ̄ M, ln2 τt−2 ,
(59)
Γ(M )
, (52)
where ∆ is defined in (20). Note that if we let M = 1, (51) and
(52) reduce to the denoisers for the single-antenna case in (15)
and (19). As mentioned before, we can also pre-compute and
store the functions νM (·) and ξM (·) in (51) as table lookup.
After the AMP algorithm has converged, we use the likelihood ratio test to perform the user activity detection. Recall
that R̃t = R + Ut where Ut follows complex Gaussian
distribution. For the case where the large-scale fading is unknown, based on (40) and pR̃t derived in (82) in Appendix G,
the likelihood probabilities given that the user is inactive and
active are, respectively
exp −kr̃tn k22 τt−2
pR̃t |R (r̃tn |R = 0) =
,
(53)
π M τt2M
Z ∞
agn−γ Q(gn )
pR̃t |R (r̃tn |R 6= 0) =
M
π (gn2 + τt2 )M
0
−kr̃tn k22
exp
dgn .
(54)
gn2 + τt2
For the case where the large-scale fading coefficient is known,
noting that R follows a Beroulli-Gaussian distribution, and
R̃t follows a mixed Gaussian distribution, then the likelihood
probabilities can be computed as, respectively
exp −kr̃tn k22 τt−2
t
,
(55)
pR̃t |R,G (r̃n |R = 0, G = gn ) =
π M τt2M
exp −kr̃tn k22 (τt2 + gn2 )−1
pR̃t |R,G (r̃n |R 6= 0, G = gn ) =
.
π M (τt2 + gn2 )M
(56)
For both cases, we immediately obtain the LLRs as, respectively
Z ∞ 2M −γ
aτt gn Q(gn )
exp kr̃tn k22 ∆ dgn , (57)
LLR = log
2
2
M
(gn + τt )
0
τt2M
LLR(gn ) = log
exp kr̃tn k22 ∆ .
(58)
2
2
M
(τt + gn )
Observing that LLRs are monotonic in kr̃tn k2 , we can set a
threshold ln on kr̃tn k2 to perform the detection. When the
∞
agn−γ Q(gn )
π M (gn2 + τt2 )M
kr̃tn k2 <ln 0
−kr̃tn k22
exp
dgn dr̃tn
gn2 + τt2
Z ∞ −γ
agn Q(gn )
(b)
=
γ̄ M, ln2 (gn2 + τt2 )−1 dgn , (60)
Γ(M
)
0
Z
Z
PM =
where γ̄(·, ·) is the lower incomplete Gamma function, and (a)
and (b) are simply obtained by noticing that the integral of r̃tn
can be interpreted as the cumulative distribution function (cdf)
of a χ2 distribution with 2M degrees of freedom since kr̃tn k22
can be regarded as a sum of the squares of 2M identical real
Gaussian random variables. Using the same approach, when
the large scale fading is known to the BS, the probabilities of
false alarm and missed detection can be evaluated as
Z
exp −kr̃tn k22 τt−2
dr̃tn ,
PF (gn ) =
π M τt2M
kr̃tn k2 >ln
1
γ̄ M, ln2 τt−2 ,
(61)
=1−
Γ(M )
Z
exp −kr̃tn k22 (τt2 + gn2 )−1
PM (gn ) =
dr̃tn
π M (τt2 + gn2 )M
kr̃tn k2 <ln
1
γ̄ M, ln2 (gn2 + τt2 )−1 .
(62)
=
Γ(M )
It is easy to verify that when M = 1, (59), (60), (61) and (62)
reduce to (26), (27), (31) and (32), respectively.
Based on (59), (60), (61) and (62), we can design the
threshold ln to achieve a trade-off between the probability of
false alarm and probability of missed detection. The proposed
thresholding strategy in the single-antenna case can still be
used in multiple-antenna scenario.
B. User Activity Detection by Parallel AMP-MMV
The outline of the parallel AMP-MMV algorithm is as
presented in Algorithm 1. In this section, we adopt the
parallel AMP-MMV algorithm for our problem setup; we
present the expression of the denoiser, η(·, g, t, i, m) ,
[ηt,i,m (·, g1 ), · · · , ηt,i,m (·, gN )]T and the probability of a device being active based on the decision at the mth AMP-SMV
*
stage, π nm . Here we only discuss the case where the largescale fading is known. The extension to the scenario where
the large-scale fading is unknown is similar.
Since the parallel AMP-MMV algorithm employs M parallel AMP-SMVs, the expression of the scalar denoiser for
each AMP-SMV is in the form of the MMSE denoiser for the
single-antenna case in (19). However, instead of using the prior
λ as the probability of being active for each user, the algorithm
has access to a better estimate for the probability of activities
10
(
p(x̃T,i
nm |X = 0, G = gn ) =
p(x̃T,i
nm |X 6= 0, G = gn ) =
2 −2
exp(−|x̃T,i
nm | τT,i )
2
πτT,i
,
10 0
10 -1
10 -2
PM
as π nm as algorithm proceeds. Therefore, the expression for
the MMSE denoiser can be written as
2 −1 t,i
) x̃nm
gn2 (gn2 + τt,i
2 t,i 2 ,
ηt,i,m (x̃t,i
,
g
)
=
(
nm n
2 +τ 2
g
−gn |x̃nm |
π nm n t,i
exp
1 + 1−
(
2
2 (g 2 +τ 2 )
τt,i
τt,i
π nm
n
t,i
(63)
where x̃t,i
and
x
are
the
elements
in
the
nth
row
and
the
nm
nm
mth column of X̃t,i and X, respectively. At the end of the ith
outer iteration, i.e., t = T , the likelihood probabilities given
that the user is inactive or active can be written as
10 -3
(64)
10 -4
2 2
2 −1
exp(−|x̃T,i
)
nm | (gn + τT,i )
.
2 + g2 )
π(τT,i
n
(65)
Further, using equations (64) and (65), the probability that
user n is active based on the decision at the mth AMP-SMV
can be calculated as
p(x̃T,i
*
nm |X 6= 0, G = gn )
π nm =
T,i
p(x̃nm |X 6= 0, G = gn ) + p(x̃T,i
nm |X = 0, G = gn )
!!−1
2
2
+ gn2
τT,i
−gn2 |x̃T,i
nm |
= 1+
exp
. (66)
2
2 (g 2 + τ 2 )
τT,i
τT,i
n
T,i
After the parallel AMP-MMV is terminated, we use likelihood ratio test to perform the user activity detection. It can be
shown that the LLR for user n can be calculated as
!!
P
2M
2
τT,I
−gn2 m |x̃T,I
nm |
.
LLR(gn ) = log
2 + g 2 )M exp
2 + g 2 )τ 2
(τT,I
(τT,I
n
n T,I
(67)
It can be seen that the LLR expression for AMP-MMV
algorithm is in a similar form as in (57). Therefore, with
the same discussion, we can show that the probabilities of
false alarm and missed detection can be further simplified in a
form similar to (61) and (62), respectively. To have complete
performance prediction analysis, we also need to determine
2
τT,I
in the parallel AMP-MMV algorithm. However, due to
the soft information exchange between the antennas, deriving
2
an analytic state evolution for τt,i
is very challenging. The numerical experiments in Section VI show that the performance
of parallel AMP-MMV is very similar to AMP with vector
2
denoiser. This observation suggests that the parameter τT,I
for the AMP-MMV algorithm should be similar to the final
value of τt2 in AMP with vector denoiser.
We briefly discuss the complexities of AMP with vector denoiser and AMP-MMV. For both algorithms the computational
complexities mainly lie in the matched filtering and residual
calculation, which depend on the problem size as O(N LM ),
at each iteration. The advantage of AMP-MMV is that parallel
computation is allowed due to the division of MMV problem
into several SMV problems.
VI. S IMULATION R ESULTS
We evaluate the performance of the proposed method in
a cell of radius R = 1000m with potential N = 4000 users
among which 200 are active, i.e., λ = 0.05. The channel fading
10 -5
AMP simulated, 5dBm
AMP predicted, 5dBm
lower bound, 5dBm
AMP simulated, 15dBm
AMP predicted, 15dBm
lower bound, 15dBm
AMP simulated, 25dBm
AMP predicted, 25dBm
lower bound, 25dBm
10 -4
10 -3
10 -2
10 -1
10 0
PF
Fig. 2. Performance of AMP based user activity detection with only statistical
knowledge of the large-scale fading.
parameters are α = 15.3, β = 37.6 and σSF = 8, and the
background noise is −169dBm/Hz over 10MHz.
We first consider the single-antenna case with only statistical knowledge of large-scale fading. Fig. 2 shows the tradeoff
between the probabilities of missed detection and the false
alarm of AMP with MMSE denoiser when the pilot sequence
length is set as L = 800 and the transmit power is set as
5dBm, 15dBm, and 25dBm. We see that the predicted PM
and PF match the analysis very well. We also plot a lower
bound using τ∞ = σw . The lower bounds are very close to
the actual performance, indicating that after convergence AMP
is able to almost completely eliminate multiuser interference;
the remaining error is dominated by the background noise.
Fig. 3 shows the performance of the AMP algorithm with
MMSE denoiser when the exact large-scale fading coefficients are known. For comparison, the performance with
only statistical knowledge of the large-scale fading is also
demonstrated (only simulated performance is included since
the predicted performance is almost the same as depicted in
Fig. 2). Fig. 3 shows that the predicted curves match the
simulated curves very well. More interestingly, it indicates
that the performance improvement for knowing the exact
large-scale fading coefficients is negligible which suggests
that knowing the distribution of the large-scale fading (rather
than the exact value) is already enough for good user activity
detection performance.
The next simulation compares the AMP algorithm with
MMSE denoiser with two other algorithms widely used in
compressed sensing: CoSaMP [23], and AMP but with soft
thresholding denoiser [3]. Compare to AMP, CoSaMP is based
on the matching pursuit technique. Compare to AMP with
MMSE denoiser, AMP with soft thresholding does not exploit
the statistical knowledge of xn . Fig. 4 shows that AMP with
MMSE denoiser outperforms both CoSaMP and AMP with
soft thresholding denoiser. This is partly due to the fact that
both CoSaMP and AMP with soft thresholding denoiser do
11
10 0
10 0
10 -1
10 -1
PM = PF
PM
10 -2
AMP simulated w/o LS, 5dBm
AMP simulated with LS, 5dBm
10 -3
10 -2
AMP predicted with LS, 5dBm
AMP simulated w/o LS, 15dBm
AMP simulated with LS, 15dBm
10
-4
MMSE denoiser, L = 300
MMSE denoiser, L = 400
MMSE denoiser, L = 800
soft thresholding denoiser, L = 600
soft thresholding denoiser, L = 800
10 -3
AMP predicted with LS, 15dBm
AMP simulated w/o LS, 25dBm
AMP simulated with LS, 25dBm
AMP predicted with LS, 25dBm
10 -4
10 -3
10 -2
10 -1
10 0
PF
10 0
PM
10 -1
soft thresholding denoiser, 5dBm
CoSaMP, 5dBm
MMSE denoiser, 5dBm
soft thresholding denoiser, 15dBm
10
-3
CoSaMP, 15dBm
MMSE denoiser, 15dBm
soft thresholding denoiser, 25dBm
CoSaMP, 25dBm
MMSE denoiser, 25dBm
10 -4
10 -4
10 -3
10 -2
5
10
15
20
25
30
35
40
45
Tx power (dBm)
Fig. 3. Performance of AMP based user activity detection with knowledge
of the large-scale fading.
10 -2
10 -4
10 -1
10 0
PF
Fig. 4. Performance comparison of AMP with MMSE denoiser, AMP with
soft threshoding denoiser, and CoSaMP.
not exploit the the statistical knowledge of xn . Note that
AMP with soft thresholding implicitly solves the LASSO
problem [14], [24], i.e., the sparse signal recovery problem
as an `1 -penalized least squares optimization. Therefore, the
results in Fig. 4 indicate that AMP with MMSE denoiser also
outperforms LASSO.
Fig. 5 compares the performance of the AMP algorithm
with MMSE denoiser and the AMP algorithm with soft
thresholding denoiser as function of transmit power and pilot
length. For convenience, we set PF = PM by properly
choosing the threshold l. We observe first that the MMSE
denoiser outperforms soft thresholding denoiser significantly,
but more importantly, we observe that the minimum L needed
to drive PF and PM to zero as transmit power increases is
between 300 and 400 for the MMSE denoiser, whereas the
Fig. 5. Impact of transmit power and length of pilot on user activity detection
performance: MMSE denoiser vs. soft threholding denoiser.
minimum L is between 600 and 800 for the soft threshloding
denoiser, indicating the clear advantage of accounting for
channel statistics in user activity detector design.
Finally, we consider the multiple-antenna case assuming the
knowledge of large-scale fading coefficients. Fig. 6 illustrates
the probabilities of false alarm and missed detection under
different numbers of antennas for both AMP with vector
denoiser and parallel AMP-MMV algorithms. For comparison,
the single-antenna M = 1 case is also included. Fig. 6 shows
that for AMP with vector denoiser, the simulated results match
the predicted results very well. Further, it shows that the
performances of AMP with vector denoiser and parallel AMPMMV are approximately the same, indicating that although
these two algorithms employ different strategies, they both
exploit the statistical knowledge of the channel in the same
way, resulting in similar performances.
The impact of the pilot length L and the number of antennas
M on the probabilities of false alarm and missed detection as
the transmit power increases is shown in Fig. 7. We set L =
300, 600, and M = 1, 2, 4. We make PF = PM for convenient
comparison by properly choosing the threshold l. Note that in
the scenario where the exact large-scale fading coefficients
are known, different users have the same probabilities of false
alarm but different probabilities of missed detection. Thus we
have to use the average probability of missed detection over all
users. Fig. 7 shows that increasing L or M brings significant
improvement. Specifically, when L = 300, M = 1, PF and
PM tend to remain unchanged as the transmit power increases.
However, by either increasing L or increasing M , PF and PM
can be driven to zero as the transmit power increases. In other
words, the minimum L required to drive PF and PM to zeros
can be reduced by increasing M .
VII. C ONCLUSION
This work shows that compressed sensing is a viable
strategy for sporadic device activity detection for massive
12
for both the single-antenna case and the multiple-antenna
case. Simulation results validate the analysis, and show that
exploiting the statistics of the channel in AMP denoiser
design can significantly improve the detection threshold, and
further deploying multiple antennas at the BS can also bring
significant performance improvement.
10 0
10 -1
Tx power = 5dB
PM
10 -2
10 -3
10 -4
10 -5
A PPENDIX
AMP simulated, M=1
AMP predicted, M=1
bound, M=1
vAMP simulated, M=2
AMP-MMV simulated, M=2
vAMP predicted, M=2
bound, M=2
vAMP simulated, M=4
AMP-MMV simulated, M=4
vAMP predicted, M=4
bound, M=4
10 -4
A. Proof of Proposition 1
Tx power = 25dB
10 -3
10 -2
10 -1
10 0
PF
Fig. 6. Performance of AMP with vector denoiser (vAMP) and AMP-MMV
for user activity detection in the multiple-antenna case.
Denote d, x, y, z as the distance from a user to the BS,
the shadowing in dB, the shadowing in linear scale, and the
path-loss, respectively. Note that in appendices we slightly
abuse some notations appeared in the paper due to the limited
alphabet. However this should not cause confusion since they
only used for derivations in the appendices. Denote g , yz as
the large-scale fading. We derive the pdfs of d, x, y, z and g
as follows.
Assuming that all users are uniformly distributed in the cell
with radius R, the pdf of d is
pD (d) =
10 0
10
PF=PM
0 < d < R.
(68)
Since x = −20 log10 (y) follows Gaussian distribution with
2
, the pdf of y can be derived as
zero mean and variance σSF
!
20
200 ln2 (y)
√
pY (y) =
exp − 2
.
(69)
2
ln (10)σSF
ln(10) 2πσSF y
-1
10 -2
By using z = 10−(α+β log10 d)/20 , we get the pdf of z as
10 -3
10
2d
,
R2
pZ (z) =
vAMP, simulated, L=300, M=1
-4
40
10−2α/β z −40/β−1 ,
R2 β
(70)
vAMP, simulated, L=300, M=2
where z > 10−(α+β log10 R)/20 . The pdf of g is
Z
pG (g) = |y|−1 pY (y)pZ y −1 g dy , ag −γ Q(g),
vAMP, simulated, L=300, M=4
10
-5
10
-6
vAMP, simulated, L=600, M=1
vAMP, simulated, L=600, M=2
vAMP, simulated, L=600, M=4
5
10
15
20
25
30
35
40
45
Tx power (dBm)
(71)
R∞
where Q(g) , b ln g+c exp(−s2 )ds, γ , 40/β + 1, and a, b
and c are constants given in the proposition.
Fig. 7. Impact of length of pilot, number of antennas, and transmit power.
B. Proof of Proposition 2
connectivity applications with random non-orthogonal signature sequences. Specifically, we propose an AMP-based user
activity detection algorithm by exploiting the statistics of the
wireless channel for the uplink of a cellular system with a
large number of potential users but only a small fraction of
them are active at any time slot. We show that by using the
state evolution, a performance characterization in terms of
the probabilities of false alarm and missed detection can be
accurately predicted. In particular, we consider both cases in
which the BS is equipped with a single antenna or multiple
antennas. We present the designs of the MMSE denoisers in
the scenarios where the large-scale fading is either available
exactly or when only its statistics is available at the BS. For
the multiple-antenna case, we adopt two AMP algorithms,
AMP with vector denoiser and parallel AMP-MMV, to tackle
the detection problem. We derive a performance analysis
Denote g as the large-scale fading following (11), and x =
xR +jxI as the Rayleigh fading, where xR and xI are the real
and imaginary parts of x, respectively. Denote h = hR +jhI as
the channel coefficient accounting for both large-scale fading
g and Rayleigh fading x, i.e., h = gx. Then the pdf of h is
Z
pH (h) = pHR ,HI ,G (hR , hI , g)dg
Z
∂(xR , xI , g)
hR hI
,
dg
= pG (g)pXR ,XI
g g
∂(hR , hI , g)
Z ∞
a
−|h|2
=
g −γ−2 Q(g) exp
dg,
(72)
π 0
g2
where pHR ,HI ,G (hR , hI , g) and pXR ,XI (xR , xI ) are the joint
∂(xR ,xI ,g)
pdfs, and | ∂(h
| = g −2 is the Jacobian determinant.
R ,hI ,g)
When g is known, pH|G (h|g) follows the complex Gaussian
distribution with zero mean and variance g 2 .
13
C. Proof of Proposition 3
We omit superscript t and subscript n for notation simplicity. The conditional expectation of X given X̃ = x̃ can be
expressed as
E[X|X̃ = x̃]
Z
pX (x)
= x
p
(x̃|x)dx
pX̃ (x̃) X̃|X
ZZ ∞
−|x̃ − x|2
−|x|2
+
dgdx
=
f (g)x exp
g2
τ2
0
ZZ ∞
−|x̃|2
−|x − δ x̃|2
f (g)x exp
=
+
dgdx
g2 + τ 2
δτ 2
0
Z ∞
−|x̃|2
Q(g)g 2−γ
(a) λax̃
=
exp
dg,
(73)
pX̃ (x̃)π 0
g 2 + τ 2 (g 2 + τ 2 )2
where f (g) , λaQ(g)/(pX̃ (x̃)π 2 τ 2 g γ+2 ), δ , g 2 /(g 2 + τ 2 ),
(a) is obtained by using Gaussian integral of x. By substituting
the expression of pX̃ (x̃) derived in Appendix D, we get
E[X|X̃ = x̃] as in (15).
D. Proof of Proposition 4
We omit superscript t and subscript n in the following. Note
that pX̃|X (x̃|X = 0) = pW (x̃) where random variable W is
defined as W , τ V , and pX̃|X (x̃|X 6= 0) = pY (x̃) where
random variable Y is defined as Y , H +W with H following
the distribution in (12). Since V follows complex Gaussian
distribution with zero mean and unit variance, we get
1
−|x̃|2
pX̃|X (x̃|X = 0) =
exp
,
(74)
πτ 2
τ2
To compute pX̃|X (x̃|X 6= 0), we derive pY (y) as follows
(a)
Z
pH (y − w)pW (w)dw
Z ∞
ag −γ Q(g)
−|y|2
(b)
=
exp
dg,
π(g 2 + τ 2 )
g2 + τ 2
0
where f (g) is defined in Appendix C, (a) is obtained by using
2 4
Gaussian integral of x, fˆ(g) , τ 2 g 2 /(g 2 +τ 2 )2 +|x̃|
g /(g 2 +
R
2 3
τ ) . By using (77), (76), (73), and MSE(τ ) = Var(X|X̃ =
x̃)pX̃ (x̃)dx̃ with some algebraic manipulations, we obtain
(34).
F. Proof of Proposition 6
Similar to Appendix E, we first
derive the conditional
expectation E |X|2 |X̃ = x̃, G = g as
λfˆ(g) exp(−|x̃|2 /(g 2 + τ 2 ))
E |X|2 |X̃ = x̃, G = g =
,
pX̃|G (x̃|g)π
(79)
where fˆ(g) is defined in Appendix
E. Then based on (19)
and Var(X|X̃ = x̃, G = g) = E |X|2 X̃ = x̃, G = g −
2
E[X|X̃ = x̃, G = g] , we have
ZZ
MSE =
Var(X|X̃ = x̃, G = g)pX̃|G (x̃|g)pG (g)dx̃dg
!
Z ∞
λg 4 1 − ϕ1 (g 2 τ −2 )
λg 2 τ 2
(a)
pG (g)dg,
+
=
g2 + τ 2
g2 + τ 2
0
(80)
where (a) is obtained by using Gaussian integral over x̃. After
plugging pG (g), we finally get (36).
pY (y) =
(75)
where (a) is obtained by pY,W (y, w) = pH (y − w)pW (w),
and (b) is obtained by substituting (72). Combine the results
in (74) and (75) with pX̃|X (x̃|X 6= 0) = pY (x̃), we get
pX̃ (x̃) =
Since we have derived E[X|X̃ = x̃] in (73), we only need to
derive E |X|2 X̃ = x̃ , which can be expressed as
E |X|2 X̃ = x̃
Z
|x|2 pX (x)
=
p
(x̃|x)dx
p (x̃) X̃|X
Z Z ∞ X̃
−|x̃ − x|2
−|x|2
=
f (g)|x|2 exp
+
dgdx
g2
τ2
0
ZZ ∞
−|x − δ x̃|2
−|x̃|2
+
dgdx
=
f (g)|x|2 exp
g2 + τ 2
δτ 2
0
Z ∞
λaQ(g)fˆ(g)
−|x̃|2
(a)
dg,
(78)
=
exp
g γ pX̃ (x̃)π
g2 + τ 2
0
1−λ
−|x̃|2
exp
πτ 2
τ2
Z ∞
λaQ(g) exp −|x̃|2 /(g 2 + τ 2 )
dg. (76)
+
πg γ (g 2 + τ 2 )
0
E. Proof of Proposition 5
We omit superscript t and subscript n for notation simplicity. The conditional variance of X given X̃ = x̃ is
2
Var(X|X̃ = x̃) = E |X|2 X̃ = x̃ − E[X|X̃ = x̃] . (77)
G. Proof of Proposition 8
We omit superscript t and subscript n. For the case where
the large scale fading is unknown, the proof is similar to
Appendix C except that we need to deal with random vectors
rather than random scalars. By using
Z
pR (r)
E[R|R̃ = r̃] = r
p
(r̃|r)dr,
(81)
pR̃ (r̃) R̃|R
where pR̃ (r̃) can be derived as
1−λ
−kr̃k22
pR̃ (r̃) =
exp
(πτ 2 )M
τ2
Z ∞
λaQ(g) exp −kr̃k22 /(g 2 + τ 2 )
+
dg. (82)
π M g γ (g 2 + τ 2 )M
0
By plugging pR̃ (r̃) and pR (r) into (81), and using multivariate
Gaussian integral of r with some algebraic manipulations, we
can obtain (42). For the case where the large-scale fading is
known, the conditional expectation can be found in [18].
14
H. Proof of Proposition 9
The state evolution in (10) is then simplified to
Since the proofs for both cases are similar, in the following
we focus on the case where the large-scale fading is known.
We omit subscript n, and use t as subscript instead of
superscript for convenience. We use induction by assuming
Σt = τt2 I holds. To evaluate the right hand side of (10), we
first derive the distribution of R̃t based on (41) as
λ exp − kr̃t k22 (g 2 + τt2 )−1
φ(r̃t ), (83)
pR̃t |G (r̃t |g) =
π M (g 2 + τt2 )M
N
2
EG [CI] , τt+1
I,
L
which completes the induction.
where φ(r̃t ) , 1 + (1 − λ)(1 + g 2 τt−2 )M exp(−∆kr̃t k22 )/λ,
and ∆ is defined in (20). We then compute the conditional
covariance matrix of Rt given R̃t = r̃t and G = g as
Cov =E[RTt (RTt )∗ |R̃t = r̃t , G = g]
∗
− E[RTt |R̃t = r̃t , G = g] E[RTt |R̃t = r̃t , G = g]
φ−1 (r̃t ) − φ−2 (r̃t ) T T ∗
g 2 τt2 φ−1 (r̃t )
r̃t (r̃t ) . (84)
I+
=
2
2
g + τt
g −4 (g 2 + τt2 )2
Then by taking the expectation over R̃t , we obtain
Z
ER̃t |G [Cov] = pR̃t |G (r̃t |g)Covdr̃t ,
(85)
which is a diagonal matrix due to fact that the off-diagonal
element is an integral of an odd function over a symmetric
interval, which is zero. Furthermore, it is easy to observe
from the integral that the diagonal elements of ER̃t |G [Cov]
are identical. Note that when MMSE denoiser is employed,
∗
the right hand side
of (10) can be rewritten as E Dt Dt∗ =
EG ER̃t |G [Cov] , which leads to the result that E Dt Dt is
also a diagonal matrix with identical diagonal elements.
In the following, we derive an explicit expression of
E Dt D∗t . To this end, we first compute ER̃t |G [Cov]. Denote
ci as the ith diagonal entry of ER̃t |G [Cov], and r̃t,i is the ith
entry of r̃t . Based on (84) and (85), we have
Z
λ exp − kr̃t k22 /(g 2 + τt2 )
g 2 τt2
ci =
·
dr̃t
g 2 + τt2
π M (g 2 + τt2 )M
Z
|r̃t,i |2 1 − φ−1 (r̃t )
+
·
g −4 (g 2 + τt2 )2
λ exp − kr̃t k22 /(g 2 + τt2 )
dr̃t
π M (g 2 + τt2 )M
λg 2 τt2
λg 4
ϕM (g 2 τt−2 )
= 2
+ 2
1−
,
(86)
g + τt2
g + τt2
Γ(M + 1)
where the fist term of the last step is obtained by using
Gaussian integral, the second term is obtained by integrating in spherical coordinates instead of Cartesian coordinates,
and function ϕi (s) is defined in (37). As expected, ci does
not depend on i, indicating that the diagonal elements are
indeed
identical.
By replacing g by G and ci by C, we get
E Dt D∗t = EG [C]I, where C is a random variable depends
on G, and
Z ∞
aQ(g) λg 2 τt2
EG [C] =
· 2
dg
gγ
g + τt2
0
Z ∞
λg 4
ϕM (g 2 τt−2 )
aQ(g)
+
· 2
1−
dg. (87)
gγ
g + τt2
Γ(M + 1)
0
2
Σt+1 = σw
I+
(88)
R EFERENCES
[1] J. G. Andrews, S. Buzzi, W. Choi, S. V. Hanly, A. Lozano, A. C. Soong,
and J. C. Zhang, “What will 5G be?” IEEE J. Sel. Areas Commun.,
vol. 32, no. 6, pp. 1065–1082, June 2014.
[2] W. Yu, “On the fundamental limits of massive connectivity,” in Information Theory and Application (ITA) Workshop, San Diego, CA, USA,
Feb. 2017.
[3] D. Donoho, A. Maleki, and A. Montanari, “Message-passing algorithms
for compressed sensing,” Proc. Nat. Acad. Sci., vol. 106, no. 45, pp.
18 914–18 919, Nov. 2009.
[4] X. Chen, T.-Y. Chen, and D. Guo, “Capacity of Gaussian many-access
channels,” IEEE Trans. Inf. Theory, vol. 63, no. 6, pp. 3516–3539, June
2017.
[5] H. F. Schepker, C. Bockelmann, and A. Dekorsy, “Exploiting sparsity in
channel and data estimation for sporadic multi-user communication,” in
Inter. Symp. Wireless Commun. Sys. (ISWCS), Ilmenau, Germany, Aug.
2013, pp. 1–5.
[6] X. Xu, X. Rao, and V. K. N. Lau, “Active user detection and channel
estimation in uplink CRAN systems,” in IEEE Inter. Conf. Commun.
(ICC), London, UK, June 2015, pp. 2727–2732.
[7] G. Wunder, P. Jung, and M. Ramadan, “Compressive random access
using a common overloaded control channel,” in IEEE Globecom
Workshops, San Diego, CA, USA, Dec. 2015, pp. 1–6.
[8] G. Wunder, H. Boche, T. Strohmer, and P. Jung, “Sparse signal processing concepts for efficient 5G system design,” IEEE Access, vol. 3, pp.
195–208, Feb. 2015.
[9] H. Zhu and G. B. Giannakis, “Exploiting sparse user activity in multiuser
detection,” IEEE Trans. Commun., vol. 59, no. 2, pp. 454–465, Feb.
2011.
[10] H. F. Schepker and A. Dekorsy, “Compressive sensing multi-user detection with block-wise orthogonal least squares,” in IEEE Veh. Technol.
Conf. (VTC Spring), Yokohama, Japan, May 2012, pp. 1–5.
[11] J. Luo and D. Guo, “Neighbor discovery in wireless ad hoc networks
based on group testing,” in Allerton Conf. on Commun., Control, and
Computing, Urbana-Champaign, IL, USA, Sep. 2008, pp. 791–797.
[12] L. Zhang, J. Luo, and D. Guo, “Neighbor discovery for wireless
networks via compressed sensing,” Performance Evaluation, vol. 70,
no. 7, pp. 457–471, July 2013.
[13] M. Bayati and A. Montanari, “The dynamics of message passing on
dense graphs, with applications to compressed sensing,” IEEE Trans.
Inf. Theory, vol. 57, no. 2, pp. 764–785, Feb. 2011.
[14] A. Maleki, L. Anitori, Z. Yang, and R. G. Baraniuk, “Asymptotic
analysis of complex LASSO via complex approximate message passing
(CAMP),” IEEE Trans. Inf. Theory, vol. 59, no. 7, pp. 4290–4308, July
2013.
[15] S. Rangan, “Generalized approximate message passing for estimation
with random linear mixing,” in IEEE Inter. Symp. Inf. Theory (ISIT), St.
Petersburg, Russia, July 2011, pp. 2168–2172.
[16] D. L. Donoho, A. Maleki, and A. Montanari, “Message passing algorithms for compressed sensing: I. motivation and construction,” in IEEE
Inf. Theory Workshop (ITW), Cairo, Egypt, Jan. 2010, pp. 1–5.
[17] P. Schniter, “Turbo reconstruction of structured sparse signals,” in
Annual Conf. on Information Sciences and Systems (CISS), Princeton,
NJ, USA, Mar. 2010, pp. 1–6.
[18] J. Kim, W. Chang, B. Jung, D. Baron, and J. C. Ye, “Belief propagation for joint sparse recovery,” [Online] available:
http://arxiv.org/abs/1102.3289v1, 2011.
[19] J. Ziniel and P. Schniter, “Efficient high-dimensional inference in the
multiple measurement vector problem,” IEEE Trans. Signal Process.,
vol. 61, no. 2, pp. 340–354, Jan. 2013.
[20] G. Hannak, M. Mayer, A. Jung, G. Matz, and N. Goertz, “Joint channel
estimation and activity detection for multiuser communication systems,”
in IEEE Inter. Conf. Commun. (ICC) Workshop, London, UK, June 2015,
pp. 2086–2091.
[21] D. L. Donoho, I. Johnstone, and A. Montanari, “Accurate prediction
of phase transitions in compressed sensing via a connection to minimax
denoising,” IEEE Trans. Inf. Theory, vol. 59, no. 6, pp. 3396–3433, June
2013.
15
[22] A. Montanari, “Graphical models concepts in compressed sensing,”
in Compressed Sensing: Theory and Applications, Y. C. Eldar and
G. Kutyniok, Eds. New York: Cambridge University Press, 2012, ch. 9,
pp. 394–438.
[23] D. Needell and J. A. Tropp, “CoSaMP: Iterative signal recovery from
incomplete and inaccurate samples,” Appl. Comp. Harmonic Anal.,
vol. 26, no. 3, pp. 301–321, May 2009.
[24] D. L. Donoho, A. Maleki, and A. Montanari, “The noise-sensitivity
phase transition in compressed sensing,” IEEE Trans. Inf. Theory,
vol. 57, no. 10, pp. 6920–6941, Oct. 2011.
Zhilin Chen (S’14) received the B.E. degree in electrical and information engineering and the M.E. degree in signal and information processing from Beihang University (BUAA), Beijing, China, in 2012
and 2015, respectively. He is currently pursuing the
Ph.D. degree at the University of Toronto, Toronto,
ON, Canada. His main research interests include
wireless communication and signal processing.
Foad Sohrabi (S’13) received his B.A.Sc. degree in
2011 from the University of Tehran, Tehran, Iran,
and his M.A.Sc. degree in 2013 from McMaster
University, Hamilton, ON, Canada, both in Electrical
and Computer Engineering. Since September 2013,
he has been a Ph.D student at University of Toronto,
Toronto, ON, Canada. Form July to December 2015,
he was a research intern at Bell-Labs, AlcatelLucent, in Stuttgart, Germany. His main research
interests include MIMO communications, optimization theory, wireless communications, and signal
processing. He received an IEEE Signal Processing Society Best Paper Award
in 2017.
Wei Yu (S’97-M’02-SM’08-F’14) received the
B.A.Sc. degree in Computer Engineering and Mathematics from the University of Waterloo, Waterloo,
Ontario, Canada in 1997 and M.S. and Ph.D. degrees
in Electrical Engineering from Stanford University,
Stanford, CA, in 1998 and 2002, respectively. Since
2002, he has been with the Electrical and Computer Engineering Department at the University of
Toronto, Toronto, Ontario, Canada, where he is now
Professor and holds a Canada Research Chair (Tier
1) in Information Theory and Wireless Communications. His main research interests include information theory, optimization,
wireless communications and broadband access networks.
Prof. Wei Yu currently serves on the IEEE Information Theory Society
Board of Governors (2015-20). He serves as the Chair of the Signal Processing
for Communications and Networking Technical Committee of the IEEE Signal
Processing Society (2017-18). He was an IEEE Communications Society
Distinguished Lecturer (2015-16). He currently serves as an Area Editor of
the IEEE T RANSACTIONS ON W IRELESS C OMMUNICATIONS (2017-19). He
served as an Associate Editor for IEEE T RANSACTIONS ON I NFORMATION
T HEORY (2010-2013), as an Editor for IEEE T RANSACTIONS ON C OMMUNI CATIONS (2009-2011), as an Editor for IEEE T RANSACTIONS ON W IRELESS
C OMMUNICATIONS (2004-2007), and as a Guest Editor for a number of
special issues for the IEEE J OURNAL ON S ELECTED A REAS IN C OMMUNI CATIONS and the EURASIP J OURNAL ON A PPLIED S IGNAL P ROCESSING .
He was a Technical Program co-chair of the IEEE Communication Theory
Workshop in 2014, and a Technical Program Committee co-chair of the
Communication Theory Symposium at the IEEE International Conference
on Communications (ICC) in 2012. Prof. Wei Yu received the IEEE Signal
Processing Society Best Paper Award in 2017 and in 2008, a J OURNAL OF
C OMMUNICATIONS AND N ETWORKS Best Paper Award in 2017, a Steacie
Memorial Fellowship in 2015, an IEEE Communications Society Best Tutorial
Paper Award in 2015, an IEEE ICC Best Paper Award in 2013, the McCharles
Prize for Early Career Research Distinction in 2008, the Early Career Teaching
Award from the Faculty of Applied Science and Engineering, University of
Toronto in 2007, and an Early Researcher Award from Ontario in 2006. Prof.
Wei Yu is recognized as a Highly Cited Researcher. He is a Fellow of the
Canadian Academy of Engineering.
| 7cs.IT
|
Distributed Colour Reduction Revisited
Jukka Kohonen · [email protected] · University of Helsinki
Janne H. Korhonen · [email protected] · Aalto University
Christopher Purcell · [email protected] · Aalto University
Jukka Suomela · [email protected] · Aalto University
arXiv:1709.00901v1 [cs.DC] 4 Sep 2017
Przemyslaw Uznański · [email protected] · ETH Zürich
Abstract. We give a new, simple distributed algorithm for graph colouring in paths and
cycles. Our algorithm is fast and self-contained, it does not need any globally consistent
orientation, and it reduces the number of colours from 10100 to 3 in three iterations.
1
Introduction
We present a very fast and simple distributed algorithm for colouring paths and cycles.
Algorithms for colouring paths or cycles are key primitives that are used as subroutines
in many other distributed and parallel algorithms, and they are also material that is
typically covered in introductory courses on distributed algorithms [1, 7, 10, 11]. Yet the
best currently known algorithms tend to be inefficient, restricted, or inelegant.
1.1
Problem setting
We will focus on iterative colour reduction algorithms in paths and cycles: we have a path
that is properly coloured with k colours, and the algorithm will relabel the nodes so that
the path is properly coloured with f (k) < k colours. Naturally, we can then iterate the
algorithm to find a colouring with f (f (k)) colours, f (f (f (k))) colours, etc., until we reach
a fixed point. We use the notion k ⊲ f (k) for colour reduction from k to f (k).
We are interested in one-round algorithms [4], in which all nodes are re-coloured based
on the current colours of their immediate neighbours only. Such an algorithm can be
interpreted as a function
A : [k] × [k] × [k] → [c]
that maps three old colours to a new colour. Here A(x, y, z) is the new colour of a node
that was coloured y, and its two neighbours had colours x and z. We want to keep the
algorithm symmetric so that we can apply it even if we do not have a well-defined global
orientation:
A(x1 , x2 , x3 ) = A(x3 , x2 , x1 ) for all x1 , x2 , x3 .
(1)
And naturally we require that the algorithm produces a proper colouring as output, assuming we have a proper colouring as input:
A(x1 , x2 , x3 ) 6= A(x2 , x3 , x4 ) for all x1 6= x2 6= x3 6= x4 .
1
(2)
1.2
A simple example
Here is a colour reduction algorithm A : 4 ⊲ 3 that reduces the number of colours from 4
to 3:
A(x, 1, z) = 1,
A(x, 2, z) = 2,
A(x, 3, z) = 3,
A(x, 4, z) = min({1, 2, 3} \ {x, z}).
For example, if we have a properly 4-coloured path with the input colours
(1, 2, 1, 4, 3, 4, 3),
and we apply function A in all local neighbourhoods, we get a properly 3-coloured path
(1, 2, 1, 2, 3, 1, 3).
At the endpoints we can apply the standard trick: pretend that the nodes near the endpoints have a neighbour of some different colour.
1.3
Our contribution
We present a simple approach that enables us to do iterative colour reduction e.g. as
follows:
10100 ⊲ 12 ⊲ 4 ⊲ 3.
That is, in only 3 steps, we can reduce the number of colours from astronomical numbers
to only 3. Naturally getting below 3 colours is not possible with any local rule; 2-colouring
paths is an inherently global problem.
We are not aware of any prior algorithm that is equally fast and that does not need to
assume e.g. a global orientation. As we will see in the next section, our algorithm is also
much simpler than prior algorithms. Our algorithm is also completely self-contained, so
one can safely skip the next section.
2
Prior work
A typical theoretical presentation of this topic gives a number of algorithms, all of which
have a complexity of O(log∗ k) or 12 log∗ (k) + O(1) iterations for a colour reduction from
k to 3. However, to highlight the differences between the algorithms, we will use here
concrete numbers; our goal is to reduce the number of colours from at least M = 2128 to 3
(note that M is chosen so that we can use e.g. IPv6 addresses as colours). As we will see,
the seemingly innocent O(1) part actually dominates here.
Most of the prior algorithms proceed in two steps. First, we develop a colour reduction
algorithm assuming that there is a well-defined global orientation (each node has at most
one predecessor and at most one successor). Put otherwise, we develop an algorithm that
satisfies (2) but not necessarily (1). Then, using such an asymmetric algorithm as a black
box, we design a symmetric algorithm that also satisfies (1).
2
2.1
Asymmetric algorithms
The classical example of an asymmetric algorithm is the algorithm by Cole and Vishkin
from 1986 [2]. In its modern form, it is a colour reduction algorithm 2k ⊲ 2k. Iterating
the rule, we get
2128 ⊲ 256 ⊲ 16 ⊲ 8 ⊲ 6,
at which point the algorithm is stuck. At 6 colours we can switch to a naive algorithm
that eliminates one colour per round, and overall we obtain
2128 ⊲ 256 ⊲ 16 ⊲ 8 ⊲ 6 ⊲ 5 ⊲ 4 ⊲ 3.
In 7 steps we reduce the number of colours from M to 3. This is where a typical introductory lecture on the topic stops.
However, we can do better. A particularly elegant approach is toresort to the algorithm
by Naor and Stockmeyer [6], which gives a colour reduction of 2k
k ⊲ 2k. This rule has a
fixed point at 4 colours, after which we can do one round of the naive algorithm. Overall,
we get
1050000 ⊲ 184756 ⊲ 20 ⊲ 6 ⊲ 4 ⊲ 3.
Only 5 steps from way beyond M to 3 colours.
There is also a simple trick that is applicable to both Cole–Vishkin and Naor–Stockmeyer
algorithms. These algorithms only use the colour of the predecessor, and ignore the colour
of the successor, that is,
A(x, y, z) = A(x′ , y, z) for all x, x′ .
We can simulate two steps of any such algorithm in one iteration; for Cole–Vishkin we
obtain
2128 ⊲ 16 ⊲ 6 ⊲ 5 ⊲ 4 ⊲ 3
and for Naor–Stockmeyer we obtain
1050000 ⊲ 20 ⊲ 4 ⊲ 3.
(3)
We are down to only 3 iterations for colour reduction from beyond M to 3.
Here 3 rounds is optimal for these values of the parameters. Also, the last step 4 ⊲ 3 is
optimal; a simple computer search shows that there is no 5 ⊲ 3 algorithm [8, 9]. However,
the step 20 ⊲ 4 leaves some room for improvement; there is a 24 ⊲ 4 algorithm [8].
2.2
From asymmetric to symmetric algorithms
Equipped with an efficient asymmetric algorithm, for example (3), we can then apply it
in undirected paths in a somewhat ad-hoc fashion. Here is a concrete example:
1. Use the local minima and local maxima of the colours to split the path in fragments
that consist of strictly increasing colours.
2. Orient each fragment in the direction of increasing colours.
3. Apply an efficient asymmetric algorithm to each such fragment to iteratively find a
3-colouring. Now everything except the local minima and local maxima are properly
coloured.
3
4. Label the local minima with colour 4 and local maxima with colour 5, to obtain a
proper 5-colouring.
5. Run two rounds of the naive colour reduction algorithm to get back to 3 colours.
Some care is needed to make sure we do not lose too many rounds in each step; for
example, we want to do useful colour reduction already in the first round, in addition to
just identifying the local minima and maxima. With a little bit of thought, combined with
(3), we obtain something along the lines of
1050000 ⊲ 20 + 2 ⊲ 4 + 2 ⊲ 3 + 2 ⊲ 4 ⊲ 3,
where “+2” refers to the local minima and local maxima that will use two additional
colours. This is 5 rounds for M to 3. Our new algorithm will only need 3 rounds, and it
is much more streamlined.
2.3
Symmetric algorithms by design
There are very few efficient algorithms that are symmetric by design; most available algorithms are derived from asymmetric algorithms by following a scheme similar to the one
sketched above.
However, there are some colour reduction algorithms that are designed for general
undirected graphs, and we can naturally apply them also in the case of graphs of maximum
degree 2. The classical example is Linial’s [4, 5] algorithm. The algorithm is based on socalled cover-free set families [3], and we obtain slightly different algorithms based on the
specific choice of the set family. Unfortunately, all of the constructions give at best an
exponential colour reduction k ⊲ O(log k) per round; for example, one of the constructions
used by Linial gives k ⊲ 5⌈4 log k⌉.
Our new algorithm is much faster; it provides a doubly-exponential colour reduction
k ⊲ O(log log k) per round, which is asymptotically optimal for both symmetric and asymmetric algorithms [5].
3
Our algorithm
Fix a target number of colours c. We will consider
• subsets X ⊆ [c],
• families X that consist of such subsets, and
• collections A that consist of such families.
It turns out that certain collections can be directly interpreted as colour reduction algorithms; we will call such collections colourful collections.
Section 3.1 gives the definition of a colourful collection. Section 3.2 shows that any
colourful collection A gives a colour reduction algorithm k ⊲ c for k = |A|. Section 3.3
shows that the converse is also true: any colour reduction algorithm k ⊲ c gives a colourful
collection A of size k. Finally, Section 3.4 shows how to construct large colourful collections.
4
3.1
Colourful collections
We say that a collection A is colourful if the following holds:
(P1) For each family X ∈ A, for all subsets X, Y ∈ X , we have X ∩ Y 6= ∅.
(P2) For all families X 6= Y, there are X ∈ X and Y ∈ Y such that X ∩ Y = ∅.
Here is an example of a colourful collection, for c = 3 and |A| = 4:
n
o
A=
{1} , {2} , {3} , {1, 2}, {1, 3}, {2, 3} .
A simple example.
To keep the examples easier to read, we leave out two innermost levels of brackets and
commas and write simply
A = 1, 2, 3, 12 13 23 .
3.2
From colourful collections to algorithms
We now show how to use a colourful collection A to do colour reduction k ⊲ c for k =
|A|. The algorithm proceeds as follows. Here it will be convenient to imagine that each
undirected edge {u, v} is a pair of directed edges.
High-level plan:
1. Label each node u with a family L(u) ∈ A.
2. Using the labels of the nodes t and u, label the directed edge (t, u) with a subset
L(t, u) ∈ L(u).
3. Using the labels of the incoming edges (t, u) and (v, u), label the node u with a
colour ℓ(u) ∈ L(t, u) ∩ L(v, u).
Detailed description:
1. Apply a bijection [k] → A to label each node u with a family L(u) ∈ A.
2. Consider an edge {t, u}. By (P2), there are X ∈ L(t) and Y ∈ L(u) with X ∩ Y = ∅.
Assign the labels L(t, u) = Y and L(u, t) = X.
3. Consider a node u with the incoming edges (t, u) and (v, u). Note that L(t, u), L(v, u) ∈
L(v). By (P1), there is a c ∈ L(t, u) ∩ L(v, u). Set ℓ(u) = c.
Correctness. We already argued above that the algorithm is well-defined, assuming
that A is colourful. Let us now show that it indeed produces a proper colouring with c
colours. If not, there would be an edge {u, v} with ℓ(u) = ℓ(v). But by construction, we
have ℓ(u) ∈ L(v, u), ℓ(v) ∈ L(u, v), and L(u, v) ∩ L(v, u) = ∅, a contradiction.
5
3.3
From algorithms to colourful collections
We will now show that colourful collections are not only sufficient in the design of oneround colour reduction algorithms, but they are also necessary: given any colour reduction
algorithm A : k ⊲ c, we can construct a colourful collection A of size k. We proceed as
follows:
1. For any two colours x, y ∈ [k], x 6= y, define the subset Fx,y = {A(x, y, z) : z 6= y}.
2. For any colour y ∈ [k], define the family Fy = {Fx,y : x 6= y}.
3. Define the collection A = {Fy : y ∈ [k]}.
Correctness. First, consider a family Fy for some y. Let X, Y ∈ Fy . Then there are
x 6= y 6= z such that X = Fx,y and Y = Fz,y . By definition, we have A(x, y, z) ∈ Fx,y and
A(z, y, x) ∈ Fz,y , and by (1) also A(x, y, z) = A(z, y, x). Hence there is a common element,
A(x, y, z), in X ∩ Y , and (P1) holds.
Second, let x 6= y, and consider the families Fx and Fy . We have Fy,x ∈ Fx and
Fx,y ∈ Fy . If there was a common element a ∈ Fx,y ∩ Fy,x , there would also exist some
s and t with A(x, y, s) = A(y, x, t) = A(t, x, y), which contradicts (2); if we had a path
with the colours (t, x, y, s) in this order, algorithm A would fail to colour the middle nodes
properly. Therefore we have X ∈ Fx and Y ∈ Fy with X ∩ Y = ∅. This shows that (P2)
holds.
Together with (P1), the above argument also shows that Fx 6= Fy for x 6= y, and hence
the size of the collection is indeed |A| = k.
3.4
Explicit constructions of colourful collections
We have already seen in Section 3.1 an example of a colourful collection for c = 3 and
|A| = 4. Now we present the general scheme, for an even c. We will use the case of c = 4
c
as a running example. To simplify the notation, let s = c/2
.
1. Take all subsets of size c/2 (there are s such subsets):
12, 13, 14, 23, 24, 34 .
2. Split these in s/2 pairs of a subset and its complement:
(12, 34), (13, 24), (14, 23) .
3. Collection A0 contains all 2s/2 families that we can form by picking one half of each
pair:
A0 = 12 13 14,
12 13 23,
12 24 14,
12 24 23,
34 13 14,
34 13 23,
34 24 14,
34 24 23 .
6
4. Collection A1 is formed from A0 by augmenting each family with all subsets of size
c − 1:
A1 = 12 13 14 123 124 134 234,
12 13 23 123 124 134 234,
12 24 14 123 124 134 234,
12 24 23 123 124 134 234,
34 13 14 123 124 134 234,
34 13 23 123 124 134 234,
34 24 14 123 124 134 234,
34 24 23 123 124 134 234 .
5. Then form a collection A2 that contains all c families that contain just one subset of
size 1:
A2 = 1,
2,
3,
4 .
6. Finally, set A = A1 ∪ A2 .
Note that in the running example, we have c = 4 and |A| = 12, and hence this is a colour
reduction algorithm 12 ⊲ 4, assuming that the collection is indeed colourful.
Correctness. Collection A0 satisfies both (P1) and (P2): Consider some X , Y ∈ A, and
let X ∈ X , and let X̄ = [c] \ X be the complement of X. Now X̄ is the only set of size
h with X ∩ X̄ = ∅, and we have X̄ ∈
/ X , which satisfies (P1), and X̄ ∈ Y, which satisfies
(P2).
The augmented collection A1 still satisfies both (P1) and (P2): we only added subsets
that intersect with everything, satisfying (P1), and for the purposes of (P2) we can ignore
the extra subsets.
The trivial collection A2 clearly also satisfies both (P1) and (P2).
Finally, we need to argue that A also satisfies both (P1) and (P2). Property (P1)
follows directly from the fact that A1 and A2 satisfy it. What remains to be checked is
that (P2) holds even if we pick X ∈ A1 and Y ∈ A2 (or vice versa). But this is trivial,
as Y contains just one subset Y ∈ Y of one element, and its complement is contained in
every X ∈ A1 .
Analysis.
By construction, for every even c we have a colour reduction scheme
2s/2 + c ⊲ c,
c
where s = c/2
. As a concrete example, we obtain the following three-round colour
reduction scheme:
2462 + 12 ⊲ 12 ⊲ 4 ⊲ 3.
Asymptotically, the scheme is doubly-exponential; we have k ⊲ O(log log k).
7
Remarks. If we are doing colour reduction k ⊲ c, we can assign the input colours
1, 2, . . . , c to the singleton families in A2 . This way the colour reduction algorithm has an
additional nice property: the nodes that already have a colour at most c do not change their
colours. This is a natural property, but prior efficient algorithms (e.g. Naor–Stockmeyer
and Cole–Vishkin) do not guarantee this.
References
[1] Leonid Barenboim and Michael Elkin. Distributed Graph Coloring: Fundamentals and
Recent Developments. Synthesis Lectures on Distributed Computing Theory. Morgan
& Claypool Publishers, 2013.
[2] Richard Cole and Uzi Vishkin. Deterministic coin tossing with applications to optimal
parallel list ranking. Information and Control, 70(1):32–53, 1986.
[3] P. Erdős, P. Frankl, and Z. Füredi. Families of finite sets in which no set is covered
by the union of r others. Israel Journal of Mathematics, 51(1):79–89, Dec 1985.
[4] Fabian Kuhn and Roger Wattenhofer. On the complexity of distributed graph coloring.
In Proc. PODC 2006, pages 7–15, 2006.
[5] Nathan Linial. Locality in distributed graph algorithms. SIAM Journal on Computing,
21(1):193–201, 1992.
[6] Moni Naor and Larry Stockmeyer. What can be computed locally? SIAM Journal
on Computing, 24(6):1259–1277, 1995.
[7] David Peleg. Distributed Computing: A Locality-sensitive Approach. Society for
Industrial and Applied Mathematics, Philadelphia, PA, USA, 2000.
[8] Joel Rybicki. Exact bounds for distributed graph colouring. Master’s thesis, Department of Computer Science, University of Helsinki, May 2011.
[9] Joel Rybicki and Jukka Suomela. Exact bounds for distributed graph colouring. In
Proc. SIROCCO 2015, pages 46–60, 2015.
[10] Jukka Suomela. Distributed algorithms, 2016. Online textbook.
[11] Roger Wattenhofer. Lecture notes on principles of distributed computing, 2017. Online lecture notes.
8
| 8cs.DS
|
Logical Methods in Computer Science
Vol. 8 (2:12) 2012, pp. 1–27
www.lmcs-online.org
Submitted
Published
Jan. 7, 2011
Jun. 19, 2012
GENERIC FIBRATIONAL INDUCTION ∗
NEIL GHANI, PATRICIA JOHANN, AND CLÉMENT FUMEX
University of Strathclyde, Glasgow G1 1XH, UK
e-mail address: {Neil.Ghani, Patricia.Johann, Clement.Fumex}@cis.strath.ac.uk
Abstract. This paper provides an induction rule that can be used to prove properties of
data structures whose types are inductive, i.e., are carriers of initial algebras of functors.
Our results are semantic in nature and are inspired by Hermida and Jacobs’ elegant algebraic formulation of induction for polynomial data types. Our contribution is to derive,
under slightly different assumptions, a sound induction rule that is generic over all inductive types, polynomial or not. Our induction rule is generic over the kinds of properties to
be proved as well: like Hermida and Jacobs, we work in a general fibrational setting and
so can accommodate very general notions of properties on inductive types rather than just
those of a particular syntactic form. We establish the soundness of our generic induction
rule by reducing induction to iteration. We then show how our generic induction rule can
be instantiated to give induction rules for the data types of rose trees, finite hereditary
sets, and hyperfunctions. The first of these lies outside the scope of Hermida and Jacobs’
work because it is not polynomial, and as far as we are aware, no induction rules have
been known to exist for the second and third in a general fibrational framework. Our
instantiation for hyperfunctions underscores the value of working in the general fibrational
setting since this data type cannot be interpreted as a set.
1. Introduction
Iteration operators provide a uniform way to express common and naturally occurring
patterns of recursion over inductive data types. Expressing recursion via iteration operators makes code easier to read, write, and understand; facilitates code reuse; guarantees
properties of programs such as totality and termination; and supports optimising program transformations such as fold fusion and short cut fusion. Categorically, iteration
operators arise from initial algebra semantics of data types, in which data types are regarded as carriers of initial algebras of functors. Lambek’s Lemma ensures that the carrier
of the initial algebra of F is its least fixed point µF , and initiality ensures that, given
any F -algebra h : F A → A, there is a unique F -algebra homomorphism, denoted f old h,
from the initial algebra in : F (µF ) → µF to that algebra. For each functor F , the map
1998 ACM Subject Classification: F.3.2, D.3.1.
Key words and phrases: Induction, algebraic semantics of data types, fibrations, category theory.
∗
This paper is revised and significantly expanded version of [9].
This research is partially supported by EPSRC grant EP/C0608917/1.
l
LOGICAL METHODS
IN COMPUTER SCIENCE
c
DOI:10.2168/LMCS-8 (2:12) 2012
CC
N. Ghani, P. Johann, and C. Fumex
Creative Commons
2
N. GHANI, P. JOHANN, AND C. FUMEX
fold : (F A → A) → µF → A is the iteration operator for the data type µF . Initial algebra
semantics thus provides a well-developed theory of iteration which is...
• ...principled, in that it is derived solely from the initial algebra semantics of data types.
This is important because it helps ensure that programs have rigorous mathematical
foundations that can be used to ascertain their meaning and correctness.
• ...expressive, and so is applicable to all inductive types — i.e., to every type which is the
carrier of an initial algebra of a functor — rather than just to syntactically defined classes
of data types such as polynomial data types.
• ...correct, and so is valid in any model — set-theoretic, domain-theoretic, realisability,
etc. — in which data types are interpreted as carriers of initial algebras.
Because induction and iteration are closely linked — induction is often used to prove properties of functions defined by iteration, and the correctness of induction rules is often established by reducing it to that of iteration — we may reasonably expect that initial algebra
semantics can be used to derive a principled, expressive, and correct theory of induction
for data types as well. In most treatments of induction, given a functor F together with a
property P to be proved about data of type µF , the premises of the induction rule for µF
constitute an F -algebra with carrier Σx : µF. P x. The conclusion of the rule is obtained by
supplying such an F -algebra as input to the iteration operator for µF . This yields a function from µF to Σx : µF. P x from which a function of type ∀x : µF. P x can be obtained. It
has not, however, been possible to characterise F -algebras with carrier Σx : µF. P x without
additional assumptions on F . Induction rules are thus typically derived under the assumption that the functors involved have a certain structure, e.g., that they are polynomial.
Moreover, taking the carriers of the algebras to be Σ-types assumes that properties are
represented as type-valued functions. So while induction rules derived as described above
are both principled and correct, their expressiveness is limited along two dimensions: with
respect to the data types for which they can be derived and the nature of the properties
they can verify.
A more expressive, yet still principled and correct, approach to induction is given by
Hermida and Jacobs [10]. They show how to lift each functor F on a base category of types
to a functor F̂ on a category of properties over those types, and take the premises of the
induction rule for the type µF to be an F̂ -algebra. Hermida and Jacobs work in a fibrational
setting and the notion of property they consider is, accordingly, very general. Indeed, they
accommodate any notion of property that can be suitably fibred over the category of types,
and so overcome one of the two limitations mentioned above. On the other hand, their
approach gives sound induction rules only for polynomial data types, so the limitation on
the class of data types treated remains in their work.
This paper shows how to remove the restriction on the class of data types treated.
Our main result is a derivation of a sound generic induction rule that can be instantiated
to every inductive type, regardless of whether it is polynomial or not. We think this is
important because it provides a counterpart for induction to the existence of an iteration
operator for every inductive type. We take Hermida and Jacobs’ approach as our point of
departure and show that, under slightly different assumptions on the fibration involved, we
can lift any functor on the base category of a fibration to a functor on the total category of
the fibration. The lifting we define forms the basis of our generic induction rule.
GENERIC FIBRATIONAL INDUCTION
3
The derivation of a generic, sound induction rule covering all inductive types is clearly
an important theoretical result, but it also has practical consequences:
• We show in Example 2 how our generic induction rule can be instantiated to the families
fibration over Set (the fibration most often implicitly used by type theorists and those
constructing inductive proofs with theorem provers) to derive the induction rule for rose
trees that one would intuitively expect. The data type of rose trees lies outside the scope
of Hermida and Jacobs’ results because it is not polynomial. On the other hand, an
induction rule for rose trees is available in the proof assistant Coq, although it is neither
the one we intuitively expect nor expressive enough to prove properties that ought to be
amenable to inductive proof. Indeed, if we define rose trees in Coq by
Node : list rose -> rose
then Coq generates the following induction rule
rose_ind : forall P : rose -> Prop,
(forall l : list rose, P (Node l)) ->
forall r : rose, P r
But to prove a property of a rose tree Node l, we must prove that property assuming only
that l is a list of rose trees, and without recourse to any induction hypothesis. There is,
of course, a presentation of rose trees by mutual recursion as well, but this doesn’t give
the expected induction rule in Coq either. Intuitively, what we expect is an induction
rule whose premise is
forall [r_0, ..., r_n] : list rose,
P(r_0) -> ... -> P(r_n) -> P(Node [r_0, ..., r_n])
The rule we derive for rose trees is indeed the expected one, which suggests that our
derivation may enable automatic generation of more useful induction rules in Coq, rather
than requiring the user to hand code them as is currently necessary.
• We further show in Example 3 how our generic induction rule can be instantiated, again
to the families fibration over Set, to derive a rule for the data type of finite hereditary
sets. This data type is defined in terms of quotients and so lies outside most current
theories of induction.
• Finally, we show in Example 7 how our generic induction rule can be instantiated to the
subobject fibration over ωCP O⊥ to derive a rule for the data type of hyperfunctions.
Because this data type cannot be interpreted as a set, a fibration other than the families
fibration over Set is required; in this case, use of the subobject fibration allows us to
derive an induction rule for admissible subsets of hyperfunctions. The ability to treat the
data type of hyperfunctions thus underscores the importance of developing our results
in the general fibrational framework. Moreover, the functor underlying the data type of
hyperfunctions is not strictly positive [7], so the ability to treat this data type also underscores the advantage of being able to handle a very general class of functors going beyond
simply polynomial functors. As far as we know, induction rules for finite hereditary sets
and hyperfunctions have not previously existed in the general fibrational framework.
4
N. GHANI, P. JOHANN, AND C. FUMEX
Although our theory of induction is applicable to all inductive functors — i.e., to all
functors having initial algebras, including those giving rise to nested types [15], GADTs [21],
indexed containers [1], dependent types [19], and inductive recursive types [6] — our examples show that working in the general fibrational setting is beneficial even if we restrict our
attention to strictly positive data types. We do, however, offer some preliminary thoughts
in Section 5 on the potentially delicate issue of instantiating our general theory with fibrations appropriate for deriving induction rules for specific classes of higher-order functors
of interest. It is also worth noting that the specialisations of our generic induction rule to
polynomial functors in the families fibration over Set coincide exactly with the induction
rules of Hermida and Jacobs. But the structure we require of fibrations generally is slightly
different from that required by Hermida and Jacobs, so while our theory is in essence a
generalisation of theirs, the two are, strictly speaking, incomparable. The structure we
require of our fibrations is, nevertheless, certainly present in all standard fibrational models
of type theory (see Section 4). Like Hermida and Jacobs, we prove our generic induction
rule correct by reducing induction to iteration. A more detailed discussion of when our
induction rules coincide with those of Hermida and Jacobs is given in Section 4.
We take a purely categorical approach to induction in this paper, and derive our generic
induction rule from only the initial algebra semantics of data types. As a result, our work
is inherently extensional. Although translating our constructions into intensional settings
may therefore require additional effort, we expect the guidance offered by the categorical
viewpoint to support the derivation of induction rules for functors that are not treatable at
present. Since we do not use any form of impredicativity in our constructions, and instead
use only the weaker assumption that initial algebras exist, this guidance will be widely
applicable.
The remainder of this paper is structured as follows. To make our results as accessible
as possible, we illustrate them in Section 2 with a categorical derivation of the familiar
induction rule for the natural numbers. In Section 3 we derive an induction rule for the
special case of the families fibration over Set. We also show how this rule can be instantiated
to derive the one from Section 2, and the ones for rose trees and finite hereditary sets
mentioned above. Then, in Section 4 we present our generic fibrational induction rule,
establish a number of results about it, and illustrate it with the aforementioned application
to hyperfunctions. The approach taken in this section is completely different from the
corresponding one in the conference version of the paper [9], and allows us to improve upon
and extend our previous results. Section 5 concludes, discusses possible instantiations of
our generic induction rule for higher-order functors, and offers some additional directions
for future research.
When convenient, we identify isomorphic objects of a category and write = rather
than ≃. We write 1 for the canonical singleton set and denote its single element by · . In
Sections 2 and 3 we assume that types are interpreted as objects in Set, so that 1 also
denotes the unit type in those sections. We write id for identity morphisms in a category
and Id for the identity functor on a category.
2. A Familiar Induction Rule
Consider the inductive data type Nat, which defines the natural numbers and can be specified in a programming language with Haskell-like syntax by
data Nat = Zero | Succ Nat
GENERIC FIBRATIONAL INDUCTION
5
The observation that Nat is the least fixed point of the functor N on Set — i.e., on the
category of sets and functions — defined by N X = 1 + X can be used to define the
following iteration operator:
foldNat
foldNat z s Zero
foldNat z s (Succ n)
=
=
=
X → (X → X) → Nat → X
z
s (foldNat z s n)
The iteration operator foldNat provides a uniform means of expressing common and naturally occurring patterns of recursion over the natural numbers.
Categorically, iteration operators such as foldNat arise from the initial algebra semantics
of data types, in which every data type is regarded as the carrier of the initial algebra of
a functor F . If B is a category and F is a functor on B, then an F -algebra is a morphism
h : F X → X for some object X of B. We call X the carrier of h. For any functor F , the
collection of F -algebras itself forms a category Alg F which we call the category of F -algebras.
In Alg F , an F -algebra morphism between F -algebras h : F X → X and g : F Y → Y is a
map f : X → Y such that the following diagram commutes:
FA
Ff
/ FB
g
h
A
f
/B
When it exists, the initial F -algebra in : F (µF ) → µF is unique up to isomorphism and
has the least fixed point µF of F as its carrier. Initiality ensures that there is a unique
F -algebra morphism fold h : µF → X from in to any F -algebra h : F X → X. This gives
rise to the following iteration operator fold for the inductive type µF :
fold
fold h (in t)
:
=
(F X → X) → µF → X
h (F (fold h) t)
Since fold is derived from initial algebra semantics it is principled and correct. It is also
expressive, since it can be defined for every inductive type. In fact, fold is a single iteration
operator parameterised over inductive types rather than a family of iteration operators, one
for each such type, and the iteration operator foldNat above is the instantiation to Nat of
the generic iteration operator fold .
The iteration operator foldNat can be used to derive the standard induction rule for Nat
which coincides with the standard induction rule for natural numbers, i.e., with the familiar
principle of mathematical induction. This rule says that if a property P holds for 0, and if
P holds for n + 1 whenever it holds for a natural number n, then P holds for all natural
numbers. Representing each property of natural numbers as a predicate P : Nat → Set
mapping each n : Nat to the set of proofs that P holds for n, we wish to represent this rule
at the object level as a function indNat with type
∀(P : Nat → Set). P Zero → (∀n : Nat. P n → P (Succ n)) → (∀n : Nat. P n)
Code fragments such as the above, which involve quantification over sets, properties, or
functors, are to be treated as “categorically inspired” within this paper. This is because
quantification over such higher-kinded objects cannot be interpreted in Set. In order to
give a formal interpretation to code fragments like the one above, we would need to work
in a category such as that of modest sets. While the ability to work with functors over
categories other than Set is one of the motivations for working in the general fibrational
6
N. GHANI, P. JOHANN, AND C. FUMEX
setting of Section 4, formalising the semantics of such code fragments would obscure the
central message of this paper. Our decision to treat such fragments as categorically inspired
is justified in part by the fact that the use of category theory to suggest computational constructions has long been regarded as fruitful within the functional programming community
(see, e.g., [2, 3, 18]).
A function indNat with the above type takes as input the property P to be proved, a
proof φ that P holds for Zero, and a function ψ mapping each n : Nat and each proof that P
holds for n to a proof that P holds for Succ n, and returns a function mapping each n : Nat
to a proof that P holds for n, i.e., to an element of P n. We can write indNat in terms
of foldNat — and thus reduce induction for Nat to iteration for Nat — as follows. First
note that indNat cannot be obtained by instantiating the type X in the type of foldNat
to a type of the form P n for a specific n because indNat returns elements of the types
P n for different values n and these types are, in general, distinct from one another. We
therefore need a type containing all of the elements of P n for every n. Such a type can
informally be thought of as the union over n of P n, and is formally given by the dependent
type Σn : Nat. P n comprising pairs (n, p) where n : Nat and p : P n.
The standard approach to defining indNat is thus to apply foldNat to an N -algebra
with carrier Σn : Nat. P n. Such an algebra has components α : Σn : N at. P n and β : Σn :
N at. P n → Σn : N at. P n. Given φ : P Zero and ψ : ∀n. P n → P (Succ n), we choose
α = (Zero, φ) and β (n, p) = (Succ n, ψ n p) and note that f oldN at α β : N at → Σn :
N at. P n. We tentatively take indNat P φ ψ n to be p, where f oldN at α β n = (m, p).
But in order to know that p actually gives a proof for n itself, we must show that m = n.
Fortunately, this follows easily from the uniqueness of foldNat α β. Indeed, we have that
1 + N at
/ 1 + Σn : N at. P n
/ 1 + N at
[α,β]
in
N at
foldNat α β
in
/ Σn : N at. P n
λ(n,p). n
/ Nat
commutes and, by initiality of in, that (λ(n, p). n)◦(f oldN at α β) is the identity map. Thus
n = (λ(n, p). n)(f oldN at α β n) = (λ(n, p). n)(m, p) = m
πP′
Letting
be the second projection on dependent pairs involving the predicate P , the
induction rule for Nat is thus
indNat
: ∀(P : Nat → Set). P Zero → (∀n : Nat. P n → P (Succ n))
→ (∀n : Nat. P n)
′
indNat P φ ψ = πP ◦ (foldNat (Zero, φ) (λ(n, p). (Succ n, ψ n p)))
As expected, this induction rule states that, for every property P , to construct a proof that
P holds for every n : Nat, it suffices to provide a proof that P holds for Zero, and to show
that, for any n : Nat, if there is a proof that P holds for n, then there is also a proof that
P holds for Succ n.
The use of dependent types is fundamental to this formalization of the induction rule
for Nat, but this is only possible because properties to be proved are taken to be set-valued
functions. The remainder of this paper uses fibrations to generalise the above treatment
of induction to arbitrary inductive functors and arbitrary properties which are suitably
fibred over the category whose objects interpret types. In the general fibrational setting,
properties are given axiomatically via the fibrational structure rather than assumed to be
(set-valued) functions.
GENERIC FIBRATIONAL INDUCTION
7
3. Induction Rules for Predicates over Set
The main result of this paper is the derivation of a sound induction rule that is generic over
all inductive types and which can be used to verify any notion of property that is fibred
over the category whose objects interpret types. In this section we assume that types are
modelled by sets, so the functors we consider are on Set and the properties we consider are
functions mapping data to sets of proofs that these properties hold for them. We make these
assumptions because it allows us to present our derivation in the simplest setting possible,
and also because type theorists often model properties in exactly this way. This makes the
present section more accessible and, since the general fibrational treatment of induction can
be seen as a direct generalisation of the treatment presented here, Section 4 should also be
more easily digestible once the derivation is understood in this special case. Although the
derivation of this section can indeed be seen as the specialisation of that of Section 4 to the
families fibration over Set, no knowledge of fibrations is required to understand it because
all constructions are given concretely rather than in their fibrational forms.
We begin by considering what we might naively expect an induction rule for an inductive
data type µF to look like. The derivation for Nat in Section 2 suggests that, in general, it
should look something like this:
ind : ∀P : µ F → Set. ??? → ∀ x : µF. P x
But what should the premises — denoted ??? here — of the generic induction rule ind be?
Since we want to construct, for any term x : µF , a proof term of type P x from proof terms
for x’s substructures, and since the functionality of the fold operator for µF is precisely
to compute a value for x : µF from the values for x’s substructures, it is natural to try to
equip P with an F -algebra structure that can be input to fold to yield a mapping of each
x : µF to an element of P x. But this approach quickly hits a snag. Since the codomain
of every predicate P : µF → Set is Set itself, rather than an object of Set, F cannot be
applied to P as is needed to equip P with an F -algebra structure. Moreover, an induction
rule for µF cannot be obtained by applying fold to an F -algebra with carrier P x for any
specific x. This suggests that we should try to construct an F -algebra not for P x for each
term x, but rather for P itself.
Such considerations led Hermida and Jacobs [10] to define a category of predicates P
and a lifting for each polynomial functor F on Set to a functor F̂ on P that respects the
structure of F . They then constructed F̂ -algebras with carrier P to serve as the premises of
their induction rules. The crucial part of their construction, namely the lifting of polynomial
functors, proceeds inductively and includes clauses such as
(F\
+ G) P = F̂ P + ĜP
and
(F\
× G) P = F̂ P × ĜP
The construction of Hermida and Jacobs is very general: they consider functors on bicartesian categories rather than just on Set, and represent properties by bicartesian fibrations
over such categories instead of using the specific notion of predicate from Definition 3.2
below. On the other hand, they define liftings for polynomial functors.
The construction we give in this section is in some sense orthogonal to Hermida and
Jacobs’: we focus exclusively on functors on Set and a particular category of predicates, and
show how to define liftings for all inductive functors on Set, including non-polynomial ones.
8
N. GHANI, P. JOHANN, AND C. FUMEX
In this setting, the induction rule we derive properly extends Hermida and Jacobs’, thus
catering for a variety of data types that they cannot treat. In the next section we derive
analogous results in the general fibrational setting. This allows us to derive sound induction
rules for initial algebras of functors defined on categories other than Set which can be used
to prove arbitrary properties that are suitably fibred over the category interpreting types.
We begin with the definition of a predicate.
Definition 3.1. Let X be a set. A predicate on X is a function P : X → Set mapping
each x ∈ X to a set P x. We call X the domain of P .
We may speak simply of “a predicate P ” if the domain of P is understood. A predicate
P on X can be thought of as mapping each element x of X to the set of proofs that P holds
for x. We now define our category of predicates.
Definition 3.2. The category of predicates P has predicates as its objects. A morphism
from a predicate P : X → Set to a predicate P ′ : X ′ → Set is a pair (f, f ∼ ) : P → P ′ of
functions, where f : X → X ′ and f ∼ : ∀x : X. P x → P ′ (f x). Composition of predicate
morphisms is given by (g, g∼ ) ◦ (f, f ∼ ) = (g ◦ f, λxp. g ∼ (f x)(f ∼ xp)).
Diagrammatically, we have
f
/ X′
④
∼
❉❉
④
❉❉ f
④④
❉❉
④/ ④P ′
④
P
}④
"
X ❉
❉
Set
As the diagram indicates, the notion of a morphism from P to P ′ does not require the sets
of proofs P x and P ′ (f x), for any x ∈ X, to be equal. Instead, it requires only the existence
of a function f ∼ which maps, for each x, each proof in P x to a proof in P ′ (f x). We denote
by U : P → Set the forgetful functor mapping each predicate P : X → Set to its domain
X and each predicate morphism (f, f ∼ ) to f .
An alternative to Definition 3.2 would take the category of predicates to be the arrow
category over Set, but the natural lifting in this setting does not indicate how to generalise
liftings to other fibrations. Indeed, if properties are modelled as functions, then every
functor can be applied to a property, and hence every functor can be its own lifting. In the
general fibrational setting, however, properties are not necessarily modelled by functions,
so a functor cannot, in general, be its own lifting. The decision not to use arrow categories
to model properties is thus dictated by our desire to lift functors in a way that indicates
how liftings can be constructed in the general fibrational setting.
We can now give a precise definition of a lifting.
Definition 3.3. Let F be a functor on Set. A lifting of F from Set to P is a functor F̂ on
P such that the following diagram commutes:
P
F̂
/P
U
U
Set
F
/ Set
We can decode the definition of F̂ as follows. The object part of F̂ must map each predicate
P : X → Set to a predicate F̂ P : F X → Set, and thus can be thought of type-theoretically
GENERIC FIBRATIONAL INDUCTION
9
as a function ∀(X : Set). (X → Set) → F X → Set. Of course, F̂ must also act on
morphisms in a functorial manner.
We can now use the definition of a lifting to derive the standard induction rule from
Section 2 for Nat as follows.
Example 1. The data type of natural numbers is µN where N is the functor on Set defined
by N X = 1 + X. A lifting N̂ of N can be defined by sending each predicate P : X → Set
b P : N X → Set given by
to the predicate N
N̂ P (inl ·)
N̂ P (inr n)
=
=
1
Pn
An N̂ -algebra with carrier P : Nat → Set can be given by in : 1 + Nat → Nat and
in∼ : ∀t : 1 + Nat. N̂ P t → P (in t). Since in (inl ·) = 0 and in (inr n) = n + 1, we see that
in∼ consists of an element h1 : P 0 and a function h2 : ∀n : Nat. P n → P (n + 1). Thus, the
second component in ∼ of an N̂ -algebra with carrier P : Nat → Set and first component in
gives the premises of the familiar induction rule in Example 1.
The notion of predicate comprehension is a key ingredient of our lifting. It begins to
explain, abstractly, what the use of Σ-types is in the theory of induction, and is the key
construct allowing us to define liftings for non-polynomial, as well as polynomial, functors.
Definition 3.4. Let P be a predicate on X. The comprehension of P , denoted {P }, is the
type Σx : X. P x comprising pairs (x, p) where x : X and p : P x. The map taking each
predicate P to {P }, and taking each predicate morphism (f, f ∼ ) : P → P ′ to the morphism
{(f, f ∼ )} : {P } → {P ′ } defined by {(f, f ∼ )}(x, p) = (f x, f ∼ x p), defines the comprehension
functor {−} from P to Set.
We are now in a position to define liftings uniformly for all functors:
Definition 3.5. If F is a functor on Set, then the lifting F̂ is the functor on P given as
follows. For every predicate P on X, F̂ P : F X → Set is defined by F̂ P = (F πP )−1 ,
where the natural transformation π : {−} → U is given by πP (x, p) = x. For every predicate
morphism f : P → P ′ , F̂ f = (k, k ∼ ) where k = F U f , and k∼ : ∀y : F X. F̂ P y → F̂ P ′ (k y)
is given by k∼ y z = F {f }z.
In the above definition, note that the inverse image f −1 of f : X → Y is indeed a predicate
P : Y → Set. Thus if P is a predicate on X, then πP : {P } → X and F πP : F {P } → F X.
Thus F̂ P is a predicate on F X, so F̂ is a lifting of F from Set to P. The lifting F̂ captures
an “all” modality, in that it generalises Haskell’s all function on lists to arbitrary data
types. A similar modality is given in [17] for indexed containers.
The lifting in Example 1 is the instantiation of the construction in Definition 3.5 to
the functor N X = 1 + X on Set. Indeed, if P is any predicate, then N̂ P = (N πP )−1 ,
i.e., N̂ P = (id + πP )−1 . Then, since the inverse image of the coproduct of functions is the
coproduct of their inverse images, since id −1 1 = 1, and since πP−1 n = {(n, p) | p : P n} for
all n, we have N̂ P (inl ·) = 1 and N̂ P (inr n) = P n. As we will see, a similar situation
to that for Nat holds in general: for any functor F on Set, the second component of an
F̂ -algebra whose carrier is the predicate P on the data type µF and whose first component
is in gives the premises of an induction rule that can be used to show that P holds for all
data of type µF .
10
N. GHANI, P. JOHANN, AND C. FUMEX
The rest of this section shows that F -algebras with carrier {P } are interderivable with
F̂ -algebras with carrier P , and then uses this result to derive our induction rule.
Definition 3.6. The functor K1 : Set → P maps each set X to the predicate K1 X = λx :
X. 1 on X and each f : X → Y to the predicate morphism (f, λx : X. id).
The predicate K1 X is called the truth predicate on X. For every x : X, the set K1 Xx of
proofs that K1 X holds for x is a singleton, and thus is non-empty. We intuitively think of
a predicate P : X → Set as being true if P x is non-empty for every x : X. We therefore
consider P to be true if there exists a predicate morphism from K1 X to P whose first
component is id X . For any functor F , the lifting F̂ is truth-preserving, i.e., F̂ maps the
truth predicate on any set X to that on F X.
Lemma 3.7. For any functor F on Set and any set X, F̂ (K1 X) ∼
= K1 (F X).
Proof. By Definition 3.5, F̂ (K1 X) = (F πK1 X )−1 . We have that πK1 X is an isomorphism
because there is only one proof of K1 X for each x : X, and thus that F πK1 X is an isomorphism as well. As a result, (F πK1 X )−1 maps every y : F X to a singleton set, and therefore
F̂ (K1 X) = (F πK1 X )−1 ∼
= λy : F X. 1 = K1 (F X).
The fact that K1 is a left-adjoint to {−} is critical to the constructions below. This is
proved in [10]; we include its proof here for completeness and to establish notation. The
description of comprehension as a right adjoint can be traced back to Lawvere [14].
Lemma 3.8. K1 is left adjoint to {−}.
Proof. We must show that, for any predicate P and any set Y , the set P(K1 Y, P ) of
morphisms from K1 Y to P in P is in bijective correspondence with the set Set(Y, {P })
of morphisms from Y to {P } in Set. Define maps (−)† : Set(Y, {P }) → P(K1 Y, P ) and
(−)# : P(K1 Y, P ) → Set(Y, {P }) by h† = (h1 , h2 ) where hy = (v, p), h1 y = v and h2 y = p,
and (k, k∼ )# = λ(y : Y ). (ky, k∼ y). These give a natural isomorphism between Set(Y, {P })
and P(K1 Y, P ).
Naturality of (−)† ensures that (g ◦f )† = g † ◦K1 f for all f : Y ′ → Y and g : Y → {P }.
Similarly for (−)# . Moreover, id† is the counit, at P , of the adjunction between K1 and
{−}. These observations are used in the proof of Lemma 3.10. Lemmas 3.9 and 3.10 are the
key results relating F -algebras and F̂ algebras, i.e., relating iteration and induction. They
are special cases of Theorem 4.8 below, but we include their proofs to ensure continuity of
our presentation and to ensure that this section is self-contained.
We first we show how to construct F̂ -algebras from F -algebras
Lemma 3.9. There is a functor Φ : Alg F → Alg F̂ such that if k : F X → X, then
Φk : F̂ (K1 X) → K1 X.
Proof. For an F -algebra k : F X → X define Φk = K1 k, and for two F -algebras k : F X → X
and k′ : F X ′ → X ′ and an F -algebra morphism h : X → X ′ between them define the F̂ algebra morphism Φh : Φk → Φk ′ by Φh = K1 h. Then K1 (F X) ∼
= F̂ (K1 X) by Lemma 3.7,
so that Φk is an F̂ -algebra and K1 h is an F̂ -algebra morphism. It is easy to see that Φ
preserves identities and composition.
GENERIC FIBRATIONAL INDUCTION
11
We can also construct F -algebras from F̂ -algebras.
Lemma 3.10. The functor Φ has a right adjoint Ψ such that if j : F̂ P → P , then Ψj :
F {P } → {P }.
Proof. We construct the adjoint functor Ψ : Alg F̂ → Alg F as follows. Given an F̂ -algebra
j : F̂ P → P , we use the fact that F̂ (K1 {P }) ∼
= K1 (F {P }) by Lemma 3.7 to define
†
#
Ψj : F {P } → {P } by Ψj = (j ◦ F̂ id ) . To specify the action of Ψ on an F̂ -algebra
morphism h, define Ψh = {h}. Clearly Ψ preserves identity and composition.
Next we show Φ ⊣ Ψ, i.e., for every F -algebra k : F X → X and F̂ -algebra j : F̂ P → P
with P a predicate on X, there is a natural isomorphism between F -algebra morphisms
from k to Ψj and F̂ -algebra morphisms from Φk to j. We first observe that an F -algebra
morphism from k to Ψj is a map from X to {P }, and an F̂ -algebra morphism from Φk
to j is a map from K1 X to P . A natural isomorphism between such maps is given by the
adjunction K1 ⊣ {−} from Lemma 3.8. We must check that f : X → {P } is an F -algebra
morphism from k to Ψj iff f † : K1 X → P is an F̂ -algebra morphism from Φk to j.
To this end, assume f : X → {P } is an F -algebra morphism from k to Ψj, i.e., assume
f ◦k = Ψj ◦F f . We must prove that f † ◦ Φk = j ◦ F̂ f † . By the definition of Φ in Lemma 3.9,
this amounts to showing f † ◦ K1 k = j ◦ F̂ f † . Now, since (−)† is an isomorphism, f is an F algebra morphism iff (f ◦k)† = (Ψj◦F f )† . Naturality of (−)† ensures that (f ◦k)† = f † ◦K1 k
and that (Ψj ◦ F f )† = (Ψj)† ◦ K1 (F f ), so the previous equality holds iff
f † ◦ K1 k = (Ψj)† ◦ K1 (F f )
(3.1)
But
j ◦ F̂ f †
= j ◦ F̂ id † ◦ K1 f )
= (j ◦ F̂ id† ) ◦ F̂ (K1 f )
= (Ψj)† ◦ K1 (F f )
= f † ◦ K1 k
by naturality of (−)† and f = id ◦ f
by the functoriality of F̂
by the definition of Ψ, the fact that (−)† and (−)#
are inverses, and Lemma 3.7
by Equation 3.1
Thus, f † is indeed an F̂ -algebra morphism from Φk to j.
Lemma 3.10 ensures that F -algebras with carrier {P } are interderivable with F̂ -algebras
with carrier P . For example, the N -algebra [α, β] with carrier {P } from Section 2 can be
derived from the N̂ -algebra with carrier P given in Example 1. Since we define a lifting
F̂ for any functor F , Lemma 3.10 thus shows how to construct F -algebras with carrier
Σx : µF. P x for any functor F and predicate P on µF .
Corollary 3.11. For any functor F on Set, the predicate K1 (µF ) is the carrier of the
initial F̂ -algebra.
Proof. Since Φ is a left adjoint it preserves initial objects, so applying Φ to the initial
F -algebra in : F (µF ) → µF gives the initial F̂ -algebra. By Lemma 3.9, Φ in has type
F̂ (K1 (µF )) → K1 (µF ), so the carrier of the initial F̂ -algebra is K1 (µF ).
12
N. GHANI, P. JOHANN, AND C. FUMEX
We can now derive our generic induction rule. For every predicate P on X and every F̂ algebra (k, k∼ ) : F̂ P → P , Lemma 3.10 ensures that Ψ constructs from (k, k∼ ) an F -algebra
with carrier {P }. Applying the iteration operator to this algebra gives a map
fold (Ψ (k, k∼ )) : µF → {P }
This map decomposes into two parts: φ = πP ◦ fold (Ψ (k, k∼ )) : µF → X and ψ : ∀(t :
µF ). P (φ t). Initiality of in : F (µF ) → µF , the definition of Ψ, and the naturality of πP
ensure φ = fold k. Recalling that πP′ is the second projection on dependent pairs involving
the predicate P , this gives the following sound generic induction rule for the type X, which
reduces induction to iteration:
genind
: ∀ (F : Set → Set) (P : X → Set) ((k, k ∼ ) : (F̂ P → P )) (t : µF ).
P (fold k t)
genind F P = πP′ ◦ fold ◦ Ψ
Notice this induction rule is actually capable of dealing with predicates over arbitrary sets
and not just predicates over µF . However, when X = µF and k = in, initiality of in further
ensures that φ = fold in = id, and thus that genind specialises to the expected induction
rule for an inductive data type µF :
ind
∀ (F : Set → Set) (P : µF → Set) ((k, k ∼ ) : (F̂ P → P )).
(k = in) → ∀(t : µF ). P t
= πP′ ◦ fold ◦ Ψ
:
ind F P
This rule can be instantiated to familiar rules for polynomial data types, as well as to ones
we would expect for data types such as rose trees and finite hereditary sets, both of which
lie outside the scope of Hermida and Jacobs’ method.
Example 2. The data type of rose trees is given in Haskell-like syntax by
data Rose = Node(List Rose)
The functor underlying Rose is F X = List X and its induction rule is
indRose
:
indRose F P
Calculating F̂ P = (F πP
list xs, we have that
)−1
=
∀ (P : Rose → Set) ((k, k ∼ ) : (F̂ P → P )).
(k = in) → ∀(x : Rose). P x
′
πP ◦ fold ◦ Ψ
: F Rose → Set, and writing xs !! k for the k th component of a
F̂ P rs
= {z : F {P } | F πp z = rs}
= {cps : List {P } | List πP cps = rs}
= {cps : List {P } | ∀k < length cps. πP (cps !! k) = rs !! k}
An F̂ -algebra whose underlying F -algebra is in : F Rose → Rose is thus a pair of functions
(in, k∼ ), where k∼ has type
= ∀rs : List Rose.
{cps : List {P } | ∀k < length cps. πP (cps !! k) = rs !! k} → P (Node rs)
= ∀rs : List Rose. (∀k < length rs. P (rs !! k)) → P (Node rs)
GENERIC FIBRATIONAL INDUCTION
13
The last equality is due to surjective pairing for dependent products and the fact that
length cps = length rs. The type of k ∼ gives the hypotheses of the induction rule for rose
trees.
Although finite hereditary sets are defined in terms of quotients, and thus lie outside
the scope of previously known methods, they can be treated with ours.
Example 3. Hereditary sets are sets whose elements are themselves sets, and so are the
core data structures within set theory. The data type HS of finitary hereditary sets is µPf
for the finite powerset functor Pf . We can derive an induction rule for finite hereditary
sets as follows. If P : X → Set, then Pf πP : Pf (Σx : X.P x) → Pf X maps each set
{(x1 , p1 ), . . . , (xn , pn )} to the set {x1 , . . . , xn }, so that (Pf πP )−1 maps a set {x1 , . . . , xn } to
the set P x1 × . . . × P xn . A Pˆf -algebra with carrier P : HS → Set and first component in
therefore has as its second component a function of type
∀({s1 , . . . , sn } : Pf (HS)). P s1 × . . . × P sn → P (in{s1 , . . . , sn })
The induction rule for finite hereditary sets is thus
indHS :: (∀({s1 , . . . , sn } : Pf (HS)). P s1 × . . . × P sn → P (in{s1 , . . . , sn }))
→ ∀(s : HS).P s
4. Generic Fibrational Induction Rules
We can treat more general notions of predicates using fibrations. We motivate the use of
fibrations by observing that i) the semantics of data types in languages involving recursion
and other effects usually involves categories other than Set; ii) in such circumstances, the
notion of a predicate can no longer be taken as a function with codomain Set; and iii)
even when working in Set there are reasonable notions of “predicate” other than that in
Section 3. (For example, a predicate on a set X could be a subobject of X). Moreover,
when, in future work, we consider induction rules for more sophisticated classes of data
types such as indexed containers, inductive families, and inductive recursive families (see
Section 5), we will not want to have to develop an individual ad hoc theory of induction for
each such class. Instead, we will want to appropriately instantiate a single generic theory of
induction. That is, we will want a uniform axiomatic approach to induction that is widely
applicable, and that abstracts over the specific choices of category, functor, and predicate
giving rise to different forms of induction for specific classes of data types.
Fibrations support precisely such an axiomatic approach. This section therefore generalises the constructions of the previous one to the general fibrational setting. The standard
model of type theory based on locally cartesian closed categories does arise as a specific
fibration — namely, the codomain fibration over Set — and this fibration is equivalent
to the families fibration over Set. But the general fibrational setting is far more flexible.
Moreover, in locally cartesian closed models of type theory, predicates and types coexist in
the same category, so that each functor can be taken to be its own lifting. In the general fibrational setting, predicates are not simply functions or morphisms, properties and types do
not coexist in the same category, and a functor cannot be taken to be its own lifting. There
is no choice but to construct a lifting from scratch. A treatment of induction based solely
on locally cartesian closed categories would not, therefore, indicate how to treat induction
in more general fibrations.
14
N. GHANI, P. JOHANN, AND C. FUMEX
Another reason for working in the general fibrational setting is that this facilitates a
direct comparison of our work with that of Hermida and Jacobs [10]. This is important,
since their approach is the most closely related to ours. The main difference between
their approach and ours is that they use fibred products and coproducts to define provably
sound induction rules for polynomial functors, whereas we use left adjoints to reindexing
functors to define provably sound induction rules for all inductive functors. In this section
we consider situations when both approaches are possible and give mild conditions under
which our results coincide with theirs when restricted to polynomial functors.
The remainder of this section is organised as follows. In Section 4.1 we recall the
definition of a fibration, expand and motivate this definition, and fix some basic terminology
surrounding fibrations. We then give some examples of fibrations, including the families
fibration over Set, the codomain fibration, and the subobject fibration. In Section 4.2 we
recall a useful theorem from [10] that indicates when a truth-preserving lifting of a functor
to a category of predicates has an initial algebra. This is the key theorem used to prove
the soundness of our generic fibrational induction rule. In Section 4.3 we construct truthpreserving liftings for all inductive functors. We do this first in the codomain fibration, and
then, using intuitions from its presentation as the families fibration over Set, as studied
in Section 3, in a general fibrational setting. Finally, in Section 4.4 we establish a number
of properties of the liftings, and hence of the induction rules, that we have derived. In
particular, we characterise the lifting that generates our induction rules.
4.1. Fibrations in a Nutshell. In this section we recall the notion of a fibration. More
details about fibrations can be found in, e.g., [12, 20]. We begin with an auxiliary definition.
Definition 4.1. Let U : E → B be a functor.
(1) A morphism g : Q → P in E is cartesian over a morphism f : X → Y in B if U g = f ,
and for every g′ : Q′ → P in E for which U g′ = f ◦ v for some v : U Q′ → X there exists
a unique h : Q′ → Q in E such that U h = v and g ◦ h = g ′ .
(2) A morphism g : P → Q in E is opcartesian over a morphism f : X → Y in B if U g = f ,
and for every g′ : P → Q′ in E for which U g′ = v ◦ f for some v : Y → U Q′ there exists
a unique h : Q → Q′ in E such that U h = v and h ◦ g = g ′ .
It is not hard to see that the cartesian morphism fP§ over a morphism f with codomain U P
is unique up to isomorphism, and similarly for the opcartesian morphism f§P . If P is an
GENERIC FIBRATIONAL INDUCTION
15
object of E, then we write f ∗ P for the domain of fP§ and Σf P for the codomain of f§P . We
can capture cartesian and opcartesian morphisms diagrammatically as follows.
E
′
Q′ ◗◗◗
◗◗◗ ′
◗◗◗g
◗◗◗
h
◗◗◗
◗◗(
∗
/P
f P
§
♠6 Q
♠♠♠ O
g ′ ♠♠♠♠
♠
h
♠♠♠
♠♠♠
♠
♠
♠
/ Σf P
P
P
U Q′ ◗◗
◗◗◗
◗◗U◗ g′
◗◗◗
v
◗◗◗
◗◗(
/Y
X
❧6 U Q
O
❧❧❧
❧
❧
❧
❧
v
❧❧
❧❧❧
❧
❧
❧
/Y
X
fP
f§
U
B
f
′
U g′
f
Cartesian morphisms (opcartesian morphisms) are the essence of fibrations (resp., opfibrations). We introduce both fibrations and their duals now since the latter will prove useful
later in our development. Below we speak primarily of fibrations, with the understanding
that the dual observations hold for opfibrations.
Definition 4.2. Let U : E → B be a functor. Then U is a fibration if for every object P
of E, and every morphism f : X → U P in B there is a cartesian morphism fP§ : Q → P in
E above f . Similarly, U is an opfibration if for every object P of E, and every morphism
f : U P → Y in B there is an opcartesian morphism f§P : P → Q in E above f . A functor
U a bifibration if it is simultaneously a fibration and an opfibration.
If U : E → B is a fibration, we call B the base category of U and E the total category
of U . Objects of the total category E can be thought of as properties, objects of the base
category B can be thought of as types, and U can be thought of as mapping each property
P in E to the type U P of which P is a property. One fibration U can capture many different
properties of the same type, so U is not injective on objects. We say that an object P in
E is above its image U P under U , and similarly for morphisms. For any object X of B, we
write EX for the fibre above X, i.e., for the subcategory of E consisting of objects above X
and morphisms above id. If f : X → Y is a morphism in B, then the function mapping
each object P of E to f ∗ P extends to a functor f ∗ : EY → EX . Indeed, for each morphism
k : P → P ′ in EY , f ∗ k is the morphism satisfying k ◦ fP§ = fP§ ′ ◦ f ∗ k. The universal
property of fP§ ′ ensures the existence and uniqueness of f ∗ k. We call the functor f ∗ the
reindexing functor induced by f . A similar situation ensures for opfibrations, and we call
the functor Σf : EX → EY which extends the function mapping each object P of E to Σf P
the opreindexing functor.
Example 4. The functor U : P → Set defined in Section 3 is called the families fibration
over Set. Given a function f : X → Y and a predicate P : Y → Set we can define a
cartesian map fP§ whose domain f ∗ P is P ◦ f , and which comprises the pair (f, λx : X. id).
The fibre PX above a set X has predicates P : X → Set as its objects. A morphism in PX
from P : X → Set to P ′ : X → Set is a function of type ∀x : X. P x → P ′ x.
Example 5. Let B be a category. The arrow category of B, denoted B → , has the morphisms,
or arrows, of B as its objects. A morphism in B → from f : X → Y to f ′ : X ′ → Y ′ is a pair
16
N. GHANI, P. JOHANN, AND C. FUMEX
(α1 , α2 ) of morphisms in B such that the following diagram commutes:
α1
X
/ X′
f′
f
Y
α2
/ Y′
i.e., such that α2 ◦ f = f ′ ◦ α1 . It is easy to see that this definition indeed gives a category.
The codomain functor cod : B → → B maps an object f : X → Y of B → to the object
Y of B and a morphism (α1 , α2 ) of B → to α2 . If B has pullbacks, then cod is a fibration,
called the codomain fibration over B. Indeed, given an object f : X → Y in the fibre above
Y and a morphism f ′ : X ′ → Y in B, the pullback of f along f ′ gives a cartesian morphism
above f ′ as required. The fibre above an object Y of B has those morphisms of B that map
into Y as its objects. A morphism in (B → )Y from f : X → Y to f ′ : X ′ → Y is a morphism
α1 : X → X ′ in B such that f = f ′ ◦ α1 .
Example 6. If B is a category, then the category of subobjects of B, denoted Sub(B), has
monomorphisms in B as its objects. A monomorphism f : X ֒→ Y is called a subobject of
Y . A morphism in Sub(B) from f : X ֒→ Y to f ′ : X ′ ֒→ Y ′ is a pair of morphisms (α1 , α2 )
in B such that α2 ◦ f = f ′ ◦ α1 .
The map U : Sub(B) → B sending a subobject f : X ֒→ Y to Y extends to a functor. If
B has pullbacks, then U is a fibration, called the subobject fibration over B; indeed, pullbacks
again give cartesian morphisms since the pullback of a monomorphism is a monomorphism.
The fibre above an object Y of B has as objects the subobjects of Y . A morphism in
Sub(B)Y from f : X ֒→ Y to f ′ : X ′ ֒→ Y is a map α1 : X → X ′ in B such that f = f ′ ◦ α1 .
If such a morphism exists then it is, of course, unique.
4.2. Lifting, Truth, and Comprehension. We now generalise the notions of lifting,
truth, and comprehension to the general fibrational setting. We prove that, in such a setting,
if an inductive functor has a truth-preserving lifting, then its lifting is also inductive. We
then see that inductiveness of the lifted functor is sufficient to guarantee the soundness
of our generic fibrational induction rule. This subsection is essentially our presentation of
pre-existing results from [10]. We include it because it forms a natural part of our narrative,
and because simply citing the material would hinder the continuity of our presentation.
Recall from Section 3 that the first step in deriving an induction rule for a datatype
interpreted in Set is to lift the functor whose fixed point the data type is to the category P of
predicates. More specifically, in Definition 3.3 we defined a lifting of a functor F : Set → Set
to be a functor F̂ : P → P such that U F̂ = F U . We can use these observations to generalise
the notion of a lifting to the fibrational setting as follows.
Definition 4.3. Let U : E → B be a fibration and F be a functor on B. A lifting of F with
respect to U is a functor F̂ : E → E such that the following diagram commutes:
E
F̂
/E
U
U
B
F
/B
GENERIC FIBRATIONAL INDUCTION
17
In Section 3 we saw that if P : X → Set is a predicate over X, then F̂ P is a predicate over
F X. The analogous result for the general fibrational setting observes that if F̂ is a lifting
of F and X is an object of B, then F̂ restricts to a functor from EX to EF X .
By analogy with our results from Section 3, we further expect that the premises of a
fibrational induction rule for a datatype µF interpreted in B should constitute an F̂ -algebra
on E. But in order to construct the conclusion of such a rule, we need to understand how
to axiomatically state that a predicate is true. In Section 3, a predicate P : X → Set is
considered true if there is a morphism in P from K1 X, the truth predicate on X, to P that
is over id X . Since the mapping of each set X to K1 X is the action on objects of the truth
functor K1 : Set → P (cf. Definition 3.6), we actually endeavour to model the truth functor
for the families fibration over Set axiomatically in the general fibrational setting.
Modeling the truth functor axiomatically amounts to understanding its universal property. Since the truth functor in Definition 3.6 maps each set X to the predicate λx : x. 1, for
any set X there is therefore exactly one morphism in the fibre above X from any predicate
P over X to K1 X. This gives a clear categorical description of K1 X as a terminal object
of the fibre above X and leads, by analogy, to the following definition.
Definition 4.4. Let U : E → B be a fibration. Assume further that, for every object
X of B, the fibre EX has a terminal object K1 X such that, for any f : X ′ → X in B,
f ∗ (K1 X) ∼
= K1 X ′ . Then the assignment sending each object X in B to K1 X in E, and
§
each morphism f : X ′ → X in B to the morphism fK
in E defines the (fibred) truth
1X
functor K1 : B → E.
The (fibred) truth functor is sometimes called the (fibred) terminal object functor. With
this definition, we have the following standard result:
Lemma 4.5. K1 is a (fibred) right adjoint for U .
The interested reader may wish to consult the literature on fibrations for the definition
of a fibred adjunction, but a formal definition will not be needed here. Instead, we can
simply stress that a fibred adjunction is first and foremost an adjunction, and then observe
that the counit of this adjunction is the identity, so that U K1 = Id . Moreover, K1 is full
and faithful. One simple way to guarantee that a fibration has a truth functor is to assume
that both E and B have terminal objects and that U maps a terminal object of E to a
terminal object of B. In this case, the fact that reindexing preserves fibred terminal objects
ensures that every fibre of E indeed has a terminal object.
The second fundamental property of liftings used in Section 3 is that they are truthpreserving. This property can now easily be generalised to the general fibrational setting
(cf. Definition 3.7).
Definition 4.6. Let U : E → B be a fibration with a truth functor K1 : B → E, let F be
a functor on B, and let F̂ : E → E be a lifting of F . We say that F̂ is a truth-preserving
lifting of F if, for any object X of B, we have F̂ (K1 X) ∼
= K1 (F X).
The final algebraic structure we required in Section 3 was a comprehension functor
{−} : P → Set. To generalise the comprehension functor to the general fibrational setting
we simply note that its universal property is that it is right adjoint to the truth functor K1
(cf. Definition 3.8). We single out for special attention those fibrations whose truth functors
have right adjoints.
18
N. GHANI, P. JOHANN, AND C. FUMEX
Definition 4.7. Let U : E → B be a fibration with a truth functor K1 : B → E. Then U is
a comprehension category with unit if K1 has a right adjoint.
If U : E → B is a comprehension category with unit, then we call the right adjoint to K1
the comprehension functor and denote it by {−} : E → B. With this machinery in place,
Hermida and Jacobs [10] show that if U is a comprehension category with unit and F̂ is a
truth-preserving lifting of F , then F̂ is inductive if F is and, in this case, the carrier µF̂ of
the initial F̂ -algebra is K1 (µF ). This is proved as a corollary to the following more abstract
theorem.
Theorem 4.8. Let F : B → B, G : A → A, and S : B → A be functors. A natural
transformation α : GS → SF , i.e., a natural transformation α such that
AO
S
B
G
/A
O
❈❈❈❈❈❈α
%
S
/B
F
induces a functor
AlgF
Φ
/ AlgG
given by Φ (f : F X → X) = S f ◦ αX . Moreover, if α is an isomorphism, then a right
adjoint T to S induces a right adjoint
AlgF l
Φ
⊤
,
AlgG
Ψ
given by Ψ(g : GX → X) = T g ◦βX , where β : F T → T G is the image of G ǫ◦α−1
T : SF T →
∼
G under the adjunction isomorphism Hom(S X, Y ) = Hom(X, T Y ), and ǫ : ST → id is
the counit of this adjunction.
We can instantiate Theorem 4.8 to generalise Lemmas 3.9 and 3.10.
Theorem 4.9. Let U : E → B be a comprehension category with unit and F : B → B be a
functor. If F has a truth-preserving lifting F̂ then there is an adjunction Φ ⊣ Ψ : Alg F →
Alg F̂ . Moreover, if f : F X → X then Φf : F̂ (K1 X) → K1 X, and if g : F̂ P → P then
Ψg : F {P } → {P }.
Proof. We instantiate Theorem 4.8, letting E be A, F̂ be G, and K1 be S. Then α is an
isomorphism since F̂ is truth-preserving, and we also have that K1 ⊣ {−}. The theorem thus
ensures that Φ maps every F -algebra f : F X → X to an F̂ -algebra Φf : F̂ (K1 X) → K1 X,
and that Ψ maps every F̂ -algebra g : F̂ P → P to an F -algebra Ψg : F {P } → {P }.
Corollary 4.10. Let U : E → B be a comprehension category with unit and F : B → B be
a functor which has a truth-preserving lifting F̂ . If F is inductive, then so is F̂ . Moreover,
µF̂ = K1 (µF ).
Proof. The hypotheses of the corollary place us in the setting of Theorem 4.9. This theorem
guarantees that Φ maps the initial F -algebra in F : F (µF ) → µF to an F̂ -algebra with
carrier K1 (µF ). But since left adjoints preserve initial objects, we must therefore have that
the initial F̂ -algebra has carrier K1 (µF ). Thus, µF̂ exists and is isomorphic to K1 (µF ).
GENERIC FIBRATIONAL INDUCTION
19
Theorem 4.11. Let U : E → B be a comprehension category with unit and F : B → B
be an inductive functor. If F has a truth-preserving lifting F̂ , then the following generic
fibrational induction rule is sound:
genfibind
genfibind F P
: ∀ (F : B → B) (P : E). (F̂ P → P ) → (µF̂ → P )
= fold
An alternative presentation of genfibind is
genfibind
genfibind F P
: ∀ (F : B → B) (P : E). (F̂ P → P ) → (µF → {P })
= fold ◦ Ψ
We call genfibind F the generic fibrational induction rule for µF .
In summary, we have generalised the generic induction rule for predicates over Set
presented in Section 3 to give a sound generic induction rule for comprehension categories
with unit. Our only assumption is that if we start with an inductive functor F on the
base of the comprehension category, then there must be a truth-preserving lifting of that
functor to the total category of the comprehension category. In that case, we can specialise
genfibind to get a fibrational induction rule for any datatype µF that can be interpreted in
the fibration’s base category.
The generic fibrational induction rule genfibind does, however, look slightly different
from the generic induction rule for set-valued predicates. This is because, in Section 3, we
used our knowledge of the specific structure of comprehensions for set-valued predicates
to extract proofs for particular data elements from them. But in the fibrational setting,
predicates, and hence comprehensions, are left abstract. We therefore take the return type
of the general induction scheme genfibind to be a comprehension with the expectation that,
when the general theory of this section is instantiated to a particular fibration of interest, it
may be possible to use knowledge about that fibration to extract from the comprehension
constructed by genfibind further proof information relevant to the application at hand.
As we have previously mentioned, Hermida and Jacobs provide truth-preserving liftings
only for polynomial functors. In Section 4.3, we define a generic truth-preserving lifting for
any inductive functor on the base category of any fibration which, in addition to being a
comprehension category with unit, has left adjoints to all reindexing functors. This gives a
sound generic fibrational induction rule for the datatype µF for any functor F on the base
category of any such fibration.
4.3. Constructing Truth-Preserving Liftings. In light of the previous subsection, it is
natural to ask whether or not truth-preserving liftings exist. If so, are they unique? Or, if
there are many truth-preserving liftings, is there a specific truth-preserving lifting to choose
above others? Is there, perhaps, even a universal truth-preserving lifting? We can also ask
about the algebraic structure of liftings. For example, do truth-preserving liftings preserve
sums and products of functors?
Answers to some of these questions were given by Hermida and Jacobs, who provided
truth-preserving liftings for polynomial functors. To define such liftings they assume that
the total category and the base category of the fibration in question have products and
coproducts, and that the fibration preserves them. Under these conditions, liftings for
polynomial functors can be defined inductively. In this section we go beyond the results of
Hermida and Jacobs and construct truth-preserving liftings for all inductive functors. We
employ a two-stage process, first building truth-preserving liftings under the assumption
20
N. GHANI, P. JOHANN, AND C. FUMEX
that the fibration of interest is a codomain fibration, and then using the intuitions of
Section 3 to extend this lifting to a more general class of fibrations. In Section 4.4 we
consider the questions from the previous paragraph about the algebraic structure of liftings.
4.3.1. Truth-Preserving Liftings for Codomain Fibrations. Recall from Example 5 that if B
has pullbacks, then the codomain fibration over B is the functor cod : B → → B. Given a
functor F : B → B, it is trivial to define a lifting F → : B → → B → for this fibration. We can
define the functor F → to map an object f : X → Y of B → to F f : F X → F Y , and to map
a morphism (α1 , α2 ) to the morphism (F α1 , F α2 ). That F → is a lifting is easily verified.
If we further verify that codomain fibrations are comprehension categories with unit,
and that the lifting F → is truth-preserving, then Theorem 4.11 can be applied to them. For
the former, we first observe that the functor K1 : B → B → mapping an object X to id and
a morphism f : X → Y to (f, f ) is a truth functor for this fibration. (In fact, we can take
any isomorphism into X as K1 X; we will use this observation below.) If we let B → (U, V )
denote the set of morphisms from an object U to an object V in B → , then the fact that K1
is right adjoint to cod can be established via the natural isomorphism
B → (f : X → Y, K1 Z) = {(α1 : X → Z, α2 : Y → Z) | α1 = α2 ◦ f } ∼
= B(Y, Z) = B(cod f, Z)
We next show that the functor dom : B → → B mapping an object f : X → Y of B → to
X and a morphism (α1 , α2 ) to α1 is a comprehension functor for the codomain fibration.
That dom is right adjoint to K1 is established via the natural isomorphism
B → (K1 Z, f : X → Y ) = {(α1 : Z → X, α2 : Z → Y ) | α2 = f ◦ α1 } ∼
= B(Z, X) = B(Z, dom f )
Finally, we have that F → is truth-preserving because
F → (K1 Z) = F → id = F id = id = K1 (F Z)
A lifting is implicitly given in [16] for functors on a category with display maps. Such
a category is a subfibration of the codomain fibration over that category, and the lifting
given there is essentially the lifting for the codomain fibration restricted to the subfibration
in question.
4.3.2. Truth-Preserving Liftings for the Families Fibration over Set. In Section 3 we defined, for every functor F : Set → Set, a lifting F̂ which maps the predicate P to (F πP )−1 .
Looking closely, we realise this lifting decomposes into three parts. Given a predicate P ,
we first consider the projection function πP : {P } → U P . Next, we apply the functor F
to πP to obtain F πP : F {P } → F U P . Finally, we take the inverse image of F πP to get a
predicate over F U P as required.
Note that π is the functor from P to Set→ which maps a predicate P to the projection
function πP : {P } → U P (and maps a predicate morphism (f, f ∼ ) from a predicate P : X →
Set to P ′ : X ′ → Set to the morphism ({(f, f ∼ )}, f ) from πP to πP ′ ; cf. Definition 3.4). If
I : Set→ → P is the functor sending a function f : X → Y to its “inverse” predicate f −1
(and a morphism (α1 , α2 ) to the predicate morphism (α2 , ∀y : Y. λx : f −1 y. α1 x)), then
each of the three steps of defining F̂ is functorial and the relationships indicated by the
GENERIC FIBRATIONAL INDUCTION
21
following diagram hold:
π
P ❇l
❇❇
❇❇
❇❇
U ❇!
⊤
I
Set
,
Set→
✇✇
✇✇
✇
✇
{ ✇ cod
✇
Note that the adjunction I ⊣ π is an equivalence. This observation is not, however,
necessary for our subsequent development; in particular, it is not needed for Theorem 4.14.
The above presentation of the lifting F̂ of a functor F for the families fibration over Set
uses the lifting of F for the codomain fibration over Set. Indeed, writing F → for the lifting
of F for the codomain fibration over Set, we have that F̂ = IF → π. Moreover, since π and
I are truth-preserving (see the proof of Lemma 3.7), and since we have already seen that
liftings for codomain fibrations are truth-preserving, we have that F̂ is truth-preserving
because each of its three constituent functors is. Finally, since we showed in Section 3 that
the families fibration over Set is a comprehension category with unit, Theorem 4.11 can be
applied to it.
Excitingly, as we shall see in the next subsection, the above presentation of the lifting
of a functor for the families fibration over Set generalises to many other fibrations!
4.3.3. Truth-Preserving Liftings for Other Fibrations. We now turn our attention to the
task of constructing truth-preserving liftings for fibrations other than codomain fibrations
and the families fibration over Set. By contrast with the approach outlined in the conference
paper [9] on which this paper is based, the one we take here uses a factorisation, like that
of the previous subsection, through a codomain fibration. More specifically, let U : E → B
be a comprehension category with unit. We first define functors I and π, and construct an
adjunction I ⊣ π between E and B → such that the relationships indicated by the following
diagram hold:
π
E ❄k
❄❄
❄❄
❄
U ❄❄
⊤
I
,
B→
⑤
⑤⑤
⑤
⑤⑤
}⑤⑤ cod
B
We then use the adjunction indicated in the diagram to construct truth-preserving a lifting
for U from that for the codomain fibration over B.
To define the functor π : E → B → we generalise the definition of π : P → Set→
from Sections 3 and 4.3.2. This requires us to work with the axiomatic characterisation in
Definition 4.7 of the comprehension functor {−} : E → B as the right adjoint to the truth
functor K1 : B → E. The counit of the adjunction K1 ⊣ {−} is a natural transformation
ǫ : K1 {−} → Id. Applying U to ǫ gives the natural transformation U ǫ : U K1 {−} → U , but
since U K1 = Id, in fact we have that U ǫ : {−} → U . We can therefore define π to be U ǫ.
Then π is indeed a functor from E to B → , its action on an object P is πP , and its action
on a morphism (f, f ∼ ) is ({(f, f ∼ )}, f ).
We next turn to the definition of the left adjoint I to π. To see how to generalise the
inverse image construction to more general fibrations we first recall from Example 4 that, if
f : X → Y is a function and P : Y → Set, then f ∗ P = P ◦ f . We can extend this mapping
22
N. GHANI, P. JOHANN, AND C. FUMEX
to a reindexing functor f ∗ : EY → EX by defining f ∗ (id , h∼ ) = (id , h∼ ◦ f ). If we define the
action of Σf : EX → EY on objects by
]
Σf P = λy.
Px
{x|f x=y}
U
where denotes the disjoint union operator on sets, and its action on morphisms by taking
Σf (id , α∼ ) to be (id , ∀(y : Y ). λ(x : X, p : f x = y, t : P x). (x, p, α∼ x t)), then Σf is left
adjoint to f ∗ . Moreover, if we compute
]
Σf (K1 X) = λy.
K1 Xx
{x | f x=y}
and recall that, for any x : X, the set K1 X x is a singleton, then Σf (K1 X) is clearly
equivalent to the inverse image of f .
The above discussion suggests that, in order to generalise the inverse image construction
to a more general fibration U : E → B, we should require each reindexing functor f ∗ to have
the opreindexing functor Σf as its left adjoint. As in [10], no Beck-Chevalley condition is
required on these adjoints. The following result, which appears as Proposition 2.3 of [11],
thus allows us to isolate the exact class of fibrations for which we will have sound generic
induction rules.
Theorem 4.12. A fibration U : E → B is a bifibration iff for every morphism f in B the
reindexing functor f ∗ has left adjoint Σf .
Definition 4.13. A Lawvere category is a bifibration which is also a comprehension category
with unit.
We construct the left adjoint I : B → → E of π for any Lawvere category U : E → B as
follows. If f : X → Y is an object of B → , i.e., a morphism of B, then we define I f to be the
object Σf (K1 X) of E. To define the action of I on morphisms, let (α1 , α2 ) be a morphism
in B → from f : X → Y to f ′ : X ′ → Y ′ in B → . Then (α1 , α2 ) is a pair of morphisms in B
such that the following diagram commutes:
X
α1
/ X′
f′
f
Y
α2
/ Y′
We must construct a morphism from Σf (K1 X) to Σf ′ (K1 X ′ ) in E. To do this, notice
′
that f§′K1 X ◦ K1 α1 : K1 X → Σf ′ (K1 X ′ ) is above f ′ ◦ α1 , and that it is also above α2 ◦ f
′
since f ′ ◦ α1 = α2 ◦ f . We can then consider the morphism f§′K1 X ◦ K1 α1 and use the
universal property of the opcartesian morphism f§K1 X to deduce the existence of a morphism
h : Σf (K1 X) → Σf ′ (K1 X ′ ) above α2 . It is not difficult, using the uniqueness of the
morphism h, to prove that setting this h to be the image of the morphism (α1 , α2 ) makes
I a functor. In fact, since cod ◦ π = U , Result (i) on page 190 of [11] guarantees that,
for any Lawvere category U : E → B the functor I : B → → E exists and is left adjoint to
π : E → B→ .
We can now construct a truth-preserving lifting for any Lawvere category U : E → B
and functor F on B.
GENERIC FIBRATIONAL INDUCTION
23
Theorem 4.14. Let U : E → B be a Lawvere category and, for any functor F on B, define
the functor F̂ on E by
F̂ : E → E
F̂ = IF → π
Then F̂ is a truth-preserving lifting of F .
Proof. It is trivial to check that F̂ is indeed a lifting. To prove that it is truth-preserving, we
need to prove that F̂ (K1 X) ∼
= K1 (F X) for any functor F on B and object X of B. We do
this by showing that each of π, F → , and I preserves fibred terminal objects, i.e., preserves
the terminal objects of each fibre of the total category which is its domain. Then since
K1 X is a terminal object in the fibre EX , we will have that F̂ (K1 X) = I(F → (π(K1 X))) is
a terminal object in EF X , i.e., that F̂ (K1 X) ∼
= K1 (F X) as desired.
We first show that π preserves fibred terminal objects. We must show that, for any
object X of B, πK1 X is a terminal object of the fibre of B → over X, i.e., is an isomorphism
with codomain X. We prove this by observing that, if η : Id → {−}K1 is the unit of
the adjunction K1 ⊣ {−}, then πK1 X is an isomorphism with inverse ηX . Indeed, if ǫ
is the counit of the same adjunction, then the facts that U K1 = Id and that K1 is full
and faithful ensure that K1 ηX is an isomorphism with inverse ǫK1 X . Thus, ǫK1 X is an
isomorphism with inverse K1 ηX , and so πK1 X = U ǫK1 X is an isomorphism with inverse
U K1 ηX , i.e., with inverse ηX . Since K1 X is a terminal object in EX and πK1 X is a terminal
object in the fibre of B → over X, we have that π preserves fibred terminal objects.
It is not hard to see that F → preserves fibred terminal objects: applying the functor F
to an isomorphism with codomain X — i.e., to a terminal object in the fibre of B → over
X — gives an isomorphism with codomain F X — i.e., a terminal object in the fibre of B →
over F X.
Finally, if f : X → Y is an isomorphism in B, then Σf is not only left adjoint to f ∗ ,
but also right adjoint to it. Since right adjoints preserve terminal objects, and since K1 X
is a terminal object of EX , we have that If = Σf (K1 X) is a terminal object of EY . Thus I
preserves fibred terminal objects.
We stress that, to define our lifting, the codomain functor over the base B of the Lawvere
category need not be a fibration. In particular, B need not have pullbacks; indeed, all that
is needed to construct our generic truth-preserving lifting F̂ for a functor F on B is the
existence of the functors I and π (and F → , which always exists). We nevertheless present
the lifting F̂ as the composition of π, F → , and I because this presentation shows it can be
factored through F → . This helps motivate our definition of F̂ , thereby revealing parallels
between it and F → that would otherwise not be apparent. At the same time it trades the
direct, brute-force presentation of F̂ from [9] for an elegant modularly structured one which
makes good use, in a different setting, of general results about comprehension categories
due to Jacobs [11].
We now have the promised sound generic fibrational induction rule for every inductive
functor F on the base of a Lawvere category. To demonstrate the flexibility of this rule, we
now derive an induction rule for a data type and properties on it that cannot be modelled
in Set. Being able to derive induction rules for fixed points of functors in categories other
than Set is a key motivation for working in a general fibrational setting.
Example 7. The fixed point Hyp = µF of the functor F X = (X → Int) → Int is the data
type of hyperfunctions. Since F has no fixed point in Set, we interpret it in the category
24
N. GHANI, P. JOHANN, AND C. FUMEX
ωCP O⊥ of ω-cpos with ⊥ and strict continuous monotone functions. In this setting, a
property of an object X of ωCP O⊥ is an admissible sub-ωCP O⊥ P of X. Admissibility
means that the bottom element of X is in P and P is closed under least upper bounds of
ω-chains in X. This structure forms a Lawvere category [11, 12]; in particular, it is routine
to verify the existence of its opreindexing functor. In particular, Σf P is constructed for a
continuous map f : X → Y and an admisible predicate P ⊆ X, as the intersection of all
admissible Q ⊆ Y with P ⊆ f −1 (Q). The truth functor maps X to X, and comprehension
maps a sub-ωCP O⊥ P of X to P . The lifting F̂ maps a sub-ωCP O⊥ P of X to the least
admissible predicate on F X containing the image of F P . Finally, the derived induction
rule states that if P is an admissible sub-ωCP O⊥ of Hyp, and if F̂ (P ) ⊆ P , then P = Hyp.
4.4. An Algebra of Lifting. We have proved that in any Lawvere category U : E → B,
any functor F on B has a lifting F̂ on E which is truth-preserving, and thus has the following
associated sound generic fibrational induction rule:
genfibind
genfibind F P
: ∀ (F : B → B) (P : E). (F̂ P → P ) → (µF → {P })
= fold ◦ Ψ
In this final subsection of the paper, we ask what kinds of algebraic properties the lifting
operation has. Our first result concerns the lifting of constant functors.
Lemma 4.15. Let U : E → B be a Lawvere category and let X be an object of B. If FX is
the constantly X-valued functor on B, then Fc
X is isomorphic to the constantly K1 X-valued
functor on E.
Proof. For any object P of E we have
→
Fc
X P = (I(FX ) π)P = I(FX πP ) = ΣF
X πP
K1 FX {P } = Σid K1 X ∼
= K1 X
The last isomorphism holds because id ∗ ∼
= Id and Σid ⊣ id ∗ .
Our next result concerns the lifting of the identity functor. It requires a little additional
structure on the Lawvere category of interest.
Definition 4.16. A full Lawvere category is a Lawvere category U : E → B such that
π : E → B → is full and faithful.
b ∼
Lemma 4.17. In any full Lawvere category, Id
= Id
Proof. By the discussion following Definition 4.13, I ⊣ π. Since π is full and faithful, the
counit of this adjunction is an isomorphism, and so IπP ∼
= P for all P in E. We therefore
have that
bP
P ∼
= IπP = Σπ K1 {P } = ΣId π K1 (Id {P }) = (I Id → π)P = Id
p
p
bP ∼
i.e., that Id
= P for all P in E. Because these isomorphisms are clearly natural, we
b ∼
therefore have that Id
= Id.
GENERIC FIBRATIONAL INDUCTION
25
We now show that the lifting of a coproduct of functors is the coproduct of the liftings.
Lemma 4.18. Let U : E → B be a Lawvere category and let F and G be functors on B.
Then F\
+G∼
= F̂ + Ĝ.
Proof. We have
(F\
+ G)P = I((F + G)→ πP ) = I(F → πP + G→ πP ) ∼
= I(F → πP ) + I(G→ πP ) = F̂ P + ĜP
The third isomorphism holds because I is a left adjoint and so preserves coproducts.
Note that the statement of Lemma 4.18 does not assert the existence of either of the
two coproducts mentioned, but rather that, whenever both do exist, they must be equal.
Note also that the lemma generalises to any colimit of functors. Unfortunately, no result
analogous to Lemma 4.18 can yet be stated for products.
Our final result considers whether or not there is anything fundamentally special about
the lifting we have constructed. It is clearly the “right” lifting in some sense because it
gives the expected induction rules. But other truth-preserving liftings might also exist and,
if this is the case, then we might hope our lifting satisfies some universal property. In
fact, under a further condition, which is also satisfied by all of the liftings of Hermida and
Jacobs, and which we therefore regard as reasonable, we can show that our lifting is the only
truth-preserving lifting. Our proof uses a line of reasoning which appears in Remark 2.13
in [10].
Lemma 4.19. Let U : E → B be a full Lawvere category and let F be a truth-preserving
lifting of a functor F on B. If F preserves Σ-types — i.e., if (F )(Σf P ) ∼
= ΣF f (F )P
∼
— then F = F̂ .
Proof. We have
(F )P
∼
=
∼
=
∼
=
∼
=
b P)
(F )(Id
(F )(ΣπP K1 {P })
ΣF πP (F )K1 {P }
ΣF πP K1 F {P }
= F̂ P
Finally, we can return to the question of the relationship between the liftings of polynomial
functors given by Hermida and Jacobs and the liftings derived by our methods. We have seen
that for constant functors, the identity functor, and coproducts of functors our constructions
agree. Moreover, since Hermida and Jacobs’ liftings all preserve Σ-types, Lemma 4.19
guarantees that in a full Lawvere category their lifting for products also coincides with
ours.
5. Conclusion and future work
We have given a sound induction rule that can be used to prove properties of data structures
of inductive types. Like Hermida and Jacobs, we give a fibrational account of induction,
but we derive, under slightly different assumptions on fibrations, a generic induction rule
that can be instantiated to any inductive type rather than just to polynomial ones. This
rule is based on initial algebra semantics of data types, and is parameterised over both
the data types and the properties involved. It is also principled, expressive, and correct.
26
N. GHANI, P. JOHANN, AND C. FUMEX
Our derivation yields the same induction rules as Hermida and Jacobs’ when specialised to
polynomial functors in the families fibration over Set and in other fibrations, but it also
gives induction rules for non-polynomial data types such as rose trees, and for data types
such as finite hereditary sets and hyperfunctions, for which no fibrational induction rules
have previously been known to exist.
There are several directions for future work. The most immediate is to instantiate our
theory to give induction rules for more sophisticated data types, such as nested types. These
are exemplified by the data type of perfect trees given in Haskell-like syntax as follows:
data PTree a : Set where
PLeaf : a → PTree a
PNode : PTree (a, a) → PTree a
Nested types arise as least fixed points of rank-2 functors; for example, the type of perfect
trees is µH for the functor H given by HF = λX. X + F (X × X). An appropriate fibration
for induction rules for nested types thus takes B to be the category of functors on Set,
E to be the category of functors from Set to P, and U to be postcomposition with the
forgetful functor from Section 3. A lifting Ĥ of H is given by Ĥ P X (inl a) = 1 and
Ĥ P X (inr n) = P (X × X) n. Taking the premise to be an Ĥ-algebra gives the following
induction rule for perfect trees:
indP T ree : ∀ (P : Set → P).
(U P = PTree) → (∀(X : Set)(x : X). P (PLeaf x)) →
(∀(X : Set)(t : PTree (X × X). P (X × X) t → P (PNode t))) →
∀(X : Set)(t : PTree X). P X t
This rule can be used to show, for example, that P T ree is a functor.
Extending the above instantiation for the codomain fibration to so-called “truly nested
types” [15] and fibrations is current work. We expect to be able to instantiate our theory for
truly nested types, GADTs, indexed containers, dependent types, and inductive recursive
types, but initial investigations show care is needed. We must ascertain which fibrations can
model predicates on such types, since the codomain fibration may not give useful induction
rules, as well as how to translate the rules to which these fibrations give rise to an intensional
setting.
Matthes [15] gives induction rules for nested types (including truly nested ones) in an
intensional type theory. He handles only rank-2 functors that underlie nested types (while
we handle any functor of any rank with an initial algebra), but his insights may help guide
choices of fibrations for truly nested types. These may in turn inform choices for GADTs,
indexed containers, and dependent types.
Induction rules can automatically be generated in many type theories. Within the
Calculus of Constructions [4] an induction rule for a data type can be generated solely
from the inductive structure of that type. Such generation is also a key idea in the Coq
proof assistant [5]. As far as we know, generation can currently be done only for syntactic
classes of functors rather than for all inductive functors with initial algebras. In some
type theories induction schemes are added as axioms rather than generated. For example,
attempts to generate induction schemes based on Church encodings in the Calculus of
Constructions proved unsuccessful and so initiality was added to the system, thus giving the
Calculus of Inductive Constructions. Whereas Matthes’ work is based on concepts such as
impredicativity and induction recursion rather than initial algebras, ours reduces induction
GENERIC FIBRATIONAL INDUCTION
27
to initiality, and may therefore help lay the groundwork for extending implementations of
induction to more sophisticated data types.
Acknowledgement
We thank Robert Atkey, Pierre-Evariste Dagand, Peter Hancock, and Conor McBride for
many fruitful discussions.
References
[1] T. Altenkirch and P. Morris. Indexed Containers. Proceedings, Logic in Computer Science, pp. 277–285,
2009.
[2] R. S. Bird and O. De Moor. Algebra of Programming. Prentice Hall, 1997.
[3] R. Bird and L. Meertens. Nested Datatypes. Proceedings, Mathematics of Program Construction, pp.
52–67, 1998.
[4] T. Coquand and G. Huet. The Calculus of Constructions. Information and Computation 76 (2-3), pp.
95–120, 1988.
[5] The Coq Proof Assistant. Available at coq.inria.fr
[6] P. Dybjer. A General Formulation of Simultaneous Inductive-Recursive Definitions in Type Theory.
Journal of Symbolic Logic 65 (2), pp. 525–549, 2000
[7] N. Ghani, M. Abbott, and T. Altenkirch. Containers - Constructing Strictly Positive Types. Theoretical
Computer Science 341 (1), pp. 3–27, 2005.
[8] N. Ghani and P. Johann. Foundations for Structured Programming with GADTs. Proceedings, Principles
of Programming Languages, pp. 297–308, 2008.
[9] N. Ghani and P. Johann and C. Fumex. Fibrational Induction Rules for Initial Algebras. Proceedings,
Computer Science Logic, pp. 336–350, 2010.
[10] C. Hermida and B. Jacobs. Structural Induction and Coinduction in a Fibrational Setting. Information
and Computation 145 (2), pp. 107–152, 1998.
[11] B. Jacobs. Comprehension Categories and the Semantics of Type Dependency. Theoretical Computer
Science 107, pp. 169–207, 1993.
[12] B. Jacobs. Categorical Logic and Type Theory. North Holland, 1999.
[13] P. Johann and N. Ghani. Initial Algebra Semantics is Enough! Proceedings, Typed Lambda Calculus
and Applications, pp. 207–222, 2007.
[14] F. W. Lawvere. Equality in Hyperdoctrines and Comprehension Scheme as an Adjoint Functor. Applications of Categorical Algebra, pp. 1–14, 1970.
[15] R. Matthes. An Induction Principle for Nested Datatypes in Intensional Type Theory. Journal of Functional Programming 19 (3&4), pp. 439–468, 2009.
[16] N. P. Mendler. Predicative type universes and primitive recursion. Proceedings, Logic in Computer
Science, pp. 173–184, 1991.
[17] P. Morris. Constructing Universes for Generic Programming. Dissertation, University of Nottingham,
2007.
[18] E. Moggi. Notations of Computation and Monads. Information and Computation 93 (1), pp. 55–92,
1991.
[19] B. Nordström, K. Petersson, and J. Smith. Programming in Martin-Löf ’s Type Theory. Oxford University Press, 1990.
[20] D. Pavlovič. Predicates and Fibrations. Dissertation, University of Utrecht, 1990.
[21] T. Sheard. Languages of the Future. SIGPLAN Notices 39 (10), pp. 116–119, 2004.
This work is licensed under the Creative Commons Attribution-NoDerivs License. To view
a copy of this license, visit http://creativecommons.org/licenses/by-nd/2.0/ or send a
letter to Creative Commons, 171 Second St, Suite 300, San Francisco, CA 94105, USA, or
Eisenacher Strasse 2, 10777 Berlin, Germany
| 6cs.PL
|
Logical Methods in Computer Science
Vol. 5 (4:2) 2009, pp. 1–48
www.lmcs-online.org
Submitted
Published
Mar. 3, 2009
Dec. 17, 2009
A STEP-INDEXED SEMANTICS OF IMPERATIVE OBJECTS ∗
CĂTĂLIN HRIŢCU a AND JAN SCHWINGHAMMER b
a
Department of Computer Science, Saarland University, Saarbrücken, Germany
e-mail address: [email protected]
b
Programming Systems Lab, Saarland University, Saarbrücken, Germany
e-mail address: [email protected]
Abstract. Step-indexed semantic interpretations of types were proposed as an alternative
to purely syntactic proofs of type safety using subject reduction. The types are interpreted
as sets of values indexed by the number of computation steps for which these values are
guaranteed to behave like proper elements of the type. Building on work by Ahmed, Appel
and others, we introduce a step-indexed semantics for the imperative object calculus of
Abadi and Cardelli. Providing a semantic account of this calculus using more ‘traditional’,
domain-theoretic approaches has proved challenging due to the combination of dynamically
allocated objects, higher-order store, and an expressive type system. Here we show that,
using step-indexing, one can interpret a rich type discipline with object types, subtyping,
recursive and bounded quantified types in the presence of state.
1. Introduction
The imperative object calculus of Abadi and Cardelli is a very small, yet very expressive
object-oriented language [2]. Despite the extreme simplicity of its syntax, the calculus
models many important concepts of object-oriented programming, as well as the often subtle
interaction between them. In particular it raises interesting and non-trivial questions with
respect to typing.
In contrast to the more common class-based object-oriented languages, in the imperative object calculus every object comes equipped with its own set of methods that can be
updated at run-time. As a consequence, the methods need to reside in the store, i.e., the
store is higher-order. Moreover, objects are allocated dynamically and aliasing is possible.
Dynamically-allocated, higher-order store is present in different forms in many practical
programming languages (e.g., pointers to functions in C and general references in SML),
but it considerably complicates the construction of adequate semantic models in which one
can reason about the behaviour of programs (as pointed out for instance by Reus [40]).
Purely syntactic arguments such as subject reduction suffice for proving the soundness
of traditional type systems. However, once such type systems are turned into powerful
1998 ACM Subject Classification: D.3.1, F.3.2.
Key words and phrases: Formal calculi, objects, type systems, programming language semantics.
∗
A preliminary version of this paper was presented at the International Workshop on Foundations of
Object-Oriented Languages (FOOL’08), 13 January 2008, San Francisco, California.
l
LOGICAL METHODS
IN COMPUTER SCIENCE
c
DOI:10.2168/LMCS-5 (4:2) 2009
CC
C. Hriţcu and J. Schwinghammer
Creative Commons
2
C. HRIŢCU AND J. SCHWINGHAMMER
specification languages, like the logic of objects of Abadi and Leino [4] or the hybrid type
system of Flanagan et al. [24], purely syntactic arguments seem no longer appropriate. The
meaning of assertions is no longer obvious, since they have to describe the code on the heap.
We believe that specifications of program behaviour should have a meaning independent of
the particular proof system on which syntactic preservation proofs rely, as also argued by
Benton [13] and by Reus and Schwinghammer [41].
In the case of specifications one would ideally prove soundness with respect to a semantic model that makes a clear distinction between semantic validity and derivability using
the syntactic rules. However, building such semantic models is challenging, and there is
currently no fully satisfactory semantic account of the imperative object calculus:
Denotational semantics: Domain-theoretic models have been employed in proving the
soundness of the logic of Abadi and Leino [41, 42]. However, the existing techniques fall
short of providing convincing models of typed objects: Reus and Streicher [42] consider an
untyped semantics, and the model presented by Reus and Schwinghammer [41] handles
neither second-order types, nor subtyping in depth. Due to the dynamically-allocated
higher-order store present in the imperative object calculus, the models rely on techniques
for recursively defined domains in functor categories [31, 36]. This makes them complex,
and establishing properties even for specific programs often requires a substantial effort.
Equational reasoning: Gordon et al. [25] develop reasoning principles for establishing
the contextual equivalence of untyped objects, and apply them to prove correctness of
a compiler optimization. Jeffrey and Rathke [29] consider a concurrent variant of the
calculus and characterize may-testing equivalence in terms of the trace sets generated
by a labeled transition system. In both cases the semantics is limited to equational
reasoning, i.e., establishing contextual equivalences between programs. In theory, this
can be used to verify a program by showing it equivalent to one that is trivially correct
and acts as a specification. However, this can be more cumbersome in practice than using
program logics, the established formalism for specifying and proving the correctness of
programs.
Translations: Abadi et al. [3] give an adequate encoding of the imperative object calculus into a lambda calculus with records, references, recursive and existential types and
subtyping. Together with an interpretation of this target language, an adequate model
for the imperative object calculus could, in principle, be obtained. However, we are
not aware of any worked-out adequate domain-theoretic models for general references
and impredicative second-order types. Even if such a model was given, it would still be
preferable to have a self-contained semantics for the object calculus, without the added
complexity of the (non-trivial) translation.
A solution to the problem of finding adequate models of objects could be the step-indexed
semantic models of types, introduced by Appel and McAllester [10] as an alternative to
subject reduction proofs. Such models are based directly on the operational semantics, and
are more easy to construct than the existing domain-theoretic models. The types are simply
interpreted as sets of syntactic values indexed by a number of computation steps. Intuitively,
a term belongs to a certain type if it behaves like an element of that type for any number of
steps. Every type is built as a sequence of increasingly accurate semantic approximations,
which allows one to easily deal with recursion. Type safety is an immediate consequence
of this interpretation of types, and the semantic counterparts of the usual typing rules
are proved as independent lemmas, either directly or by induction on the index. Ahmed
A STEP-INDEXED SEMANTICS OF IMPERATIVE OBJECTS
3
et al. [6, 9] successfully applied this generic technique to a lambda calculus with general
references, impredicative polymorphism and recursive types.
In this paper we further extend the semantics of Ahmed et al. with object types and
subtyping, and we use the resulting interpretation to prove the soundness of an expressive
type system for the imperative object calculus. The main contribution of our work is the
novel semantics of object types. We extend this semantics in two orthogonal ways. First,
we adapt it to self types, i.e., recursive object types that validate the usual subtyping rules
as well as strong typing rules with structural assumptions. Second, we study a natural
generalization of object types that results in simpler and more expressive typing rules.
Even though in this paper we are concerned with the safety of a type system, the
step-indexing technique is not restricted to types, and has already been used for equational
reasoning [5, 7, 10] and for proving the soundness of Hoare-style program logics of low-level
languages [13, 14]. We expect therefore that it will eventually become possible to use a stepindexed model to prove the soundness of more expressive program logics for the imperative
object calculus.
Outline. The next section introduces the syntax, operational semantics, and type system
that we consider for the imperative object calculus. In Section 3 we present a step-indexed
semantics for this calculus. In particular, we define the interpretations of types and establish their semantic properties. In Section 4 these properties are used to prove the soundness
of the type system. Section 5 studies self types, while Section 6 discusses a natural generalization of object types. Section 7 gives a comparison to related work and Section 8
concludes. The Appendix presents the proofs of the most interesting typing and subtyping
lemmas for object types, while an earlier technical report contains additional proofs [28].
2. The Imperative Object Calculus
We recall the syntax of the imperative object calculus with recursive and second-order types,
and introduce a small-step operational semantics for this calculus that is equivalent to the
big-step semantics given by Abadi and Cardelli [2].
2.1. Syntax. Let Var, TVar and Meth be pairwise disjoint, countably infinite sets of variables, type variables and method names, respectively. Let x, y range over Var, X, Y range
over TVar, and let m range over Meth. Figure 1 defines the syntax of the types and terms
of the imperative object calculus.
Objects are unordered collections of named methods, written as [md =ς(xd :A)bd ]d∈D .
In a method m = ς(x:A)b, ς is a binder that binds the ‘self’ argument x in the method
body b. The self argument can be used inside the method body for invoking the methods of
the containing object. Methods with arguments other than self can be obtained by having
a procedure as the method body. The methods of an object can be invoked or updated,
but no new methods can be added, and the existing methods cannot be deleted. The type
of objects with methods named md that return results of type Ad , for d in some set D, is
written as [md :νd Ad ]d∈D , where ν ∈ {◦, +, −} is a variance annotation that indicates if
the method is considered invoke-only (+), update-only (−), or if it may be used without
restriction (◦).
While procedural abstractions are sometimes defined in the imperative object calculus
using an additional let construct, we include them as primitives. We write procedures
4
C. HRIŢCU AND J. SCHWINGHAMMER
A, B, C ::= X | Top | Bot | A → B
(type expressions)
| [md :νd Ad ]d∈D | µ(X)A
| ∀(X6A)B | ∃(X6A)B
ν ::= ◦ | + | −
a, b ::= x
(variance annotations)
(variable)
| [md =ς(xd :A)bd ]d∈D
(object creation)
| a.m
(method invocation)
| a.m := ς(x:A)b
(method update)
| clone a
(shallow copy)
| λ(x:A)b
(procedure)
| ab
(application)
| foldA b
(recursive folding)
| unfoldA b
(recursive unfolding)
| Λ(X6A)b
(type abstraction)
| a[A]
(type application)
| pack X6A = C in a :B
(existential package)
| open a as X6A, x:B in b :C
(package opening)
Figure 1: Syntax of types and terms
with type A → B as λ(x:A)b and applications as a b, respectively. We use foldA and
unfoldA to denote the isomorphism between a recursive type µ(X)B and its unfolding
{{X 7→ µ(X)B}}(B). Finally, we consider bounded universal and existential types ∀(X6A)B
and ∃(X6A)B along with their introduction and elimination forms [21].
The set of free variables of a term a is denoted by fv(a), and similarly the free type
variables in a type A by fv(A). We identify types and terms up to the consistent renaming of
bound variables. We use {{t 7→ r}} to denote the singleton map that maps t to r. For a finite
map σ from variables to terms, σ(a) denotes the result of capture-avoiding substitution of
all x ∈ fv(a) ∩ dom(σ) by σ(x). The same notation is used for the substitution of type
variables. Generally, for any function f , the notation f [t := r] denotes the function that
maps t to r, and otherwise agrees with f .
2.2. Operational Semantics. Let Loc be a countably infinite set of heap locations ranged
over by l. We extend the set of terms by run-time representations of objects {md =ld }d∈D ,
associating heap locations to a set of method names. Values are given by the grammar:
v ∈ Val ::= {md =ld }d∈D | λ(x:A)b | foldA v | Λ(X6A)b | pack X6A = C in v :B
Apart from run-time objects, values consist of procedures, values of recursive type, type
abstractions and existential packages as in the call-by-value lambda calculus. We often only
consider terms and values without free variables, and denote the set of these closed terms
A STEP-INDEXED SEMANTICS OF IMPERATIVE OBJECTS
5
E[·] ::= [·] | E.m | E.m := ς(x:A)b | clone E | E b | v E | foldA E | unfoldA E
| E[A] | pack X6A = C in E :B | open E as X6A, x:B in b :C
Figure 2: Evaluation contexts
(Red-Obj)
hh, [md =ς(xd :A)bd ]d∈D i → hh [ld := λ(xd :A)bd ]d∈D , {md =ld }d∈D i
where ∀d ∈ D. ld ∈
/ dom(h)
(Red-Inv)
(Red-Upd)
(Red-Clone)
hh, {md =ld }d∈D .me i → hh, h(le ) {md =ld }d∈D i, if e∈D
hh, {md =ld }d∈D .me := ς(x:A)bi → hh [le := λ(x:A)b], {md =ld }d∈D i, if e∈D
hh, clone {md =ld }d∈D i → hh ld′ := h(ld ) d∈D , md =ld′ d∈D i
where ∀d ∈ D. ld′ ∈
/ dom(h)
hh, (λ(x:A)b) vi → hh, {{x 7→ v}}(b)i
(Red-Beta)
(Red-Unfold)
hh, unfoldA (foldB v)i → hh, vi
hh, (Λ(X6A)b)[B]i → hh, {{X 7→ B}}(b)i
(Red-TBeta)
(Red-Open)
hh, open v as X6A, x:B in b :Ci → hh, {{x 7→ v ′ , X 7→ C ′ }}(b)i
where v ≡ pack X ′ 6A′ = C ′ in v ′ :B ′
Figure 3: One-step reduction relation
and closed values by CTerm and CVal, respectively. A program is a closed term that does
not contain any locations, and we denote the set of all programs by Prog. A heap h is a
finite map from Loc to CVal1, and we write Heap for the set of all heaps.
Figure 2 defines the set of evaluation contexts, formalizing a left-to-right, call-by-value
strategy. We write E[a] for the term obtained by plugging a into the hole [·] of E. The
one-step reduction relation → is defined as the least relation on configurations hh, ai ∈
Heap × CTerm generated by the rules in Figure 3 and closed under the following context
rule:
hh, ai → hh ′ , a′ i =⇒ hh, E[a]i → hh ′ , E[a′ ]i
(Red-Ctx)
The methods are actually stored in the heap as procedures. Object construction allocates new heap storage for these procedures and returns a record of references to them
(Red-Obj). Upon method invocation the corresponding stored procedure is retrieved from
the heap and applied to the enclosing object (Red-Inv). The self parameter is thus passed
just like any other procedure argument. Identifying methods and procedures makes the
‘self-application’ semantics of method invocation explicit, while technically it allows us to
use the step-indexed model of Ahmed et al. [6, 9] with only few modifications.
While variables are immutable identifiers, methods can be updated destructively. Such
updates only modify the heap and leave the run-time object unchanged (Red-Upd). Object
1In fact, for the purpose of modelling the imperative object calculus it would suffice to regard procedures
as the only kind of storable value.
6
C. HRIŢCU AND J. SCHWINGHAMMER
Γ⊢A6B
Subtyping
Γ⊢A
Γ⊢A6A
(SubRefl)
(SubTop)
′
Γ⊢A
Γ ⊢ A 6 Top
(SubTrans)
(SubBot)
(SubProc)
′
Γ⊢A6A Γ⊢A 6B
Γ⊢A6B
Γ⊢A
Γ ⊢ Bot 6 A
(SubVar)
Γ1 , X6A, Γ2 ⊢ ⋄
Γ1 , X6A, Γ2 ⊢ X 6 A
Γ ⊢ A′ 6 A Γ ⊢ B 6 B ′
Γ ⊢ A → B 6 A′ → B ′
E⊆D
(SubObj)
∀e∈E. (νe ∈ {+, ◦} ⇒ Γ ⊢ Ae 6 Be )
∧ (νe ∈ {−, ◦} ⇒ Γ ⊢ Be 6 Ae )
Γ ⊢ [md :νd Ad ]d∈D 6 [me :νe Be ]e∈E
∀d ∈ D. νd = ◦ ∨ νd = νd′
(SubObjVar)
Γ ⊢ [md :νd Ad ]d∈D 6 [md :νd′ Ad ]d∈D
(SubRec)
Γ ⊢ µ(X)A
Γ ⊢ µ(Y )B Γ, Y 6Top, X6Y ⊢ A 6 B
Γ ⊢ µ(X)A 6 µ(Y )B
(SubUniv)
Γ ⊢ A′ 6 A Γ, X6A′ ⊢ B 6 B ′
Γ ⊢ ∀(X6A)B 6 ∀(X6A′ )B ′
(SubExist)
Γ ⊢ A 6 A′ Γ, X6A ⊢ B 6 B ′
Γ ⊢ ∃(X6A)B 6 ∃(X6A′ )B ′
Figure 4: Subtyping
cloning generates a shallow copy of an object in the heap (Red-Clone). The last four rules
in Figure 3 are as in the lambda calculus.
For k ∈ N, →k denotes the k-step reduction relation. We write hh, ai9 if the configuration hh, ai is irreducible (i.e., there exists no configuration hh ′ , a′ i such that hh, ai → hh ′ , a′ i).
Note that reduction is not deterministic, due to the arbitrarily chosen fresh locations
in (Red-Obj) and (Red-Clone). However, we still have that there is always at most one,
uniquely determined redex. This has the important consequence that the reduction order
is fixed. For example, if there is a reduction sequence beginning with a method invocation
and ending in an irreducible configuration: hh1 , a.mi →k hh2 , bi9, then this sequence can
be split into
hh1 , a.mi →i hh1′ , a′ .mi →k−i hh2 , bi
where hh1 , ai →i hh1′ , a′ i9 for some i ≥ 0. Similar decompositions into subsequences hold
for reductions starting from the other term forms.
It is easy to see that the operational semantics is independent of the type annotations
inside terms. Also the semantic types that we define in Section 3 will not depend on the
syntactic type expressions in the terms. In order to reduce the notational overhead and to
prevent confusion between the syntax and semantics of types we will omit type annotations
A STEP-INDEXED SEMANTICS OF IMPERATIVE OBJECTS
Γ⊢a:A
Subsumption and axioms
(Sub)
7
Γ⊢a:A Γ⊢A6B
Γ⊢a:B
(Var)
Γ1 , x:A, Γ2 ⊢ ⋄
Γ1 , x:A, Γ2 ⊢ x : A
Procedure types
(Lam)
Object types
Γ, x:A ⊢ b : B
Γ⊢a:B→A Γ⊢b:B
(App)
Γ ⊢ λ(x:A)b : A → B
Γ⊢ab:A
(where A ≡ [md :νd Ad ]d∈D )
(Obj)
∀d∈D. Γ, xd :A ⊢ bd : Ad
Γ ⊢ [md =ς(xd :A)bd ]d∈D : A
(Inv)
(Upd)
(Clone)
Γ⊢a:A
Γ ⊢ clone a : A
Γ ⊢ a : A e ∈ D νe ∈ {+, ◦}
Γ ⊢ a.me : Ae
Γ ⊢ a : A e ∈ D Γ, x:A ⊢ b : Ae νe ∈ {−, ◦}
Γ ⊢ a.me := ς(x:A)b : A
Recursive types
(Unfold)
Γ ⊢ a : µ(X)A
Γ ⊢ unfoldµ(X)A a : {{X 7→ µ(X)A}}(A)
(Fold)
Γ ⊢ a : {{X 7→ µ(X)A}}(A)
Γ ⊢ foldµ(X)A a : µ(X)A
Bounded quantified types
(TAbs)
Γ, X6A ⊢ b : B
Γ ⊢ Λ(X6A)b : ∀(X6A)B
(Pack)
(Open)
(TApp)
Γ ⊢ a : ∀(X6A)B Γ ⊢ A′ 6 A
Γ ⊢ a[A′ ] : {{X 7→ A′ }}(B)
Γ ⊢ C 6 A Γ ⊢ {{X 7→ C}}(a) : {{X 7→ C}}(B)
Γ ⊢ (pack X6A = C in a :B) : ∃(X6A)B
Γ ⊢ a : ∃(X6A)B Γ ⊢ C Γ, X6A, x:B ⊢ b : C
Γ ⊢ (open a as X6A, x:B in b :C) : C
Figure 5: Typing of terms
when presenting the step-indexed semantics. For example, instead of the type application
a[A] we will merely write a[].
2.3. Type System. The type system we consider features procedure, object, iso-recursive
and (impredicative, bounded) quantified types, as well as subtyping, and corresponds to
FOb<:µ from [2]. It is fairly standard and consists of four inductively defined typing
judgments:
• Γ ⊢ ⋄, describing well-formed typing contexts,
• Γ ⊢ A, defining well-formed types,
8
C. HRIŢCU AND J. SCHWINGHAMMER
• Γ ⊢ A 6 B, for subtyping between well-formed types, and
• Γ ⊢ a : A, for typing terms.
The typing context Γ is a list containing type bindings for the (term) variables x:A and
upper bounds for the type variables X6A. A typing context is well-formed if it does not
contain duplicate bindings for (term or type) variables and all types appearing in it are
well-formed. A type is well-formed with respect to a well-formed context Γ if all its type
variables appear in Γ.
Figure 4 defines the subtyping relation. For the object types it allows subtyping in
width: an object type with more methods is a subtype of an object type with fewer methods, as long as the types of the common methods agree. For the invoke-only (+) and
update-only methods (−) in object types, covariant respectively contravariant subtyping in
depth is allowed (SubObj). Furthermore, the unrestricted methods (◦) can be regarded, by
subtyping, as either invoke-only or update-only (SubObjVar). Since the annotations can
be conveniently chosen at creation time (Obj) this brings much flexibility. As explained by
Abadi and Cardelli [2], this allows us to distinguish in the type system between the invocations and updates done through the self argument, and the ones done from the outside.
The main idea is to type an object creation with an object type where all methods are
considered invariant, so that all invocations and updates through the self argument (internal) are allowed, but have to be type preserving. Then rules (Sub) and (SubObjVar) are
applied and some of the methods can become invoke-only, some others update-only. This
enables the subsequent weakening of the types of these methods using (SubObj). In effect,
this allows for safe and flexible subtyping of methods, at the price of restricting update
and invocation of the methods from the outside. Nevertheless, the internal updates and
invocations remain unrestricted.
Figure 5 defines the typing relation. The applicability of the rules for method invocation
(Inv), and for method update (Upd), depends on the variance annotation. Also notice that
only type-preserving updates are allowed in (Upd). Finally, it is important to note that we
do not give types to heap locations, since the type system is only used to check programs, and
programs do not contain locations. In contrast, a proof of type safety using the preservation
and progress properties would require the syntactic judgement to also depend on a heap
typing since partially evaluated terms would also need to be typed.
3. A Step-indexed Semantics of Objects
Modelling higher-order store is necessarily more involved than the treatment of firstorder storage since the semantic domains become mutually recursive. Recall that heaps
store values that may be procedures. These in turn can be modeled as functions that take
a value and the initial heap as input, and return a value and the possibly modified heap
upon termination. This suggests the following semantic domains for values and heaps, respectively:
DVal = (DHeaps × DVal ⇀ DHeaps × DVal ) + . . .
(3.1)
DHeaps = Loc ⇀fin DVal
A simple cardinality argument shows that there are no set-theoretic solutions (i.e., where
D ⇀ E denotes the set of all partial functions from D to E) satisfying the equations in (3.1).
A possible solution is to use a domain-theoretic approach, as done for the imperative object
calculus by Reus and Streicher [42], building on earlier work by Kamin and Reddy [30].
A STEP-INDEXED SEMANTICS OF IMPERATIVE OBJECTS
9
In a model of a typed calculus one also wants to interpret the types. But naively taking
a collection Type of subsets τ ⊆ DVal as interpretations of syntactic types does not work,
since values generally depend on the heap and a typed model should guarantee that all heap
access operations are type-correct. We are led to the following approach: first, in order to
ensure that updates are type-preserving, we also consider heap typings. Heap typings are
partial maps Ψ ∈ HeapTyping = Loc ⇀fin Type that track the set of values that may be
stored in each heap location. Second, the collection of types is refined to take heap typings
into account: a type will now consist of values paired with heap typings that describe the
necessary requirements on heaps. These ideas suggest that we take
Type = P(HeapTyping × DVal )
(3.2)
HeapTyping = Loc ⇀fin Type
Again, a cardinality argument shows the impossibility of defining these sets.
A final obstacle to modelling the object calculus, albeit independent of the higher-order
nature of heaps, is due to dynamic allocation in the heap. This results in heap typings that
may vary in the course of a computation, reflecting the changing ‘shape’ of the heap.
However, as is the case for many high-level languages, the object calculus is well-behaved
in this respect:
• inside the language, there is no possibility of deallocating heap locations; and
• only weak (i.e., type-preserving) updates are allowed.
As a consequence, extensions are the only changes that need to be considered for heap
typings. Intuitively, values that rely on heaps with typing Ψ will also be type-correct for
extended heaps, with an extended heap typing Ψ′ ⊒ Ψ. For this reason, semantic models
of dynamic allocation typically lend themselves to a Kripke-style presentation, where all
semantic entities are indexed by possible worlds drawn from the set of heap typings, partially
(pre-) ordered by heap typing extension [31, 33, 34, 37, 39].
Rather than trying to extend the already complex domain-theoretic models to heap
typings and dynamic allocation, we will use the step-indexing technique. Since this technique is based directly on the operational semantics, it provides an alternative that has less
mathematical overhead. In particular, there is no need to find semantic domains satisfying
(3.1); we can simply have DVal be the set of closed values and use syntactic procedures in
place of set-theoretic functions. Moreover, it is relatively easy to also model impredicative
second-order types in the step-indexed model of Ahmed et al. [6, 9], which is crucial for the
interpretation of object types we develop below. Although recently there has been progress
in finding domain-theoretic models of languages that combine references and polymorphic
types [15, 16, 17], the constructions are more involved.
The circularity in (3.2) is resolved by considering a stratification based on a notion of
‘k-step execution safety’. The central idea is that a term has type τ with approximation k
if this assumption cannot be proved wrong (in the sense of reaching a stuck state) in any
context by executing fewer than k steps. The key insight for constructing the sets satisfying
(3.2) is that all operations on the heap consume one step. Thus, in order to determine
whether a pair hΨ, vi, where Ψ is a heap typing and v a value, belongs to a type τ with
approximation k it is sufficient to know the types of the stored values on which v relies (as
recorded by Ψ) only up to level k − 1. The true meaning of types and heap typings is then
obtained by taking the limit over all such approximations.
For instance, if a heap typing Ψ asserts that a Bool-returning procedure is stored at
location l, i.e., Ψ(l) = [m:Bool] → Bool, then it is certainly not safe to assume that the
10
C. HRIŢCU AND J. SCHWINGHAMMER
pair hΨ, λ(y){m=l}.mi belongs to the type of Int-returning procedures. However, it is not
possible to contradict this assumption by taking only two reduction steps: the first step is
consumed by the beta reduction, the second one by the method selection {m=l}.m in the
procedure body, which involves a heap access. In this case, there are no steps left to observe
that the result of the computation is a boolean rather than an integer. Consequently, the
value λ(y){m=l}.m is in the type of Int-returning procedures for two computation steps,
even though it does not actually return an integer. One can of course distinguish such ‘false
positives’ by taking more reduction steps.
The preceding considerations are now formalized, building on the model originally developed by Ahmed et al. for an ML-like language with general references and impredicative
second-order types [6, 9]. Apart from some notational differences, the definitions in Section 3.1 are the same as in [6]. Section 3.2 adds subtyping, while Section 3.3 deals with
procedure types, and Section 3.4 revisits reference types. The semantics of object types is
presented in Section 3.5 and constitutes the main contribution of this paper. We further
deviate from [6] by adding bounds to the second-order types in Section 3.6, and by using
iso-recursive instead of equi-recursive types in Section 3.7.
3.1. The Semantic Model. To make the (circular) definition of types and heap typings
from (3.2) work, the step-indexed semantics considers triples with an additional natural
number component, representing the step index, rather than just pairs. First, we inductively define two families (PreTypek )k∈N of pre-types, and (HeapPreTypingk )k∈N of heap
pre-typings, by
τ ∈ PreType0 ⇔ τ = ∅
S
τ ∈ PreTypek+1 ⇔ τ ∈ P(N × ( j≤k HeapPreTypingj ) × CVal)
∧ ∀hj, Ψ, vi ∈ τ. j ≤ k ∧ Ψ ∈ HeapPreTypingj
where HeapPreTypingk = Loc ⇀fin PreTypek . That is, each τ ∈ PreTypek is a set of
triples hj, Ψ, vi where the set HeapPreTypingj from which the heap pre-typing Ψ is drawn
depends on the index j < k. Clearly PreTypek ⊆ PreTypek+1 and thus HeapPreTypingk ⊆
HeapPreTypingk+1 for all k. Now it is possible to set
S
τ ∈ PreType ⇔ τ ∈ P(N × ( j HeapPreTypingj ) × CVal)
∧ ∀hj, Ψ, vi ∈ τ. Ψ ∈ HeapPreTypingj
We call the elements of this set pre-types, rather than types, since there will be a further
condition that proper types must satisfy (this is done in Definition 3.4 below). From now
on, when writing hk, Ψ, vi, we always implicitly assume that Ψ ∈ HeapPreTypingk . By
HeapPreTyping we denote the set Loc ⇀fin PreType of finite maps into pre-types.
Each pre-type τ is a union of sets τk ∈ PreTypek where the index appearing in elements of τk is bounded by k. This is made explicit by the following notion of semantic
approximation and the stratification invariant below.
Definition 3.1 (Semantic approximation). For any pre-type τ we call ⌊τ ⌋k the k-th approximation of τ and define it as the subset containing all elements of τ that have an index
strictly less than k: ⌊τ ⌋k = {hj, Ψ, vi ∈ τ | j < k}. This definition is lifted pointwise to the
(partial) functions in HeapPreTyping: ⌊Ψ⌋k = λl ∈ dom(Ψ). ⌊Ψ(l)⌋k .
Proposition 3.2
S (Stratification). For all τ ∈ PreType and k ∈ N, ⌊τ ⌋k ∈ PreTypek .
Moreover, τ = k ⌊τ ⌋k .
A STEP-INDEXED SEMANTICS OF IMPERATIVE OBJECTS
11
So in particular, if hk, Ψ, vi ∈ τ and l ∈ dom(Ψ) then Ψ(l) ∈ PreTypej for some j ≤ k.
This is captured by the following ‘stratification invariant’, which will be satisfied by all
the constructions on (pre-) types, and which ensures the well-foundedness of the whole
construction:
Stratification invariant. For all pre-types τ , ⌊τ ⌋k+1 cannot depend on any
pre-type beyond approximation k.
As indicated above, in order to take dynamic allocation into account we consider a possible
worlds model. Intuitively we think of a pair (k, Ψ) as describing the state of a heap h,
where Ψ lists locations in h that are guaranteed to be allocated, and contains the types of
the stored values up to approximation k. In the course of a computation, there are three
different situations where the heap state changes:
• New objects are allocated on the heap, which is reflected by a heap pre-typing Ψ′ with
additional locations compared to Ψ. This operation does not affect any of the previously
stored objects, so Ψ′ will be an extension of Ψ.
• The program executes for k − j steps, for some j ≤ k, without accessing the heap.
This is reflected by a heap state (j, ⌊Ψ⌋j ) that ‘forgets’ that we have a more precise
approximation, and guarantees that the heap is safe only for j execution steps.
• The heap is updated, but in such a way that all typing guarantees of Ψ are preserved.
Thus updates will be reflected by an information forgetting extension, as in the previous
case. However, because of the step taken by the update itself, in this case we necessarily
have that j < k.
The following definition of state extension captures these possible evolutions of a state.
Definition 3.3 (State extension). State extension ⊑ is the relation on N × HeapPreTyping
defined by
(k, Ψ) ⊑ (j, Ψ′ ) ⇔ j ≤ k ∧ dom(Ψ) ⊆ dom(Ψ′ )
∧ ∀l ∈ dom(Ψ). Ψ′ j (l) = ⌊Ψ⌋j (l)
The step-indexing technique relies on the approximation of the ‘true’ set of values that
constitute a type, by all those values that behave accordingly unless a certain number of
computation steps are taken. Limiting the number of available steps, we will only be able to
make fewer distinctions. Moreover, if for instance a procedure relies on locations in the heap
as described by a state (k, Ψ), we can safely apply the procedure after further allocations.
In fact, if we are only interested in safely executing the procedure for j < k steps, a heap
described by state (j, ⌊Ψ⌋j ) will suffice. These conditions are captured precisely by state
extension, so we require our semantic types to be closed under state extension:
Definition 3.4 (Semantic types and heap typings). The set Type of semantic types is the
subset of PreType defined by
τ ∈ Type ⇔ ∀k, j ≥ 0. ∀Ψ, Ψ′ . ∀v ∈ CVal.
(k, Ψ) ⊑ (j, Ψ′ ) ∧ hk, Ψ, vi ∈ τ ⇒ hj, Ψ′ , vi ∈ τ
We also define the set HeapTyping = Loc ⇀fin Type of heap typings, ranged over by Ψ in
the following, as the subset of heap pre-typings that map to semantic types.
As explained by Ahmed [6], this structure may be viewed as an instance of Kripke
models of intuitionistic logic where states are the possible worlds, state extension is the
12
C. HRIŢCU AND J. SCHWINGHAMMER
reachability relation between worlds, and where closure under state extension corresponds
to Kripke monotonicity.
Next we define when a particular heap h conforms to the requirements expressed by a
heap typing Ψ. This is done with respect to an approximation index.
Definition 3.5 (Well-typed heap). A heap h is well-typed with respect to Ψ with approximation k, written as h :k Ψ, if dom(Ψ) ⊆ dom(h) and
∀j < k. ∀l ∈ dom(Ψ). hj, ⌊Ψ⌋j , h(l)i ∈ Ψ(l)
Semantic types only contain values, but we also need to associate types with terms that
are not values. We do this in two steps, first for closed terms, then for arbitrary ones. A
closed term has a certain type to approximation k with respect to some heap typing Ψ, if
in all heaps that are well-typed with respect to Ψ the term behaves like an element of the
type for k computation steps. In general, before reducing to a value the term will execute
for j steps, and possibly allocate some new heap locations in doing so. The state describing
the final heap will therefore be an extension of the state describing the initial heap, and it
only needs to be safe for the remaining k − j steps. Similarly, the final value needs to be in
the original type only for another k − j steps. The next definition makes this precise.
Definition 3.6 (Closed term has semantic type). We say that a closed term a has type τ
with respect to the state (k, Ψ), denoted as a :k,Ψ τ , if and only if
∀j < k, h, h ′ , b. (h :k Ψ ∧ hh, ai →j hh ′ , bi ∧ hh ′ , bi9)
⇒ ∃Ψ′ . (k, Ψ) ⊑ (k − j, Ψ′ ) ∧ h ′ :k−j Ψ′ ∧ hk − j, Ψ′ , bi ∈ τ
Even though the terms we evaluate are closed, when type-checking their subterms we
also have to reason about open terms. Typing open terms is done with respect to a semantic
type environment Σ that maps variables to semantic types. We reduce typing open terms to
typing their closed instances obtained by substituting all free variables with appropriately
typed, closed values. This is done by a value environment σ (a finite map from variables to
closed values) that agrees with the type environment.
Definition 3.7 (Value environment agrees with type environment). We say that value
environment σ agrees with semantic type environment Σ, with respect to the state (k, Ψ), if
∀x ∈ dom(Σ). σ(x) :k,Ψ Σ(x). We denote this by σ :k,Ψ Σ.
Definition 3.8 (Semantic typing judgement). We say that a term a (possibly with free
variables, but not containing locations), has type τ with respect to a semantic type environment Σ, written as Σ |= a : τ , if after substituting well-typed values for the free variables
of a, we obtain a closed term that has type τ for any number of computation steps. More
precisely:
Σ |= a : τ
⇔ fv(a) ⊆ dom(Σ) ∧ ∀k ≥ 0. ∀Ψ. ∀σ :k,Ψ Σ. σ(a) :k,Ψ τ
By construction, the semantic typing judgment enforces that all terms that are typable
with respect to it do not produce type errors when evaluated.
Definition 3.9 (Safe for k steps). We call a configuration hh, ai safe for k steps, if the
term a does not get stuck in less than k steps when evaluated in the heap h, i.e., we define
the set of all such configurations by
Safek = {hh, ai | ∀j < k. ∀h ′ , b. hh, ai →j hh ′ , bi ∧ hh ′ , bi9 ⇒ b ∈ Val}
A STEP-INDEXED SEMANTICS OF IMPERATIVE OBJECTS
13
Definition 3.10 (Safety).
T We call a configuration safe if it does not get stuck in any number
of steps, and let Safe = k∈N Safek .
Theorem 3.11 (Safety). For all programs a such that ∅ |= a : τ and for all heaps h we
have that hh, ai ∈ Safe.
Proof. One first easily shows that, if a :k,Ψ τ and h :k Ψ, then hh, ai ∈ Safek . The theorem
then follows by observing that any h is well-typed with respect to the empty heap typing,
to any approximation k.
This is much more direct than a subject reduction proof [46]. However, unlike with
subject reduction, the validity of the typing rules still needs to be proved with respect to
the semantics. We do this in two steps. In the remainder of this section we introduce
the specific semantic interpretations of types, and prove that they satisfy certain semantic
typing lemmas. These proofs are similar in spirit to proving the ‘fundamental theorem’ of
Kripke logical relations [32]. Then, in Section 4 we prove the soundness of the rules of the
initial type system with respect to these typing lemmas.
Even though the semantic typing lemmas are constructed so that they directly correspond to the rules of the original type system, there is a big difference between the two.
While the semantic typing lemmas allow us to logically derive valid semantic judgments
using other valid judgments as premises, the typing rules are just syntax that is used in the
inductive definitions of the typing and subtyping relations.
3.2. Subtyping. Since types in the step-indexed interpretation are sets (satisfying some
additional constraints), the natural subtyping relation is set inclusion. This subtyping
relation forms a complete lattice on semantic types, where infima and suprema are given
by set-theoretic intersections and unions, respectively. The least element is ⊥ = ∅, while
the greatest is
⊤ = {hj, Ψ, vi | j ∈ N, Ψ ∈ HeapTypingj , v ∈ CVal}.
Obviously ⊥ and ⊤ satisfy both the stratification invariant (i.e., they are pre-types) and
the closure under state extension condition, so they are indeed semantic types.
We can easily show the standard subsumption property
Lemma 3.12 (Subsumption). If Σ |= a : α and α ⊆ β then Σ |= a : β.
While it is very easy to define subtyping in this way, the interaction between subtyping
and the other features of the type system, in particular the object types, is far from trivial.
This point will be discussed further in Section 3.5.
3.3. Procedure Types. Intuitively, a procedure has type α → β for k computation steps
if, when applied to any well-typed argument of type α, it produces a result that has type β for
another k − 1 steps. This is because the procedure application itself takes one computation
step, and the only way to use a procedure is by applying it to some argument.
Additionally, we have to take into account that the procedure can also be applied after
some computation steps that extend the heap. So, for every j < k and for every heap
typing Ψ′ such that (k, Ψ) ⊑ (j, Ψ′ ), when applying the procedure to a value in type α
for j steps with respect to Ψ′ , the result must have type β for j steps with respect to Ψ′ .
This computational intuition nicely fits the possible worlds reading of procedure types as
intuitionistic implication.
14
C. HRIŢCU AND J. SCHWINGHAMMER
Σ[x := α] |= b : β =⇒ Σ |= λx. b : α → β
(Σ |= a : β → α ∧ Σ |= b : β =⇒ Σ |= a b : α
α′ ⊆ α ∧ β ⊆ β ′ =⇒ α → β ⊆ α′ → β ′
(SemLam)
(SemApp)
(SemSubProc)
Figure 6: Typing lemmas: procedure types
Definition 3.13 (Procedure types). If α and β are semantic types, then α → β consists of
those triples hk, Ψ, λx. bi such that for all j < k, heap typings Ψ′ and closed values v:
((k, Ψ) ⊑ (j, Ψ′ ) ∧ hj, Ψ′ , vi ∈ α) ⇒ {{x 7→ v}}(b) :j,Ψ′ β
Proposition 3.14. If α and β are semantic types, then α → β is also a semantic type.
Figure 6 contains the semantic typing lemmas associated with procedure types. The
procedure type constructor is of course contravariant in the argument type and covariant
in the result type.
Lemma 3.15 (Procedure types). The three semantic typing lemmas shown in Figure 6 are
valid implications.
Proof sketch. The validity of (SemApp) and (SemLam) is proved in [6]. Verifying (SemSubProc) is simply a matter of unfolding the definitions.
3.4. Revisiting Reference Types. While our calculus does not have references syntactically, we will use the model of references from [6, 9] in our construction underlying object
types. In order to interpret the variance annotations in object types, we additionally introduce readable reference types and writable reference types, with covariant and contravariant
subtyping, respectively [35, 43].
A heap typing associates with each allocated location the precise type that can be used
when reading from it and writing to it. So all heap locations support both reading and
writing at a certain type, and we do not have read-only or write-only locations. Intuitively,
for the readable reference types and the writable ones the precise type of the locations is
only partially known, so that without additional information only one of the two operations
is safe at a meaningful type.
We first recall the definition of reference types from [6, 9].
Definition 3.16 (Reference types). If τ is a semantic type then
ref◦ τ = {hk, Ψ, li | ⌊Ψ(l)⌋k = ⌊τ ⌋k }
According to this definition, a location l has type ref◦ τ if the type associated with l
by the heap typing Ψ is approximately τ . Semantic approximation is used to satisfy the
stratification invariant, and is operationally justified by the fact that reading from a location
or writing to it takes one computation step. So, l has type ref◦ τ for k steps if all values
that are read from l or written to l have type τ for k − 1 steps.
The readable reference type ref+ τ is similar to ref◦ τ , but poses less constraints on
the heap typing Ψ: it only requires that Ψ(l) is a subtype of τ , as before up to some
approximation.
A STEP-INDEXED SEMANTICS OF IMPERATIVE OBJECTS
15
α⊆β
=⇒
ref+ α ⊆ ref+ β
(SemSubCovRef)
β⊆α
=⇒
ref− α ⊆ ref− β
(SemSubConRef)
ref◦ α ⊆ refν α, where ν ∈ {◦, +, −}
(SemSubVarRef)
Figure 7: Subtyping reference types
Definition 3.17 (Readable reference types). If τ is a semantic type then
ref+ τ = {hk, Ψ, li | ⌊Ψ⌋k (l) ⊆ ⌊τ ⌋k }
The value stored at location l also has type τ by subsumption, and therefore can be
read and safely used as a value of type τ . However, the true type of location l is in general
unknown, so writing any value to it could be unsafe (the true type of l might be the empty
type ⊥). Nevertheless, knowing that a location has type ref+ τ does not mean that we
cannot write to it: it simply means that we do not know the type of the values that can be
written to it, so in the absence of further information no writing can be guaranteed to be
type safe2.
Dually, the type ref− τ of writable references contains all those locations l whose type
associated by Ψ is a supertype of τ .
Definition 3.18 (Writable reference types). If τ is a semantic type then
ref− τ = {hk, Ψ, li | ⌊τ ⌋k ⊆ ⌊Ψ⌋k (l)}
We can safely write a value of type τ to a location of type ref− τ , since this value also
has the real type of location l by subsumption. However, the real type of such locations
can be arbitrarily general. In particular it can be ⊤, the type of all values. Thus a location
about which we only know that it has type ref− τ can only be read safely at type ⊤.
With these definitions in place, the usual reference type from Definition 3.16 can be
recovered as the intersection of a readable and a writable reference type:
ref◦ τ = ref+ τ ∩ ref− τ
Hence ref+ τ and ref− τ are both supertypes of ref◦ τ . It can also be easily shown that the
readable reference type constructor is covariant, the writable reference type constructor is
contravariant (Figure 7), while the usual reference types are obviously invariant. For a
variance annotation ν ∈ {◦, +, −} we use refν to stand for the reference type constructor
with this variance.
Note that, strictly speaking, the set refν τ is not a semantic type since for our calculus
locations are not values (although locations appear in object values {md =ld }d∈D ; see Section 2.2). In fact, the definition of object types (Definition 3.20 in the next section) will not
depend on refν τ being a semantic type. However, in order for the object type constructor
to yield semantic types, it is crucial that refν τ is closed under state extension.
Proposition 3.19. If τ is a semantic type, then refν τ is closed under state extension.
2This is conceptually different from the immutable reference types modeled in [6] using singleton types.
16
C. HRIŢCU AND J. SCHWINGHAMMER
3.5. Object Types. Giving a semantics to object types is much more challenging than for
the other types. The typing rules from Section 2 indicate why this is the case. First, an
adequate interpretation of object types must permit subtyping both in width and in depth,
taking the variance annotations into account. Second, in contrast to all the other types we
consider that have just a single elimination rule, once constructed, objects support three
different operations: invocation, update, and cloning. The definition of object types must
ensure the consistent use of an object through all possible future operations. That is, all
the requirements on which invocation, update or cloning rely must already be established
at object creation time.
Before defining the object types, it is instructive to consider some simpler variants that
do not fulfill all the requirements we have for object types.
Our decision to store methods in the heap as procedures, together with the ‘selfapplication’ semantics of method invocation (Red-Inv in Figure 3), suggest that object
types are somewhat similar to recursive types of records of references holding procedures
that take the enclosing record as argument:
?
[md : τd ]d∈D = µ(α).{md : ref◦ (α → τd )}d∈D
However, the invariance of the reference type constructor blocks any form of subtyping,
even in width. A look at typing rules for subtyping recursive types, such as Cardelli’s
Amber rule [20] (which appears as rule SubRec in Figure 4), suggests that the position
of the recursion variable should be covariant. For instance, when attempting to establish
the subtyping [m1 : τ1 , m2 : τ2 ] ⊆ [m1 : τ1 ] by the Amber rule one needs to show that
ref◦ (α → τ1 ) ⊆ ref◦ (β → τ1 ), for any α and β such that α ⊆ β. Clearly this does not
hold. Even in a simpler setting without the reference types (e.g., for the functional object
calculus) the contravariance of the procedure type constructor in its first argument would
cause subtyping to fail.
A combination of type recursion and an existential quantifier that uses the recursion
variable as bound would allow us to enforce covariance for the positions of the recursion
variable, and thus have subtyping in width:
?
[md : τd ]d∈D = µ(α).∃α′ ⊆α.{md : ref◦ (α′ → τd )}d∈D
Intuitively α′ can be viewed as the ‘true’ (i.e., most precise) type of the object, while α is
a more general type that can be given to it by subtyping. This is essentially the idea of the
encodings of object types explored by Abadi et al. [2, 3].
For subtyping in depth with respect to the variance annotations we simply use the
readable and writable reference types we defined in the previous section:
?
[md :νd τd ]d∈D = µ(α).∃α′ ⊆α.{md : refνd (α′ → τd )}d∈D
Still, by keeping α′ abstract, neither the typing rule for method invocation (Inv in Figure 5),
nor the one for object cloning (Clone) is validated.
By explicitly enforcing in the definition of object types that the object value itself in
fact belongs to this existentially quantified α′ , the assumptions become sufficiently strong
to repair the invocation case. This is consistent with seeing α′ as the ‘true’ type of the
object. Semantically, we can express this using an intersection of types:
?
[md :νd τd ]d∈D = µ(α).∃α′ ⊆α.({md : refνd (α′ → τd )}d∈D ∩ α′ )
A STEP-INDEXED SEMANTICS OF IMPERATIVE OBJECTS
17
Forcing not only the current object value to be in α′ , but also all the ‘sufficiently similar’
values (maybe not even created yet), covers the case of cloning. The following definition
formalizes this construction.
Definition 3.20 (Object types). Let α = [md :νd τd ]d∈D be defined as the set of all triples
hk, Ψ, {me =le }e∈E i such that D ⊆ E and
∃α′ . α′ ∈ Type ∧ α′ k ⊆ ⌊α⌋k
(Obj-1)
∧ (∀d ∈ D. hk, Ψ, ld i ∈ refνd (α′ → τd ))
∧ (∀j < k. ∀Ψ′ . ∀ me =le′ e∈E .
(k, Ψ) ⊑ (j, Ψ′ ) ∧ (∀e ∈ E. Ψ′ j (le′ ) = ⌊Ψ⌋j (le ))
⇒ hj, Ψ′ j , me =le′ e∈E i ∈ α′ )
(Obj-2)
(Obj-3)
The condition stating that D ⊆ E ensures that all values in an object type provide
at least the required methods listed by this type, but can also provide more. Clearly this
is necessary for subtyping in width. Condition (Obj-1) postulates the existence of a more
specific type α′ , the ‘true’ type of the object {me =le }e∈E (up to approximation k), and the
subsequent conditions are all stated in terms of α′ rather than α. Condition (Obj-2) states
the requirements for the methods in terms of the reference type constructors introduced in
Section 3.4. Since the existentially quantified α′ might equal α, one must take care that
(Obj-2) does not introduce a circularity. However, due to the use of approximation in the
definition of the reference type constructors, the condition only depends on ⌊α′ ⌋k , rather
than α′ . This will ensure the well-foundedness of the construction.
As explained above, in order to invoke methods we must know that {me =le }e∈E belongs
to the more specific type α′ for j < k steps (which suffices since application consumes a
step). In the particular case where Ψ′ is Ψ and {me =le′ }e∈E is {me =le }e∈E condition (Obj3) states exactly this. We need the more general formulation in order to ensure that the
clones of the considered object also belong to the same type α′ . Therefore we enforce that no
matter how an object value {me =le′ }e∈E is constructed, it belongs to type α′ provided that
it satisfies the same typing assumptions as {me =le }e∈E , with respect to a possibly extended
heap typing Ψ′ . Allowing for state extension is necessary since cloning itself allocates new
locations not present in the original Ψ, and also because cloning can be performed after
some intermediate computation steps that result in further allocations.
We show that this definition of object types actually makes sense, in that it defines a
semantic type. This is not immediately obvious because of the recursion.
Proposition 3.21. If τd ∈ Type for all d ∈ D, then we also have that [md :νd τd ]d∈D ∈ Type.
Proof sketch. We must show (1) that [md :νd τd ]d∈D is well-defined, i.e., that the recursive
definition is well-founded, and (2) that it is closed under state extension.
To prove the well-definedness one can use general results about recursive types in stepindexed semantics [10], since the
S object type constructor is ‘contractive’. Alternatively,
from the observation that
τ
=
⌋k for all types τ , it
argue that
k ⌊τ
suffices to directly
Definition 3.20 defines [md :νd τd ]d∈D k only in terms of [md :νd τd ]d∈D j for j < k. The
closure under state extension follows from the corresponding property of the types α′ → τd
(Proposition 3.14) and of the sets refνd (α′ → τd ) (Proposition 3.19), and from the transitivity
of state extension.
18
C. HRIŢCU AND J. SCHWINGHAMMER
Let α = [md :νd τd ]d∈D .
(∀d ∈ D. Σ[xd := α] |= bd : τd )
=⇒
Σ |= [md =ς(xd )bd ]d∈D : α
(SemObj)
(Σ |= a : α ∧ e ∈ D ∧ νe ∈ {+, ◦})
=⇒
Σ |= a.me : τe
(SemInv)
∧ Σ[x := α] |= b : τe )
=⇒
Σ |= a.me := ς(x)b : α
Σ |= a : α
=⇒
Σ |= clone a : α
(Σ |= a : α ∧ e ∈ D ∧ νe ∈ {−, ◦}
(SemUpd)
(E ⊆ D ∧ (∀e ∈ E. νe ∈ {+, ◦} ⇒ αe ⊆ βe )
(SemClone)
(SemSubObj)
∧ (∀e ∈ E. νd ∈ {−, ◦} ⇒ βe ⊆ αe ))
=⇒
[md :νd αd ]d∈D ⊆ [me :νe βe ]e∈E
νd′ )
=⇒
[md :νd αd ]d∈D ⊆ [md :νd′ αd ]d∈D
(SemSubObjVar)
(∀d ∈ D. νd = ◦ ∨ νd =
Figure 8: Typing lemmas: object types
Figure 8 presents the semantic typing and subtyping lemmas for object types.
Lemma 3.22 (Object types). All the semantic typing lemmas shown in Figure 8 are valid
implications.
Proof sketch. The semantic typing lemmas are proved independently. We sketch this for
(SemObj). A detailed proof, as well as the proofs of the other typing lemmas are given in
the Appendix.
For α = [md :νd τd ]d∈D and assuming Σ[xd := α] |= bd : τd for all d ∈ D, we must
show that Σ |= [md =ς(xd )bd ]d∈D : α. So let k ≥ 0, σ and Ψ be such that σ :k,Ψ Σ. By
Definition 3.8 (Semantic typing judgement) we must prove that σ([md =ς(xd )bd ]d∈D ) :k,Ψ α,
or equivalently (after suitable α-renaming), that [md =ς(xd )σ(bd )]d∈D :k,Ψ α holds. Now let
h, h′ and b′ be such that h :k Ψ and
hh, [md =ς(xd )σ(bd )]d∈D i →j hh′ , b′ i
for some j < k, and assume that hh′ , b′ i is irreducible. From the operational semantics it is
clear that j = 1, b′ ≡ {md =ld }d∈D and that, for some locations ld ∈
/ dom(h),
h′ = h [ld := λ(xd )σ(bd )]d∈D
Choosing Ψ′ = Ψ [ld := (α → τd )]d∈D k−1 it is easily seen that (k, Ψ) ⊑ (k − 1, Ψ′ ). Furthermore, from the hypothesis by (SemLam) we have that Σ |= λ(xd )bd : α → τd for all
d ∈ D. From this and the assumption that h :k Ψ it follows that h′ :k−1 Ψ′ .
By Definition 3.6 it remains to establish that hk − 1, Ψ′ , {md =ld }d∈D i ∈ α. This is
achieved by proving the following more general claim by induction on j0 :
Claim 3.23. For all j0 ≥ 0, Ψ∗ and {md =ld∗ }d∈D we have that
(k − 1, Ψ′ ) ⊑ (j0 , Ψ∗ ) ∧ (∀d∈D. ⌊Ψ∗ ⌋j0 (ld∗ ) = Ψ′ j0 (ld )) ⇒ hj0 , ⌊Ψ∗ ⌋j0 , {md =ld∗ }d∈D i ∈ α
The key step is in choosing α′ equal to ⌊α⌋j0 , then verifying the three conditions of Definition 3.20 (Object types), where the inductive hypothesis is used for showing (Obj-3).
A STEP-INDEXED SEMANTICS OF IMPERATIVE OBJECTS
19
Remark 3.24. In the above proof, establishing hk−1, Ψ′ , {md =ld }d∈D i ∈ α directly does
not seem possible, and the generalization to Claim 3.23 arises naturally from a failed proof
attempt: in order to prove the three conditions of Definition 3.20 a sensible choice for α′ is
α, and for E is D, after which (Obj-1) and (Obj-2) follow easily. But (Obj-3) requires
us to show that hj, ⌊Ψ′′ ⌋j , {md =ld′ }d∈D i ∈ α, for any j < k, any ld′ , and any extension Ψ′′
of Ψ′ with ⌊Ψ′′ ⌋j (ld′ ) = ⌊Ψ′ ⌋j (ld ). This is just what Claim 3.23 states.
The fact that there is an inductive argument hidden in this proof does not come as a
surprise: the induction on the step index j0 resolves the recursion that is inherent to objects
due to the self application semantics of method invocation.
3.6. Bounded Quantified Types. Impredicative quantified types were previously studied in a step-indexed setting by Ahmed et al. [6, 9] for a lambda-calculus with general
references, and we follow their presentation. However, unlike in the work of Ahmed et al.
our quantifiers have bounds, and we are also studying subtyping. It is important to note
that the impredicative second-order types were the reason why a semantic stratification of
types was needed in the presence of general references [6], as opposed to a syntactic one
based on the nesting of reference types [8]. In the setting we consider in this paper we need
the semantic stratification not only to explicitly accommodate quantified types, but also
because our interpretation of object types uses existential types implicitly.
As in Appel and McAllester’s work [10], a type constructor F (i.e., a function from
semantic types to semantic types) is non-expansive if in order to determine whether a term
has type F (τ ) with approximation k, it suffices to know the type τ only to approximation
k. As we will later show (Lemma 3.33), all the type constructors we define in this paper
are non-expansive.
Definition 3.25 (Non-expansiveness). A type constructor F : Type → Type is nonexpansive if for all types τ and for all k ≥ 0 we have that ⌊F (τ )⌋k = ⌊F (⌊τ ⌋k )⌋k .
The definitions of second-order types require that ∀ and ∃ are only applied to nonexpansive type constructors. The non-expansiveness condition ensures that in order to
determine level k of a universal or existential type, quantification over the types in PreTypek
suffices. This helps avoid the circularity that is otherwise introduced by the impredicative
quantification.
Definition 3.26 (Bounded universal types). If F : Type → Type is non-expansive and
α ∈ Type, then we define ∀α F by hk, Ψ, Λ. ai ∈ ∀α F if and only if
∀j, Ψ′ . ∀τ. (k, Ψ) ⊑ (j, Ψ′ ) ∧ τ ∈ Type ∧ ⌊τ ⌋j ⊆ ⌊α⌋j ⇒ ∀i < j. a :i,⌊Ψ′ ⌋i F (τ )
Definition 3.27 (Bounded existential types). For all non-expansive F : Type → Type and
α ∈ Type, the set ∃α F is defined by hk, Ψ, pack vi ∈ ∃α F if and only if
∃τ.τ ∈ Type ∧ ⌊τ ⌋k ⊆ ⌊α⌋k ∧ ∀j < k. hj, ⌊Ψ⌋j , vi ∈ F (τ )
Proposition 3.28. If α ∈ Type and F : Type → Type is non-expansive, then ∀α F and ∃α F
are also types.
Proof sketch. The proofs are minor modifications of those given in [6], to additionally take
the bounds into account.
20
C. HRIŢCU AND J. SCHWINGHAMMER
For all non-expansive F, G : Type → Type,
(∀τ ∈ Type. τ ⊆ α ⇒ Σ |= a : F (τ ))
=⇒
Σ |= Λ. a : ∀α F
(SemTAbs)
(Σ |= a : ∀α F ∧ τ ∈ Type ∧ τ ⊆ α)
=⇒
Σ |= a[] : F (τ )
(SemTApp)
(∃τ ∈ Type. τ ⊆ α ∧ Σ |= a : F (τ ))
=⇒
Σ |= pack a : ∃α F
(SemPack)
(Σ |= a : ∃α F ∧ ∀τ ∈Type.
(SemOpen)
τ ⊆ α ⇒ Σ[x := F (τ )] |= b : β)
=⇒
Σ |= open a as x in b : β
(β ⊆ α ∧ ∀τ ∈ Type. τ ⊆ β ⇒ F (τ ) ⊆ G(τ ))
=⇒
∀α F ⊆ ∀β G
(SemSubUniv)
(α ⊆ β ∧ ∀τ ∈ Type. τ ⊆ α ⇒ F (τ ) ⊆ G(τ ))
=⇒
∃α F ⊆ ∃β G
(SemSubExist)
Figure 9: Typing lemmas: bounded quantified types
Lemma 3.29 (Bounded quantified types). All the semantic typing lemmas shown in Figure 9 are valid implications.
Proof sketch. The first four implications are proved as in [6]; the additional precondition
τ ⊆ α in (SemTApp) and (SemPack) serves to establish the requirements for the bounds.
The two subtyping lemmas (SemSubUniv) and (SemSubExist) are easily proved by just
unfolding the definitions.
3.7. Recursive Types. In contrast to most previous work on step-indexed models, we
consider iso-recursive rather than equi-recursive types, so folds and unfolds are explicit in our
syntax and consume computation steps. Iso-recursive types have been previously considered
by Ahmed for a step-indexed relational model of the lambda calculus [7]. Iso-recursion is
simpler, and sufficient for our purpose. As a consequence, we require type constructors to
be only non-expansive, as opposed to the stronger ‘contractiveness’ requirement [10].
Definition 3.30 (Recursive types). Let F : Type → Type be a non-expansive function. We
define the set µF by
hk, Ψ, fold vi ∈ µF ⇔ ∀j < k. hj, Ψ′ , vi ∈ F (µF )
Proposition 3.31. For all non-expansive F : Type → Type, µF ∈ Type is well-defined.
Proof sketch. The well-definedness follows from the observation that ⌊µF ⌋k is defined only
in terms of ⌊F (µF )⌋j for j < k, which by the non-expansiveness of F means that ⌊µF ⌋k
relies only on ⌊µF ⌋j . The closure under state extension is established by an induction,
proving that for each k ≥ 0, ⌊µF ⌋k ∈ Type.
Figure 10 presents the semantic typing lemmas for recursive types. As a consequence,
we have the expected fixed point property |= a : F (µF ) ⇔ |= fold a : µF .
Lemma 3.32 (Recursive types). All the semantic typing lemmas shown in Figure 10 are
valid implications.
Proof sketch. The validity of (SemFold) and (SemUnfold) are easy consequences of Definition 3.30. For (SemSubRec), one shows by induction on k that ⌊µF ⌋k ⊆ ⌊µG⌋k , using
the precondition of the rule and the non-expansiveness of F and G.
A STEP-INDEXED SEMANTICS OF IMPERATIVE OBJECTS
21
For all non-expansive F, G : Type → Type,
Σ |= a : µF =⇒ Σ |= unfold a : F (µF )
(SemUnfold)
Σ |= a : F (µF ) =⇒ Σ |= fold a : µF
(SemFold)
(∀α, β. α ⊆ β ⇒ F (α) ⊆ G(β)) =⇒ µF ⊆ µG
(SemSubRec)
Figure 10: Typing lemmas: recursive types
i
h
y
q
[md :νd Ad ]d∈D η = md :νd JAd Kη
JXKη = η(X)
JBotKη = ⊥
JTopKη = ⊤
d∈D
Jµ(X)AKη = µ(λα∈Type. JAKη[X:=α] )
JA → BKη = JAKη → JBKη
J∀(X6A)BKη = ∀JAKη (λα∈Type. JBKη[X:=α] )
J∃(X6A)BKη = ∃JAKη (λα∈Type. JBKη[X:=α] )
Figure 11: Interpretation of types
Lemma 3.33 (Non-expansiveness). All the considered type constructors are non-expansive.
Proof sketch. It is easily seen that the definition of ⌊α → β⌋k uses only ⌊α⌋j and ⌊β⌋j for
j < k, and therefore that ⌊α → β⌋k = ⌊⌊α⌋k → ⌊β⌋k ⌋k . A similar statement holds for the
k-th approximation of quantified types ∀α F and ∃α F , since their definition only depends
on ⌊α⌋j and ⌊F ⌋j for j < k. In the case of object and recursive types, the properties
[md :νd τd ]d∈D k = [md :νd ⌊τd ⌋k ]d∈D k and ⌊µF ⌋k = ⌊µ ⌊F ⌋k ⌋k can be established by
induction on k, using the non-expansiveness of F in the latter case.
4. Semantic Soundness
In order to prove that well-typed terms are safe to evaluate we relate the syntactic types
to their semantic counterparts, and then use the fact that the semantic typing judgement
enforces safety by construction (Theorem 3.11). This approach is standard in denotational
semantics. In fact, none of the main statements or proofs in this section mentions stepindices explicitly.
Definition 4.1 (Interpretation of types and typing contexts). Let η be a total function
from type variables to semantic types.
(1) The interpretation JAKη of a type A is given by the structurally recursive meaning
function defined in Figure 11.
(2) The interpretation of a well-formed typing context Γ with respect to η is given by the
function that maps x to JAKη , for every x:A ∈ Γ.
Note that in Figure 11 the type constructors used on the left-hand sides of the equations
are simply syntax, while those on the right hand-sides refer to the corresponding semantic
constructions, as defined in the previous section.
Recall that non-expansiveness is a necessary precondition for some of the semantic
typing lemmas. In particular, the well-definedness of JAKη depends on non-expansiveness,
22
C. HRIŢCU AND J. SCHWINGHAMMER
due to the use of µ, ∀(·) and ∃(·) in Figure 11. So we begin by showing that the interpretation
of types is a non-expansive map.
Lemma 4.2 (Non-expansiveness). JAKη is non-expansive in η.
k
k
j
j
holds by induction on the structure of
Proof sketch. We show that JAKη = JAK⌊η⌋
k k
k
A, relying on Lemma 3.33 for the non-expansiveness of the semantic type constructions.
Definition 4.3 (η |= Γ). Let Γ be a well-formed typing context. We say that η satisfies
Γ, written as η |= Γ, if η(X) ⊆ JAKη holds for all X6A appearing in Γ.
We show the soundness of the subtyping relation.
Lemma 4.4 (Soundness of subtyping). If Γ ⊢ A 6 B and η |= Γ then JAKη ⊆ JBKη .
Proof sketch. By induction on the derivation of Γ ⊢ A 6 B and case analysis on the
last applied rule. Each case is immediately reduced to one of the subtyping lemmas from
Section 3.
Finally, we prove the semantic soundness of the syntactic type system with respect to
the model.
Theorem 4.5 (Semantic soundness). Whenever Γ ⊢ a : A and η |= Γ it follows that
JΓKη |= a : JAKη .
Proof sketch. By induction on the derivation of Γ ⊢ a : A and case analysis on the last rule
applied. Each case is easily reduced to one of the semantic typing lemmas from Section 3,
using a standard type substitution lemma for derivations ending with an application of
(Fold), (Unfold), (TApp), or (Pack).
By Theorems 4.5 (Semantic soundness) and 3.11 (Safety), we have a proof of safety for
the type system from Section 2.3.
Corollary 4.6 (Type safety). Well-typed terms are safe to evaluate.
5. Self Types
Self types have been proposed by Abadi and Cardelli [2] as a means to reconcile recursive
object types with ‘proper’ subtyping. Self types are interesting because they allow us to
type methods that return the possibly modified host object, or a clone of it. For instance,
a type of list nodes, with a filter method that produces the sublist of all elements satisfying
a given predicate, is
ListA = Obj(X)[hd :◦ A, tl :◦ X+Unit, filter :+ (A → Bool) → X, . . .]
Note that the similar recursive type
µ(X)[hd :◦ A, tl :◦ X+Unit, filter :+ (A → Bool) → X, . . .]
does not satisfy the usual subtyping for object types because of the invariance of the hd
and tl fields.
A STEP-INDEXED SEMANTICS OF IMPERATIVE OBJECTS
23
Let α = [md :νd Fd ]d∈D and β = [me :νe Ge ]e∈E with E ⊆ D.
(∀d ∈ D. Σ[xd := α] |= bd : Fd (α))
=⇒
Σ |= [md =ς(xd )bd ]d∈D : α
(SemObj-Self)
(Σ |= a : α ∧ e ∈ D ∧ νe ∈ {+, ◦})
=⇒
Σ |= a.me : Fe (α)
(Σ |= a : α ∧ e ∈ D ∧ νe ∈ {−, ◦}
(SemInv-Self)
(SemUpd-Self)
∧ ∀ξ ⊆ α. Σ[x := ξ] |= b : Fe (ξ))
=⇒
Σ |= a.me := ς(x)b : α
Σ |= a : α
=⇒
Σ |= clone a : α
(∀e ∈ E. (νe ∈ {+, ◦} ⇒ ∀ξ ⊆ α. Fe (ξ) ⊆ Ge (ξ))
(SemClone-Self)
(SemSubObj-Self)
∧ (νe ∈ {−, ◦} ⇒ ∀ξ ⊆ α. Ge (ξ) ⊆ Fe (ξ)))
=⇒
α⊆β
(∀d ∈ D. νd = ◦ ∨ νd = νd′ )
=⇒
[md :νd Fd ]d∈D ⊆ [md :νd′ Fd ]d∈D
(SemSubObjVar-Self)
Figure 12: Typing lemmas: self types
5.1. Semantics of Self Types. Abadi and Cardelli [2, Ch. 15] show how self types can be
understood in terms of recursive and existentially quantified object types via an encoding.
More precisely, the type Obj(X)[md :νd Bd ]d∈D where X may occur positively in Bd , stands
for the recursive type µ(Y )∃(X6Y )[md :νd Bd ]d∈D . The bounded existential quantifier introduced by this encoding gives rise to the desired subtyping in width and depth, despite
the type recursion.
Since our type system features recursive and bounded existential types, self types could
be accommodated via this encoding. However a treatment of self types can be achieved
even more directly, without relying on the encoding. In fact, almost everything is in place
already: recall that the semantics of object types (Definition 3.20) employs recursion and
an existential quantification to refer to the ‘true’ type of an object. Condition Obj-2 in
Definition 3.20 can be changed to take advantage of this type:
Definition 5.1 (Self types). Assume Fd : Type → Type are monotonic and non-expansive
type constructors, for all d ∈ D. Then let α = [md :νd Fd ]d∈D be defined as the set of all
triples hk, Ψ, {me =le }e∈E i such that D ⊆ E and
∃α′ . α′ ∈ Type ∧ α′ k ⊆ ⌊α⌋k
(Obj-1)
∧ (∀d ∈ D. hk, Ψ, ld i ∈ refνd (α′ → Fd (α′ )))
∧ (∀j < k. ∀Ψ′ . ∀ me =le′ e∈E .
(k, Ψ) ⊑ (j, Ψ′ ) ∧ (∀e ∈ E. Ψ′ j (le′ ) = ⌊Ψ⌋j (le ))
⇒ hj, Ψ′ j , me =le′ e∈E i ∈ α′ )
(Obj-2-self)
(Obj-3)
As in Section 3.5 one shows that Definition 5.1 uniquely determines a type. In this
proof,
the non-expansiveness
of the type functions Fd isnecessary in order to ensure that
[md :νd Fd ]d∈D k is defined in terms of [md :νd Fd ]d∈D j for j < k only. Moreover, the
proofs of the typing lemmas for object types (see Section A.2 in the Appendix) carry over
with minor modifications, to show that the semantic typing lemmas in Figure 12 hold. Most
24
C. HRIŢCU AND J. SCHWINGHAMMER
cases are obtained by replacing the result type τd by Fd (α′ ) throughout the proof, for α′
the existentially quantified type from condition (Obj-1) of Definition 5.1. The proof of
(SemInv-Self) uses the monotonicity of Fe , to conclude that the result of the invocation
has type Fe (α) from the fact that it has type Fe (α′ ), as given by condition (Obj-2-Self). In
the proofs of (SemUpd-Self) and (SemSubObj-Self), the universally quantified ξ from
the respective assumptions is instantiated by α′ . Since (Obj-1) only gives that ⌊α′ ⌋k ⊆ ⌊α⌋k
but not necessarily α′ ⊆ α, these three proofs also use the non-expansiveness of Fd in an
of each Fd , an induction shows that
given the non-expansiveness
essential way. Finally,
[md :νd Fd ]d∈D k = [md :νd ⌊Fd ⌋k ]d∈D k for all k. In other words, [md :νd Fd ]d∈D , viewed
as a type constructor, is non-expansive and Lemma 3.33 still holds.
The interpretation of syntactic type expressions given in Figure 11 extends straightforwardly to self types using the new type constructor:
i
h
y
q
Obj(X)[md :νd Ad ]d∈D η = md :νd λ(α∈Type) JAd Kη[X:=α]
d∈D
With this interpretation and the semantic typing lemmas from Figure 12, the soundness
theorem from Section 4 should extend to a syntactic type system for objects with self types
similar to the one derived by Abadi and Cardelli [2, Ch. 15] for their encoding (but also
including variance annotations and a typing rule for cloning).
5.2. Limitations. Note that with the exception of SemUpd-Self, all the semantic typing
lemmas for self types are stronger than their counterparts from Figure 8. This is already
enough to typecheck many examples involving self types [2, Ch. 15].
However, as for the encoding of Abadi and Cardelli, when updating methods one usually
does not have full information about the precise self type α′ of the host object, which may be
a proper subtype of α. Therefore the statement (SemUpd-Self) about method update in
Figure 12 includes a quantification over all subtypes ξ of the known type α of the object a, to
ensure that the updated method also works correctly for the precise type. As a consequence
the new method body b must be sufficiently parametric in the type of its self parameter x,
which can be overly restrictive. Abadi and Cardelli [2, Ch. 17] demonstrate this limitation
with an example of objects that provide a backup and a retrieve method:
Bk = Obj(X)[retrieve :◦ X, backup :◦ X, . . .]
A sensible definition of the backup method updates the retrieve method so that a subsequent
invocation of retrieve yields a clone of the current object x:
backup(x) = let z = clone x in x.retrieve := ς(y)z
Here, the ‘ let z = a in b ’ stands for the usual syntactic sugar (λ(z)b) a. Let β = JBkKη
denote the interpretation of the syntactic type Bk. While the backup method has the correct
operational behaviour, to typecheck the method update to x in its body using (SemUpdSelf) we would need the statement Σ [x:=β, z:=β, y:=ξ] |= z : ξ. But this statement does
not hold for an arbitrary subtype ξ ⊆ β. Therefore the semantic typing lemmas stated
above are not strong enough to prove that Σ [x:=β, z:=β] |= x.retrieve := ς(y)z : β, and
thus that Σ [x:=β] |= backup(x) : β holds for the method body. This prevents us from
typing an object that contains this method (e.g., [backup = ς(x)backup(x), . . .]) to type β
using the semantic typing lemmas.
A STEP-INDEXED SEMANTICS OF IMPERATIVE OBJECTS
25
Abadi and Cardelli [2, Ch. 17] address this lack of expressiveness by modifying the
calculus in two respects. First, they introduce a new syntax for method update, a.m :=
(y, z = c)ς(x)b. Operationally this new construct behaves just like
let y = a in let z = c in y.m := ς(x)b
(5.1)
but its typing rule is more powerful than the one induced by this encoding. When typing
c and the method body b, y can be assumed to have the precise type of the object:
A ≡ Obj(Y )[md :νd Ad ]d∈D
Γ⊢a:A
e∈D
νe ∈ {−, ◦}
Γ, X6A, y:X ⊢ c : C
Γ, X6A, y:X, z:C, x:X ⊢ b : Ae
(Upd-Self)
Γ ⊢ a.me := (y, z = c)ς(x)b : A
Second, in order to propagate this information, typing rules with ‘structural’ assumptions
are introduced. For instance, the inference rule for object cloning takes the form
(Clone-Str)
A6Obj(Y )[md :νd Ad ]d∈D
Γ ⊢ clone a : A
Γ⊢a:A
thus applying also in the case where A is a type variable. In this modified system, the body
of the backup method can be rewritten as
backupmod (x) = x.retrieve := (y, z = clone y)ς(x)z
(5.2)
and the judgement Γ, x:Bk ⊢ backupmod (x) : Bk is derivable.
Even in the purely syntactic setting, the ad hoc character of the syntax extension is
not entirely satisfactory, but for the step-indexed semantics of types both modifications are
in fact problematic. First, although it seems reasonable to expect that all the semantic
typing lemmas from Section 3 continue to hold, a change of the calculus and its operational
semantics would require us to recheck the proofs about object types in detail. Fortunately,
the syntax extension does not seem necessary from the semantic typing point of view; we
can already prove the semantic soundness of rule (Upd-Self) with respect to the encoding
of the new method update construct from (5.1):
If α = [md :νd Fd ]d∈D , e ∈ D, and νe ∈ {−, ◦} then
Σ |= a : α ∧ ∀ξ ⊆ α. Σ[y := ξ] |= c : γ ∧ ∀ξ ⊆ α. Σ [y := ξ, z := γ, x := ξ] |= b : Fe (ξ)
=⇒
Σ |= let y = a in let z = c in y.m := ς(x)b : α
However, by itself this rule does not help in typing the body of the backup method, and the
introduction of rules with structural assumptions presents a more severe difficulty. Soundness of these rules relies on the fact that every subtype of an object type is another object
type. In other words, in the (Clone-Str) rule the type A is assumed to range only over
object types. Such structural assumptions are usually not valid in semantic models, and
they are certainly not justified with respect to our semantically defined subtype relation,
which is just set inclusion.
5.3. Self Types with Structural Assumptions. To sum up the previous subsection, the
problem is that the semantic typing lemmas from Figure 12 are too weak to type certain
examples such as the body of the backup method, but the usual way to strengthen these rules
in a syntactic setting is not semantically sound in our model. Still, Σ[x := β] |= backup(x) : β
is a valid typing judgement about the method body. This can be seen by taking a closer
26
C. HRIŢCU AND J. SCHWINGHAMMER
Let α = [md :νd Fd ]d∈D .
(∀d∈D. ∀ξ∈Type. ξ ⊳ α ⇒ Σ[xd := ξ] |= bd : Fd (ξ))
=⇒
Σ |= [md =ς(xd )bd ]d∈D : α
(SemObj-Str)
(α′ ⊳ α ∧ Σ |= a : α′ ∧ e ∈ D ∧ νe ∈ {+, ◦})
=⇒
Σ |= a.me : Fe (α′ )
(α′ ⊳ α ∧ Σ |= a : α′ ∧ e ∈ D ∧ νe ∈ {−, ◦}
(SemInv-Str)
(SemUpd-Str)
∧ Σ[x := α ] |= b : Fe (α ))
=⇒
Σ |= a.me := ς(x)b : α′
α′ ⊳ α ∧ Σ |= a : α′
=⇒
Σ |= clone a : α′ (SemClone-Str)
Σ |= a : α ∧ (∀ξ∈Type. ξ ⊳ α ⇒ Σ[x := ξ] |= b : β)
=⇒
Σ |= let x = a in b : β
(SemLet-Str)
′
′
Figure 13: Typing lemmas with structural assumptions: self types
look at the semantic definition of the self type β = JBkKη : essentially, if for a suitable
substitution σ :k,Ψ Σ[x := β] the substitution instance
σ(backup(x)) = let z = clone σ(x) in σ(x).retrieve := ς(y)z
becomes irreducible in less than k steps, then σ(x) must be an object value v = {md =ld }d∈D
such that hk, Ψ, vi ∈ β. Property (Obj-1) of β asserts the existence of a type α′ such that
⌊α′ ⌋k ⊆ ⌊β⌋k , and property (Obj-3) entails that z becomes bound to a value v ′ of this type
α′ . Thus by (Obj-2-self) the eventual update of the retrieve field of v is valid, since the
new method λ(y)v ′ has the expected type α′ → α′ to sufficient approximation.
Similar ‘manual’ reasoning seems possible in other cases, but a more principled approach will let us use typing lemmas that are strong enough and avoid explicit reasoning
about the operational semantics and step indices. To facilitate this, we develop a semantic
counterpart to the structural assumptions that appear in the syntactic type system of Abadi
and Cardelli [2, Ch. 17]. More precisely, we introduce a relation α′ ⊳ α between semantic
types that strengthens the subtype relation: intuitively α′ is the precise, recursive type of
some collection of object values from the object type α. The type α acts as an interface
that lists the permitted operations on these object values.
Definition 5.2 (Self type exposure). For α = [md :νd Fd ]d∈D and α′ ∈ Type the relation
α′ ⊳ α holds if and only if α′ ⊆ α and for all E ⊇ D and hk, Ψ, {me =le }e∈E i ∈ α′ ,
(Obj-2-self)
(∀d ∈ D. hk, Ψ, ld i ∈ refνd (α′ → Fd (α′ )))
′
′
∧ (∀j < k. ∀Ψ . ∀ me =le e∈E .
(Obj-3)
′ ′
′
(k, Ψ) ⊑ (j, Ψ ) ∧ (∀e ∈ E. Ψ j (le ) = ⌊Ψ⌋j (le ))
⇒ hj, Ψ′ j , me =le′ e∈E i ∈ α′ )
Notice that α′ ⊳ α essentially states that α′ is a type that can take the place of the
existentially quantified ‘self type’ in an object type (see Definition 5.1). It is immediate
from this definition that α′ ⊳ α implies α′ ⊆ α. Note however that ⊳ is not reflexive: in
general, α′ is not an object type (e.g., α′ could be empty). Intuitively, the object type α is
obtained as a union of such α′ .
A STEP-INDEXED SEMANTICS OF IMPERATIVE OBJECTS
27
Figure 13 lists new typing lemmas for self types that exploit the relation ⊳. Compared
to (SemInv-Self) and (SemClone-Self) from Figure 12, the typing lemmas (SemInvStr) and (SemClone-Str) use the additional assumptions α′ ⊳ α and Σ |= a : α′ to
establish a more precise typing for the result. Similarly, while (SemUpd-Self) universally
quantifies over all ξ ⊆ α in its premise, (SemUpd-Str) limits this to those ξ ∈ Type for
which ξ ⊳ α holds. Finally, (SemLet-Str) lets us use an object a within b with the more
precise type ξ where ξ ⊳ α, and similarly (SemObj-Str) lets us type the method bodies
under the more informative assumption that ξ ⊳ α. The latter two are the key lemmas to
introduce an assumption α′ ⊳ α in proofs using these semantic typing lemmas.
As an illustration, consider the example of the backup method again. In order to
construct objects with the backup method, we will establish that
∀ξ∈Type. ξ ⊳ β ⇒ Σ[x := ξ] |= backup(x) : ξ
(5.3)
holds, where β = JBkKη and backup(x) abbreviates ‘let z = clone x in x.retrieve := ς(y)z’ as
before. From this, by (SemObj-Str) it will follow that
Σ |= [backup = ς(x)backup(x), retrieve = . . .] : β
If we desugar the let construct in backup(x) and apply lemmas (SemApp) and (SemLam),
we notice that in order to show (5.3) it suffices to prove that Σ[x:=ξ] |= clone x : ξ and
Σ[x:=ξ, z:=ξ] |= x.retrieve := ς(y)z : ξ. Using ξ ⊳ β and Σ[x:=ξ] |= x : ξ, the validity
of the former judgement is immediate by (SemClone-Str). Similarly, the latter follows
by (SemUpd-Str) from the fact that ξ ⊳ β and since the retrieve method is listed with
variance annotation ‘◦’ in β.
Lemma 5.3 (Self types: lemmas with structural assumptions). All the semantic typing
lemmas shown in Figure 13 are valid implications.
Proof sketch. The proofs of (SemInv-Str), (SemClone-Str), and (SemUpd-Str) are
straightforward adaptations of those for (SemInv), (SemClone), and (SemUpd). As an
example, we give the proof of (SemUpd-Str) as Lemma A.10 in the Appendix. More
interestingly, (SemLet-Str) relies on the following property of object types α:
hk, Ψ, vi ∈ α
=⇒
∃α′ ∈ Type. α′ ⊳ α ∧ hk − 1, ⌊Ψ⌋k−1 , vi ∈ α′
In the proof of (SemLet-Str), this α′ is used to instantiate the universally quantified ξ
in the premise ξ ⊳ α ⇒ Σ[x := ξ] |= b : β. The full proof is given as Lemma A.12 in the
Appendix.
The proof of (SemObj-Str) is similar to the one of (SemObj)
(Lemma A.4 in the Ap
pendix), except that we use the heap typing extension Ψ′ = Ψ [ld := (β → Fd (β))]d∈D k−1 ,
where β is a recursive record type satisfying conditions (Obj-2-self) and (Obj-3), but not
validating any subtyping property. In verifying that the extended heap is well-typed with
respect to this Ψ′ , one uses that β ⊳ α, in order to instantiate the assumptions on the
method bodies and obtain Σ[xd := β] |= bd : Fd (β). Finally, to show that the generated
object value has type α, Claim 3.23 is strengthened to show that the object value is in fact
in β, which is a subtype of α since β ⊳ α (see Proposition A.14 and Lemma A.15 in the
Appendix for the full proof).
28
C. HRIŢCU AND J. SCHWINGHAMMER
Remark 5.4. The implication Σ |= a : α =⇒ ∃α′ ∈ Type. α′ ⊳ α ∧ Σ |= a : α′
for α = [md :νd Fd ]d∈D may appear reasonable (and would entail both (SemLet-Str) and
(SemObj-Str)), but we do not believe that it holds. The problem is that, while the premise
Σ |= a : α guarantees for each k ≥ 0 the existence of a type α′k that satisfies the requirement
α′k ⊳ α, it is in general not possible to construct a type α′ ‘in the limit’ from this sequence.
For the same reason, the implication Σ |= pack a : ∃α F =⇒ ∃α′ ⊆ α. Σ |= a : F (α′ ) is not
valid. The typing lemma (SemLet-Str) avoids this problem since α′ is only needed up to
a fixed approximation, and so the choice of α′k for sufficiently large k suffices (cf. proof of
Lemma A.12 in the Appendix). On the other hand, the (SemObj-Str) lemma avoids the
problem by instantiating ξ with a particular type β, for which β ⊳ α is already known.
In this section we showed that our semantics of object types naturally extends to self
types, while avoiding any change to the syntax and operational semantics of the calculus.
We proved a first set of typing lemmas that are natural and apply to many examples
(Figure 12). These lemmas are however not sufficient to typecheck self-returning methods.
To achieve this, we developed a second set of typing lemmas that involve the object’s precise
type, through the relation α′ ⊳ α (Figure 13). Note that these latter lemmas do not fully
subsume the former ones, since the ⊳ relation is not reflexive. We leave open the problem
of relating the lemmas in Figure 13 to a syntactic type system.
6. Generalizing Reference and Object Types
The semantics described in this paper generalizes the reference types from [6, 9] to readable
and writable reference types. This can be generalized even further. We can have a reference
type constructor that takes two types as arguments: one that represents the most general
type that can be used when writing to the reference, and another for the most specific type
that can be read from it [38]. This can be easily expressed using our readable and writable
reference types together with intersection types:
ref(τ w , τ r ) = ref− τ w ∩ ref+ τ r
After unfolding the definitions, this yields
ref(τ w , τ r ) = {hk, Ψ, li | ⌊τ w ⌋k ⊆ ⌊Ψ(l)⌋k ⊆ ⌊τ r ⌋k }.
As one would expect, this generalized reference type constructor is contravariant in the first
argument and covariant in the second one:
β w ⊆ αw
∧
αr ⊆ β r
=⇒
ref(αw , αr ) ⊆ ref(β w , β r )
(SemSubRef-Gen)
Note that, if one takes these generalized reference types as primitive, then the three reference
types from Section 3.4 are obtained as special cases:
ref◦ τ = ref(τ, τ ),
ref+ τ = ref(⊥, τ ),
ref− τ = ref(τ, ⊤),
and the subtyping properties from Figure 7 are still valid.
Figures 14 and 15 give a graphical representation of the different reference type constructors. In both figures the horizontal axis represents the type at which a reference can be
read, while the vertical one gives the type at which it can be written. Notice that because
of the different variance the read axis goes from ⊥ to ⊤ while the write axis from ⊤ to ⊥.
Figure 14 represents the usual, as well as the readable, and the writable reference types
as points on the three edges of a triangle. Notice that the usual references can be read
and written at the same type. Without additional information, the readable references
A STEP-INDEXED SEMANTICS OF IMPERATIVE OBJECTS
Figure 14: Readable/writable reference types
29
Figure 15: Generalized reference types
Let α = [md : (τdw , τdr ) ]d∈D and α′ = [md : (τd , τd ) ]d∈D .
(∀d ∈ D. Σ[xd := α′ ] |= bd : τd )
=⇒
Σ |= [md =ς(xd )bd ]d∈D : α′
(SemObj-Gen)
(Σ |= a : α ∧ e ∈ D)
=⇒
Σ |= a.me :
τer
(SemInv-Gen)
(Σ |= a : α ∧ e ∈ D ∧ Σ[x := α] |= b : τew )
=⇒
Σ |= a.me := ς(x)b : α
Σ |= a : α
=⇒
Σ |= clone a : α
r
r
(E ⊆ D ∧ (∀e ∈ E. βew ⊆ αw
e ∧ αe ⊆ βe ))
=⇒
r
w
r
[md : (αw
d , αd ) ]d∈D ⊆ [me : (βe , βe ) ]e∈E
(SemSubObj-Gen)
(SemUpd-Gen)
(SemClone-Gen)
Figure 16: Typing lemmas: generalized object types
can only be written safely at type ⊥, and the writable ones can only be read at type ⊤.
Subtyping is represented by arrows: covariant on the edge of the readable reference types
and contravariant on the writable reference types’ edge. An invariant reference type can
only be subtyped either to a readable or to a writable reference type.
Figure 15 illustrates that our generalization of reference types is indeed very natural.
When generalizing, we take not only the points on the edges of the triangle, but also the
points inside it to be reference types. Furthermore, instead of having three different kinds of
reference types, we only have one. Subtyping is also more natural: the set of all supertypes
of a reference type cover the area of a rectangle which goes from the point corresponding
to this reference type to the ‘top’ reference type ref(⊥, ⊤). For instance, the dark gray
rectangle in Figure 15 contains all supertypes of ref(τ, τ ).
Applying this idea in the context of the imperative object calculus leads not only to
more expressive subtyping but also to simplifications, since the variance annotations are no
longer needed. The extended object type [md : (τdw , τdr ) ]d∈D has two types for each method
md : τdw is the most general type that can be used to update the given method, and τdr is
the most specific type that can be expected as a result when invoking the method. When
defining the semantics of these generalized object types, the only difference with respect
to Definition 3.20 (Object types) is that condition (Obj-2) is changed to use an extended
reference type:
∀d ∈ D. hk, Ψ, ld i ∈ ref(α′ → τdw , α′ → τdr ).
(Obj-2-Gen)
30
C. HRIŢCU AND J. SCHWINGHAMMER
r
′
Let A = [md : (Aw
d , Ad ) ]d∈D and A = [md : (Ad , Ad ) ]d∈D .
(Obj-Gen)
(Inv-Gen)
∀d ∈ D. Γ, xd : A′ ⊢ bd : Ad
Γ ⊢ md =ς(xd :A′ )bd d∈D : A′
Γ⊢a:A e∈D
Γ ⊢ a.me : Are
(SubObj-Gen)
(Upd-Gen)
(Clone-Gen)
Γ⊢a:A
Γ ⊢ clone a : A
Γ ⊢ a : A e ∈ D Γ, x : A ⊢ b : Aw
e
Γ ⊢ a.me := ς(x:A)b : A
E ⊆ D ∀e ∈ E. Γ ⊢ Bew 6 Aw
∀e ∈ E. Γ ⊢ Are 6 Ber
e
r
w
r
Γ ⊢ [md : (Aw
d , Ad ) ]d∈D 6 [me : (Be , Be ) ]e∈E
Figure 17: The typing rules for generalized object types
Figure 16 presents the semantic typing lemmas that are validated by this definition of
object types, while Figure 17 gives the corresponding syntactic typing rules. Note that the
complex and seemingly ad-hoc rules for subtyping object types given in Figure 4 or in [2]
are replaced by only one rule (SubObj-Gen).
Lemma 6.1 (Generalized object types). All the semantic typing lemmas shown in Figure 16
are valid implications.
Proof sketch. The proof of the subtyping lemma (SemSubObj-Gen) follows easily from the
lemma for subtyping generalized reference types (SemSubRef-Gen above), and is therefore
significantly simpler than when variance annotations are involved (see Lemma A.16 in the
Appendix). For all the other semantic typing lemmas the proofs are basically unchanged
(see Section A.2 in the Appendix).
Note that the generalization of object types presented in this section is orthogonal to
the extension to self types from the previous section. The generalized object types lead to
a type system that is both simpler and more expressive than the usual type systems for
objects [2]. Our generalized object types directly correspond to the split types of Bugliesi
and Pericás-Geertsen [19], who have shown that these types are strictly more expressive
than object types with variance annotations [19, Example 4.3].
7. Comparison to Related Work
7.1. Domain-theoretic Models. Abadi and Cardelli give a semantic model for the functional object calculus in [1, 2]. Their type system is comparable to the one we consider
here. Types are interpreted as certain partial equivalence relations over an untyped domaintheoretic model of the calculus. No indication is given on how to adapt this to the imperative
execution model.
Based on earlier work by Kamin and Reddy [30], Reus et al. [41, 42, 44] construct
domain-theoretic models for the imperative object calculus, with the goal of proving soundness for the logic of Abadi and Leino [4]. The higher-order store exhibited by the object
calculus requires defining the semantic domains by mixed-variant recursive equations. The
dynamic allocation is then addressed by interpreting specifications of the logic as Kripke
relations, indexed by store specifications, which are similar to the heap typings used here.
A STEP-INDEXED SEMANTICS OF IMPERATIVE OBJECTS
31
Building on work by Levy [31], an ‘intrinsically typed’ model of the imperative object
calculus is presented in the second author’s PhD thesis [44], by solving the domain equations
in a suitable category of functors. However, only first-order types are considered.
Compared to these domain-theoretic models, the step-indexed model we present not
only soundly interprets a richer type language, but is also easier to work with. The way it
is based on the operational semantics eliminates the need for explicit continuity conditions,
and the admissibility conditions are replaced by the closure under state extension, which
is usually very easy to check. All that is needed for the definition of iso-recursive and
second-order types are non-expansiveness and the stratification invariant. What is missing
from our model is a semantic notion of equality that approximates program equivalence.
For reasoning about program equivalence in an ML-like language, Ahmed et al. [5] have
recently developed a relational step-indexed model, and it could be interesting to adapt
their work to an object-oriented setting.
Recently proposed models for polymorphism and general references [15, 16, 17] suggest
that an adequate semantics for imperative objects with expressive typing could in principle
be developed also in a domain-theoretic setting. A detailed comparison between stepindexed semantics and domain-theoretic models would be useful, to make the similarities
and differences between the two approaches more precise.
It is interesting to see how the object construction rule (Obj) is proved correct in
each of the models described above. In the domain-theoretic self-application models [41,
42], it directly corresponds to a recursive predicate whose well-definedness (i.e., existence
and uniqueness) must be established. This proof exploits properties of the underlying,
recursively defined domain, and imposes some further restrictions on the semantic types:
besides admissibility, types appearing in the defining equation of a recursive predicate need
to satisfy an analogue of the contractiveness property [36]. In the typed functor category
model [44], object construction is interpreted using a recursively defined function, and
correspondingly (Obj) is proved by fixed point induction. In the step-indexed case, the
essence of the proof is a more elementary induction on the step index, with a suitably
generalized induction hypothesis (see Claim 3.23 in the proof sketch of Lemma 3.22 on page
18, or the full proof in the Appendix).
7.2. Interpretations of Object Types. Our main contribution in this paper is the novel
interpretation of object types in the step-indexed model. The step-index-induced stratification permits the construction of mixed-variance recursive as well as impredicative, secondorder types. Both are key ingredients in our interpretation of object types. The use of
recursive and existentially quantified types is in line with the type-theoretic work on object
encodings, which however has mainly focused on object calculi with a functional execution
model [18].
Closest to our work is the encoding of imperative objects into an imperative variant
of system F6µ with updatable records, proposed by Abadi et al. [3]. There, objects are
interpreted as records containing references to the procedures that represent the methods.
As in our case, these records have a recursive and existentially quantified record type. The
difference is that two additional record fields are included in order to achieve invocation and
cloning, and uninitialized fields are used to construct this recursive record. Subtyping in
depth is considered in [3] only for the encoding of the functional object calculus. However,
if one added to the target language the readable and writable reference types we use in this
32
C. HRIŢCU AND J. SCHWINGHAMMER
paper, the encoding of the imperative object calculus would extend to subtyping in depth
as well.
In the typing rules for self types, the structural assumptions about the subtype relation
play an important role [2]. In Section 5 we developed a semantic counterpart to such typing
rules with structural assumptions, in order to deal with the polymorphic update of selfreturning methods. This is, however, tailored specifically to object types. Hofmann and
Pierce [26] investigate the metatheory of subtyping with structural assumptions in general,
and give elementary presentations of two encodings of functional objects in a variant of
System F≤ with type destructors. It may be interesting to see if a step-indexed model of
this variant of System F≤ can be found.
7.3. Step-indexed Models. Step-indexed semantic models were introduced by Appel et
al. in the context of foundational proof-carrying code. Their goal was to construct more
elementary and modular proofs of type soundness that can be easily checked automatically.
They were primarily interested in low-level languages, however they also applied their technique to a pure λ-calculus with recursive types [10]. Later Ahmed et al. successfully
extended it to general references and impredicative polymorphism [6, 9]. The step-indexed
semantic model we present extends the one by Ahmed et al. with object types and subtyping. In order to achieve this, we refine the reference types from [6] to readable and writable
reference types.
Subtyping in a step-indexed semantic model was previously considered by Swadi who
studied Typed Machine Language [45]. Our setup is however much different. In particular,
the subtle issues concerning the subtyping of object types are original to our work.
The previous work on step-indexing focuses on ‘semantic type systems’, i.e., the semantic typing lemmas can directly be used for type-checking programs [9, 10, 12]. However,
when one considers more complex type systems with subtyping, recursive types or polymorphism, the semantic typing lemmas no longer directly correspond to the usual syntactic
rules. These discrepancies can be fixed, but usually at the cost of more complex models, like
the one developed by Swadi to track type variables [12, 45]. In Swadi’s model an additional
‘semantic kind system’ is used to track the contractiveness and non-expansiveness of types
with free type variables. We avoid having a more complex model (e.g., one that tracks type
variables) by considering iso-recursive rather than equi-recursive types. An equi-recursive
type is well-defined if its argument is contractive, and some of the type constructors are not
contractive in general (e.g., the identity as well as the equi-recursive type constructor itself). On the other hand, an iso-recursive type is well-defined under the weaker assumption
that the argument is non-expansive, and all our type constructors are indeed non-expansive
(see Lemma 3.33). It is then relatively straightforward to use the semantic typing lemmas
in order to prove the soundness of the standard, syntactic type system we consider (see
Theorem 4.5).
7.4. Type Safety Proofs. Abadi and Cardelli use subject reduction to prove the safety
of several type systems very similar to the one considered in this paper [2]. Those purely
syntactic proofs are very different from the ‘semantic’ type safety proof we present (for
detailed discussions about the differences see [10, 46]). Since type safety is built into the
model, our safety proof neither relies on a preservation property, nor can preservation be
concluded from it.
A STEP-INDEXED SEMANTICS OF IMPERATIVE OBJECTS
33
Constructing a step-indexed semantics is more challenging than proving progress and
preservation. However, for our particular semantics we could reuse the model by Ahmed
et al. and extend it to suit our needs, even though the calculus we are considering is
quite different. So one would expect that once enough general models are constructed
(e.g., [6, 10, 11]), it will become easier to build new models just by mixing and matching.
Assuming the existence of an adequate step-indexed model, the effort needed to prove
the semantic typing lemmas using ‘pencil-and-paper’ is somewhat comparable to the one
required for a subject reduction proof. Since each of the semantic typing lemmas is proved in
isolation, the resulting type soundness proof is more modular; the extensions we consider in
Sections 5 and 6 illustrate this aspect rather well. According to Appel’s original motivation,
the advantages of step-indexing should become even more apparent when formalizing the
proofs in a proof assistant [10].
7.5. Generalized Reference and Object Types. The readable and the writable reference types we define in Section 3.4 and use for modeling object types in Section 3.5 are
similar to the reference types in the Forsythe programming language [43] and to the channel
types of [22, 35]. The generalization to a reference type constructor taking two arguments
described in Section 6 is quite natural, and also appeared in Pottier’s thesis [38], where
it facilitated type inference by allowing meets and joins to distribute over reference types.
This idea has recently been applied by Craciun et al. for inferring variant parametric types
in Java [23].
The generalized object types we introduce in Section 6 directly correspond to the split
types of Bugliesi and Pericás-Geertsen [19]. Split types are also motivated by type inference, since they guarantee the existence of more precise upper and lower bounds. In
particular, Bugliesi and Pericás-Geertsen show that split types are strictly more expressive
than first-order object types with variance annotations [19, Example 4.3]. They establish
the soundness of a type system with split types by subject reduction, with respect to a
functional semantics of the object calculus.
7.6. Functional Object Calculus. Our initial experiments on the current topic were done
in the context of the functional object calculus [27]. Even though in the functional setting
the semantic model is much simpler, both models satisfy the same semantic typing lemmas.
Even more, the syntactic type system we considered for the functional calculus is exactly the
same as the one in this paper, so all the results in Section 4 directly apply to the functional
object calculus: well-typed terms do not get stuck, no matter whether they are evaluated in
a functional or an imperative way. It would not be possible to directly prove such a result
using subject reduction, since for subject reduction the syntactic typing judgment for the
imperative calculus would also depend on a heap typing, and thus be different from the
judgment for the functional calculus. However, since we are not using subject reduction,
we do not need to type-check partially evaluated terms that contain heap locations.
8. Conclusion
We have presented a step-indexed semantics for Abadi and Cardelli’s imperative object
calculus, and used it to prove the safety of a type system with object types, recursive and
second-order types, as well as subtyping. We showed how this semantics can be extended
to self types and typing lemmas with structural assumptions; and generalized in a way that
34
C. HRIŢCU AND J. SCHWINGHAMMER
eliminates the need for variance annotations and at the same time simplifies the subtyping
rules for objects.
The step-indexing technique is however not limited to type safety proofs, and has already been employed for more general reasoning about programs. Based on previous work
by Appel and McAllester [10], Ahmed built a step-indexed partial equivalence relation model
for the lambda calculus with recursive and impredicative quantified types, and showed that
her relational interpretation of types is sound for proving contextual equivalences [7]. Recently, this was extended significantly to reason about program equivalence in the presence
of general references [5]. Benton also used step-indexing as a technical device, together with
a notion of orthogonality relating expressions to contexts, to show the soundness of a compositional program logic for a simple stack-based abstract machine [13]. He also employed
step-indexing in a Floyd-Hoare-style framework based on relational parametricity for the
specification and verification of machine code programs [14].
We hope that our work paves the way for similarly compelling, semantic investigations
of program logics for the imperative object calculus: using a step-indexed model it should
be possible to prove the soundness of more expressive program logics for this calculus.
Acknowledgements
We express our gratitude to the anonymous reviewers for their detailed and constructive
comments on the preliminary versions of this article. We also thank Andreas Rossberg for
pointing us to the work of John C. Reynolds on Forsythe. Cătălin Hriţcu is supported by a
fellowship from Microsoft Research and the International Max Planck Research School for
Computer Science.
References
[1] Martı́n Abadi and Luca Cardelli. A semantics of object types. In Proceedings Ninth Annual IEEE
Symposium on Logic in Computer Science, LICS’94, pages 332–341. IEEE Computer Society Press,
1994.
[2] Martı́n Abadi and Luca Cardelli. A Theory of Objects. Springer, 1996.
[3] Martı́n Abadi, Luca Cardelli, and Ramesh Viswanathan. An interpretation of objects and object types.
In Proceedings 23rd Symposium on Principles of Programming Languages, POPL’96, pages 396–409.
ACM Press, 1996.
[4] Martı́n Abadi and K. Rustan M. Leino. A logic of object-oriented programs. In Nachum Dershowitz,
editor, Verification: Theory and Practice. Essays Dedicated to Zohar Manna on the Occasion of his
64th Birthday, Lecture Notes in Computer Science, pages 11–41. Springer, 2004.
[5] Amal Ahmed, Derek Dreyer, and Andreas Rossberg. State-dependent representation independence. In
Zhong Shao and Benjamin C. Pierce, editors, Proceedings 36th Symposium on Principles of Programming
Languages, POPL’09, pages 340–353, 2009.
[6] Amal J. Ahmed. Semantics of types for mutable state. PhD thesis, Princeton University, 2004.
[7] Amal J. Ahmed. Step-indexed syntactic logical relations for recursive and quantified types. In Peter
Sestoft, editor, Proceedings 15th European Symposium on Programming, ESOP’06, volume 3924 of
Lecture Notes in Computer Science, pages 69–83. Springer, 2006.
[8] Amal J. Ahmed, Andrew W. Appel, and Roberto Virga. A stratified semantics of general references
embeddable in higher-order logic. In Proceedings 17th Annual IEEE Symposium Logic in Computer
Science, LICS’02, pages 75–86. IEEE Computer Society Press, 2002.
[9] Amal J. Ahmed, Andrew W. Appel, and Roberto Virga. An indexed model of impredicative polymorphism and mutable references. Princeton University, January 2003.
A STEP-INDEXED SEMANTICS OF IMPERATIVE OBJECTS
35
[10] Andrew W. Appel and David McAllester. An indexed model of recursive types for foundational proofcarrying code. ACM Transactions on Programming Languages and Systems, 23(5):657–683, September
2001.
[11] Andrew W. Appel, Paul-André Melliès, Christopher D. Richards, and Jérôme Vouillon. A very modal
model of a modern, major, general type system. In Proceedings 34th Symposium on Principles of Programming Languages, POPL’07, pages 109–122, 2007.
[12] Andrew W. Appel, Christopher Richards, and Kedar Swadi. A kind system for typed machine language.
Technical report, Princeton University, September 2002.
[13] Nick Benton. A typed, compositional logic for a stack-based abstract machine. In Zoltán Ésik, editor,
Proceedings Asian Symposium on Programming Languages and Systems, APLAS’05, volume 3780 of
Lecture Notes in Computer Science, pages 182–196. Springer, 2005.
[14] Nick Benton. Abstracting allocation: the new new thing. In Zoltán Ésik, editor, Proceedings Computer
Science Logic, CSL’06, volume 4207 of Lecture Notes in Computer Science, pages 364–380. Springer,
2006.
[15] Lars Birkedal, Kristian Støvring, and Jacob Thamsborg. Realizability semantics of parametric polymorphism, general references, and recursive types. In Proceedings Foundations of Software Science and
Computation Structures, FOSSACS’09, volume 5504 of Lecture Notes in Computer Science, pages 456–
470. Springer, 2009.
[16] Nina Bohr. Advances in Reasoning Principles for Contextual Equivalence and Termination. PhD thesis,
IT University of Copenhagen, 2007.
[17] Nina Bohr and Lars Birkedal. Relational reasoning for recursive types and references. In Naoki
Kobayashi, editor, Proceedings Asian Symposium on Programming Languages, APLAS’06, volume 4279
of Lecture Notes in Computer Science, pages 79–96. Springer, 2006.
[18] Kim B. Bruce, Luca Cardelli, and Benjamin C. Pierce. Comparing object encodings. Information and
Computation, 155(1/2):108–133, November 1999.
[19] Michele Bugliesi and Santiago M. Pericás-Geertsen. Type inference for variant object types. Information
and Computation, 177(1):2–27, 2002.
[20] Luca Cardelli. Amber. In Guy Cousineau, Pierre-Louis Curien, and Bernard Robinet, editors, Combinators and Functional Programming Languages, volume 242 of Lecture Notes in Computer Science,
pages 21–47. Springer, 1985.
[21] Luca Cardelli. Type systems. In Allen B. Tucker, editor, The Computer Science and Engineering Handbook, chapter 103, pages 2208–2236. CRC Press, 1997.
[22] Giuseppe Castagna, Rocco De Nicola, and Daniele Varacca. Semantic subtyping for the π-calculus.
Theoretical Computer Science, 398(1-3):217–242, 2008. Essays in honour of Mario Coppo, Mariangiola
Dezani-Ciancaglini and Simona Ronchi della Rocca.
[23] Florin Craciun, Wei-Ngan Chin, Guanhua He, and Shengchao Qin. An interval-based inference of variant
parametric types. In Giuseppe Castagna, editor, Proceedings 18th European Symposium on Programming, ESOP ’09, volume 5502 of Lecture Notes in Computer Science, pages 112–127. Springer, 2009.
[24] Cormac Flanagan, Stephen Freund, and Aaron Tomb. Hybrid types, invariants, and refinements for
imperative objects. In Workshop on Foundations and Developments of Object-Oriented Languages,
FOOL/WOOD’06, 2006.
[25] Andrew D. Gordon, Paul D. Hankin, and Søren B. Lassen. Compilation and equivalence of imperative
objects. In S. Ramesh and G. Sivakumar, editors, Proceedings 17th Conference on Foundations of
Software Technology and Theoretical Computer Science, FST+TCS’97, volume 1346 of Lecture Notes
in Computer Science, pages 74–87. Springer, 1997.
[26] Martin Hofmann and Benjamin C. Pierce. Type destructors. Information and Computation, 172(1):29–
62, 2002.
[27] Cătălin Hriţcu. A step-indexed semantic model of types for the functional object calculus. Master’s
thesis, Programming Systems Lab, Saarland University, May 2007.
[28] Cătălin Hriţcu and Jan Schwinghammer. A step-indexed semantics of imperative objects. Extended
version, Programming Systems Lab, Saarland University, February 2008.
[29] Alan Jeffrey and Julian Rathke. A fully abstract may testing semantics for concurrent objects. Theoretical Computer Science, 338(1-3):17–63, 2005.
36
C. HRIŢCU AND J. SCHWINGHAMMER
[30] Samuel N. Kamin and Uday S. Reddy. Two semantic models of object-oriented languages. In Carl A.
Gunter and John C. Mitchell, editors, Theoretical Aspects of Object-Oriented Programming: Types,
Semantics, and Language Design, pages 464–495. MIT Press, 1994.
[31] Paul Blain Levy. Possible world semantics for general storage in call-by-value. In Julian Bradfield, editor,
Proceedings Computer Science Logic, CSL’02, volume 2471 of Lecture Notes in Computer Science, pages
232–246. Springer, 2002.
[32] John C. Mitchell and Eugenio Moggi. Kripke-style models for typed lambda calculus. Annals of Pure
and Applied Logic, 51(1–2):99–124, 1991.
[33] Eugenio Moggi. An abstract view of programming languages. Technical Report ECS-LFCS-90-113,
Laboratory for Foundations of Computer Science, University of Edinburgh, 1990.
[34] Frank J. Oles. Type algebras, functor categories, and block structure. In Maurice Nivat and John C.
Reynolds, editors, Algebraic Methods in Semantics. Cambrige University Press, 1985.
[35] Benjamin C. Pierce and Davide Sangiorgi. Typing and subtyping for mobile processes. Mathematical
Structures in Computer Science, 6(5), 1996.
[36] Andrew M. Pitts. Relational properties of domains. Information and Computation, 127:66–90, 1996.
[37] Andrew M. Pitts and Ian D. B. Stark. Operational reasoning for functions with local state. In Andrew D.
Gordon and Andrew M. Pitts, editors, Higher-Order Operational Techniques in Semantics, Publications
of the Newton Institute, pages 227–273. Cambridge University Press, 1998.
[38] François Pottier. Type inference in the presence of subtyping: from theory to practice. Research Report
3483, INRIA, September 1998.
[39] Uday S. Reddy and Hongseok Yang. Correctness of data representations involving heap data structures.
Science of Computer Programming, 50(1–3):129–160, March 2004.
[40] Bernhard Reus. Modular semantics and logics of classes. In Matthias Baatz and Johann A. Makowsky,
editors, Proceedings Computer Science Logic, CSL’03, volume 2803 of Lecture Notes in Computer Science, pages 456–469. Springer, 2003.
[41] Bernhard Reus and Jan Schwinghammer. Denotational semantics for a program logic of objects. Mathematical Structures in Computer Science, 16(2):313–358, April 2006.
[42] Bernhard Reus and Thomas Streicher. Semantics and logic of object calculi. Theoretical Computer
Science, 316:191–213, 2004.
[43] John C. Reynolds. Design of the programming language Forsythe. Technical Report CMU-CS-96-146,
Carnegie Mellon University, June 1996. Reprinted in O’Hearn and Tennent, ALGOL-like Languages,
vol. 1, pages 173-233, Birkhäuser, 1997.
[44] Jan Schwinghammer. Reasoning about Denotations of Recursive Objects. PhD thesis, Department of
Informatics, University of Sussex, 2006.
[45] Kedar N. Swadi. Typed Machine Language. PhD thesis, Princeton University, July 2003.
[46] Andrew K. Wright and Matthias Felleisen. A syntactic approach to type soundness. Information and
Computation, 115(1):38–94, 1994.
Appendix A.
A.1. Auxiliary Propositions.
Proposition A.1 (Preorder). The state extension relation, ⊑, is reflexive and transitive.
Proposition A.2 (Information-forgetting extension). If j ≤ k then (k, Ψ) ⊑ (j, ⌊Ψ⌋j ).
Proposition A.3 (Relation between hk, Ψ, vi ∈ τ and v :k,Ψ τ ). Let v be a closed value.
(1) If hk, Ψ, vi ∈ τ then v :k,Ψ τ .
(2) If v :k,Ψ τ , k > 0, and there exists some h such that h :k Ψ, then hk, Ψ, vi ∈ τ .
A STEP-INDEXED SEMANTICS OF IMPERATIVE OBJECTS
37
A.2. Typing Lemmas for Object Types.
Lemma A.4 (SemObj: Object construction). For all object types α = [md :νd τd ]d∈D , if
for all d ∈ D we have Σ[xd := α] |= bd : τd , then Σ |= [md =ς(xd )bd ]d∈D : α.
Proof. Let α = [md :νd τd ]d∈D and assume that ∀d ∈ D. Σ[xd := α] |= bd : τd . We must
show that Σ |= [md =ς(xd )bd ]d∈D : α. Thus, let k ≥ 0, σ be a value environment and Ψ be
a heap typing such that σ :k,Ψ Σ. By the definition of the semantic typing judgement (Definition 3.8) we need to show that σ([md =ς(xd )bd ]d∈D ) :k,Ψ α. Equivalently (after suitable
α-renaming), we show that
[md =ς(xd )σ(bd )]d∈D :k,Ψ α
′
′
Suppose j < k, h, h and b are such that the following three conditions are fulfilled:
h :k Ψ
∧
hh, [md =ς(xd )σ(bd )]d∈D i →j hh′ , b′ i
∧
hh′ , b′ i9
(A.1)
By the operational semantics Red-Obj is the only rule that applies, which means that
necessarily j = 1 and for some distinct ld 6∈ dom(h) we have b′ = {md =ld }d∈D and
We choose
and show that
h′ = h [ld := λ(xd )σ(bd )]d∈D
(A.2)
Ψ′ = Ψ [ld := (α → τd )]d∈D k−1
(A.3)
(k, Ψ) ⊑ (k − 1, Ψ′ )
∧
h′ :k−1 Ψ′
∧
hk − 1, Ψ′ , b′ i ∈ α
(A.4)
Ψ′
That the first conjunct of (A.4) holds is immediate from the construction of
(A.3).
In order to show the second conjunct, by Definition 3.5 (Well-typed heap) we first need
to show that dom(Ψ′ ) ⊆ dom(h′ ). From the first conjunct of (A.1) and Definition 3.5 it
is clear that dom(Ψ) ⊆ dom(h). Thus from the shape of h′ (A.2) and the definition of Ψ′
(A.3) we obtain the required inclusion.
Next, let i < k − 1 and l ∈ dom(Ψ′ ). To establish h′ :k−1 Ψ′ in (A.4) we now need to
show that hi, ⌊Ψ′ ⌋i , h′ (l)i ∈ Ψ′ (l). We distinguish two cases:
• Case l = ld for some d ∈ D. From (A.2) and (A.3) respectively we get that
h′ (l) = λ(xd )σ(bd )
∧
Ψ′ (l) = ⌊α → τd ⌋k−1
Thus we need to show that
hi, Ψ′ i , λ(xd )σ(bd )i ∈ ⌊α → τd ⌋k−1
(A.5)
∀d ∈ D. λ(xd )σ(bd ) :k,Ψ α → τd
(A.6)
By SemLam in Figure 6 (Lemma 3.15) and the assumption Σ[xd := α] |= bd : τd we
already know that Σ |= λ(xd )bd : α → τd for all d ∈ D. From this and σ :k,Ψ Σ by
Definition 3.8 (Semantic typing judgement) we obtain
Since k > 1 and from (A.1) h :k Ψ, Proposition A.3 shows that (A.6) implies
∀d ∈ D. hk, Ψ, λ(xd )σ(bd )i ∈ α → τd
1, Ψ′ )
(i, ⌊Ψ′ ⌋
(A.7)
By Proposition A.2 we get that (k −
⊑
i ), which together with the first
conjunct of (A.4) and the transitivity of ⊑ yields (k, Ψ) ⊑ (i, ⌊Ψ′ ⌋i ). Since each α → τd
is closed under state extension, the latter property and (A.7) imply the required (A.5).
38
C. HRIŢCU AND J. SCHWINGHAMMER
• Case l ∈ dom(Ψ). From (A.2) and (A.3) respectively we get that h′ (l) = h(l) and
Ψ′ (l) = ⌊Ψ(l)⌋k−1 , so we actually need to show that hi, ⌊Ψ′ ⌋i , h(l)i ∈ ⌊Ψ(l)⌋k−1 . From
h :k Ψ (A.1) by Definition 3.5 we get that hk −1, ⌊Ψ⌋k−1 , h(l)i ∈ Ψ(l). Since Ψ(l) is closed
under state extension and (k − 1, ⌊Ψ⌋k−1 ) ⊑ (i, ⌊Ψ′ ⌋i ), we obtain hi, ⌊Ψ′ ⌋i , h(l)i ∈ Ψ(l).
Finally, we need to show the third conjunct of (A.4), i.e., hk − 1, Ψ′ , {md =ld }d∈D i ∈ α.
To this end, we prove the following more general claim:
Claim: For all j0 ≥ 0, for all Ψ0 and for all {md =ld′ }d∈D
(k − 1, Ψ′ ) ⊑ (j0 , Ψ0 ) ∧ (∀d ∈ D. ⌊Ψ0 ⌋j0 (ld′ ) = Ψ′ j0 (ld ))
⇒ hj0 , ⌊Ψ0 ⌋j0 , md =ld′ d∈D i ∈ α (A.8)
From this and ⌊Ψ′ ⌋k−1 = Ψ′ , (A.4) follows by taking j0 = k − 1, Ψ0 = Ψ′ , and ld′ = ld for
all d ∈ D, and by observing that ⊑ is reflexive (Proposition A.1).
The claim is proved by complete induction on j0 . So assume j0 ≥ 0 and Ψ0 are such
that
(k − 1, Ψ′ ) ⊑ (j0 , Ψ0 )
Moreover, for all d ∈ D let
ld′
(A.9)
∈ dom(Ψ0 ) such that
∀d ∈ D. ⌊Ψ0 ⌋j0 (ld′ ) = Ψ′ j0 (ld )
(A.10)
We show that hj0 , ⌊Ψ0 ⌋j0 , {md =ld′ }d∈D i ∈ α, by checking that all the conditions obtained
by unfolding the definition of α = [md :νd τd ]d∈D hold. Choosing α′ = ⌊α⌋j0 yields (Obj-1):
∃α′ .α′ ∈ Type ∧ α′ j0 ⊆ ⌊α⌋j0
(A.11)
Next, by the construction of Ψ′ in (A.3), together with (A.9), (A.10), and the nonexpansiveness of procedure types, it follows that for all d ∈ D
j
k
⌊Ψ0 ⌋j0 (ld′ ) = Ψ′ j0 (ld ) = ⌊α → τd ⌋j0 = ⌊α⌋j0 → τd
(A.12)
= α′ → τd j0
j0
By the definition of reference types (Definition 3.16) this implies that
∀d ∈ D. hj0 , ⌊Ψ0 ⌋j0 , ld′ i ∈ ref◦ (α′ → τd )
(A.13)
By the lemma for subtyping variance annotations (SemSubVarRef in Figure 7) we then
obtain property (Obj-2):
∀d ∈ D. hj0 , ⌊Ψ0 ⌋j0 , ld′ i ∈ refνd (α′ → τd )
(A.14)
Finally, we must prove (Obj-3), i.e., that for all j < j0 , Ψ1 and
{md =ld′′ }d∈D
(j0 , Ψ0 ) ⊑ (j, Ψ1 ) ∧ (∀d ∈ D. ⌊Ψ1 ⌋j (ld′′ ) = ⌊Ψ0 ⌋j (ld′ ))
⇒ hj, ⌊Ψ1 ⌋j , md =ld′′
d∈D
i ∈ α (A.15)
Note that this last condition holds trivially in the base case of the induction, when j0 = 0.
So assume j < j0 and Ψ1 and ld′′ are such that (j0 , Ψ0 ) ⊑ (j, Ψ1 ) and ⌊Ψ1 ⌋j (ld′′ ) = ⌊Ψ0 ⌋j (ld′ )
for all d ∈ D. Now j < j0 and assumption (A.10) yield that for all d ∈ D
j
k
j
k
⌊Ψ1 ⌋j (ld′′ ) = ⌊Ψ0 ⌋j (ld′ ) = ⌊Ψ0 ⌋j0 (ld′ ) = Ψ′ j0 (ld ) = Ψ′ j (ld )
j
j
A STEP-INDEXED SEMANTICS OF IMPERATIVE OBJECTS
39
Moreover, from (k − 1, Ψ′ ) ⊑ (j0 , Ψ0 ) (A.9) and (j0 , Ψ0 ) ⊑ (j, Ψ1 ), by the transitivity of ⊑
we have that (k − 1, Ψ′ ) ⊑ (j, Ψ1 ). Since j < j0 , the induction hypothesis of the claim gives
hj, ⌊Ψ1 ⌋j , md =ld′′ d∈D i ∈ α
and we have established (A.15).
By Definition 3.20 applied to the object type α = [md :νd τd ]d∈D the properties D ⊆ D,
(A.11), (A.14), and (A.15) establish that indeed hj0 , ⌊Ψ0 ⌋j0 , {md =ld′ }d∈D i ∈ α. This finishes
the inductive proof of claim (A.8), and the proof of the lemma.
Lemma A.5 (SemInv: Method invocation). For all object types α = [md :νd τd ]d∈D and
for all e ∈ D, if Σ |= a : α and νe ∈ {+, ◦}, then Σ |= a.me : τe .
Proof. Let α = [md :νd τd ]d∈D . We assume that e ∈ D, νe ∈ {+, ◦}, and Σ |= a : α and show
that Σ |= a.me : τe . To this end, let k ≥ 0, σ and Ψ such that σ :k,Ψ Σ. From Σ |= a : α by
Definition 3.8 we get that
σ(a) :k,Ψ α
(A.16)
′
We need to show that σ(a).me :k,Ψ τe . Thus, let j < k, and consider heaps h and h and a
term b′ such that the following three conditions are fulfilled:
∧
h :k Ψ
hh, σ(a).me i →j hh ′ , b′ i
∧
hh ′ , b′ i9
(A.17)
From the second and third conjunct of (A.17) by the operational semantics we have that
for some i ≤ j, h ∗ and b∗
hh, σ(a)i →i hh∗ , b∗ i9
∧
hh ∗ , b∗ .mi →j−i hh ′ , b′ i
(A.18)
From the first conjunct together with (A.16) and the first conjunct of (A.17), by Definition 3.6 it follows that there exists a heap typing Ψ∗ such that
(k, Ψ) ⊑ (k − i, Ψ∗ )
∧
h ∗ :k−i Ψ∗
∧
hk − i, Ψ∗ , b∗ i ∈ α = [md :νd τd ]d∈D
By the definition of object types, the latter shows that there exists C and
b∗ = {mc =lc }c∈C , D ⊆ C and (Obj-1) and (Obj-2) hold:
′
α′ ∈ Type ∧
α k−i ⊆ ⌊α⌋k−i
α′
∀d ∈ D. hk − i, Ψ∗ , ld i ∈ refνd (α′ → τd )
as well as (Obj-3): for all j0 < k − i, all Ψ′ and all {mc =lc′ }c∈C ,
((k − i, Ψ∗ ) ⊑ (j0 , Ψ′ ) ∧ ∀c ∈ C. Ψ′ j0 (lc′ ) = ⌊Ψ∗ ⌋j0 (lc )) ⇒ hj0 , Ψ′ j0 , mc =lc′
(A.19)
such that
(A.20)
(A.21)
i ∈ α′
(A.22)
c∈C
From e ∈ D and νe ∈ {+, ◦} using (A.21) we deduce that ⌊Ψ∗ ⌋k−i (le ) ⊆ ⌊α′ → τe ⌋k−i . So
by expanding the definition of h ∗ :k−i Ψ∗ from (A.19) for k − i − 1 < k − i we have
(A.23)
hk − i − 1, ⌊Ψ∗ ⌋k−i−1 , h ∗ (le )i ∈ α′ → τe k−i
By the definition of the procedure type α′ → τe this means in particular that h ∗ (le ) must
be an abstraction, i.e., for some x and a′ , h ∗ (le ) = λ(x)a′ . Thus, since {mc =lc }c∈C ∈ CVal
40
C. HRIŢCU AND J. SCHWINGHAMMER
and e ∈ D ⊆ C, by (A.18), Red-Ctx, Red-Inv, Red-Beta and the operational semantics,
we obtain a reduction sequence of the form
hh, σ(a).me i →i hh ∗ , {mc =lc }c∈C .me i
→ hh∗ , (λ(x)a′ ) {mc =lc }c∈C i
→ hh∗ , {{x 7→ {mc =lc }c∈C }}(a′ )i
(A.24)
→j−i−2 hh′ , b′ i
Since k − i − 2 < k − i by Proposition A.2 we have that
(k − i, Ψ∗ ) ⊑ (k − i − 2, ⌊Ψ∗ ⌋k−i−2 )
(A.25)
We can now use the property (Obj-3) of the object type α: we instantiate (A.22) with
lc′ = lc , j0 = k − i − 2, and Ψ′ = ⌊Ψ∗ ⌋k−i−2 to obtain
hk − i − 2, ⌊Ψ∗ ⌋k−i−2 , {mc =lc }c∈C i ∈ α′
(A.26)
From this using (A.23) and from the definition of procedure types, it follows that
{{x 7→ {mc =lc }c∈C }}(a′ ) :k−i−2,⌊Ψ∗ ⌋k−i−2 τe
(A.27)
On the other hand, the second conjunct of (A.19) implies
h∗ :k−i−2 ⌊Ψ∗ ⌋k−i−2
(A.28)
by Definition 3.5, Proposition A.2 and the closure of types under state extension. Moreover,
by (A.24), hh ∗ , {{x 7→ {mc =lc }c∈C }}(a′ )i →j−i−2 hh ′ , b′ i, which by (A.17) is irreducible. This,
combined with (A.28) and (A.27), by Definition 3.6, means that there exists Ψ′′ such that
(k − i − 2, ⌊Ψ∗ ⌋k−i−2 ) ⊑ (k − j, Ψ′′ )
∧
h′ :k−j Ψ′′
∧
hk − j, Ψ′′ , b′ i ∈ τe
(A.29)
From the first conjunct above, the first conjunct in (A.19), and (A.25), using the transitivity
of state extension we obtain
(k, Ψ) ⊑ (k − j, Ψ′′ )
(A.30)
From (A.17), (A.30), and the second and third conjuncts of (A.29), by Definition 3.6 we
can conclude that σ(a).me :k,Ψ τe holds. This is what we needed to show.
Lemma A.6 (SemUpd: Method update). For all object types α = [md :νd τd ]d∈D and for
all e ∈ D, if Σ |= a : α and Σ[x := α] |= b : τe and νe ∈ {−, ◦}, then Σ |= a.me := ς(x)b : α.
Proof sketch. The proof is similar to that of Lemma A.5 (Method invocation). The existence
of some Ψ∗ such that
hh, σ(a).me := ς(x)σ(b)i →j hh∗ , {me =le }e∈E .me := ς(x)σ(b)i
with (k, Ψ) ⊑ (k − j, Ψ∗ ), h∗ :k−j Ψ∗ and hk − j, Ψ∗ , {me =le }e∈E i ∈ α follows from the
existence of a corresponding reduction sequence hh, σ(a)i →j hh∗ , {me =le }e∈E i. Since the
only reduction from hh∗ , {me =le }e∈E .me := ς(x)σ(b)i is by (Red-Upd) and results in the
configuration hh′ , {me =le }e∈E i where
h′ = h∗ [le := λ(x)σ(b)]
the proof of the lemma is essentially a matter of showing that h′ :k−j−1 ⌊Ψ∗ ⌋k−j−1 .
(A.31)
A STEP-INDEXED SEMANTICS OF IMPERATIVE OBJECTS
41
First, note that dom(Ψ∗ ) ⊆ dom(h′ ) = dom(h∗ ) holds, by h∗ :k−j Ψ∗ . Next, let
i < k − j − 1, and let l ∈ dom(Ψ∗ ). It remains to show that
hi, ⌊Ψ∗ ⌋i , h′ (l)i ∈ ⌊Ψ∗ (l)⌋k−j−1
Note that by the definition of hk − j, Ψ∗ , {me =le }e∈E i ∈ α (Definition 3.20)
there exists α′ ∈ Type such that ⌊α′ ⌋k−j ⊆ ⌊α⌋k−j , that D ⊆ E, and that
∀d ∈ D. hk − j, Ψ∗ , ld i ∈ refνd (α′ → τd )
(A.32)
it follows that
(A.33)
We now prove (A.32) by a case distinction on the location l:
• Case l = le . From (A.31) we have that h′ (le ) = λ(x)σ(b). Since e ∈ D ⊆ E and
νe ∈ {−, ◦} by assumption, (A.33) yields ⌊α′ → τe ⌋k−j ⊆ ⌊Ψ∗ ⌋k−j (le ). Since ⌊α′ ⌋k−j ⊆
⌊α⌋k−j , the subtyping lemma (SemSubProc) and the non-expansiveness of procedure
types yield ⌊α → τe ⌋k−j ⊆ ⌊Ψ∗ ⌋k−j (le ). The monotonicity of semantic approximation
therefore entails
⌊α → τe ⌋k−j−1 ⊆ ⌊Ψ∗ ⌋k−j−1 (le )
(A.34)
Additionally, the assumption Σ[x := α] |= b : τe gives λ(x)σ(b) :k−j,Ψ∗ α → τe . Since
0 ≤ i < k − j and h∗ :k−j Ψ∗ , Proposition A.3 yields hk − j, Ψ∗ , λ(x)σ(b)i ∈ α → τe . By
the closure under state extension, this implies hi, ⌊Ψ∗ ⌋i , λ(x)σ(b)i ∈ α → τe , from which
(A.32) follows by (A.34).
• Case l 6= le . This case is easier since the value in the heap does not change for this
location, i.e., h′ (l) = h∗ (l), so the result follows from the closure under state extension of
Ψ∗ (l).
Lemma A.7 (SemClone: Object cloning). For all object types α = [md :νd τd ]d∈D , if
Σ |= a : α then Σ |= clone a : α.
Proof sketch. The proof is similar to that of Lemma A.6 (Method update). Assuming
σ :k,Ψ Σ and h :k Ψ such that hh, clone σ(a)i halts in fewer than k steps, by appealing to
the operational semantics and the assumption that Σ |= a : α one obtains the existence of
some Ψ∗ such that
hh, clone σ(a)i →j hh∗ , clone {me =le }e∈E i → hh′ , b′ i
with (k, Ψ) ⊑ (k − j, Ψ∗ ), h∗ :k−j Ψ∗ and hk − j, Ψ∗ , {me =le }e∈E i ∈ α. Since the only reduction from hh∗ , clone {me =le }e∈E i is by (Red-Clone) it is clear that for some (distinct)
/ dom(h∗ ) we have
le′ ∈
h′ = h∗ le′ := h∗ (le ) e∈E ∧ b′ = me =le′ e∈E
(A.35)
If we set Ψ′ = Ψ∗ [le′ := Ψ∗ (le )]e∈E k−j−1 then it follows that (k, Ψ) ⊑ (k − j − 1, Ψ′ ), and
to establish the lemma it suffices to prove
h′ :k−j−1 Ψ′ ∧ hk − j − 1, Ψ′ , me =le′ e∈E i ∈ α
(A.36)
Observing that dom(Ψ′ ) ⊆ dom(h′ ) is satisfied, the first conjunct is proved by showing
that hi, ⌊Ψ′ ⌋i , h′ (l)i ∈ ⌊Ψ′ (l)⌋k−j−1 holds for all i < k − j − 1 and all l ∈ dom(Ψ′ ). This is
done by a case distinction on whether l ∈ dom(Ψ∗ ) or l = le′ for some e. In both cases, the
relation follows from h∗ :k−j Ψ∗ and the closure under state extension of types.
As for the second conjunct of (A.36), we note that Ψ′ is constructed from Ψ∗ such
that ⌊Ψ′ (le′ )⌋k−j−1 = ⌊Ψ∗ (le )⌋k−j−1 holds for all e ∈ E. Therefore, by unfolding the
definition of the object types for hk − j, Ψ∗ , {me =le }e∈E i ∈ α, condition (Obj-3) allows
42
C. HRIŢCU AND J. SCHWINGHAMMER
us to conclude that hk − j − 1, ⌊Ψ′ ⌋k−j−1 , {me =le′ }e∈E i ∈ ⌊α⌋k−j−1 . Then the required
hk − j − 1, Ψ′ , {me =le′ }e∈E i ∈ α follows from the fact that ⌊Ψ′ ⌋k−j−1 = Ψ′ holds by definition of Ψ′ , and that ⌊α⌋k−j−1 ⊆ α.
A.3. Subtyping Lemmas for Object Types.
Lemma A.8 (SemSubObj: Subtyping object types). E ⊆ D and for all e ∈ E if
νe ∈ {+, ◦} then αe ⊆ βe and if νe ∈ {−, ◦} then βe ⊆ αe imply that [md :νd αd ]d∈D ⊆
[me :νe βe ]e∈E .
Proof. We denote α = [md :νd αd ]d∈D and β = [me :νe βe ]e∈E . We assume that E ⊆ D and
∀e ∈ E. (νe ∈ {+, ◦} ⇒ αe ⊆ βe ) ∧ (νe ∈ {−, ◦} ⇒ βe ⊆ αe ),
(A.37)
and prove that for all heap typings Ψ, for all values v and all k ≥ 0, if hk, Ψ, vi ∈ α then
hk, Ψ, vi ∈ β, by complete induction on k. The induction hypothesis is that for all j < k if
hj, Ψ, vi ∈ α then hj, Ψ, vi ∈ β, or equivalently ⌊α⌋k ⊆ ⌊β⌋k .
If we assume that hk, Ψ, vi ∈ α, then by the definition of α (Definition 3.20) we have
that v = {mc =lc }c∈C , D ⊆ C and there exists α′ ∈ Type such that ⌊α′ ⌋k ⊆ ⌊α⌋k and
∀d ∈ D. hk, Ψ, ld i ∈ refνd (α′ → αd )
(A.38)
Moreover, condition (Obj-3) holds with respect to α, i.e., for all j < k, all Ψ′ and all
{me =le′ }e∈E such that (k, Ψ) ⊑ (j, Ψ′ ),
(∀e ∈ E. Ψ′ j (le′ ) = ⌊Ψ⌋j (le )) ⇒ hj, Ψ′ j , me =le′ e∈E i ∈ α′
(A.39)
From E ⊆ D and D ⊆ C by transitivity E ⊆ C. From ⌊α′ ⌋k ⊆ ⌊α⌋k and the induction
hypothesis ⌊α⌋k ⊆ ⌊β⌋k we get that ⌊α′ ⌋k ⊆ ⌊β⌋k , i.e., (Obj-1) holds. Moreover, (A.39)
entails that condition (Obj-3) also holds with respect to the object type β. So in order
to conclude that hk, Ψ, vi ∈ β, and therefore that α ⊆ β, all that remains to be proven is
condition (Obj-2):
∀e ∈ E. hk, Ψ, le i ∈ refνe (α′ → βe )
For this, we choose some e in E and do a case analysis on the variance annotation νe :
• Case νe = +. By (A.37) we deduce that αe ⊆ βe , thus by the covariance of the procedure
type constructor in its second argument (SemSubProc in Figure 6) we get that α′ →
αe ⊆ α′ → βe . But since E ⊆ D from (A.38) we know that hk, Ψ, le i ∈ ref+ (α′ → αe ).
Since the type constructor ref+ is covariant (SemSubCovRef in Figure 7) this implies
hk, Ψ, le i ∈ ref+ (α′ → βe ).
• Case νe = −. Similarly to the previous case, (A.37) gives us that βe ⊆ αe . Again by
the covariance of λξ. α′ → ξ (SemSubProc) we infer that α′ → βe ⊆ α′ → αe . From
(A.38) hk, Ψ, le i ∈ ref− (α′ → αe ), so by the contravariance of ref− (SemSubConRef in
Figure 7) we get that hk, Ψ, le i ∈ ref− (α′ → βe ).
• νe = ◦. Now (A.37) entails that αe = βe . Since hk, Ψ, le i ∈ ref◦ (α′ → αe ) by (A.38) we
immediately obtain that also hk, Ψ, le i ∈ ref◦ (α′ → βe ).
A STEP-INDEXED SEMANTICS OF IMPERATIVE OBJECTS
43
Lemma A.9 (SemSubObjVar: Subtyping object variances). If for all d ∈ D we have
νd = ◦ or νd = νd′ then [md :νd τd ]d∈D ⊆ [md :νd′ τd ]d∈D .
Proof. The proof proceeds similarly to the proof of Lemma A.8.
[md :νd τd ]d∈D and α′ = [md :νd′ τd ]d∈D . We assume that
Let us denote α =
∀d ∈ D. νd = ◦ ∨ νd = νd′
(A.40)
Let Ψ and v be arbitrary. We prove that for all k ≥ 0, if hk, Ψ, vi ∈ α then hk, Ψ, vi ∈ α′ ,
by complete induction on k. The induction hypothesis is that for all j < k if hj, Ψ, vi ∈ α
then hj, Ψ, vi ∈ α′ , or equivalently ⌊α⌋k ⊆ ⌊α′ ⌋k .
Assume that hk, Ψ, vi ∈ α, then by the definition of α we have that v = {me =le }e∈E ,
D ⊆ E, and there exists a type α′′ such that ⌊α′′ ⌋k ⊆ ⌊α⌋k and
∀d ∈ D. hk, Ψ, ld i ∈ refνd (α′′ → τd )
(A.41)
Moreover, condition (Obj-3) holds.
From ⌊α′′ ⌋k ⊆ ⌊α⌋k and the induction hypothesis ⌊α⌋k ⊆ ⌊α′ ⌋k by transitivity we get
that ⌊α′′ ⌋k ⊆ ⌊α′ ⌋k . This choice of α′′ also shows that (Obj-3) holds for hk, Ψ, vi with
respect to α′ . So in order to show that hk, Ψ, vi ∈ α′ , and therefore that α ⊆ α′ , all that
remains to be proven is that:
∀d ∈ D. hk, Ψ, ld i ∈ refνd′ (α′′ → τd )
We show this by case analysis on the disjunction in (A.40). Both cases are trivial:
• Case νd = ◦. From (A.41) and ref◦ (α′′ → τd ) ⊆ refνd′ (α′′ → τd ) (SemSubVarRef in
Figure 7) it is immediate that hk, Ψ, ld i ∈ refνd′ (α′′ → τd ).
• Case νd = νd′ , then the required statement is the same as (A.41).
A.4. Typing Lemmas with Structural Assumptions for Self Types.
Lemma A.10 (SemUpd-Str: Method update with structural assumptions). For all object
types α = [md :νd Fd ]d∈D and all α′ ∈ Type such that α′ ⊳ α, if e ∈ D, νe ∈ {−, ◦} and
Σ |= a : α′ and Σ[x := α′ ] |= b : Fe (α′ ), then Σ |= a.me := ς(x)b : α′ .
Proof. The proof is an adaptation of the proof given for Lemma A.6 (Method update)
above. Assume α = [md :νd Fd ]d∈D , e ∈ D, and νe ∈ {−, ◦}, and let α′ ∈ Type such that
α′ ⊳ α. Moreover assume that Σ |= a : α′ and Σ[x := α′ ] |= b : Fe (α′ ) hold. We show that
Σ |= a.me := ς(x)b : α′ .
Let k ≥ 0, σ be a value environment and Ψ be a heap typing such that σ :k,Ψ Σ. We
must prove that σ(a.me := ς(x)b) :k,Ψ α′ , so let h and j < k be such that
h :k Ψ
hh, σ(a).m e := ς(x)σ(b)i →j hh ′ , a′ i
∧
∧
hh ′ , a′ i9
→i
(A.42)
hh ′′ , a′′ i
By the operational semantics, this sequence is induced by hh, σ(a)i
for i ≤ j
′′
′′
′
′′
and some h and a , and by the assumption Σ |= a : α there exists some Ψ such that
(k, Ψ) ⊑ (k − i, Ψ′′ )
∧
h ′′ :k−i Ψ′′
∧
hk − i, Ψ′′ , a′′ i ∈ α′ ⊆ [md :νd Fd ]d∈D
(A.43)
In particular, a′′ is of the form {me =le }e∈E for some E ⊇ D, and by the operational semantics hh ′′ , a′′ .me := ς(x)σ(b)i → hh ′ , a′′ i. In particular, a′ is a′′ and h′ is h′′ [le := λ(x)σ(b)].
By choosing Ψ′ = ⌊Ψ′′ ⌋k−j , the first and last conjuncts of (A.43) yield
(k, Ψ) ⊑ (k − j, Ψ′ )
∧
hk − j, Ψ′ , a′′ i ∈ α′
44
C. HRIŢCU AND J. SCHWINGHAMMER
by Proposition A.2 and transitivity, and by closure under state extension of α′ . To establish
the lemma, it remains to show that h′ :k Ψ′ . For l ∈ dom(Ψ′ ) − {le } this follows from the
second conjunct of (A.43) by the closure under state extension. The interesting case is when
l = le and we must prove h′ (l) = λ(x)σ(b) :k−j Ψ′ (l). Since α′ ⊳ α and hk − j, Ψ′ , a′′ i ∈
α′ , condition (Obj-2-self) in Definition 5.2 (Self type exposure) yields hk − j, Ψ′ , le i ∈
refνe (α′ → Fe (α′ )). By assumption, νe ∈ {−, ◦} so Ψ′ (l) ⊇ ⌊α′ → Fe (α′ )⌋k−j holds by the
definition of refνe . Hence it suffices to prove that
λ(x)σ(b) :k−j,Ψ′ α′ → Fe (α′ )
which follows from the assumption Σ[x := α′ ] |= b : Fe (α′ ).
Proposition A.11 (Self type exposure). Let α be a self type and suppose hk, Ψ, vi ∈ α.
Then there exists α′′ ∈ Type such that α′′ ⊳ α and hk − 1, ⌊Ψ⌋k−1 , vi ∈ α′′ .
Proof. Suppose hk, Ψ, vi ∈ α = [md :νd Fd ]d∈D . By Definition 5.1 (Self types), this means
that there exists α′ ∈ Type such that ⌊α′ ⌋k ⊆ ⌊α⌋k and conditions (Obj-2-Self) and (Obj3) are satisfied. Choosing α′′ = ⌊α′ ⌋k , it is clear that α′′ ⊳ α since all the conditions only
rely on α′ to approximation k. Moreover, by instantiating Ψ′ = Ψ and {me =le′ }e∈E = v in
(Obj-3) we obtain that hk − 1, ⌊Ψ⌋k−1 , vi ∈ α′′ . This proves the proposition.
Lemma A.12 (SemLet-Str: Introducing structural assumptions). Let α = [md :νd Fd ]d∈D
and suppose that Σ |= a : α and that Σ[x := ξ] |= b : β for all ξ ∈ Type with ξ ⊳ α. Then
Σ |= let x = a in b : β.
Proof. Let α = [md :νd Fd ]d∈D and suppose that Σ |= a : α and that Σ[x := ξ] |= b : β for all
ξ ∈ Type with ξ ⊳ α. We must show that Σ |= let x = a in b : β. Thus, let k ≥ 0, Ψ and σ
be such that σ :k,Ψ Σ. By the definition of the semantic typing judgement (Definition 3.8)
we must show that σ(let x = a in b) :k,Ψ β, or equivalently (after suitable α-renaming and
removing the syntactic sugar) that
(λ(x)σ(b)) σ(a) :k,Ψ β
Suppose j < k, h, h′ and b′ are such that
h :k Ψ
∧
hh, (λ(x)σ(b)) σ(a)i →j hh′ , b′ i
∧
hh′ , b′ i9
(A.44)
From the second and third conjunct of (A.44) by the operational semantics we have that
for some i ≤ j < k, some h′′ and some Ψ′′ ,
hh, σ(a)i →i hh′′ , a′′ i9
hh ′′ , (λ(x)σ(b)) a′′ i →j−i hh ′ , b′ i
∧
(A.45)
From the first conjunct together with the assumption Σ |= a : α and the first conjunct of
(A.44), by Definition 3.6 it follows that there exists a heap typing Ψ′′ such that
(k, Ψ) ⊑ (k − i, Ψ′′ )
∧
h ′′ :k−i Ψ′′
∧
hk − i, Ψ′′ , a′′ i ∈ α = [md :νd Fd ]d∈D
(A.46)
In particular, a′′ ∈ Val and the operational semantics gives
hh, (λ(x)σ(b)) (σ(a))i →i hh′′ , (λ(x)σ(b)) a′′ i → hh ′′ , σ x := a′′ (b)i →j−i−1 hh ′ , b′ i (A.47)
From the third conjunct of (A.46) by Proposition A.11 there exists α′ ∈ Type such that
α′ ⊳ α and hk − i − 1, ⌊Ψ′′ ⌋k−i−1 , a′′ i ∈ α′ . From σ :k,Ψ Σ, the first conjunct of (A.46),
Propositions A.1 and A.2 and the closure under state extension, this yields
σ x := a′′ :k−i−1,⌊Ψ′′ ⌋k−i−1 Σ x := α′
(A.48)
A STEP-INDEXED SEMANTICS OF IMPERATIVE OBJECTS
45
Since α′ ⊳ α, by instantiating the universally quantified type ξ in the hypothesis on b we
obtain that Σ[x := α′ ] |= b : β. Therefore, (A.48) gives σ[x := a′′ ](b) :k−i−1,⌊Ψ′′ ⌋k−i−1 β.
Clearly h′′ :k−i−1 ⌊Ψ′′ ⌋k−i−1 by the second conjunct of (A.46), so that the second conjunct
of (A.45) shows that there is some Ψ′ such that (k, Ψ) ⊑ (k −i−1, ⌊Ψ′′ ⌋k−i−1 ) ⊑ (k − j, Ψ′ ),
h ′ :k−j Ψ′ and hk−j, Ψ′ , b′ i ∈ β, by Definition 3.6. This establishes that σ(let x = a in b) :k,Ψ
β holds as required.
We next define a recursive type of records {| md :νd Fd |}d∈D , which is the type arising
from the recursive record interpretation of (imperative) objects [18]. While this type does
not give rise to non-trivial subtyping, we will show that it satisfies {| md :νd Fd |}d∈D ⊳
[md :νd Fd ]d∈D .
Definition A.13. Assume Fd : Type → Type are monotonic and non-expansive type constructors, for all d ∈ D. Then let β = {| md :νd Fd |}d∈D be defined as the set of all triples
hk, Ψ, {md =ld }d∈D i such that
(∀d ∈ D. hk, Ψ, ld i ∈ refνd (β → Fd (β)))
∧ (∀j < k. ∀Ψ′ . ∀ me =le′ e∈E .
(k, Ψ) ⊑ (j, Ψ′ ) ∧ (∀d ∈ D. Ψ′ j (ld′ ) = ⌊Ψ⌋j (ld )) ⇒ hj, Ψ′ j , md =ld′
(Rec-1)
(Rec-2)
d∈D
i ∈ β)
Note that the recursive specification of β is well-founded, i.e., β is well-defined. Moreover,
β is a type, i.e., it is closed under state extension.
Proposition A.14. For all self types [md :νd Fd ]d∈D we have that
{| md :νd Fd |}d∈D ⊳ [md :νd Fd ]d∈D
Proof. Let α = [md :νd Fd ]d∈D and β = {| md :νd Fd |}d∈D for some arbitrary monotonic
and non-expansive type constructors Fd . It is clear that for all hk, Ψ, {md =ld }d∈D i ∈ β,
conditions (Obj-2-Self) and (Obj-3) from Definition 5.2 are satisfied, by the definition of
β (Definition A.13). It remains to prove that β ⊆ α. We establish this by showing that for
all k ≥ 0, ⌊β⌋k ⊆ ⌊α⌋k , by complete induction on k. Let hk, Ψ, {md =ld }d∈D i ∈ β; we need
to show that hk, Ψ, {md =ld }d∈D i ∈ α. We have that D ⊆ D and we choose α′ = β which
is a type and fulfills ⌊β⌋k ⊆ ⌊α⌋k (Obj-1) by the induction hypothesis. The conditions
(Obj-2-Self) and (Obj-3) in Definition 5.1 (Self types) are exactly the same as conditions
(Rec-1) and (Rec-2) in the definition of β (Definition A.13), which concludes the proof.
Lemma A.15 (SemObj-Str: Object construction with structural assumptions). Let α =
[md :νd Fd ]d∈D and suppose that for all d ∈ D and all ξ ∈ Type with ξ ⊳ α, Σ[x := ξ] |= bd :
Fd (ξ). Then Σ |= [md =ς(xd )bd ]d∈D : α.
Proof. Let α = [md :νd Fd ]d∈D and assume that
∀d∈D. ∀ξ∈Type. ξ ⊳ α ⇒ Σ[xd := ξ] |= bd : Fd (ξ)
(A.49)
We must show that Σ |= [md =ς(xd )bd ]d∈D : α. Thus, let k ≥ 0, σ be a value environment
and Ψ be a heap typing such that σ :k,Ψ Σ. By Definition 3.8 we need to show that
σ([md =ς(xd )bd ]d∈D ) :k,Ψ α. Equivalently (after suitable α-renaming), we show that
[md =ς(xd )σ(bd )]d∈D :k,Ψ α
Suppose j <
k, h, h′
and
h :k Ψ
∧
b′
are such that the following three conditions are fulfilled:
hh, [md =ς(xd )σ(bd )]d∈D i →j hh′ , b′ i
∧
hh′ , b′ i9
(A.50)
46
C. HRIŢCU AND J. SCHWINGHAMMER
By the operational semantics Red-Obj is the only rule that applies, which means that
necessarily j = 1 and for some distinct ld 6∈ dom(h) we have b′ = {md =ld }d∈D and
h′ = h [ld := λ(xd )σ(bd )]d∈D
(A.51)
Let β = {| md :νd Fd |}d∈D , as in Definition A.13. We choose
Ψ′ = Ψ [ld := (β → Fd (β))]d∈D k−1
(A.52)
and show that
(k, Ψ) ⊑ (k − 1, Ψ′ )
∧
h′ :k−1 Ψ′
∧
hk − 1, Ψ′ , b′ i ∈ α
(A.53)
Ψ′
The first conjunct of (A.53) holds by the construction of
(A.52). In order to show the
second conjunct, let i < k − 1 and l ∈ dom(Ψ′ ). We now need to show that hi, ⌊Ψ′ ⌋i , h′ (l)i ∈
Ψ′ (l). In case l ∈ dom(Ψ) the proof proceeds exactly as for Lemma A.4, so we only consider
the case when l = ld for some d ∈ D. From (A.51) and (A.52) we get that
h′ (l) = λ(xd )σ(bd )
∧
Ψ′ (l) = ⌊β → Fd (β)⌋k−1
Thus we need to show that
hi, Ψ′ i , λ(xd )σ(bd )i ∈ ⌊β → Fd (β)⌋k−1
(A.54)
By Proposition A.14 we obtain that β ⊳ α, so we can instantiate the universally quantified
ξ in (A.49) with β and obtain that
Σ[xd := β] |= bd : Fd (β)
By SemLam in Figure 6 (Lemma 3.15) this gives us that
Σ |= λ(xd )bd : β → Fd (β)
From this and σ :k,Ψ Σ by Definition 3.8 we obtain
λ(xd )σ(bd ) :k,Ψ β → Fd (β)
(A.55)
Since k > 1 and from (A.50) h :k Ψ, Proposition A.3 shows that (A.55) implies
hk, Ψ, λ(xd )σ(bd )i ∈ β → Fd (β)
(A.56)
By Proposition A.2 we get that (k−1, Ψ′ ) ⊑ (i, ⌊Ψ′ ⌋i ), which together with the first conjunct
of (A.53) and the transitivity of ⊑ yields (k, Ψ) ⊑ (i, ⌊Ψ′ ⌋i ). Since each β → Fd (β) is closed
under state extension, the latter property and (A.56) imply the required (A.54).
Finally, we need to show the third conjunct of (A.53), i.e., hk − 1, Ψ′ , {md =ld }d∈D i ∈ α.
To this end, we prove the following more general claim:
Claim: For all j0 ≥ 0, for all Ψ0 and for all {md =ld′ }d∈D
(k − 1, Ψ′ ) ⊑ (j0 , Ψ0 ) ∧ (∀d ∈ D. ⌊Ψ0 ⌋j0 (ld′ ) = Ψ′ j0 (ld ))
⇒ hj0 , ⌊Ψ0 ⌋j0 , md =ld′ d∈D i ∈ β (A.57)
From this and ⌊Ψ′ ⌋k−1 = Ψ′ , the last conjunct of (A.53) follows by taking j0 = k − 1,
Ψ0 = Ψ′ , and ld′ = ld for all d ∈ D, and by observing that ⊑ is reflexive (Proposition A.1)
and β ⊆ α (since β ⊳ α).
The claim above is proved by complete induction on j0 . So assume j0 ≥ 0 and Ψ0 are
such that
(k − 1, Ψ′ ) ⊑ (j0 , Ψ0 )
(A.58)
A STEP-INDEXED SEMANTICS OF IMPERATIVE OBJECTS
47
Moreover, for all d ∈ D let ld′ ∈ dom(Ψ0 ) such that
⌊Ψ0 ⌋j0 (ld′ ) = Ψ′ j0 (ld )
(A.59)
We show that hj0 , ⌊Ψ0 ⌋j0 , {md =ld′ }d∈D i ∈ β, by checking that the two conditions from the
definition of β (Definition A.13) are satisfied. By the construction of Ψ′ in (A.52), together
with (A.58) and (A.59), it follows that for all d ∈ D
⌊Ψ0 ⌋j0 (ld′ ) = Ψ′ j0 (ld ) = ⌊β → Fd (β)⌋j0
(A.60)
By the definition of reference types (Definition 3.16) this implies that
∀d ∈ D. hj0 , ⌊Ψ0 ⌋j0 , ld′ i ∈ ref◦ (β → Fd (β))
(A.61)
By the lemma for subtyping reference types (SemSubVarRef in Figure 7) we then obtain
property (Rec-1):
∀d ∈ D. hj0 , ⌊Ψ0 ⌋j0 , ld′ i ∈ refνd (β → Fd (β))
(A.62)
Second, we must prove (Rec-2), i.e., that for all j < j0 , Ψ1 and {md =ld′′ }d∈D
(j0 , Ψ0 ) ⊑ (j, Ψ1 ) ∧ (∀d ∈ D. ⌊Ψ1 ⌋j (ld′′ ) = ⌊Ψ0 ⌋j (ld′ ))
⇒ hj, ⌊Ψ1 ⌋j , md =ld′′
d∈D
i ∈ β (A.63)
Note that this last condition holds vacuously in the base case of the induction, when j0 = 0.
So assume j < j0 and Ψ1 and ld′′ are such that (j0 , Ψ0 ) ⊑ (j, Ψ1 ) and ⌊Ψ1 ⌋j (ld′′ ) = ⌊Ψ0 ⌋j (ld′ )
for all d ∈ D. Now j < j0 and assumption (A.59) yield that for all d ∈ D
k
j
j
k
⌊Ψ1 ⌋j (ld′′ ) = ⌊Ψ0 ⌋j (ld′ ) = ⌊Ψ0 ⌋j0 (ld′ ) = Ψ′ j0 (ld ) = Ψ′ j (ld )
j
j
1, Ψ′ )
Moreover, from (k −
⊑ (j0 , Ψ0 ) (A.58) and (j0 , Ψ0 ) ⊑ (j, Ψ1 ), by the transitivity of ⊑
we have that (k − 1, Ψ′ ) ⊑ (j, Ψ1 ). Since j < j0 , the induction hypothesis of the claim gives
hj, ⌊Ψ1 ⌋j , md =ld′′ d∈D i ∈ β
and we have established (A.63).
By Definition A.13 applied to the type β = {| md :νd Fd |}d∈D the properties (A.62),
and (A.63) establish that indeed hj0 , ⌊Ψ0 ⌋j0 , {md =ld′ }d∈D i ∈ β. This finishes the inductive
proof of claim (A.57), and the proof of the lemma.
A.5. Subtyping Lemma for Generalized Object Types.
Lemma A.16 (SemSubGen-Obj: Subtyping generalized object types). If E ⊆ D and for
r
r
w
r
w
r
all e ∈ E we have that βew ⊆ αw
e and αe ⊆ βe then [md : (αd , αd ) ]d∈D ⊆ [me : (βe , βe ) ]e∈E .
w
r
r
Proof. Denote α = [md : (αw
d , αd ) ]d∈D , β = [me : (βe , βe ) ]e∈E , and assume E ⊆ D and
r
r
∀e ∈ E. (βew ⊆ αw
e ∧ αe ⊆ βe ).
(A.64)
We prove that for all heap typings Ψ, for all values v and all k ≥ 0, if hk, Ψ, vi ∈ α then
hk, Ψ, vi ∈ β, by complete induction on k. The induction hypothesis is that ⌊α⌋k ⊆ ⌊β⌋k .
If we assume that hk, Ψ, vi ∈ α, then by the definition of α (Definition 3.20 with
condition (Obj-2-Gen) instead of (Obj-2)) we have that v = {mc =lc }c∈C , D ⊆ C and
there exists α′ ∈ Type such that ⌊α′ ⌋k ⊆ ⌊α⌋k and
′
r
∀d ∈ D. hk, Ψ, ld i ∈ ref(α′ → αw
d , α → αd )
(A.65)
48
C. HRIŢCU AND J. SCHWINGHAMMER
Moreover, condition (Obj-3) holds with respect to α, i.e., for all j < k, all Ψ′ and all
{me =le′ }e∈E such that (k, Ψ) ⊑ (j, Ψ′ ),
(∀e ∈ E. Ψ′ j (le′ ) = ⌊Ψ⌋j (le )) ⇒ hj, Ψ′ j , me =le′ e∈E i ∈ α′
(A.66)
From E ⊆ D and D ⊆ C by transitivity E ⊆ C. From ⌊α′ ⌋k ⊆ ⌊α⌋k and the induction
hypothesis ⌊α⌋k ⊆ ⌊β⌋k we get that ⌊α′ ⌋k ⊆ ⌊β⌋k , i.e., (Obj-1) holds. Moreover, (A.66)
entails that condition (Obj-3) also holds with respect to the object type β. So in order to
conclude that hk, Ψ, vi ∈ β, all that remains to be proven is condition (Obj-2-Gen):
∀e ∈ E. hk, Ψ, le i ∈ ref(α′ → βew , α′ → βer )
(A.67)
Let e ∈ E. Since the procedure type constructor is covariant in the result type (SemSubProc in Figure 6) assumption A.64 implies that
′
r
′
r
α′ → βew ⊆ α′ → αw
e ∧ α → αe ⊆ α → βe
From this by (SemSubRef-Gen) we get that
′
r
′
w
′
r
ref(α′ → αw
e , α → αe ) ⊆ ref(α → βe , α → βe )
This together with A.65 and E ⊆ D directly implies A.67, which concludes the proof.
This work is licensed under the Creative Commons Attribution-NoDerivs License. To view
a copy of this license, visit http:// reative ommons.org/li enses/by-nd/2.0/ or send a
letter to Creative Commons, 171 Second St, Suite 300, San Francisco, CA 94105, USA, or
Eisenacher Strasse 2, 10777 Berlin, Germany
| 6cs.PL
|
Submitted to IEEE Transactions on Signal Processing
Efficient L1-Norm Principal-Component Analysis
via Bit Flipping
Panos P. Markopoulos†, Sandipan Kundu▽ , Shubham Chamadia‡ , and Dimitris A. Pados‡∗
arXiv:1610.01959v1 [cs.DS] 6 Oct 2016
†
Department of Electrical and Microelectronic Engineering
Rochester Institute of Technology
Rochester, NY 14623 USA
E-mail: [email protected]
▽
Qualcomm Technologies, Inc.
San Jose, CA 95110 USA
E-mail: [email protected]
‡
Department of Electrical Engineering
State University of New York at Buffalo
Buffalo, NY 14260 USA
E-mail: {shubhamc, pados}@buffalo.edu
EDICS: MLR-ICAN, MLR-LEAR, MLR-PATT, MDS-ALGO, SSP-SSAN
Submitted: September 22, 2016
Abstract
It was shown recently that the K L1-norm principal components (L1-PCs) of a real-valued data matrix X ∈
RD×N (N data samples of D dimensions) can be exactly calculated with cost O(2N K ) or, when advantageous,
O(N dK−K+1 ) where d = rank(X), K < d [1], [2]. In applications where X is large (e.g., “big” data of large N
and/or “heavy” data of large d), these costs are prohibitive. In this work, we present a novel suboptimal algorithm
for the calculation of the K < d L1-PCs of X of cost O(N Dmin{N, D} + N 2 (K 4 + dK 2 ) + dN K 3 ), which
is comparable to that of standard (L2-norm) PC analysis. Our theoretical and experimental studies show that the
proposed algorithm calculates the exact optimal L1-PCs with high frequency and achieves higher value in the L1-PC
optimization metric than any known alternative algorithm of comparable computational cost. The superiority of the
calculated L1-PCs over standard L2-PCs (singular vectors) in characterizing potentially faulty data/measurements is
demonstrated with experiments on data dimensionality reduction and disease diagnosis from genomic data.
Index Terms — Dimensionality reduction, data analytics, eigen-decomposition, L1 norm, L2 norm, machine
learning, outlier resistance, principal component analysis, subspace signal processing.
∗
Corresponding author.
A part of this paper was presented in IEEE Intern. Conf. on Acoust., Speech, and Signal Processing (ICASSP), Florence, Italy, May 2014
[3]. This work was supported in part by the National Science Foundation under Grant ECCS-1462341 and the Office of the Vice President
for Research of Rochester Institute of Technology.
I. I NTRODUCTION
Principal-Component Analysis (PCA) [4] has been a “mainstay” of signal processing, machine learning, pattern
recognition, and classification [5]–[7] for more than a century. Numerous important applications of PCA can
be found in the fields of wireless communications, computer networks, computer vision, image processing, bioinformatics/genomics, and neuroscience, to name a few. Broadly speaking, PCA seeks to calculate orthogonal
directions that define a subspace wherein data presence is maximized. Traditionally, PCA quantifies data presence
by the Frobenius (L2) norm of the projected data onto the subspace, or, equivalently, the Euclidean distance of the
original data from their subspace representations (L2-norm of residual error). Herein, we will be referring to standard
PCA as L2-PCA. Key strengths of L2-PCA are (i) its low-complexity implementation (quadratic in the number of
data points) by means of singular-value decomposition (SVD) of the data matrix and (ii) the reliable approximation
that it offers to the nominal principal data subspace, when calculated over sufficiently many clean/nominal or
benign-noise corrupted data points.
In the advent of the big-data era, datasets often include grossly corrupted, highly deviating, irregular data points
(outliers), due to a variety of causes such as transient sensor malfunctions, errors in data transmission/transcription,
errors in training data labeling, and bursty-noise data corruption, to name a few [8], [9]. Regretfully, standard
L2-PCA is well-known to be fragile in the presence of such faulty data, even when they appear in a vanishingly
small fraction of the training set [10]. The reason is that the L2-norm objective of standard PCA (minimization
of error variance or maximization of squared projection magnitude) gives squared importance on the magnitude of
every datum, thus overemphasizing peripheral data points.
To remedy the impact of outliers, researchers from the fields of data analysis and signal processing have long
focused on calculating subspaces of minimum absolute error deviations, instead of minimum error variances.
Important early theoretical studies date back to the 1940s [11]–[14]. In the past decade, there have been several
L1-norm-based principal component calculators under the general label “L1-PCA” [3], [15]–[37]. In these works,
researchers seek either to (i) minimize the aggregate absolute data representation error or (ii) maximize the aggregate
absolute magnitude of the projected data points. For approach (i), it has been shown that the error surface is
non-smooth and the problem non-convex, resisting attempts to reach an exact solution even with exponential
computational cost [38]. Therefore, only suboptimal algorithms exist in the literature [15]–[18]. For approach (ii),
a problem known as maximum-projection L1-PCA, the works in [1], [2] proved that maximum-projection L1PCA is not NP-hard for fixed data dimension D and offered the first two optimal algorithms in the literature
for exact calculation. Specifically, for a data record of size N , X ∈ RD×N , the first optimal algorithm solves
2
L1-PCA exactly with complexity O(2N K ); the second optimal algorithm solves L1-PCA exactly with complexity
O(N rank(X)K−K+1 ) where K < rank(X) is the desired number of L1-PCs. Before the results of [1], [2], several
low-complexity suboptimal algorithms for L1-PCA existed in the literature [19]–[22]. However, in lack of the
optimal solution, no absolute performance evaluation of those algorithms with respect to the L1-PCA metric could
be conducted until [1], [2] made the optimal value available. Today, optimal-solution-informed experimental studies
indicate that the existing low-cost approximate algorithms for L1-PCA yield non-negligible performance degradation,
in particular when more than one principal component is calculated.
In this present work, we introduce L1-BF, a bit-flipping based algorithm for the calculation of the K L1-PCs
of any rank-d data matrix X ∈ RD×N with complexity O(N Dmin{N, D} + N 2 (K 4 + dK 2 ) + N dK 3 ). The
proposed algorithm is accompanied by: (i) Formal proof of convergence and theoretical analysis of converging
points; (ii) detailed asymptotic complexity derivation; (iii) theoretically proven performance guarantees; (iv) L1PCA performance comparisons with state-of-the-art counterparts of comparable computational cost [19]–[22]; and
(v) outlier-resistance experiments on dimensionality reduction, foreground motion detection in surveillance videos,
and disease diagnosis from genomic data. Our studies show that L1-BF outperforms all suboptimal counterparts of
comparable cost with respect to the L1-PCA metric and retains high outlier-resistance similar to that of optimal
L1-PCA. Thus, the proposed algorithm may bridge the gap between computationally efficient and outlier-resistant
principal component analysis.
The rest of the paper is organized as follows. Section II presents the problem statement and reviews briefly
pertinent technical background. Section III is devoted to the development and analysis of the proposed algorithm.
Extensive experimental studies are presented in Section IV. A few concluding remarks are drawn in Section V.
II. P ROBLEM S TATEMENT
AND
BACKGROUND
Given a data matrix X = [x1 , x2 , . . . , xN ] ∈ RD×N of rank d ≤ min{D, N }, we are interested in calculating a
low-rank data subspace of dimensionality K < d in the form of an orthonormal basis QL1 ∈ RD×K that solves
QL1 =
argmax
Q=[q1 ,...,qK ]∈RD×K
Q⊤ Q=IK
K
X
k=1
X⊤ qk
1
.
(1)
In (1), k·k1 denotes the element-wise L1-norm of the vector/matrix argument that returns the sum of the absolute
values of the individual entries. [1], [2] and [22] presented different proofs that L1-PCA in the form of (1) is
formally NP-hard problem in jointly asymptotic N , d. Suboptimal algorithms (with non-negligible performance
degradation) for approximating the solution in (1) were proposed in [19]–[22]. [1], [2] showed for the first time
3
that for fixed data dimension D , (1) is not NP-hard and presented two optimal algorithms for solving it.
Below, we review briefly the optimal solution to (1) and the suboptimal algorithms that exist in the literature.
A. Optimal Solution
△
For any matrix A ∈ Rm×n , m > n, that admits SVD A = UΣn×n VT , define U (A) = UV⊤ . Let k · k∗ denote
nuclear norm. In [1], [2] it was shown that, if
Bopt = argmax kXBk∗ ,
(2)
QL1 = U (XBopt )
(3)
B∈{±1}N ×K
then
is a solution to (1). In addition, Q⊤
L1 X
1
= kXBopt k∗ and Bopt = sgn(X⊤ QL1 ) [1], [2].
For K = 1, (2) takes the binary quadratic form1
bopt = argmax kXbk22
(4)
b∈{±1}N
and the L1-principal component of X is given by
qL1 = U (Xbopt ) = Xbopt kXbopt k−1
2 .
In addition, X⊤ qL1
1
(5)
= kXbopt k∗ = kXbopt k2 and bopt = sgn(X⊤ qL1 ).
In view of (2), (3), the first optimal algorithm in [1], [2] performs an exhaustive search over the size-2N K
candidate set {±1}N ×K to obtain a solution Bopt to (2); then QL1 is returned by SVD on XBopt .2 The second
optimal algorithm in [1], [2], of polynomial complexity, constructs and searches inside a subset of {±1}N ×K
wherein a solution to (2) is proven to exist. Importantly, for d constant with respect to N , the cost to construct and
search exhaustively within this set is O(N dK−K+1 ).
From (2)-(5), it is seen that, in contrast to L2-PCA, in L1-PCA the scalability principle does not hold. That
is, qL1 =
argmax
q∈RD×1 ;
kqk2 =1
kX⊤ qk1 is not (in general) a column of QL1 =
argmax
Q∈RD×K ;
Q⊤ Q=IK
kX⊤ Qk1 for some
For every a ∈ Rd , the nuclear and euclidean norm trivially coincide, i.e., kak∗ = kak2 .
In practice, the exhaustive-search optimal algorithm takes advantage of the nuclear-norm invariability to negations and permutations of
N −1
the columns of the matrix argument and searches exhaustively in a size- 2 K+K−1 subset of {±1}N×K wherein a solution to (2) is
guaranteed to exist.
1
2
4
K > 1. Therefore, the size-(K > 1) L1-PCA problem cannot be translated into a series of size-(K = 1) L1-PC
problems simply by projecting the data-matrix onto the null-space of the previous solutions.
B. State-of-the-art Approximate Algorithms
1) Fixed-point Iterations with Successive Nullspace Projections [19], [20]: Kwak et al. [19], [20] made an
important early contribution to the field by proposing a fixed point (FP) iteration to approximate the K = 1 L1-PC
solution. Following our formulation and notation in Section II, the algorithm has the iterative form
b(t) = sgn X⊤ Xb(t−1) , t = 2, 3, . . . ,
(6)
where b(1) is an arbitrary initialization point in {±1}N . For any initialization, (6) is guaranteed to converge to a fixed
point of sgn X⊤ Xb in Φ(X) = {b ∈ {±1}N : b = sgn(X⊤ Xb)}. Then, the L1-PC can be approximated by
(t)
qfp = Xbfp kXbopt k−1
2 where bfp is the converging point of the iteratively generated sequence {b }. For K > 1,
L1-PCs are approximated sequentially in a greedy fashion. That is, the kth L1-PC qfp,k , k > 1, is calculated by the
above procedure, having replaced X by its projection onto the nullspace of the previously calculated components,
Pk−1
qfp,i q⊤
(ID − i=1
fp,i )X. The complexity of this algorithm is O(M N DK) where M is the maximum number of
iterations per component. Considering M to be bounded by a linear function of N , or practically terminating the
iterations at most at N , the complexity of this algorithm can be expressed as O(N 2 DK).
2) Iterative Alternating Optimization (Non-greedy) [21]: Nie et al. [21] offered a significant advancement for
the case K > 1. Following, again, our formulation and notation, they proposed a converging iterative algorithm
that, initialized at an arbitrary orthonormal matrix Q(1) ∈ RD×K , calculates
B(t) = sgn X⊤ Q(t−1)
and
Q(t) = U XB(t) , t = 2, 3, . . . .
(7)
The solution to (1) is approximated by the convergence point of the generated sequence {Q(t) }, say Qao . Notice
that, for K = 1, the algorithm coincides with the one in [19]. The computational complexity of (7) is O(T (N D +
K 2 )) where T is the maximum number of iterations per component. Considering T to be bounded by a linear
function of N K , or practically terminating the iterations at most at N K , the complexity of the algorithm becomes
O(N 2 DK + N K 3 ).
5
3) SDP with Successive Nullspace Projections [22]: McCoy and Tropp [22] suggested a novel semi-definite
programming (SDP) view of the problem. The binary-optimization problem in (4) can be rewritten as
maximize
N
Z∈S+ , [Z]n,n =1 ∀n
rank(Z)=1
Tr ZX⊤ X
(8)
N the set of positive semi-definite matrices in RN ×N . Specifically, if Z
where S+
opt is the solution to (8), then any
column of Zopt is a solution to (4). Then, the algorithm relaxes the non-convex rank constraint in (8) and finds
instead the solution Zsdp to the convex semi-definite program
maximize
N
Z∈S+ , [Z]n,n =1 ∀n
Tr ZX⊤ X .
(9)
In this present paper’s notation, to obtain an approximation to bopt , say bsdp , the algorithm factorizes Zsdp =
WW⊤ , W ∈ RN ×N , and calculates L instances of b = sgn W⊤ a for vectors a drawn from N (0N , IN ) (L-
instance Gaussian randomization). bsdp is chosen to be the instance that maximizes kXbk2 and the solution to (1)
is approximated by qsdp = Xbsdp kXbsdp k−1
2 . For K > 1, similar to [19], [22] follows the method of sequential
Pk−1
qsdp,i q⊤
nullspace projections calculating the kth L1-PC qsdp,k as the L1-PC of (ID − i=1
sdp,i )X. The complexity
to solve within ǫ accuracy the SDP in (9) is O(N 3.5 log(1/ǫ)) [39]. Thus, the overall computational cost of the
algorithm is O(KN 3.5 log(1/ǫ) + KL(N 2 + DN )).
III. P ROPOSED A LGORITHM
We begin our new algorithmic developments with the following Proposition.
SVD
⊤
Proposition 1. Assume that data matrix X admits compact singular-value decomposition X = UD×d Σd×d VN
×d ∈
RD×N where d = rank(X) ≤ min{D, N } and define
△
Y = [y1 , y2 , . . . , yN ] = ΣV⊤ ∈ Rd×N .
(10)
Then, for any B ∈ {±1}N ×K , kXBk∗ = kYBk∗ .
Proof: Notice that
X⊤ X = VΣU⊤ UΣV⊤ .
√
(11)
√
B⊤ X⊤ XB = Tr
B⊤ Y ⊤ YB = kYBk∗ where for any positive semi-definite
√ √
symmetric matrix A ∈ Rm×m , A A = A.
Thus, kXBk∗ = Tr
6
By Proposition 1, the proposed algorithm will attempt to find a solution to (2) by solving instead the reduced-size
problem
Bopt = argmax kYBk∗ ,
(12)
bopt = argmax kYbk22 .
(13)
B∈{±1}N ×K
which for K = 1 takes the form
b∈{±1}N ×1
This problem-size reduction, with no loss of optimality so far, contributes to the low cost of the proposed algorithm.
In the sequel, we present separately the cases K = 1 and K > 1.
A. Calculation of the L1-Principal Component (K = 1)
First, we attempt to find efficiently a quality approximation to bopt in (13), say bbf , by a bit-flipping search
procedure. Then, per (5),
qbf = Xbbf kXbbf k−1
2
(14)
will be the suggested L1-principal component of the data. Proposition 2, presented herein for the first time, bounds
the L1 error metric of interest by L2-norm differences.
Proposition 2. For any b ∈ {±1}N and corresponding q = Xb kXbk−1
2 ,
X⊤ qL1
1
− X⊤ q
1
≤ kYbopt k2 − kYbk2
(15)
with equality if b = sgn(Y ⊤ Yb)=sgn(X⊤ Xb).
Proof: By (2)-(5) and Proposition 1,
X⊤ qL1
1
− X⊤ q
1
= kXbopt k2 − X⊤ Xb
= kYbopt k2 − Y ⊤ Yb
1
1
kXbk−1
2
kYbk−1
2
= kYbopt k2 − (maxz∈{±1}N z⊤ Y ⊤ Yb) kYbk−1
2
≤ kYbopt k2 − b⊤ Y ⊤ Yb kYbk−1
2
= kYbopt k2 − kYbk2 .
(16)
Equality holds when b = sgn(Y ⊤ Yb) = argmaxz∈{±1}N z⊤ Y ⊤ Yb.
7
In particular, by Proposition 2, the performance degradation in the L1-PCA metric kX⊤ qk1 when q = Xb kXbk−1
2
is used instead of the optimal qL1 is upper bounded by the performance degradation in the metric of (13) when b
is used instead of the optimal bopt .
In view of Proposition 2, the proposed algorithm operates as follows. The algorithm initializes at a binary vector
b(1) and employs bit-flipping (BF) iterations to generate a sequence of binary vectors {b(t) }. At each iteration
step, the algorithm browses all bits that have not been flipped before kept in an index-set3 L. Then, the algorithm
negates (flips) the single bit in L that, when flipped, offers the highest increase in the metric of (13). If no bit
exists in L the negation of which increases the quadratic metric (13), then L is reset to {1, 2, . . . , N } and all N
bits become eligible for flipping consideration again. The iterations terminate when metric (13) cannot be further
increased by any single-bit flipping.
Mathematically, at tth iteration step, t ≥ 1, the optimization objective (13) is
Yb(t)
2
2
X
= kYk2F +
(t) ⊤
b(t)
n bm yn ym
(17)
n∈{1,2,...,N }
m∈{1,2,...,N }\n
where kYkF is the Frobenius (Euclidean) norm of Y . Factorizing (17), we observe that the contribution of the nth
(t)
bit of b(t) , bn , to (17) is
α b(t) , n = 2b(t)
n
X
m∈{1,2,...,N }\n
(t) ⊤
⊤
(t)
2
bm
yn ym = 2 b(t)
y
Yb
−
ky
k
n
n n
2 .
That is, for any n ∈ {1, 2, . . . , N }, (17) can be written as
Yb(t)
2
2
= α b(t) , n + β b(t) , n
(18)
(t)
(t)
where β b(t) , n is a constant with respect to bn . Then, if we flip the nth bit by setting b(t+1) = b(t) − 2bn en,N
(where en,N is the nth column of the size-N identity matrix IN ), the change to the quadratic metric in (13) is
Y(b(t) − 2bn(t) en,N )
2
2
− Yb(t)
2
2
= −2α b(t) , n .
(19)
(t)
Thus, if the contribution of the nth bit to the quadratic of (17), α b(t) , n , is negative, flipping bn will increase
the quadratic by 2|α b(t) , n |, whereas if α b(t) , n is positive, flipping bn will decrease the quadratic by
2|α b(t) , n |.
In view of this analysis, L1-BF is initialized at an arbitrary binary antipodal vector b(1) and L = {1, 2, . . . , N }.
3
The index-set L is used to restrain the “greediness” of unconstrained single-bit flipping iterations.
8
The initial bit contributions α b(1) , 1 , . . . , α b(1) , N are calculated by (18). Then, at iteration step t ≥ 1, we
find
n = argmin α b(t) , m .
(20)
m∈L
(t)
(t)
If α b(t) , n < 0, bn is flipped by setting b(t+1) = b(t) − 2bn en,N , α b(t+1) , 1 , . . ., α b(t+1) , N are recalcu
lated, and L is updated to L \ n. If, otherwise, α b(t) , n ≥ 0, a new solution n to (20) is obtained after resetting
(t)
L to {1, 2, . . . , N }. If, for the new solution n, α b(t) , n < 0, then bn is flipped, α b(t+1) , 1 , . . . , α b(t+1) , N
are recalculated, and L is updated to L \ n. Otherwise, the iterations terminate and the algorithm returns bbf = b(t) ,
qbf = Ybbf kYbbf k−1
2 .
(t)
Notice that the contribution factors at the end of tth iteration step when bn has been flipped can be directly
updated by
α b(t+1) , n = −α b(t) , n > 0
(21)
(t) ⊤
α b(t+1) , m = α b(t) , m − 4b(t)
m bn ym yn
(22)
and
for every m 6= n. Given the data correlation matrix Y ⊤ Y , updating the contributions by (21) and (22) costs only
O(N ). For ease in reference, complete code for L1-BF, K = 1, is provided in Fig. 1.
Convergence Characterization
The termination condition of the BF iterations (convergence) is summarized to
α b(t) , n ≥ 0 ∀n ∈ {1, 2, . . . , N }.
(23)
That is, when (23) is met, the algorithm returns bbf = b(t) and terminates. For any initialization point, (23) will be
met and the BF iterations will converge in a finite number of steps, since (i) the binary quadratic form maximization
in (4) has a finite upper bound and (ii) at every step of the presented BF iterations the quadratic value increases.
Thus, the proposed iterations converge globally4 both in argument and in value.
Characterization of Point of Converngence and Relationship to Fixed Points of [19], [21]
The last point of the generated sequence, bbf , satisfies the termination condition α (bbf , n) ≥ 0 ∀n, which in
view of (18) can be equivalently rewritten as bbf,n yn⊤ Ybbf ≥ ||yn ||22 ∀n. Therefore, for every initialization point,
4
We say that a sequence converges globally if it convergences for any initialization.
9
the BF iterations converge in the set
Ω(Y) = {b ∈ {±1}N ; bn yn⊤ Yb ≥ ||yn ||22 ∀n}.
(24)
By (11), Y ⊤ Y = X⊤ X and, thus, Ω(Y) = Ω(X). Importantly, the following proposition holds true.
Proposition 3. Every optimal solution to the binary-quadratic maximization in (13) belongs to Ω(Y).
Proof: Let b be a solution to (13) outside Ω(Y). Then, there exists at least one entry in b, say the nth entry,
for which the contribution to the objective value is negative; i.e., α b(t) , n < 0. Consider now the binary vector
b′ ∈ {±1}N derived by negation of the nth entry of b. By (19), kYb′ k22 = kYbk22 − 2|αn | > kYbk22 . Hence, b
is not a maximizer in (4).
By Proposition 3, being an element of Ω(Y) is a necessary condition for a binary vector to be a solution
to (13). Since any point in Ω(Y) is reachable by the BF iterations upon appropriate initialization, L1-BF may
indeed calculate bopt and, through (14), return the optimal L1-PC solution to (1), qL1 . The following Proposition
establishes an important relationship between the set of convergence points of the proposed BF iterations, Ω(Y),
and the set of convergence points of the fixed-point iterations in [19], [21], say Φ(Y).
Proposition 4. Every element of Ω(Y) lies also in the fixed-point set Φ(Y); that is, Ω(Y) ⊆ Φ(Y).
Proof: Consider an arbitrary b ∈ Ω(Y). Then, bn yn⊤ Yb ≥ ||yn ||22 , for all n ∈ {1, 2, . . . , N }. Assume that
bn = +1 for some n. Then, yn⊤ Yb ≥ kyn k22 ≥ 0 and sgn(yn⊤ Yb) = +1 = bn . Otherwise, bn = −1. Then,
yn⊤ Yb ≤ − kyn k22 ≤ 0 and sgn(yn⊤ Yb) = −1 = bn . Hence, for every vector b ∈ Ω(Y), bn = sgn(yn⊤ Yb) for all
n and, thus, b ∈ Φ(Y).
If we denote by B(Y) the set of optimal solutions to (4), Propositions 3 and 4 can be summarized as
B(Y) ⊆ Ω(Y) ⊆ Φ(Y) ⊆ {±1}N
(25)
or, in terms of set sizes, 2 ≤ |B(Y)| ≤ |Ω(Y)| ≤ |Φ(Y)| ≤ 2N .
For illustration, in Fig. 2, we demonstrate experimentally the relationship between the cardinalities of Φ(Y),
Ω(Y), and B(Y). We generate 1000 independent instances of a full-rank data matrix Y ∈ R(d=2)×N for N =
2, 3, . . . , 7, with each element of Y drawn independently from the Gaussian N (0, 1) distribution. Then, we plot the
average observed cardinality of Φ(Y), Ω(Y), and B(Y) versus the number of data points N . We see that as N
increases Ω(Y) remains a tight super-set of B(Y), while the cardinality of Φ(Y) (number of possible converging
points of the FP iterations) increases at a far higher rate. Fig. 2 offers insight why the proposed BF iterations are
10
expected to return the optimal solution to (4) much more often than the FP iterations of [19]–[21].
Performance Bounds
The performance attained by qbf is formally lower bounded by Proposition 5 whose proof is given in the
Appendix.
Proposition 5. The performance of qbf in the metric of (1), calculated upon any initialization, is lower bounded
by X⊤ qbf
1
≥ kXkF .
Moreover, for the optimal L1-PC qL1
X⊤ qL1
1
= kXbopt k2 =
√
N
max
1
z∈{± √N }N
kXzk2 ≤
√
N
max
z∈RN ; kzk2 =1
kXzk2 =
√
N σmax
(26)
where σmax is the maximum singular value of the data matrix X. Then, the loss with respect to the L1-PCA metric
in (1) experienced by qbf is bounded by
X⊤ qL1
1
− X⊤ qbf
1
≤
√
N σmax − kXkF .
(27)
In Fig. 3, we execute the proposed BF iterations on an arbitrary data matrix X ∈ R(d=4)×(N =32) (elements drawn
independently from N (0, 1)) and plot the binary quadratic metric kXb(t) k2 and L1-PCA metric kX⊤ q(t) k1 per
iteration t. For reference, we also plot the optimal line kX⊤ qL1 k1 = kXbopt k2 and lower bound kXkF . Fig. 3
offers a vivid numerical illustration of Proposition 5 and eq. (27).
Initialization
Next, we consider initializations that may both expedite convergence and offer superior L1-PC approximations. In
the trivial d = 1 case, Y = σv⊤ for some σ ∈ R+ and v ∈ RN ×1 , kvk2 = 1, and bopt = arg maxb∈{±1}N kYbk22 =
arg maxb∈{±1}N |v⊤ b|22 = ±sgn(v). Motivated by this special case, we initialize the BF iterations to the sign of the
right singular vector of Y that corresponds to the highest singular value. We call this initialization choice sv-sign
initialization. Evidently, by the definition of Y in (10),
b(1) = sgn([Y ⊤ ]:,1 ).
(28)
Experimental studies that compare the efficiency of the proposed sv-sign initialization with that of equiprobable
random initializations (b(1) takes any value in {±1}N with probability 2−N ) show that sv-sign initialization not
only leads to superior L1-PC approximations with respect to (1), but also reduces significantly the actual execution
time of the proposed algorithm by reducing the number of bit-flips needed for convergence. As an illustration, we
11
generate 1000 realizations of the data matrix X ∈ R3×20 with entries drawn independently from N (0, 1). On each
realization we run bit-flipping iterations with both sv-sign and equiprobable binary initialization. We observe that
81.6% of the time the L1-PC obtained with sv-sign initialization attains value in the metric of (1) greater than or
equal to L1-PC obtained by random initialization. Also, in Fig. 4 we plot the empirical CDF of the number of
bit-flips needed until convergence for the two initializations. We observe that 50% of the time sv-sign initialization
is already in the convergence set Ω(X) and no bit-flips take place. Moreover, with sv-sign initialization no more
than 6 bit-flips are needed for convergence with empirical probability 1. On the contrary, random initialization may
need with non-zero probability up to 16 bit-flips for convergence.
Complexity Analysis
Prior to bit-flipping iterations, L1-BF calculates Y in (10) and chooses the initial binary vector b(1) with complexity O(N Dmin{N, D}) (SVD). Then, the algorithm calculates and stores the correlation matrix Y ⊤ Y ∈ RN ×N
with complexity O(N 2 d) and, given b(1) and Y ⊤ Y , evaluates by (18) the initial contribution factors α b(1) , n
for all n ∈ {1, 2, . . . , N } with cost O(N 2 ). Therefore, initialization has total cost O(N Dmin{N, D} + N 2 (d + 1)).
At iteration t ≥ 1, finding n by (20) costs O(N ). In case α b(t) , n > 0, repetition of the maximization over the
entire index set {1, 2, . . . , N } costs an additional O(N ). Therefore, each bit flip costs O(N ). Denoting by M the
number of bit flips until the termination criterion is met, we find that the computational cost for calculating bbf is
O(N Dmin{N, D} + N 2 (d + 1) + M N ).
Finally, taking into account the computation of qbf from bbf , the proposed L1-BF algorithm costs O(N D
min{N, D} +N 2 (d + 1) + (M + D)N ). According to our experiments, M is with empirical probability 1 upper
bounded by N ; that is, no more than N bits need to be flipped to reach bbf .5 Therefore, the total complexity of
the proposed algorithm becomes O(N Dmin{N, D} + N 2 (d + 2) + N D). Keeping only the dominant complexity
terms, L1-BF has complexity O(N Dmin{N, D} + N 2 d), i.e. quadratic in N and either linear, or quadratic in D
(depending on the cost of the initial SVD). Thus, the proposed L1-BF algorithm has cost comparable to that of
standard PCA (SVD). A summary of the calculated complexity is provided in Table I.
Multiple Initializations
(1)
For further L1-PCA metric improvement, we may also run BF on L distinct initialization points, b1
(1)
sgn([Y ⊤ ]:,1 ) and bl
= sgn(Y ⊤ al ), with al ∼ N (0d , Id ), l = 2, . . . , L, to obtain L corresponding con(L)
vergence points bbf,1 , bbf,2 , . . . , bbf,L . Then, if qbf,l = Xbbf,l kXbbf,l k−1
2 , l = 1, . . . , N , we return qbf
argmaxq∈{qbf,1 ,...,qbf,L } X⊤ q
5
=
1
=
. Certainly, multiple initializations will increase the complexity of the algorithm
In practice, a brute-force termination M ≤ N can be imposed to the algorithm.
12
by a constant factor L.
B. Calculation of K > 1 L1-Principal Components
In contrast to L2-PCA, L1-PCA in (1) is not a scalable problem [2]. Therefore, successive-nullspace-projection
approaches [19], [22] fail to return optimal L1-PC bases. Optimal L1-PCA demands joint computation of all K
principal components of Y , a procedure that increases exponentially in K the computational complexity of optimal
algorithms. In this section, we generalize the BF algorithm presented in Section III.A to calculate K > 1 L1-PCs
of any given data matrix X ∈ RD×N of rank d ≥ K .
Given the reduced-size, full-row-rank data matrix Y ∈ Rd×N (see Proposition 1), the proposed algorithm first
attempts to approximate the solution to (12) Bopt . To do so, it begins from an initial matrix B(1) and employs
bit-flipping iterations to reach an approximation to Bopt , say Bbf . Finally, per (29) the algorithm returns
Qbf = U (XBbf ) .
(29)
Evidently, if Bbf is an exact solution to (2), then Qbf is an exact solution to the L1-PCA problem in (1).
Calculation of Bbf via Bit-flipping Iterations
The algorithm initializes at a binary matrix B(1) and employs BF iterations to generate a sequence of binary
matrices B(t) , t = 2, 3, . . . , such that each matrix B(t) attains the highest possible value in the maximization metric of
(12) given that it differs from B(t−1) in exactly one entry. Similar to the K = 1 case, the indices of bits that have not
been flipped throughout the iterations are kept in a memory index-set L initialized at {1, 2, . . . , N K} (the (n, k)th
bit of the binary matrix argument in (12) has corresponding single-integer index (k − 1)N + n ∈ {1, 2, . . . , N K}).
Mathematically, at the tth iteration step, t ≥ 1, we search for the single bit in the current binary argument B(t)
whose index is not in L and its flipping increases the nuclear-norm metric kYB(t) k∗ the most. We notice that if
(t)
we flip the (n, k)th bit of B(t) setting B(t+1) = B(t) − 2Bn,k en,N e⊤
k,K , then
(t)
YB(t+1) = YB(t) − 2Bn,k yn e⊤
k,K .
(30)
Therefore, at the tth iteration step, the algorithm obtains the index pair
(n, k) =
(t)
argmax
(m,l)∈{1,2,...,N }×{1,2,...,K}
(l−1)N +m∈L
If
(t)
YB(t) − 2Bn,k yn e⊤
k,K
∗
>
YB(t)
YB(t) − 2Bm,l ym e⊤
l,K
(t)
∗
∗
.
(31)
(t)
, the algorithm flips Bn,k setting B(t+1) = B(t) − 2Bn,k en,N e⊤
k,K
13
(t)
YB(t) − 2Bn,k yn e⊤
k,K
and updates L to L \ {(k − 1)N + n}. If, otherwise,
∗
YB(t)
≤
∗
, the algorithm
obtains a new solution (n, k) to (31) after resetting L to {1, 2, . . . , N K}. If now, for this new pair (n, k),
(t)
YB(t) − 2Bn,k yn e⊤
k,K
∗
YB(t)
>
(t)
∗
, the algorithm sets B(t+1) = B(t) − 2Bn,k en,N e⊤
k,K and updates L =
L \ {(k − 1)N + n}. Otherwise, the iterations terminate and the algorithm returns Bbf = B(t) and Qbf = ÛV̂⊤ ,
SVD
⊤
where XBbf = ÛD×K Σ̂K×K V̂K×K
.
(t)
We note that to solve (31) exhaustively, one has to calculate YB(t) − 2Bm,l ym e⊤
l,K
∗
for all (m, l) ∈ {1, 2, . . . , N }×
{1, 2, . . . , K} such that (l − 1)N + m ∈ L. In worst case, L = {1, 2, . . . , N K} and this demands N K in-
dependent singular-value/nuclear-norm calculations. A simplistic, but computationally inefficient, approach would
(t)
2
be perform an SVD on YB(t) − 2Bm,l ym e⊤
l,K from scratch with cost O(dK ). This would yield a worst case
total cost of O(N dK 3 ) to find (n, k) in (31). Below, we present an alternative method to solve (31) with cost
O(dK 2 + N (K 3 + dK)).
Reduced-cost Nuclear-norm Evaluations for Optimal Bit-flipping
At the beginning of the tth iteration step, we perform the singular-value decomposition
SVD
(t)
(t)⊤
(t)
YB(t) = Ud×K SK×K VK×K
(32)
(t)
with cost O(dK 2 ). Then, for any (m, l) ∈ {1, 2, . . . , N }×{1, 2, . . . , K}, the singular values of YB(t) −2Bm,l ym e⊤
l,K
(t)
(t)
⊤
(t) − 2B
⊤
are the square roots of the eigenvalues of (YB(t) − 2Bm,l ym e⊤
l,K ) (YB
m,l ym el,K ) (found with constant
(t)
⊤
⊤
(t)
cost), which due to rotation invariance are equal to the eigenvalues of A = V(t) (YB(t) −2Bm,l ym e⊤
l,K ) (YB −
(t)
(t)
⊤
⊤
(t) U(t) y ] ∈ RK×2 and the eigenvalue decomposition
(t)
(t) ⊤
2Bm,l ym e⊤
m
l,K )V . Defining W = [[V ]l,: , −2Bm,l S
EVD
⊤
(EVD) QDQ⊤ = kym k22 e1,2 e⊤
1,2 + [e2,2 , e1,2 ] where D = diag([d1 , d2 ] ), it is easy to show that
⊤
A = S(t) S(t) + d1 Wq1 (Wq1 )⊤ + d2 Wq2 (Wq2 )⊤ .
(33)
Consider now the eigenvalue decomposition
EVD
⊤
ZK×K PZ⊤ = S(t) S(t) + d1 Wq1 (Wq1 )⊤ .
(34)
(t)
Then, again by eigenvalue permutation invariance, the singular values of YB(t) − 2Bm,l ym e⊤
l,K are equal to the
square roots of the eigenvalues of
Z⊤ AZ = P + d2 Z⊤ Wq2 (Z⊤ Wq2 )⊤ .
⊤
Notice that matrix S(t) U(t)
⊤
(35)
needed for the calculation of W is constant for all candidate-bit evaluations of
14
the tth iteration and, given the SVD of YB(t) , can be found with cost O(K 2 ). Thus, this computational cost is
absorbed in the SVD of YB(t) . Then, for bit-flip examination of all K bits of the mth row of B(t) it suffices to
⊤
⊤
⊤
calculate the quantities Q, d, and S(t) U(t) ym , appearing in S(t) S(t) + d1 Wq1 (Wq1 )⊤ and d2 Wq2 , just once
with cost O(dK). Employing the algorithm proposed in [40] for the EVD of a diagonal positive semidefinite matrix
⊤
perturbated by a rank-1 symmetric matrix, EVD of S(t) S(t) + d1 Wq1 (Wq1 )⊤ (i.e., calculation of Z and P) costs
O(K 2 ). Then, defining d2 Z⊤ Wq2 (Z⊤ Wq2 )⊤ and finding the eigenvalues of P + d2 Z⊤ Wq2 (Z⊤ Wq2 )⊤ (again
per [40]) costs an additional O(K 2 ). Thus, including the initial SVD, the total cost for bit-flipping examination of
all bits in B(t) (worst case scenario) is O(dK 2 ) + N O(dK) + N KO(K 2 ) ≡ O(dK 2 + N (K 3 + dK)).
Termination Guarantee
Similarly to the K = 1 case, the termination condition
(t)
YB(t) − 2Bm,l ym e⊤
l,K
∗
≤ YB(t)
∗
∀(m, l) ∈ {1, 2, . . . , N } × {1, 2, . . . , K}
(36)
is guaranteed to be met after a finite number of iterations because (i) binary nuclear-norm maximization in (2) has
finite upper bound and (ii) at every step of the BF iterations the nuclear-norm value increases.
Initialization
Generalizing the initialization idea for the K = 1 case, we choose B(1) = sgn([Y ⊤ ]:,1 )1⊤
K where 1K is the allone vector of length K . Again, our motivation here lies in the fact that for d = 1 and any K , Bopt = sgn(vopt )1⊤
K.
For this particular initialization, YB(1)
∗
= K Ysgn([Y ⊤ ]:,1 )
2
. Pseudocode of the proposed L1-BF algorithm
for K ≥ 1 is offered in Fig. 5. Of course, for K = 1 the L1-PCs generated by the algorithms of Fig. 1 and Fig. 5
coincide.
Complexity Analysis
Before the BF iterations, the algorithm expends O(N Dmin{N, D}) to calculate Y ∈ Rd×N through SVD of X ∈
RD×N . By the same SVD, we calculate also the initial binary matrix. Then, for the proposed initialization YB(1) and
YB(1)
∗
are calculated in the beginning of the first iteration with complexity O(N dK). To find a solution to (31),
we incur worst case cost O(dK 2 +N (K 3 +dK)). Setting the maximum number of BF iterations run by the algorithm
to N K , the total complexity to calculate Bbf is in worst case O(N Dmin{N, D} + N dK 3 + N 2 (K 4 + dK 2 )).
Then, to calculate Qbf from Bbf costs an additional O(N Dmin{N, D} + N DK). Therefore, the total complexity
of the proposed algorithm is O(N Dmin{N, D} + N 2 (K 4 + dK 2 ) + N dK 3 ); that is, quadratic complexity in N
and at most quadratic complexity in D . Notice that for K = 1 the asymptotic complexity of the algorithm becomes
O(N Dmin(N, D) + N 2 d), i.e. equal to that of the algorithmic operations of Fig. 1. A summary of the calculated
15
complexity is provided in Table II.
Multiple Initializations
Similar to the K = 1 case, we can further increase performance of the proposed L1-PCA scheme by running
(1)
(1)
(1)
(1)
BF iterations on L distinct initialization matrices B1 , B2 , . . . , BL . We choose B1 = sgn([Y ⊤ ]:,1 )1⊤
K (sv-sign
(1)
initialization) and for l ∈ {2, 3, . . . , L} we initialize randomly to Bl
= sgn(al )1⊤
K , al ∼ N (0N , IN ), to obtain
the L corresponding convergence points Bbf,1 , Bbf,2 , . . . , Bbf,L and, through (29), the associated L1-PC bases
(L)
Qbf,1 , Qbf,2 , . . . , Qbf,L . Then, we return Qbf = argmaxQ∈{Qbf ,1 ,Qbf,2 ,...,Qbf,L } X⊤ Q
1
.
IV. E XPERIMENTAL S TUDIES
A. Comparison of L1-BF with Current State-of-the-art Algorithms
We first compare the performance of L1-BF with that of algorithms proposed in [19], [21], [22].6 In our
experiment, we generate 1000 arbitrary matrices of size (D = d = 4) × (N = 16) with entries drawn independently
from N (0, 1) and calculate the primary (K = 1) L1-PC by L1-BF iterations (L = 1 initialization), the fixed-point
algorithm of [19], [21] (L = 1 initialization), and the SDP approach in [22] (L = 1 Gaussian randomization). We
measure the performance degradation experienced by a unit-length vector q on the metric of (1) by
∆(q; X) =
kX⊤ qL1 k1 − kX⊤ qk1
kX⊤ qL1 k1
(37)
and in Fig. 6(a) we plot the empirical cumulative distribution function (CDF) of ∆(qbf ; X), ∆(qfp ; X), and
∆(qsdp ; X). We observe that 86% of the time, L1-BF returns the exact, optimal L1-PC of X. In addition, the
performance degradation attained by the proposed algorithm is, with empirical probability 1, less than 0.09. On the
other hand, FP iterations [19], [21] and SDP [22] attain optimal performance, with empirical probabilities 0.3 and
0.5, respectively. The performance degradation for FP [19], [21] and SDP [22] can be as much as 0.25 and 0.55,
respectively.
We repeat the same experiment after increasing the number of initializations for the proposed algorithm and the
FP [19], [21] to L = N = 16 and the number of Gaussian randomizations in SDP [22] to L = N = 16. In Fig.
6(b), we plot the corresponding empirical CDFs of ∆(qbf ; X), ∆(qfp ; X), and ∆(qsdp ; X). L1-BF outperforms the
two counterparts, returning the exact L1-PC with empirical probability 1.
Next, we examine the performance of the four algorithms (FP iterations with successive nullspace projections [19],
alternating optimization [21], and SDP with successive nullspace projections [22], and proposed BF iterations) when
Regarding the SDP approach of [22], the total computational cost employing L Gaussian randomization instances is O(N 3.5 log(1/ǫ) +
L(N 2 + DN )), which is similar to running the proposed BF iterations over N distinct initialization points.
6
16
K = 2 L1-PCs are sought. We generate 1000 arbitrary data matrices of size (D = 3) × (N = 8) (all entries drawn
independently from N (0, 1)) and plot in Fig. 7(a) the empirical CDF of the corresponding performance degradation
ratios, ∆(Qfp ; X), ∆(Qao ; X), ∆(Qsdp ; X), and ∆(Qbf ; X). All iterative methods are run on a single initialization
point (one Gaussian-randomization instance for SDP). We observe that 83% of the time, L1-BF returns the exact,
optimal L1-PCs of X. The performance degradation attained by the proposed PCs is, with empirical probability 1,
less than 0.09. On the other hand, the PCs calculated by [19], [21], [22] attain optimal performance, with empirical
probabilities of above 0.31, 0.05, and 0.1, respectively, while the performance degradation for the algorithms in
[19], [21] and [22] can be as much as 30%. Evidently, the nullspace projections of [19], [22] that “violate” the
non-scalability principle of the L1-PCA problem, have significant impact on their performance. Next, we repeat
the above experiment with L = N K = 16 initializations (L = 16 Gaussian-randomization instances for SDP). In
Fig. 7(b), we plot the new performance degradation ratios for the four algorithms. This time, the proposed method
returns the exact/optimal L1-PCs of X with probability 1. High, but inferior, performance is also attained by the
iterative algorithm of [21]. On the other hand, the algorithms in [19], [22] that follow the greedy approach of
solving a sequence of nullspace-projected K = 1 problem instances experience performance degradation values
similar to those of Fig. 7(a).
B. Line-fitting Experiment
For visual evaluation of the outlier resistance of the proposed L1-PCs, we consider N = 100 2-dimensional
4 10
points drawn from the nominal zero-mean Gaussian distribution N 02 , R =
organized in matrix
10 29
form Xnom = [x1 , x2 , . . . , x100 ]. We use the nominal data points to extract an estimate of the true maximumvariance/minimum-mean-squared-error subspace (line) of our data distribution by means of L2-PCA (standard SVD
method). We do the same by means of L1-PCA through the proposed L1-BF method. In Fig. 8(a), we plot on the
2-dimensional plain the 100 training data points in Xnom and the lines described by qL2 (Xnom ) and qbf (Xnom ). For
reference, we also plot alongside the true maximum-variance line. All three lines visually coincide, i.e. qL2 (Xnom )
and qbf (Xnom ) are excellent estimates of the true maximum-variance line.
Next, we assume that instead of the clean matrix of nominal training data points Xnom , we are given the outliercorrupted data matrix Xcor = [Xnom , O] ∈ R2×104 , for estimating qopt . Here, in addition to the previous 100
nominal data points in Xnom , Xcor also contains the 4 outliers in O = [o1 , o2 , o3 , o4 ] ∈ R2×4 as seen in Fig. 8(b)
that do not follow the nominal distribution. Similar to our treatment of Xnom , we calculate the L2-PC and L1-PC
(by L1-BF) of Xcor , qL2 (Xcor ) and qbf (Xcor ), respectively. In Fig. 8(b) we plot the two lines, against the true
17
maximum-variance line of the nominal data. This time, the lines of qL2 (Xcor ) and qbf (Xcor ) differ significantly. In
sharp contrast to L1-PC, the L2-PC of Xcor is strongly attracted by the four outliers and drifting away from the true
maximum-variance line. The outlier-resistance and superior line-fitting performance of the proposed approximate
L1-PC in the presence of the faulty data is clearly illustrated.
C. Classification of Genomic Data – Prostate Cancer Diagnosis
In this experiment, we conduct subspace-classification of MicroRNA (miRNA) data for prostate cancer diagnosis
[41], [42]. Specifically, we operate on the expressions of D = 9 miRNAs (miR-26a, miR-195, miR-342-3p, miR126*, miR-425*, miR-34a*, miR-29a*, miR-622, miR-30d) that have been recently considered to be differentially
expressed between malignant and normal human prostate tissues [43]. For each of the above miRNAs, we obtain one sample expression from one malignant and one normal prostate tissue from N = 19 patients (patient
indexes per [43]: 1, 2, . . . , 10, 12, . . . , 20) of the of the Swedish Watchful Waiting cohort [44].7 The miRNA
expressions from malignant and normal tissues are organized in XM = [xM,1 , xM,2 , . . . , xM,N ] ∈ RD×N and
XN = [xN,1 , xN,2 , . . . , xN,N ] ∈ RD×N , respectively. The classification/diagnosis experiment is conducted as
follows.
We collect Ntrain = 10 points from XM in XM,tr = [XM ]:,IM with IM ⊂ {1, 2, . . . , N } and |IM | = Ntrain ,
and Ntrain points from XN in XN,tr = [XN ]:,IN with IN ⊂ {1, 2, . . . , N } and |IN | = Ntrain . Then, we calculate
the zero-centered malignant and normal tissue training datasets X̃M,tr = XM,tr − mM 1⊤
Ntrain where mM =
1
Ntrain XM,tr 1Ntrain
and X̃N,tr = XN,tr − mN 1⊤
Ntrain where mN =
1
Ntrain XN,tr 1Ntrain ,
respectively. Next, we
find K = 3 L2-PCs or L1-PCs by L1-BF of the zero-centered malignant and normal datasets, QM ∈ RD×K
and QN ∈ RD×K , respectively, and conduct subspace-classification [45], [46] of each data-point x that has not
c
c
c ∪ {x
c where I
been used for PC training (i.e., every x ∈ {xM,i }i∈IM
N,i }i∈IN
M = {1, 2, . . . , N } \ IM and IN =
{1, 2, . . . , N } \ IN ) as
malignant
kQ⊤
M (x
−
mM )k22
>
<
normal
2
kQ⊤
N (x − mN )k2 + λ
(38)
where λ is the classification/detection bias term. We run the above classification experiment over 2000 distinct
training-dataset configurations and calculate the receiver operating characteristic (ROC) curve of the developed
L2-PCA or L1-PCA classifier identifying as “detection” the event of classifying correctly a malignant tissue and
as “false alarm” the event of classifying erroneously a normal tissue (Fig. 9, “p = 0” curves).
7
More information about the studied data can be found in [43].
18
Next, we consider the event of training on faulty/mislabeled data. That is, we recalculate the above ROC curves
having exchanged p = 2 or 4 data points between XM,tr and XN,tr . In Fig. 9, we plot the ROC curves for the
designed L2-PCA and L1-PCA malignant tissue detectors, for K = 3 and p = 0, 2, 4. We observe that for p = 0
(no mislabeling) the the two classifiers perform almost identically, attaining frequency of detection (FD) close to
1 for frequency of false alarm (FFA) less than 0.15. However, in the event of training data mislabeling, the strong
resistance of the proposed L1-PCs against dataset contamination is apparent. The L1-PCA classifier outperforms
significantly L2-PCA, for every examined value of p. For instance, for p = 2, L1-PCA achieves FD of .95 for FFA
0.23, while L2-PCA achieves the same FD for FFA .65.
V. C ONCLUSIONS
In this work, we presented a novel, near-optimal, bit-flipping based algorithm named L1-BF, to calculate K
L1-PCs of any rank-d D × N data matrix with complexity O(N Dmin{N, D} + N 2 (K 4 + dK 2 ) + N dK 3 ). Formal
proof of convergence, theoretical analysis of converging points, and detailed asymptotic complexity derivation were
carried out. Our numerical experiments show that the proposed algorithm outperforms in the optimization metric
all other L1-PCA calculators at cost comparable to L2-PCA. L1-PCA and L2-PCA have almost indistinguishable
behavior on clean, nominal data. L1-PCA shows remarkable relative resistance to faulty data contamination. Thus,
the proposed algorithm, retaining the robustness of L1-PCA at a cost close to that of L2-PCA, may bridge the gap
between outlier resistant and computationally efficient principal-component analysis.
A PPENDIX
A. Proof of Proposition 5
Since bbf ∈ Ω(Y), (24) and (11) imply
kYbbf k22 =
≥
X
bn yn⊤ Yb
X
kyn k22 = kYk2F = kXk2F .
n∈{1,2,...,N }
n∈{1,2,...,N }
19
(39)
In addition, by Proposition 4, bbf lies in Φ(Y) as well and satisfies bbf = sgn(Y ⊤ Ybbf ). Therefore, for qbf =
Xbbf kXbbf k−1
2 ,
X⊤ qbf
1
= X⊤ Xbbf kXbbf k−1
2
1
= sgn(X⊤ Xbbf )⊤ X⊤ Xbbf kXbbf k−1
2
= sgn(Y ⊤ Ybbf )⊤ X⊤ Xbbf kXbbf k−1
2
(39)
= kXbbf k2 = kYbbf k2 ≥ kXkF .
20
(40)
R EFERENCES
[1] P. P. Markopoulos, G. N. Karystinos, and D. A. Pados, “Some options for L1-subspace signal processing,” in Proc. 10th Intern. Symp.
Wireless Commun. Syst. (ISWCS), Ilmenau, Germany, Aug. 2013, pp. 622–626.
[2] P. P. Markopoulos, G. N. Karystinos, and D. A. Pados, “Optimal algorithms for L1-subspace signal processing,” IEEE Trans. Signal
Process., vol. 62, pp. 5046–5058, Oct. 2014.
[3] S. Kundu, P. P. Markopoulos, and D. A. Pados, “Fast computation of the L1-principal component of real-valued data,” in Proc. IEEE
Int. Conf. Acoust., Speech, Signal Process. (ICASSP), Florence, Italy, May 2014, pp. 8028–8032.
[4] K. Pearson, “On lines and planes of closest fit to systems of points in space,” Philosophical Mag., vol. 2, pp. 559–572, 1901.
[5] I. T. Jolliffe, Principal Component Analysis. New York, NY: Springer-Verlag, 1986.
[6] C. M. Bishop, Pattern Recognition and Machine Learning. New York, NY: Springer, 2006.
[7] R. O. Duda, P. E. Hart, and D. G. Stork. Pattern Classification, 2nd ed. New York, NY: Wiley, 2001.
[8] V. Barnett and T. Lewis, Outliers in Statistical Data, New York, NY: John Wiley and Sons, 1994.
[9] J. W. Tukey, “The future of data analysis,” Ann. Math. Stat., pp. 1–67, 1962.
[10] E. J. Candès, X. Li, Y. Ma, and J. Wright, “Robust principal component analysis?,” J. ACM, vol. 58, art. 11, May 2011.
[11] R. R. Singleton, “A method for minimizing the sum of absolute values of deviations,” Ann. Math. Stat., vol. 11, pp. 301–310, Sep.
1940.
[12] O. J. Karst, “Linear curve fitting using least deviations,” J. American Stat. Assoc., vol. 53, pp. 118–132, Mar. 1958.
[13] I. Barrodale, “L1 approximation and the analysis of data,” J. Royal Stat. Soc., App. Stat., vol. 17, pp. 51–57, 1968.
[14] I. Barrodale and F. D. K. Roberts, “An improved algorithm for discrete l1 linear approximation,” SIAM J. Num. Anal., vol. 10, pp.
839–848, Oct. 1973.
[15] Q. Ke and T. Kanade, “Robust subspace computation using L1 norm,” Internal Technical Report Computer Science Dept., Carnegie
Mellon Univ., Pittsburgh, PA, USA, CMU-CS-03-172, 2003.
[16] Q. Ke and T. Kanade, “Robust L1 norm factorization in the presence of outliers and missing data by alternative convex programming,”
in Proc. IEEE Conf. Comp. Vision Patt. Recog. (CVPR), San Diego, CA, Jun. 2005, pp. 739–746.
[17] J. P. Brooks and J. H. Dulá, “The L1-norm best-fit hyperplane problem,” Appl. Math. Lett., vol. 26, pp. 51–55, Jan. 2013.
[18] J. P. Brooks, J. H. Dulá, and E. L. Boone, “A pure L1-norm principal component analysis, J. Comput. Stat. Data Anal., vol. 61, pp.
83–98, May 2013.
[19] N. Kwak, “Principal component analysis based on L1-norm maximization,” IEEE Trans. Patt. Anal. Mach. Intell., vol. 30, pp. 1672–
1680, Sep. 2008.
[20] N. Kwak and J. Oh, “Feature extraction for one-class classification problems: Enhancements to biased discriminant analysis,” Patt.
Recog., vol. 42, pp. 17–26, Jan. 2009.
[21] F. Nie, H. Huang, C. Ding, D. Luo, and H. Wang, “Robust principal component analysis with non-greedy L1-norm maximization,” in
Proc. Int. Joint Conf. Artif. Intell. (IJCAI), Barcelona, Spain, Jul. 2011, pp. 1433–1438.
[22] M. McCoy and J. A. Tropp, “Two proposals for robust PCA using semidefinite programming,” Electron. J. Stat., vol. 5, pp. 1123–1160,
Jun. 2011.
[23] A. Eriksson and A. v. d. Hengel, “Efficient computation of robust low-rank matrix approximations in the presence of missing data
using the L1 norm,” in Proc. IEEE Conf. Comput. Vision Patt. Recog. (CVPR), San Francisco, CA, USA, Jun. 2010, pp. 771–778.
21
[24] R. He, B.-G. Hu, W.-S. Zheng, and X.-W. Kong, “Robust principal component analysis based on maximum correntropy criterion,”
IEEE Trans. Image Process., vol. 20, pp. 1485–1494, Jun. 2011.
[25] L. Yu, M. Zhang, and C. Ding, “An efficient algorithm for L1-norm principal component analysis,” in Proc. IEEE Int. Conf. Acoust.
Speech, Signal Process. (ICASSP), Kyoto, Japan, Mar. 2012, pp. 1377–1380.
[26] X. Li, Y. Pang, and Y. Yuan, “L1-norm-based 2DPCA,” IEEE Trans. Syst., Man. Cybern. B, Cybern., vol. 40, pp. 1170–1175, Aug.
2009.
[27] Y. Pang, X. Li, and Y. Yuan, “Robust tensor analysis with L1-norm,” IEEE Trans. Circuits Syst. Video Technol., vol. 20, pp. 172–178,
Feb. 2010.
[28] C. Ding, D. Zhou, X. He, and H. Zha, “R1 -PCA: Rotational invariant L1-norm principal component analysis for robust subspace
factorization,” in Proc. Int. Conf. Mach. Learn., Pittsburgh, PA, 2006, pp. 281–288.
[29] P. P. Markopoulos, S. Kundu, and D. A. Pados, “L1-fusion: Robust linear-time image recovery from few severely corrupted copies,”
in Proc. IEEE Int. Conf. Image Process. (ICIP), Quebec City, Canada, Sep. 2015, pp.1225–1229.
[30] N. Funatsu and Y. Kuroki, “Fast parallel processing using GPU in computing L1-PCA bases,” in Proc. IEEE TENCON 2010, Fukuoka,
Japan, Nov. 2010, pp. 2087–2090.
[31] D. Meng, Q. Zhao, and Z. Xu, “Improve robustness of sparse PCA by L1-norm maximization,” Patt. Recog., vol. 45, pp. 487–497,
Jan. 2012.
[32] H. Wang, Q. Tang, and W. Zheng, “L1-norm-based common spatial patterns,” IEEE Trans. Biomed. Eng., vol. 59, pp. 653–662, Mar.
2012.
[33] H. Wang, “Block principal component analysis with L1-norm for image analysis,” Patt. Recog. Lett., vol. 33, pp. 537-542, Apr. 2012.
[34] M. Johnson and A. Savakis, “Fast L1-eigenfaces for robust face recognition,” in Proc. IEEE West. New York Image Signal Process.
Workshop (WNYISPW), Rochester, NY, Nov. 2014, pp. 1–5.
[35] P. P. Markopoulos, “Reduced-rank filtering on L1-norm subspaces,” in Proc. IEEE Sensor Array Multichan. Signal Process. Workshop
(SAM), Rio de Janeiro, Brazil, Jul. 2016.
[36] P. P. Markopoulos, N. Tsagkarakis, D. A. Pados, and G. N. Karystinos, “Direction finding with L1-norm subspaces,” in Proc. Comp.
Sens. Conf., SPIE Def., Security, Sens. (DSS), Baltimore, MD, May 2014, pp. 91090J-1–91090J-11.
[37] N. Tsagkarakis, P. P. Markopoulos, D. A. Pados, “Direction finding by complex L1-principal component analysis,” in Proc. IEEE 16th
Int. Workshop Signal Process. Adv. Wireless Commun. (SPAWC), Stockholm, Sweden, Jun. 2015, pp. 475–479.
[38] Y. Nesterov and A. Nemirovski, “On first-order algorithms for L1/nuclear norm minimization,” Acta Numerica, vol. 22, pp. 509–575,
May 2013.
[39] Z.-Q. Luo, W.-K. Ma, A.-C. So, Y. Ye, and S. Zhang, “Semidefinite relaxation of quadratic optimization problems,” IEEE Signal
Process. Mag., vol. 27, no. 3, pp. 20-34, May 2010.
[40] G. H. Golub, “Some modified matrix eigenvalue problems,” SIAM Rev., vol. 15, pp. 318-334, 1973.
[41] J. Lu, G. Getz, E. A. Miska, E. Alvarez-Saavedra, J. Lamb, D. Peck, A. Sweet- Cordero, B. L. Ebert, R. H. Mak, A. A. Ferrando, J.
R. Downing, T. Jacks, H. R. Horvitz, and T. R. Golub, “MicroRNA expression profiles classify human cancers,” Nature, vol. 435, pp.
834-838, Jun. 2005.
[42] A. Gaur, D. A. Jewell, Y. Liang, D. Ridzon, J. H. Moore, C. Chen, V. R. Ambros, and M. A. Israel, “Characterization of microRNA
expression levels and their biological correlates in human cancer cell lines,” Cancer Res., vol. 67, pp. 2456-2468, Mar. 2007.
22
[43] J. Carlsson, S. Davidsson, G. Helenius, M. Karlsson, Z. Lubovac, O. Andrén, B. Olsson, and K. Klinga-Levan, “A miRNA expression
signature that separates between normal and malignant prostate tissues,” Cancer Cell Intern., vol. 11, pp. 1-10, May 2011.
[44] J. E. Johansson, O. Andrén, S. O. Andersson SO, P. W. Dickman, L. Holmberg, A. Magnuson, and H. O. Adami, “Natural history of
early, localized prostate cancer,” J. Amer. Med. Assoc., vol. 291, pp. 2713-2719, Jun. 2004.
[45] A. Jain, R. Duin, and J. Mao, “Statistical pattern recognition: A review,” IEEE Trans. Pattern Anal. Mach. Intell., vol. 22, no. 1, pp.
4-37, Jan. 2000.
[46] Y. Liu, S. Ge, C. Li, and Z. You, “k-ns: A classifier by the distance to the nearest subspace,” IEEE Trans. Neural Net., vol. 22, pp.
1256- 1268, Aug. 2011.
23
L1-BF for the calculation of the first L1-norm principal component (K = 1)
Input: XD×N
1: (U, Σ, V) ← csvd(X), Y ← ΣV⊤
2: bbf ← bf Y⊤ Y, sgn([Y]:,1 )
3: qbf ← Xbbf / kXbbf k2
Output: qbf
Function bf(A, b)
1:
2:
3:
4:
5:
6:
7:
8:
9:
10:
N ← length(b), L ← {1, 2, . . . , N }
P
αi ← 2bi m6=i bm [A]i,m ∀i ∈ {1, 2, . . . , N }
while true (or terminate at N BFs)
n ← argminm∈L αm
if an ≤ 0,
bn ← −bn , L ← L \ {n}
an ← −an , αi ← αi − 4bi bn [A]i,n ∀i 6= n
elseif an > 0 and |L| < N , L ← {1, 2, . . . , N }
else, break
Return b
Fig. 1. The proposed L1-BF algorithm for the calculation of the L1-principal component of a rank-d data matrix XD×N of N samples of
dimension D; csvd(·) returns the compact SVD of the argument.
24
6
Φ(Y)
Ω(Y)
B(Y)
Average Cardinality
5
4
3
2
1
2
3
4
5
6
7
Number of data points N
Fig. 2. Average cardinality of fixed-point set Φ(Y) [19], [20], bit-flipping convergence set Ω(Y), and set of optimal points B(Y) versus
number of data points N (d = 2).
25
70
60
50
40
30
kXT qL1 k1 = kXbopt k2
kXT q(t) k1
kXb(t) k2
kXkF
20
10
0
2
4
6
8
10
12
14
16
18
Iteration index t
Fig. 3. Binary quadratic metric kX⊤ b(t) k2 and L1-PCA metric kX⊤ q(t) k1 per iteration. We plot along the upper bound line kX⊤ qL1 k1 =
kXbopt k2 and lower bound line kXkF of Proposition 5.
26
1
Empirical CDF
0.8
0.6
0.4
0.2
random initialization
sv-sign initialization
0
0
2
4
6
8
10
12
14
16
18
Number of bit-flips until convergence
Fig. 4. Empirical CDF of number of bit-flips needed until convergence for sv-sign and random initialization (D = 3, N = 20, 1000
experiments).
27
TABLE I: Computational cost of L1-BF for the calculation of the principal component (K = 1) of a D × N real data matrix.
Computational Task
Computational Cost
Y and b(1)
O(N D min{N, D})
Y⊤Y
One bit flip
O(N 2 d)
O(N )
N BF iterations
O(N 2 )
qbf from bbf
O(N D)
Total
O(N D min{N, D} + N 2 d)
28
L1-BF for the calculation of K > 1 L1-norm principal components
Input: Data matrix XD×N of rank d, K ≤ d
1: (U, Σd×d , V) ← csvd(X)
2: Y ← ΣV⊤ , v ← [V]:,1 , B = sgn(v1⊤
K)
3:
Bbf ← bfK (Y, B, K, )
4:
ÛD×K , Σ̂K×K , V̂K×K ← svd(XBbf )
5: Qbf ← ÛV̂⊤
Output: Qbf
Function bfK(Yd×N , BN ×K , K ≤ d)
1:
2:
3:
4:
5:
6:
7:
8:
9:
10:
11:
12:
13:
14:
15:
16:
17:
ω ← KkY[B]:,1 k2
L ← {1, 2, . . . , K}
while true (or terminate at N K BFs)
(U, SK×K , V⊤ ) ← svd(YB), F ← US
for x ∈ L, m ← mod(x, N ), l ← (x − m)/N + 1
([q1 , qh2 ], D) ← evd(kym k22 ei1,2 e⊤
1,2 + [e2,2 , e1,2 ])
⊤
W ← [V]⊤
l,: , −2Bm,l F ym
(Z, diag(p)) ← fevd(S⊤ S + d1 Wq1 (Wq1 )⊤ )
(∼, Λ) ← fevd(diag(d) + d2 Z⊤ Wq2 (Z⊤ Wq2 )⊤ )
p
P
al,k ← j=1:K λj
(n, k) ← argmaxm,l: (l−1)N +m∈L am,l
if ω < an,k ,
Bn,k ← −Bn,k , ω ← an,k ,
L ← L \ {(k − 1)N + n}
elseif ω ≥ an,k and |L| < N K, L ← {1, 2, . . . , N K}
else, break
Return B
Fig. 5. The proposed L1-BF algorithm for the calculation of K > 1 L1-principal components of a rank-d data matrix XD×N of N samples
of dimension D; csvd(·) returns the compact SVD of the argument; fevd(·) returns the EVD of the argument (diagonal positive semidefinite
matrix perturbated by a rank-1 symmetric) by the algorithm of [40].
29
TABLE II: Computational complexity of L1-BF for the calculation of K > 1 principal components of a D × N real data matrix of rank d.
Computational Task
Computational Cost
Y and B(1)
O(N Dmin{N, D})
One bit flip
O(dK 2 + N (K 3 + dK))
N K BF iterations
O(N 2 (K 4 + dK 2 )N dK 3 )
Qbf by Bbf
O(N Dmin{N, D} + N DK)
Total
O(N Dmin{N, D} + N 2 (K 4 + dK 2 ) + N dK 3 )
30
(a)
Empirical CDF
1
0.8
0.6
0.4
L1-BF (proposed)
SDP [22]
FP iterations [19] / Alternating optimization [21]
0.2
0
0.05
0.1
0.15
0.2
0.25
0.3
0.35
0.4
Performance degradation ratio
(b)
Empirical CDF
1
0.98
0.96
0.94
L1-BF (proposed)
SDP [22]
FP iterations [19] / Alternating optimization [21]
0.92
0
0.01
0.02
0.03
0.04
0.05
0.06
Performance degradation ratio
Fig. 6. Empirical CDF of ∆(qbf ; X), ∆(qfp ; X), and ∆(qsdp ; X) (D = 4, N = 16) for (a) L = 1 and (b) L = N = 16 initializations.
31
(a)
1
Empirical CDF
0.8
0.6
0.4
L1-BF (proposed)
SDP [22]
FP iterations [19]
Alternating optimization [21]
0.2
0
0
0.05
0.1
0.15
0.2
0.25
0.3
0.35
0.4
Performance degradation ratio
(b)
1
Empirical CDF
0.9
0.8
0.7
0.6
L1-BF (proposed)
SDP [22]
FP iterations [19]
Alternating optimization [21]
0.5
0.4
0
0.02
0.04
0.06
0.08
0.1
0.12
Performance degradation ratio
Fig. 7. Empirical CDF of ∆(Qao ; X), ∆(Qfp ; X), ∆(Qsdp ; X), and ∆(Qbf ; X) for (a) L = 1 and (b) L = N K = 16 initializations.
32
(a)
20
15
10
5
0
-5
Nominal training data
L2-PC (by SVD)
PC by L1-BF (proposed)
True max-variance line
-10
-15
-20
-40
-30
-20
-10
0
10
20
30
40
(b)
20
15
10
5
0
-5
Nomining training data
Outliers
L2-PC (by SVD)
PC by L1-BF (proposed)
True max-variance line
-10
-15
-20
-40
-30
-20
-10
0
10
20
30
40
Fig. 8. L2-PC and L1-PC (by L1-BF) trained over (a) the clean data matrix X ∈ R2×100 (100 nominal data points) and (b) the outlier
corrupted data matrix Xcor ∈ R2×104 (same 100 nominal data points plus 4 outliers).
33
Frequency of detection
1
0.8
p=0
0.6
p=2
p=4
0.4
0.2
L2-PCA
L1-PCA (proposed)
0
0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
1
Frequency of false alarm
Fig. 9. ROC curve of L2-PCA and L1-PCA (L1-BF) malignant tissue detector for p = 0, 2, 4 mislabeled points per training dataset (N = 19,
Ntrain = 10, D = 9, K = 3).
34
| 8cs.DS
|
arXiv:1609.09000v1 [cs.DB] 28 Sep 2016
StruClus: Structural Clustering of Large-Scale
Graph Databases
Till Schäfer
Petra Mutzel
TU Dortmund University
Dept. of Computer Science
44227 Dortmund, Germany
Email: [email protected]
TU Dortmund University
Dept. of Computer Science
44227 Dortmund, Germany
Email: [email protected]
Abstract—We present a structural clustering algorithm for
large-scale datasets of small labeled graphs, utilizing a frequent
subgraph sampling strategy. A set of representatives provides
an intuitive description of each cluster, supports the clustering process, and helps to interpret the clustering results. The
projection-based nature of the clustering approach allows us to
bypass dimensionality and feature extraction problems that arise
in the context of graph datasets reduced to pairwise distances
or feature vectors. While achieving high quality and (human)
interpretable clusterings, the runtime of the algorithm only grows
linearly with the number of graphs. Furthermore, the approach is
easy to parallelize and therefore suitable for very large datasets.
Our extensive experimental evaluation on synthetic and real
world datasets demonstrates the superiority of our approach over
existing structural and subspace clustering algorithms, both, from
a runtime and quality point of view.
I. I NTRODUCTION
Molecules, protein interaction networks, XML documents,
social media interactions, and image segments all have in common that they can be modeled by labeled graphs. The ability to
represent topological and semantic information causes graphs
to be among the most versatile data structures in computer
science. In the age of Big Data, huge amounts of graph data
are collected and the demand to analyze them increases with
its collection. We focus on the special case of clustering
large sets of small labeled graphs. Our main motivation stems
from the need to cluster large-scale molecular databases for
drug discovery, such as PubChem1 , ChEMBL2 , ChemDB3 or
synthetically constructed de-novo databases [1], which contain
up to a billion molecules. However, the presented approach is
not limited to this use case.
Clustering techniques aim to find homogeneous subsets in a
set of objects. Classical approaches do not interpret the objects
directly, but abstract them by utilizing some intermediate
representation, such as feature vectors or pairwise distances.
While the abstraction over pairwise distances is beneficial in
terms of generality, it can be disadvantageous in the case of
intrinsic high dimensional datasets [2, 3]. In this case the
concentration effect may cause the pairwise distances to loose
their relative contrast; i.e. the distances converge towards a
1 https://pubchem.ncbi.nlm.nih.gov/
2 https://www.ebi.ac.uk/chembl/
3 http://cdb.ics.uci.edu/
common value [4]. The concentration effect is closely related
to a bad clusterability [5].
Sets of graphs are usually clustered by transforming the
graphs to feature vectors or by using graph theoretic similarity
measures.
Typical feature extraction methods for graphs are: counting
graphlets, that is, small subgraphs [6, 7], counting walks [8],
and using eigenvectors of adjacency matrices (spectral graph
theory) [9]. The enumeration of all subgraphs is considered
intractable even for graphs of moderate size, because there
exist up to exponentially many subgraphs wrt the graph size.
Many efficient clustering algorithms have been proposed for
vector space. Therefore, the transformation to feature vectors might look beneficial in the first place. However, the
above mentioned feature extraction methods tend to produce
a large amount of distinct and often unrelated features. This
results in datasets with a high intrinsic dimensionality [10,
11]. Additionally, the extracted features only approximate the
graph structure, which implies that feature vectors cannot be
transformed back into a graph. Hence, the interpretability of
clustering algorithms which perform vector modifications (e.g.
calculating centroids) is limited.
Besides the utilization of various feature extraction techniques, it is possible to compare graphs directly using graph
theoretic distances such as (maximum) common subgraph
derived distances [12–14] or the graph edit distance [15].
The computation of the previously mentioned graph theoretic
distances is NP-hard and as shown in [11] their application
results in datasets with a high intrinsic dimensionality as well.
High quality clustering methods for arbitrary (metric) and
high dimensional datasets furthermore require a superlinear
number of exact distance computations. These factors render
graph theoretic distance measures in combination with generic
clustering algorithms infeasible for large-scale datasets.
Subspace and projected clustering methods tackle high
dimensional datasets by identifying subspaces in which well
separated clusters exist. However, generic subspace algorithms
come with a high runtime burden and are often limited to an
euclidean vector space.
Our structural projection clustering algorithm approaches
the dimensionality problems by explicitly selecting cluster
representatives in form of common subgraphs. Consider a
feature mapping to binary feature vectors that contain one
feature for each subgraph that is found in the complete dataset.
For each graph, its feature vector has binary entries encoding
the presence of the associated substructure graph. Selecting a
common subgraph S as cluster representative is another way
of selecting a subspace in these feature vectors consisting of
all features associated with subgraphs of S.
Our main contributions in this paper are: We present a
novel structural projection clustering algorithm for datasets of
small labeled graphs which scales linearly with the dataset
size. A set of representatives provides an intuitive description
of each cluster, supports the clustering process, and helps to
interpret the clustering results. Up to our knowledge, this is
the first approach actively selecting representative sets for each
cluster based on a new ranking function. The candidates for
the representatives are constructed using frequent subgraph
sampling. In order to speed up the computation, we suggest
a new error bounded sampling strategy for support counting
in the context of frequent subgraph mining. Our experimental
evaluation shows that our new approach outperforms competitors in its runtime and quality.
The paper is structured as follows: Section II provides an
overview of related clustering algorithms. Basic definitions are
given in Section III. Section IV presents the main algorithm
and a runtime analysis. Our experimental evaluation in which
we compare our new algorithm to SCAP [16], PROCLUS [17]
and Kernel K-Means [18] is presented in Section V.
II. R ELATED W ORK
Several clustering algorithms for graph and molecule data
have been proposed in the last years. Tsuda and Kudo [10]
presented an EM algorithm using a binomial mixture model
over very high dimensional binary vectors, indicating the presence of all frequent substructures. Two years later, Tsuda and
Kurihara [19] presented a similar graph clustering algorithm
using a Dirichlet Process mixture model and pruning the set
of frequent substructures to achieve smaller feature vectors. A
K-Median like graph clustering algorithm has been presented
by Ferrer et al. [20], that maps each graph into the euclidean
space utilizing the edit distance to some pivot elements. A
median graph is then selected based on the distance to the
euclidean median. Furthermore, a parallel greedy overlapping
clustering algorithm has been presented by Seeland et al. [21].
It adds a graph to a cluster whenever a common substructure
of a user-defined minimum size exists. However, none of
the previously mentioned algorithms are suitable for large
datasets as a result of their high computational complexity.
XProj [22] uses a projection-based approach by selecting all
(enumerated) frequent substructures of a fixed size as cluster
representatives. The approach scales well with the dataset
size but is limited to trees. A generalization to graphs would
result in a huge performance degradation. Furthermore, there
exist some hybrid approaches, that pre-cluster the dataset by
using a vector-based representation and refine the results using
structural clustering algorithms. The most relevant with respect
to large-scale datasets is the SCAP algorithm proposed by
Seeland et al. [16].
Of course many subspace algorithms are also applicable after using a feature extraction method. Giving a comprehensive
overview on subspace techniques is out of the scope of this
article. However, there are two algorithms that are of special
interest for this work: In [23] it is shown that frequent pattern
mining can be used for feature selection in vector space.
This relates to XProj and our approach in the way, that the
selection of a graph representative—with the help of frequent
substructure mining—is another way of selecting a subspace
in the feature space of substructures. In the later evaluation
we will compare ourself to the PROCLUS [17] algorithm. It
is a fast projected clustering algorithm with noise detection,
that selects features by minimizing variance. The algorithm
has been studied intensively and performed well in various
subspace clustering comparisons [24, 25].
III. P RELIMINARIES
An undirected labeled graph G = (V, E, l) consists of
a finite set of vertices V (G) = V , a finite set of edges
E(G)U= E ⊆ {{u, v} ⊆ V | u 6= v} and a labeling function
l : V E → L, where L is a finite set of labels. |G| is used
as a short term for |V (G)| + |E(G)|. A path of length n is
a sequence of vertices (v0 , . . . , vn ) such that {vi , vi+1 } ∈ E
and vi 6= vj for i 6= j. Let G and H be two undirected
labeled graphs. A (label preserving) subgraph isomorphism
from G to H is an injection ψ : V (G) → V (H), where ∀v ∈
V (G) : l(v) = l(ψ(v)) and ∀u, v ∈ V (G) : {u, v} ∈ E(G) ⇒
{ψ(u), ψ(v)} ∈ E(H)∧l({u, v}) = l({ψ(u), ψ(v)}). Iff there
exists a subgraph isomorphism from G to H we say G is
supported by H, G is a subgraph of H, H is a supergraph
of G or write G ⊆ H. If there exists a subgraph isomorphism
from G to H and from H to G, the two graphs are isomorphic.
A common subgraph of G and H is a graph S, that is subgraph
isomorphic to G and H. Furthermore, the support sup(G, G)
of a graph G over a set of graphs G is the fraction of graphs
in G, that support G. G is said to be frequent, iff its support is
larger or equal than a minimum support threshold minSup. A
frequent subgraph G is maximal, iff there exists no frequent
supergraph of G. For a set of graphs G, we write F(G)
for the set of all frequent subgraphs and M(G) for the set
of all maximal frequent subgraphs. A clustering of a graph
dataset X is a partition C = {C1 , . . . , Cn } of X . Each cluster
C ∈ C consists of a set of graphs and is linked to a set of
cluster representatives R(C) = {R1 , . . . , Rk } which are itself
undirected labeled graphs. Please note, that we consider each
graph in our dataset to be a distinct object. As a result, it is
possible to have isomorphic graphs in a single set.
IV. T HE StruClus A LGORITHM
A high level description of the StruClus algorithm is given
in Algorithm 1. Initially, it partitions the dataset using a
lightweight pre-clustering algorithm. Afterwards, the clustering is refined using an optimization loop similar to the KMeans algorithm. In order to fit the number of clusters to the
dataset structure (i.e., to achieve a good cluster separation and
homogeneity, see Sections IV-B and IV-D), we use a cluster
splitting and merging strategy in each iteration.
Algorithm 1 StruClus Algorithm
1: apply pre-clustering {Section IV-E}
2: while not convergent {Section IV-F} do
3:
split clusters {Section IV-D}
4:
merge clusters {Section IV-D}
5:
update representatives {Section IV-B}
6:
assign graphs to closest cluster {Section IV-C}
7: end while
∀G, H ∈ G⊆ : G ⊆ H ⇒ sup(G, G) ≥ sup(H, G)
An important ingredient of our algorithm is the set of
representatives R(C) for each cluster C ∈ C. Representatives
serve as an intuitive description of the cluster and define the
substructures over which intra cluster similarity is measured.
The set R(C) is chosen such that for every graph G ∈ C there
exists at least one representative R ∈ R(C) which is subgraph
isomorphic (i.e., supported by G). With the exception of a
single noise cluster the following invariant holds after each
iteration:
∀C ∈ C : ∀G ∈ C : ∃R ∈ R(C) : R ⊆ G.
S 0 or otherwise removed. If no further extension is possible
without violating the minimum support threshold, a maximal
frequent subgraph S has been found. This process is justified
by the monotonicity property of subgraphs G⊆ of graphs in G:
(1)
The representative set R(C) of a cluster C ∈ C is
constructed using maximal frequent subgraphs of C (see
Section IV-A). Having a representative set instead of a single
representative has the advantage, that graphs composed of
multiple common substructures can be represented. In order
to be meaningful and human interpretable, the cardinality of
R(C) is limited by a user defined value Rmax . Figure 1 shows
two example clusters and their representatives of a real world
molecular dataset generated with StruClus.
A. Stochastic Representative Mining
We construct the representative set R(C) of a cluster C ∈ C
using maximal frequent subgraphs of C. Since the set M(C)
may have exponential size wrt the maximal graph size in
C, we restrict ourselves to a subset of candidate representatives S(C) ⊆ M(C) using a randomized maximal frequent
connected subgraph sampling technique from ORIGAMI [26]
combined with a new stochastic sampling strategy for support
counting. In a second step, the final representative set R(C) ⊆
S(C) is selected using a ranking function (see Section IV-B1).
ORIGAMI constructs a maximal frequent connected subgraph S ∈ S(G) over a set of graphs G by extending a
random frequent vertex with frequent paths of length one
leading to a graph S 0 . In a first step, all frequent vertices
N(G) and all frequent paths of length one P(G) are enumerated
with a single scan of G. Then, for each extension, a random
vertex of S 0 is chosen and a random, label preserving path
p ∈ P(G) is connected to it in a forward (creating a new
vertex) or backward (connecting two existing vertices) fashion.
After each extension, the support sup(S 0 , G) is evaluated by
solving a subgraph isomorphism test for all graphs in G. If
sup(S 0 , G) ≥ minSup, the extension is permanently added to
(2)
While ORIGAMI greatly improves performance in comparison
with enumeration algorithms, the |G| subgraph isomorphism
tests for each extension remain a major performance bottleneck for StruClus. For this reason, we have added a stochastic
sampling strategy for support counting.
Initially, we draw a random sample H ⊆ G. Then θ̂ =
sup(S 0 , H) is an estimator for the parameter θ of a binomial distribution B(·, θ), where θ = sup(S 0 , G) is the true
probability of the underlying Bernoulli distribution. We are
interested if the true value of sup(S 0 , G) is smaller than the
minimum support threshold. Without loss of generality, let us
focus on the case θ̂ < minSup in the following. We can take
advantage of a binomial test under the null hypothesis, that
θ ≥ minSup and thereby determine the probability of an
error, if we assume sup(S 0 , G) < minSup. With a predefined
significance level α we can decide if the sample gives us
enough confidence to justify our assumption. If we cannot
discard our null hypothesis, we continue by doubling the
sample size |H| and repeat the process. In the extreme case
we will therefore calculate the exact value of sup(S 0 , G).
The statistical test is repeated for each extension and each
sample size doubling. As a consequence, a multiple hypothesis
testing correction is necessary to bound the real error for S to
be a maximal frequent substructure of G.
Proposition IV-A.1. Let G be a set of undirected labeled
graphs, |Hmin | the minimal sample size, P(G) the set of all
frequent paths of length one, minSup a minimum support
threshold, and Vmax the (1 − minSup)-quantile of the sorted
(increasing order) graph sizes of each graph in G. Then the
maximal number of binomial tests to construct a maximal
frequent substructure over G is bounded by:
|G|
2
(Vmax
+ Vmax ) |P(G)|
log
|Hmin |
m
l
times if the
Proof. The sample size is doubled log |H|G|
|
min
test never reaches the desired significance level. The size of
some S ∈ M(G) is bounded by the size of each supporting
graph. In the worst case S is supported by the (|G|minSup)largest graphs of G. The graph size of the smallest supporting
graph is then equal to the (1 − minSup)-quantile of the sorted
graph sizes in increasing order. The number of backward
extensions is bounded by the number of vertex pairs times
the number of applicable extensions p ∈ P(G). Additionally,
we need to check |P(G)| forward extensions for each vertex
to conclude that S is maximal.
Finally, with the value of Proposition IV-A.1 we are able
to apply a Bonferroni correction to our significance level. We
can afford a relative high error, because the selection of the
Fig. 1. Two real world clusters generated with StruClus. The grey boxes show the cluster representatives.
final representatives will filter out bad candidates. However,
the used significance level has an influence on the runtime.
On the one hand a high error leads to many bad candidates
and we need to increase the number of candidates to mine. On
the other hand a low error will lead to larger sample sizes to
reject the null hypothesis. A maximal error of 50% has turned
out to be a good choice during our experimental evaluation.
B. Update Representatives
1) Representative Selection: In its role as a cluster description, a good representative R ∈ R(C) explains a large
portion of its cluster. Accordingly, it should (a) be supported
by a large fraction of C and (b) cover a large fraction of
vertices and edges of each graph G ∈ C supporting R. We
define the coverage of a graph G by a representative R as
|R|
cov(R, G) := |G|
. The two criteria are closely related to the
cluster homogeneity. A uniform cluster, that is, a cluster that
contains only isomorphic graphs, can achieve optimal values
for both criteria. Vice versa, the monotonicity property (2)
implies non-optimal values for inhomogeneous clusters for at
least one of the two criteria. As homogeneous clusters are
desired, we use a product of the two criteria for our ranking
function.
In order to discriminate clusters from each other, a cluster
representative should be cluster specific as well. Thus, its
support in the rest of the dataset should be low. For this reason,
we use the following ranking function for a dataset X , cluster
C and representative R ∈ R(C):
CR := {G ∈ C | R ⊆ G}
|CR | |R|
(sup(R, C) − sup(R, X ))
rank(R) := P
G∈CR |G|
(3)
Finally, we select the Rmax highest ranked sampled subgraphs
from S(C) as cluster representatives R(C).
2) Balancing Cluster Homogeneity: Besides the representative selection, the choice of the minimum support threshold
for representative mining has an influence on the cluster
homogeneity. The fraction of unsupported graphs for a cluster C after updating the cluster representatives is bounded
by (1 − minSup). The bound clearly relates to criteria (a)
from the representative ranking (see Section IV-B1). However,
unsupported graphs are removed from the cluster and will be
assigned to a different cluster at the end of the current iteration
(see Section IV-C). Due to the monotonicity property (2), this
process of sorting out graphs by choosing a minimum support
threshold below 1, will increase the size of the representatives
and therefore our coverage value, i.e criteria (b). A decrease
of the minimum support threshold will lead to an increase in
the size of the representatives. Subsequently, this process will
also reduce the cluster cardinality. Therefore, increasing the
homogeneity to an optimal value will result in a clustering
with uniform or singleton clusters, which is clearly not the
desired behavior.
To get around this, we will aim towards a similar homogeneity for all clusters and choose the minimum support threshold
cluster specific. However, choosing a fixed homogeneity level
a priori is not an easy task, as an appropriate value depends on
the dataset. Therefore we will calculate an average coverage
score over all clusters and use this as a baseline adjustment.
For the ease of computation, we choose a slightly simplified
coverage approximation:
aCov(C) =
relCov(C, C) =
1
|R(C)|
1
|C|
1
|C|
P
R∈R(C) |R|
P
G∈C |G|
aCov(C)
P
0
C 0 ∈C aCov(C )
(4)
(5)
Finally, we can define a linear mapping from the relative coverage relCov to a cluster specific minimum support threshold
with the help of two predefined tuples (ls, lr ) and (hs, hr ),
where ls < hs and lr < hr . The parameter ls (hs) denotes the
lowest (highest) support value and lr (hr ) the relative coverage
value mapped to the lowest (highest) minimum support:
hs − ls
hs − ls
f(C, C) := relCov(C, C)
+ (ls − lr
)
hr − lr
hr − lr
if relCov(C, C) < lr
ls
minSup(C, C) := hs
(6)
if relCov(C, C) > hr
f(C, C) otherwise.
To result in a minimum support threshold of 1 for all clusters
(i.e. stopping the process of sorting out graphs if the clustering
is balanced), we will set the values of the parameters hs very
close or equal to 1 and hr < 1.
C. Cluster Assignment
Each graph G in the dataset is assigned to its most similar
cluster in the assignment phase. As a measure for similarity,
we are summing up the squared sizes of the representatives of
a cluster, which are subgraph isomorphic to G. This choice of
similarity is once more justified by the representative ranking
criteria. We square the representative sizes, to prefer a high
coverage over a high number of representatives to be subgraph
isomorphic to the assigned graph.
As mentioned in Section IV-B2 it is possible that a graph
G ∈ C will no more be supported by any representative R ∈
R(C) of its cluster C after updating the representatives. In this
situation we will create a single noise cluster, where all graphs
(of all clusters) that are not supported by any representative
are collected. As the minimum support threshold is bounded
by the fixed value ls (see Section IV-B2) and the number
of representatives is limited by Rmax , it is not guaranteed
that we can find an appropriate set of representatives for this
most likely largely inhomogeneous noise cluster. It is therefore
excluded from the invariant (1).
The problem of finding all subgraph isomorphic graphs in a
graph database is also known as the subgraph search problem
and was extensively studied in the past. We apply the fingerprint pre-filtering technique CT-Index [27], which has emerged
from this research topic, to speed up the assignment phase.
CT-Index enumerates trees and circles up to a specified size
for a given graph and hashes the presence of these subgraphs
into a binary fingerprint of fixed length. If the fingerprint of
a graph G has a bit set that is unset in the fingerprint of
a graph H we can conclude that no subgraph isomorphism
from G to H exists, because G contains a subgraph that is not
present in H. We calculate a fingerprint for each graph and
representative and only perform a subgraph isomorphism test
in our assignment phase if the fingerprint comparison cannot
rule out the presence of a subgraph isomorphism.
D. Cluster Splitting and Merging
Without the operation of cluster splitting, the above mentioned process of creating noise clusters would create at most
one extra cluster in each iteration of our main loop. A large
difference in the initial and final number of clusters therefore
would lead to a slow convergence of the StruClus algorithm
towards its final result. As mentioned before, it is also possible
that no representative is found at all for the noise cluster,
and therefore the process of sorting out graphs from the noise
cluster to increase its homogeneity is stopped completely in the
worst case. A similar situation can occur for regular clusters.
For example, if a cluster is composed of uniform sets T ∈ T
of graphs, we will require a minimum support threshold less
1
than or equal to 1 − minT ∈T
{|T |} to sort the smallest possible
number of graphs out. For this reason, a cluster splitting step is
necessary (see Algorithm 1). In this step, all clusters that have
a relative coverage value below an a priori specified threshold
relCovmin will be merged into a single set of graphs and
the pre-clustering algorithm is applied on them. The resulting
clusters are added back to the clustering.
On the contrary to cluster splitting, which focuses on cluster
homogeneity, cluster merging ensures a minimum separation
between clusters. Separation can be measured on different levels. Many classical measures define separation as the minimum
distance between two cluster elements. However, this type of
definition is not suitable for projected clustering algorithms,
because the comparison does not take the cluster specific
subspace into account. As mentioned in the introduction,
the cluster representatives in StruClus define the subspace
of the cluster. Additionally, they serve as a description of
the graphs inside the cluster itself. A high coverage value
leads to an accurate cluster description. Thus, we will define
separation between two clusters C and C 0 solely over the
representatives sets R(C) and R(C 0 ). This definition is also
beneficial from a runtime perspective, as separation calculation
is independent of the cluster size. Without cluster merging it
is possible that clusters with very similar representatives do
exist. Although the pre-clustering will ensure that the initial
clusters will have dissimilar representatives (see Section IV-E)
it may happen that two clusters converge towards each other or
that newly formed clusters are similar to an already existing
one. Therefore, we will merge two clusters whenever their
representatives are too similar.
To compare two single representatives R, R0 we calculate
the size of their maximum common subgraph (MCS) and use
its relative size as similarity:
sim(R, R0 ) :=
|mcs(R, R0 )|
max{|R|, |R0 |}
(7)
The maximum of the representatives sizes is chosen as denominator to support different clusters with subgraph isomorphic
representatives, which differ largely in size. Finally, we will
merge two clusters C and C 0 if the following condition holds:
|{(R, R0 ) ∈ R(C) × R(C 0 ) | sim(R, R0 )
≥ sim min }| ≥ sim num
(8)
where sim num is a minimum number of representative pairs
which have a similarity greater than or equal to sim min .
Note that the calculated MCS between two representatives
R ∈ R(C), R0 ∈ R(C 0 ) is supported by all the graphs
G ∈ C ∪ C 0 , that support either R or R0 , because the
subgraph isomorphism relation is transitive. The coverage for
these graphs in the merged cluster is furthermore bounded
by max{cov(G, R), cov(G, R0 )} sim(R, R0 ) if we reuse the
MCS as representative. For this reason, we recommend to set
sim num close to the number of representatives per cluster to
support a large fraction of graphs in the merged cluster. The
parameter sim min is furthermore an intuitive knob to adjust
the granularity of the clustering.
Finally, the representatives for the merged clusters are updated. We do not use the calculated MCSs as representatives,
because better representatives may exist and the decision
problem “Does an MCS larger than some threshold exist?”
is computationally less demanding than calculating the MCS
itself.
E. Pre-Clustering
The pre-clustering serves as an initial partitioning of the
dataset. A random partitioning of all graphs would be problematic as representatives may not be found for all partitions and
the found representatives are most likely not cluster specific.
This will result in a high number of clusters to be merged to
a few inhomogeneous clusters and in a slow convergence of
the StruClus algorithm.
To pre-cluster the dataset X , we compute maximal frequent
subgraphs S(X ) ⊆ M(X ), as described in Section IV-A with
a fixed minimum support. These frequent subgraphs serve as
representative candidates for the initial clusters. To avoid very
similar representatives we will first greedily construct maximal
sets of dissimilar graphs. As a measure of similarity we reuse the similarity (7) and the threshold sim min from cluster
merging. In other words, we are picking all graphs G from
S(X ) in a random order and add G to our dissimilar set D, if
@H ∈ D with sim(G, H) < sim min . This process is repeated
several times and the largest set Dmax is used to create one
cluster for each H ∈ D with H as single representative.
Afterwards we run a regular assignment phase as described in
Section IV-C. As a result of re-using sim min we can expect,
that we have well separated cluster and no cluster merging is
necessary in the first iteration of the main loop (excluding the
noise cluster).
F. Convergence
StruClus optimizes cluster homogeneity while maintaining
a minimal cluster separation. As described in Section IV-B2,
homogeneity is defined by two criteria wrt the cluster representatives. With an exception of the noise cluster, the
representative support criteria (a) can be dismissed, because
not represented graphs will be sorted out of the cluster. It
was further discussed, that an optimal homogeneity can be
achieved for singleton clusters. However, the later introduced
cluster separation constraint will limit the granularity of the
clustering, because two similar representatives will be merged.
Nevertheless, there might exist several clusterings with different granularity that respect the separation constraint. For this
reason the objective function balances the coverage criteria (b)
of the homogeneity and the granularity of the clustering. The
parameter α adjusts this granularity:
P
|C| aCov(C)
(9)
z(C) := C∈C
α
|C|
As a consequence of the cluster splitting and merging, the
objective function will fluctuate and contain local optima. We
will therefore smooth the objective function. Let Ci be the
clustering after the i-th iteration. Algorithm 1 will terminate
after the first iteration c for which the following condition
holds:
P
c−w<i≤c z(Ci )
≤1+β
(10)
c≥s+w∧ P
c−s−w<i≤c−s z(Ci )
where w is the averaging width and β is the minimum relative
increase of the objective function in s iterations.
G. Runtime Analysis
The subgraph isomorphism problem and the maximum
common subgraph problem are both NP-complete even in their
decision variants [28]. StruClus solves these problems in order
to calculate the support and to decide if clusters need to be
merged. Furthermore, the size of a representative can be linear
in the sizes of the supporting graphs. Thus, StruClus scales
exponentially wrt the size of the graphs in the dataset X .
Nevertheless, these problems can be solved sufficiently fast
for small graphs with a few hundred vertices and edges, e.g.,
molecular structures. We will therefore consider the graph size
a constant in the following analysis and focus on the scalability
wrt the dataset size. Let Vmax be the maximal number of
vertices for a graph in X and Cmax the maximal number of
clusters during the clustering process.
Representative Mining: As described in Proposition IV-A.1 the number of extensions to mine a single maximal
frequent subgraph from a set of graphs G ⊆ X is bounded
2
by (Vmax
+ Vmax ) |P(G)|. The variable Vmax is obviously
a constant. Also |P(G)| is only bounded by Vmax and the
lowest minimum support threshold ls, as a fraction of dlsV 2|G|e
max
of all edges in G must be isomorphic to be frequent. Thus,
2
Vmax
there exist at most ls frequent paths of length one. Note,
that the number of distinct labels does not have an influence
on this theoretical bound. For each mined maximal frequent
subgraph we need to calculate the support over G and this
involves O(|G|) many subgraph isomorphism tests. Therefore,
the runtime to mine a single maximal frequent subgraph is
bounded by O(|G|).
Representative Update: Updating the representatives involves two steps: representative mining and representative
selection. Representative mining includes the calculation of
the cluster specific minimum support with the help of relCov.
The relCov values for each cluster can be calculated in
O(Cmax ) if we maintain a sum of the graph sizes of each
cluster and update this value during cluster assignment. Since
the cluster sizes sum up to |X | the runtime of the actual
frequent subgraph mining is in O(X ). To rank a representative
candidate R ∈ R(C), the set CR , the value sup(R, C), and the
value sup(R, X ) can be calculated by a single scan over X , i.e.
calculating whether R is subgraph isomorphic to each graph
in X . As the number of candidates per cluster is a constant,
the runtime of the ranking is in O(Cmax |X |), which is also
the overall runtime of the representative update.
Other Parts: The runtimes of the cluster assignment and
the pre-clustering are in O(Cmax |X |). Cluster splitting is com2
putable in O(Cmax + |X |), cluster merging in O(Cmax
+ |X |),
and converge in O(Cmax ) time.
Overall Runtime: A single iteration of Algorithm 1 has a
runtime of O(Cmax |X |), and hence is linear. This is justified
2
) ≤ O(Cmax |X |). Let m be
by the observation that O(Cmax
the number of iterations after Algorithm 1 terminates. Then,
the overall runtime is in O(m Cmax |X |).
V. E VALUATION
In the following evaluation we compare StruClus to SCAP
[16], PROCLUS [17] and Kernel K-Means [18] wrt their
runtime and the clustering quality. Furthermore, we evaluate
the influence of our sampling strategy for support counting
and evaluate the parallel scaling of our implementation.
Hardware & Software: All tests were performed on a
dual socket NUMA system (Intel Xeon E5-2640 v3) with 128
GiB of RAM. The applications were pinned to a single NUMA
domain (i.e. 8 cores + HT / 64 GiB RAM) to eliminate random
memory effects. Turbo Boost was deactivated to minimize
external runtime influences. Ubuntu Linux 14.04.1 was used
as operating system. The Java implementations of StruClus,
PROCLUS and Kernel K-Means were running in an Oracle
Java Hotspot VM 1.8.0 66. SCAP was compiled with GCC
4.9.3 and -O3 optimization level. StruClus and SCAP are
shared memory parallelized.
Test Setup & Evaluation Measures: The tests were
repeated 30 times if the runtime was below 2 hours and
15 times otherwise. Quality is measured by ground truth
comparisons. We use the Normalized Variation of Information
(NVI) [29] and Fowlkes-Mallows (FW) [30] measures for
StruClus, PROCLUS and Kernel K-Means. While NVI and
FW are established quality measures, they are not suited for
overlapping clusterings as produces by SCAP. Thus, Purity
[31] was used for comparison with SCAP.
Datasets: We evaluate the algorithms on synthetic
datasets of different sizes and three real world datasets. The
synthetic datasets have ≈ 35 vertices and ≈ 51 edges on
average with 10 vertex and 3 edge labels (weights drawn
from an exponential distribution). They contain 100 clusters
and 5% random noise graphs. For each cluster 3 seed patterns
were randomly generated with a Poisson distributed number of
vertices (mean 10) and an edge probability of 10%. Additionally, we created a common seed pool with 100 (noise) seeds.
The cluster specific graphs were then generated by connecting
the cluster specific seeds with 0 to 2 common noise seeds.
The first real world dataset (AnchorQuery) is a molecular
de-novo database. Each molecule is the result of a chemical
reaction of multiple purchasable building blocks. We have
used 11 reaction types, i.e., class labels, from AnchorQuery4 .
The second real world dataset (Heterocyclic) is similar to
the first one, but contains heterocyclic compounds and 39
distinct reaction types. The third real world dataset (ChemDB)
contains 5 million molecules from ChemDB [32]. ChemDB
is a collection of purchasable molecules from 150 chemical
vendors. The dataset has no ground truth and will mainly
serve as proof that we are able to cluster large real world
datasets. Table I shows additional statistics about these real
world datasets.
To cluster the datasets with PROCLUS the graph data was
transformed to vectors by counting distinct subgraphs of size 3,
resulting in 7 000 to 10 000 features on the synthetic datasets.
The application to the AnchorQuery and Heterocyclic datasets
results in much lower feature counts of 274 and 133. The same
vectors were used as features space representation for Kernel K-Means. Additionally, we have evaluated various other
graph kernels with explicit feature mapping for PROCLUS
and Kernel K-Means, such as the Weisfeiler-Lehman shortest
path, Weisfeiler-Lehman subtree and fixed length random walk
kernels [33]. However, we observed feature vectors with even
higher dimensionality and clustering results with lower quality.
For this reason, the results of these graph kernels are omitted
here.
Algorithm Configuration: The parameters of StruClus
were set as follows: maximal error for support counting: 50%;
number of representative candidates: 25; minimum support
threshold calculated with: lr = 0; ls = 0.4; hr = 0.9; hs =
0.99; splitting threshold: aCov ≤ 0.6; minimum separation
sim min = 0.3; sim num = 3; convergence determined with:
α = 1, β = 0.01, w = 3, s = 3. SCAP has two parameters:
(a) the minimum size for a common substructure, that must be
present in a single cluster and (b) a parameter that influences
the granularity of the fingerprint-based pre-partitioning. (a)
was set to 80% of the mean seed size for the synthetic dataset,
otherwise to 8. (b) was set to the highest number that results
in a reasonable clustering quality, as the clustering process
is faster with fine grained pre-partitioning, but cannot find
any clusters over the pre-partition boundary. For the synthetic
dataset this was 0.2. PROCLUS has two parameters as well:
(a) the number of clusters and (b) the average dimensionality
of the cluster subspaces. (a) was set to the number of clusters
in the ground truth. The ChemDB dataset, which has no
ground truth, was too large for PROCLUS. (b) was set to 20,
for which we got best quality results. The number of clusters
4 http://anchorquery.csb.pitt.edu/reactions/
TABLE I
R EAL WORLD DATASET STATISTICS . C UMULATED VALUES ARE GIVEN AS TRIPLE MIN / MAX / AVERAGE .
dataset
size
AnchorQuery
Heterocyclic
ChemDB
65 700
10 000
5 000 000
classes
# vertices
# edges
# vertex labels
# edge labels
11
39
–
11 / 90 / 79.19
9 / 69 / 42.99
1 / 684 / 50.74
11 / 99 / 86.02
10 / 79 / 47.35
0 / 745 / 53.20
6
25
86
5
5
5
Scaling
is the only parameter of Kernel K-Means and was set in exact
the same manner as for PROCLUS. Please note, that the a
priory selection of this optimal value gives PROCLUS and
Kernel K-Means an advantage over their competitors during
the evaluation.
Results: As shown in Table II, StruClus outperforms
all competitors in terms of quality on the synthetic datasets.
The differences of all corresponding mean quality values are
always larger than small multiples of the standard deviations.
For small synthetic datasets, SCAP was the fastest algorithm.
This changes for larger data set sizes. StruClus outperforms
SCAP by a factor of ≈ 13 at size 500 000. Furthermore,
we were unable to cluster the largest dataset with SCAP in
less than 2 days. PROCLUS and Kernel K-Means were the
slowest algorithms with a huge gap to their next competitor.
The sublinear growth of StruClus’s runtime for the smaller
datasets sizes is caused by our sampling strategy for support
counting. Running StruClus with exact support counting yields
a linear growth in runtime. It is important to mention that the
clustering quality is not significantly influenced by the support
counting strategy.
Our implementation of StruClus scales well with the number
of cores as shown in Fig. 2. With 8 cores we get a speedup of
7.15. Including the Hyper-Threading cores, we get a speedup
of 9.11.
10
8
6
4
2
00
Optimal Scaling
Scaling StruClus (CV < 0.02)
1
2
4
Cores
8
8+HT
Fig. 2. Parallel scaling: synthetic dataset of size 10 000
The evaluation results wrt the real world datasets are summarized in Table III. We were unable to cluster the AnchorQuery and the ChemDB datasets with PROCLUS, Kernel KMeans and SCAP (even for high values of parameter (b)) in
less than 2 days. For the chemical reaction datasets, StruClus
also needs more time compared to a synthetic dataset of similar
size. This runtime increase can be partially explained by the
larger graph sizes. However, other hidden parameters of the
datasets, such as the size of maximal common substructures
have an influence on the runtime as well. StruClus had
some runtime outliers on the Heterocyclic dataset. They were
caused by small temporary clusters with large (> 60 vertices)
representatives. The high runtime of the SCAP algorithm on
the AnchorQuery dataset is a bit surprising, as the common
substructures processed by SCAP are limited in their size
(maximum 8 vertices). We consider a larger frequent pattern
search space to be the reason for this runtime increase. Kernel
K-Means was surprisingly slow on the Heterocyclic dataset
and took more than 24 hours for a single run. We have
therefore created a random subset with a size of 5000 graphs
for it.
StruClus always outperforms the competitors wrt the quality
scores. For the AnchorQuery dataset StruClus created 26
clusters on average. The high score for the Purity measure
shows, that StruClus splitted some of the real clusters, but
keeps a well inter cluster separation. ChemDB was clustered
by StruClus in ≈ 19 hours with ≈ 117 clusters. As a
consequence of the high runtime of the ChemDB measurement
and the lack of competitors, we repeated the test only 3 times.
The aCov value for the final clustering was 0.49 on average.
This highlights the ability of StruClus to cluster large-scale
real-world datasets.
VI. C ONCLUSION
In this paper we have presented a new structural clustering
algorithm for large scale datasets of small labeled graphs. With
the help of explicitly selected cluster representatives, we were
able to achieve a linear runtime in the worst case wrt the
dataset size. A novel support counting sampling strategy with
multiple hypothesis testing correction was able to accelerate
the algorithm significantly without lowering the clustering
quality. We have furthermore shown, that cluster homogeneity
can be balanced with a dynamic minimum support strategy
for representative mining. A cluster merging and splitting
step was introduced to achieve a well separated clustering
even in the high dimensional pattern space. Our experimental
evaluation has shown that our new approach outperforms the
competitors wrt clustering quality, while attaining significantly
lower runtimes for large scale datasets. Although we have
shown that StruClus greatly improves the clustering performance compared to its competitors, de-novo datasets with
several billion of molecules are still outside the scope of
this work. For this reason, we consider the development of a
distributed variant of the algorithm to be future work. Another
consideration to further improve the quality of the algorithm
is to integrate a discriminative frequent subgraph miner for
representative mining. The integration of the discriminative
property into the mining process has the advantage, that higher
TABLE II
R ESULTS FOR THE SYNTHETIC DATASETS . RUNTIMES ARE GIVEN IN HOURS . B EST RESULTS ARE MARKED BOLD . CV IS THE COEFFICIENT OF
VARIATION . T HE CV VALUE IS GIVEN AS THE MAXIMUM FOR EACH COLUMN . A LL OTHER VALUES ARE AVERAGED . Q UALITY MEASURES ARE :
N ORMALIZED VARIATION OF I NFORMATION (NVI), F OWLKES -M ALLOWS I NDEX (FW), AND P URITY.
Size
Runtime
CV < 0.08
1 000
5 000
10 000
50 000
100 000
500 000
1 000 000
0.05
–
0.19
0.33
0.47
1.35
2.73
StruClus
NVI
FM
Purity
StruClus (Exact Support)
Runtime NVI
FM
Runtime
SCAP
Purity
Runtime
PROCLUS
NVI
FM
Kernel K-Means
Runtime
NVI
FM
< 0.03
< 0.07
< 0.03
< 0.07
< 0.02
< 0.04
< 0.06
< 0.02
< 0.32
< 0.11
< 0.22
< 0.01
< 0.01
< 0.01
0.90
–
0.95
0.94
0.93
0.93
0.91
0.75
–
0.87
0.87
0.86
0.86
0.84
0.90
–
0.99
0.99
0.99
0.99
0.98
0.05
–
0.59
2.69
–
–
–
0.90
–
0.93
0.92
–
–
–
0.77
–
0.85
0.85
–
–
–
7.5 × 10−4
–
0.03
0.38
1.21
18.15
–
0.83
–
0.83
0.83
0.86
0.83
–
0.13
4.99
11.91
–
–
–
–
0.58
0.50
0.49
–
–
–
–
0.26
0.24
0.24
–
–
–
–
0.02
2.87
10.32
–
–
–
–
0.77
0.84
0.86
–
–
–
–
0.54
0.67
0.78
–
–
–
–
TABLE III
R ESULTS FOR THE REAL WORLD DATASETS . RUNTIMES ARE GIVEN IN HOURS . B EST RESULTS ARE MARKED BOLD . CV IS THE COEFFICIENT OF
VARIATION . T HE CV VALUE IS GIVEN AS THE MAXIMUM FOR EACH COLUMN . A LL OTHER VALUES ARE AVERAGED . T HE C HEM DB MEASUREMENT WAS
REPEATED ONLY 3 TIMES . W E HAVE NOT CALCULATED THE — OTHERWISE MEANINGLESS — CV FOR IT. R ESULTS FOR K ERNEL K-M EANS ARE ARE
GIVEN FOR A RANDOM SUBSET OF THE H ETEROCYLCIC DATASET. Q UALITY MEASURES ARE : N ORMALIZED VARIATION OF I NFORMATION (NVI),
F OWLKES -M ALLOWS I NDEX (FW), AND P URITY.
dataset
Runtime
CV < 2.91
AnchorQuery
Heterocyclic
ChemDB
2.47
1.07
≈ 19
StruClus
NVI
FM
Purity
SCAP
Runtime Purity
Runtime
< 0.09
< 0.12
< 0.08
< 0.02
< 0.01
< 0.03
< 0.05
< 0.08
< 0.01
< 0.01
< 0.01
0.44
0.46
–
0.63
0.53
–
0.89
0.66
–
–
0.01
–
–
0.58
–
–
0.01
–
–
0.29
–
–
0.29
–
–
3 .03 (Subset)
–
–
0.27
–
–
0 .29
–
quality representative candidates are mined. This will result in
a lower number of necessary candidate patterns, which has a
positive effect on the runtime. Furthermore, it allows to mine
highly discriminant, non-maximal subgraphs. However, it is
non-trivial to extend the support counting sampling strategy
to such miners. Additionally, discriminative scores are usually
non-monotonic on the subgraph lattice [34, 35], which imposes
another runtime burden.
ACKNOWLEDGMENT
This work was supported by the German Research Foundation (DFG), priority programme Algorithms for Big Data
(SPP 1736). We would like to thank Nils Kriege for providing
a fast subgraph isomorphism and Madeleine Seeland, Andreas
Karwath, and Stefan Kramer for providing their SCAP implementation.
R EFERENCES
[1]
[2]
[3]
C. Kalinski, M. Umkehrer, L. Weber, J. Kolb, C. Burdack, and G. Ross,
“On the industrial applications of mcrs: molecular diversity in drug
discovery and generic drug synthesis,” English, Molecular Diversity,
vol. 14, no. 3, pp. 513–522, 2010.
E. Chávez and G. Navarro, “A probabilistic spell for the curse
of dimensionality,” in Algorithm Engineering and Experimentation,
Third International Workshop, ALENEX 2001, Washington, DC, USA,
January 5-6, 2001, Revised Papers, A. L. Buchsbaum and J. Snoeyink,
Eds., ser. Lecture Notes in Computer Science, vol. 2153, Springer,
2001, pp. 147–160.
A. Gupta, R. Krauthgamer, and J. R. Lee, “Bounded geometries,
fractals, and low-distortion embeddings,” in Foundations of Computer
Science, 2003. Proceedings. 44th Annual IEEE Symposium on, IEEE,
2003, pp. 534–543.
[4]
[5]
[6]
[7]
[8]
[9]
[10]
[11]
[12]
[13]
[14]
PROCLUS
NVI
FM
Runtime
Kernel K-Means
NVI
FM
K. S. Beyer, J. Goldstein, R. Ramakrishnan, and U. Shaft, “When is
”nearest neighbor” meaningful?,” in Proceedings of the 7th International Conference on Database Theory, ser. ICDT ’99, London, UK,
UK: Springer-Verlag, 1999, pp. 217–235.
M. Ackerman and S. Ben-David, “Clusterability: A theoretical study,”
in Proceedings of the Twelfth International Conference on Artificial
Intelligence and Statistics, AISTATS 2009, Clearwater Beach, Florida,
USA, April 16-18, 2009, D. A. V. Dyk and M. Welling, Eds., ser. JMLR
Proceedings, vol. 5, JMLR.org, 2009, pp. 1–8.
N. Wale, I. A. Watson, and G. Karypis, “Comparison of descriptor
spaces for chemical compound retrieval and classification,” Knowl. Inf.
Syst., vol. 14, no. 3, pp. 347–375, 2008.
N. Shervashidze, S. V. N. Vishwanathan, T. Petri, K. Mehlhorn, and K.
M. Borgwardt, “Efficient graphlet kernels for large graph comparison,”
in Proceedings of the Twelfth International Conference on Artificial
Intelligence and Statistics, AISTATS 2009, Clearwater Beach, Florida,
USA, April 16-18, 2009, D. A. V. Dyk and M. Welling, Eds., ser. JMLR
Proceedings, vol. 5, JMLR.org, 2009, pp. 488–495.
S. V. N. Vishwanathan, N. N. Schraudolph, R. I. Kondor, and K. M.
Borgwardt, “Graph kernels,” Journal of Machine Learning Research,
vol. 11, pp. 1201–1242, 2010.
P. Foggia, G. Percannella, and M. Vento, “Graph matching and learning
in pattern recognition in the last 10 years,” IJPRAI, vol. 28, no. 1, 2014.
K. Tsuda and T. Kudo, “Clustering graphs by weighted substructure
mining,” in Proceedings of the 23rd international conference on
Machine learning, ACM, 2006, pp. 953–960.
N. Kriege, P. Mutzel, and T. Schäfer, “Practical sahn clustering for
very large data sets and expensive distance metrics,” Journal of Graph
Algorithms and Applications, vol. 18, no. 4, pp. 577–602, 2014.
H. Bunke and K. Shearer, “A graph distance metric based on the
maximal common subgraph,” Pattern Recognition Letters, vol. 19, no.
3-4, pp. 255–259, 1998.
W. D. Wallis, P. Shoubridge, M. Kraetzl, and D. Ray, “Graph distances
using graph union,” Pattern Recognition Letters, vol. 22, no. 6/7,
pp. 701–704, 2001.
M.-L. Fernández and G. Valiente, “A graph distance metric combining
maximum common subgraph and minimum common supergraph,”
Pattern Recognition Letters, vol. 22, no. 6–7, pp. 753 –758, 2001.
[15]
[16]
[17]
[18]
[19]
[20]
[21]
[22]
[23]
[24]
[25]
[26]
[27]
[28]
[29]
[30]
[31]
[32]
[33]
[34]
H. Bunke, “On a relation between graph edit distance and maximum common subgraph,” Pattern Recognition Letters, vol. 18, no. 8,
pp. 689–694, 1997.
M. Seeland, A. Karwath, and S. Kramer, “Structural clustering of
millions of molecular graphs,” in Symposium on Applied Computing,
SAC 2014, Gyeongju, Republic of Korea - March 24 - 28, 2014,
Y. Cho, S. Y. Shin, S. Kim, C. Hung, and J. Hong, Eds., ACM, 2014,
pp. 121–128.
C. C. Aggarwal, C. M. Procopiuc, J. L. Wolf, P. S. Yu, and J. S.
Park, “Fast algorithms for projected clustering,” in SIGMOD 1999,
Proceedings ACM SIGMOD International Conference on Management
of Data, June 1-3, 1999, Philadelphia, Pennsylvania, USA., A. Delis, C.
Faloutsos, and S. Ghandeharizadeh, Eds., ACM Press, 1999, pp. 61–72.
M. A. Girolami, “Mercer kernel-based clustering in feature space,”
IEEE Trans. Neural Networks, vol. 13, no. 3, pp. 780–784, 2002.
K. Tsuda and K. Kurihara, “Graph mining with variational dirichlet
process mixture models,” in Proceedings of the SIAM International
Conference on Data Mining, SDM 2008, April 24-26, 2008, Atlanta,
Georgia, USA, SIAM, 2008, pp. 432–442.
M. Ferrer, E. Valveny, F. Serratosa, I. Bardajı́, and H. Bunke, “Graphbased k-means clustering: A comparison of the set median versus
the generalized median graph,” in Computer Analysis of Images
and Patterns, 13th International Conference, CAIP 2009, Münster,
Germany, September 2-4, 2009. Proceedings, X. Jiang and N. Petkov,
Eds., ser. Lecture Notes in Computer Science, vol. 5702, Springer,
2009, pp. 342–350.
M. Seeland, S. A. Berger, A. Stamatakis, and S. Kramer, “Parallel
structural graph clustering,” in Machine Learning and Knowledge
Discovery in Databases - European Conference, ECML PKDD 2011,
Athens, Greece, September 5-9, 2011, Proceedings, Part III, 2011,
pp. 256–272.
C. C. Aggarwal, N. Ta, J. Wang, J. Feng, and M. J. Zaki, “Xproj: a
framework for projected structural clustering of xml documents,” in
Proceedings of the 13th ACM SIGKDD International Conference on
Knowledge Discovery and Data Mining, San Jose, California, USA,
August 12-15, 2007, P. Berkhin, R. Caruana, and X. Wu, Eds., ACM,
2007, pp. 46–55.
M. L. Yiu and N. Mamoulis, “Frequent-pattern based iterative projected clustering,” in Data Mining, 2003. ICDM 2003. Third IEEE
International Conference on, 2003, pp. 689–692.
E. Müller, S. Günnemann, I. Assent, and T. Seidl, “Evaluating clustering in subspace projections of high dimensional data,” in Proc. 35th
International Conference on Very Large Data Bases (VLDB 2009),
Lyon, France, PVLDB Journal, Vol. 2, No. 1,, VLDB Endowment,
2009, pp. 1270–1281.
A. Patrikainen and M. Meila, “Comparing subspace clusterings,” IEEE
Trans. Knowl. Data Eng., vol. 18, no. 7, pp. 902–916, 2006.
M. A. Hasan, V. Chaoji, S. Salem, J. Besson, and M. J. Zaki,
“ORIGAMI: mining representative orthogonal graph patterns,” in Proceedings of the 7th IEEE International Conference on Data Mining
(ICDM 2007), October 28-31, 2007, Omaha, Nebraska, USA, IEEE
Computer Society, 2007, pp. 153–162.
K. Klein, N. Kriege, and P. Mutzel, “CT-Index: fingerprint-based graph
indexing combining cycles and trees,” in Data Engineering (ICDE),
2011 IEEE 27th International Conference on, 2011, pp. 1115–1126.
M. R. Garey and D. S. Johnson, Computers and Intractability: A Guide
to the Theory of NP-Completeness. W. H. Freeman, 1979.
M. Meilă, “Comparing clusterings—an information based distance,”
Journal of multivariate analysis, vol. 98, no. 5, pp. 873–895, 2007.
E. B. Fowlkes and C. L. Mallows, “A method for comparing two hierarchical clusterings,” Journal of the American Statistical Association,
vol. 78, no. 383, pp. 553–569, 1983.
X. Hui and L. Zhongmon, “Clustering validation measures,” in Data
clustering: algorithms and applications, C. C. Aggarwal and C. K.
Reddy, Eds., New York: CRC Press, 2013, ch. 23, pp. 571–605.
J. H. Chen, E. Linstead, S. J. Swamidass, D. Wang, and P. Baldi,
“Chemdb update—full-text search and virtual chemical space,” Bioinformatics, vol. 23, no. 17, pp. 2348–2351, 2007.
N. Shervashidze, P. Schweitzer, E. J. van Leeuwen, K. Mehlhorn,
and K. M. Borgwardt, “Weisfeiler-lehman graph kernels,” Journal of
Machine Learning Research, vol. 12, pp. 2539–2561, 2011.
X. Yan, H. Cheng, J. Han, and P. S. Yu, “Mining significant graph
patterns by leap search,” in Proceedings of the ACM SIGMOD
International Conference on Management of Data, SIGMOD 2008,
[35]
[36]
Vancouver, BC, Canada, June 10-12, 2008, J. T. Wang, Ed., ACM,
2008, pp. 433–444.
M. Thoma, H. Cheng, A. Gretton, J. Han, H. Kriegel, A. J. Smola, L.
Song, P. S. Yu, X. Yan, and K. M. Borgwardt, “Discriminative frequent
subgraph mining with optimality guarantees,” Statistical Analysis and
Data Mining, vol. 3, no. 5, pp. 302–318, 2010.
D. A. V. Dyk and M. Welling, Eds., Proceedings of the Twelfth International Conference on Artificial Intelligence and Statistics, AISTATS
2009, Clearwater Beach, Florida, USA, April 16-18, 2009, vol. 5, ser.
JMLR Proceedings, JMLR.org, 2009.
| 8cs.DS
|
On Sound Compilation of Reals
Eva Darulova
Viktor Kuncak
EPFL
[email protected]
EPFL
[email protected]
arXiv:1309.2511v1 [cs.PL] 10 Sep 2013
Abstract
In this paper we advocate a natural but ambitious alternative:
source code programs should be expressed in terms of ideal, mathematical real numbers. In our system, the programmer writes a program using a Real data type and states the desired postconditions,
then specifies uncertainties explicitly in pre- and postconditions as
well as the desired target precision on the return values. It is then
up to our trustworthy compiler to check, taking into account all uncertainties and their propagation, that the desired precision can be
soundly realized in a finite precision implemenation. If so, the compiler chooses and emits one such finite-precision implementation.
Following this model, we derive verification conditions that encode reasoning about floating-point roundoff errors into reasoning about real numbers. Our verification conditions explicitly separate the ideal program without external uncertainties and roundoffs from the actual program, which is actually executed in finite
precision with possibly noisy inputs. Additionally, our constraints
are fully parametric in the floating-point precision and can thus be
used with different floating-point hardware configurations, as well
as used for determining the minimum precision needed.
To summarize, the view of source code as functions over real
numbers has several advantages:
Writing accurate numerical software is hard because of many
sources of unavoidable uncertainties, including finite numerical
precision of implementations. We present a programming model
where the user writes a program in a real-valued implementation
and specification language that explicitly includes different types
of uncertainties. We then present a compilation algorithm that generates a conventional implementation that is guaranteed to meet
the desired precision with respect to real numbers. Our verification
step generates verification conditions that treat different uncertainties in a unified way and encode reasoning about floating-point
roundoff errors into reasoning about real numbers. Such verification conditions can be used as a standardized format for verifying
the precision and the correctness of numerical programs. Due to
their often non-linear nature, precise reasoning about such verification conditions remains difficult. We show that current state-of-the
art SMT solvers do not scale well to solving such verification conditions. We propose a new procedure that combines exact SMT
solving over reals with approximate and sound affine and interval
arithmetic. We show that this approach overcomes scalability limitations of SMT solvers while providing improved precision over
affine and interval arithmetic. Using our initial implementation we
show the usefullness and effectiveness of our approach on several
examples, including those containing non-linear computation.
1.
• Programmers can, for the most part, reason in real arithmetic
and not floating-point arithmetic. We thus achieve separation of
the design of algorithms (which may still be approximate) from
their realization using finite precision computations.
• We can verify the ideal meaning of programs using techniques
Introduction
developed to reason over real numbers, which are more scalable
than techniques that directly deal with floating point arithmetic.
Writing numerical programs is difficult, in part because the programmer needs to deal not only with the correctness of the algorithm but also with unavoidable uncertainties. Program inputs
may not be exactly known because they come from physical experiments or were measured by an embedded sensor. The computation itself suffers from roundoff errors at each step, because of the
use of finite-precision arithmetic. In addition, resources like energy
may be scarce so that only a certain number of bits are available
for the numerical datatype. At the same time, the computed results
can have far-reaching consequences if used to control, for example,
a vehicle or a nuclear power plant. It is therefore of great importance to improve confidence in numerical code [34]. One of the first
challenges in doing this is that most of our automated reasoning
tools work with real arithmetic, whereas the code is implemented in
finite-precision (typically, floating point) arithmetic. Many current
approaches to verifying numerical programs start with the floatingpoint implementation and then try to verify the absense of (runtime)
errors. However, the absence of errors by itself does not guarantee
that program behavior matches the desired specification expressed
using real numbers. Fundamentally, the source code semantics is
expressed in terms of data types such as floating points. This is
further problematic for compiler optimizations, because, e.g., the
associativity law is unsound with respect to such source code semantics.
• The approach allows us to quantify the deviation of implemen-
tation outputs from ideal ones, instead of merely proving e.g.
range bounds of floating-point variables which is used in simpler static analyses.
• The compiler for reals is free to do optimizations as long as they
preserve the precision requirements. This allows the compiler
to apply, for example, associativity of arithmetic, or even select
different approximation schemes for transcendental functions.
• In addition to roundoff errors, the approach also allows the
developer to quantify program behavior in the face of external
uncertainties such as input measurement errors.
Using our verification condition generation approach, the correctness and the precision of compilation for small programs can
be directly verified using an SMT solver such as Z3 [16] (see Section 5.3); this capability will likely continue to improve as the
solvers advance. However, the complexity of the generated verification conditions for larger programs is still out of reach of such
solvers and we believe that specialized techniques are and will continue to be necessary for this task. This paper presents two specialized techniques that improve the feasibility of the verification task.
The first technique performs local approximation and is effective
1
2013/9/11
different selection of numerical algorithms. To support programming of larger code fragments our system also supports a modular verification technique, which handles functions through inlining
function bodies or using their postconditions. We thus expect that
our technique is applicable to larger code bases as well, possibly
through refactoring code into multiple smaller and annotated functions. Even on the benchmarks that we release, we are aware of no
other available system that would provide the same guarantees with
our level of automation.
even in benchmarks containing nonlinear arithmetic; the second
technique specifically handles conditional expressions.
1.1
Solving Non-Linear Constraints
Nonlinear arithmetic poses a significant challenge for verification
because it cannot directly be handled using Simplex-like algorithms
embedded inside SMT solvers. Although interesting relevant fragments are decidable and are supported by solvers, their complexity is much higher. Unfortunately, non-linear arithmetic is ubiquitous in numerical software. Furthermore, our verification conditions add roundoff errors to arithmetic expressions, so the resulting
constraints grow further in complexity, often becoming out of reach
of solvers. An alternative to encoding into SMT solver input is to
use a sound and overapproximating arithmetic model such as interval or affine arithmetic [15]. However, when used by itself on
non-linear code, these approaches yield too pessimistic results to
be useful.
We show that we can combine range arithmetic computation
with SMT solving to overcome the limitations of each of the individual techniques when applied to non-linear arithmetic. We obtain
a sound, precise, and somewhat more scalable procedure. During
range computation, our technique also checks for common problems such as overflow, division by zero or square root of a negative
number, emitting the corresponding warnings. Additionally, because the procedure is a forward computation, it is suitable for automatically generating function summaries containing output ranges
and errors of a function. From the point of view of the logical
encoding of the problem, range arithmetic becomes a specialized
method to perform approximate quantifier elimination of bounded
variables that describe the uncertainty.
1.2
1.4
• We present a real-valued implementation and specification lan-
guage for numerical programs with uncertainties; we define its
semantics in terms of verification constraints that they induce.
We believe that such verification conditions can be used as a
standardized format for verifying the precision and the correctness of numerical programs.
• We develop an approximation procedure for computing precise
range and error bounds for nonlinear expressions which combines SMT solving with interval arithmetic. We show that such
an approach significantly improves computed range and error
bounds compared to standard interval arithmetic, and scales
better than SMT solving alone. Our procedure can also be used
independently as a more precise alternative to interval arithmetic, and thus can perform forward computation without having the desired postconditions.
• We describe an approach for soundly computing error bounds
in the presence of branches and uncertainties, which ensures
soundness of compilation in case the function defined by a
program is not known to be continuous.
Sound Compilation of Conditional Branches
In the presence of uncertainties, conditional braches become another verification challenge. The challenge is that the ideal execution may follow one branch, but, because of input or roundoff errors, the actual execution follows another. This behaviour may be
acceptable, however, if we can show that the error on the output
remains within required bounds. Therefore, our approach would
directly benefit from automated analysis of continuity, which was
advocated previously [48]. For such continuous functions, our analysis can be done separately for each program path. In the absence
of knowledge of continuity, we present a new method to check that
different paths taken by real-valued and floating point versions of
the program still preserves the desired precision specification. Our
check does not require continuity (which becomes difficult to prove
for non-linear code). Instead, it directly checks that the difference
between the two values on different branches meets the required
precision. This technique extends our method for handling nonlinear arithmetic, so it benefits from the combination of range arithmetic and SMT solving.
1.3
Summary of Contributions
Our overall contribution is an approach for sound compilation of
real numbers. Specifically:
• We have implemented our framework and report our experience
on a set of diverse benchmarks, including benchmarks from
physics, biology, chemistry, and control systems. The results
show that our technique is effective and that it achieves a synergy of the techniques on which it relies.
2.
Example
We demonstrate some aspects of our system on the example in Figure 1. The methods triangle and triangleSorted compute the area of
a triangle with side lengths a, b and c. The notation a.in(1.0, 9.0)
is short for 1.0 < a && a < 9.0. We consider a particular application where the user may have two side lengths given, and may
vary the third. She has two functions available to do the computation and wants to determine whether either or both satisfy the
precision requirement of 1e-11 on line 9. Our tool determines that
such requirement needs at least double floating point precision; the
challenge then is to establish that this precision is sufficient to ensure these bounds, given that errors in floating code accumulate and
grow without an a priori bound.
Our tool verifies fully automatically that the method triangleSorted
indeed satisfies the postcondition and generates the source code
with the Double datatype which also includes a more precise and
complete postcondition:
Implementation and Evaluation
We have implemented our compilation and verification procedure,
including the verification condition generation, analysis of possibly non-linear expressions, and the handling of conditionals. Our
system is implemented as an extension of a verifier for functional
Scala programs. The implementation relies on a range arithmetic
implementation for Scala as well as on the Z3 SMT solver.
We have evaluated the system on a number of diverse benchmarks, obtaining promising results. We are releasing our benchmarks (and making them available as supplementary material for
reviewers). We hope that the benchmarks can be used to compare future tools for error quantification, help the development of
nonlinear solvers, and also present challenges for more aggressive
compilation schemes, with different number representations and
0.01955760939159717 ≤ res ∧ res ≤ 12.519984025578283 ∧
res +/− 8.578997409317759e−12
To achieve this result, our tool first checks that the precondition
of the function call is satisfied using the Z3 solver. Then, it inlines
the body of the function triangleSorted and computes a sound bound
on the result’s uncertainty with our approximation procedure and
uses it to show that the postcondition is satisfied. The error compu-
2
2013/9/11
2
4
6
8
and the user annotates them with pre- and postconditions that explicitly talk about uncertainties. Real represents ideal real numbers without any uncertainty. We allow arithmetic expressions over
√
Reals with the standard arithmetic operators {+, −, ∗, /, }, and
together with conditionals and function calls they form the body
of methods. Our tool also supports immutable variable declarations
as val x = .... This language allows the user to define a computation
over real numbers. Note that this specification language is not executable.
The precondition allows the user to provide a specification of
the environment. A complete environment specification consists of
lower and upper bounds for all method parameters and an upper
bound on the uncertainty or noise. Range bounds are expressed
with regular comparison operators. Uncertainty is expressed with
the predicate such as x +/− 1e−6, which denotes that the variable
x is only known up to 1e − 6. Alternatively, the programmer can
specify the relative error as x +/− 1e−7 ∗ x. If no noise except for
roundoff is present, roundoff errors are automatically added to
input variables.
The postcondition can specify the range and the maximum uncertainty accepted on the output. In addition to the language allowed in the precondition, the postcondition may reference the errors on inputs directly in the following way: res +/− 3.5 ∗ !x, which
says that the maximum acceptable error on the output variable res
is bounded from above by 3.5 times the initial error on x. Whereas
the precondition may only talk about the ideal values, the postcondition can also reference the actual value directly via ∼x. This allows us to assert that runtime values will not exceed a certain range,
for instance.
def mainFunction(a: Real, b: Real, c: Real): Real = {
require(4.500005 <= a && a <= 6.5)
val b = 4.0
val c = 8.5
val area = triangleSorted(a, b, c)
//val area = triangleUnstable(a, b, c)
area
} ensuring(res => res +/− 1e−11)
10
12
def triangle(a: Real, b: Real, c: Real): Real = {
require(a.in(1.0, 9.0) && b.in(1.0, 9.0) && c.in(1.0, 9.0) &&
a + b > c + 1e−6 && a + c > b + 1e−6 && b + c > a + 1e−6)
14
val s = (a + b + c)/2.0
sqrt(s ∗ (s − a) ∗ (s − b) ∗ (s − c))
16
}
18
20
22
def triangleSorted(a: Real, b: Real, c: Real): Real = {
require(a.in(1.0, 9.0) && b.in(1.0, 9.0) && c.in(1.0, 9.0) &&
a + b > c + 1e−6 && a + c > b + 1e−6 && b + c > a + 1e−6 &&
a < c && b < c)
if (a < b) {
sqrt((c+(b+a)) ∗ (a−(c−b)) ∗ (a+(c−b)) ∗ (c+(b−a)))/4.0
} else {
sqrt((c+(a+b)) ∗ (b−(c−a)) ∗ (b+(c−a)) ∗ (c+(a−b))) / 4.0
}
24
26
28
}
Figure 1. Computing the area of a triangle.
Floating-point arithmetic Our tool and technique support in the
generated target code any floating-point precision and in particular, single and double floating-point precision as defined by the
IEEE 754 floating-point standard [46]. It is also straightforward to
combine it with techniques to generate fixed-point arithmetic [13],
which similarly relies on knowing ranges of variables. We assume
rounding-to-nearest rounding mode and that basic arithmetic oper√
ations {+, −, ∗, /, } are rounded correctly, which means that the
result from any such operation must be the closest representable
floating-point number. Hence, provided there is no overflow, the
result of a binary operation ◦F satisfies
tation takes into account in a sound way the input uncertainty (here
an initial roundoff error on the inputs), its propagation and roundoff
errors commited at each arithmetic operation. Additionally, due to
the initial roundoff error the comparison on line 24 is not exact, so
that some floating-point computations will take a different branch
than their corresponding real-valued computation. More precisely,
the total error when computing the condition is 7.22e − 16, as
computed by our tool. That is, floating-point values that satisfy
a < b + 7.22e − 16 may take the else branch, even though the
corresponding real values would follow the then branch, and similarly in the opposite direction. Our tool verifies that the difference
in the computed result in two branches, due to this divergence between real arithmetic and floating point arithmetic code, remains
within the precision requirement.
Finally, our tool uses our novel range computation procedure
to also compute a more precise output range than we could have
obtained in interval arithmetic. It then includes this more complete postcondition in the generated floating-point code. In fact, interval arithmetic alone computes the ranges [0.0138, 16.163] and
[0.0169, 14.457] for using the methods triangle and triangleSorted
respectively, seemingly suggesting that two methods perform entirely different computations. With our technique, the tool computes the same range [0.0195, 12.52] for both methods, but shows a
difference in the absolute error of the computation. For the method
triangle, the verification fails, as desired, because the computed error (2.3e − 11) exceeds the required precision bound. This result
is expected—the textbook formula for triangles is known to suffer
from imprecision for flat triangles [31], which is somewhat rectified
in the method triangleSorted.
3.
x ◦F y = (x ◦R y)(1 + δ), |δ| ≤ M , ◦ ∈ {+, −, ∗, /}
(1)
where ◦R is the ideal operation in real numbers and M is the
machine epsilon that determines the upper bound on the relative
error. This model provides a basis for our roundoff error estimates.
4.
Compiling Reals to Finite Precision
Given a specification or program over reals and a possible target
floating-point precision, our tool generates code over floating-point
numbers that satisfy the given pre- and postconditions. Figure 2
presents a high-level view of our compilation algorithm. Our tool
first analyses the entire specification and generates one verification
condition for each postcondition to be proven. To obtain a modular
algorithm, the tool also generates verification conditions that check
that at each function call the precondition of the called function
is satisfied. The methods are then sorted by occuring function
calls. This allows us to re-use already computed postconditions
of function calls in a modular analysis. If the user specified a
target precision, the remaining part of the compilation process is
perfomed with respect to this precision, if not or in the case the
user specified several possible precisions, our tool will iteratively
select the next more precise precision and attempt to compile the
entire program. If compilation fails due to unsatisfied assertions,
the next precision is selected. Thus, one task of our algorithm is to
Programs with Reals
Each program to be compiled consists of one top-level object with
methods written in a functional subset of the Scala programming
language [41]. All methods are functions over the Real datatype
3
2013/9/11
x ∈ [a, b]
x◦ = x + errx ∧ errx ∈ [−k, k]
x +/− m ∗ x
x◦ = x + errx ∧ errx ∈ [−|mx|, |mx|]
x◦
∼x
!x
errx
xy
(x y) ∧ (x◦ y◦ )(1 + δ1 )
sqrt(x)
sqrt(x) ∧ sqrt(x◦ )(1 + δ2 )
z = x ∧ z◦ = x◦
val z = x
if (c(x)) e1(x)
((c(x) ∧ e1(x)) ∨ (¬c(x) ∧ e2(x)))∧
else e2(x)
((c◦ (x◦ ) ∧ e1◦ (x◦ )) ∨ (¬c◦ (x◦ ) ∧ e2◦ (x◦ )))
g(x)
g(x) ∧ g◦ (x◦ )
∈ {+, −, ∗, /}
−m ≤ δi ∧ δi ≤ m , all δ are fresh
cond◦ and e◦ denote functions with roundoff errors at each step
a <= x && x <= b
Input: spec: specification over Reals, prec: candidate precisions
for fnc ← spec.fncs
fnc.vcs = generateVCs(fnc)
x +/− k
spec.fncs.sortBy((f1, f2) => f1 ⊆ f2.fncCalls)
while prec 6= ∅ and notProven(spec.fncs)
precision = prec.nextPrecise
for fnc ← spec.fncs
for vc ← fnc.vcs
while vc.hasNextApproximation ∧ notProven(vc)
approx = getNextApproximation(vc, precision)
vc.status = checkWithZ3(approx)
generateSpec(fnc)
generateCode(spec)
Output: floating−point code with generated postconditions
Figure 2. Compilation algorithm.
Table 1. Semantics of our specification language.
automatically determine the necessary least floating-point precision
to ensure the specification is met. Currently, each precision iteration
is computed separately, which is not a big issue due to a small
number alternative floating-point targets. We did identify certain
shared computations between iterations; we can exploit them in the
future for more efficient compilation. In order for the compilation
process to succeed, the specification has to be met with respect
to a given floating-point precision, thus the principal part of our
algorithm is spent in verification, which we describe in Section 5.
We envision that in the future the compilation task will also
include automatic precision-preserving code optimizations, but in
this paper we concentrate on the challenging groundwork of verifying the precision of code.
Our tool currently generates Scala code over single, double, double-double and quad-double floating-point arithmetic. For
double-double and quad-double precision, which were implemented in software by [3], we provide a Scala interface with the
generated code. In case the verification part of compilation fails,
we nonetheless generate code (together with a failure report) with
the best postconditions our tool was able to compute. The user can
then use the generated specifications to gain insight why and where
her program does not satify requirements.
Our techniques are not restricted to these floating-point precisions, however, and will work for any precision that follows the
abstraction given in Equation (1). Furthermore, while we have implemented our tool to accept specifications in a domain specific language embedded in Scala and generate code in Scala, all our techniques apply equally to all programming languages and hardware
that follow the floating-point abstraction we assume (Equation 1).
5.
where ~
x, res,
~ ~
y denote the input, output and local variables respectively.
Table 1 summarizes how verification constraints are generated
from our specification language. Each variable x in the specification
corresponds to two real-valued variables x, x◦ , the ideal one in the
absence of uncertainties and roundoff errors and the actual one,
computed by the compiled program. Note that the ideal and actual
variables are related only through the error bounds in the pre- and
postconditions, which allows for the ideal and actual executions to
take different paths.
In the method body we have to take into account roundoff
errors from arithmetic operations and the propagation of existing
√
errors. Our system currently supports operations {+, −, ∗, /, },
but these can be in principle extended to elementary functions, for
instance by encoding them via Taylor expansions [37].
Note that the resulting verification conditions are parametric in
the machine epsilon.
5.2
In order to give feedback to developers and to facilitate automatic
modular analysis, our tool also provides automatic specification
generation. By this we mean that the programmer still needs to
provide the environment specification in form of preconditions, but
our tool automatically computes a precise postcondition.
Formally, we can rewrite the constraint (*) as
∀~
x, res. (∃~
y . P (~
x) ∧ body(~
x, ~
y , res)) → Q(~
x, res)
where Q is now unknown. We obtain the most precise postcondition Q by applying quantifier elimination (QE) to P (~
x) ∧
body(~
x, ~
y , res) and eliminate ~
y . The theory of arithmetic over reals admits QE so it is theoretically possible to use this approach.
We do not currently use a full QE procedure for specification generation, as it is expensive and it is not clear whether the returned
expressions would be of a suitable format. Instead, we use our approximation approach which computes ranges and maximum errors
in a forward fashion and allows us to compute an (over) approximation of a postcondition of the form res ∈ [a, b] ∧ res ± u.
Verifying Real Programs
We will now describe the verification part of our compilation algorithm. In the following we will call the ideal computation the
computation in the absence of any uncertainties and implemented
in a real arithmetic, and the actual computation the computation
that will finally be executed in finite-precision and with potentially
uncertain inputs.
5.1
Verification Conditions for Loop-Free Programs
5.3
Each method with its precondition P and postcondition Q implies
the following verification condition:
∀~x, res,
~ ~
y . P (~
x) ∧ body(~
x, ~
y , res)
~ → Q(~
x, res)
~
Specification Generation
Difficulty of Simple Encoding into SMT solvers
For small functions we can already prove interesting properties by
using the exact encoding of the problem just described and discharging the verification constraints with Z3. Consider the follow-
(*)
4
2013/9/11
approximation pipeline. We support inlining of both user-provided
postconditions and postconditions computed by our own specification generation procedure. If this still is not precise enough, we
inline the entire function body.
Postcondition inlining is implemented by replacing the function
call with a fresh variable and constraining it with the postcondition. Thus, if verification suceeds with inlining the postcondition,
we avoid having to consider each path of the inlined function separately and can perform modular verification avoiding a potential
path explosion problem. Such modular verification is not feasible
when postconditions are too imprecise and we plan to explore the
generation of more precise postconditions in the future. One step
in this direction is to allow postconditions that are parametric in
the initial errors, for example with the operator !x introduced in
Section 3. While our tool currently supports postcondition inlining
with such postconditions, we do not yet generate these automatically.
def getNextApproximation(vc, precision):
getPathError!
inlining postcondition
inlining full function
functions
getRange!
evalWithError!
path-wise
approx. error
merging
approx. error & range
paths
arithmetic
first approximation
Figure 3. Approximation pipeline.
ing code a programmer may write to implement the third B-spline
basic function which is commonly used in signal processing [29].
def bspline3(u: Real): Real = {
require(0 ≤ u && u ≤ 1 && u ± 1e−13)
−u∗u∗u / 6.0
} ensuring (res ≥ −0.17 ≤ res && res ≤ 0.05 && res ± 1e−11)
Arithmetic The arithmetic part of the verification constraints generated by Table 1 can be essentially divided into the ideal part and
the actual part, which includes roundoff errors at each computation
step. The ideal part determines whether the ideal range constraints
in the postcondition are satisfied and the actual part determines
whether the uncertainty part of the postcondition is satisfied. We
can use our procedure presented in Section 6 to compute a sound
approximation of both the result’s range as well as its uncertainty.
Based on this, our tool constructs two approximations. For the first,
the ideal part of the constraint is left untouched and the actual part
is replaced by the computed uncertainty bound. This effectively removes a large number of variables and many times suffiently simplifies the constraint for Z3 to succeed. If this fails, our tool additionally replaces the ideal part by the computed range constraint.
Note that the second approximation may not have enough information to prove a more complex postconditions, as correlation information is lost. We note that the computation of ranges and errors
is the same for both approximations and thus trying both does not
affect efficiency significantly.
Functions and the correspoding verification conditions of this complexity are already within the possibilities of the nonlinear solver
within Z3. For more complex functions however, Z3 does not
(yet) provide an answer in a reasonable time, or returns unknown.
Whether alternative techniques in SMT solvers can help in such
cases remains to be seen [7, 30]. We here provide an approach
based on step-wise approximation that addresses the difficulty of
general-purpose constraint solving.
5.4
Verification with Approximations
In order to nontheless verify interesting programs, we have developed an approximation procedure that computes a sound overapproximation of the range of an expression and of the uncertainty
on the output. This procedure is a forward computation and we
also use it to generate specifications automatically. We describe the
approximation procedure in detail in Section 6, for now we will assume that it exists and, given a precondition P and an expression
expr, computes a sound bound on the output range and its associated uncertainty:
Paths In the case of several paths through the program, we have
the option to consider each path separately or to merge results at
each join in the control flow graph. This introduces a tradeoff between efficiency and precision, since on one hand, considering each
path separately leads to an exponential number of paths to consider.
On the other hand, merging at each join looses correlation information between variables which may be necessary to prove certain
properties. Our approximation pipeline chooses merging first, before resorting to a path-by-path verification in case of failure. We
believe that other techniques for exploring the path space could also
be integrated into our tool [9, 32]. Another possible improvement
are heuristics that select a different order of approximations depending on particular characteristics of the verification condition.
([a, b], err) = evalWithError(P, expr) ⇔
∀~
x, x~◦ , res, res◦ .P (~
x, x~◦ ) ∧ res = expr(~
x) ∧ res◦ = expr◦ (x~◦ )
→ ¬(res < a ∨ b < res) ∧ |res − res◦ | < err
We have identified three possibiities for approximation: nonlinear arithmetic, function calls, and paths and each can be approximated at different levels. We have observed in our experiments, that
“one size does not fit all” and a combination of different approximations is most successful in proving the verification conditions we
encountered. For each verification condition we thus construct approximations until Z3 is able to prove one, or until we run out of approximations where we report the verification as failed. We can thus
view verification as a stream of approximations to be proven. We illustrate the pipeline that computes the different approximations in
Figure 3. The routines getPathError, getRange and evalWithError are
described in the following sections in more detail.
The first approximation (indicated by the long arrow in Figure 3) is to use Z3 alone on the entire constraint constructed by the
rules in Table 1. This is indeed an approximation, as all function
calls are treated as uninterpreted functions in this case. As noted
before, this approach only works in very simple cases or when no
uncertainties and no functions are present.
Example We illustrate the verification algorithm on the example in Figure 4. The functions sineTaylor and sineOrder3 are verified first since they do not contain function calls. Verification with
the full verification constraint fails. Next, our tool computes the errors on the output and Z3 suceeds to prove the resulting constraint
with the ideal part untouched. From this approximation our tool
directly computes a new, more precise postcondition, in particular
it can narrow the resulting error to 1.63e-15. Next, our tool considers the comparison function. Inlining only the postcondition is
not enough in this case, but computing the error approximation on
the inlined functions suceeds in verifying the postcondition. Z3 is
again able to verify the real-valued portion independently. Finally,
the tool verifies that the preconditions of the function calls are satisfied by using Z3 alone. Verification of the function comparisonInvalid
fails with all approximations so that our tool reports unknown. It also
Function calls If the verification constraint contains function
calls and the first approximation failed, our tool will attempt to inline postconditions and pass on the resulting constraint down the
5
2013/9/11
sents possible values of variables as affine forms
def comparisonValid(x: Real): Real = {
require(−2.0 < x && x < 2.0)
val z1 = sineTaylor(x)
val z2 = sineOrder3(x)
z1 − z2
} ensuring(res => ∼res <= 0.1)
x̂ = x0 +
where x0 denotes the central value (of the represented interval) and
each noise symbol i is a formal variable denoting a deviation from
the central value, intended to range over [−1, 1]. The maximum
magnitude of each noise term is given by the corresponding xi .
Note that the sign of xi does not matter in isolation, it does, however, reflect the relative dependence between values. For example,
take x = x0 + x1 1 , then
x − x = x0 + x1 1 − (x0 + x1 1 )
= x0 − x0 + x1 1 − x1 1 = 0
def sineTaylor(x: Real): Real = {
require(−2.0 < x && x < 2.0)
x − (x∗x∗x)/6.0 + (x∗x∗x∗x∗x)/120.0 − (x∗x∗x∗x∗x∗x∗x)/5040.0
} ensuring(res => −1.0 < res && res < 1.0 && res +/− 1e−14)
If we subtracted x = x0 −x1 1 instead, the resulting interval would
have width 2 ∗ x1 and not zero.
The range represented by an affine form is computed as
def sineOrder3(x: Real): Real = {
require(−2.0 < x && x < 2.0)
0.954929658551372 ∗ x − 0.12900613773279798∗(x∗x∗x)
} ensuring(res => −1.0 < res && res < 1.0 && res +/− 1e−14)
[x̂] = [x0 − rad(x̂), x0 + rad(x̂)],
n
X
|xi |
A general affine operation αx̂ + β ŷ + ζ consists of addition, subtraction, addition of a constant (ζ) or multiplication by a constant
(α, β). Expanding the affine forms x̂ and ŷ we get
αx̂ + β ŷ + ζ = (αx0 + βy0 + ζ) +
provides the counterexample it obtains from Z3 (x = 1.0), which
in this case is a valid counterexample.
n
X
(αxi + βyi )i
i=1
An additional motivation for using affine arithmetic is that different contributions to the range it represents remain, at least partly,
separated. This information can be used for instance to help identify the major contributor of a result’s uncertainty or to separate
contributions from external uncertainties from roundoff errors.
Soundness
Our procedure is sound because our constraints overapproximate
the actual errors. Furthemore, even in the full constraint as generated from Table 1, roundoff errors are overapproximated since we
assume the worst-case error bound at each step. While this ensures
soundness, it also introduces incompleteness, as we may fail to validate a specification because our overapproximation is too large.
This implies that counterexamples reported by Z3 are in general
only valid, if they disprove the ideal real-valued part of the verification constraint. In general, this is easy to distinguish from the case
where verification fails due to too large error bounds, as Z3 prefers
simpler counterexamples over complex ones and thus chooses all
error variables to be zero if the ideal part is being violated.
6.1
Range Computation
The goal of this procedure is to perform a forward-computation to
determine the real-valued range of a nonlinear arithmetic expression given ranges for its inputs. Two common possibilities are interval and affine arithmetic, but they tend to overapproximate the
resulting range, especially if the input intervals are not sufficiently
small (order 1). Affine arithmetic improves over interval arithmetic
somewhat by tracking linear correlations, but in the case of nonlinear expressions the results can become actually worse than for
interval arithmetic.
Solving Nonlinear Constraints
Observation: A nonlinear theorem prover such as the one that
comes with Z3 can decide with relatively good precision whether
a given bound is sound or not. That is we can check with a prover
whether for an expression e the range [a, b] is a sound interval enclosure. This observation is the basis of our range computation.
Having given an overview of the approximation pipeline, we now
describe the computation of the approximation for nonlinear arithmetic, which corresponds to the last box in Figure 3. For completeness of presentation, we first review interval and affine arithmetic
which are common choices for performing sound arithmetic computations and which we also use as part of our technique. We then
present our novel procedure for computing the output range of a
nonlinear expression given ranges for its inputs that can be a more
precise substitute for interval or affine arithmetic. Finally, we continue with a procedure that computes a sound overapproximation
of the uncertainty on the result of a nonlinear expression.
One possibility to perform guaranteed computations is to use
standard interval arithmetic [39]. Interval arithmetic computes a
bounding interval for each basic operation as
x ◦ y = [min(x ◦ y), max(x ◦ y)]
rad(x̂) =
i=1
Figure 4. Different polynomial approximations of sine.
6.
xi i
i=1
def comparisonInvalid(x: Real): Real = {
require(−2.0 < x && x < 2.0)
val z1 = sineTaylor(x)
val z2 = sineOrder3(x)
z1 − z2
} ensuring(res => ∼res <= 0.01) // counterexample: 1.0
5.5
n
X
The input to our algorithm is a nonlinear expression expr and a
precondition P on its inputs, which specifies, among possibly other
constraints, ranges on all input variables ~
x. The output is an interval
[a, b] which satisfies the following:
[a, b] = getRange(P, expr) ⇔
∀~
x, res.P (~
x) ∧ res = expr(~
x) → ¬(res < a ∨ b < res)
The algorithm for computing the lower bound of a range is given
in Figure 5. The computation for the upper bound is symmetric. For
each range to be computed, our tool first computes an initial sound
estimate of the range with interval arithmetic. It then performs an
initial quick check to test whether the computed first approximation
bounds are already tight. If not, it uses the first approximation as the
starting point and then narrows down the lower and upper bounds
◦ ∈ {+, −, ∗, /}
and analogously for square root. Affine arithmetic was originally
introduced in [15] and addresses the difficulty of interval arithmetic
in handling correlations between variables. Affine arithmetic repre-
6
2013/9/11
other additional constraints on the ideal variables. The uncertainty
specification is necessary, as it related the ideal and actual variables.
The idea of our procedure is to “execute” a computation while
keeping track of the output range of the current expression and
its associated errors. At each arithmetic operation, we propagate
existing errors, compute an upper bound on the roundoff error
and add it to the overall errors. Since the roundoff error depends
proportionally on the range of values, we also need to keep track of
the ranges as precisely as possible.
Our procedure is build on the abstraction that a computation is
an ideal computation plus or minus some uncertainty. The abstraction of floating-point roundoff errors that we chose also follows this
separation:
def getRange(expr, precondition, precision, maxIterations):
z3.assertConstraint(precondition)
[aInit, bInit] = evalInterval(expr, precondition.ranges);
//lower bound
if z3.checkSat(expr a + precision) == UNSAT
a = aInit
b = bInit
numIterations = 0
while (b−a) < precision ∧ numIterations < maxIterations
mid = a + (b − a) / 2
numIterations++
z3.checkSat(expr mid) match
case SAT ⇒ b = mid
case UNSAT ⇒ a = mid
case Unknown ⇒ break
aNew = a
else
aNew = aInit
f l(x y) = (x y)(1 + δ) = (x y) + (x y)δ
for δ ∈ [−m , m ] and ∈ {+, −, ∗, /}. This allows us to treat all
uncertainties in a unified manner.
Our procedure builds on the idea of the SmartFloat datatype [12],
which uses affine arithmetic to track both the range and the errors.
For nonlinear operations, however, the so computed ranges become
very pessimistic quickly and the error computation may also suffer
from this imprecision. We observed that since the errors tend to be
relatively small, this imprecision does not affect the error propagation itself to such an extent. If the initial errors are small (less than
one), multiplied nonlinear terms tend to be even smaller, whereas
if the affine terms are larger than one, the nonlinear terms grow. We
thus concentrate on improving the ideal range of values and use our
novel range computation procedure for this part and leave the error
propagation with affine arithmetic as in [12].
In our adaptation, we represent every variable and intermediate
computation result as a datatype with the following components:
bNew = ... //upper bound symmetrically
return: [aNew, bNew]
Figure 5. Algorithm for computing the range of an expression.
using a binary search. At each step of the binary search our tool
uses Z3 to confirm or reject the newly proposed bound.
The search stops when either Z3 fails, i.e. returns unknown for
a query or cannot answer within a given timeout, the difference
between subsequent bounds is smaller than a precision threshold,
or the maximum number of iterations is reached. This stopping
criterion can be set dynamically.
x : (range : Interval, err
ˆ : Af f ineF orm)
where range is the range of this variable, computed as described in
Section 6.1 and err
ˆ is the affine form representing the errors. The
(overapproximation) of the actual range including all uncertainties
is then given by totalRange = range + [err],
ˆ where err
ˆ denotes
the interval represented by the affine form.
Additional constraints In addition to the input ranges, the precondition may also contain further constraints on the variables. For
example consider again the method triangle in Figure 1. The precondition bounds the inputs as a, b, c ∈ [1, 9], but the formula is
useful only for valid triangles, i.e. when every two sides together
are longer than the third. If not, we will get an error at the latest when we try to take the square root of a negative number. In
interval-based approaches we can only consider input intervals that
satisfy this constraint for all values, and thus have to check several (and possibly many) cases. In our approach, since we are using
Z3 to check the soundness of bounds, we can assert the additional
constraints up-front and then all subsequent checks are performed
with respect to all additional and initial contraints. This allows us to
avoid interval subdivisions due to imprecisions or problem specific
constraints such as those in the triangle example. This becomes especially valuable in the presence of multiple variables, where we
may need an exponential number of subdivisions.
6.2
Roundoff error computation
each computation step as
Roundoff errors are computed at
ρ = δ ∗ maxAbs(totalRange)
where δ is the machine epsilon, and added to err
ˆ as a fresh noise
term. Note that this roundoff error computation makes our error
computation parametric in the floating-point precision.
Error propagation For affine operations addition, subtraction,
and multiplication by a constant factor the propagated errors are
computed term-wise and thus as for standard affine arithmetic. We
refer the reader to [12, 15] for further details and describe here
only the propagation for nonlinear arithmetic. For multiplication,
division and square root, the magnitude of errors also depends
on the ranges of variables. Since our ranges are not affine terms
themselves, propagation has to be adjusted. In the following, we
denote the range of a variable x by [x] and its associated error by
the affine form err
ˆ x . When we write [x] ∗ err
ˆ y we mean that the
interval [x] is converted into an affine form and the multiplication
is performed in affine arithmetic.
Multiplication is computed as
Error Approximation
We now describe our approximation procedure which, for a given
expression expr and a precondition P on the inputs, computes
the range and error on the output. More formally, our procedure
satisfies the following:
([a, b], err) = evalWithError(P, expr) ⇔
∀~
x, x~◦ , res, res◦ .P (~
x, x~◦ ) ∧ res = expr(~
x) ∧ res◦ = expr◦ (x~◦ )
x ∗ y = ([x] + err
ˆ x )([y] + err
ˆ y)
→ ¬(res < a ∨ b < res) ∧ |res − res◦ | < err
= [x] ∗ [y] + [x] ∗ err
ˆ y + [y] ∗ err
ˆ x + err
ˆ x ∗ err
ˆy +ρ
where ρ is the new roundoff error. Thus the first term contributes
to the ideal range and the remaining three to the error affine form.
The larger the factors [x] and [y] are, the larger the finally computed
where expr◦ represents the expression evaluated in floating-point
arithmetic and ~x, x~◦ are the ideal and actual variables. The precondition specifies the ranges and uncertainties of initial variables and
7
2013/9/11
errors will be. In order to keep the overapproximation as small as
possible, we evaluate [x] and [y] with our new range computation.
Division is computed as
1
x
= x ∗ = ([x] + err
ˆ x )([1/y] + errˆ1/y )
y
y
1
1
= [x] ∗ [ ] + [x] ∗ err
ˆ 1 + [ ] ∗ err
ˆ x + err
ˆ x ∗ err
ˆ1 +ρ
y
y
y
y
2
4
6
8
For square root, we first compute an affine approximation of square
root as in [12]
√
x=α∗x+ζ +θ
and then perform the affine multiplication term wise.
10
12
Overflows and NaN Our procedure allows us to detect potential
overflows, division by zero and square root of a negative value, as
our tool computes ranges of all intermediate values. We currently
report these issues as warnings to the user.
6.3
14
def computePathError(pre, c, f1, f2):
([c], errc ) = evalWithError(pre, c)
([f2]f loat , errf loat ) =
evalWithError(pre ∧ c(x) ∈ [0, errc ], f2)
[f1]real =
getRange(pre ∧ c(x) ∈ [−errc , 0], f1)
return: max |[f1]real − ([f2]f loat + errf loat )|
Figure 6. Computing error due to diverging paths.
Limitations
conditional of the form c(x) == 0 would yield different results
for the ideal and actual computation for nearly any input, so we do
not allow it in our specification language.
The actual computation commits a certain error when computing the condition of the branch and it is this error that causes some
executions to follow a different branch than the corresponding ideal
one would. Consider the case where the ideal computation evaluates f1, but the actual one evaluates f2. Algorithm 6 gives the computation of the path error in this case. The idea is to compute the
ranges of f1 and f2, but only for the inputs that could be diverging. The final error is then the maximum difference of these value.
The algorithm extends naturally to several variables. In the case of
several paths through the program, this error has to be, in principle, computed for each combination of paths. We use Z3 to rule out
infeasible paths up front so that the path error computation is only
performed for those paths that are actually feasible.
We have currently implemented this approach in our tool for
the case when we use merging to handle paths in order to avoid
having to consider an exponential number of path combinations.
We also use a higher default precision and number of iterations
threshold during the binary search in the range computation as this
computation requires in general very tight intervals for each path.
We identify two challenges for performing this computation:
The limitation of this approach is clearly the ability of Z3 to check
our constraints. We found its capabilities satisfactory, although we
expect the performance to still significantly improve. To emphasize
the difference to the constraints that are defined by Table 1, the
constraints we use here do not add errors at each step and thus
the number of variables is reduced significantly. We also found
several transformations helpful, such as rewriting powers (e.g. x ∗
x ∗ x to x3 ), multiplying out products and avoiding non-strict
comparisons in the precondition, although the benefits were not
entirely consistent. Note that at each step of our error computation,
our tool computes the current range. Thus, even if Z3 fails to tighten
the bound for some expressions, we still compute more precise
bounds than interval arithmetic overall in most cases, as the ranges
of the remaining subexpressions have already been computed more
precisely.
7.
def getPathError:
Input: pre (x ∈ [a, b] ∧ x ± n)
program (if (cond(x) < 0) f1(x) else f2(x))
val pathError1 = computePathError(pre, cond, f1, f2)
val pathError2 = computePathError(pre, ¬ cond, f2, f1)
return max (pathError1, pathError2)
Conditional Statements
In this Section we consider the difference between the ideal and
actual computation due to uncertainties on computing branch conditions and the resulting different paths taken. We note that the
full constraint constructed according to Section 3 automatically includes this error. Recall that the ideal and actual computations are
independent except for the initial conditions, so that it is possible
that they follow different paths through the program.
In the case of approximation, however, we compute the error
on individual paths and have to consider the error due to diverging
paths separately. If a method encodes a continous function in the
usual mathematical sense then we note that we only need to quantify errors for each path separately. Thus, if have a method [9] to
determine whether a function with conditional statements is continuous, then our approach described so far is sufficient to provide
sound error bounds. For the case where such a procedure does not
exist or fails to provide an answer, for example due to nonlinearity, or the function simply is not continuous, we propose the following algorithm to explicitly compute the difference between the
ideal and the actual computation across paths. Note that we do not
assume continuity, i.e. the algorithm allows us to compute error
bounds even in the case on non-continous functions.
For simplicity, we present here the algorithm for the case of one
conditional statement:
1. As soon as the program has multiple variables, the inputs for the
different branches are not two-dimensional intervals anymore,
which makes an accurate evaluation of the individual paths
difficult in standard interval arithmetic.
2. The inputs for the two branches are inter-dependent. Thus,
simply evaluating the two branches with inputs that are in the
correct ranges, but are not correlated, yields pessimistic results
when computing the final difference (line 16).
We overcome the first challenge with our range computation
which takes into account additional constraints. For the second
challenge, we use our range computation as well, however unfortunately Z3 fails to tighten the final range to a satisfactory precision
due to timeouts. We still obtain much better error estimates than
with interval arithmetic alone, as the ranges of values for the individual paths are already computed much more precisely. We report
in Section 8 on the type of programs whose verification is already
in our reach today.
8.
if (c(x) < 0) f1(x)
else f2(x)
Experiments
The examples in Figure 1 and 4 and Section 5.3 provide an idea
of the type of programs our tool is currently able to verify fully
automatically. The B-spline example from Section 5.3 is the largest
It generalizes readily to more complex expressions. W.l.o.g. we
assume that the condition is of the form c(x) < 0. Indeed, any
8
2013/9/11
Benchmark
doppler1*
doppler1
doppler2*
doppler2
doppler3*
doppler3
rigidBody1*
rigidBody1
rigidBody2*
rigidBody2
jetEngine*
jetEngine
turbine1*
turbine1
turbine2*
turbine2
turbine3*
turbine3
verhulst*
verhulst
predatorPrey*
predatorPrey
carbonGas*
carbonGas
Sine (single)
Sine
Sqrt (single)
Sqrt
Sine, order 3 (single)
Sine, order 3
Our error (IA only)
2.36e-6
4.92e-13 (4.95e-13)
6.21e-5
1.29e-12
1.23e-4
2.03e-13 (2.05e-13)
9.21e-7
5.08e-13
1.51e-4
6.48e-11
0.15
(-)
1.62e-8 (-)
4.86e-6
1.25e-13 (1.38e-13)
8.05e-6
1.76e-13 (1.96e-13)
3.35e-6
8.50e-14 (9.47e-14)
2.82e-4
6.82e-16
9.22e-5
2.94e-16 (2.96e-16)
2114297.84
4.64e-8 (5.04e-8)
1.03e-6 (1.57e-6)
9.57e-16 (1.46e-15)
9.03e-7 (9.52e-7)
8.41e-16 (8.87e-16)
1.19e-6 (1.55e-6)
1.11e-15 (1.44e-15)
variables in preconditions and path conditions. Table 2 compares
results of our range computation procedure described in Section 6
against ranges obtained with standard interval arithmetic. Interval
arithmetic is one of the methods used for step-wise range estimation; an alternative being affine arithmetic. We have also experimented with an affine arithmetic implementation [12]. However, we
found that affine arithmetic gives more pessimistic results for computing ranges for non-linear benchmarks. We believe that this is
due to imprecision in computing nonlinear operations. Note, however, that we still use affine arithmetic to estimate errors given the
computed ranges.
We set the default precision threshold to 1e−10 and maximum
number of iterations for the binary search to 50. To obtain an
idea about the ranges of our functions, we have also computed
a lower bound on the range using simulations with 107 random
inputs and with exact rational arithmetic evaluation of expressions.
We observe that our range computation can significantly improve
over standard interval bounds. The jetEngine benchmark is a notable
example, where interval arithmetic yields the bound [−∞, ∞],
but our procedure can still provide bounds that are quite close to
the true range. Running times are below 7 seconds for the most
complex benchmarks, except for jetEngine which runs in about 1
minute due to timeouts from Z3 for some intermediate ranges.
Simulated error
5.97e-7
7.11e-14
1.85-5
1.14e-13
5.96e-5
4.27e-14
8.24e-7
2.28e-13
1.25e-4
2.19e-11
3.58e-5
5.46e-12
3.71e-7
1.07e-14
7.66e-7
1.43e-14
1.04e-6
5.33e-15
2.40e-4
2.23e-16
8.61e-5
1.12e-16
168874.70
3.73e-9
1.79e-7
4.45e-16
2.45e-7
4.45e-16
2.12e-7
3.34e-16
Error computation Table 3 compares uncertainties computed by
our tool against maximum uncertainties obtained through extensive simulation with 107 random inputs. We ran the simulation in
parallel with rational and their corresponding floating-point value
and obtained the error by taking the difference in the result. Benchmarks marked with (*) have added initial uncertainties. Unless otherwise indicated, we used double floating-point precision. To our
knowledge this is the first quantitative comparison of an error computation precision with (an approximation) of the true errors on
such benchmarks. Except for the benchmarks jetEngine∗ our computed uncertainties are within an order and many times even closer
to the underapproximation of the true errors provided by simulation. In the case of the jetEngine∗ benchmark, we believe that the
imprecision is mainly due to its complexity and subsequent failures
of Z3. The values in parentheses in the second column indicate errors computed if ranges at each arithmetic operation are computed
using interval arithmetic alone. While we have not attempted to improve the affine arithmetic-based error computation from [12], we
can see that in some cases a more precise range computation can
gain us improvements. The full effect of the imprecision of standard range computation appears when, due to this imprecision, we
obtain possible errors such as division-by-zero or square root of a
negative number errors. The first case happens in the case of the
non-linear jetEngine benchmark, so with interval arithmetic alone
we would therefore not obtain any meaningful result. Similarly, for
the triangle example from Section 2, without being able to constrain the inputs to form valid triangles, we cannot compute any
error bound, because the radicand becomes possibly negative.
Table 4 presents another relevant experiment, evaluating the
ability to use additional constraints during our range computation.
We use the triangle example from Section 2 with additional constraints allowing increasingly flat triangles by setting the threshold
on line 13 (a + b > c + 1e−6) to the different values given in the first
column. As the triangles become flatter, we observe an expected increase in uncertainty on the input since the formula becomes more
prone to roundoff errors. At threshold 1e−10 our range computation
fails to provide the necessary precision and the radicand becomes
possibly negative. (We used double precision in this example as
well.)
Table 3. Comparison of errors computed with our procedure
against simulated errors. Simulations were performed with 107 random inputs. (*) indicates that inputs have external uncertainties associated.
meaningful example we were able to find that Z3 alone could verify
in the presence of uncertainties. For all other cases, it was necessary
to use our approximation methods.
8.1
Evaluating Effectiveness on Nonlinear Expressions
To evaluate our range and error computation technique we have
chosen several nonlinear expressions commonly used in physics,
biology and chemistry [40, 43, 49] as benchmark functions, as well
as benchmarks used in control systems [1] and suitable benchmarks
from [18].Experiments were performed on a desktop computer
running Ubuntu 12.04.1 with a 3.5GHz i7 processor and 16GB of
RAM. Running times highly depend on the timeout used for Z3.
Our default setting is 1 second; we did not find much improvement
in the success rate above this threshold.
Range computation Stepwise estimation of errors crucially depends on the estimate of the ranges of variables. The strength of
using a constraint solver such as Z3 is that it can perform such estimation while taking into account the precise dependencies between
9
2013/9/11
Benchmark
doppler1
doppler2
doppler3
rigidBody1
rigidBody2
jetEngine
turbine1
turbine2
turbine3
verhulst
predatorPrey
carbonGas
Sine
Sqrt
Sine (order 3 approx.)
Our range
[-137.639, -0.033951]
[-230.991, -0.022729]
[-83.066, -0.50744]
[-705.0, 705.0]
[-56010.1, 58740.0]
[-1987.022, 5099.243]
[-18.526, -1.9916]
[-28.555, 3.8223]
[0.57172, 11.428]
[0.31489, 1.1009]
[0.039677, 0.33550]
[4.3032 e6, 1.6740 e7]
[-1.0093, 1.0093]
[1.0, 1.3985]
[-1.0001, 1.0001]
interval arithmetic
[-158.720, -0.029442]
[-276.077, -0.019017]
[-96.295, -0.43773]
[-705.0, 705.0]
[-58740.0, 58740.0]
[−∞, ∞]
[-58.330, -1.5505]
[-29.437, 80.993]
[0.46610, 40.376]
[0.31489, 1.1009]
[0.037277, 0.35711]
[2.0974 e6, 3.4344 e7]
[-2.3012, 2.3012]
[0.83593, 1.5625]
[-2.9420, 2.9420]
Simulated range
[-136.346, -0.035273]
[-227.841,-0.023235]
[-82.624, -0.51570]
[-697.132, 694.508]
[-54997.635, 57938.052]
[-1779.551, 4813.564]
[-18.284, -1.9946]
[-28.528, 3.8107]
[0.61170, 11.380]
[0.36685,0.94492]
[0.039669,0.33558]
[4.1508 e6, 1.69074 e7]
[-1.0093, 1.0093]
[1.0, 1.3985]
[-1.0, 1.0]
Table 2. Comparison of ranges computed with out procedure against interval arithmetic and simulation. Simulations were performed with
107 random inputs. Ranges are rounded outwards.
Benchmark
triangle1 (0.1)
triangle2 (1e-2)
triangle3 (1e-3)
triangle4 (1e-4)
triangle5 (1e-5)
triangle6 (1e-6)
triangle7 (1e-7)
triangle8 (1e-8)
triangle9 (1e-9)
triangle10 (1e-10)
Range
[0.29432, 35.0741]
[0.099375, 35.0741]
[3.16031e-2, 35.0741]
[9.9993e-3, 35.0741]
[3.1622e-3, 35.0741]
[9.9988e-4, 35.0741]
[3.1567e-4, 35.0741]
[9.8888e-5, 35.0741]
[3.0517e-5, 35.0741]
-
Max. abs. error
2.72e-11
8.04e-11
2.53e-10
7.99e-10
2.53e-9
7.99e-9
2.54e-8
8.08e-8
2.62e-7
-
2
4
6
def cav10(x: Real): Real = {
require(x.in(0, 10))
if (x∗x − x >= 0)
x/10
else
x∗x + 2
} ensuring(res => 0 <= res && res <= 3.0 && res +/− 3.0)
8
10
12
14
def squareRoot3(x: Real): Real = {
require( x.in(0,10) && x +/− 1e−10 )
if (x < 1e−5)
1 + 0.5 ∗ x
else
sqrt(1 + x)
} ensuring( res => res +/− 1e−10)
16
Table 4. Ranges and errors for increasingly flat triangles. All values are rounded outwards. Interval arithmetic alone fails to provide
any result.
18
def smartRoot(a: Real, b: Real, c: Real): Real = {
require(3 <= a && a <= 3 && 3.5 <= b && b <= 3.5 &&
c.in(−2, 2) && b∗b − a ∗ c ∗ 4.0 > 0.1)
20
22
8.2
Evaluating Errors across Program Paths
Figure 7 presents several examples to evaluate our error computation procedure across different paths from Section 7. The first
method cav10 [22] has been used before as a benchmark function
for computing the output range. Our tool can verify the given postcondition immediately. Note that the error on the result is actually
as large as the result itself, since the method is non-continuous,
an aspect that has been ignored in previous work, but that our
tool detects automatically. The method squareRoot3 is also an noncontinous function that computes the square root of 1 + x using an
approximation for small values and the regular library method otherwise. Note the additional uncertainty on the input, which could
occur for instance if this method is used in an embedded controller.
Our tool can verify the given spefication. If we change the condition
on line 10 to x < 1e−4 however, verification fails. In this fashion,
we can use our tool to determine the appropriate branch condition
to meet the precision requirement. The above examples verify all
in under 5 seconds. Finally, the smartRoot method computes one
root of a quadratic equation using the well-known more precise
24
26
28
30
val discr = b∗b − a ∗ c ∗ 4.0
if(b∗b − a∗c > 10.0) {
if(b > 0.0) c ∗ 2.0 /(−b − sqrt(discr))
else if(b < 0.0) (−b + sqrt(discr))/(a ∗ 2.0)
else (−b + sqrt(discr))/(a ∗ 2.0)
}
else {
(−b + sqrt(discr))/(a ∗ 2.0)
}
} ensuring (res => res +/− 6e−15)
Figure 7. Path error computation examples.
method from [23]. We are currently not aware of an automatic tool
to prove programs continuous in the presence of nonlinear arithmetic, so that we need to compute the error across different paths
as well. Our tool succeeds in verifying the postcondition in about
25s. The rather long running time is due to the complexity of the
conditions when computing the error across paths, and thus Z3’s
longer response time, and a number of Z3 timeouts (Z3 timeout
here means merely that some ranges have not been tightend to the
best precision). In the future, we envision that optimizations that
10
2013/9/11
arithmetic [30] will make many more verification problems feasible
with our techniques. An alternative to this approach is using linear
approximations to solve polynomial constraints [7]. We believe that
such advances are largely orthogonal to our use of range arithmetic
and complement each other.
select alternative approximations can be performed automatically
by a trustworthy compiler.
9.
Related work
Current approaches for verifying floating-point code include abstract interpretation, interactive theorem proving and decision procedures, which we survey in this section. We are not aware of work
that would automatically integrate reasoning about uncertainties.
Testing Symbolic execution is a well-known technique for generating test inputs. [6] use a combination of meta-heuristic search and
interval constraint solving to solve the floating-point constraints
that arise, whereas [33] combine random search and evolutionary
techniques. [47] test numerical code for precision by perturbing
low-order bits of values and rewriting expressions. The idea is to
exagerate initial errors and thus make imprecisions more visible.
Probabilistic arithmetic [45] is a similar approach but it does the
perturbation by using different rounding modes. [4] also propose
a testing produce to detect accuracy problems by instrumentring
code to perform a higher-precision computation side by side with
the regular computations. While these approaches are sound with
respect to floating-point arithmetic, they only generate or can check
individual inputs and are thus not able to verify or compute output
ranges or their roundoff errors.
Abstract interpretation (AI) Abstract domains that are sound
with respect to floating-point computations can prove bounds on
the ranges of variables [5, 11, 20, 28, 38]. The only work in this area
that can also quantify roundoff errors is the tool Fluctuat[17, 24].
These techniques use interval or affine arithmetic and together with
the required join and meet operations may yield too pessimistic
results. [42] improves the precision of Fluctuat by refining the input domains with a constraint solver. Our approach can be viewed
as approaching the problem from a different end, starting with an
exact constraint and then using approximation until the solver succeeds. Unlike AI tools in general, our system currently handles
only functional code, in particular it does not handle loops. If the
user can provide inductive postconditions, then we can still prove
the code correct, but we do not in general discover these ourselves.
Our focus lies on proving precise bounds on the ranges in the presence of nonlinear computations and the quantification of roundoff
errors and other uncertainties.
Robustness analysis [27] combines abstract interpretation with
model checking to check programs for stability by tracking the evolution of the width of the interval representing a single input. [36]
use concolic execution to find inputs which, given maximum deviations on inputs, maximize the deviation on the outputs. These two
works however, use a testing approach and cannot provide sound
guarantees. [9] presents a framework for continuity analysis of programs along the mathematical −δ definition of continuity and [10]
builds on this work and presents a sound robustness analysis. This
framework provides a syntactic proof of robustness for programs
over reals and thus does not consider floating-points. Our approach
describes a quantitative measure of robustness for nonlinear programs with floating-point numbers and other uncertainties, and we
believe that it can complement the cited framework.
Theorem proving The Gappa tool [14, 35] generates a proof
checkable by the interactive theorem prover Coq from source code
with specifications. It can reason about properties that can be reduced to reasoning about ranges and errors, but targets, very precise
properties of specialized functions, such as software implementations of elementary functions. The specification itself requires expertize and the proofs human intervention. A similar approach is
taken by [2] which generate verification conditions that are discharged by various theorem provers. Harisson has also done significant work on proving floating-point programs in the HOL Light
theorem prover [26].
Our approach makes a different compromise on the precision
vs. automation tradeoff, by being less precise, but automatic. The
Gappa approach can be used complementary to ours, in that if we
detect that more precision is needed, Gappa is employed by an
expert user on selected methods, and the results are then used by
our tool instead of automatically computed specifications.
10.
Conclusion
We have presented a programming model for numerical programs
that decouples the mathematical problem description from its realization in finite precision. The model uses a Real data type that
corresponds to mathematical real numbers. The developer specifies
the program using reals and indicates the target precision; the compiler chooses a floating point representation while checking that the
desired precision targets are met. We have described the soundness
criteria by translating programs with precision requirements into
verification conditions over mathematical reals. The resulting verification conditions, while a natural description of the problem being
solved, are difficult to solve using a state-of-the art SMT solver Z3.
We therefore developed an algorithm that combines SMT solving
with range computation. Our notion of soundness incorporates full
input/output behavior of functions, taking into account that, due to
conditionals, small differences in values can lead to different paths
being taken in the program. For such cases our approach estimates
a sound upper bound on the total error of the computation.
We have evaluated our techniques on a number of benchmarks
from the literature, including benchmarks from physics, biology,
chemistry, and control systems. We have found that invocation of
SMT solver alone is not sufficient to handle these benchmarks due
to scalability issues, whereas the use of range arithmetic by itself
is not precise enough. By combining these two techniques we were
able to show that a floating point version of the code conforms to
the real-valued version with reasonable precision requirements.
We believe that our results indicate that it is reasonable to introduce Reals as a data type, following a list of previously introduced mathematical abstractions in programming languages, such
Range computation The Gappa tool and most constraint solvers
internally use interval arithmetic for sound range computations,
whose limitations are well-known. [19] describes an arithmetic
based on function enclosures and [37] use an arithmetic based
on taylor series as an alternative. This approach is useful when
checking a constraint, but is not suitable for a forward computation
of ranges and errors.
Decision procedures An alternative approach to verification via
range computation are floating-point decision procedures. Bitprecise constraints, however, become very large quickly. [8] addresses this problem by using a combination of over- and underapproximations. [25] present an alternative approach in combining
interval constraint solving with a CDCL algorithm and [21] is a
decision procedure for nonlinear real arithmetic combining interval constraint solving with an SMT solver for linear arithmetic.[44]
formalizes the floating-points for the SMT-LIB format. While these
approach can check ranges on numeric variables, they do not handle
roundoff errors or other uncertainties and cannot compute specifications automatically.
Our techniques rely on the performance of Z3. We hope that
an integration of the recent new improved solver for nonlinear
11
2013/9/11
as unbounded integers, rationals, and algebraic data types. The feasibility of verified compilation of our benchmarks suggests that it
is realistic to decouple the verification of executable mathematical
models over reals from their sound compilation. We therefore expect that this methodology will help advance rigorous formal verification of numerical software and enable us to focus more on highlevel correctness properties as opposed to run-time errors alone.
[24] E. Goubault, S. Putot, and F. Védrine. Modular static analysis with
zonotopes. In SAS, 2012.
[25] L. Haller, A. Griggio, M. Brain, and D. Kroening. Deciding floatingpoint logic with systematic abstraction. In FMCAD, 2012.
[26] J. Harrison. Floating-Point Verification using Theorem Proving. In
Formal Methods for Hardware Verification, 6th International School
on Formal Methods for the Design of Computer, Communication, and
Software Systems, SFM, 2006.
[27] F. Ivancic, M. Ganai, S. Sankaranarayanan, and A. Gupta. Numerical
stability analysis of floating-point computations using software model
checking. In MEMOCODE, 2010.
[28] B. Jeannet and A. Miné. Apron: A Library of Numerical Abstract
Domains for Static Analysis. In CAV, 2009.
[29] J. Jiang, W. Luk, and D. Rueckert. FPGA-Based Computation of FreeForm Deformations. In Field-Programmable Logic and Applications,
2003.
References
[1] A. Anta and P. Tabuada. To Sample or not to Sample: Self-Triggered
Control for Nonlinear Systems. IEEE Transactions on Automatic
Control, 55(9), 2010.
[2] A. Ayad and C. Marché. Multi-prover verification of floating-point
programs. In IJCAR, 2010.
[3] D. H. Bailey, Y. Hida, X. S. Li, and B. Thompson. C++/Fortran90 double-double and quad-double package.
http://crdlegacy.lbl.gov/˜dhbailey/mpdist/, 2013.
[4] F. Benz, A. Hildebrandt, and S. Hack. A dynamic program analysis to
find floating-point accuracy problems. In PLDI, 2012.
[5] B. Blanchet, P. Cousot, R. Cousot, J. Feret, L. Mauborgne, A. Miné,
D. Monniaux, and X. Rival. A static analyzer for large safety-critical
software. In PLDI, pages 196–207, 2003.
[6] M. Borges, M. d’Amorim, S. Anand, D. Bushnell, and C. S. Pasareanu.
Symbolic Execution with Interval Solving and Meta-heuristic Search.
In ICST, 2012.
[7] C. Borralleras, S. Lucas, A. Oliveras, E. Rodrı́guez-Carbonell, and
A. Rubio. Sat modulo linear arithmetic for solving polynomial constraints. J. Automated Reasoning, 48(1), 2012.
[8] A. Brillout, D. Kroening, and T. Wahl. Mixed abstractions for floatingpoint arithmetic. In FMCAD, pages 69–76, 2009.
[9] S. Chaudhuri, S. Gulwani, and R. Lublinerman. Continuity analysis
of programs. In POPL, 2010.
[10] S. Chaudhuri, S. Gulwani, R. Lublinerman, and S. Navidpour. Proving
Programs Robust. In ESEC/FSE, 2011.
[11] L. Chen, A. Miné, J. Wang, and P. Cousot. Interval Polyhedra: An
Abstract Domain to Infer Interval Linear Relationships. In SAS, 2009.
[12] E. Darulova and V. Kuncak. Trustworthy numerical computation in
Scala. In OOPSLA, 2011.
[13] E. Darulova, V. Kuncak, R. Majumdar, and I. Saha. Synthesis of fixedpoint arithmetic code. In EMSOFT, 2013.
[14] F. de Dinechin, C. Lauter, and G. Melquiond. Certifying the FloatingPoint Implementation of an Elementary Function Using Gappa. IEEE
Trans. Comput., 2011.
[15] L. H. de Figueiredo and J. Stolfi. Self-Validated Numerical Methods
and Applications. IMPA/CNPq, Brazil, 1997.
[16] L. De Moura and N. Bjørner. Z3: an efficient SMT solver. In TACAS,
2008.
[17] D. Delmas, E. Goubault, S. Putot, J. Souyris, K. Tekkal, and
F. Vedrine. Towards an industrial use of FLUCTUAT on safety-critical
avionics software. In FMICS, 2009.
[18] V. D’Silva, L. Haller, D. Kroening, and M. Tautschnig. Numeric
Bounds Analysis with Conflict-driven Learning. In TACAS, 2012.
[19] J. A. Duracz and M. Konecny. Polynomial function enclosures and
floating point software verification. Constraints in Formal Verification/IJCAR, 2008.
[20] J. Feret. Static Analysis of Digital Filters. In the 13th European
Symposium on Programming - ESOP, 2004.
[21] S. Gao, M. Ganai, F. Ivancic, A. Gupta, S. Sankaranarayanan, and
E. Clarke. Integrating ICP and LRA solvers for deciding nonlinear
real arithmetic problems. In FMCAD, 2010.
[22] K. Ghorbal, E. Goubault, and S. Putot. A Logical Product Approach
to Zonotope Intersection. In CAV, 2010.
[23] D. Goldberg. What every computer scientist should know about
floating-point arithmetic. ACM Comput. Surv., 23(1), 1991.
[30] D. Jovanović and L. de Moura. Solving Non-linear Arithmetic. In
IJCAR 2012, 2012.
[31] W. Kahan. Miscalculating Area and Angles of a Needle-like Triangle.
Technical report, University of California Berkeley, 2000.
[32] V. Kuznetsov, J. Kinder, S. Bucur, and G. Candea. Efficient State
Merging in Symbolic Execution. In PLDI, 2012.
[33] K. Lakhotia, N. Tillmann, M. Harman, and J. de Halleux. FloPSy Search-Based Floating Point Constraint Solving for Symbolic Execution. In Testing Software and Systems. Springer Berlin / Heidelberg,
2010.
[34] X. Leroy. Verified squared: does critical software deserve verified
tools? In POPL, 2011.
[35] M. D. Linderman, M. Ho, D. L. Dill, T. H. Meng, and G. P. Nolan. Towards program optimization through automated analysis of numerical
precision. In CGO, 2010.
[36] R. Majumdar, I. Saha, and Z. Wang. Systematic Testing for Control
Applications. In MEMOCODE, 2010.
[37] K. Makino and M. Berz. Taylor Models and Other Validated Functional Inclusion Methods. International Journal of Pure and Applied
Mathematics, 4, 2003.
[38] A. Miné. Relational Abstract Domains for the Detection of FloatingPoint Run-Time Errors. In ESOP, 2004.
[39] R. Moore. Interval Analysis. Prentice-Hall, 1966.
[40] J. D. Murray. Mathematical Biology, I. An Introduction. Springer,
2002.
[41] M. Odersky, L. Spoon, and B. Venners. Programming in Scala: A
Comprehensive Step-by-step Guide. Artima Incorporation, 2008.
[42] O. Ponsini, C. Michel, and M. Rueher. Refining Abstract Interpretation Based Value Analysis with Constraint Programming Techniques.
In CP, 2012.
[43] A. Quarteroni, F. Saleri, and P. Gervasio. Scientific Computing with
MATLAB and Octave. Springer, 3rd edition, 2010.
[44] P. Rümmer and T. Wahl. An SMT-LIB Theory of Binary FloatingPoint Arithmetic. In Informal proceedings of 8th International Workshop on Satisfiability Modulo Theories (SMT) at FLoC, 2010.
[45] N. Scott, F. Jézéquel, C. Denis, and J.-M. Chesneaux. Numerical
‘health check’ for scientific codes: the CADNA approach. Computer
Physics Communications, 2007.
[46] I. C. Society. IEEE Standard for Floating-Point Arithmetic. IEEE Std
754-2008, 2008.
[47] E. Tang, E. Barr, X. Li, and Z. Su. Perturbing numerical calculations
for statistical analysis of floating-point program (in)stability. In ISSTA,
2010.
[48] E. M. Westbrook and S. Chaudhuri. A Semantics for Approximate
Program Transformations. CoRR, 2013.
[49] C. Woodford and C. Phillips. Numerical Methods with Worked Examples, volume 2nd. Springer, 2012.
12
2013/9/11
| 6cs.PL
|
Fundamental concepts in the Cyclus nuclear fuel cycle simulation framework
Kathryn D. Huffa,∗, Matthew J. Giddenb , Robert W. Carlsenb , Robert R. Flanaganc , Meghan B. McGarryb ,
Arrielle C. Opotowskyb , Erich A. Schneiderc , Anthony M. Scopatzd , Paul P.H. Wilsonb
a University of California - Berkeley, Department of Nuclear Engineering Berkeley, CA, 94720
of Wisconsin - Madison, Department of Nuclear Engineering and Engineering Physics, Madison, WI 53706
c University of Texas - Austin, Department of Mechanical Engineering, Nuclear and Radiation Engineering Program, Austin,
TX 78758
d University of South Carolina, Nuclear Engineering Program, Columbia, SC 29201
arXiv:1509.03604v2 [cs.SE] 11 Mar 2016
b University
Abstract
As nuclear power expands, technical, economic, political, and environmental analyses of nuclear fuel cycles
by simulators increase in importance. To date, however, current tools are often fleet-based rather than
discrete and restrictively licensed rather than open source. Each of these choices presents a challenge to
modeling fidelity, generality, efficiency, robustness, and scientific transparency. The Cyclus nuclear fuel
cycle simulator framework and its modeling ecosystem incorporate modern insights from simulation science
and software architecture to solve these problems so that challenges in nuclear fuel cycle analysis can be
better addressed. A summary of the Cyclus fuel cycle simulator framework and its modeling ecosystem are
presented. Additionally, the implementation of each is discussed in the context of motivating challenges in
nuclear fuel cycle simulation. Finally, the current capabilities of Cyclus are demonstrated for both open
and closed fuel cycles.
Keywords: nuclear fuel cycle, simulation, agent based modeling, nuclear engineering, object orientation,
systems analysis
1. Introduction
As nuclear power expands, technical, economic,
political, and environmental analyses of nuclear fuel
cycles by simulators increase in importance. The
merits of advanced nuclear technologies and fuel
cycles are shaped by myriad physical, nuclear, chemical, industrial, and political factors. Nuclear fuel
cycle simulators must therefore couple complex models of nuclear process physics, facility deployment,
and material routing.
Indeed, the cardinal purpose of a dynamic nuclear
fuel cycle simulator is to calculate the time- and
facility-dependent mass flow through all or part
the fuel cycle. Dynamic nuclear fuel cycle analysis
more realistically supports a range of simulation
goals than static analysis [1]. Historically, dynamic
nuclear fuel cycle simulators have calculated fuel
cycle mass balances and performance metrics derived
∗ Corresponding
Author
Email address: [email protected] (Kathryn D. Huff)
Preprint submitted to Advances in Engineering Software
from them using software ranging from spreadsheetdriven flow calculators to highly specialized system
dynamics modeling platforms.
To date, current tools are typically distributed
under restrictive rather than open source licenses,
having been developed in industrial contexts or using
commercial software platforms. Additionally, having
often been developed for customized applications,
many possess inflexible architectures, never having
been designed to enable new features or extensions.
Finally, many model only fleet-level dynamics of facilities and materials rather than discrete resolution
of those individual agents and objects. When the
DOE-NE Fuel Cycles Technologies Systems Analysis Campaign developed requirements necessary in
a next generation fuel cycle simulator, three main
failure modes were associated with those software
choices. First, they discourage targeted contribution and collaboration among experts. Next, they
hobble efforts to directly compare modeling methodologies. Finally, they over-specialize, rendering most
tools applicable to only a subset of desired simulaMarch 14, 2016
1.1. Background
Nuclear fuel cycle simulators drive research development and design (RD&D) by calculating ‘metrics’,
quantitative measures of performance that can be
compared among fuel cycle options. The feasibility of the technology development and deployment
strategies which comprise a fuel cycle option, the operational features of nuclear energy systems, the dynamics of transitions between fuel cycles, and many
other measures of performance can be expressed
in terms of these metrics. For example, economic
feasibility is often measured in levelized cost of electricity (LCOE), a combination of fuel and operating
costs normalized by electricity generation, while
environmental performance might be measured by
spent fuel volume, radiotoxicity, or mined uranium
requirements. A meta-analysis of fuel cycle systems
studies identified over two dozen unique quantitative
metrics spanning economics and cost, environmental sustainability and waste management impacts,
safety, security and nonproliferation, resource adequacy and utilization, among others [11]. With few
exceptions, these metrics are derived from mass balances and facility operation histories calculated by
a fuel cycle simulator. For example, where nuclear
waste repository burden is derived from ejected fuel
masses, water pollution or land use can be derived
from facility operational histories (as in [12]).
However, methods for calculating those metrics
vary among simulators. Some model the system of
facilities, economics, and materials in static equilibrium, while other simulators capture the dynamics
of the system. Similarly, while some simulators discretely model batches of material and individual
facilities, others aggregate facilities into fleets and
materials into streams. Some simulators were designed to model a single aspect of the fuel cycle
in great detail while neglecting others. For example, a simulator created for policy modeling might
have excellent capability in economics while capabilities for tracking transformations in material isotopics and the effects of isotopics on technology
performance are neglected. The Code for Advanced
Fuel Cycles Assessment (CAFCA)[13] simulator
is problem-oriented in this way, having elected to
neglect isotopic resolution in favor of integral effects.
Historically, domestic national laboratories have
driven development and regulated the use of their
own tools: the Verifiable Fuel Cycle Simulation
Model (VISION)[14], Dynamic Model of Nuclear
Development (DYMOND)[15], and Nuclear Fuel
Cycle Simulator (NFCSim)[16, 17]. Internationally,
tion fidelities, scales, and applications. Those three
constraints were identified as presenting significant
challenges to modeling fidelity, generality, efficiency,
robustness, and scientific transparency in the field
of fuel cycle analysis [2].
The Cyclus nuclear fuel cycle simulator framework and its modeling ecosystem, the suite of agents
and other physics plug-in libraries compatible with
it, incorporate modern insights from simulation science and software architecture to solve these problems. These modern methods simultaneously enable
more efficient, accurate, robust, and validated analysis. This next-generation fuel cycle simulator is the
result of design choices made to:
• support access to the tool by fuel cycle analysts
and other users,
• encourage developer extensions,
• enable plug-and-play comparison of modeling
methodologies,
• and address a range of analysis types, levels of
detail, and analyst sophistication.
Cyclus is a dynamic, agent-based model, which
employs a modular architecture, an open development process, discrete agents, discrete time, and
arbitrarily detailed isotopic resolution of materials. Experience in the broader field of systems
analysis indicates that agent-based modeling enables more flexible simulation control than system
dynamics, without loss of generality [3]. Furthermore, openness allows cross-institutional collaboration, increases software robustness [4, 5], improves
the strength and quality of results through peer review [6, 7, 8, 9, 10], and cultivates an ecosystem of
modeling options. This ecosystem is modular, being
comprised of dynamically loadable, interchangeable,
plug-in libraries of fuel cycle component process
physics that vary in their scope, depth, and fidelity.
This modularity allows users and developers to customise Cyclus to analyze the cases that are of
interest to them rather than any custom application
the simulator was originally developed to address.
Additionally, that customizability allows users and
developers to address those cases at the level of
fidelity necessary for their application. The fundamental concepts of the Cyclus nuclear fuel cycle
simulator capture these modern insights so that
novel challenges in nuclear fuel cycle analysis can
be better addressed.
2
other laboratories have created their own as well,
such as Commelini-Sicard (COSI)[18, 19, 20, 21]
and ORION[22]. Finally, some simulators initiated
in a national lab setting have been continued as
propriety, industry-based simulators, such as Dynamic Analysis of Nuclear Energy System Strategies
(DANESS)[23]. Outside the national laboratories,
researchers have created new nuclear fuel cycle simulation tools when existing tools were not available
or not sufficiently general to calculate their metrics of interest. With limited access to the national
laboratory tools and a need to customize them for
research purposes, universities and private industry
researchers have “reinvented the wheel” by developing tools of their own from scratch and tailored
to their own needs. Examples include CAFCA[24]
and Dynamic Analysis of Nuclear Energy Systems
Strategies (DESAE)[25, 26, 17].
Cyclus emerged from a line of tools seeking to
break this practice. Its precursor, Global Evaluation
of Nuclear Infrastructure Utilization Scenarios (GENIUS) Version 1 [27, 28], originated within Idaho
National Laboratory (INL) and sought to provide
generic regional capability. Based on lessons learned
from GENIUS Version 1, the GENIUS Version 2
[29, 30] simulator sought to provide more generality
and an extensible interface to facilitate collaboration. The Cyclus project then improved upon the
GENIUS effort by implementing increased modularity and encapsulation. The result is a dynamic
simulator that treats both materials and facilities discretely, with an architecture that permits multiple
and variable levels of fidelity. Using an agent-based
framework, the simulator tracks the transformation
and trade of resources between autonomous regional
and institutional entities with customizable behavior and objectives. Each of these concepts (agentbased, resource tracking, and regional as well as
institutional entities) will be described in their own
sections (sections 2.2, 2.3.1, and 2.2.2 respectively).
Together, they provide a capability for extension
and reuse beyond that pursued by any existing fuel
cycle simulator.
This essential capability is absent in previous simulators where user customization and extensibility were
not design objectives. While the modular and open
architecture of Cyclus is necessary to meet these
goals, it is not sufficient. Agent interchangeability is
also required to facilitate direct comparison of alternative modeling methodologies and facility concepts.
With this concept at its core, Cyclus provides a
platform for users to quickly develop the capabilities
at a level of detail and validation necessary for their
unique applications. Finally, Cyclus is applicable
to a broader range of fidelities, scales, and applications than other simulators, due to the flexibility
and generality of its agent-based modeling (ABM)
paradigm and discrete, object-oriented approach.
This structure recognizes that specialists should
utilize their time and resources in modeling the specific process associated with their area of expertise
(e.g., reprocessing and advanced fuel fabrication),
without having to create a model of the entire fuel
cycle to serve as its host. Cyclus supports them
by separating the problem of modeling a physicsdependent supply chain into two distinct components: a simulation kernel and archetypes that interact with it. The kernel is responsible for supporting
the deployment and interaction logic of entities in
the simulation. Physics calculations and customized
behaviors of those entities are implemented within
archetype classes.
Ultimately, modeling the evolution of a physicsdependent, international nuclear fuel supply chain
is a multi-scale problem which existing tools cannot
support. They have either focused on macro effects,
e.g., the fleet-level stocks and flows of commodities,
or micro effects, e.g., the used-fuel composition of
fast reactor fuel. Each focus has driven the development of specialized tools, rendering the task of
answering questions between the macro and micro
levels challenging within a single tool. In contrast,
the open, extensible architecture and discrete object
tracking of Cyclus allow the creation and interchangeability of custom archetypes at any level of
fidelity and by any fuel cycle analyst.
1.2. Motivation
The Cyclus paradigm enables targeted contribution and collaboration within the nuclear fuel cycle
analysis community to achieve two important goals:
lower the barrier for users to include custom nuclear
technologies and facility types in their fuel cycle
analyses while improving the ability to compare simulations with and without those custom concepts.
1.2.1. Open Access and Development Practices
The proprietary concerns of research institutions
and security constraints of data within fuel cycle
simulators often restrict access. Use of a simulator is therefore often limited to its institution of
origin, necessitating effort duplication at other institutions and thereby squandering broader human
resources. License agreements and institutional ap3
proval are required for most current simulators (e.g.
COSI6, DANESS, DESAE, EVOLCODE, FAMILY21, NFCSim)[31], including ORION, and VISION. Even when the source code is unrestricted,
the platform on which it relies (e.g. VENSIM) is
often restricted or costly. The MIT CAFCA software, for example, relies on the commercially licensed VENSIM product as a platform. Cyclus,
on the other hand, is written in C++ for which
freely available development tools and an open standard are available. Further, Cyclus relies only on
open source, freely available libraries. As such, it
provides fully free and open access to all users and
developers, foreign and domestic.
Moreover, both technical and institutional aspects
of the software development practices employed
by the Cyclus community facilitate collaboration.
Technically, Cyclus employs a set of tools commonly used collaborative software development that
reduce the effort required to comment on, test and
ultimately merge individual contributions into the
main development path. For many of the simulation platforms adopted by previous simulators, there
were technical obstacles that impeded this kind of
collaboration. Institutionally, Cyclus invites all
participants to propose, discuss and provide input to
the final decision making for all important changes.
as they enter and leave the simulation over time and
seek to trade materials whose specific composition
may not be known a priori.
1.2.3. Discrete Facilities and Materials
Many fuel cycle phenomena have aggregate
system-level effects which can only be captured
by discrete material tracking [2]. Cyclus tracks
materials as discrete objects. Some current fuel
cycle simulation tools such as COSI [26, 32, 24],
FAMILY21[26], GENIUS version 1, GENIUS version 2, NFCSim, and ORION also possess the ability
to model discrete materials. However, even among
these, the ability to model reactor facilities individually is not equivalent to the ability to model distinct
activities. COSI, for example, has some support for
modeling reactors individually, but according to a
recent benchmark [33], it models many reactors operating in sync. That is, refuelling and discharging
occur simultaneously for all reactors. While Cyclus
allows this type of fleet-based aggregation of reactor
behavior, Cyclus also enables operations in each
facility to vary independently of any others in the
simulation.
Similarly, the ability to model disruptions (i.e.
facility shutdowns due to insufficient feed material or
insufficient processing and storage capacity) is most
readily captured by software capable of tracking
the operations status of discrete facilities [2]. Fleetbased models (e.g. VISION) are unable to capture
this gracefully, since supply disruptions are modeled
as a reduction in the capacity of the whole fleet.
All of the software capable of discrete materials
have a notion of discrete facilities, however not all
handle disruption in the same manner. DESAE, for
example, does not allow shutdown due to insufficient
feedstock. In the event of insufficient fissile material
during reprocessing, DESAE borrows material from
storage, leaving a negative value [26]. The Cyclus
framework does not dictate such heuristics. Rather,
it provides a flexible framework on which either
method is possible.
A final benefit of the discreteness of facilities and
materials is their power when combined. The ability to track a material’s history as it moves from
one facility to another is unique to Cyclus. While
some current simulators track materials in discrete
quanta, they do not necessarily preserve the identity
of each quantum as the materials move around the
fuel cycle. When coupled with Cyclus’ individual facility modeling, this capacity becomes distinct
from what other fuel cycle simulators are able to do.
1.2.2. Modularity and Extensibility
Modularity is a key enabler of extending the scope
of fuel cycle analysis within the Cyclus framework.
Changes that are required to improve the fidelity of
modeling a particular agent, or to introduce entirely
new agents, are narrowly confined and place no new
requirements on the Cyclus kernel. Furthermore,
there are few assumptions or heuristics that would
otherwise restrict the algorithmic complexity that
can be used to model the behavior of such agents.
For example, most current simulators describe
a finite set of acceptable cycle constructions (once
through, single-pass, multi-pass). That limits the
capability to create novel material flows and economic scenarios. The Cyclus simulation logic relies
on a market paradigm, parameterized by the user,
which flexibly simulates dynamic responses to pricing, availability, and other institutional preferences.
This minimal set of mutual dependencies between
the kernel and the agents is expressed through the
dynamic resource exchange (DRE) that provides a
level of flexibility that does not exist in other fuel
cycle simulators. It creates the potential for novel
agent archetypes to interact with existing archetypes
4
So, while FAMILY21 and COSI can identify whether
a batch being discharged from a reactor originated
in mixed oxide (MOX) fabrication rather than fresh
uranium oxide (UOX) fabrication, Cyclus can go
further, tracking which of the fresh batches contained material from a particular discharged batch.
By extension, Cyclus can also report which individual facilities the batch passed through and in
which it originated. The ability to track a single
material through the simulation, though it might
be split, transmuted, or merged with other materials along the way, allows Cyclus to answer more
data-rich questions that previous simulators have
been unable to ask. For example, it allows precise
tracking of specific material diversions, so queries
about nonproliferation robustness in a facility can
be levied either in the context of a single event or a
series of nefarious acts.
This encapsulated ‘plug-in’ design choice provides
two major benefits. First, analysts can take advantage of the simulation logic API and archetype
ecosystem when they apply Cyclus to their specific
problem. A modeler can focus on creating or customizing nuclear facility, institution, resource, and
toolkit models within their specific area of technical
expertise. Second, because Cyclus uses a modular
archetype approach, comparing two archetypes is
straightforward. For example, if an analyst would
like to compare the effect of using different models to determine the input fuel composition for fast
reactors, fuel fabrication archetypes can be developed and interchanged while keeping the rest of the
models used in the simulation fixed.
2.1.1. Cluster-Ready Software
Many fuel cycle simulators rely on commercial, offthe-shelf (COTS) and Windows-only software that
limits performance on and compatibility with resource computing infrastructures (e.g. cluster or
cloud computing). This constrains the possible
scope of simulations and increases the wall-clock
time necessary to conduct parameterized sensitivity
analyses and other multi-simulation studies. For
example, large scale sensitivity analyses to quantify
the dependence of fuel cycle outcomes are only feasible in a massively parallelized environment. To
enable such research, Cyclus, is primarily written
in C++ and relies on libraries supported by Linux
and UNIX (including Ubuntu and OSX) platforms,
which are flexible and support parallelization.
Cyclopts [34], a proof-of-principle Cyclusenabled application on such a large compute system,
uses UW-Madison’s HTCondor high-throughput
computing (HTC) infrastructure to study DRE performance and outcomes. For example, an investigation of the effect of solution degeneracy, a commonly observed phenomenon in scenarios with individual facilities and basic (e.g., commodity-only)
preference definitions, was performed for three different fuel cycles: once-through, MOX fast reactorthermal reactor cycles, and MOX-thorium oxide
(ThOX) recycle in fast reactors-thermal reactor cycles. Objective coefficients were generated based on
two factors: commodity-facility pairings and facility
location. The population of the possible values of
commodity-facility pairings is always small (O(10)).
The size of the population of possible values of the
location effect was investigated for values of zero,
ten, and infinity (i.e., any real number). The study
additionally included an investigation of problem-
2. Methodology and Implementation
A modular, agent-based modeling (ABM) approach is ideal for modeling the coupled, physicsdependent supply chain problems involving material
routing, facility deployment, and regional and institutional hierarchies which arise in fuel cycle analysis.
Additionally, the choice to build Cyclus on open
source libraries in modern programming languages
enables both remote and multiprocess execution
on a number of platforms. This section begins by
describing the general design features that make
Cyclus both flexible and powerful: cluster-ready
software and dynamically loadable libraries. The
ABM framework is then described, focusing on its
implementation and benefits in a fuel cycle context. A discussion of the time-dependent treatment
of discrete resources follows, focusing on the DRE.
Support for users and developers via the Cycamore
library of archetypes and the experimental toolkit
are also presented. Lastly, the methods for quality
assurance are outlined.
2.1. Modular Software Architecture
The architecture of Cyclus allows developers to
define nuclear fuel cycle processes independent of
the simulation logic. To achieve this, agents are developed which represent facilities, institutions, and
regions comprising the nuclear fuel cycle. These
agents are created using the Cyclus framework
application programming interface (API), a set of
functions and protocols which assist in agent development and specify how agents should be defined.
5
scaling behavior in order to quantify the magnitude
and rate-of-increase of DRE solution times as a function of the simulation-entity population for each fuel
cycle[35]. In total, Cyclopts has run over 105 jobs,
comprising more than 60,000 compute hours. The
HTC infrastructure has separately been utilized to
run and collect information from full Cyclus simulations running in parallel on 103 machines reliably
for order 105 simulations.
2.1.2. Dynamically Loadable Libraries
The Cyclus architecture encourages efficient, targeted contribution to the ecosystem of archetype
libraries. With Cyclus, a researcher can focus on
generating an archetype model within their sphere
of expertise while relying on the contributions of
others to fill in the other technologies in the simulation. Similarly, individual developers may explore
different levels of complexity within their archetypes,
including wrapping other simulation tools as loadable libraries within Cyclus.
Cyclus achieves this behavior by implementing
generic APIs and a modular architecture via a suite
of dynamically loadable plug-in libraries (pictured
in Figure 2.1.2). By anticipating the possible classes
of information required by the simulation kernel,
the Cyclus APIs facilitate information passing between the plug-in agents and the core framework.
Though common in modern software architecture,
such a plug-in paradigm has not previously been
implemented in a nuclear fuel cycle simulator. It
allows the core Cyclus framework to operate independently from the plug-in libraries, and the dynamically loadable plug-ins to be the primary mechanism
for extending Cyclus’ capabilities independent of
the core.
An additional benefit is the ability for contributors to choose different distribution and licensing
strategies for their contributions. Users and modelers control the accessibility of their archetypes and
data sets (See Figure 2.1.2). In particular, since the
clean plug-in architecture loads libraries without any
modifications to the Cyclus kernel, closed-source
archetypes can be used with the simulator alongside
open source archetypes without transfer of sensitive
information. This architecture allows closed-source
libraries (e.g., those representing sensitive nuclear
processes and subject to export control) to be developed and licensed privately.
Finally, dynamically loadable libraries enable Cyclus to easily handle varying levels of simulation
complexity. Hence a single simulation engine can be
Figure 1: The Cyclus core provides APIs that abstract away
the details in the kernel and allow the archetypes to be loaded
into the simulation in a modular fashion.
Figure 2: The Cyclus framework enables fully open, partially
open, and fully closed collaborations[36].
6
used by both users interested in big-picture policy
questions as well as users focused on detailed technical analyses. They merely choose their preferred
level of modeling depth from among the available
libraries in the ecosystem.
Critically, this novel functionality enables the comparison between agent implementations. For example, an archetype that inherits the Region interface,
as in Figure 3, is interchangeable with any other
Region agent.
cyclus::StateWrangler
2.2. Agent-based Paradigm
Cyclus implements an agent-based modeling
paradigm. ABM enables model development to
take place at an agent level rather than a system
level. In the nuclear fuel cycle context, for example,
an analyst can design a reactor agent that is entirely
independent from a fuel fabrication agent. Defining
the behavior of both agents according to the API
contract is sufficient for them to interact with one
another as bona fide agents in the simulation. The
two archetype libraries can be used in the same
simulation without any shared knowledge, allowing
modelers to construct a simulation from building
blocks of many types and origins.
Furthermore, the ABM paradigm is more flexible and intuitive (from both a developer and user
perspective) than the system dynamic approach
used in current simulators. System dynamics is a
popular approach for modeling nuclear fuel cycles
[37, 23, 13, 24]. Formally however, system dynamics
models are simply a strict subset of agent-based
models [3]. That is, any system dynamics model
can be translated into an agent-based model. ABM
techniques therefore enable a broader range of simulations in a more generic fashion.
cyclus::Ider
cyclus::Agent
cyclus::Facility
cyclus::Agent
cyclus::Region
cyclus::Dummy
Figure 3: Inheriting Cyclus class interfaces, such as the
Agent, Facility, Institution, and Region classes, abstracts
away unnecessary details while exposing powerful functionality. In the above example, the Dummy archetype simply
inherits from Region in order to become a bona fide Regiontype Agent.
In this way, a researcher can directly compare
two different reactor modeling implementations
(perhaps the imaginary classes DetailedReactor
and SimpleReactor) simply by exchanging the two
corresponding archetypes. That is, two reactor
archetypes both inheriting from the Facility class
are indistinguishable from a simulation perspective.
This can be done with any agent type, where agents
can be “Regions,” “Institutions,” or “Facilities.”
2.2.2. Regions, Institutions, and Facilities
Cyclus provides a novel representation of entities in the nuclear fuel cycle that reflects the reality
in international nuclear power: facilities implementing individual fuel cycle technologies, institutions
managing those facilities, and regions providing geographical and political context for institutions and
facilities. While few simulators have provided any
notion of static regional effects [2, 31], Cyclus allows for both regions and institutions to be firstclass agents in simulated fuel cycles. The fundamental interactions for each entity are implemented
in a corresponding archetype class in Cyclus, i.e.,
the Region class, Institution class, and Facility class. Archetype developers can then build on
the provided functionality by inheriting from the
appropriate class. Cyclus implements a RegionInstitution-Facility (RIF) relationship through a
parent-child hierarchy where regions are the parents
of institutions which are, in turn, the parents of
2.2.1. Agent Interchangeability
ABM is inherently object-oriented because agents
represent discrete, independently acting objects.
Figure 2.1.2 illustrates the modular nature of Cyclus archetypes. The core of the Cyclus simulator creates a set of classes on which agent plugins are based. Agent plug-ins utilize the generic
core APIs that define agent-to-agent interaction as
well as agent-to-environment interaction. For example, they use the resource exchange paradigm
API for trading resources with one another. For
the archetype developer, these interfaces provide
enormous power simply. The API abstracts away
details unnecessary to specifying the archetype behavior, while providing all necessary functionality
for interacting with the Cyclus simulation kernel.
For the user, multiple archetypes that inherit the
same APIs are interchangeable in a simulation.
7
2.3.1. Resources and Materials
Another such use case seeks to capture system
vulnerability to material diversion. Provenance and
trade-history of distinct materials is the fundamental
information unit in such studies, and so this type
of analysis requires discrete simulation of a target
facility and the individual materials modified within
it. Material risk analysis, therefore, demands that
both facilities and materials should be discretely
modeled objects like those in Cyclus.
In Cyclus, agents can transfer discrete resource
objects among one another. Cyclus supports two
types of resources:
facilities. In other words, RIF hierarchies form a
directed acyclic graph (DAG)1 , with regions as root
nodes and facilities as leaf nodes.
Two primary consequences arise from this structure. First, institutions are nominally responsible
for deploying and decommissioning facilities. Accordingly, advanced logic regarding building and decommissioning can be implemented on top of those
behaviors inherited from the Institution interface.
Second, the Facility class implements the Trader
interface to participate in resource exchange, and
institutions and regions, respectively, can adjust the
resource flow preferences of their managed facilities.
Importantly, this capability allows for the modeling
of preferential regional trading of resources (e.g.,
tariffs) as well as preferential institutional trading
(e.g., long-term contracts).
• materials: these represent typical nuclear materials with nuclide compositions;
• products: these can represent any user-defined
measure: carbon credits, build permits, employees, etc..
2.3. Discrete Objects
Cyclus models facilities, institutions, regions,
and resources as discrete objects. A discrete resource model allows for a range of modeling granularity. In the macroscopic extreme, it is equivalent
to time-stepped continuous flow. In the microscopic
extreme, the model is capable of representing arbitrarily small material objects at isotopic resolution.
In this way, Cyclus is applicable across the full
range of modeling fidelity.
Fleet-based, lumped-material models do not distinguish between discrete facility entities or materials. However, some questions require resolution at
the level of individual facilities and materials. As
a result, many detailed performance metrics cannot be captured with previously existing fleet-based
models. For example, meaningful models of spent
nuclear fuel storage transport, and disposal strategies, require representation of discrete casks and
their varying isotopic compositions.
For all of the reasons that the ABM paradigm
in 2.2 enables novel simulations, multiple use cases
require that these agents, such as the regions, institutions, and facilities in Cyclus, must be represented
as discrete objects. For instance, tracking of individual shipments is only viable if materials and
resources are tracked as discrete objects.
All operations performed on every resource object (splitting, combining, decay, etc.) are tracked
in detail as they are performed. This information
includes the agent that created each resource when
it was introduced into the simulation. The parentage of each resource is also tracked. This makes it
possible to follow the history of resources as they
are transferred between agents.
The Cyclus kernel has built-in experimental support for decay calculations. Materials store the time
since their last decay and agents are free to invoke
the decay function on them as desired to decay them
to the current simulation time. Cyclus can operate
in 3 decay modes, with 1 additional mode likely to
be added in a future release:
• “manual” (currently implemented) is the default mode where agents decay materials when
requested by an archetype,
• “never” (currently implemented) globally turns
off all decay. The Material decay function does
nothing,
• “lazy” (currently implemented) decays material
automatically whenever its composition is observed (e.g. when an agent queries information
about a material’s 239 Pu content),
1 DAGs are common graph theoretic structures whose
most important feature is a lack of cycles, i.e., there is a
single path from the root node to any other node in the
graph.
• “periodic” (future) automatically decays all
materials in a simulation with some fixed frequency.
8
When decay is invoked, a material checks to see
if it contains any nuclides with decay constants
that are significant with respect to the time change
since the last decay operation. If none of the decay
constants are significant, no decay calculation is performed and the material remains unchanged. This
error does not accumulate because the next time
the material’s decay function is invoked, the time
change will be larger.
Cyclus has no notion of “tracked” versus “untracked” nuclides. In Cyclus, the composition of a
material is represented by an arbitrarily large list
(potentially thousands) of nuclides. Agents are free
to treat nuclides present in materials any way they
please - including ignoring them. It is the responsibility of archetype developers to choose how to
handle potentially full-fidelity compositions.
In large simulations, many material objects may
change frequently. Material decay can also contribute significantly to such changes. In order to
help avoid unnecessary runtime performance and
database size impacts, compositions in Cyclus have
some special features. In particular, compositions
are immutable once created. This allows multiple material objects to hold references to the same
composition safely. Additionally, new compositions
resulting from decay are cached and used to avoid
redundant decay calculations. Figure 4 illustrates
how this decay history cache works. Composition
immutability in concert with decay history caching
help eliminate many redundant calculations in addition to reducing the total number of composition
entries recorded in the database.
A
dt=1
decay A +1
B
dt=2
decay A +4
decay B +3
decay A +3
decay B +2
C
dt=1
D
Figure 4: A simple decay line cache holding compositions
A, B, C, and a yet-uncomputed composition D. B comes
from decaying A 1 time step. C comes from decaying B 2
time steps, etc. Black represents the cache for this particular
composition chain. Blue indicates decay operations that can
be satisfied by cache lookups. If A needs to be decayed 1 time
step (A +1), a quick lookup returns the previously computed
B. Decaying B 3 time steps requires a decay calculation to
compute a new composition D, but subsequent requests such
as decaying A 4 time steps will not require any recalculation.
question, i.e., the same procedure is used for both
Materials and Products.
The information-gathering step begins by polling
potential consumers. Agents define both the quantity of a commodity they need to consume as well
as the target isotopics, or quality, by posting their
demand to the market exchange as a series of requests. Users may optionally parameterize the agent
to associate a collection of demand constraints with
each collection of requests. Collections of requests
may be grouped together, denoting mutual requests
that represent demand for some common purpose.
For example, a reactor may request UOX and MOX
fuel mutually, indicating that either will satisfy its
demand for fuel.
Suppliers then respond to the series of requests
with a bid. A bid supplies a notion of the quantity
and quality of a resource to match a request. Suppliers may add an arbitrary number of constraints
to accompany bids. For example, an enriched UOX
supplier may be constrained by its current inventory
of natural uranium or its total capacity to provide
enrichment in Separative Work Units (SWUs). It
attaches such constraints to its bids.
Any potential resource transfer, i.e., a bid or a
request, may be denoted as exclusive. An exclusive
2.3.2. Dynamic Resource Exchange (DRE)
The Cyclus simulation paradigm allows discrete
agents, based on archetypes about which the kernel
has no knowledge, to enter the simulation at arbitrary times and trade in discrete resources. These
resources are not defined a priori. Therefore, the
logic engine defining resource interaction mechanisms among agents is crucial. The DRE, described
in detail in [35], is that critical logic engine on which
Cyclus simulations are built. Supporting the general Cyclus philosophy, facilities are treated as
black boxes and a supply-demand communication
framework is defined.
The DRE consists of three steps: supply-demand
information gathering, resource exchange solution,
and trade execution. Importantly, each step is agnostic with respect to the exchange of resources in
9
transfer excludes partial fulfilment; it must either be
met fully or not at all. This mode supports concepts
such as the trading of individual reactor assemblies.
In combination with the notion of mutual requests,
complex instances of supply and demand are enabled.
Finally, requesting facilities, institutions and
regions may apply preferences to each potential
request-bid pairing based on the proposed resource
transfer. Facilities can apply arbitrary complex logic
to rank the bids that they have received, whether
based on the quantity available in each bid or on the
quality of each bid, and the consequent implications
of the physics behavior of that facility. In addition,
an institution can apply a higher preference to a
partner to which it is congenial; similarly, a region
may negate any transfers of material which have a
higher uranium enrichment than is allowable.
[35]:
min z = c> x
(1)
s.t. Ax ≤ b
(2)
x
xi,j ∈ [0, x̃j ] ∀i ∈ I, ∀j ∈ J
(3)
where
I = set of supply nodes
J = set of request nodes
ci,j = unit cost of flow from i to j
xi,j = flow from node i to node j
aki,j = coefficient for constraint k between i and j
bks|r = RHS for constraint k for a given supplier or requester
x̃j = requested quantity of request node j.
In practice, LPs solve much faster than MILPs.
However, both capabilities exist in Cyclus in order
to provide users with the desired level of fidelity.
Trades between agents are initiated by the Cyclus kernel after a solution to the DRE is found.
For each trade, the supplying agent is notified of
its matched request and provides an associated resource to the exchange. All supplied resources are
then sent to the corresponding requesting agents.
In Cyclus, the DRE is executed at each time
step. Therefore, if a facility’s request for a resource
is not met at a given time step, it may offer a request
in the following time step. Because agent behavior
may change arbitrarily, the exchange executed at
any given time step may be unique in a simulation.
The DRE is a novel simulation concept in the
nuclear fuel cycle domain. It provides a flexible
supply-demand infrastructure, supporting dynamic
flows of resources between agents, even as those
agents enter and leave the simulation, and even when
those agents are defined by archetypes of arbitrary
complexity. Trading between agents can be affected
by both the proposed quality of a resource and
agent relationships through the use of preferences.
Accordingly, a wide range of possible effects can
be modeled, from capacity-limited fuel supply to
international trade agreements.
Figure 5: The flow graph representing three suppliers (left),
two requesters (right), and the potential flows of various
commodities among them. The second consumer makes
two different requests. Meanwhile, the second supplier can
supply the commodities requested by both consumers and
has provided two bids accordingly [35].
Given a full definition of supply and demand, represented in Figure 5 as a flow graph, the DRE may
be solved either optimally using a mathematical
program or approximately by a simulation-based
heuristic [35]. If any trade is denoted as exclusive,
e.g., if an analyst desires an assembly-fidelity model,
then either a heuristic must be used or exchanges
must be represented as a mixed-integer linear program (MILP). If no exclusive trades exist, a linear
program (LP) model may be used. The LP formulation in Cyclus is of the form given by Gidden
2.4. Simulation Support
So that users and developers can build working simulations in the shortest time possible, the
10
Cyclus ecosystem provides fundamental building
blocks: basic archetypes and a toolkit of commonly
needed functions. The Cycamore library provides
a suite of fundamental Region, Institution, and Facility archetypes, while the Cyclus toolkit provides
assistance to developers.
classes provide the basis for in-simulation deployment determination: CommodityProducer, CommodityProducerManager, Builder, BuildingManager.
The CommodityProducer class provides an interface
for querying the prototypes which have the capacity
to produce a given commodity. The CommodityProducerManager provides an interface for registering
CommodityProducers and querying the current capacity (supply) of a commodity. The Builder class
provides an interface for querying which prototypes
can be built and interacts with the BuildingManager, which orders prototypes to be built. The
BuildingManager uses a simple minimum cost algorithm to determine how many of each prototype,
yi , to build given a demand (Φ), capacities (φi ),
and costs (ci ). Here i indexes I available prototypes which perform a similar function, and the
demand, capacity and cost carry prototype-specific
units which are defined by the developer.
2.4.1. Cycamore
Cycamore [38], the Cyclus additional module
repository, provides a fundamental set of dynamically loadable libraries providing agent archetypes
for basic simulation functionality within Cyclus.
Since Cyclus relies on external archetypes to represent the agents within a simulation, Cycamore
provides the basic archetypes a new user needs
to get started running simple simulations. These
archetypes support a minimal set of fuel cycle simulation goals and provide, by example, a guide to
new developers who would seek to contribute their
own archetypes outside of Cycamore.
As of version 1.0, Cycamore contains one region
archetype, two institution archetypes, and four facility archetypes. Three additional facilities were
added in version 1.3 . Short descriptions of these
functions can be found in Table 1.
Section 3.2 will demonstrate how the current
Cycamore release provides basic functionality enabling simple fuel cycle analyses. As future contributions are vetted, the capabilities in Cycamore
may grow.
min
N
X
ci yi
i=1
s.t.
N
X
(4)
φi yi ≥ Φ
i=1
yi ∈ [0, ∞) ∀i ∈ I, yi integer
Nuclear Engineering Tools. The Cyclus toolkit
provides two useful interfaces for querying physical
parameters of Material objects. First, the MatQuery tool provides a basic querying API, including
the atom and mass fractions of nuclides, the number of moles of a nuclide in a material, etc. (i.e., a
Composition), in a material. The Enrichment tool
provides an API for determining enrichment-related
parameters of a material, including the SWU and
natural uranium required to enrich a material provided knowledge of feed, product, and tails assays.
2.4.2. Toolkit
In addition to the core functionality of the Cyclus kernel, which is focused on the set of capabilities needed to implement an agent-based simulation
with DRE, a toolkit is provided to assist developers
and users with related simulation and nuclear engineering tasks. The toolkit is an actively developed
part of Cyclus, with a primarily forward-looking focus on supporting interesting in situ metric analysis
tools.
Toolkit Extensions. In addition to those that already exist, new tools will emerge from the
archetypes developed by the community. As these
tools gain adoption between projects and demonstrate their utility to the developer community, they
will be considered for screening and adoption into
the kernel as toolkit extensions. Likely extensions
include
Simulation Tools. A series of utility classes are provided to support demand-constrained agent facility
deployment. For example, symbolic function representations of linear, exponential, and piecewise
functions are supported via the SymbFunctionFactory class. Such functions are used with other
toolkit classes to determine commodity demand
(e.g., power demand) from user input. Four mix-in
• fuel cycle metrics calculators,
• supportive data tables,
• policy models,
11
Entity
Facility
Facility
Facility
Archetype
EnrichmentFacility
Fab∗
Reactor∗
Facility
Facility
Separations∗
Sink
Facility
Source
Institution
ManagerInst
Institution
DeployInst
Region
GrowthRegion
Functionality
This facility enriches uranium at a specified capacity.
This facility fabricates fuel material based on separated streams.
A reactor model that handles batch refueling, based on pre-determined
recipes of compositions. It requests any number of input fuel types
and transmutes them to static compositions upon discharge.
This facility separates an incoming material into specified streams.
This facility is capable of accepting a finite or infinite quantity of some
commodity produced in the simulation.
This facility generates material of the composition and commodity
type specified as input.
The manager institution manages production of commodities among
its facilities by building new ones as needed.
This institution deploys specific facilities as defined explicitly in the
input file.
This region determines whether there is a need to meet a certain
capacity (as defined via input) at each time step.
Table 1: The Archetypes in Cycamore seek to cover a large range of simple simulation use cases [38]. Facilities added in version
1.3 are marked with ∗ .
• and economic models.
Verficiation of Cyclus implementation relies on
software development best practices such as version
control, testing, continuous integration, documentation, and style guidelines to ensure reliability and reproducibility. Sections 2.5.1-2.5.3 discuss in greater
detail the software development components that
comprise the Cyclus verification strategy, which
allows the simulation kernel and physics models to
be tested explicitly and separately. Each of these approaches on its own is a valuable addition to QA but
it cannot be the entire answer to the requirements
imposed by NQA-1. Taken together and strictly adhered to, they present a fortress to protect against
poorly designed or otherwise undesirable code.
2.5. Quality Assurance
To garner the trust of a broad user and archetype
developer community, the Cyclus project must implement a strategy to assure the ongoing quality of
the software it provides. Multiple strategies, collectively known as quality assurance (QA), have been
developed by the scientific software development
community to mitigate structural and algorithmic
errors in software. These include verification and
validation (V&V) [39], testing, and others.
Nuclear engineering software quality is often governed by Nuclear Quality Assurance - 1 (NQA-1), an
American Society of Mechanical Engineers (ASME)
specification whose latest revision appeared in 2009
[40]. Cyclus has adopted an agile development process [41], interpreting NQA-1 in a manner similar
to the process adopted by the Department of Energy (DOE) within Nuclear Engineering Advanced
Modeling and Simulation (NEAMS) [42] or by the
PyNE toolkit [43].
Validation of simulators like Cyclus, that are
intended to give forecasts of system behavior in
uncertain futures, is not well defined as there are
no experimental benchmarks upon which to rely.
Instead, code-to-code comparisons with fuel cycle
simulators that use other modeling paradigms are
underway as the best approach to establish confidence that Cyclus produces correct answers. [44]
2.5.1. Version Control
Automated version control, provided by git, a
well-established distributed version control tool [45],
is at the heart of the QA strategy because it records
the full history of any change, along with metadata
such as the author, a timestamp, and an accompanying message. This makes it possible to identify when
changes were introduced, how they are related to
other changes, and who made those changes. If necessary, it also facilitates reversing individual changes
from a long and intricate set of changes. Version
control also enables a code review process descried
in section 2.5.3 below.
12
2.5.2. Testing
Automated software testing is the first line of
defense against errors in implementation. Testing
directly compares the actual results of running software versus the expected behavior of the software.
In Cyclus, three categories of tests are defined:
unit tests, integration tests, and regression tests.
Before a proposed code change is allowed into Cyclus, the change must be covered by a test, either
new or existing, and all tests must pass.
version of the Cyclus source code that includes
the requested changes and runs the complete test
suite, reporting back whether or not those steps
were successful. If CI was not successful, the code
author(s) must first identify and resolve the problems. These steps are performed for all incoming
code prior to inclusion, so broken code never enters
the main software development branch.
2.5.3. Code Review
Unit Tests. Unit tests verify behavior of the smallest code unit, typically a single function or a class.
Cyclus uses the Google Test framework [46] as a
harness for running unit tests. Sufficient unit tests
are required for any proposed change to the Cyclus
code base. Currently, Cyclus implements over 450
unit tests and Cycamore implements 85. These
cover approximately 65% of their respective code
bases, and these numbers are expected to grow over
time.
Automated testing including CI is a necessary
but not sufficient component of the Cyclus QA
system: it keeps bad code out of Cyclus. However,
Cyclus will always require human eyes and hands
to let good code in.
The main Cyclus repository is hosted remotely
and publicly on the GitHub website [49], in part
because it provides tools that facilitate a culture
of frequent and thorough code review. A number
of policies exist to ensure that a proposed set of
changes, known as a pull request, adhere to the
projects QA standards. Every pull request must be
reviewed and accepted by a member of the Cyclus
core team that was not involved in authoring the
changes. In addition to reviewing the algoirthmic
design in the source code, the reviewer relies on tests
to ensure correct behavior, and requires the authors
to adhere to the style guide and provide sufficient
documentation. Only after the QA standards have
been met are the proposed changes merged into the
software. This step has been repeatedly shown to
improve code quality [4].
Once CI is successful, a code reviewer inspects not
only the changes that are proposed to the software
itself, but also the changes that have been proposed
to tests.
Because of their long term benefit to the maintainability of the project, documentation and coding
style are also reviewed as part of this process. The
API must be documented as required by the Cyclus QA policy. In Cyclus, this information is
aggregated together into static websites with the
Doxygen [50] and Sphinx [51] tools, and can be
accessed at http://fuelcycle.org. Cyclus also
strictly enforces the use of the Google C++ Style
Guide [52] for all software contributions. This means
that all developers of Cyclus write Cyclus code
in the same way. This homogenization may be a
hurdle to new developers but ultimately improves
code legibility and, therefore, robustness [4].
Integration Tests. Integration tests combine multiple elements of the Cyclus interface and test that
they work correctly with each other. In Cyclus
and Cycamore, integration tests are performed by
running sample simulations for scenarios verifying
that results match predictions. A set of standard
input files are run, then the output is inspected and
compared via Nose [47], a Python test framework.
Regression Tests. Regression tests ensure significant
unintended changes do not occur over the course
of Cyclus development. Regression tests are implemented similarly to integration tests. In this
category, however, the comparison is done against
the output of the same input file when run with a
previous version of Cyclus, typically the last released version. In some sense, regression tests are
‘dumb’ in that they do not care about the contents
of a simulation being correct, only whether or not
it changed.
Continuous Integration. Continuous integration
(CI) is the idea that software should be automatically tested and validated as it is being developed,
rather than as a final stage in a longer development
cycle. The results of this testing are shown to code
reviewers prior to reviewing the software changes
themselves. The Cyclus project uses the CircleCI
[48] service for continuous integration. When a code
change is submitted for review, CircleCI builds a
13
3. Demonstration
Already, the ecosystem is growing. Early crossinstitutional contributions to the ecosystem demonstrate a significant achievement by the Cyclus
framework and provide the basis for a communitydriven development model.
The success of the Cyclus simulator can be measured in many ways. The most compelling are
early successes of the community-driven development model and demonstrations of its fundamental simulation capabilities. Promising growth of
the Cyclus ecosystem at multiple institutions indicates that a fuel cycle simulator can advance in
this community-driven development paradigm. Additionally, simulation results for both once-through
and more complex recycling scenarios demonstrate
that Cyclus possesses the fundamental fuel cycle
simulation capabilities to contribute to the field.
3.1.1. Supplementary Projects
A number of projects and tools outside of the
core simulation kernel have been developed to improve the scope and the diversity of the capabilities
in the Cyclus ecosystem. Table 2 lists the tools
and projects developed under close integration with
the Cyclus kernel. These tools are used to ease
development and simulation design (Cycstub, Cycic, Ciclus), data visualization and analysis (Cyclist,
Cyan), and remote execution (Cloudlus).
3.1. Ecosystem
The Cyclus community-driven software development model seeks to break the proliferation of
specialized simulators. It instead leverages the collective expertise of fuel cycle researchers toward a
single, more extensible, tool. Through the targeted
contributions of those researchers, an ecosystem of
capabilities should emerge. The Cyclus ‘Ecosystem’ is the collection of tools, calculation libraries,
archetypes, data, and input files intended for use
with the Cyclus simulator. Members of the ecosystem include:
3.1.2. Archetype Contributions
It is expected that the most common type of
contribution to Cyclus will be contributions of new
archetypes. Researchers will be driven to create a
new archetype when a need arises, such as to improve
the fidelity of simulation, or to represent a novel
reactor type, an innovative reprocessing strategy, or
a particular governmental or institutional behavior.
The real-world utility of Cyclus can in part be
measured by the breadth and diversity of archetypes
being developed in this way.
Early progress has been promising.
Many
archetypes external to the Cycamore library (Table 3) have been [59, 60] or are being [61, 62, 63]
developed for contribution to the Cyclus ecosystem. These archetypes provide the first examples of
developer-contributed capabilities. They add to the
fundamental Cycamore archetypes by providing
physics-based reactors, separations logic, fuel fabrication processing, storage facilities, and expanded
institutional paradigms. The existence and diversity
of these contributed archetypes illustrate the power
and potential of the community-based development
approach that Cyclus has taken.
• the archetypes provided in the Cycamore [38]
repository
• the archetypes created by researchers
• isotopic composition data
• historical facility deployment data
• the Cyclus graphical user interface (GUI) tool
Cyclist
• fundamental analysis tools in the Cyclus
toolkit
• tools for Cyclus optimization, parallelization,
and development
3.2. Simulations
To illustrate the flexibility of Cyclus, this section will discuss the creation and results of a range
of representative fuel cycle simulations. Previous
benchmarks between Cyclus and system dynamics
simulators for more complex problems, including
transition analyses, have been reported elsewhere
and Cyclus has demonstrated satisfactory performance [64, 65, 35, 66] . The examples presented here
Taken together, these form an ecosystem of capabilities. Over time, this ecosystem will grow as
archetype developers, kernel developers, and even
users contribute capabilities developed for their own
needs. Indeed, the long-term vision for the Cyclus
framework predicts an ever-expanding ecosystem of
both general and specialized capability extensions.
14
Name
Cycic
Cyclist
Ciclus
Cycstub
Cyan
Cloudlus
Description
Input control embedded in Cyclist
Interactive data exploration environment
Continuous integration scripts for Cyclus
Skeleton for clean slate module development
Cyclus analysis tool
Tools for running Cyclus in a cloud environment
Ref.
[53]
[54]
[55]
[56]
[57]
[58]
Table 2: Many tools have been developed outside of the scope of the Cyclus kernel for improved user, developer, and analyst
experiences with Cyclus.
Name
Bright-lite
Nuclear Fuel Inventory Model
CommodConverter
MktDrivenInst
SeparationsMatrix
StreamBlender
Description
A physics-based reactor archetype and fuel fabrication archetype
A flexible, ORIGEN-based, reactor analysis module
A simple commodity converting storage facility archetype
An institution that controls deployment based on commodity
availability
A facility for elemental separations of used fuel
A facility for fuel fabrication from reprocessed streams
Ref.
[61]
[62]
[60]
[63]
[59]
[59]
Table 3: A diverse set of archetypes under development reflect the diverse needs of researchers at various institutions. These
archetypes, contributed outside of the Cyclus core and Cycamore libraries are the first demonstration of community-driven
development in a fuel cycle simulator.
are not the limit of Cyclus’ capabilities, which are
extended with each new addition to the ecosystem,
as discussed above. Rather, these simulations are
designed to illustrate that Cyclus matches the capabilities of any recipe-based simulator and that
variations on a single Cyclus simulations can be
run with small changes to the input specification.
For simplicity of the current demonstration, many
simplifying assumptions have been made with respect to material compositions, fuel transmutation,
among others. The three fuel cycles examined are:
1. No Recycle (once through)
2. 1-pass MOX Recycle
3. Infinite-pass MOX Recycle
For each of these fuel cycles, a 1,100 month singlereactor Cyclus simulation was run in addition to a
10-reactor simulation with staggered refueling times.
As the number of staggered-cycle reactors increases
the system converges toward continuous material
flow results. Cyclus flexibility allows this transition
to be examined.
15
Facility
LWR
cycamore::Reactor
Spent Fuel Fabrication
cycamore::Fab
Separations
cycamore::Separations
Repository
cycamore::Sink
UOX Fabrication
cycamore::Source
DU Source
cycamore::Source
Parameter
type
batches
batch size
cycle length
refuelling time
requests
requests
requests
requests
offers
requests
offers
efficiency
Pu separation capacity
requests
capacity
offers
capacity
offers
capacity
Value
LWR
3 [batches/cycle]
20 [MTHM]
18 [months]
2 [months]
recycled MOX (1st preference)
enriched UOX (2nd preference)
depleted uranium
separated fissile material
recycled MOX
all spent fuel types
separated fissile material
0.99
6.0E4 [kg/month]
all waste
∞
UOX
∞
depleted U
∞
Table 4: Facility Configurations for Example Simulations
16
As described in Table 4, the facilities used in these
simulations are:
oped by the French Commissariat à l’Énergie
Atomique et aux Énergies Alternatives (CEA).
• Separations
(cycamore::Separations):
This facility takes in all kinds of spent fuel and
separates it into plutonium and waste streams
with some efficiency (0.99 was used for these
simulations). Up to 60,000 kg of Pu can be
separated per month.
• Repository (cycamore::Sink): This is an
infinite-capacity facility that can take in all
types of material including separations waste
streams and spent reactor fuel.
• Reactor (cycamore::Reactor): This is a reactor model that requests any number of input
fuel types and transmutes them to static compositions when they are burnt and discharged
from the core. The reactor was configured to
model a light water reactor with a 3 batch core
operating on an 18 month cycle with a 2 month
refuel time. The batches in the core each contain 20 metric tons of heavy metal (MTHM),
where heavy metals are actinide elements like
thorium, uranium, and plutonium. The reactor
was also configured to take in either enriched
UOX fuel or recycled MOX fuel.
• UOX Fabrication (cycamore::Source):
This facility provides enriched UOX fuel as
requested. This facility has infinite production
capacity.
• DU Source (cycamore::Source): This facility provides depleted uranium as requested.
This facility has infinite production capacity.
To model each of the three fuel cycles, only simple
adjustments to the input file specification were necessary. Specifically, only the commodity types and
trade preferences for each of the facilities needed to
be altered from one simulation to another.
For all cases, the reactor was configured to request recycled MOX fuel with a higher preference
than new UOX fuel. For the No Recycle case, the
Reactor was set to offer all spent fuel as waste. For
the 1-pass recycle case, the Reactor offered spent
UOX into a spent fuel market, but spent MOX is
still offered as waste. In the Infinite-Pass Recycle
case, the Reactor offers all burned fuel into a spent
fuel market. The separations facility requests spent
fuel with a higher preference than the repository resulting in preferential recycle. All these preferences
are easy to adjust and Cyclus dynamically handles
supply constraints and non-uniform preference resolution. It is notable that separations and recycle
fuel fabrication capacity are deployed identically in
all simulations. In the once through case, the recycling loop never acquires material, and so reactors
always receive fresh UOX fuel. The DRE ensures
everything operates smoothly in all cases.
Cyclus’ discrete materials make single-pass recycle straightforward to implement. The reactors
keep track of fuel as discrete batches. A reactor
remembers where it received each batch from. If a
batch was received from the recycled fuel fabrication facility, it does not offer it to separations and
• Spent Fuel Fabrication (cycamore::Fab):
This facility requests depleted uranium and
separated fissile material and mixes them together into recycled MOX fuel that it offers to
requesters. The two streams are mixed in a
ratio in order to match simple neutronics properties of the target fuel as closely as possible.
The method used is based on a variation “equivalence method” originally developed by Baker
and Ross [67]. This technique has also been
used in the COSI fuel cycle simulator devel17
DU Source
DU Source
depleted U
(7.11e+05 kg)
depleted U
(7e+05 kg)
Spent Fuel Fab
Spent Fuel Fab
Enrichment
Enrichment
MOX fuel
(1.2e+05 kg)
MOX fuel
(1e+05 kg)
UOX fuel
(1.06e+06 kg)
Pu stream
(1.38e+04 kg)
Pu stream
(1.08e+04 kg)
LWR
LWR
spent fuel
(1.08e+06 kg)
spent Fuel
(1e+06 kg)
waste
(0 kg)
Separations
waste
(8e+04 kg)
Separations
UOX fuel
(1.04e+06 kg)
waste
(1.07e+06 kg)
Repository
waste
(9.89e+05 kg)
Repository
Figure 7: Full MOX recycle (multi-pass) fuel cycle material
flows.
Figure 6: 1-pass MOX recycle scenario material flows.
and infinite-pass cases, enough separated fissile material accumulates in the fuel fabrication facility to
generate a full recycled batch. When this batch is
transmuted, more plutonium is burned than created.
This results in a drop in the total fuel cycle system
plutonium inventory. This pattern repeats roughly
every 10 cycles (200 months) for the Single-Pass
Recycle case and every 9 cycles (180 months) for
the Infinite-Pass Recycle case. Because the 1-pass
recycle scenario does not re-recycle material, it takes
the fabrication facility 2 cycles longer to accumulate
a full batch of fissile material.
Because facilities are represented individually and
transact discrete materials as discrete events, realistic non-uniform patterns in facility behavior that
affect total system behavior are observed using Cyclus.
Figure 9 shows plutonium buildup for the 10reactor simulations of the No Recycle, Single-Pass
Recycle, and Infinite-Pass Recycle scenarios. As
the number of reactors (with staggered refueling)
increases, the behavior of the system approaches
a more steady average reminiscent of continuous
material flow models.
The fundamental capabilities of demonstrated in
these simulations qualify Cyclus and its ecosystem
to model the breadth of scenario types expected of
instead offers it as a waste commodity which is only
requested by the repository. The material flows for
the Single-Pass Recycle case are shown in Figure 6.
Changing the scenario from a 1-pass fuel cycle
to an infinite-pass fuel cycle requires only a oneword change in the input file regarding the output
commodity for the spent MOX fuel of the Reactor:
< fuel >
< incommodity > mox </ incommodity >
< outcommodity > waste </ outcommodity >
+
< outcommodity > spent_fuel </ outcommodity >
< inrecipe > mox_fresh_fuel </ inrecipe >
< outrecipe > mox_spent_fuel </ outrecipe >
</ fuel >
This results in the material flows in Figure 7. A
similarly trivial change was used to switch from
the No Recycle to a 1-pass fuel cycle. Note that
because the reactors always transmute fuel into fixed
compositions, the error in isotopic compositions is
larger for the Infinite-Pass Recycle case.
Figure 8 shows the full-system plutonium buildup
for No Recycle (once through), Single-Pass Recycle, and Infinite-Pass Recycle variations of the onereactor scenario described above. The figure was
generated directly from Cyclus output data. After
several batch cycles (near month 300) in the 1-pass
18
once through
1-pass recycle
infinite-pass recycle
10
Pu (metric tonnes)
transparency and community oversight. Furthermore, the object-oriented ABM simulation paradigm
ensures more generic simulation capability. It allows
Cyclus to address common analyses in a more flexible fashion and enables analyses that are impossible
with system dynamics simulators.
Similarly, the fidelity-agnostic, modular Cyclus
architecture facilitates simulations at every level of
detail. Simulations relying on arbitrarily complex
isotopic compositions are possible in Cyclus, as
are simulations not employing any physics at all.
Physics is introduced through optional system-wide
radioactive decay of materials and through the use
of physics-enabled facility libraries. To support calculation of physical processes in nuclear facilities,
the Cycamore library provides models employing
basic physics for core fuel cycle facilities and extension libraries from the community support more
detailed simulations. Indeed, agents of such varying
fidelity can even exist in the same simulation. Researchers no longer need to reinvent the underlying
simulator framework in order to model a simulation
focused on the aspects of the fuel cycle relevant to
their research.
Furthermore, when the capabilities within Cyclus, Cycamore, and the rest of the ecosystem
are insufficient, adding custom functionality is simplified by a modular, plug-in architecture. A clean,
modern API simplifies customization and independent archetype development so that researchers can
create models within their domain of expertise without modifying the core simulation kernel. Throughout the Cyclus infrastructure, architecture choices
have sought to enable cross-institutional collaboration and sustainable, community-driven development. The ecosystem of capabilities, already growing, may someday reflect the full diversity of use
cases in the nuclear fuel cycle simulation domain.
Plutonium Buildup: 1 Reactor
12
8
6
4
2
0
0
200
400
600
Time (month)
800
1000
1200
Figure 8: System plutonium buildup with one reactor.
Plutonium Buildup: 10 Reactors
120
once through
1-pass recycle
infinite-pass recycle
Pu (metric tonnes)
100
80
60
40
20
0
0
200
400
600
Time (month)
800
1000
1200
Figure 9: System plutonium buildup with staggered refueling
for many reactors.
nuclear fuel cycle simulator tools. These examples
further show the flexibility provided by the DRE
logic within the Cyclus framework and provide an
example of the resolution made possible by discretefacility and discrete-material modeling in fuel cycle
simulation.
5. Acknowledgements
This work has been supported by a number of
people. The authors would like to thank software
contributors Zach Welch and Olzhas Rakhimov. Additionally, the authors are grateful to the many enthusiastic members of the Cyclus community and
our Nuclear Energy University Programs (NEUP)
collaboration partners at the University of Texas,
University of Utah, and University of Idaho.
Direct support for the Cyclus project has been
received from the Nuclear Regulatory Commission
(NRC) Faculty Development program, the NEUP
4. Conclusions
The Cyclus nuclear fuel cycle framework
presents a more generic and flexible alternative to existing fuel cycle simulators. Where previous nuclear
fuel cycle simulators have had limited distribution,
constrained simulation capabilities, and restricted
customizability, Cyclus emphasizes an open strategy for access and development. This open strategy
not only improves accessibility, but also enables
19
Research & Development Program, the National
Nuclear Security Administration (NNSA) Consortium for Verification Technology, and the University
of Wisconsin (UW). In addition students contributing to Cyclus have received fellowship support
from the Argonne National Laboratory (ANL) Lab
Grad program, the NEUP Fellowship program, the
National Science Foundation (NSF) Graduate Fellowship program, and the Department of Homeland
Security (DHS) Nuclear Forensics Graduate Fellowship program.
[10] M. Petre, G. Wilson, Code Review For and By Scientists,
arXiv:1407.5648 [cs]00000 arXiv: 1407.5648.
URL http://arxiv.org/abs/1407.5648
[11] M. Flicker, E. A. Schneider, P. Campbell, Evaluation
Criteria for Analyses of Nuclear Fuel Cycles, in: Transactions of the American Nuclear Society, Vol. 111 of Fuel
Cycle Options Analysis – III, American Nuclear Society,
La Grange Park, IL 60526, United States, Anaheim, CA,
United States, 2014, pp. 233–236.
URL http://epubs.ans.org/?a=36342
[12] C. Poinssot, S. Bourg, N. Ouvrier, N. Combernoux,
C. Rostaing, M. Vargas-Gonzalez, J. Bruno, Assessment of the environmental footprint of nuclear energy systems. Comparison between closed
and open fuel cycles, Energy 69 (2014) 199–211.
doi:10.1016/j.energy.2014.02.069.
URL
http://www.sciencedirect.com/science/
article/pii/S0360544214002035
[13] L. Guerin, M. Kazimi, Impact of Alternative Nuclear
Fuel Cycle Options on Infrastructure and Fuel Requirements, Actinide and Waste Inventories, and Economics,
Technical Report MIT-NFC-TR-111, Massachusetts Institute of Technology, Cambridge, MA, United States
(Sep. 2009).
[14] J. Jacobson, A. Yacout, G. Matthern, S. Piet, D. Shropshire, R. Jeffers, T. Schweitzer, VERIFIABLE FUEL
CYCLE SIMULATION MODEL (VISION): A TOOL
FOR ANALYZING NUCLEAR FUEL CYCLE FUTURES, Nuclear Technology 172 (2010) 157–178.
[15] A. M. Yacout, J. J. Jacobson, G. E. Matthern, S. J.
Piet, A. Moisseytsev, Modeling the Nuclear Fuel Cycle,
in: The 23rd International Conference of the System
Dynamics Society,” Boston, 2005.
URL
http://www.inl.gov/technicalpublications/
Documents/3169906.pdf
[16] E. A. Schneider, C. G. Bathke, M. R. James, NFCSim:
A Dynamic Fuel Burnup and Fuel Cycle Simulation Tool,
Nuclear Technology 151 (1) (2005) 35–50.
[17] C. Allan, International Atomic Energy Agency, International Project on Innovative Nuclear Reactors and Fuel
Cycles, Guidance for the application of an assessment
methodology for Innovative Nuclear Energy Systems
INPRO Manual: overview of the Methodology., rev.1
Edition, no. IAEA-TECDOC-1575 in IAEA-TECDOC,
International Atomic Energy Agency, Vienna, 2008.
URL http://www-pub.iaea.org/MTCD/Publications/
PDF/TE_1575_web.pdf
[18] L. Boucher, J. P. Grouiller, ”COSI” : A Simulation
Software for a Pool of Reactors and Fuel Cycle Plants, in:
Fuel Cycle and High Level Waste Management, Beijing,
China, 2005.
[19] L. Boucher, J. P. Grouiller, ”COSI”: The Complete
Renewal of the Simulation Software for The Fuel Cycle
Analysis, in: Fuel Cycle and High Level Waste Management, Vol. 1, ASME, New York, NY, USA, Miami, FL,
United states, 2006.
[20] M. Meyer, L. Boucher, New Developments on COSI
6, the Simulation Software for Fuel Cycle Analysis, in:
Proceedings of GLobal 2009, Paris, France, 2009.
[21] C. Coquelet-Pascal, M. Meyer, R. Girieud, R. Eschbach, C. Chabert, C. Garzenne, P. Barbrault, L. Van
Den Durpel, T. Duquesnoy, M. Caron-Charles, B. Carlier, J. C. Lefevre, Comparison of Different Scenarios
for the Deployment of Fast Reactors in France - Results
Obtained with COSI, in: Proceedings of GLobal 2011,
References
[1] S. J. Piet, B. W. Dixon, J. J. Jacobson, G. E. Matthern,
D. E. Shropshire, Dynamic Simulations of Advanced
Fuel Cycles, Nuclear Technology 173 (3) (2011) 227–
238.
URL http://epubs.ans.org/?a=11658
[2] K. Huff, B. Dixon, Next Generation Fuel Cycle Simulator Functions and Requirements Document, Tech. Rep.
fcrd-sysa-2010-000110, Idaho National Laboratory (Jul.
2010).
[3] C. M. Macal, To agent-based simulation from system
dynamics, in: Simulation Conference (WSC), Proceedings of the 2010 Winter, IEEE, 2010, pp. 371–382.
URL
http://ieeexplore.ieee.org/xpls/abs_all.
jsp?arnumber=5679148
[4] J. Cohen, Modern Code Review, Making Software:
What Really Works, and Why We Believe It (2010)
329–336.
URL
http://books.google.com/books?hl=
en&lr=&id=DxuGi5h2-HEC&oi=fnd&pg=PA329&dq=
cohen+modern+code+review&ots=0VrtsgOdqP&sig=
-XrFj4mAA2MB-20qZXOFx8Felzo
[5] M. Fagan, Design and code inspections to reduce
errors in program development, in: Software pioneers,
Springer, 2002, pp. 575–607.
URL
http://link.springer.com/chapter/10.1007/
978-3-642-59412-0_35
[6] D. L. Donoho, A. Maleki, I. U. Rahman, M. Shahram,
V. Stodden, Reproducible Research in Computational
Harmonic Analysis, Computing in Science & Engineering
11 (1) (2009) 8–18, 00110. doi:10.1109/MCSE.2009.15.
URL
http://scitation.aip.org/content/aip/
journal/cise/11/1/10.1109/MCSE.2009.15
[7] D. C. Ince, L. Hatton, J. Graham-Cumming, The case
for open computer programs, Nature 482 (7386) (2012)
485–488, 00142. doi:10.1038/nature10836.
URL
http://www.nature.com/doifinder/10.1038/
nature10836
[8] V. Stodden, The Scientific Method in Practice: Reproducibility in the Computational Sciences, SSRN Electronic Journal00037. doi:10.2139/ssrn.1550193.
URL http://www.ssrn.com/abstract=1550193
[9] J. M. Wicherts, M. Bakker, D. Molenaar, Willingness
to Share Research Data Is Related to the Strength
of the Evidence and the Quality of Reporting of
Statistical Results, PLoS ONE 6 (11) (2011) e26828.
doi:10.1371/journal.pone.0026828.
URL
http://dx.doi.org/10.1371/journal.pone.
0026828
20
[32] G. Grasso, S. Monti, M. Sumini, NEA-WPFC/FCTS
benchmark for fuel cycle scenarios study with COSI6,
Tech. rep., Report RSE/2009/136 (2009).
URL
http://www.enea.it/it/Ricerca_sviluppo/
documenti/ricerca-di-sistema-elettrico/
nucleare-fissione/studi-accordi/rse136.pdf
[33] L. Boucher, Benchmark Study on Nuclear Fuel Cycle Transition Scenarios Analysis Codes, Tech. Rep.
NEA/NSC/WPFC/DOC(2012)16, OECD, Nuclear Energy Agency (Jun. 2012).
Gidden,
Cyclopts
0.0.10,
Figshare[34] M.
Http://dx.doi.org/10.6084/m9.figshare.1288959. doi:
http://dx.doi.org/10.6084/m9.figshare.1288959.
URL
http://figshare.com/articles/cyclopts/
1288959
[35] M. J. Gidden, An Agent-Based Modeling Framework
and Application for the Generic Nuclear Fuel Cycle,
Ph.D. thesis, University of Wisconsin, Madison, WI,
United States (Mar. 2015).
[36] R. W. Carlsen, M. Gidden, K. Huff, A. C.
Opotowsky,
O. Rakhimov,
A. M. Scopatz,
Z. Welch, P. Wilson, Cyclus v1.0.0, FigshareHttp://dx.doi.org/10.6084/m9.figshare.1041745. doi:
http://dx.doi.org/10.6084/m9.figshare.1041745.
URL http://figshare.com/articles/Cyclus_v1_0_0/
1041745
[37] J. J. Jacobson, R. F. Jeffers, G. E. Matthern, S. J. Piet,
B. A. Baker, J. Grimm, VISION User Guide-VISION
(Verifiable Fuel Cycle Simulation) Model, Tech. rep.,
Idaho National Laboratory (INL) (2009).
URL
http://www.inl.gov/technicalpublications/
Documents/4363869.pdf
[38] R. W. Carlsen, M. Gidden, K. Huff, A. C.
Opotowsky,
O.
Rakhimov,
A.
M.
Scopatz, P. Wilson, Cycamore v1.0.0, FigshareHttp://figshare.com/articles/Cycamore v1 0 0/1041829.
doi:http://figshare.com/articles/Cycamore_v1_0_
0/1041829.
URL http://figshare.com/articles/Cycamore_v1_0_
0/1041829
[39] B. Boehm, Software risk management, Springer, 1989.
[40] ASME, NQA-1a-2009, Addenda to ASME NQA-1-2008,
Quality Assurance Requirements for Nuclear Facility
Applications, no. NQA-1a-2009 in Nuclear Quality Assurance, American Society of Mechanical Engineers, New
York, NY, 2009.
[41] C. Larman, Agile and Iterative Development: A Manager’s Guide, Addison-Wesley Professional, 2004.
[42] NEAMS, NUCLEAR ENERGY ADVANCED MODELING AND SIMULATION (NEAMS) SOFTWARE
VERIFICATION AND VALIDATION (V&V) PLAN
REQUIREMENTS, Tech. rep., U.S. Deptartment of
Energy, Office of Nuclear Energy (Jul. 2013).
URL
http://www.energy.gov/sites/prod/files/
2013/09/f2/NEAMS%20Software%20Verification%
20and%20Validation%20Plan%20Requirements%
20Version%200.pdf
[43] E. Biondo, A. Scopatz, M. Gidden, R. Slaybaugh, P. P.
Bates, Cameron Wilson, Quality Assurance within the
PyNE Open Source Toolkit, in: American Nuclear Society 2014 Winter Meeting, ANS, 2014.
[44] K. D. Huff, M. Fratoni, H. Greenberg, Extensions to
the Cyclus Ecosystem In Support of Market-Driven
Transition Capability, in: Transactions of the American
Nuclear Society, American Nuclear Society, Anaheim,
Makuhari, Japan, 2011.
[22] A. Worrall, R. Gregg, Scenario Analyses of Future UK
Fuel Cycle Options, Journal of Nuclear Science and Technology 44 (3) (2007) 249–256. doi:10.1080/18811248.
2007.9711279.
URL http://www.tandfonline.com/doi/abs/10.1080/
18811248.2007.9711279
[23] L. Van Den Durpel, A. Yacout, D. Wade, T. Taiwo,
U. Lauferts, DANESS V4.2: Overview of Capabilities
and Developments, in: Proceedings of Global 2009, Paris,
France, 2009.
[24] L. Guerin, B. Feng, P. Hejzlar, B. Forget, M. S. Kazimi, L. Van Den Durpel, A. Yacout, T. Taiwo, B. W.
Dixon, G. Matthern, L. Boucher, M. Delpech, R. Girieud,
M. Meyer, A Benchmark Study of Computer Codes for
System Analysis of the Nuclear Fuel Cycle, Technical
Report, Massachusetts Institute of Technology. Center
for Advanced Nuclear Energy Systems. Nuclear Fuel
Cycle Program, electric Power Research Institute (Apr.
2009).
URL http://dspace.mit.edu/handle/1721.1/75245
[25] E. A. Andrianova, V. D. Davidenko, V. F. Tsibul’skii,
Desae program for systems studies of long-term growth
of nuclear power, Atomic energy 105 (6) (2008) 385–390.
URL
http://www.springerlink.com/index/
l581v3n5111475u8.pdf
[26] K. A. McCarthy, B. Dixon, Y.-J. Choi, L. Boucher,
K. Ono, F. Alvarez-Velarde, E. M. Gonzalez, B. Hyland,
Benchmark Study on Nuclear Fuel Cycle Transition
Scenarios-Analysis Codes.
URL
http://www.iaea.org/inis/collection/
NCLCollectionStore/_Public/44/089/44089401.pdf
[27] M. L. Dunzik-Gougar, C. A. Juchau, K. Pasamehmetoglu, P. P. H. Wilson, K. M. Oliver, P. J. Turinsky,
H. S. Abdel-Khalik, R. Hays, T. E. Stover, Global Evaluation of Nuclear Infrastructure Utilization Scenarios
(GENIUS), in: GLOBAL 2007: Advanced Nuclear Fuel
Cycles and Systems, September 9, 2007 - September
13, 2007, GLOBAL 2007: Advanced Nuclear Fuel Cycles and Systems, American Nuclear Society, Boise, ID,
United states, 2007, pp. 1604–1611.
[28] R. Jain, P. P. H. Wilson, Transitioning to global optimization in fuel cycle system study tools, in: 2006
Winter Meeting of the American Nuclear Society, Nov
12 - 16 2006, Vol. 95 of Transactions of the American Nuclear Society, American Nuclear Society, Albuquerque,
NM, United states, 2006, pp. 162–163.
[29] K. M. Oliver, P. P. Wilson, A. Reveillere, T. W. Ahn,
K. Dunn, K. D. Huff, R. A. Elmore, Studying international fuel cycle robustness with the GENIUSv2 discrete
facilities/materials fuel cycle systems analysis tool, in:
Proceedings of GLOBAL 2009, GLOBAL 2009: Advanced Nuclear Fuel Cycles and Systems, Paris, France,
2009.
[30] K. D. Huff, K. M. Oliver, P. P. Wilson, T. W. Ahn,
K. Dunn, R. Elmore, GENIUSv2 Discrete Facilities/Materials Modeling of International Fuel Cycle Robustness,
in: Transactions of the American Nuclear Society, Vol.
101 of Nuclear Fuel Cycle Codes and Applications, American Nuclear Society, Washington D.C., United States,
2009, pp. 239–240.
[31] C. A. Juchau, M. L. Dunzik-Gougar, J. J. Jacobson,
Modeling the Nuclear Fuel Cycle, Nuclear Technology
171 (2) (2010) 136–141.
URL http://epubs.ans.org/?a=10778
21
CA, United States, 2014.
[45] Software Freedom Conservancy, Git, Software Freedom
Conservancy, 2014.
URL http://git-scm.com
[46] G. Inc, googletest - Google C++ Testing Framework,
Technical Report, Google Inc. (2008).
URL https://code.google.com/p/googletest/
[47] J. Pellerin, Nose is nicer testing for python, Technical
Report, nose (2007).
URL https://nose.readthedocs.org/en/latest/
[48] P. Biggar, CircleCI, https://circleci.com/ (2015).
URL https://github.com/circleci/frontend
[49] L. Dabbish, C. Stuart, J. Tsay, J. Herbsleb, Social coding
in GitHub: transparency and collaboration in an open
software repository, in: Proceedings of the ACM 2012
conference on Computer Supported Cooperative Work,
ACM, 2012, pp. 1277–1286.
[50] D. van Heesch, Doxygen: Source code documentation
generator tool, URL: http://www.doxygen.org.
[51] G. Brandl, Sphinx Documentation.
URL http://sphinx-doc.org/sphinx.pdf
[52] B. Weinberger, C. Silverstein, G. Eitzmann, M. Mentovai, T. Landray, Google C++ style guide, h ttp://googlestyleguide. googlecode. com/svn/trunk/cppguide. xml.
[53] R. Flanagan, E. A. Schneider, Input Visualization for
the Cyclus Nuclear Fuel Cycle Simulator: Cyclus Input
Control, in: Proceedings of GLOBAL 2013, Salt Lake
City, UT, United States, 2013.
[54] Y. Livnat, H. Gur, K. Potter, R. Flanagan, Cyclist,
github.com/SCI-Utah/cnome (Oct. 2014).
URL github.com/SCI-Utah/cnome
[55] A. Scopatz, Z. Welch, M. Gidden, Ciclus,
https://github.com/cyclus/ciclus (Oct. 2014).
URL https://github.com/cyclus/ciclus
[56] R. W. Carlsen, M. Gidden, K. Huff, A. C. Opotowsky,
O. Rakhimov, A. M. Scopatz, Z. Welch, P. Wilson, CycStub, https://github.com/cyclus/cycstub (Jun. 2014).
URL https://github.com/cyclus/cycstub
W.
Carlsen,
CyAn,
Figshare[57] R.
Http://dx.doi.org/10.6084/m9.figshare.1041836. doi:
http://dx.doi.org/10.6084/m9.figshare.1041836.
URL http://dx.doi.figshare.com/articles/CyAn_
Cyclus_Analysis_Tools/1041836
[58] R.
W.
Carlsen,
Cloudlus,
https://github.com/rwcarlsen/cloudlus (Oct. 2014).
URL https://github.com/rwcarlsen/cloudlus
[59] K.
D.
Huff,
StreamBlender,
FigshareHttp://dx.doi.org/10.6084/m9.figshare.1134648. doi:
http://dx.doi.org/10.6084/m9.figshare.1134648.
URL http://figshare.com/articles/StreamBlender/
1134648
[60] K.
D.
Huff,
CommodConverter,
FigshareHttp://dx.doi.org/10.6084/m9.figshare.1134647. doi:
http://dx.doi.org/10.6084/m9.figshare.1134647.
URL
http://figshare.com/articles/
CommodConverter/1134647
[61] R. Flanagan, Bright-lite, https://github.com/brightdev/bright-lite (Oct. 2014).
URL https://github.com/bright-dev/bright-lite
[62] S. Skutnik, Development of an ORIGEN-based
Reactor Analysis Module for Cyclus, FigshareHttp://dx.doi.org/10.6084/m9.figshare.1291144. doi:
http://dx.doi.org/10.6084/m9.figshare.1291144.
URL http://figshare.com/articles/Development_
of_an_ORIGEN_based_Reactor_Analysis_Module_for_
Cyclus/1291144
[63] K.
D.
Huff,
MktDrivenInst,
FigshareHttp://dx.doi.org/10.6084/m9.figshare.1134650. doi:
http://dx.doi.org/10.6084/m9.figshare.1134650.
URL http://figshare.com/articles/MktDrivenInst/
1134650
[64] D. Djokic, A. M. Scopatz, H. R. Greenberg, K. D. Huff,
R. P. Nibbelink, M. Fratoni, The Application of CYCLUS to Fuel Cycle Transition Analysis, in: Proceedings of Global 2015, LLNL-CONF-669315, Paris, France,
2015, p. 5061.
[65] A. M. Scopatz, Non-judgemental Dynamic Fuel Cycle Benchmarking, arXiv:1511.09095 [physics]ArXiv:
1511.09095.
URL http://arxiv.org/abs/1511.09095
[66] M. Gidden, P. Wilson, K. Huff, R. Carlsen, OnceThrough Benchmarks with C YCLUS, a Modular, OpenSource Fuel Cycle Simulator, in: Proceedings of the 2012
ANS Winter Conference, San Diego, CA, 2012.
[67] A. Baker, R. Ross, Comparison of the Value of Plutonium and Uranium Isotopes in Fast Reactors, in: Proc.
Conf. on Breeding, Economics and Safety in Large Fast
Breeder Reactors, ANL-6972, 1963, pp. 7–10.
22
| 5cs.CE
|
Optimization-Based Autonomous Racing
of 1:43 Scale RC Cars
Alexander Liniger, Alexander Domahidi and Manfred Morari
arXiv:1711.07300v1 [math.OC] 20 Nov 2017
Abstract
This paper describes autonomous racing of RC race cars based on mathematical optimization. Using a dynamical
model of the vehicle, control inputs are computed by receding horizon based controllers, where the objective is to
maximize progress on the track subject to the requirement of staying on the track and avoiding opponents. Two
different control formulations are presented. The first controller employs a two-level structure, consisting of a path
planner and a nonlinear model predictive controller (NMPC) for tracking. The second controller combines both
tasks in one nonlinear optimization problem (NLP) following the ideas of contouring control. Linear time varying
models obtained by linearization are used to build local approximations of the control NLPs in the form of convex
quadratic programs (QPs) at each sampling time. The resulting QPs have a typical MPC structure and can be solved
in the range of milliseconds by recent structure exploiting solvers, which is key to the real-time feasibility of the
overall control scheme. Obstacle avoidance is incorporated by means of a high-level corridor planner based on
dynamic programming, which generates convex constraints for the controllers according to the current position of
opponents and the track layout. The control performance is investigated experimentally using 1:43 scale RC race
cars, driven at speeds of more than 3 m/s and in operating regions with saturated rear tire forces (drifting). The
algorithms run at 50 Hz sampling rate on embedded computing platforms, demonstrating the real-time feasibility
and high performance of optimization-based approaches for autonomous racing.
I. I NTRODUCTION
Autonomous car racing is a challenging task for automatic control systems due to the need for handling
the vehicle close to its stability limits and in highly nonlinear operating regimes. In addition, dynamically
changing racing situations require advanced path planning mechanisms with obstacle avoidance executed
in real-time. Fast dynamics constrain the sampling time to be in the range of a few tens of milliseconds at
most, which severely limits the admissible computational complexity of the algorithms. This situation is
even more challenging if the autonomous algorithms shall be executed on simple, low-power embedded
computing platforms.
In this paper, we investigate optimization-based control strategies for the task of racing an autonomous
vehicle around a given track. We focus on methods that can be implemented to run in real-time on
embedded control platforms, and present experimental results using 1:43 scale Kyosho dnano RC race
cars that achieve top speeds of more than 3 m/s, which corresponds to an upscaled speed of about 465
km/h. For high performance, the proposed controllers operate the car at its friction limits, far beyond
the linear region of what is typically used in other autonomous driving systems. This challenging task is
generally mastered only by expert drivers with lots of training. In contrast, our approach requires merely
a map of the track and a dynamical model of the car; in particular, we use a bicycle model with nonlinear
tire forces, neglecting load transfer and coupling of lateral and longitudinal slip.
Two model-based control schemes are presented in this paper. First, we describe a hierarchical two-level
control scheme consisting of a model-based path planner generating feasible trajectories for an underlying
nonlinear model predictive control (NMPC) trajectory tracking controller. Our second approach combines
both path planning and path following into one formulation, resulting in one NMPC controller that is based
on a particular formulation known from contouring control [1], [2]. The objective of both approaches is
to maximize the progress on the track, measured by a projection of the vehicle’s position onto the center
line of the track. Linear time varying (LTV) models obtained by linearization are employed to construct a
The authors are with the Automatic Control Laboratory, ETH Zurich, 8092 Zürich, Switzerland;
Emails: liniger|domahidi|[email protected]
tractable convex optimization problem to be solved at each sampling time. Efficient interior point solvers
for embedded systems, generated by FORCES [3], [4], are employed to solve the resulting optimization
problems, which makes the approaches presented in this paper amenable for use on embedded systems.
We furthermore demonstrate that both schemes can easily be extended to incorporate obstacle avoidance
by adjusting the constraints of the resulting optimization problems according to the racing situation. With
this mechanism in place, the avoidance trajectory is optimized to yield maximal overall progress, which
translates into highly effective overtaking maneuvers that are automatically planned and executed.
A. Related work
Safe autonomous driving at moderate speeds has been demonstrated for example in the context of
Autonomous Highway Systems (AHS) in 1997 [5] within the Californian PATH programme, or in contests
such as the DARPA Grand Challenge in 2005 [6] and the Urban Challenge in 2007 [7] by numerous
research groups. In these projects, the main goal was do develop autonomous driving systems for public
traffic situations. Autonomous racing has received less attention, but impressive advances have been made
recently also in this discipline. For instance, a fully autonomous Audi TTS has been reported to race
at high speeds in [8] using a trajectory tracking controller, which follows a pre-computed trajectory.
However, to the best knowledge of the authors, fully automatic racing including obstacle avoidance in
dynamic racing situations has not been demonstrated yet in practice.
From a control perspective, the fundamental building blocks for automatic racing can be grouped into
three categories: drift control, reference tracking, and path planning. Research associated with the first
group focuses on gaining a better understanding of the behavior of the car near the friction limit for
designing safety control systems that try to prevent loss-of-control accidents. By analyzing the nonlinear
car model, it can be shown that steady state motions corresponding to drifting can be generated, see
e.g. [9] and [10]. As these approaches are not designed for trajectory tracking, the controlled car drifts
in a circle, which represents a steady state drift equilibrium.
Reference tracking controllers usually are designed to operate the vehicle within the linear tire force
region for maximum safety and robustness. Approaches based on NMPC, which allows one to incorporate
the latter constraint in a systematic manner, deal with the nonlinearities of the model either directly by
a nonlinear programming (NLP) solver [11], or use LTV [12] or piece-wise affine (PWA) approximations [13], resulting in a convex quadratic program (QP) or a mixed-integer QP, respectively. Approaches
without optimization in real-time include nonlinear robust control [14] and flatness-based control [15],
which however lack the ability to incorporate constraints. One approach that is explicitly designed for
operating with saturated tire forces is [8], where the center of percussion (CoP) is used to design a state
feedback steering controller for reference tracking. At the CoP, the rear tire forces do not influence the
dynamics by definition, which allows for a simple linear steering controller.
In order to avoid obstacles, reference tracking schemes rely on a higher-level path planner. A simple
point mass model in the high-level path planner is used in [16] to limit the computational complexity. This
can be problematic if the generated trajectories are infeasible with respect to the car’s physical dynamics,
which can lead to preventable collisions [17]. The latter reference suggests to use a library of steadystate movements (trims) with parametrizable length, which are generated from a more complex dynamical
model, and to connect them by maneuvers which allow the transition from one trim to another, similar to
the idea of [18]. This path planning has the advantage that it generates feasible reference trajectories for
the low level tracking controller, and that drifting trims can be included to increase the agility of the car.
However, the resulting optimization problem is a mixed-integer program that is too complex to be solved
within the typical sampling times on embedded platforms. Consequently, the approach is not well suited
for real-time autonomous racing.
In order to avoid relying on the feasibility of the path planner, one-level approaches have been investigated e.g. in [19], where optimal trajectories and the associated open-loop control inputs for rally
racing maneuvers are calculated numerically in simulation. In [16], obstacle avoidance is incorporated into
the tracking controller by using an additional cost term that penalizes coming too close to the obstacle.
However, the controller is reported to perform worse than its two-level counterpart, especially if the car
has to diverge from the reference trajectory. A one-level approach similar to [16] is studied in [20] in
simulation, where obstacle avoidance is incorporated into the problem formulation by using a spatial
reformulation and imposing constraints on the transformed variables. While in [19], [16] the solution to
the NLP is obtained by a standard nonlinear solver with prohibitively long solve times, [20] uses real-time
iterations [21] to achieve the low computation times needed for a real-time implementation. However, it
is assumed that there is only one side where the obstacle can be overtaken, which may not be the case
in practical racing situations.
An interesting alternative to optimization-based methods are sampling-based methods. For example,
rapidly exploring random trees (RRTs) have been investigated in [22] and [23] to generate time optimal
trajectories through a 180◦ curve. The advantage is that such algorithms tend to quickly find feasible
trajectories. In [23], the differential flatness of the system is exploited to speed up the convergence of
the algorithm, generating obstacle-free trajectories even in complex situations. However, the reported
computation times are not yet low enough to allow for a real-time implementation.
B. Contribution
In this paper, we describe two novel autonomous, optimization-based racing controllers that incorporate
obstacle avoidance, track constraints, actuator limitation and the ability for controlled drift in a systematic
and straightforward way. Real-time feasibility at 50 Hz sampling rate is demonstrated in experimental
results with fast RC cars. To the best knowledge of the authors, this is the first implementation of
autonomous racing controllers with obstacle avoidance on an experimental testbed.
The fundamental idea of both approaches is to use a receding horizon controller, which maximizes
progress on the track within the horizon as a performance measure. This is closely related to a time
optimality criterion and allows for a systematic incorporation of obstacles and other constraints. It is
particularly effective for overtaking opponents, as our controllers seek for a progress-optimal solution
around the obstacles. In order to deal with track constraints and obstacle avoidance situations, we represent
the feasible set for the position of the car at any time instance by a slab defined by two parallel linear
inequalities. This ensures tractability of the subproblems by convex programming on one hand, and
incorporation of track and obstacle constraints in a systematic manner on the other hand. To deal with
non-convex situations such as overtaking an obstacle on the left or right, a high-level corridor planning
algorithm supplies a set of appropriate convex constraints to the controllers. This constraint set is the
result of a shortest path problem solved by dynamic programming.
The first, hierarchical control approach presented in this paper is in principle similar to [17], but its
complexity is low enough to be fully implemented in real-time. Our path planner also uses trims, but only
one is selected for the whole prediction horizon from a set of trajectories, which represent steady-state
cornering conditions of the nonlinear model, gridded for velocity and steering angle. The path planner
selects the trajectory with the largest progress that does not leave the track or hits any obstacle. Moving
obstacles can inherently be dealt with, as new trims are generated every sampling time. A low level
nonlinear reference tracking MPC controller tracks the selected trim, as in [17], but we additionally use
obstacle avoidance constraints as described above, which turns out to be very effective in practice.
The second controller is based on a model predictive contouring control (MPCC) formulation [1], [2],
combining path generation and path tracking into one problem. The MPCC essentially plans a progressoptimal path by taking into account the (nonlinear) projection of the vehicle’s position onto the center
line. The resulting controller is able to plan and to follow a path which is similar to the time-optimal path
in [19] when the horizon is chosen long enough.
Both NMPC problems are approximated locally by linearizing the continuous nonlinear system dynamics
around a state trajectory to obtain an LTV model. In the hierarchical approach, we linearize around the
trajectory obtained from the path planner, while in the one-level approach the linearization points are given
by the shifted trajectory from the previous sampling time. We then discretize by the matrix exponential, and
solve the resulting quadratic program (QP) by a tailored interior point solver, generated by FORCES [3],
[4]. The computation time of the solvers scale linearly with the horizon length, hence we make use of
long horizons of up to 40 time steps in the MPCC, for example. After the QP has been solved, the first
control input is applied to the system, and the process is repeated at the next sampling time. This approach
for solving the NMPC problems corresponds to the basic version of the real-time iterations from [21].
C. Outline
The paper is organized as follows. In Section II, the dynamical model of the car and some of its basic
properties are described. In Section III-A, we present the hierarchical control approach with separate path
planner and path tracking, while the combined approach is presented in Section III-B. In Section III-C,
we briefly discuss our method for selecting convex constraints for the two controllers, enabling obstacle
avoidance in dynamic racing situations. Section IV presents the testbed setup and experimental results for
both controllers.
N OTATION
The following notation is used throughout the paper. The set of reals is denoted by R, and the set of
real column vectors of length n is denoted by Rn . The set of nonnegative and strictly positive reals is
denoted by R+ and R++ , respectively. The set of positive definite matrices of size n is denoted by Sn++
and the set of positive semi-definite matrices of size n by Sn+ . We define kxk2P , xT P x for a vector
x ∈ Rn and some positive definite matrix P ∈ Sn++ . If a ∈ Rn and b ∈ Rm are column vectors or scalars,
we denote their concatenation by (a, b) , [aT , bT ]T ∈ Rn+m .
II. C AR M ODEL
A. Bicycle Model
The RC cars are modeled using a bicycle model as done in [9], [10], where the car is modeled as one
rigid body with a mass m and an inertia Iz , and the symmetry of the car is used to reduce it to a bicycle.
Only the in-plane motions are considered, i.e. the pitch and roll dynamics as well as load changes are
neglected. As the used cars are rear wheel driven and do not have active brakes, the longitudinal force
on the front wheel is neglected. The resulting model is shown in Figure 1 together with the resulting
differential equations (1) defining the model:
Ẋ = vx cos(ϕ) − vy sin(ϕ) ,
(1a)
Ẏ = vx sin(ϕ) + vy cos(ϕ) ,
ϕ̇ = ω ,
1
v̇x = (Fr,x − Ff,y sin δ + mvy ω) ,
m
1
v̇y = (Fr,y + Ff,y cos δ − mvx ω) ,
m
1
ω̇ = (Ff,y lf cos δ − Fr,y lr ) .
Iz
(1b)
(1c)
(1d)
(1e)
(1f)
The equation of motion is derived around the center of gravity (CoG), where the states are the position
X, Y of the CoG in the inertial frame and the angle of the car relative to the inertial frame, ϕ. This
is the kinematic part of the model. The kinetic part of the model is derived around a body-fixed frame
centered at the CoG, where the states are the longitudinal and lateral velocity of the car, vx and vy , and
finally the yaw rate ω. The control inputs are the PWM duty cycle d of the electric drive train motor and
αf
δ
Ffy(αf)
vy
Y
ω
vx
φ
Fry(αr)
αr
lf
lr
Frx(d)
X
Fig. 1: Schematic drawing of the car model
the steering angle δ. The subscripts x and y indicate longitudinal and lateral forces or velocities, while
r and f refer to the front and rear tires, respectively. Finally, lf and lr are the distance from the CoG to
the front and the rear wheel.
The tire forces F model the interaction between the car and the road and are the most important part
of the dynamics. As the goal is for the cars to race, the model of the tire forces has to be realistic enough
to represent the car at high speeds and its handling limits. The lateral forces Ff,y and Fr,y are modeled
using a simplified Pacejka Tire Model [24], see (2a) and (2b), which is a good approximation to measured
friction curves in practice:
ωlf + vy
Ff,y = Df sin(Cf arctan(Bf αf )) where αf = − arctan
+δ,
(2a)
vx
ωlr − vy
Fr,y = Dr sin(Cr arctan(Br αr )) where αr = arctan
,
(2b)
vx
Fr,x = (Cm1 − Cm2 vx )d − Cr − Cd vx2 .
(2c)
The parameters B, C and D define the exact shape of the semi-empirical curve. The longitudinal force of
the rear wheel Fr,x is modeled using a motor model for the DC electric motor as well as a friction model
for the rolling resistance and the drag, cf. (2c). Due to the size of the cars, it is currently not feasible to
measure the wheel speed, which makes it impossible to use combined slip models such as [9], [19].
The model is identified using a combination of stationary and dynamic identification experiments, as
reported in [10]. This allows for identifying the rear wheel combined slip effects into the lateral tire
friction model (2b). Thus the identified model is suitable to represent the car also at its handling limits,
when the tire forces are saturated or close to saturation.
B. Stationary Velocity Analysis
For a better understanding of the model, the stationary velocities of the model are investigated. A
similar, more elaborate analysis is presented in [9], [10].
The objective is to find points in the model where all accelerations are zero. This is done for different
constant forward velocities v̄x and constant steering angles δ̄, thus the problem becomes a nonlinear
algebraic system of equations, with two equations (v̇y = 0, ω̇ = 0) and two unknowns (v̄y , ω̄). By finding
a solution to the algebraic equations for different steering angles δ̄ and different initial conditions, the
stationary velocities for one forward velocity (v̄y , ω̄) can be calculated.
2
2
Normal Driving
Understeering
Oversteering
4
1
−0.5
2
0
y
0
−0.5
−2
−1
4
0.5
v [m/s]
ω [rad/s]
y
0
6
1
2
0.5
v [m/s]
Normal Driving
Understeering
Oversteering
1.5
ω [rad/s]
1.5
6
−2
−1
−4
−1.5
−4
−1.5
−6
−2
0
−0.2
0
δ [rad]
0.2
−6
−0.2
0
δ [rad]
0.2
−2
−0.2
0
δ [rad]
0.2
−0.2
0
0.2
δ [rad]
Fig. 2: Stationary velocities for v̄x = 1.5 m/s (plots on the left) and v̄x = 2 m/s (plots on the right),
parameterized by the steering angle δ̄.
The resulting stationary velocities are shown for v̄x = 1.5 m/s and v̄x = 2 m/s in Figure 2. The resulting
stationary velocities can be categorized into three different regions. The blue line is the normal driving
region, which is characterized by the linear dependency of the yaw rate and the steering angle, as well
as small lateral velocities. The other two regions in red and green correspond to over- and understeering
points in the model. In the understeering region (green curves), the saturated front tire force law prevents
the car from achieving larger yaw rates, and thus the car can not take a sharper turn. In the oversteering
region (red curves), the rear tire force law is fully saturated and the car drives with a high lateral velocity,
which is usually called drift.
The dynamics of the model in the different regions are different, for example in the oversteering region
where the car is drifting, the steering angle can be of opposite sign than the curvature the car is driving,
known as counter steering, which does not occur in the case of the normal and the understeering mode. A
stability analysis of the lateral velocity model around the different stationary velocity points also shows
the difference in the dynamics: While the normal driving region and the understeering region are stable,
the drifting region is unstable [10].
From Figure 2 it is visible that the normal steering region is getting smaller when the car is driving
faster. Thus the maximal and minimal steering angle for which the tire forces are not saturated is reduced
with increasing velocities, and at the same time the maximal yaw rate decreases.
Because all velocities are constant, the movement of the car is a uniform circular movement, for which
the relationship between the yaw rate, the radius R and the curvature κ is known,
q
v
|ω| =
= vκ where v = vx2 + vy2 .
(3)
R
Consequently, all stationary velocity points correspond to the car driving in a circle with a constant radius,
i.e. the model can predict drifting along a circle.
III. AUTONOMOUS R ACING C ONTROL
In this main part of the paper, we present two optimization-based formulations for the task of autonomous racing for RC cars which are based on the model described in Section II. First, the hierarchical
two-level approach is presented in Section III-A, followed by the one-level scheme in Section III-B. We
then briefly discuss a high-level method for obstacle avoidance that is the same for both controllers in
Section III-C. A summary of the algorithms is given in Section III-D.
TABLE I: Part of the Path Planning Library
Description
vx [m/s]
vy [m/s]
ω [rad/s]
δ [rad]
Left
1.75
0.1185
-2.7895
-0.1000
Straight
1.75
0
0
0
Slight Right
1.75
-0.0486
1.3849
0.05
Hard Right
1.75
-0.3163
4.4350
0.16
Hard Left Drift
1.75
1.1860
-4.8709
0.2
A. Hierarchical Receding Horizon Controller
In our hierarchical control approach, a high-level path planner finds a progress-optimal, feasible trajectory within a finite number of possibilities and for a horizon of N sampling times. This trajectory is then
tracked by an MPC controller employing soft-constraints to ensure feasibility of the low-level optimization
at all times. This process is repeated in a receding-horizon fashion, hence we call the controller Hierarchical
Receding Horizon Controller (HRHC) in what follows. Details on the individual components are given
in the following.
1) Path Planning.: The purpose of the path planner is to generate a trajectory with maximal progress
which is feasible for the car’s dynamics. It is based on gridding the stationary velocities of the nonlinear
system described in Section II-B. By solving (1) for the stationary lateral velocity points given a set of
different longitudinal velocities v̄x and steering angles δ̄, a library of zero acceleration points is generated.
This includes stationary velocities from the normal as well as from the drifting region. Table I shows an
excerpt of such a library. The library used in our experiments consists of Nv̄ = 95 stationary points, out
of which 26 correspond to drifting equilibria. The stationary points are selected by uniformly gridding the
longitudinal velocity vx between 0.5 and 3.5 m/s, with steps of 0.25 m/s. For each v̄x , about five to nine
points are selected in the normal driving region, and up to four drifting points are selected for forward
velocities between 1.5 and 2.25 m/s. Note that the library does not change during run-time, hence it can
be generated offline.
To generate trajectories for the whole horizon, the stationary points are integrated. This is computationally cheap under the assumption that the stationary velocity can be reached within one time step, since
in this case all accelerations are zero. This procedure allows us to generate reference trajectories also for
unstable drifting points. The integration of the reasonable stationary velocity points leads to a countable
set of possible trajectories with a constant turning radius over the horizon. A visualization of such a set
of candidate trajectories is shown in Figure 3.
We now proceed with finding the best trajectory among the candidates obtained by gridding and
integration, which is equivalent to solving the following optimization problem:
θ? , max P XNj , YNj ,
(4a)
j∈1,...,Nv̄
s.t. xj0 = x ,
xjk+1
xjk ∈
j
=
fkm (xjk , v̄ j ) ,
Xtrack ,
v̄ ∈ V(v) ,
(4b)
k = 1, . . . , N
(4c)
k = 1, . . . , N
(4d)
(4e)
where xk , (Xk , Yk , ϕk ) is the position and orientation at time step k generated by integrating the
kinematic part of the bicycle model using the stationary velocity v̄ j , (v̄xj , v̄yj , ω̄ j ) from the library of
stationary points. The integration is denoted by the discrete time version of the model, fkm in (4c). The
objective (4a) is to find the trajectory with the largest progress on the track, which is measured by the
projection operator P : R2 → [0, L] that calculates the scalar projection of the position Xkj , Ykj onto the
piecewise linear center line parameterized by the arc length θ ∈ [0, L], where L is the length of the center
θ*
Fig. 3: Left: Generation of reference trajectories for two different longitudinal velocities (black: slow,
green: fast). Right: Solution of the path planner. The red trajectory has the largest progress but leaves the
track, so that the orange one is the optimal trajectory maximizing θN .
line. All positions of the trajectory xjk , k = 0, . . . , N, need to lie in the set Xtrack ⊂ R2 (4d) which is the
set of all admissible positions inside the track boundaries. Finally, we demand that the stationary velocity
v̄ j lies in the set V ⊂ R3 (4e), which depends on the current velocity measurement v , (vx , vy , ω) and
enforces that the planned velocity is reasonably close to the measured velocity, which helps to deal with
the assumption that the stationary velocity can be reached within one time step:
(
{(v̄x , v̄y , ω̄) ∈ R3 | ρ ≥ |vx − v̄x |}
if |v̄y | ≤ ν
V(v) ,
, (5)
3
{(v̄x , v̄y , ω̄) ∈ R | σx ≥ |vx − v̄x | , σy ≥ |vy − v̄y | , σω ≥ |ω − ω̄|} otherwise
where the two cases correspond to selecting non-drift or drift trajectories, respectively, based on the
lateral steady state velocity v̄y . The tuning parameters ν, ρ, σx , σy , σω are used to determine the behavior
of the path planner in terms of selecting candidate trajectories. Choosing parameters that lead to too
loose constraints in (5) result typically in physically impossible trajectories, which can result in crashes.
Conversely, constraints that are too tight limit the performance of the controller, as the planned trajectories
are too conservative.
Problem (4) is an integer program with the decision variable j and solved by enumeration in practice
by checking the discrete positions against local convex inner approximations of the track set, which is
reasonable due to the relatively small number of trajectories (about one hundred in our case). The result
of the optimization problem is visualized in Figure 3. The path planning allows to efficiently handle drift,
as a drifting trajectory is selected if it yields the largest progress.
The path planner has some limitations. Firstly, on one hand, the horizon length has to be quite short,
otherwise the planned trajectories become circles and always yield infeasibilities with respect to the track
constraints, even if the selected velocity would be well suited for the current state. A horizon which is too
short, on the other hand, leads to poor performance as the path planner recognizes a forthcoming curve
too late. The second problem is that the whole horizon has the same velocity, which can lead to problems
in complicated curve combinations such as chicanes.
2) Model Predictive Reference Tracking.: The reference trajectory from the path planner is generated
under simplifying assumptions, thus it is not possible to directly apply the controls which correspond to
the stationary velocity point. In order to efficiently track the reference the following MPC formulation is
used, penalizing the deviation from the reference trajectory using a quadratic function:
min kxN −
x,u,s
2
xref
N kP
+ pksN k∞ +
N
−1
X
s.t. x0 = x ,
xk+1 = Ak xk + Bk uk + gk ,
F k xk ≤ f k + s k ,
sk ≥ 0 ,
x ≤ xk ≤ x̄ ,
u ≤ uk ≤ ū ,
2
ref 2
kxk − xref
k kQ + qksk k∞ + kuk − uk kR
(6a)
k=0
k = 0, ..., N − 1
k = 1, ..., N
k = 1, ..., N
k = 1, ..., N
k = 0, ..., N − 1
(6b)
(6c)
(6d)
(6e)
(6f)
(6g)
where Q ∈ S6++ , R ∈ S2++ , P ∈ S6++ are tuning matrices and p, q ∈ R++ are penalties for the soft
constraints (6d), which are discussed below. To capture as much of the nonlinear model as possible while
maintaining computational tractability, the model is linearized around the reference trajectory, cf. (6c). It
is necessary to linearize around a trajectory rather than a single operating point, since both position and
orientation can vary significantly over the horizon. The reference trajectory is by construction feasible with
respect to the track. In order to also guarantee that the trajectory of the optimal MPC solution satisfies
track constraints, the X and Y states are limited to stay inside the track if possible. This is achieved by
limiting each point in the horizon to lie within two parallel half spaces (6d), which leads to two affine
inequality constraints per time step (one for the right and one for the left border), resulting in a convex
feasible set. Figure 4 depicts these border constraints (6d), which we formulate as soft constraints in
order to avoid infeasibility problems in practice. By using an exact penalty function (an infinity norm
in our case), we ensure that the resulting optimization problem is always feasible, and that the original
solution of the hard constrained problem is recovered in case it would admit a solution [25], [26]. Finally,
the control inputs are limited within their physical constraints, and all other states are limited within
reasonable bounds to avoid convergence problems of the solver due to free variables (6f), (6g).
Problem (6) can be reformulated as a convex QP and thus efficiently solved. In this work FORCES [3],
[4], is employed to generate tailored C-code that solves instances of (6) where parameters are the initial
state x, the matrices defining the dynamics along the reference trajectory (Ak , Bk , gk ) as well as the
changing halfspace constraints Fk , fk .
B. Model Predictive Contouring Control
Contouring control is used in various industrial applications such as machine tools for milling and
turning [27] or laser profiling [28]. In these applications, the challenge is to compute inputs that control
the movement of the tool along a reference path. The latter is given only in spatial coordinates; the
associated velocities, angles, etc. are calculated and imposed by the control algorithm. This is different
from tracking controllers in that the controller has more freedom to determine the state trajectories to
follow the given path, for example to schedule the velocity, which in tracking is defined by the reference
trajectory.
The contour following problem can be formulated in a predictive control framework to incorporate
constraints, see e.g. [1]. In this work, we adapt the particular formulation of the model predictive contouring
control (MPCC) framework from [2] to obtain a high-performance controller for autonomous racing, which
maximizes the travelled distance on the reference path within the prediction horizon. In our experiments,
we use the center line as a reference path, but employ it merely as a measure of progress by selecting
low weights on the tracking (contouring) error. As a result, the driven trajectory is very similar to those
driven by expert drivers. The advantage of this approach is that path planning and path tracking can be
combined into one nonlinear optimization problem, which can be solved in real-time by approximating
the NLP using local convex QP approximations at each sampling time [21]. The resulting QPs can be
Fig. 4: Left: Halfspace constraints (red lines) of one point in the horizon (large red dot). Right: Resulting
halfspaces (red lines) of the whole horizon, approximating the track. The green points on the center line
correspond to the closest point on the discrete center line, for each point in the horizon. The points on
the borders are the projection of the center line points on the borders.
formulated with multistage structure, which is exploited by FORCES [3], [4] to obtain solution times in
the range of a few milliseconds.
Before posing the MPCC problem, a few preliminaries are introduced such as the parameterization of
the reference path and the definition of useful error measures.
1) Parameterization of Reference Trajectory.: The reference path is parameterized by its arc length
θ ∈ [0, L] using third order spline polynomials, where L is the total length. The splines are obtained by
an offline fitting of the center line. Using this parameterization, we can obtain any point X ref (θ), Y ref (θ)
on the center line by evaluating a third order polynomial for its argument θ. The angle of the tangent to
the path at the reference point with respect to the X-axis,
∂Y ref (θ)
,
(7)
Φ(θ) , arctan
∂X ref (θ)
is also readily available. This parameterization leads to an accurate interpolation within the known points
of the reference path, and it is more accurate than the piece-wise linear parameterization employed for
the HRHC in Section III-A, which only uses a linear interpolation between the points.
2) Error measures.: In order to formulate the MPCC problem, error measures are needed that define
the deviation of the car’s current position X, Y from the desired reference point X ref (θ), Y ref (θ). We use
the same definitions as in [2], but give another derivation here. Let P : R2 → [0, L] be a projection
operator on the reference trajectory defined by
P(X, Y ) , arg min(X − X ref (θ))2 + (Y − Y ref (θ))2 .
θ
(8)
For brevity, we define θP , P(X, Y ). The orthogonal distance of the car from the reference path is then
given by the contouring error
ec (X, Y, θP ) , sin (Φ(θP )) X − X ref (θP ) − cos (Φ(θP )) Y − Y ref (θP ) ,
(9)
where Φ(·) is defined in (7). The contouring error is depicted in the left picture in Figure 5.
The projection operator (8) is not well suited for use within online optimization algorithms, as it
resembles an optimization problem itself. Thus an approximation θA of θP is introduced, which is an
independent variable determined by the controller. For this approximation to be useful, it is necessary to
link θA to θP via the lag error
el (X, Y, θA ) , |θA − θP | ,
(10)
(X,Y)
θP
(X,Y)
ec
ec
(Xref(θP ),Yref(θP ))
ф(θP )
êl
el
(Xref(θP ),Yref(θP ))
êc
ref
ref
(X (θA),Y (θA))
ф(θA )
Fig. 5: Contouring error ec (left) and lag error el (right) with linear approximations êc and êl .
which measures the quality of the approximation. The lag error is depicted in the right picture in Figure 5
(red segment on reference path).
In order to be independent of the projection operator (8), the contouring (9) and the lag error (10)
can be approximated as a function of the position X, Y and the approximate projection θA , which are
variables that can be controlled by the MPCC controller:
ec ≈ êc (X, Y, θA ) , sin (Φ(θA )) X − X ref (θA ) − cos (Φ(θA )) Y − Y ref (θA ) ,
(11a)
el ≈ êl (X, Y, θA ) , − cos (Φ(θA )) X − X ref (θA ) − sin (Φ(θA )) Y − Y ref (θA ) .
(11b)
The approximate contouring error êc (11a) and the approximate lag error êl (11b) are defined as the
orthogonal and tangential component of the error between X ref (θA ), Y ref (θA ) and the position X, Y, see
the right picture in Figure 5.
3) MPCC problem.: With these error measures in place, we now formulate the model predictive
contouring control problem for autonomous racing. The objective is a trade-off between the quality of
path-following (low contouring error) and the amount of progress achieved over a finite horizon of N
sampling times, subject to model dynamics, track- and input constraints:
min
N
X
keck (Xk , Yk , θP )k2qc − γθP,N
(12a)
k=1
s.t. x0 = x ,
xk+1 = f (xk , uk ) ,
Fk xk ≤ fk ,
x ≤ xk ≤ x̄ ,
u ≤ uk ≤ ū ,
k = 0, ..., N − 1
k = 1, ..., N
k = 1, ..., N
k = 0, ..., N − 1
(12b)
(12c)
(12d)
(12e)
(12f)
where Xk , Yk is the position of the car at time step k determined by the nonlinear model f in (12c),
which is the discrete-time version of (1) with piece-wise constant control inputs. The contouring error
eck (Xk , Yk , θP ) in the objective (12a) is defined in (9), and θP,k is the associated path parameter such that
X ref (θP,k ), Y ref (θP,k ) is the orthogonal projection of Xk , Yk onto the reference path. The two objectives
(maximum progress and tight path following) are traded off by the weights γ ∈ R++ and qc ∈ R++ ,
respectively. Constraints (12d) are the parallel half space constraints for containing the position in the
allowed corridor as described in Section III-A2, while (12e), (12f) corresponds to (6f), (6g), limiting states
and inputs to physically admissible values.
Due to the implicit dependency of the objective (12a) on the projection operator (8), optimization
problem (12) is a bi-level NLP, which is too complex to solve in real-time. Thus the idea is to use
the approximate projection θA,k instead of θP,k as introduced in Section III-B2, and to control the
approximation quality by adding a cost on the lag error (10) to the objective. In order to allow for
forming the lag error at each time step in the prediction horizon, it is necessary to introduce an integrator
state with dynamics θA,k+1 = θA,k + vk /Ts , where vk can be interpreted as the projected velocity, and θA,k
as the state of progress at time k, respectively. This approximation reduces (12) to an optimal control
problem in form of an NLP that is amenable for a real-time implementation:
min
N
X
kêck (Xk , Yk , θA,k )k2qc + kêlk (Xk , Yk , θA,k )k2ql − γvk Ts + k∆uk k2Ru + k∆vk k2Rv − γv0 Ts
(13a)
k=1
s.t. (12b) ,
θ0 = θ ,
vk
,
(12c) , θA,k+1 = θA,k +
Ts
(12d) , (12e) , 0 ≤ θk ≤ L ,
(12f) , 0 ≤ vk ≤ v̄ ,
(13b)
k = 0, ..., N − 1
k = 1, ..., N
k = 0, ..., N
(13c)
(13d)
(13e)
where ∆uk , uk −uk−1 and ∆vk , vk −vk−1 . Note that the approximate contouring error êck (Xk , Yk , θA,k )
(11a) is used in theP
objective (13a). Furthermore, we have replaced the maximization of the final progress
−1
measure, θP,N , by N
k=0 vk Ts , which is equivalent if the approximation is accurate. Sensible lower and
upper bounds on θk and vk are imposed to avoid spurious solutions of the NLP, with v̄ denoting the largest
possible progress per sampling time. The cost on the lag error êlk (Xk , Yk , θA,k ) in (13a) links the state
of progress to the dynamics of the car. To ensure an accurate progress approximation and thus a strong
coupling between the cost function and the car model, the weight on the lag error ql ∈ R++ is chosen
high as suggested in [2]. Furthermore, a cost term on the rate of change of the inputs is added to the
objective (13a) in order to penalize fast changing controls, which helps to obtain smooth control inputs
in order to prevent amplifying unmodeled dynamics.
4) Solving the MPCC problem.: In order to solve the nonlinear optimal control problem (13) in realtime, local convex approximations of (13) in form of the following QPs are built at each sampling time
by linearization of nonlinear terms:
T
T
N
X
xk
xk
xk
∆uk
∆uk
T
Γk
+ ck
min
− γvk Ts +
R
+ qksk k∞ − γv0 Ts
(14a)
θA,k
θA,k
θA,k
∆vk
∆vk
x,u,θ,v,s
k=1
s.t. x0 = x , θA,0 = θ ,
xk+1 = Ak xk + Bk uk + gk ,
vk
θA,k+1 = θA,k +
,
Ts
F k xk ≤ f k + s k ,
sk ≥ 0 ,
x ≤ xk ≤ x̄ , 0 ≤ θA,k ≤ L ,
u ≤ uk ≤ ū , 0 ≤ vk ≤ v̄ ,
k = 0, ..., N − 1
(14b)
(14c)
k = 0, ..., N − 1
(14d)
k = 1, ..., N
k = 1, ..., N
k = 1, ..., N
k = 0, ..., N
(14e)
(14f)
(14g)
(14h)
where Γk ∈ S7+ is formed by the quadratic part of the linearized contouring and lag error cost function
from (13a) and ck ∈ R7 stems from the linear part, respectively. In order to keep the linearization error
small, we use an LTV approximation of the dynamics (1) as well as of the contouring and lag errors (11).
Each nonlinear function is linearized around the output of the last QP iteration shifted by one stage. The
measurement x = x0 is used as the first linearization point, and the last input of the previous iteration is
kept constant to generate a new last input. The linearization point for the terminal state xN is calculated
by simulating the nonlinear model for one time step. The track or corridor constraints (14e) are formulated
by two half space constraints tangential to the track per time step, as described in Section III-A. These
Fig. 6: Left: Combinatorial nature of the overtaking problem, as it is possible to overtake each opponent
on the left or the right side. Center: Spatial-temporal grid for the shortest path problem solved by dynamic
programming, where it is not allowed to visit the blue grid points. Right: Optimal path of the DP (in
black) and the resulting corridor, which is issued to the HRHC (Section III-A) or MPCC (Section III-B).
constraints are formulated as soft constraints, with slack variables sk ∈ R2 and a corresponding infinitynorm penalty in the objective (14a) weighted by q ∈ R++ , which is chosen quite high to recover the
behavior of the hard constrained problem whenever possible [25], [26].
Note that the described re-linearization scheme is essentially the basic real-time iteration from [21],
where a nonlinear continuous-time optimal control problem is solved using a multiple shooting technique,
with an exponential map as integrator, and in every time step only one step of a sequential QP (SQP)
scheme is performed. Hence the convergence properties that are known for the real-time iteration [29]
hold also for the MPCC problem (13).
The local QP (14) is solved using FORCES [3], [4]. Due to the rate costs, the problem has to be lifted
to obtain the multistage structure, for which FORCES has been designed, by introducing a copy of the
previous control input at each stage.
C. Obstacle Avoidance
The two controllers presented in this section plan a path within the track, or a corridor, by taking into
account constraints (4d), (6f) or (14e) on the position of the car. Thus it is possible to include obstacle
avoidance by adapting this corridor online depending on the current constellation of opponents. Generating
the optimal corridor based on the position of the obstacles is in general hard, because of the fundamental
problem that the decision on which side to overtake is non-convex and of combinatorial nature if more
than one opponent is involved. This situation is depicted in Figure 6 in the left picture.
In this work, we employ a high level path planner based on dynamic programming (DP) to decide
on which side to overtake the obstacles. Based on the result of the DP, the adapted corridor is given
to the controllers of Section III-A or Section III-B. The high level path planner solves a shortest path
problem [30] on a spatial-temporal grid, cf. Figure 6. To incorporate the planned path of the lower level
controller into the DP, the grid is built up based on the last output of the path planner in the case of the
HRHC or the last QP solution in the case of the MPCC. The DP minimizes the travelled distance, and
additionally the deviation from the last plan of the lower level controller, such that the DP path is close
to the state prediction of the lower level controller. This ensures that the linearizations stay approximately
valid most of the time. Based on the optimal trajectory from the DP, the corridor constraints can easily
be identified, see the right picture in Figure 6.
Due to space restrictions, we do not go further into detail, but many references exist that solve the
corridor problem with different approaches. For example, a similar problem had to be solved during the
DARPA Urban challenge where the goal was autonomous driving in city traffic. In [7] several approaches
to this problem are presented, for example based on model predictive trajectory generation algorithms [31],
modified A? -algorithms [32] or RRTs [33].
D. Summary
In this main part of the paper, we have presented two approaches for automatic racing. Both approaches
are model-based and maximize progress within a given time horizon. They directly incorporate track
and obstacle constraints by using two parallel affine inequalities (slabs) representing the feasible set of
positions at each time. The major difference between the approaches is that the first (cf. Section III-A
and Algorithm Ia) is a two-level approach with a separate path planning mechanism combined with
a reference tracking MPC controller, while the second approach (cf. Section III-B and Algorithm Ib)
formulates both tasks (path planning and path tracking) into one nonlinear optimal control problem based
on model predictive contouring control.
Algorithm 1 HRHC and MPCC algorithm
(Ia) HRHC
1:
2:
3:
4:
5:
6:
7:
8:
9:
10:
11:
12:
get current position and velocities
compute delay compensation
function B ORDER A DJUSTMENT
shift old planned path by one stage
get position of opposing cars
run DP and get new borders (Sec. III-C)
end function
run path planner (4)
linearize and discretize the model, i.e. build
problem (6)
solve QP using FORCES (6)
send first control at the end of time slot
go to step 1
(Ib) MPCC
1:
2:
3:
4:
5:
6:
7:
8:
9:
10:
11:
get current position and velocities
compute delay compensation
augment old QP output (Section III-B4)
function B ORDER A DJUSTMENT
get position of opposing cars
run DP and get new borders (Sec. III-C)
end function
linearize and discretize the model and the cost
function, i.e. build problem (14)
solve QP using FORCES (14)
send first control at the end of time slot
go to step 1
Both controllers send the new control input at the end of the sampling period. In order to compensate
for this fixed delay, the nonlinear model is simulated forward for this time interval at the beginning of
the algorithm, see line 2 of Algorithm Ia and Ib. The nonlinear model is integrated using a second-order
Runge-Kutta method with the current state estimate from the Kalman filter as the initial condition.
IV. R ESULTS
In this section the performance of both controllers is evaluated on an experimental testbed using 1:43
scale RC race cars. Details on the experimental setup, the particular implementation and closed-loop
performance are given in the following.
A. Experimental Setup
For our experiments, we use the Kyosho dnano cars, which are quite sophisticated with a front and rear
suspension and a rear axle differential. The cars are able to reach speeds of over 3 m/s on straights, which
corresponds to an upscaled speed of about 465 km/h. The wide speed range, the high maximal forward
velocity and the small scale introduce further complexity which is not present in full size testbeds.
The testbed further consists of an infrared camera tracking system, a control board and a custom built
race track of length L = 18.43 m (measured at the center line), see Figure 7. A PointGrey Flea3 camera
captures 100 frames per second with an accuracy of below 4 mm. A wide angle lens with an infrared filter
is used to capture the 4 by 4 meter track. Reflecting markers are arranged on the cars in different unique
patterns and used to identify the car and to measure the position and angle of the cars. An Extended
Kalman Filter (EKF) is used for state estimation, filtering the vision data and estimating the longitudinal
and lateral velocity as well as the yaw rate. The vision data of all cars is passed over an Ethernet connection
to a control PC, which runs one of the two control algorithms described in Section III together with the
corridor planner solving the DP problem. The calculated inputs are then sent via Bluetooth to the car. As
the original electronics of the dnano cars does not feature an open digital communication link, we have
replaced it by a custom made PCB, featuring a Bluetooth chip, current and voltage sensors, H-bridges
Infrared Vision USB3
System
Reflecting
Markers
Car Detection Ethernet
Estimation
Car
Controller
Bluetooth
Fig. 7: Visualization of the control loop and a picture of the used track
for DC motor actuation, and an ARM Cortex M4 microcontroller for the low level control loops such as
steering servo and traction motor control.
B. Implementation
The HRHC is implemented in ANSI C on an embedded platform using an ARM A9 chip (Exynos
4412) at 1.7 GHz. The chip is identical to the one used in Samsung Galaxy S3 smartphones. The runtime
environment was Ubuntu Linux version 12.11, and the code was compiled with the GCC compiler 4.6.3
with option -O2. The sampling time of the controller is Ts = 20 ms, while the discretization time of
the path planner and the tracking MPC controller is 25 ms. The horizon length is N = 14 time steps,
which gives an ahead prediction of 0.35 s. The tracking MPC controller results in a multistage QP with 9
variables per stage, i.e. 132 variables in total. We use a library of 95 candidate trajectories, out of which
26 are drifting stationary velocity points.
The MPCC does not have the limitation of the HRHC path planner, or in other words, the acceleration
is not zero over the whole horizon, thus it is possible to use longer horizons which should increase
the performance. Hence the controller runs with a horizon length of N = 40 with a sampling time of
Ts = 20 ms, which corresponds to a 0.8 s ahead prediction. With the re-linearization method presented in
Section III-B4, the discretization and sampling time have to be the same, as the re-linearization is based
on the last QP solution. The resulting QP has 14 variables per stage and 570 in total.
The MPCC is implemented on a newer gereration embedded platform using an Exynos 5410 chip based
on the ARM Cortex A15 architecture and running at 1.6 GHz. The chip is identical to the one used in the
Korean version of the Samsung Galaxy S4 smartphone. The embedded board runs Ubuntu Linux version
12.04. The ANSI C code of the MPCC controller has been compiled with GCC 4.6.3 with option -O3 .
To improve the convergence of the QP, the problem data is scaled to lie between −1 and +1, and the
objective (Hessian and linear term) is scaled by a factor of 0.1 to reduce large entries occurring due to
the lag error term. Note that these scalings do not change the minimizer. The optimality criteria of the
FORCES solver are adjusted such as to terminate once a sufficiently good solution has been found, with
the residuals of the equality and inequality constraints set to 10−5 , and with the duality gap set to 10−4 .
C. Single Car Racing
The driven trajectories for a single car are depicted in Figure 8, with the velocity profile encoded in
colors. Depicted are three laps, for which the HRHC achieves a lap time between 9.5 to 9.8 s while the
MPCC yields lap times between 8.9 and 9.2 s.
To better understand the driven trajectory of the HRHC, the path planner of the HRHC has to be
analyzed. The path planner has a comparably short prediction horizon, and all velocities over the whole
horizon are constant, which leads to turns of constant radius. Such a path planner works fine for 90◦ and
even up to 180◦ curves. However, the car tends to drive to the outer border after the curve, as this allows
driving a wider radius at a faster longitudinal velocity and thus maximizes the progress. However, this is
a limitation if the car should drive a combination of curves, where the position of the car at the end of the
first curve is essential for a low overall time. This problem can be seen in Figure 8, for example in the
chicane in the lower left corner. Furthermore, due to the short lookahead, the HRHC path planner is not
able to prevent such an ill positioning relative to a curve ahead. Due to the suboptimal position relative
to the curve, the controller has to brake more compared to a car on the ideal line and it is even possible
that the car touches the borders, see the narrow 180◦ curve in the center of the track in Figure 9.
These limitations are not present in the MPCC approach, where a comparably long horizon is employed.
Unlike the HRHC, the velocity is not fixed over the horizon but computed for each time step as the result of
the NLP. As a consequence of these features, the MPCC is able to plan a path even through a complicated
combination of curves. This results in a trajectory which is closer to an ideal line in a least curvature
sense, especially through complicated curves like the chicane in the lower left corner of Figure 8. In this
curve combination, the MPCC achieves velocities not smaller than 1 m/s, while the HRHC has to reduce
the velocity to nearly 0.5 m/s to drive through the chicane, and additionally causes the car to drive an
overall longer distance.
However, the MPCC has some disadvantages compared to the HRHC. In our current implementation,
the MPCC still tracks the center line, which can be seen on the first straight, after the top left corner where
a movement to the center line is clearly visible in Figure 8. This is clearly due to the objective function
of the MPCC, and cannot be completely avoided, even if the contouring cost is small. The HRHC on the
other hand does not have any cost related to the center line and only uses it as a measure of progress, thus
the car drives straight in the aforementioned segment of the track. The second disadvantage can be seen in
the S-curve at lower end of the track, where the MPCC has a small S shape in the driven trajectory. The
HRHC on the other hand shows that the section can be driven completely straight, which gives a speed
benefit. This can be most probably also attributed to the contouring error penalty, for which it would be
beneficial to drive an S-shape in order to follow the center line.
In summary, our results indicate that the HRHC is limited by the path planner, and the fact that the
whole horizon has the same velocity. Allowing multiple velocities within one prediction horizon would
most probably increase the performance of the controller. However, the complexity in the path planner
grows exponentially with the number of different velocity triples in the horizon. The MPCC, which does
not have this limitation, is able to plan better trajectories which also improve the closed loop performance.
However, the MPCC comes at a price, first in terms of computational cost (it is about a factor of five more
expensive than the HRHC when comparing computation times on the same platform, see Section IV-E for
more details on computation times) and, second, in terms of robustness. During our experiments, the relinearization scheme turned out to be sensitive to measurement errors and model drift. In particular, sudden
unexpected skidding of the car can cause problems in the re-linearization scheme. Such drifts are also
responsible for the high variations between different trajectories in certain areas of the track. Furthermore,
since the MPCC relies on the results from the previous iteration to compute the linearizations, sudden
jumps in the high-level corridor path planner might make the previous trajectory infeasible. In such a
situation, several time steps might be needed to recover feasibility of the planned trajectory, although by
using the soft constrained formulation sensible inputs are still provided to the car. The HRHC on the
other hand does not need any information from the last iteration as only the current position and velocity
is needed. This helps to deal efficiently with unexpected behavior and fast changing measurements or
corridors from the high-level obstacle avoidance mechanism.
D. Racing with multiple cars
Since both controllers plan a path around opponents while maximizing the progress, fast and safe
overtaking maneuvers are enabled. In the current work, we focus on static obstacles, which seems
sufficient to outperform non-expert human drivers. To robustly overtake dynamically moving obstacles
(other automatic controllers or expert drivers), it would be necessary to have a prediction of their behavior,
which is currently not available. These predictions could however be systematically incorporated into the
HRHC controller (Section III-A)
MPCC controller (Section III-B)
3.5
3.5
1.5
1.5
2.5
2
−0.5
Y [m]
0
0.5
[m/s]
Y [m]
0.5
3
1
2.5
0
2
[m/s]
3
1
−0.5
1.5
1.5
−1
−1
1
1
−1.5
−1.5
−1.5
−1
−0.5
0
0.5
1
1.5
2
2.5
0.5
−1.5
−1
−0.5
0
X [m]
0.5
1
1.5
2
2.5
0.5
X [m]
Fig. 8: The driven trajectory with velocity profile for 3 different laps. A video of the HRHC controller is
available at https://www.youtube.com/watch?v=ioKTyc9bG4c.
HRHC controller (Section III-A)
MPCC controller (Section III-B)
3.5
3.5
1.5
3
1
0.5
0
2
Y [m]
2.5
[m/s]
Y [m]
0.5
3
1
2.5
0
2
[m/s]
1.5
−0.5
−0.5
1.5
1.5
−1
−1
1
1
−1.5
−1.5
−1.5
−1
−0.5
0
0.5
X [m]
1
1.5
2
2.5
0.5
−1.5
−1
−0.5
0
0.5
1
1.5
2
2.5
0.5
X [m]
Fig. 9: Driven trajectories with static obstacles for 3 different laps. A video of the MPCC with obstacle
avoidance is available at https://www.youtube.com/watch?v=mXaElWYQKC4.
high-level corridor planner outlined in Section III-C, and would merely change the corridor issued to the
controllers presented in Section III-A and Section III-B.
An obstacle avoidance situation is shown in Figure 9, with a detailed zoom taken at three different time
points in Figure 10. These zooms show the prediction horizon and the corresponding planned trajectories.
Due to the limited path planner, the HRHC is not able to directly plan a path around all obstacles from the
beginning. Thus it avoids the first three cars, without predicting how to overtake the next two cars. Only
after the car has made enough progress and overtaken the first group of obstacles, can the controller find
a way between of the other two cars. The MPCC on the other hand plans the path around all obstacles
before it even reaches the first car due to its long horizon. This is an advantage, but if the model mismatch
is significant it can lead to problems, as the planned path is too optimistic.
E. Computation Times
The computation times of the most time-consuming components of the two controllers are given in
Tables II and III for the laps depicted in Figure 8 (single car racing) and Figure 9 (avoidance of opponents).
Note that the computation times of the individual blocks of the two controllers cannot be compared directly.
HRHC controller (Section III-A)
MPCC controller (Section III-B)
3.5
3.5
1.8
Y [m]
1.6
1.4
3
−1
−0.5
0
0.5
3
−1
1
X [m]
−0.5
0
0.5
1
X [m]
2.5
2.5
1.8
2
1.4
Y [m]
1.6
[m/s]
1.8
Y [m]
1.4
1.2
1.2
1.6
2
1.4
1.2
1.2
−1
−0.5
0
0.5
−1
1
−0.5
1.5
X [m]
0
0.5
1
1.5
X [m]
1.8
1
1.6
1.4
Y [m]
1.8
Y [m]
1.6
[m/s]
Y [m]
1.8
1
1.6
1.4
1.2
1.2
−1
−0.5
0
X [m]
0.5
1
0.5
−1
−0.5
0
0.5
1
0.5
X [m]
Fig. 10: Obstacle avoidance at 3 different time steps. Grey indicates the corridor chosen by the obstacle
avoidance algorithm.
To directly compare the computation times of the two controllers, the horizon length should be identical.
However, due to the constant velocity limitation in the HRHC path planner, a horizon length identical to
the MPCC is impossible without using multiple segments of constant velocity, which would make the path
planning combinatorial in complexity and thus currently prohibitive for a real-time implementation. Vice
versa, the MPCC cannot work robustly with the short horizon length used by the HRHC. The multiple
car case does not have a significant impact on the execution times of the lower level controller; only
the path planning in the HRHC gets slightly more complicated. However, the computation time of the
high-level corridor planner (the DP approach from Section III-C) varies significantly, depending on the
complexity of the racing situation. For the MPCC, the limiting factor with respect to the sampling time
is the computation time for solving the QP. The main bottleneck in the HRHC is the path planner, which
has a very high maximal computation time caused by back up rules in the case no feasible trajectory can
be found. The QP on the other hand is the most expensive step in the average, but does not have a large
variation in computation time, which makes it less critical.
The 20 ms sampling time was missed by the HRHC controller in only 0.07 % (one sampling instant in
three laps) and by the MPCC controller in 4.4 % (60 sampling instants in three laps) of the time. Overall,
our experiments demonstrate that both schemes can be implemented in (soft) real-time at sampling rates
of 50 Hz.
ACKNOWLEDGMENT
We gratefully acknowledge the work of our students in the course of the race car project: Kenneth
Kuchera, Samuel Zhao and Florian Perrodin for their fruitful help in implementing the MPCC approach;
Michael Janser for his contribution to the obstacle avoidance (DP) algorithm; Sandro Merkli, Nils Wenzler
and Marcin Dymczyk for the development of the embedded control software and the testbed setup;
Celestine Dünner and Sandro Merkli for building the track and low-level filtering software; Michael
TABLE II: HRHC computation times in milliseconds
Without Obstacles
Mean
Stdev
Max
With Obstacles
Mean
Stdev
Max
Border Adjustment
0.02
0.01
0.04
Border Adjustment
1.33
0.37
2.45
Path Planning
2.93
1.96
13.47
QP Generation
0.52
0.09
1.32
Path Planning
3.18
2.04
9.76
QP Generation
0.53
0.06
QP with FORCES
5.10
0.73
7.56
0.76
QP with FORCES
5.11
0.92
9.25
Mean
Stdev
Max
TABLE III: MPCC computation times in milliseconds
Without Obstacles
Mean
Stdev
Max
With Obstacles
Border Adjustment
0.34
0.11
0.64
Border Adjustment
0.70
0.24
1.05
QP Generation
1.95
0.44
2.92
QP Generation
2.34
0.37
2.79
15.03
1.61
21.16
14.43
1.40
21.46
QP with FORCES
QP with FORCES
Dahinden and Christian Stocker for the initial design of the embedded PCB board; Benjamin Keiser
for the current version of the infrared vision system. We furthermore thank Sean Summers, Nikolaos
Kariotoglou, Christian Conte and John Lygeros from the Automatic Control Laboratory for supervising
various projects and providing important comments to the methods presented in this paper. Special thanks
go to Colin Jones, who originated the idea of the race car testbed.
The research leading to these results has received funding from the EU in FP7 via EMBOCON (ICT248940).
R EFERENCES
[1] T. Faulwasser, B. Kern, and R. Findeisen, “Model predictive path-following for constrained nonlinear systems,” in Proceedings of the
48th IEEE Conference on Decision and Control, 2009 held jointly with the 2009 28th Chinese Control Conference, pp. 8642–8647,
2009.
[2] D. Lam, C. Manzie, and M. Good, “Model predictive contouring control,” in Conference on Decision and Control (CDC), pp. 6137–
6142, 2010.
[3] A. Domahidi, A. Zgraggen, M. Zeilinger, M. Morari, and C. Jones, “Efficient interior point methods for multistage problems arising
in receding horizon control,” in Conference on Decision and Control (CDC), (Maui, HI, USA), pp. 668–674, Dec. 2012.
[4] A. Domahidi, “FORCES: Fast optimization for real-time control on embedded systems.” http://forces.ethz.ch, Oct. 2012.
[5] R. Horowitz and P. Varaiya, “Control design of an automated highway system,” Proceedings of the IEEE, vol. 88, no. 7, pp. 913–925,
2000.
[6] M. Buehler, K. Iagnemma, and S. Singh, The 2005 DARPA grand challenge: the great robot race, vol. 36. Springer, 2007.
[7] M. Buehler, K. Iagnemma, and S. Singh, The DARPA urban challenge: Autonomous vehicles in city traffic, vol. 56. Springer, 2009.
[8] K. Kritayakirana and C. Gerdes, “Using the centre of percussion to design a steering controller for an autonomous race car,” Vehicle
System Dynamics, vol. 15, pp. 33–51, 2012.
[9] E. Velenis, E. Frazzoli, and P. Tsiotras, “On steady-state cornering equilibria for wheeled vehicles with drift,” in Proceedings of the
48th IEEE Conference on Decision and Control, 2009 held jointly with the 2009 28th Chinese Control Conference, pp. 3545–3550,
IEEE, 2009.
[10] C. Voser, R. Y. Hindiyeh, and J. C. Gerdes, “Analysis and control of high sideslip manoeuvres,” Vehicle System Dynamics, vol. 48,
no. S1, pp. 317–336, 2010.
[11] F. Borrelli, P. Falcone, T. Keviczky, and J. Asgari, “MPC-based approach to active steering for autonomous vehicle systems,”
International Journal of Vehicle Autonomous Systems, vol. 3, no. 2, pp. 265–291, 2005.
[12] P. Falcone, F. Borrelli, J. Asgari, H. E. Tseng, and D. Hrovat, “Predictive active steering control for autonomous vehicle systems,”
IEEE Transactions on Control Systems Technology, vol. 15, no. 3, pp. 566–580, 2007.
[13] T. Besselmann and M. Morari, “Hybrid parameter-varying model predictive control for autonomous vehicle steering,” European Journal
of Control, vol. 14, no. 5, pp. 418–431, 2008.
[14] J. Ackermann, J. Guldner, W. Sienel, R. Steinhauser, and V. I. Utkin, “Linear and nonlinear controller design for robust automatic
steering,” IEEE Transactions on Control Systems Technology, vol. 3, no. 1, pp. 132–143, 1995.
[15] S. Fuchshumer, K. Schlacher, and T. Rittenschober, “Nonlinear vehicle dynamics control - a flatness based approach,” in Conference
on Decision and Control (CDC), pp. 6492–6497, 2005.
[16] Y. Gao, T. Lin, F. Borrelli, E. Tseng, and D. Hrovat, “Predictive control of autonomous ground vehicles with obstacle avoidance on
slippery roads,” Proceedings of DSCC, 2010.
[17] A. Gray, Y. Gao, T. Lin, J. K. Hedrick, H. E. Tseng, and F. Borrelli, “Predictive control for agile semi-autonomous ground vehicles
using motion primitives,” in American Control Conference (ACC), pp. 4239–4244, IEEE, 2012.
[18] E. Frazzoli, M. Dahleh, and E. Feron, “Maneuver-based motion planning for nonlinear systems with symmetries,” IEEE Transactions
on Robotics, vol. 21, no. 6, pp. 1077–1091, 2005.
[19] E. Velenis, P. Tsiotras, and J. Lu, “Modeling aggressive maneuvers on loose surfaces: The cases of trail-braking and pendulum-turn,”
in European Control Conference, pp. 1233–1240, 2007.
[20] J. Frasch, A. Gray, M. Zanon, H. Ferreau, S. Sager, F. Borrelli, and M. Diehl, “An auto-generated nonlinear MPC algorithm for real-time
obstacle avoidance of ground vehicles,” in Proceedings of the European Control Conference, no. accepted, 2013.
[21] M. Diehl, H. Bock, J. Schlöder, R. Findeisen, Z. Nagy, and F. Allgöwer, “Real-time optimization and nonlinear model predictive control
of processes governed by differential-algebraic equations,” Journal of Process Control, vol. 12, no. 4, pp. 577–585, 2002.
[22] J. Jeon, S. Karaman, and E. Frazzoli, “Anytime computation of time-optimal off-road vehicle maneuvers using the RRT∗ ,” in Conference
on Decision and Control (CDC), pp. 3276–3282, 2011.
[23] J. Jeon, R. Cowlagi, S. Peters, S. Karaman, E. Frazzoli, P. Tsiotras, and K. Iagnemma, “Optimal motion planning with the half-car
dynamical model for autonomous high-speed driving,” in American Control Conference (ACC), pp. 188–193, 2013.
[24] E. Bakker, L. Nyborg, and H. Pacejka, “Tyre modelling for use in vehicle dynamics studies,” SAE, 1987.
[25] D. G. Luenberger, Linear and nonlinear programming. Springer, 2003.
[26] E. C. Kerrigan and J. M. Maciejowski, “Soft constraints and exact penalty functions in model predictive control,” in Proc. UKACC
International Conference (Control 2000), (Cambridge, UK), September 2000.
[27] Y. Koren, “Control of machine tools,” Transactions of the ASME Journal of Manufacturing Science and Engineering, vol. 119, 1997.
[28] R. C. Ko, M. C. Good, and S. K. Holgamuge, “Adaptive calibration of feedforward controllers for laser profiling machines,” in
Proceedings of Information, Decision and Control, 1999.
[29] M. Diehl, H. G. Bock, and J. P. Schlöder, “A real-time iteration scheme for nonlinear optimization in optimal feedback control,” SIAM
Journal on Control and Optimization, vol. 43, no. 5, pp. 1714–1736, 2005.
[30] D. P. Bertsekas, Dynamic programming and optimal control, vol. 1. Athena Scientific Belmont, 1995.
[31] D. Ferguson, T. M. Howard, and M. Likhachev, “Motion planning in urban environments,” Journal of Field Robotics, vol. 25, pp. 939–
960, 2008.
[32] M. Montemerlo, J. Becker, S. Bhat, H. Dahlkamp, D. Dolgov, S. Ettinger, D. Haehnel, T. Hilden, G. Hoffmann, B. Huhnke, and et al.,
“Junior: The stanford entry in the urban challenge,” Journal of Field Robotics, vol. 25, pp. 569–597, 2008.
[33] Y. Kuwata, G. A. Fiore, J. Teo, E. Frazzoli, and J. P. How, “Motion planning for urban driving using RRT,” in IEEE/RSJ International
Conference on Intelligent Robots and Systems, pp. 1681–1686, 2008.
| 3cs.SY
|
Polarization-Division Multiplexing Based on the
Nonlinear Fourier Transform
J AN -W ILLEM G OOSSENS , 1,2,* M ANSOOR I. YOUSEFI , 2 Y VES
J AOUËN 2 AND H ARTMUT H AFERMANN 1
1
2
Mathematical and Algorithmic Sciences Lab, Paris Research Center, Huawei Technologies France
Communications and Electronics Department, Telecom ParisTech, Paris, 75013, France
arXiv:1707.08589v1 [cs.IT] 26 Jul 2017
* [email protected]
Abstract: Polarization-division multiplexed (PDM) transmission based on the nonlinear Fourier
transform (NFT) is proposed for optical fiber communication. The NFT algorithms are generalized
from the scalar nonlinear Schrödinger equation for one polarization to the Manakov system
for two polarizations. The transmission performance of the PDM nonlinear frequency-division
multiplexing (NFDM) and PDM orthogonal frequency-division multiplexing (OFDM) are
determined. It is shown that the transmission performance in terms of Q-factor is approximately
the same in PDM-NFDM and single polarization NFDM at twice the data rate and that the
polarization-mode dispersion does not seriously degrade system performance. Compared with
PDM-OFDM, PDM-NFDM achieves a Q-factor gain of 6.4 dB. The theory can be generalized to
multi-mode fibers in the strong coupling regime, paving the way for the application of the NFT
to address the nonlinear effects in space-division multiplexing.
© 2017 Optical Society of America
OCIS codes: (060.2330) Fiber optics communications,(060.4230) Multiplexing, (060.4370) Nonlinear optics, fibers
References and links
1. A. Hasegawa and F. Tappert, “Transmission of stationary nonlinear optical pulses in dispersive dielectric fibers. I.
anomalous dispersion,” Appl. Phys. Lett. 23, 142–144 (1973).
2. A. Hasegawa and T. Nyu, “Eigenvalue communication,” J. Lightwave Technol. 11, 395–399 (1993).
3. C. S. Gardner, J. M. Greene, M. D. Kruskal, and R. M. Miura, “Method for solving the Korteweg-deVries equation,”
Phys. Rev. Lett. 19, 1095–1097 (1967).
4. M. I. Yousefi and F. R. Kschischang, “Information transmission using the nonlinear Fourier transform, part I:
Mathematical tools,” IEEE Trans. Inf. Theory 60, 4312–4328 (2014).
5. M. I. Yousefi and F. R. Kschischang, “Information transmission using the nonlinear Fourier transform, part II:
Numerical methods,” IEEE Trans. Inf. Theory 60, 4329–4345 (2014).
6. M. I. Yousefi and F. R. Kschischang, “Information transmission using the nonlinear Fourier transform, part III:
Spectrum modulation,” IEEE Trans. Inf. Theory 60, 4346–4369 (2014).
7. S. T. Le, J. E. Prilepsky, and S. K. Turitsyn, “Nonlinear inverse synthesis for high spectral efficiency transmission in
optical fibers,” Opt. Express 22, 26720–26741 (2014).
8. S. T. Le, J. E. Prilepsky, and S. K. Turitsyn, “Nonlinear inverse synthesis technique for optical links with lumped
amplification,” Opt. Express 23, 8317–8328 (2015).
9. M. I. Yousefi and X. Yangzhang, “Linear and nonlinear frequency-division multiplexing,”
https://arxiv.org/abs/1603.04389 .
10. I. Tavakkolnia and M. Safari, “Signaling on the continuous spectrum of nonlinear optical fiber,”
https://arxiv.org/abs/1704.05537 .
11. S. T. Le, I. Philips, J. Prilepsky, P. Harper, N. Doran, A. D. Ellis, and S. Turitsyn, “First experimental demonstration of
nonlinear inverse synthesis transmission over transoceanic distances,” in “Optical Fiber Communications Conference,”
(IEEE, 2016), pp. 1–3.
12. S. T. Le, I. D. Philips, J. E. Prilepsky, M. Kamalian, A. D. Ellis, P. Harper, and S. K. Turitsyn, “Achievable information
rate of nonlinear inverse synthesis based 16QAM OFDM transmission,” in “2016 64nd European Conference on
Optical Communication (ECOC),” (VDE, 2016), pp. 1–3.
13. S. T. Le, H. Buelow, and V. Aref, “Demonstration of 64x0.5Gbaud nonlinear frequency division multiplexed
transmission with 32QAM,” in “Optical Fiber Communication Conference,” (Optical Society of America, 2017), pp.
W3J–1.
14. V. Aref, H. Bülow, K. Schuh, and W. Idler, “Experimental demonstration of nonlinear frequency division multiplexed
transmission,” in “2015 41st European Conference on Optical Communication (ECOC),” (IEEE, 2015), pp. 1–3.
15. V. Aref, S. T. Le, and H. Buelow, “Demonstration of fully nonlinear spectrum modulated system in the highly
nonlinear optical transmission regime,” in “2016 42nd European Conference on Optical Communication (ECOC),”
(VDE, 2016), pp. 1–3.
16. A. Maruta and Y. Matsuda, “Polarization division multiplexed optical eigenvalue modulation,” in “2015 International
Conference on Photonics in Switching (PS),” (IEEE, 2015), pp. 265–267.
17. X. Yangzhang, M. Yousefi, A. Alvarado, D. Lavery, and P. Bayvel, “Nonlinear frequency-division multiplexing in the
focusing regime,” in “Optical Fiber Communication Conference,” (Optical Society of America, 2017), pp. Tu3D–1.
18. M. I. Yousefi, G. Kramer, and F. R. Kschischang, “Upper bound on the capacity of the nonlinear Schrödinger channel,”
https://arxiv.org/abs/1502.06455 .
19. P. Serena, N. Rossi, and A. Bononi, “Nonlinear penalty reduction induced by PMD in 112 Gbit/s WDM PDM-QPSK
coherent systems,” in “2009 35th European Conference on Optical Communication,” (2009), pp. 1–2.
20. I. Tavakkolnia and M. Safari, “Effect of PMD on the continuous spectrum of nonlinear optical fibre,” in “2017 65th
European Conference on Optical Communication (ECOC),” (2017), pp. ci–p–11.
21. C. R. Menyuk, “Pulse propagation in an elliptically birefringent Kerr medium,” IEEE J. Quantum. Electron. 25,
2674–2682 (1989).
22. P. Wai and C. R. Menyuk, “Polarization mode dispersion, decorrelation, and diffusion in optical fibers with randomly
varying birefringence,” Lightwave Technology, Journal of 14, 148–157 (1996).
23. D. Marcuse, C. Manyuk, and P. Wai, “Application of the Manakov-PMD equation to studies of signal propagation in
optical fibers with randomly varying birefringence,” J. Lightwave Technol. 15, 1735–1746 (1997).
24. C. R. Menyuk and B. S. Marks, “Interaction of polarization mode dispersion and nonlinearity in optical fiber
transmission systems,” J. Lightwave Technol. 24, 2806–2826 (2006).
25. M. J. Ablowitz and H. Segur, Solitons and the inverse scattering transform (SIAM, 1981).
26. S. V. Manakov, “On the theory of two-dimensional stationary self-focusing of electromagnetic waves,” Soviet
Physics-JETP 38, 248–253 (1974).
27. A. Degasperis and S. Lombardo, “Integrability in action: Solitons, instability and rogue waves,” in “Rogue and Shock
Waves in Nonlinear Dispersive Media,” (Springer, 2016), pp. 23–53.
28. G. P. Agrawal, Nonlinear fiber optics (Academic press, 2007).
29. S. Civelli, E. Forestieri, and M. Secondini, “Why noise and dispersion may seriously hamper nonlinear frequencydivision multiplexing,” https://arxiv.org/abs/1705.06779 .
30. L. F. Mollenauer, S. G. Evangelides, and H. A. Haus, “Long-distance soliton propagation using lumped amplifiers
and dispersion shifted fiber,” J. Lightwave Technol. 9, 194–197 (1991).
31. K. Blow and N. Doran, “Average soliton dynamics and the operation of soliton systems with lumped amplifiers,”
IEEE Photonics Technol. Lett. 3, 369–371 (1991).
32. P. Wai, C. R. Menyuk, and H. Chen, “Stability of solitons in randomly varying birefringent fibers,” Opt. Lett. 16,
1231–1233 (1991).
33. S. G. Evangelides, L. F. Mollenauer, J. P. Gordon, and N. S. Bergano, “Polarization multiplexing with solitons,” J.
Lightwave Technol. 10, 28–35 (1992).
34. M. Bertolini, N. Rossi, P. Serena, and A. Bononi, “Do’s and don’ts for a correct nonlinear PMD emulation in 100Gb/s
PDM-QPSK systems,” Opt. Fiber Technol. 16, 274–278 (2010).
35. M. Eberhard and C. Braimiotis, Numerical Implementation of the Coarse-Step Method with a Varying DifferentialGroup Delay (Springer US, Boston, MA, 2005), pp. 530–534.
36. C. D. Poole, J. H. Winters, and J. A. Nagel, “Dynamical equation for polarization dispersion,” Opt. Lett. 16, 372–374
(1991).
37. S. Mumtaz, R.-J. Essiambre, and G. P. Agrawal, “Nonlinear propagation in multimode and multicore fibers:
Generalization of the Manakov equations,” J. Lightwave Technol. 31, 398–406 (2013).
1.
Introduction
Nonlinear frequency-division multiplexing (NFDM) is an elegant method to address the nonlinear
effects in optical fiber communication. The scheme can be viewed as a generalization of
communication using fiber solitons [1] and goes back to the original idea of eigenvalue
communication [2].
In this approach, information is encoded in the nonlinear spectrum of the signal, defined
by means of the nonlinear Fourier transform (NFT), [3–5]. The evolution of the nonlinear
spectral components in fiber is governed by simple independent equations. As a result, the
combined effects of the dispersion and nonlinearity can be compensated in the digital domain by
the application of an all-pass-like filter. Furthermore, interference-free communication can be
achieved in network environments.
The nonlinear spectrum consists of a continuous part and, in the case of the anomalous dispersion
fiber, also a discrete part (solitonic component). In principle all degrees-of-freedom can be
modulated. Prior work has mostly focused on either continuous spectrum modulation [6–10],
or discrete spectrum modulation [6]. Transmission based on NFT has been experimentally
demonstrated for the continuous spectrum [11–13], discrete spectrum [14], as well as the full
spectrum [15].
Modern coherent optical fiber systems are based on polarization-division multiplexed (PDM)
transmission to improve the achievable rates. However, research on data transmission using the
NFT is limited to the scalar nonlinear Schrödinger equation, which does not take into account
polarization effects (with the exception of [16]).
In this paper we overcome this limitation by generalizing the NFDM to the Manakov system,
proposing PDM-NFDM. We develop a stable and accurate algorithm to compute the forward
and inverse NFT of a two-dimensional signal. It is shown that the PDM-NFDM based on the
continuous spectrum modulation is feasible, and that the data rate can be approximately doubled
compared to the single polarization NFDM. Compared to the PDM-OFDM, the PDM-NFDM
exhibits a peak Q-factor gain of 6.4 dB, in a system with 25 spans of 80 km standard single-mode
fiber and 16 QAM.
In this paper, we set the discrete spectrum to zero and modulate only the continuous spectrum.
Even though not all available degrees of freedom are modulated, the achievable rates reach
remarkably close to the upper bound [17, 18].
The differential group delay and randomly varying birefringence give rise to temporal pulse
broadening due to polarization-mode dispersion (PMD). The PMD might be compensated in
conventional systems, and may even be beneficial in reducing the nonlinear distortions [19]. On
the other hand, PMD changes the nonlinear interaction between signals in the two polarizations.
The impact of PMD on NFDM is not fully investigated yet [20]. In this paper, we show that linear
polarization effects can be equalized in PDM-NFDM at the receiver using standard techniques,
and that the PMD does not seriously degrade performance.
2.
Channel Model with Two Polarizations
Light propagation in two polarizations in optical fiber is modeled by the coupled nonlinear
Schrödinger equation (CNLSE) [21]. The fiber birefringence usually varies rapidly and randomly
along the fiber in practical systems (on a scale of 0.3 to 100 meters). Under this assumption, the
averaging of the nonlinearity in the CNLSE leads to the Manakov-PMD equation [22–24]:
∂A
∆β0
∆β1
∂A
=− j
σZ A −
σZ
∂Z
2
2
∂T
2
α
β2 ∂ A
8
2
(1)
− A+ j
− jγ A A.
2
2 ∂T 2
9
Here A ≡ A(Z, T) is the 2 × 1 Jones vector containing the complex envelopes A1 and A2 of the
two polarization components, Z denotes the distance along the fiber, T represents time, σZ is a
2 × 2 Pauli matrix (depending on the state of polarization at Z), and ∆β0 , ∆β1 , α, β2 and γ are
constant numbers.
The term ∂A/∂T is responsible for the PMD [24], while the second line represents loss,
chromatic dispersion and Kerr nonlinearity. The factor 8/9 in front of the nonlinear term stems
from polarization averaging.
The NFT applies to the integrable equations. However, the Manakov-PMD system (1) apparently
is not integrable in the presence of loss or PMD. The loss may be approximately compensated
using ideal distributed Raman amplification, leading to a lossless model with distributed additive
white Gaussian noise (AWGN). Discrete amplification with short spans can also be modeled
similarly, but with a modified nonlinearity coefficient (depending on loss); see [8]. Ignoring the
PMD and, for the moment, noise, (1) is reduced to
8
∂A
β2 ∂ 2 A
2
=j
− jγ A A.
∂Z
2 ∂T 2
9
(2)
It is convenient to normalize (2). Let Z0 = 1 and
T0 =
q
| β2 | Z0 /2,
A0 =
r .
8
2
γZ0 .
9
(3)
Introducing the normalized variables z = Z/Z0 , t = T/T0 and q = A/A0 , (2) is simplified to the
Manakov equation
j
∂q ∂ 2 q
2
=
− 2s q q,
∂z ∂t 2
(4)
where s = 1 in the normal dispersion (defocusing) regime and s = −1 in the anomalous dispersion
(focusing or solitonic) regime.
It is shown in [25] that (4) is integrable. In what follows, we develop the corresponding NFT.
3.
NFT of the Two-Dimensional Signals
In this section, basics of the NFT theory of the two-dimensional signals, as well as numerical
algorithms to compute the forward and inverse NFT, are briefly presented [26].
3.1.
Brief Review of the Theory
Equation (4) can be represented by a Lax pair L̂ and M̂. This means that, operators L̂ and M̂ can
be found such that (4) is in one-to-one correspondence with the Lax equation ∂ L̂/∂z = [ M̂, L̂].
The Lax pair for the Manakov equation was found by Manakov in 1974 [26, 27]. The operator L̂
is:
∂
−q1 −q2
© ∂t∗
ª
∂
L̂ = j sq1 − ∂t
(5)
0 ®.
∂
∗
0
− ∂t
«sq2
¬
To simplify the presentation, consider the focusing regime with s = −1.
The eigenvalue problem
L̂v = λv
(6)
can be solved for the Jost function v (eigenvector) assuming that the signals vanish at t = ±∞.
This gives rise to six boundary conditions for v, denoted by j± :
j±(0) → e(0) exp(− jλt),
j±(i) → e(i) exp( jλt),
i = 1, 2,
as t → ±∞,
(7)
where e(k) are unit vectors, i.e., el(k) = δkl , k, l = 0, 1, 2. Each of the boundary conditions (7) is
bounded when λ ∈ C+ or λ ∈ C− . The eigenvalue problem (6) under the boundary conditions
(7) can be solved, obtaining six Jost functions { j±(i) (t, λ)}i=0,1,2 for all t. It can be shown that
{ j+(i) (t, λ)}i=0,1,2 and { j−(i) (t, λ)}i=0,1,2 each form an orthonormal basis for the solution space of
(6). Thus we can expand j−(0) (t, λ) in the basis of { j+(i) (t, λ)}i=0,1,2 :
j−(0) (t, λ) = a(λ) j+(0) (t, λ) + b1 (λ) j+(1) (t, λ) + b2 (λ) j+(2) (t, λ),
(8)
where a(λ), b1 (λ) and b2 (λ) are called nonlinear Fourier coefficients. It can be shown that in the
focusing regime the inner product of two Jost functions corresponding to the same eigenvalue
does not depend on time. This implies that a(λ) and bi (λ) do not depend on time (as the notation
in (8) suggests). The NFT of the q = [q1, q2 ] is now defined as
NFT(q)(λ) =
q̂i =
q̃i =
bi (λ)
a(λ) ,
bi (λ j )
a0 (λ j ) ,
λ ∈ R,
λ j ∈ C+,
i = 1, 2,
(9)
where λ j , j = 1, 2, · · · , N, are the solutions of a(λ j ) = 0 in λ j ∈ C+ .
An important property of NFT(q(t, z))(λ) is that it evolves in distance according to the
all-pass-like filter
2
H(λ) = e4jsλ L .
(10)
Remark. Note that if q1 or q2 is set to zero in the Manakov equation and the associated L̂
operator (5), the equation and the operator are reduced, respectively, to the scalar NLSE and
the corresponding L̂ operator [17]. Also, the theory in this section can straightforwardly be
generalized to include any number of signals qi , i = 1, 2, ..., for instance, LP modes propagating
in a multi-mode fiber in the strong coupling regime.
Remark Since the Jost vectors are orthonormal at all times, (8) implies the unimodularity
condition
|a(λ)| 2 + |b1 (λ)| 2 + |b2 (λ)| 2 = 1,
(11)
which will be used in the algorithm.
3.2.
Forward NFT Algorithm
We develop the NFT algorithms for the continuous spectrum that is considered in this paper. The
forward and inverse NFT algorithms are, respectively, based on the Ablowitz-Ladik and discrete
layer peeling (DLP) methods. These algorithms generalize the corresponding ones in [9].
We begin by rewriting the eigenvalue problem L̂v = λv in the form ∂v/∂t = Pv, where
− jλ
©
P = −q1∗ (t)
∗
«−q2 (t)
q1 (t)
jλ
0
q2 (t)
ª
0 ®.
jλ ¬
(12)
Here and in the remainder of this section we suppress the dependence on the coordinate z. We
discretize the time interval [T0, T1 ] according to t[k] = T0 + k∆T, where ∆T = (T1 − T0 )/N such
that t[0] = T0 and t[N] = T1 . We set qi [k] = qi (T0 + k∆T) and similarly for the vector v. The
Ablowitz-Ladik method is a discretization of ∂v/∂t = Pv as follows
z 1/2
©
v[k + 1, λ] = ck −Q1 [k]∗
∗
«−Q2 [k]
Q1 [k] Q2 [k]
ª
z −1/2
0 ® v[k, λ],
−1/2
0
z
¬
(13)
where Qi [k] = qi [k]∆T, z := e−2jλ∆T and k = 0, 1, · · · , N − 1. With this discretization v
becomesp a periodic function of λ with period π/∆T. We introduced a normalization factor
ck = 1/ |det P[k]| in (13), where |det P[k]| = 1 + |Q2 [k]| 2 + |Q2 [k]| 2 , to improve numerical
stability.
The iterative equation (13) is initialized with
v[0, λ] = j−(0) (T0, λ) = e(0) zT0 /2∆T .
After N iterations, the nonlinear Fourier coefficients are obtained as projections onto the
Jost-solutions j+(i) (T1, λ),
N
T0
a[λ] = z− 2 − 2∆T v0 [N, λ],
bi [λ] = z
N
2
T0
+ 2∆T
vi [N, λ],
(14)
i = 1, 2.
The continuous component in the NFT(q)(λ) is then computed based on (9).
(15)
The algorithm can be efficiently implemented in the frequency domain. Let us write (13) as
1
Q1 [k]z−1
© ∗
z−1
V[k + 1, λ] = ck −Q1 [k]
∗
0
«−Q2 [k]
Q2 [k]z −1
ª
0
® V[k, λ],
z−1 ¬
(16)
where V[k, λ] := (A[k, λ], B1 [k, λ], B2 [k, λ])T , and
A[k, λ] = a[k, λ]
T0
Bi [k, λ] = z−N − ∆T + 2 bi [k, λ],
1
i = 1, 2.
(17)
We discretize λ on the interval [Λ0, Λ1 ] with Λ1 = −Λ0 = π/(2∆T) such that λ[k] = Λ0 + k∆Λ.
Let tilde ’∼’ denote the action of the discrete Fourier transform DFT with respect to λ[k],
e.g., Ã[., l] = DFT(A[., λ[k]]), where
l = 0, 1, , · · · , N − 1 is the discrete frequency. Note that
DFT z −1 Bi [., λ[k]] [l] = shift B̃i [l], where shift denotes circular right shift of the array by
one element. Equation (16) in the frequency domain is
Ã[k + 1, l] = ck ( Ã[k, l] + Q1 [k] shift B̃1 [k] [l]
+ Q2 [k] shift B̃2 [k] [l]),
B̃1 [k + 1, l] = ck (−Q1 [k]∗ Ã[k, l] + shift B̃1 [k] [l]),
(18)
∗
B̃2 [k + 1, l] = ck (−Q2 [k] Ã[k, l] + shift B̃2 [k] [l]).
The initial condition is given by Ã[0] = DFT a[k] [0] and B̃i [0] = 0. At k = N − 1, a and bi are
found by recovering V[N, λ] through an inverse DFT and using (17).
3.3.
Inverse NFT Algorithm
The inverse NFT algorithm consists of two steps. First, we compute v[N, λ] from the continuous
spectra q̂1,2 (λ) and invert the forward iterations (13). Second, at each iteration, we compute Q[k]
from v[k, λ].
Substituting bi (λ) = q̂i (λ)a(λ) in the unimodularity condition (11), we can compute |a(λ)|
q
|a(λ)| = 1/ 1 + | q̂1 (λ)| 2 + | q̂2 (λ)| 2 .
In the absence of a discrete spectrum, from (9), a(λ) , 0 for all λ. We have Re(log(a)) = log(|a|)
and Im(log(a)) = ∠a. Since a(λ) and therefore log(a) are analytic functions of λ we can recover
the phase of a as ∠a = H (log(|a|)), where H denotes the Hilbert transform.
The inverse iterations are obtained by inverting the matrix in (13) and dropping terms of order
Q2i ∼ ∆T 2 (to yield the same accuracy as for the forward iterations):
z −1/2 −Q1 [k] −Q2 [k]
© ∗
ª
v[k, λ] = ck Q1 [k]
(19)
z1/2
0 ® v[k + 1, λ].
∗ [k]
1/2
Q
0
z
« 2
¬
In practice, as in the forward NFT, it is better to invert the frequency domain iterations
(18).
The signal Q[k] can then be obtained from V[k + 1, l] as follows. Recall that shift B̃i [k] [l] =
B̃i [k, l − 1]. Using the initial conditions for the forward iterations Ã[0, l] = δ0,l and B̃i [0, l] = 0,
it is straightforward to show that Ã[k, l] = 0 and B̃i [k, l] = 0 for l ≥ k > 0. In particular,
B̃i [k, N − 1] = 0 for k < N. For the first element of B̃i [k], from (18) and shift B̃1 [k] [−1] =
shift B̃1 [k] [N − 1] = 0, we obtain
Ã[k + 1, 0] = ck Ã[k, 0],
B̃1 [k + 1, 0] = −ck Q1 [k]∗ Ã[k, 0],
B̃2 [k + 1, 0] = −ck Q2 [k]∗ Ã[k, 0].
1
4
^1
q
NFT[INFT(^
q)]1
^2
q
NFT[INFT(^
q)]2
0.9
0.8
^1
q
NFT[INFT(^
q)]1
^2
q
NFT[INFT(^
q)]2
3.5
3
0.7
2.5
qj (a.u.)
j^
qj (a.u.)
j^
0.6
0.5
0.4
2
1.5
0.3
1
0.2
0.5
0.1
0
0
-4
-2
0
2
4
6
-4
-2
0
6 (a.u.)
2
4
6
6 (a.u.)
Fig. 1. Comparing q̂ = [q̂1, q̂2 ] and NFT(INFT(q̂)), where q̂1 and q̂2 are displaced Gaussians.
For fixed sampling rate, the algorithm is less accurate at higher input-power.
These equations can be solved for Q∗i [k] = − B̃i [k + 1, 0]/ Ã[k + 1, 0].
3.4.
Testing the NFT Algorithms
The forward and inverse NFT numerical algorithms are tested as follows. Figure 1 compares
a signal containing two polarization components in the nonlinear Fourier domain with its
reconstruction after successive
INFT and NFT operations. We take two displaced Gaussians with
√
standard deviation σ = 2 as input signals for the two polarization components q̂i , i = 1, 2. We
set time windows to T = 64 and take 2048 samples in time and nonlinear Fourier domain. In the
discrete layer peeling method, the signal is periodic in nonlinear Fourier domain with period
π/∆T ≈ 100.
Deviations occur as the amplitude of the signal is increased. The accuracy of the NFT (INFT)
can be enhanced by increasing the sampling rate in time (nonlinear frequency), as this reduces
errors due to to the piecewise-constant approximation of the signal. We find that we roughly need
twice the number of samples to achieve the same accuracy as in the single-polarization case.
4.
PDM-NFDM Transmitter and Receiver
In this section, we describe the transmitter (TX) and receiver (RX) digital signal processing
(DSP) in PDM-NFDM and PDM-OFDM. A schematic diagram of the polarization-division
multiplexed NFDM and OFDM transmission systems is shown in Fig. 2. The TX DSP produces
digital signals for the in-phase and quadrature components of both polarization components,
which are fed into the IQ-Modulator. The modulated signals for the two polarization components
are combined in the polarization beam combiner before they enter the transmission line consisting
of multiple fiber spans. We consider the practically relevant case of lumped amplification using
Erbium-doped fiber amplifiers (EDFAs). After propagation through the fiber the two polarization
components are separated in the polarization beam splitter and fed into the polarization-diversity
intradyne coherent receiver, which provides the input to the RX DSP.
We briefly describe the TX and RX DSP in OFDM and NFDM. We first map the incoming
bits to signals taken from a QAM constellation. We then oversample the discrete-time signal in
the time domain (by introducing zeros in the frequency domain outside the support of the signal)
and then add guard intervals in the time domain. Increasing the guard interval increases the
accuracy of the INFT. Unless stated otherwise, we do not increase the guard intervals during the
computation of the INFT as the amplitude is increased, so there is a penalty due to inaccuracies
in the algorithm. In NFDM, these steps are performed in what is called the U-domain; see [9].
SSMF
I xin
Tx DSP
in
x
in
y
in
y
Q
I
Q
IQ Mod.
E xin
IQ Mod.
E yin
Tx DSP NFDM
Tx DSP OFDM
OFDM data encoder
OFDM data encoder
Oversample in time
Include guard interval
EDFA
E xout
E yout
N spans
I xout
Qxout
I yout
Q yout
Coh.
Rx
Rx DSP NFDM
Rx DSP OFDM
Separate bursts
Separate bursts
Oversample in time
Normalize
Normalize
Include guard interval
Manakov NFT
FFT
One tap channel eq.
CDC or DBP
U qˆ
Rx DSP
qˆ U
Manakov INFT
IFFT
Unnormalize
Unnormalize
Remove guard interval
Remove guard interval
Combine bursts
Combine bursts
Training sequence equalization
Training sequence equalization
Downsample in time
Downsample in time
Minimum distance decoder
Minimum distance decoder
Fig. 2. System diagram for the polarization-multiplexed NFDM and OFDM transmission
systems with processing steps in the transmitter and receiver DSP. Steps highlighted in
purple require joint processing of both polarization components. For further explanation see
text.
The signal is subsequently mapped from the U-domain to nonlinear Fourier domain through the
transformation
q
Ui (λ) =
− log(1 − | q̂i (λ)| 2 )e j∠q̂i (λ),
i = 1, 2.
(20)
This step has no analogue in OFDM. Note that contrary to the single-polarization case here
the energy of the signal in the time domain is only approximately proportional to the energy in
the U-domain. At this point, we perform the inverse NFT in the case of NFDM and an inverse
FFT in case of OFDM. Then we obtain the unnormalized signal by introducing units using (3).
Finally we combine the OFDM or NFDM bursts to obtain the signal to be transmitted. The output
of the DSP are the in-phase and quadrature components of the signals in the two polarization
components. All steps in the DSP are performed independently for the two polarizations, except
for the INFT, which requires joint processing.
At the RX we invert the steps of the TX DSP. The signal processing begins with burst
separation and signal normalization using (3), followed by the forward NFT (FFT) in case
of NFDM (OFDM). In NFDM, we subsequently equalize the channel using (10), which is a
single-tap phase compensation. Similarly, in OFDM we either compensate the dispersion with
the phase exp(− j β2 ω2 Nspan L/2), or perform digital backpropagation (DBP) for a fixed number
of steps per span. We then remove the guard intervals in time, down-sample the signal in the time
domain, and obtain the output symbols. The bit error rate (BER) is calculated using a minimum
distance decoder for symbols.
In Fig. 3 we compare the time domain signals of NFDM and OFDM at the TX for one of the
polarization components. We use 112 subcarriers over a burst duration of T0 = 2ns in both cases.
These parameters are the same as in Ref. [8] and correspond to a burst data rate of 224GBit/s per
polarization at a Baudrate of 56GBaud. In NFDM the shape of the signal in the time domain
changes as the amplitude | q̂1,2 | is increased in the nonlinear Fourier domain. At a signal power of
0.5 dBm, one can see a significant amount of broadening of the NFDM signal in the time domain.
This effect makes it difficult to control the time duration of signals in NFDM.
0.08
0.08
NFDM
OFDM
0.07
0.05
p
jqj ( W )
0.06
0.05
p
jqj ( W )
0.06
0.04
0.04
0.03
0.03
0.02
0.02
0.01
0.01
0
-20
NFDM
OFDM
0.07
-15
-10
-5
0
5
10
15
0
-20
20
t (ns)
-15
-10
-5
0
5
10
15
20
t (ns)
Fig. 3. NFDM and OFDM signals at the TX (left panel) and RX (right panel) using the ideal
model (2) at P = 0.5 dBm without noise. Only one of the two polarizations is shown at both
TX and RX. The NFDM signal at the TX is broader due to the nonlinear effects in the NFT.
At the RX, the time duration of both signals is about the same.
Parameter
Value
Length span
80 km
Number of spans
25
Number of carriers
112
Burst duration
2 ns
Guard interval
18 ns
Burst data rate
448GBit/s
Effective data rate
44.8GBit/s
Bandwidth
56 GHz
Oversampling factor
16
Constellation
16 QAM
Noise Figure
6.2 dB
Attenuation
0.2 dB/km
Dispersion
16.89 ps/km/nm
Nonlinear index
2.6◦10⁻²⁰ W/m²
Core area
80 μm²
Table 1. The system parameters used in the simulations.
Pulse broadening in the time domain due to the chromatic dispersion can be estimated by [28]
∆T = 2π| β2 |LB,
(21)
where B is the signal bandwidth. The guard time intervals for minimizing the interaction among
bursts can be approximated using (21). For the parameters in Table 1, we obtain ∆T ∼ 15ns, to
which we add a 20% margin to obtain a total symbol duration of T = T0 +Tguard = (2+18)ns= 20ns.
This corresponds to an effective data rate of 44.8Gbit/s. The right panel of Fig. 3 shows the
time domain signals in NFDM and OFDM in one polarization at the RX after propagation
over 2000 km governed by the integrable model (2) (we consider transmission over the realistic
fiber model (1) below in Sec. 5.3). We observe that the amount of the temporal broadening
of the NFDM and OFDM signals is about the same. In our simulations we do not employ
pre-compensation at the TX, which may be added to further increase the data rate [29].
5.
Simulation results
In this section, we consider the system shown in Fig. 2 and compare the PDM-OFDM and
PDM-NFDM via simulations, taking into account loss and PMD. First, we consider the model
with loss and periodic amplification, setting the PMD to zero (namely, neglecting the terms in
the first line of the Eq. 1). Next, the PMD effects are studied in Sec. 5.3. The system parameters
used are summarized in Table 1.
10 0
B2B
Lossless
Lossy
Transf. lossless
BER
10 -1
10 -2
10 -3
10 -4
5
10
15
20
OSNR (dB)
Fig. 4. BER as a function of OSNR for NFDM, for back-to-back, lossless, lossy and
transformed-lossless models.
5.1.
Effect of loss and lumped amplification
In the lumped amplification scheme, the signal is periodically amplified after each span. The
channel is described by the model (1) including the attenuation term, which is manifestly
not integrable. The effect of lumped amplification on NFT transmission has been studied
previously [8].
We illustrate the effect of the attenuation by comparing the BER as a function of the OSNR in
Fig. 4 for four models: 1) the back-to-back (B2B) configuration; 2) lossless model (2) (there is
propagation but without loss); 3) lossy model with lumped amplification; and 4) a transformedlossless model that is introduced below. In producing Fig. 4, we artificially introduced AWGN at
the receiver, to exclude the effects of the signal-noise interaction from the comparison. We fix the
power P = −3.1dBm to keep the accuracy of the NFT algorithms the same and vary the OSNR
by changing the noise power.
We see that the BER in B2B and lossless models are approximately the same. This result
verifies that the implementation of the NFT is correct. It also shows that the effect of burst
interaction due to the finite guard interval is negligible.
The BER of the lossy model is significantly higher than the BER in the B2B and lossless models
at high OSNRs. The BER of the lossy model seems to flatten out as the OSNR is increased.
Fiber loss can be taken into account in the NFT as follows. The attenuation term in (1)
α
can be eliminated by a change of variable A(z, t) = A0 (z, t)e− 2 z [30, 31]. This transforms the
non-integrable equation (1) to the integrable equation (2) (thus with zero loss) and a modified
nonlinearity parameter
γeff (L) =
1
L
∫
L
γe−αz dz = γ(1 − e−αL )/(αL).
(22)
0
We refer to the resulting model as the transformed-lossless model.
Note that the power of the signal A0 (z, t) in the transformed-lossless model is higher than the
power of the A(z, t) in the original lossy model. That is because γeff < γ, which yields higher
amplitudes according to (3).
NLSE
Manakov
12
10
Q-factor (dB)
8
6
4
2
0
-2
-4
-10
-5
0
5
P (dBm)
Fig. 5. Comparison between NFT transmission based on the NLSE (single polarization) and
the Manakov equation (polarization-multiplexed). P denotes the total power of the signal.
At low power the offset between the curves is 3dB, as the signal power doubles when two
polarization components propagate. At higher power signal-noise interactions decrease this
gap.
Figure 4 shows that the proposed scheme for loss cancellation is indeed very effective. Using
γeff (L) in the inverse and forward NFTs, the BER of the lossy and transformed-lossless models
are almost identical.
5.2.
Transmission performance PDM-NFDM
In order to assess the transmission performance of the PDM-NFDM transmission, we performed
system simulations as described in Sec. 4. For now we focus on deterministic impairments and
neglect PMD. We simulate transmission of 112 subcarrier NFDM and OFDM pulses based
on a 16QAM constellation over Nspan = 25 spans of 80km of standard single-mode fiber. We
assume ideal flat-gain amplifiers with a noise figure of NF = 6.2dB. The system parameters are
summarized in Table 1. They are comparable with those of Ref. [8] to facilitate a comparison
with the single-polarization case.
We first compare single-polarization NFT transmission based on the NLSE with polarizationmultiplexed transmission based on the Manakov equation. Figure 5 shows that in the linear regime
the two curves are offset by 3dB because polarization multiplexing doubles the transmission
power without increasing the signal to noise ratio. We note that in this calculation we increased
the guard band and thereby the sampling rate in frequency domain by a factor of two in the
computation of the PDM-NFT (the guard band in transmission is the same) compared to the
single-polarization case. This is to avoid penalties due to inaccuracies in the NFT, which here
amount to approximately 1dB at peak power if the same guard bands are used. Without this
penalty, the Q-factor at peak power is roughly the same in both cases. This implies that data rates
can approximately be doubled using polarization multiplexing. We show in Sec. 5.3 that this
conclusion still holds in presence of polarization effects.
In Fig. 6 we compare the Q-factor as a function of launch power in PDM-OFDM and
PDM-NFDM transmission. The optimum launch power in OFDM is significantly smaller than
in NFDM. OFDM transmission is limited by nonlinearity, while NFDM is mainly limited by
nonlinear signal-noise interaction. The Q-factor at optimum launch power is significantly higher
in NFDM than in OFDM. We observe an overall gain of 6.4dB in Q-factor. Figure 7 shows the
NFDM
OFDM
OFDM DBP 4 step/span
OFDM DBP 10 step/span
OFDM DBP 16 step/span
13
12
Q-factor (dB)
11
10
9
8
7
6
5
4
3
-8
-6
-4
-2
0
2
4
6
P (dBm)
Fig. 6. Comparison of polarization-multiplexed OFDM to NFDM transmission in an idealized
setting. We use the Manakov equation without PMD-effects for 25 spans of 80 km for
16QAM. NFDM performs as well as OFDM with 10step/span DBP. The full potential of
NFDM is leveraged in a multi-channel scenario, where DBP is less efficient.
Fig. 7. Received constellations for OFDM (left panel, without DBP) and NFDM (right panel)
at the respective optimal launch power. The noise in NFDM is clearly non-Gaussian.
corresponding received constellations at the respective optimal launch power.
In Fig. 6 we further report results obtained by digital backpropagation after OFDM transmission.
For a fair comparison with NFDM, we use the same sampling rate in the DBP and apply it without
prior downsampling. OFDM with DBP achieves similar performance to NFDM for around 10
DBP steps per span and exceeds it for 16 steps per span. We note that here we considered a
single-user scenario. The full potential of NFDM is leveraged in a network scenario, where DBP
is less efficient. Our results are in good qualitative agreement with those reported in Ref. [8] for
the single-polarization case, while achieving twice the data rate.
5.3.
5.3.1.
Effect of PMD
Simulation of PMD
The fiber model (1) describes light propagation in linearly birefringent fibers [32, 33]. Here ∆β1
is the difference in propagation constants for the two polarization states which are induced by
fiber imperfections or stress. In real fibers, this so-called modal birefringence varies randomly,
resulting in PMD.
We simulate PMD with the coarse-step method, in which continuous variations of the
12
11
Q-factor (dB)
10
9
8
7
6
p
0 ps= km
p
0.1 ps= km
p
0.2 ps= km
5
4
5
10
15
20
25
30
35
40
Ntaps
Fig. 8. The effect of the number of taps on the Q-factor for different values of DPMD . The
power of the signal in the simulation is 0.5 dBm.
birefringence are approximated by a large number of short fiber sections in which the birefringence
is kept constant. PMD is thus emulated in a distributed fashion [34]. We use fixed-length sections
of length 1km, larger than typical fiber correlation lengths. At the beginning of each section,
the polarization is randomly rotated to a new point on the Poincaré sphere. We apply a uniform
random phase in the new reference frame. The latter accounts for the fact that in reality the
birefringence will vary in the sections where it is assumed constant, which will lead to a random
phase relationship between the two polarization components [32]. The differential group delay
(DGD) of each section is selected randomly from a Gaussian distribution. This way artifacts in
the wavelength domain caused by a fixed delay for all sections are avoided [35].
Within each scattering section, the equation (1) without PMD terms is solved using standard
split-step Fourier integration. To speed up the simulations, we employ a CUDA/C++ based
implementation with a MEX interface to Matlab.
The resulting DGD of the fiber is Maxwell distributed [36],
32 ∆t 2
4∆t 2
p h∆t i (∆t) = 2
exp −
.
(23)
π h∆ti 3
πh∆ti 2
For the
p distribution, the mean is related to the root-mean-square (RMS) delay by
p Maxwell
h∆ti 3π/8 = h∆t 2 i. The average DGD varies with the square root of the fiber length [28],
p
√
h∆ti ∼ h∆t 2 i = DPMD L,
(24)
where DPMD is the PMD parameter. Typical√PMD values for fibers used in telecommunications
range from 0.05 in modern fibers to 0.5 ps/ km.
5.3.2.
PMD impact on NFDM
In this section, we consider polarization effects on the NFT transmission. Since the nonlinear
term in the Manakov equation is invariant under polarization rotations, the equalization can be
done in in the nonlinear Fourier domain (cf. Fig. 2). We use a simple training sequence based
equalization algorithm to compensate linear polarization effects. The samples of the first NFDM
symbol are used as the training sequence. The filter taps of the equalizer are determined using
p
0 ps= km
p
0.2 ps= km
p
0.5 ps= km
No birefringence
12
10
Q-factor (dB)
8
6
4
2
0
-10
-8
-6
-4
-2
0
2
4
6
8
10
P (dBm)
Fig. 9. The effect of PMD √
on PDM-NFDM. We used 13, 25 and 61 taps in the equalizer for
DPMD = 0, 0.2 and 0.5ps/ km, respectively.
least-squares estimation. To obtain the Q-factor we average the BER over 120 random realizations
of PMD for each data point.
Due to its statistical nature, the effects of PMD are often quantified in terms of outage
probabilities. In practice, the system is designed to tolerate a certain amount of PMD, in this case
by fixing the number of taps in the equalizer. When the DGD exceeds this margin, the system is
said to be in outage. We first determine the number of taps to achieve a given outage probability.
Fig. 8 shows the Q-factor as a function of the number of taps for different values of the PMD
parameter. For DPMD = 0 we only have random polarization rotations and no DGD. Hence there
is no interaction between nonlinearity and PMD and we use this case as the reference. The upturn
for a small number of taps is due to the fact that a couple of taps are required to fully compensate
the polarization rotations in presence of noise.
For finite PMD, the curves for different PMD values converge to the same result within error
bars. We can estimate the required number of taps from the Maxwell distribution. In order
for the system to fully compensate PMD in 98.7% of the cases (1.3% outage probability), the
equalization must cover a time interval equal √
to the mean plus 3 standard deviations of the DGD
distribution. For example, for DPMD = 0.1ps/ km this interval corresponds to 12.8ps. Using the
sampling frequency we find that we approximately need 12 taps to converge, consistent with the
figure.
We note that by construction our equalizer also corrects potential phase rotations and therefore
at least part of the nonlinear effects. In order to separate effects due to interaction of nonlinearity
and PMD, we performed a calculation where we reversed the polarization rotations and phases
exactly (instead of using the equalizer), by keeping track of the randomly generated angles and
DGD values in the fiber simulation. We find a small phase rotation consistent with but smaller
than the one reported in [20]. In our case it remains negligible at peak power.
In Fig. 9 we compare the Q-factor as a function of launch power with and without PMD
effects. Here we use the number of taps determined
√ as described above. We observe a penalty of
roughly
0.3dB
at
peak
power
for
D
=
0.2ps/
km relative to the case without PMD (DPMD =
PMD
√
0ps/ km). The penalty is therefore not serious for typical fibers used in telecommunication.
Compared to the case without PMD effects and random birefringence (labeled “no birefringence"),
we find a penalty of roughly 1.2dB at peak power for the case of zero PMD due to the equalization.
6.
Conclusions
In this paper, we have proposed polarization-division multiplexing based on the nonlinear Fourier
transform. NFT algorithms are developed based on the Manakov equations. Our simulations
demonstrate feasibility of polarization multiplexed NFDM transmission over standard single-mode
fiber. The results show that data rates can approximately be doubled in polarization-multiplexed
transmission compared to the single-polarization case. This is an important step to achieve data
rates that can exceed those of conventional linear technology.
Numerical simulations of polarization multiplexed transmission over a realistic fiber model
including randomly varying birefringence and polarization mode dispersion have shown that
penalties due to PMD in real fibers do not seriously impact system performance. Our fiber
simulations are based on the Manakov equations. As a next step, which is beyond the scope of
this paper, the results should be verified experimentally.
The equation governing light propagation of N modes in multi-mode fibers in the strong
coupling regime can be written in the form [37]
∂A
β̄2 ∂ 2 A
2
=j
− jγκ A A,
∂Z
2 ∂T 2
(25)
where A = (A1, . . . , A N )T and β̄2 denotes the average group velocity dispersion. The NFT, as
well as the algorithms presented here, generalize to this equation in a straightforward manner.
Our approach therefore paves the way to combining the nonlinear Fourier transform with
space-division multiplexing.
7.
Acknowledgements
J-W.G. would like to thank Wasyhun A. Gemechu for helpful discussions. H.H. acknowledges
useful discussions with Djalal Bendimerad, Majid Safari, Sergei K. Turitsyn and Huijian Zhang.
| 7cs.IT
|
arXiv:math/0202276v1 [math.NA] 26 Feb 2002
A numerical method for solution of ordinary
differential equations of fractional order
Jacek Leszczynski, Mariusz Ciesielski
Technical University of Czestochowa, Institute of Mathematics & Computer Science,
ul. Dabrowskiego 73, 42-200 Czestochowa, Poland
[email protected], [email protected]
Abstract. In this paper we propose an algorithm for the numerical
solution of arbitrary differential equations of fractional order. The algorithm is obtained by using the following decomposition of the differential
equation into a system of differential equation of integer order connected
with inverse forms of Abel-integral equations. The algorithm is used for
solution of the linear and non-linear equations.
1
Introduction
In opposite to differential equations of integer order, in which derivatives depend
only on the local behaviour of the function, fractional differential equations accumulate the whole information of the function in a weighted form. This is so
called memory effect and has many applications in physics [10], chemistry [5],
engineering [3], etc. For that reason we need a method for solving such equations which will be effective, easy-to-use and applied for the equations in general
form. However, known methods used for solution of the equations have more
disadvantages. Analytical methods, described in detail in [7,12], do not work in
the case of arbitrary real order. Another analytical method [9], which uses the
multivariate Mittag-Leffler function and generalizes the previous results, can be
used only for linear type of equations. On the other hand, for specific differential
equations with oscillating and periodic solution there are some specific numerical
methods [6,13,14]. Other numerical methods [2,4] allow solution of the equations
of arbitrary real order but they work properly only for relatively simple form of
fractional equations.
Let us consider an initial value problem for the fractional differential equation
a1 τ Dtα1 y(t) + a2 τ Dtα2 y(t) + . . . + an τ Dtαn y(t) + an+1 y(t) = f (t)
(1)
connected with initial conditions
y (k) (τ ) = bk ,
(2)
where 0 < t ≤ T < ∞, ai ∈ IR, a1 6= 0, αi ∈ IR+ , i = 1, . . . , n, mi −1 ≤ αi < mi ,
mi ∈ IN, αl > αl+1 for l = 1 , . . . , n−1, bk ∈ IR, k = 0 , . . . , m1 −1, f (t) is a given
function defined on the interval [0, T ], y(t) is the unknown function which is the
solution of eqn. (1).
The fractional derivative operator Dαi is defined in the Riemann-Liouville
sense [12]. We also have other definitions of the operator like Caputo [9],
Grünwald-Letnikov [9], and Weyl-Marchaud [12]. Regarding to the RiemannLiouville operator we can define it as the left-side operator
(a Dtα y)(t)
1
dm
=
Γ (m − α) dtm
Zt
y(ξ)dξ
,
(t − ξ)α−m+1
(3)
a
where a < t ≤ T < ∞, m−1 < α ≤ m, m ∈ IN. In our study, we also use integral
operators defined in the Riemann-Liouville sense [12]. When y(t) ∈ L1 (a, b) and
α > 0 then left-side integral operator is defined as
(a Itα y)(t)
1
=
Γ (α)
Zt
y(ξ)dξ
, t > a.
(t − ξ)1−α
(4)
a
More informations concerning the operators properties one can find in literature [7,9,12]. Applying definitions (3) and (4) we have the following property
(a Dtα y)(t) = Dm (a Itm−α y)(t),
(5)
where Dm represents typical derivative operator of integer order m. According
to [9] in the fractional integrals the composition rule occurs
β
α
a It (a It y(t))
= a Itα+β y(t) = a Itβ+α y(t) = a Itβ (a Itα y(t)) .
(6)
In the general case the Riemann-Liouville fractional derivatives do not commute
(a Dtα (a Dtβ y))(t) 6= (a Dtα+β )(t) 6= (a Dtβ (a Dtα y))(t) .
(7)
Extending our considerations, the integer operator commutes with the fractional
operator
(Dtm (a Dtα y))(t) = (a Dtm+α )(t) ,
(8)
but the opposite property is impossible. The mixed operators (derivatives and
integrals) commute only in the following way
(a Dtα (a Itβ y))(t) = (a Dtα−β )(t) ,
(9)
but they do not commute in the opposite way.
We turn our attention to the composition rule in two following ways. First of
all we found in literature [10,11] the fact, that authors neglect the general property of fractional derivatives given by eqn. (7). Regarding to solution of fractional
differential equations we will apply above properties in the next sections.
2
Mathematical background
In this section we concentrate on describing the method which is decomposition
of fractional differential equation into a system of one ordinary differential equation and inverse forms of Abel-integral equations. We focus on decomposition
ways of arbitrary fractional differential equation (1) with initial conditions (2). In
our considerations we classify the fractional differential equations in dependence
on the fractional derivative occurrence as:
– one-term equations, in which Dα occurs only one,
– multi-term equations, in which Dα occurs much more times.
Corollary 1. An arbitrary one-term equation
a1 τ Dtα y(t) + a2 y(t) = f (t) , 0 ≤ m − 1 ≤ α < m, m ∈ IN ,
(10)
in which a1 6= 0 and initial conditions y (k) (τ ) = bk , (k = 0 , . . . , m − 1) should
be taken into account, can decomposed into the following system
a2
1
m
Dt z(t) = − a1 y(t) + a1kf (t)
P
m−1
(t−τ )
m−α
(0)
z(t)
y (t) = k=0 bk k! k + τ Dt
P
m−1
(t−τ
)
m−α+1
(1)
.
(11)
z(t)
y (t) = k=1 bk k! + τ Dt
...
y (m−1) (t) = bm−1 (t−τ )m−1 + τ D2m−α−1 z(t)
t
(m−1)!
Proof. We use the property (5) and we introduce a new variable z(t) which we
call a temporary function. For such introduction we have
α
τ Dt y(t)
= Dtm z(t) and z(t) = (τ Itm−α y)(t).
The new variable represents the Abel integral equation which solution, found in
literature [12], is the left inverse operator
y(t) = (τ Dtm z)(t).
In dependence on the integer order m we introduce initial conditions bk which
)k
are multiplied by a term (t−τ
k! . The term issues from a kernel of the left-side
Riemman-Liouville operator (3).
⊓
⊔
In the system of eqns. (11), the first equation is the differential equation of m
integer order and the next equations represent inverse forms of Abel-integral
equations. We also found in literature [2,4] different or similar approaches for
solution of the eqn. (10).
In mathematical point of view, more interesting is the second class called
multi-term equations. Let us consider the eqn. (1) with initial conditions (2). As
in previous class we use the rule (5). Following the previous corollary we apply
a system of new variables zi (t) = (τ Itmi −αi y)(t) and we decompose eqn. (1) into
a system of the following equations
a1 Dtm1 z1 (t) + a2 Dtm2 z2 (t) + . . . + an Dtmn zn (t)+
+an+1 y(t) = f (t)
z1 (t) = (τ Itm1 −α1 y)(t)
.
(12)
z2 (t) = (τ Itm2 −α2 y)(t)
...
zn (t) = (τ Itmn −αn y)(t)
Such system is not a final form and requires additional transformations. In multiterm class we distinguish two sub-classes depending on αi or follows the decomposition on mi :
– independent subclass, in which mi are different derivatives orders of integer
type mi 6= mi+1 for i = 1, . . . , n − 1,
– dependent subclass, in which one can find some dependence between the
derivatives order of integer type, e.g. m1 = m2 = m3 or m1 = m4 , etc.
Proposition 1. Taking into consideration the independent subclass it is possible
to formulate the unique solution of eqn. (12) in the following form
m1
m2
mn
1
Dt z1 (t) = − a1 1(a2 Dt z2 (t) + . . . + an Dt zn (t)) +
+ a1 (−an+1 y(t) + f (t))
m1 −m2 −α1 +α2
z
(t)
=
D
z1 (t)
2
0
t
m1 −m3 −α1 +α3
z
(t)
=
D
z
3
0
1 (t)
t
...
.
(13)
zn (t) = 0 Dtm1 −mn −α1 +αn z1 (t)
y (0) (t) = Pm1 −1 b (t−τ )k + Dm1 −α1 z (t)
0 t
1
k
k=0
k!
Pm
(t−τ )k
m1 −α1 +1
1 −1
(1)
z1 (t)
y (t) = k=1 bk k! + 0 Dt
.
.
.
(t−τ )m1 −1
y (m1 −1) (t) = b
+ D2m1 −α1 −1 z (t)
m1 −1 (m1 −1)!
0
t
1
Proof. In this proof we try to show a dependence of temporary functions zi (t)
(i = 2, ..., n) on the temporary function z1 (t). We assume, that initial conditions
bk = 0. In such way we can easily construct the proof. We observe that the Abel
integral operators in (12) have the order 0 < mi − αi < 1. We consider the
inverse form of such operators and for two operators we have
β
τ Dt z2 (t)
= τ Dtγ z1 (t),
(14)
where 0 < β < 1 and 0 < γ < 1 respectively. Using the property (5) we obtain
Dt1 (τ It1−β z2 (t)) = Dt1 (τ It1−γ z1 (t)).
Neglecting the operator Dt1 we can obtain the following formula
1−β
z2 (t)
τ It
= τ It1−γ z1 (t) + C1,2 ,
(15)
where C1,2 is an arbitrary constant. Additionally, we multiply the operator τ Itβ
for eqn. (15) and applying the property (6) we obtain
1
τ It z2 (t)
= τ It1−γ+β z1 (t) + τ Itβ C1,2 .
(16)
Differentiating both sides by the operator Dt1 and applying the property (9) (see
literature [7,9]) we obtain the final dependence
z2 (t) = τ Dtγ−β z1 (t) + C1,2 ·
β
.
Γ (1 + β) · t1+β
(17)
In general way we state that zi (t) (i = 2, ..., n) depends on z1 (t) together with
additional term represented by the constants C1,i .
⊓
⊔
Proposition 2. All constants C1,i = 0 (i = 2, ..., n) in eqn. (17) are equaled to
zero for τ = 0.
Proof. Taking into consideration the formula (17) we can calculate the constants
C1,i =
Γ (1 + β) · t1+β
· zi (t) − τ Dtγ−β z1 (t) .
β
Additionally, putting the initial condition t = τ = 0 we obtain all C1,i = 0.
⊓
⊔
Remark 1. According to previous considerations, the initial conditions for temporary functions zi (t) satisfy the following dependence
zi (τ = 0) = 0, i = 1, . . . , n .
(18)
In practical point of view given by formula (18), we can solve ordinary differential
equation of integer type with variable coefficients under zeros initial conditions
of temporary functions zi (t).
Let us consider the next subclass of a fractional differential equation. We
assume, that some integer orders mi are the same. This may happen in the
case when some αi and αi+1 , (i = 1, . . . , n − 1) belong to the same integer
values mi = mi+1 . We introduce a parameter r which denotes a number of
temporary functions having the same derivative operator Dm1 of integer type.
If m1 = m2 = . . . = mr for different temporary functions zi (t), (i = 1, . . . , r)
then it can express the functions zi (t), (i = 2, . . . , r) through the z1 (t) function.
It may observe a relationship in eqn. (13) for z2 (t) and z1 (t) functions respectively z2 (t) = τ Dtm1 −m2 −α1 +α2 z1 (t). We assume an equalities of the integer
orders m1 = m2 . The previous relationship becomes z2 (t) = τ Dt−α1 +α2 z1 (t).
Following the property (τ Dt−α y)(t) = (τ Itα y)(t) we have zi (t) = τ Itα1 −αi z1 (t),
(i = 2, . . . , r). In the ordinary differential equation presented in the system (13)
we change all zi (t), (i = 2, . . . , r) which depends on z1 (t). Moreover, we introduce
a new temporary function in the following form
w(t) = z1 (t) +
a2 α1 −α2
ar α1 −αr
z1 (t) + . . . +
z1 (t).
τI
τI
a1 t
a1 t
(19)
We need to find an inverse form of eqn. (19) as
z1 (t) = (1 +
a2 α1 −α2
ar α1 −αr −1
+ ...+
) w(t) ,
τI
τI
a1 t
a1 t
(20)
where (1 + aa21 τ Itα1 −α2 + . . . + aar1 τ Itα1 −αr )−1 denotes the left inverse operator to
the operator (1+ aa21 τ Itα1 −α2 +. . .+ aar1 τ Itα1 −αr ). We can apply the known Laplace
transform for eqn. (19) that to find the left inverse operator in expression (20).
The formula (20) is not suitable for practical purposes and computations.
Proposition 3. Connecting expressions (13) and (19)÷(20), the solution of the
second subclass of the fractional differential equation is presented as follows:
m1
m
Dt w(t) = − a11 (ar+1 Dt r+1 zr+1 (t) + . . . +
m
n
+an Dt zn (t) + an+1 y(t)) + a11 f (t)
z1 (t) = (1 + a2 0 I α1 −α2 + . . . + ar 0 I α1 −αr )−1 w(t)
t
t
a1
a1
m1 −mr+1 −α1 +αr+1
z
(t)
z
(t)
=
D
1
r+1
0
t
...
.
(21)
zn (t) = 0 Dtm1 −mn −α1 +αn z1 (t)
k
P
m1 −1
(t−τ )
m1 −α1
(0)
z1 (t)
y (t) = Pk=0 bk k! k + 0 Dt
m
−1
(t−τ
)
m1 −α1 +1
1
(1)
+
D
z1 (t)
b
y
(t)
=
0 t
k
k=1
k!
.
.
.
(m1 −1)
)m1 −1
2m1 −α1 −1
z1 (t)
y
(t) = bm1 −1 (t−τ
(m1 −1)! + 0 Dt
We do not need to proof the above proposition because it was done in the
previous subclass of the multi-term class.
3
Numerical treatment and examples of calculations
In this section we propose an explicit numerical scheme. For numerical solution
of the differential equation of integer order we apply one-step Euler’s method [8].
There is no limits to extend above approach to the Runge-Kutta method. According to results presented by Oldham and Spanier [7] we use numerical scheme
for integral operator as
i−1
α
X
h
α
z0 (iα − (i − 1)α ) + zi +
zi−j ((j + 1)α − (j − 1)α ) ,
0 It zi =
2Γ (1 + α)
j=1
(22)
which is valid for arbitrary α > 0. We also use an algorithm given by Oldham
and Spanier [7] for numerical differentiation. The algorithm depends on α range.
For 0 ≤ α < 1 we have
i−1
−α
X
h
α
(1 − α)z0 +
(zi−j − zi−(j+1) ) ((j + 1)1−α − j 1−α ) .
0 Dt z i =
Γ (2 − α)
iα
j=0
(23)
In literature [7] one may find the algorithm for highest values.
For limited practical solution of eqn. (20) we found in literature Babenko’s
method [1]. In general form of eqn. (20) we cannot consider the left inverse operator. For that, we found only a simplified formula limited to the two fractional
orders α1 , α2 , which binomial expansion we can show as
z1i =
∞
X
(−1)j (
j=0
a2 j (α1 −α2 )j
) ·I
wi .
a1
(24)
For some comparison we select the most popular fractional differential equation in literature [9]
a ·0 Dt2 y(t) + b ·0 Dt1.5 y(t) + c · y(t) = f (t)
(25)
called the Bagley-Torvik equation. Fig. 1a presents the qualitative comparison
of the numerical solution generated by our method with the
analytical solution.
8 for 0 ≤ t < 1
We assume: a = 1, b = 0.5, c = 0.5, the function f (t) =
and
0 for 1 < t < ∞
8
4
numerical results
Podlubny's work
Mittag-Leffler function
6
3
2
K
K
K
K
4
1
2
y(t)
y(t)
0
0
-1
-2
-2
-3
-4
-4
0
2
4
6
8
10
t
12
14
16
18
20
0
2
4
6
8
10
t
12
14
16
18
20
Fig. 1. Solution of the Bagley-Torvik equation:
a) comparison of our results with the analytical solution [9] and with Podlubny’s
work [9],
b) numerical solution of the non-linear form of Bagley-Torvik equation.
initial conditions y(0) = 0, y ′ (0) = 0 respectively. The analytical solution one
may find in [9]. We also found a simple way of solution created in Podlubny’s
work [9]. Fig. 1a certifies that our numerical results behave good in comparison to
the analytical solution. The great advantage would be the apply of our approach
to the non-linear form of fractional differential equation. Let us consider the
non-linear form of the fractional differential equation
2
0 Dt y(t)
+ 0.5 · 0 Dt1.5 y(t) + 0.5 · y 3 (t) = f (t)
(26)
8 for 0 ≤ t < 1
with initial conditions y(0) = 0, y ′ (0) = 0. This
0 for 1 < t < ∞
is the Bagley-Torvik equation where non-linear term y 3 (t) is introduced. Fig. 1b
where f (t) =
shows a behaviour of the numerical solution of eqn. (26). We can see that the
step h has strong influence to the solution y(t). Non-linear fractional differential
equations need some computational tests that to choose right value of the step.
4
Conclusions
In this paper we propose a new method for the numerical solution of arbitrary
differential equations of fractional order. Following to Blank’s results [2] the main
advantage of the method is a decomposition of fractional differential equation
into a system composed with one ordinary differential equation of integer order
and the left inverse equations of the Abel-integral operator. We distinguish two
classes of such system: one-term fractional derivative and multi-term fractional
derivatives. The comparison certifies that our method gives quite good results.
Summarizing these results, we can say that the decomposition method in its
general form gives a reasonable calculations, is the effective method and easy to
use and is applied for the fractional differential equations in general form.
References
1. Babenko Yu. I.: Heat and Mass Transfer, Khimiya, Leningrad (1986) (in Russian)
2. Blank L.: Numerical Treatment of Differential Equations of Fractional Order, Numerical Analysis Report 287, Manchester Centre for Computational Mathematics,
Menchester (1996), pp. 1-16
3. Carpinteri A. and Mainardi F. (eds.): Fractals and Fractional Calculus in Continuum Mechanics, Springer Verlag, Vienna - New-York, (1997)
4. Diethelm K.: An algorithm for the numerical solution of differential equations of
fractional order, Elec. Transact. Numer. Anal. 5 (1997), pp. 1-6
5. Goto M. and Oldham K.B.: Semiintegral electroanalysis: shapes of the neopolarogeographic plateau, Anal. Chem., 61, (1975), pp. 361-365
6. Konguetsof A. and Simons T.E.: On the construction of exponentially fitted methods for the numerical solution of the Schrodinger equation, Journal of Computational Methods in Sciences and Engineering, 1, (2001), pp. 143-160
7. Oldham K. B. and Spanier J.: The fractional Calculus, Academic Press, New York
and London (1974)
8. Palczewski A.: Ordinary differential equations, WNT, Warsaw (1999)(in Polish)
9. Podlubny I.: Fractional Differential Equations, Academic Press, San Diego (1999)
10. Riewe F.: Nonconservative Lagrangian and Hamiltonian mechanics, Phys. Rev. E
53(2) (1996) pp. 1890-1899
11. Riewe F.: Mechanics with fractional derivatives, Phys. Rev. E 55(3) (1997),
pp. 3581-3592
12. Samko S. G., Kilbas A. A. i Marichev O. I.: Integrals and derivatives of fractional
order and same of their applications, Gordon and Breach, London (1993)
13. Simons T.E.: On finite difference methods for the solution of the Schrodinger equation, Computers & Chemistry 23 (1999), pp. 513-555
14. Simons T.E.: Atomic structure computations in chemical modelling: Applications
and theory (Ed.: A. Hinchliffe, UMIST) The Royal Society of Chemistry, (2000),
pp. 38-142
| 5cs.CE
|
Adaptive Computation of the Klee’s Measure
in High Dimensions
Jérémy Barbay1 , Pablo Pérez-Lantero2 , and Javiel Rojas-Ledesma1?
1
arXiv:1505.02855v2 [cs.DS] 2 Oct 2015
2
Departmento de Ciencias de la Computación, Universidad de Chile, Chile
[email protected], [email protected].
Escuela de Ingenierı́a Civil Informática, Universidad de Valparaı́so, Chile.
[email protected].
Abstract. The Klee’s Measure of n axis-parallel boxes in Rd is the
volume of their union. It can be computed in time within O(nd/2 ) in
the worst case. We describe three techniques to boost its computation:
one based on some type of “degeneracy” of the input, and two ones
on the inherent “easiness” of the structure of the input. The first technique benefits from instances where the Maxima of the input is of small
size h, and yields a solution running in time within O(n log2d−2 h +
hd/2 ) ⊆ O(nd/2 ). The second technique takes advantage of instances
where no d-dimensional axis-aligned hyperplane intersects more than k
boxes in some dimension, and yields a solution running in time within
O(n log n+nk(d−2)/2 ) ⊆ O(nd/2 ). The third technique takes advantage of
instances where the intersection graph of the input has small treewidth ω.
It yields an algorithm running in time within O(n4 ω log ω+n(ω log ω)d/2 )
in general, and in time within O(n log n + nω d/2 ) if an optimal tree decomposition of the intersection graph is given. We show how to combine
these techniques in an algorithm which takes advantage of all three configurations.
1
Introduction
The Klee’s Measure of a set of n axis-parallel boxes in Rd is defined as
the volume of the union of the boxes in the set [8]. Its computation was first
posed by Victor Klee in 1977 [20], who originally considered the measure for
intervals in the real line. Bentley [8] generalized the problem to d dimensions and
described an algorithm running in time within O(nd−1 log n). Several years later,
Overarms and Yap [23] described a solution running in time within O(nd/2 log n),
which remained essentially unbeaten for more than 20 years until 2013, when
Chan [12] presented an algorithm running in time within O(nd/2 ). We consider
that, additionally, a d-dimensional domain box Γ is given, making the objective
to compute the Klee’s Measure within Γ .
Some special cases of this problem have been studied, such as the Hypervolume problem, where boxes are orthants of the form {(x1 , . . . , xd ) ∈ Rd |
?
Corresponding Author
2
(x1 ≤ α1 ) ∧ . . . ∧ (xd ≤ αd )}, each αi being a real number, which can be solved
in time within O(nd/3 polylog n); and Cube-KMP [2,10], when the boxes are hypercubes, which can be solved in running time within O(n(d+1)/3 polylog n) [12].
Yildiz and Suri[26] considered k-Grounded-KMP, the case when the projection of the input boxes to the first k dimensions is an instance of Hypervolume. They described an algorithm to solve 2-Grounded in time within
O(n(d−1)/2 log2 n), for any dimension d ≥ 3.
The best lower bound known for the computational complexity of the Klee’s
Measure problem in the worst case over instances of size n is within Ω(n log n),
so far tight only for dimensions one and two [16] as the best known upper bound
is O(nd/2 ) in dimension d. Chan [11] conjectured that no ‘purely combinatorial’
algorithm computing the Klee’s Measure in dimension d exists running in
time within O(nd/2−ε ) for any ε > 0. He proved that if the d-dimensional Klee’s
Measure problem can be solved in time Td (n), then one can decide whether an
arbitrary n-vertex graph G = (V, E) contains a clique of size d in time within
O(Td (O(n2 ))). The current best combinatorial algorithm for finding k-cliques in
a graph, requires near-O(nk ) time, and hence the conjecture.
In an adaptive analysis, the cost of an algorithm is measured as a function of,
not just the input size, but of other parameters that capture the inherent simplicity or difficulty of an input instance [1]. An algorithm is said to be adaptive if
“easy” instances are solved faster than the “hard” ones. There are adaptive algorithms to solve classical problems such as Sorting a permutation [22], Sorting
a multiset [5], computing the Convex Hull [19] of a set of points in the plane
and in 3-space, and computing the Maxima of a set of d-dimensional vectors [18].
There are also adaptive algorithms for the Maximum Weight Box problem [6],
of particular interest since, for any dimension d ≥ 2, the Maximum Weight
Box problem can be reduced to an instance of the Klee’s Measure problem
in 2d dimensions.
Even though the asymptotic complexity of O(nd/2 ) is the best known so far
for the Klee’s Measure problem [12], there are many cases which can be solved
in time within O(n lg n) (see Figures 1 and 3 for some examples). Some of those
“easy” instances can be mere particular cases, some others can be hints of some
hidden measures of difficulty of the Klee’s Measure problem.
Hypothesis: There are such difficulty measures that gradually separate instances of the same size n into various classes of difficulty; from easy ones solvable in time within O(n log n), to difficult ones which the best known algorithm
solves in time within O(nd/2 ).
Results: We describe three techniques to boost the computation of the Klee’s
Measure on “easy” instances, and analyze each in the adaptive model. For each
technique, we identify a proper difficulty measure, which models the features
which the technique is taking advantage of. The first technique is the simplest,
taking advantage of degenerated instances, while the second and third ones are
more sophisticated.
3
x2
x2
Γ
Γ
x3
x1
(a)
x1
(b)
Fig. 1: Two ‘easy’ instances of the Klee’s Measure problem: red dashed boxes
in (a) can be removed without affecting the Klee’s Measure within the domain
Γ , yielding a much smaller instance to solve; while the instance in (b) belongs
to the family of instances which intersection graph is a tree (a path in this
particular case), that can be solved in time within O(n log n) by a divide-andconquer algorithm.
The first technique (described in Section 2) is related to a classical problem in Computational Geometry: the computation of the Maxima of a set of
vectors. A vector in a set T ⊂ Rd is called maximal if none of the remaining
vectors in T dominates it in every component. The Maxima of T (denoted by
M (T )) is the set of maximal elements in T . In 1985, Kirkpatrick and Seidel [18]
gave an output-size sensitive algorithm for this problem, running in time within
O(n logd−2 h), where h is the size of the Maxima. We extend the concept of
Maxima to the sets of boxes, and describe an algorithm computing the Klee’s
Measure in time within O(n log2d−2 h + hd/2 ) ⊆ O(nd/2 ), where h denotes the
size of the Maxima of the input set.
The second technique (described in Section 3) is based on the profile k of the
input set, which D’Amore et al. [13] defined as the minimum, over all dimensions
i, of the maximum number of boxes intersected by a same axis aligned hyperplane orthogonal to i. The algorithm described by Chan [12] to compute the
Klee’s Measure is in fact sensitive to a difficulty measure based on a slightly
weaker definition of the profile. We improve on this by describing an algorithm to
compute the Klee’s Measure in time within O(n log n+nk (d−2)/2 ) ⊆ O(nd/2 ),
where k is the profile of the input set.
The third technique (described in Section 4) is based on the treewidth of the
intersection graph of the input set. The intersection graph of a set of boxes is a
graph G where the vertices are the boxes, and where two boxes are connected
by an edge if and only if they intersect. The treewidth ω of a graph measures
how “close” the graph is to a tree. This technique yields a solution running in
time within O(n log n + nω d/2 ) if a tree decomposition of the intersection graph
of the input set is given; and a solution running in time within O(n4 ω log ω +
n(ω log ω)d/2 ) if only the boxes are given.
4
In Section 5, we discuss how to compare and combine these three techniques,
and in Section 6 we describe some potential directions for future work.
2
Maxima Filtering
Our first technique considers the Maxima of the input set of boxes to take
advantage of instances where many boxes can be “filtered out” in small time. A
box in a set B is called maximal if none of the remaining boxes in B completely
contains it. The Maxima M (B) of B is the set of maximal elements in B. One
can observe that, by definition, elements not in the Maxima of an input set
of the Klee’s Measure problem can be removed from the input set without
affecting the value of the Klee’s Measure.
Algorithm 1 takes advantage of this fact to compute the Klee’s Measure
in time sensitive to the size of the Maxima of the input set.
Algorithm 1 maxima adaptive measure
Input: A d-dimensional domain box Γ , and a set of n d-dimensional boxes B
Output: The Klee’s Measure of B within Γ
1: Compute M (B), the Maxima of B
2: return SDC(Γ, M (B))
Overmars [24] showed that if the Maxima of a set of n d-dimensional
vectors can be computed in time Td (n), then the Maxima of a set of n boxes
can be computed in time within O(T2d (n)). To prove this, Overmars [24] expressed each box bi = [li,1 , ui,1 ] × . . . × [li,d , ui,d ] as a 2d dimensional vector
b~i = (−li,1 , ui,1 , . . . , −li,1 , ui,d ). Note that if bi , bj are boxes, then bi dominates
bj if and only if b~i dominates b~j . We use this result to show in Lemma 1 that
the Klee’s Measure can be computed in running time sensitive to the size of
Maxima of the input.
Lemma 1. Let B be a set of n boxes in Rd and Γ a d-dimensional box. The
Klee’s Measure of B within Γ can be computed in time within O(n(log h)2d−2 +
hd/2 ), where h is the size of the Maxima M (B) of B.
Proof. Algorithm 1 achieves the bound given in the lemma: as a consequence
of Overmars’ result [24], the Maxima M (B) of B is computed in step one in
time within O(n log2d−2 h) using the output-size sensitive algorithm described by
Kirkpatrick and Seidel [18]; in the second step, the Klee’s Measure of M (B),
of size h, is computed in time within O(hd/2 ) using the algorithm proposed by
Chan [12]. The result follows.
t
u
Note that in the bound from the Lemma 1, the base with exponent d/2 is h,
instead of n as in the bound O(nd/2 ) for the running time of SDC. In degenerated
5
instances, where h is significantly smaller than n, the bound from Lemma 1 is
significantly better than O(nd/2 ).
One way to further improve this result is to remove dominated elements
at each recursive call of Chan’s algorithm [12]: we discuss the difficulties in
analyzing this approach in Section 6. In the next section we describe another
boosting technique, which still reduces to Chan O(nd/2 )’s algorithm, but is less
focused on degenerated instances.
3
Profile-based Partitioning
The i-th profile ki of a set of boxes B is defined as the maximum number of boxes
intersected by any hyperplane orthogonal to the i-th dimension. The profile k
of a set of boxes is defined as k = mini∈[1..d] {ki }. D’Amore et al. [13] showed
how to compute it in linear time (after sorting the boxes in each dimension). We
make the observation that Chan’s algorithm [12] for this problem is adaptive
to a measure slightly different to the profile (and weaker than it); and improve
on this result by describing a technique which yields a solution sensitive to the
profile of B.
3.1
Intrinsic Adaptivity of Chan’s Algorithm
The Simplify, Divide and Conquer algorithm (SDC for short) proposed by Chan [12]
to compute the Klee’s Measure, already behaves adaptively in the sense that
it runs faster on some large families of instances. Let the quasi-profile κ of a set
of boxes be defined as κ = max{ki | i ∈ [1..d]}, where ki denotes the i-th profile.
Observation Let B be a set of boxes having quasi-profile κ within a domain
box Γ . Algorithm SDC computes the Klee’s Measure of B within Γ in time
within O(n log n + nκ(d−2)/2 )
The proof of this observation is quite technical and long. Since the result in
next section subsumes this one, and the analysis there is considerably simpler,
we omit the proof of this observation.
An example of the class of instances with small quasi-profile is illustrated in
Figure 1b. In the next subsection, we describe a slightly modified version of the
algorithm SDC which runs in time sensitive to the profile k of B rather than its
quasi-profile κ, an improvement since k ≤ κ on all instances.
3.2
Profile-based partitioning
Let B be a set of boxes with profile k, and Γ a domain box. Given the profile k of
B, Algorithm 2 splits Γ into m ∈ O(n/k) slabs Γ1 . . . Γm , such that the measure
of B within Γ is equal to the summation of the measures of B within Γ1 , . . . , Γm ,
respectively. The algorithm performs a plane sweep by one of the dimensions with
the smallest profile and cuts the domain by a hyperplane every 2k endpoints.
6
Algorithm 2 split-domain
Input: A domain Γ , a set of n boxes B, and the profile k of B
Output: A partition of Γ into m slabs, intersecting each one O(k) boxes.
1: let i be a dimension where the i-th profile ki of B equals k
2: for j = 1, 2, . . . , m ∈ O(n/k) do
3:
let p ← (2k × j)-th endpoint of B within Γ
4:
split Γ into {ΓL , ΓR } by the hyperplane xi = p
5:
let Γj ← ΓL , and Γ ← ΓR
6: return {Γ1 , . . . , Γm }
By computing the Klee’s Measure of B within each Γi , and summing up all
those values, one can compute the Klee’s Measure of B within Γ .
Each of the slabs into which Γ is divided in Algorithm 2 can intersect at most
O(k) boxes of B: by definition of the profile, at most O(k) boxes can intersect
the boundaries of the slab, and since each slab contains at most 2k endpoints,
no more than O(k) boxes can completely lie in its interior. This can be used to
bound the running time of the computation of the Klee’s Measure of B.
Lemma 2. Let B be a set of n boxes in Rd , Γ be a d-dimensional domain box,
and k denote the profile of B. The Klee’s Measure of B within Γ can be
d−2
computed in time within O n log n + nk 2 .
Proof. Using Algorithm 2, one can split the domain into O(n/k) slabs in linear
time after sorting the input. The measure within each slab can be computed in
time within O(k d/2 ) using the algorithm SDC. The result follows.
t
u
Note that again in the bound from Lemma 2, the value with d in the exponent
is the profile k ∗ , instead of the size of the set n. Over instances with small profile
k ∗ (like the ones in the class illustrated in Figure 1b), the bound from Lemma 2
is significantly better than the upper bound O(nd/2 ) for the running time of
SDC. In the next section, we describe a technique, based on the treewidth of the
intersection graph of the input set, a measure that captures how “close” a graph
is to a tree. This technique takes advantage of inputs where the intersection
graph is of small treewidth.
4
Intersection Graph’s Treewidth
In instances such as the one described in Figure 1b, where the intersection graph
is a tree, a minor variant of SDC performs in time within O(n log n) independently
of the dimension d. The concept of treewidth was discovered independently several times under different names (for a nice introduction, see Sections 10.4 and
10.5 of Kleinberg and Tardos’ book [21]). Many graph problems that are NPhard for general graphs can be solved in polynomial time for graphs with small
treewidth. For example, Arnborg and Proskurowski [4] showed that for most
7
NP-hard problems that have linear time algorithms for trees, there are algorithms solving them in time linear in the size of the graph but exponential or
super-exponential in the treewidth. They illustrated the idea with classical optimization problems involving independent sets, dominating sets, graph coloring,
Hamiltonian circuits and network reliability.
In this section we describe how to generalize this behavior to instances with
a more general intersection graph, taking advantage of the treewidth [21] of this
intersection graph. We recall the definition and some basic results on treewidth
in Section 4.1, to apply them to the computation of the Klee’s Measurein
Section 4.2.
4.1
Preliminaries
The most widely used treewidth definition, based on tree decompositions, was
introduced by Robertson and Seymour [25].
Definition 1. A tree decomposition of a graph G = (V, E) is a pair ({Xi |i ∈
I}, T = (I, F )), with {Xi |i ∈ I} a family of subsets of V and T a tree, such that:
S
– (Node coverage) i∈I Xi = V ,
– (Edge coverage) for all hu, vi ∈ E, there is an i ∈ I with u, v ∈ Xi , and
– (Coherence) if v ∈ Xi ∩ Xj , then for all k in the simple path from i to j in
T we have v ∈ Xk .
One refers to the elements of I as nodes, and to the elements of V as vertices.
The width of a tree decomposition ({Xi |i ∈ I}, T = (I, F )) is maxi∈I |Xi | − 1.
A tree decomposition of G is called optimal if its width is the minimum width
among all tree decompositions of G. The treewidth ω of a graph G is the width
of an optimal decomposition of itself. (see Figure 2 for an illustration of tree
decompositions and treewidth).
8
5 7
1
4
3
1
6
5,6,
7,8
7
6
2
2
5
3
8
4
(a)
(b)
1,2,
5,6
3,4,
5,6,
(c)
Fig. 2: Tree decomposition: (a) a set of boxes; (b) the intersection graph of the
set; and (c) an optimal tree decomposition of the graph, with treewidth ω = 2
If Ti is a sub-graph of T , we use Vi to denote the vertices associated with the
nodes in Ti , and GVi to denote the sub-graph of G induced by Vi .
8
Property 1. [21] Let k be a node of T and suppose that T − k has components T1 , T2 , . . . , Td . Then, the sub-graphs GV1 \Xk , GV2 \Xk , . . . , GVd \Xk have no
vertices in common, and there are no edges between them.
A tree decomposition ({Xi | i ∈ I}, T = (I, F )) is nonredundant if there is
no edge hi, ji in T such that Xi ⊆ Xj . There is a simple procedure to make any
tree decomposition be nonredundant without affecting the width: if there is an
edge hi, ji in T such that Xi ⊆ Xj , one can contract the edge by ‘folding’ the
node i into j; by repeating this process as often as necessary, one ends up having
a nonredundant tree decomposition.
Property 2. [21] Any nonredundant tree decomposition of an n-vertex graph
has at most n nodes.
It is NP-hard to determine the treewidth of a given graph. Furthermore, there
is no known algorithm to compute a constant-factor approximation of an optimal
tree decomposition in polynomial time. The best polynomial
time approxima√
tion algorithms for tree decompositions are a O(ω log ω)-factor approximation
algorithm running in time within nO(1) described by Feige et al. [15]; and a
O(ω log ω)-factor approximation algorithm running in time within (n4 ω log ω)
described by Amir [3]. If the treewidth is known to be constant, then a (3ω + 4)approximation can be computed in time within O(2O(ω) +n log n), and a (5ω+4)approximation can be computed in time within O(2O(ω) + n), using two algorithms described by Bodlaender et al. [9] respectively.
4.2
An algorithm sensitive to the intersection graph’s treewidth
Here we describe an algorithm which benefits from instances that have an intersection graph with small treewidth. We will say that a set of boxes has treewidth
ω if its intersection graph has treewidth ω.
Intuitively, consider a set of n boxes with a tree T as intersection graph. Its
Klee’s Measure can be computed in a divide-and-conquer fashion: reduce the
problem to two sub-problems by dividing the intersection graph, via a vertex
removal, into two sub-trees T1 , T2 of roughly equal sizes; solve each problem
independently; and then combine their solutions. Since the intersection graph is
a tree, one can always find a vertex v that divides the tree into two forests of size
at most bn/2c; by adding v back to both forests we obtain T1 , T2 of size at most
d(n + 1)/2e. The Klee’s Measure of the original instance is the sum of the
Klee’s Measure of each sub-instance minus the measure of their intersection,
which is the volume of the box (vertex) used to split. This procedure yields an
algorithm running time within O(n log n).
This procedure can be extended to the computation of the Klee’s Measure
of an instance of tree width ω, given a tree decomposition T of its intersection
graph. The following lemma shows how the solutions of two sub-problems can
be combined into the general solution. If t is a node of T , we denote by Bt the
subset of the boxes of B corresponding to the vertices within Xt .
9
Lemma 3. Let T be a tree decomposition of the intersection graph of a set of
boxes B, and t be a node of T such that, when removed, T is split into S
two nonempty sub-forests
F
and
F
.
Let
T
=
F
∪
{t},
T
=
F
∪
{t},
B
=
1
2
L
1
R
2
L
l∈TL Bl ,
S
and BR = r∈TR Br . Then, TL and TR are tree decompositions of BL and BR
respectively; and the Klee’s Measure of B equals the Klee’s Measure of BL
plus the Klee’s Measure of BR minus the Klee’s Measure of Bt .
Proof. By Property 1 we know that F1 and F2 share no vertices, and there is
no edge between them. Hence, no box corresponding to a vertex in a node from
F1 can intersect a box corresponding to a vertex in a node from F2 . Therefore,
the intersection between BL and BR is Bt . This, and the fact that the volume of
a box (i.e., the Klee’s Measure
of a box) is a Lebesgue measure, proves that
S
the Klee’s Measure of BL BR equals the Klee’s Measure of BL plus the
Klee’s Measure of BR minus the Klee’s Measure of Bt . The result follows.
t
u
Using this lemma, we can apply the procedure described above for trees to
general tree decompositions, as in Algorithm 3. Lemma 4 provides an upper
bound for the running time of this new solution.
Algorithm 3 tw measure
Input: A domain box Γ , and set of n boxes B in Rd , and an ρ-node tree decomposition T of the intersection graph of B
Output: The Klee’s Measure of B within Γ
1: if ρ = 1 then
2:
let t ← the only node in T
3:
if the measure of t has not being computed before then
4:
measures[t] ← SDC(Γ, Bt )
5:
return measures[t]
6: else
7:
Find a node t ∈ T that when removed splits T into two sub-forests
{F1 , F2 } of sizes at most dρ/2e
8:
let T1 ← FS
∪ {t}
1 ∪ {t}, T2 ← F2 S
9:
Let BL ← t∈TL Bt , BR ← t∈TR Bt
10:
return tw measure(Γ, BL , TL ) + tw measure(Γ, BR , TR ) − measures[t]
Lemma 4. Let B be a set of n d-dimensional boxes. Let T be a tree decomposition of the intersection graph of B with ρ nodes of sizes n1 , . . . , nρ , respectively.
The Klee’s Measure of B within a given d-dimensional domain box Γ can be
Pρ
d/2
computed in time within O(ρ log ρ + i=1 ni ).
Proof. We show that Algorithm 3 runs in time within the given bound. The
recursion tree corresponding to the algorithm has ρ leaves, and since at each
10
step the size of the problem is approximately reduced by one half, the height is
within O(log ρ). The total running time at each internal level of the tree is within
O(ρ), and hence the ρ log ρ term. Moreover, the Klee’s Measure of the boxes
within each leaf is computed only once, making the total time of computing the
Pρ
d/2
measure of the leaves to be within O( i=1 ni ). The result follows.
t
u
When computing the Klee’s Measure of a set B of boxes, for whose intersection graph is given an optimal tree decomposition, the following result
follows.
Corollary 1. Let B be a set of n d-dimensional boxes, and T be an optimal
non-redundant tree decomposition of the intersection graph of B. The Klee’s
Measure of B within a given d-dimensional domain box Γ can be computed in
time within O(n log n+nω d/2 ), where ω is the treewidth of the intersection graph
of B.
Proof. Let ρ denote the number of nodes in T . By Property 2, we know that,
since T is non-redundant, ρ ≤ n. Besides, each node in the tree decomposition
has at most ω + 1 vertices. The result follows by replacing each ni by ω, for
i = [1..ρ], in Lemma 4.
t
u
Note that the bound O(n log n + nω d/2 ) in Lemma 4 is better than or equal
to O(nd/2 ) as long as ω ≤ n1−2/d . For this bound to be achieved, an optimal
tree decomposition of the intersection graph is required. This decomposition,
in general, needs to be computed, and it is not known whether this can be
performed in polynomial time [21].
Optimal tree decomposition of certain classes of graphs can be found efficiently. The Klee’s Measure of sets with intersection graph in such classes
can be computed in time depending on the treewidth. We describe two examples
of such results in Corollaries 2 and 3.
Corollary 2. Let B = {b1 , b2 , . . . , bn } be a set of n boxes in Rd , Γ be a ddimensional box, and G be the intersection graph of B. The measure of the union
Pρ
d/2
of B within Γ can be computed in time within O(n logd−1 n + i=1 ni ), where
ρ is the number of connected components of G, and n1 , . . . , nρ are the sizes of
the ρ connected components.
Proof. An algorithm described by Edelsbrunner et al. [14] computes the connected components in time within O(n logd−1 n). From the connected components one can easily obtain a tree decomposition of the graph as follows: create
a node for each connected component, and add edges between the nodes until
obtaining any arbitrary tree. By Lemma 4 the bound follows.
t
u
When the profile of the input instance is k, the same bound from Lemma 2
can be achieved by using Algorithm 3, as seen in Corollary 3.
Corollary 3. Let B be a set of n boxes in Rd , Γ be a d-dimensional box, and
k be the profile of B within Γ . The Klee’s Measure of B within Γ can be
d−2
computed in time within O(n log n + (n/k)k 2 ).
11
Proof. We can transform B into a set B 0 with the same Klee’s Measure, but
with treewidth within O(k), as follows: Split the domain into O(n/k) slabs using
Algorithm 2. Then, for each slab, add to B 0 the boxes in B that intersect the
slab, restricted to it (i.e., if a box intersects several slabs, the box will be split
into multiple boxes that intersect only one slab, and such that the union of them
is the original one)
A tree decomposition of the intersection graph of B 0 with O(n/k) nodes of
size within O(k) can be obtained as follows: create a node for each slab, and
add edges between the nodes until an arbitrary tree is obtained. Since no box
can intersect a box out of its slab, the tree decomposition is valid. The bound
follows from applying Lemma 4.
t
u
An approximation of the optimal tree decomposition of the input set could
be used, obtaining the weaker bounds described in Corollary 4.
Corollary 4. Let B be a set of n d-dimensional boxes, and Γ a d-dimensional
domain box. The Klee’s Measure of B within Γ can be computed in time
within O(n4 ω log ω + N (ω log ω)d/2 ).
Proof. The result follows by replacing the ω term in the bound of Lemma 4 by
the O(ω log ω)-factor approximation obtained by Amir’s algorithm [3] (see the
end of Section 4.1 for details).
t
u
Note that the O(n4 ω log ω +n(ω log ω)d/2 ) bound for the general case is lower
than O(nd/2 ) as long as d > 8 (because of the first term) and ω log ω ≤ n(d−2)/d .
In the following section, we compare the techniques we have described so far
and describe how to combine them.
5
Combining the Techniques
A low profile implies that the intersection graph has low treewidth (Corollary 3),
but a low treewidth does not imply a low profile: an instance of n boxes in the
class illustrated in Figure 3b has a profile within O(n), and its treewidth is one.
On the other hand, the running time of the algorithm taking advantage of the
profile is never worth than O(nd/2 ), which is not true for the running time of the
one sensitive to the treewith, even if an optimal tree decomposition is provided
to it.
The treewidth and profile measures are independent from the size of the
Maxima. For example, an instance of n boxes in the class illustrated in Figure 3a has a Maxima of size 1, and its treewidth is n − 1. With respect to the
treewidth, this is a ‘hard’ instance, but with respect to the Maxima size is easy.
On the contrary, an instance of n boxes in the class illustrated in Figure 3b has
a Maxima size of n, whiles its treewidth is one.
Since these two measures are independent, we can combine them to obtain
an algorithm sensitive to both of them, at the same time, by computing the
Maxima of the set; and finding the Klee’s Measure of the Maxima of the
remaining graph as described in Corollary 4. This way of proceeding yields an
algorithm with running time improving over the results from Lemmas 1 and 4.
12
x2
x2
Γ
Γ
x1
x1
(a)
(b)
Fig. 3: Two classes of instances of the Klee’s Measure problem that are “easy”
or “hard” depending on the measure considered: instance (a) is easy if the size of
the Maxima is considered, but difficult if the profile or treewidth are considered;
instance (b) is easy for treewidth, but hard for both Maxima size and profile.
Theorem 1. Let B be a set of n boxes in Rd , Γ a d-dimensional box, and G
be the intersection graph of B. The Klee’s Measure of B within
Γ can be
2d−2
4
d/2
computed in time within O n log
h + h ω log ω + h(ω log ω)
, where h is
the size of the Maxima M (B) of B, and w its treewidth.
No lower bound is known for this problem with respect to these measures. In
fact, we believe the results obtained here can be further improved, by considering
finer versions of the measures in order to improve the analysis. We describe
preliminary results in this direction in the next section.
6
Discussion
Each of the three boosting techniques that we analyzed can be improved, and
we describe preliminary results for each technique in those directions below, as
well as other lines of research.
The Maxima based technique (described in Section 2) yields an algorithm
running in time within O(n(log h)2d−2 +hd/2 ), where h is the size of the Maxima
of the input set. This bound can be improved. For example, if instead of filtering
items not in the Maxima only once, this is done as part of the simplification
step of the algorithm Simplify, Divide and Conquer (SDC) [12], the expression for
h
) + O(n log2d−2 h). This running time
its running time becomes T (n) = 2T ( 22/d
is still within O(nd/2 ) in the worst case, and also within O(n log2d−2 h + hd/2 ).
It is never worse (asymptotically) and it is better in many cases, but how to
formally analyze this improvement is still an open question.
The profile based
technique (described in
Section 3) yields a solution running
d−2
2
in time within O (n − k) log n−k
. The algorithm SDC [12] is already
k + nk
adaptive to the profile of the input set. It has a limitation though: it necessarily
13
cycles over the dimensions in order to ensure running in time within O(nd/2 ).
If there are few dimensions where the profile of the set is small, this technique
performs considerably better than SDC. The technique could be further improved
if, instead of considering an upper bound for the profile in each sub-problem, we
use the exact value of the profile of the subproblem. However, it is not clear how
to analyze this improvement.
The treewidth based technique (described in Section 4) yields an algorithm
running in time within O(n log n+nω d/2 ), if an optimal tree decomposition of the
intersection graph is given; and in time within O(n4 ω log ω+n(ω log ω)d/2 ) if not.
Note that these running time bounds are not always better than or equal to the
O(nd/2 ) achivied by algorithm SDC. The dependence on a tree decomposition
is the main weakness of this approach. It is not known whether the Klee’s
Measure can be computed by using a treewidth-sensitive algorithm that does
not depend explicitly on a tree decomposition of the intersection graph.
Finally, note that the techniques that we described focus on the structure of
the instance, as opposed to the order in which the instance is given. As such,
the algorithms that we described cannot beat the O(n log n) bound, even though
there are instances that can still be solved in time within O(n). If one considers
the order in which the input is given, for special pre-sorted inputs one can achieve
the O(n) bound.
Acknowledgement
All authors were partially supported by Millennium Nucleus Information and
Coordination in Networks ICM/FIC RC130003. We thank Timothy Chan for
his helpful comments, and one anonymous referee from ESA 2015 for pointing
out the relation between our techniques of analysis and the treewidth of the
intersection graph.
References
1. Afshani, P., Barbay, J., Chan, T.M.: Instance-optimal geometric algorithms. In:
Proceedings of the Annual IEEE Symposium on Foundations of Computer Science
(FOCS). pp. 129–138 (2009)
2. Agarwal, P.K.: An improved algorithm for computing the volume of the union
of cubes. In: Proceedings of the Annual Symposium on Computational Geometry
(SoCG). pp. 230–239. ACM, New York, NY, USA (2010)
3. Amir, E.: Approximation algorithms for treewidth. Algorithmica 56(4), 448–479
(2010)
4. Arnborg, S., Proskurowski, A.: Linear time algorithms for NP-hard problems restricted to partial k-trees. Journal of Discrete Applied Mathematics (JDAM) 23(1),
11 – 24 (1989)
5. Barbay, J.: From time to space: Fast algorithms that yield small and fast data
structures. In: Space-Efficient Data Structures, Streams, and Algorithms (IanFest).
pp. 97–111 (2013)
14
6. Barbay, J., Chan, T.M., Navarro, G., Pérez-Lantero, P.: Maximum-weight planar
boxes in O(n2 ) time (and better). Information Processing Letters (IPL) 114(8),
437–445 (2014)
7. Barbay, J., Pérez-Lantero, P., Rojas-Ledesma, J.: Adaptive computation of the
klee’s measure in high dimensions. CoRR abs/1505.02855 (2015), http://arxiv.
org/abs/1505.02855, (Last accessed on 2015-08-28.)
8. Bentley, J.L.: Algorithms for Klee’s rectangle problems. Unpublished notes (1977)
9. Bodlaender, H.L., Drange, P.G., Dregi, M.S., Fomin, F.V., Lokshtanov, D.,
Pilipczuk, M.: A o(cˆk n) 5-approximation algorithm for treewidth. CoRR
abs/1304.6321 (2013), http://arxiv.org/abs/1304.6321, (Last accessed on 201508-22.)
10. Bringmann, K.: An improved algorithm for Klee’s measure problem on fat boxes.
Computational Geometry: Theory and Applications (CGTA) 45(5-6), 225–233
(2012)
11. Chan, T.M.: A (slightly) faster algorithm for Klee’s measure problem. In: Proceedings of the Annual Symposium on Computational Geometry (SoCG). pp. 94–100
(2008)
12. Chan, T.M.: Klee’s measure problem made easy. In: Proceedings of the Annual
IEEE Symposium on Foundations of Computer Science (FOCS). pp. 410–419
(2013)
13. d’Amore, F., Nguyen, V.H., Roos, T., Widmayer, P.: On optimal cuts of hyperrectangles. Computing 55(3), 191–206 (1995)
14. Edelsbrunner, H., Van Leeuwen, J., Ottmann, T., Wood, D.: Computing the connected components of simple rectilinear geometrical objects in d-space. RAIRO Theoretical Informatics and Applications - Informatique Thorique et Applications
18(2), 171–183 (1984)
15. Feige, U., Hajiaghayi, M., Lee, J.R.: Improved approximation algorithms for
minimum-weight vertex separators. In: Proceedings of the annual ACM Symposium on Theory Of Computing (STOC). pp. 563–572. STOC ’05, ACM, New York,
NY, USA (2005)
16. Fredman, M.L., Weide, B.: On the complexity of computing the measure of
∪n
1 [ai , bi ]. Communications of the ACM (CACM) 21(7), 540–544 (Jul 1978)
17. Gavril, F.: The intersection graphs of subtrees in trees are exactly the chordal
graphs. Journal of Combinatorial Theory (JCT), Series B 16(1), 47 – 56 (1974)
18. Kirkpatrick, D.G., Seidel, R.: Output-size sensitive algorithms for finding maximal
vectors. In: Proceedings of the Annual Symposium on Computational Geometry
(SoCG). pp. 89–96 (1985)
19. Kirkpatrick, D.G., Seidel, R.: The ultimate planar convex hull algorithm. SIAM
Journal on Computing (JC) 15(1), 287–299 (Feb 1986)
20. Klee, V.: Can the measure of ∪n
1 [ai , bi ] be computed in less than o(n log n) steps?
The American Mathematical Monthly (AMM) 84(4), 284–285 (1977)
21. Kleinberg, J., Tardos, E.: Algorithm Design. Addison-Wesley Longman Publishing
Co., Inc., Boston, MA, USA (2005)
22. Moffat, A., Petersson, O.: An overview of adaptive sorting. Australian Computer
Journal (AJC) 24(2), 70–77 (1992)
23. Overmars, M.H., Yap, C.: New upper bounds in Klee’s measure problem. SIAM
Journal on Computing (JC) 20(6), 1034–1045 (1991)
24. Overmars, M.H.: On the equivalence of rectangle containment, rectangle enclosure and ECDF-searching. Technical report, University of Utrecht, Department of
Computer Science (01 1981)
15
25. Robertson, N., Seymour, P.: Graph minors. ii. algorithmic aspects of tree-width
7(3), 309 – 322 (1986)
26. Yildiz, H., Suri, S.: On Klee’s measure problem for grounded boxes. In: Proceedings of the Annual Symposium on Computational Geometry (SoCG). pp. 111–120.
ACM, New York, NY, USA (2012)
| 8cs.DS
|
Uplink Coverage Performance of an Underlay
Drone Cell for Temporary Events
(Invited Paper)
Xiaohui Zhou∗ , Jing Guo∗ , Salman Durrani∗ , and Halim Yanikomeroglu†
∗
arXiv:1801.05948v2 [cs.IT] 6 Mar 2018
Research School of Engineering, The Australian National University, Canberra, ACT 2601, Australia.
Emails: {xiaohui.zhou, jing.guo, salman.durrani}@anu.edu.au.
†
Department of Systems and Computer Engineering, Carleton University, Ottawa, ON K1S 5B6, Canada.
Email: [email protected].
Abstract—Using a drone as an aerial base station (ABS)
to provide coverage to users on the ground is envisaged as
a promising solution for beyond fifth generation (beyond-5G)
wireless networks. While the literature to date has examined
downlink cellular networks with ABSs, we consider an uplink
cellular network with an ABS. Specifically, we analyze the use of
an underlay ABS to provide coverage for a temporary event, such
as a sporting event or a concert in a stadium. Using stochastic
geometry, we derive the analytical expressions for the uplink
coverage probability of the terrestrial base station (TBS) and
the ABS. The results are expressed in terms of (i) the Laplace
transforms of the interference power distribution at the TBS
and the ABS and (ii) the distance distribution between the ABS
and an independently and uniformly distributed (i.u.d.) ABSsupported user equipment and between the ABS and an i.u.d.
TBS-supported user equipment. The accuracy of the analytical
results is verified by Monte Carlo simulations. Our results show
that varying the ABS height leads to a trade-off between the
uplink coverage probability of the TBS and the ABS. In addition,
assuming a quality of service of 90% at the TBS, an uplink
coverage probability of the ABS of over 85% can be achieved,
with the ABS deployed at or below its optimal height of typically
between 250 − 500 m for the considered setup.
I. I NTRODUCTION
The use of drones as aerial base stations (ABSs), to form
drone cells and to provide flexible and agile coverage, has
gained enormous interest as a potential beyond fifth generation
(beyond-5G) wireless network solution [1, 2]. In [1], eight
scenarios have been identified for drone cell deployment, with
drones providing service to: 1) rural areas, 2) urban areas,
3) users with high mobility, 4) congested urban areas, 5)
congested backhaul, 6) temporary events, 7) temporary blind
spots and 8) sensor networks. For these scenarios, drone
cells have been investigated in the literature from different
perspectives, such as drone channel modeling [3, 4, ?, 4, 6],
drone deployment and optimization of trajectories [7–9, ?–9,
11–14], and performance analysis of drone enabled cellular
systems [15–19]. In this work, we focus on the last aspect.
Motivation: The literature to date [15–19] has focused on
the downlink, generally for Scenarios 1–3. In this work, we
focus on the uplink for Scenario 6, i.e., using ABS to provide
additional coverage for temporary events, such as concerts
This work was supported in part by the Australian Research Council’s
Discovery Project Funding Scheme (Project number DP170100939) and the
Natural Sciences and Engineering Research Council of Canada (NSERC).
or sports events. Such temporary events are happening more
frequently in cities all over the world with very high number of
users gathering, which often leads to network congestion and
poor user experience. The optimal ABS height with respect
to the downlink performance does not necessarily optimize
the uplink performance. For example, the distribution of the
interference powers are different for the uplink and downlink
cases. Moreover, the uplink is the bottleneck in such events
since an increasing number of users try to share content from
the event on social media applications. Therefore, assessing the
usefulness of ABSs to provide uplink coverage for temporary
events is an important open problem in the literature which
has great practical importance as well.
Related Work: The downlink performance of a single static
drone and a single mobile drone with underlay device-todevice users was studied in [15], while [16] analyzed the
downlink coverage of a finite network formed by multiple
drones. The downlink coverage probability of a network with
multiple directional beamforming drones was investigated in
[17]. The downlink coverage performance of a network of
multiple drones in an urban environment was studied in [18].
In [19], the authors studied the downlink in a cellular network
with multiple ground base stations and a drone user equipment
(UE).
Contributions: In this paper, we analyze the use of an ABS
to provide coverage for a temporary event. To the best of our
knowledge, this scenario has not yet been considered in the
literature to date. Using stochastic geometry, we analyze the
uplink coverage probability of the terrestrial base station (TBS)
and the ABS. The novel contributions of this paper are:
• We derive the analytical expressions for uplink coverage
probability of the TBS and the ABS. The results are in
terms of (i) the Laplace transforms of the interference
power distribution at the TBS and the ABS and (ii) the
distance distribution between the ABS and an independently and uniformly distributed (i.u.d.) ABS-supported
user equipment (AsUE) and between the ABS and an
i.u.d. TBS-supported user equipment (TsUE).
• Our results show that increasing the height of the ABS
can generally improve the uplink coverage probability of
the ABS, while it degrades the uplink coverage probability of the TBS. In addition, there is an optimal ABS
height which maximizes the uplink coverage probability
of the ABS.
• We assess the feasibility of establishing an underlay
ABS to provide uplink coverage for temporary events.
Assuming a quality of service of 90% at the TBS, under
our considered system set up, uplink coverage probability
of the ABS of over 85% can be achieved, with the ABS
at or below its optimal height if the center of the stadium
is sufficiently distanced from the TBS.
II. S YSTEM MODEL
We consider the uplink communication in a two-cell network comprised of a TBS and an ABS, where the network
region S1 is a disk with radius r1 , i.e., |S1 | = πr12 and a TBS
is located at the center. We assume that there is a temporary
event held inside a stadium within the network region. The
stadium’s building area S2 is modeled as a disk with radius
r2 and its center is at a distance d from the TBS. To provide
additional resources for the event, an ABS is deployed to act
as an ABS1 . The ABS is assumed to be placed at a height
of h above the center of the stadium, as shown in Fig. 1.
There are Nc TsUEs served by the TBS, which are spatially
distributed according to a Binomial point process (BPP) inside
the network region excluding the stadium, i.e., S1 \ S2 . At
the same time, there are Nd AsUEs on the ground inside the
stadium S2 . For tractability, we assume that all AsUEs are
served by the ABS only. The location of AsUEs is modelled
as an independent BPP in S2 .
Channel Model: There are two types of communication
links in the considered system model: aerial links and terrestrial links. The link between the TsUE and the ABS and the
link between the AsUE and the ABS are aerial links. The link
between the TsUE and the TBS and the link between the AsUE
and the TBS are terrestrial links. For analytical tractability,
similar to [16, 19], we assume the aerial links experience
Nakagami-m fading. The AsUE to ABS link has a strong line
of sight (LOS) with fading parameter mAA while the TsUE
to ABS link has a weak LOS with fading parameter mTA . As
for the terrestrial links, Rayleigh fading is assumed [11].
A general power-law path-loss model is considered2 , in
which the signal power decays at a rate `−α with the propagation distance ` and α is the path-loss exponent. Due
to different characteristics of aerial link and terrestrial link,
different path-loss exponent αTBS , αTA and αAA are assumed
for terrestrial link, aerial link between the TsUE and the ABS,
and aerial link between the AsUE and the ABS respectively.
Furthermore, both the TsUE to TBS link and the AsUE to
ABS link experience additive white Gaussian noise (AWGN)
with variance σ 2 .
Uplink Network Model: For the spectrum efficiency, we assume that both the TBS and the ABS share the same spectrum
resource, i.e., in an underlay fashion. Orthogonal multiple
1 Current
drone regulations prohibit a drone from flying over stadiums or
sports events. This is expected to change in the future.
2 The validity of using this path-loss model for aerial links will be verified in
Section IV by comparing with simulations using air-to-ground channel model
for aerial links.
TBS
TsUE
ABS
AsUE
h
ZT
r1
ZA
d
r2
TBS
TsUE
ABS
AsUE
S1
dT
d
dA
r2
S2
r1
(a) 3D view.
(b) Projection on the ground.
Fig. 1. Illustration of the system model.
access technique is employed in this work. We assume that
the number of the TsUEs and the AsUEs are sufficiently high.
That is to say, there will always be one TsUE and one AsUE
to be served per each channel at the same time. Hence, there
is no interference among TsUEs (or AsUEs), but interference
exists between TsUEs and AsUEs. In this work, we focus our
analysis on one uplink channel since other channels share the
same interference statistics.
For reliable and successful uplink communication, the TsUE
controls its transmit power such that the average signal received at the TBS is equal to the threshold ρT . Power control
is deployed at the AsUE. We also set a maximum transmit
power constraint at the AsUE, to avoid the transmit power for
AsUE going to very large when the ABS is placed at a high
altitude. In other words, the AsUE compensates for the pathloss to keep the average signal power at the ABS equal to
the threshold ρA if the transmit power required for the pathloss inversion is less than Pmax . Otherwise, the AsUE tries to
establish an uplink connection with the ABS by transmitting
with a power of Pmax . Therefore, the instantaneous transmit
power for the AsUE, Pa , depends on the propagation distance
between the AsUE and the ABS and can be shown as (1) at
the top of the next page, where ZA is the Euclidean distance
between the AsUE and the ABS.
SINR: For the considered setup, the instantaneous signal-tointerference-plus-noise ratio (SINR) at the TBS is given as
SINRT =
Pt HT dT−αTBS
ρ T HT
=
,
−αTBS
TBS
2
+
σ
+ σ2
Pa HA d−α
P
H
d
a A A
A
(2)
TBS
where Pt = ρT dα
is the TsUE transmit power. HT and HA
T
are the fading power gain between the TsUE and the TBS and
between the AsUE and the TBS, respectively, which follow
exponential distribution. dT and dA are the Euclidean distance
between the TsUE and the TBS and between the AsUE and
the TBS, respectively. The transmit power of the AsUE Pa is
given in (1).
The instantaneous SINR at the ABS is given as
SINRA =
−αAA
Pa GA ZA
,
Pt GT ZT−αTA + σ 2
(3)
where GA and GT are the fading power gain between the
AsUE and the ABS and between the TsUE and the ABS,
respectively, which follow Gamma distribution. ZA and ZT
are the Euclidean distance between the AsUE and the ABS
and between the TsUE and the ABS, respectively.
Pmax ,
q
1
2
1
1
Pmax αAA
Pmax αAA
2 < h < ( Pmax ) αAA
αAA
(
||
−
r
&
Z
>
(
)
)
)
h > ( Pρmax
A
2
ρA
ρA
ρA
A
q
.
q
Pa =
2
2
1
1
αAA
Pmax αAA
Pmax αAA
Pmax αAA
2 ||
2 < h < ( Pmax ) αAA
ρ
Z
,
h
6
(
)
−
r
(
)
−
r
&
Z
<
(
)
A A
A
2
2
ρA
ρA
ρA
ρA
R √ 2 2 R
h +r2 2π
1
fZA (z)dθdz,
LIT (s, Pmax |θ, z) 2π
h
0
1
Pmax α
R ( ρA ) AA R 2π
1
LIT (s, ρA z αAA |θ, z) 2π
fZA (z)dθdz
h √
0
LIT (s) =
R h2 +r22 R 2π
1
+
LIT (s, Pmax |θ, z) 2π fZA (z)dθdz,
1
0
) αAA
√( Pρmax
A
R h2 +r22 R 2π L (s, ρ z αAA |θ, z) 1 f (z)dθdz,
IT
A
h
2π ZA
0
III. C OVERAGE P ROBABILITY
1
h > ( Pρmax
) αAA
A
q
2
1
) αAA − r22 < h < ( Pρmax
) αAA
( Pρmax
A
A
q
2
h 6 ( Pρmax
) αAA − r22
A
.
(5)
B. ABS Coverage Probability
In this section, we analyze the network performance by
adopting coverage probability as the performance metric. The
coverage probability is formally defined as
Pbcov , Pr(SINRb > γb ),
(1)
(4)
where superscript b is T for TBS and A for ABS, and γb is the
SINR threshold. SINRT and SINRA can be found in (2) and
(3), respectively. The results for the coverage probability of the
TBS and the ABS are presented in the next two subsections.
A. TBS Coverage Probability
First we present two lemmas, which are used in deriving
the coverage probability of the TBS in Theorem 1.
Lemma 1: The Laplace transform of the interference power
distribution at the TBS is given as (5) at the top of the this
page, where
1
.
(6)
LIT (s, p|θ, z) =
sp
1+
αTBS
√
(z2−h2+d2−2 z2−h2 d cos(θ)) 2
Proof: See Appendix 1.
From Lemma 1, we can see that the transmit power of the
AsUE Pa and the distance between the AsUE and the TBS
dA are related to the distance between the AsUE and the ABS
ZA . This important distance distribution is presented in the
following lemma.
Lemma 2: The probability density function (pdf) of the
distance ZA between the ABS at height h above the center of
S2 and an i.u.d. AsUE inside S2 is
q
2z
fZA (z) = 2 , h 6 z 6 r22 + h2 .
(7)
r2
Proof: See Appendix 2.
Theorem 1: Based on the system model in Section II, the
coverage probability of the TBS is
γT 2
σ
LIT (s),
PT
=
exp
−
(8)
cov
ρT
TBS
where IT = Pa HA d−α
, s = γρTT , and LIT (s) is given by
A
Lemma 1.
Proof: See Appendix 3.
Substituting (5) and (7) into (8), we can obtain the coverage
probability of the TBS.
We begin by presenting three lemmas, which will then
be used to compute the coverage probability of the ABS in
Theorem 2.
Lemma 3: The Laplace transform of the interference power
distribution at the ABS is
Z √(r1 −d)2 +h2Z 2π
LIA (s|ω, z)fΩ (ω|z)fZT (z)dωdz
LIA(s) = √
r 2 +h2
2
Z √(r1 +d)2 +h2 Z
+ √
(r1 −d)2 +h2
0
ω
b
LIA (s|ω, z)fΩ (ω|z)fZT (z)dωdz, (9)
−b
ω
where
TA
LIA (s|ω, z) = mm
mTA + sρT z −αTA
TA
−mTA
αTBS
p
2
2
2
2
2
2
.
× z −h +d −2 z −h d cos(ω)
(10)
Proof: The proof follows the same lines as Lemma 1 and
is skipped for the sake of brevity.
The pdf of the distance between the TsUE and the ABS
fZT (z), and the conditional pdf of the angle, fΩ (ω|z), between
the ground projection of ZT and dT are given in Lemma 4
and Lemma 5, respectively.
Lemma 4: The pdf of the distance ZT between the ABS
at height h above the center of S2 and an i.u.d. TsUE inside
S1 \ S2 is
( 2z
p
p
,
r22 +h2 6 z 6 (r1 −d)2 +h2
r12−r22
p
p
fZT(z) =
,
2zb
ω
,
(r1 −d)2 +h2 6 z 6 (r1 +d)2 +h2
π(r12−r22 )
(11)
√
2
2
z −h
where ω
b = arcsec( d22d
).
+z 2 −h2 −r12
Proof: See Appendix 4.
Lemma 5: The pdf of the angle, fΩ (ω|z), between the
ground projection of ZT and dT conditioned on ZT is
(
p
p
1
,
r22 +h2 6 z 6 (r1 −d)2 +h2
2π
p
p
fΩ (ω|z) = 1
.
(r1 −d)2 +h2 6 z 6 (r1 +d)2 +h2
2b
ω,
(12)
Proof: This lemma can be proved by using cosine rule
and simple trigonometry.
R √ 2 2
h +r2 A
Pcov (Pmax |z)fZA (z)dz,
h
1
Pmax α
R ( ρA ) AA A
Pcov (ρA z αAA |z)fZA (z)dz
h √
PA
=
2
2
R
cov
h +r2
PA
+
1
cov (Pmax |z)fZA (z)dz,
Pmax α
) AA
( ρ
√ A
R h2+r22 PA (ρ z αAA |z)f (z)dz,
ZA
cov A
h
1
h > ( Pρmax
) αAA
A
q
.
2
1
2 < h < ( Pmax ) αAA
αAA
)
−r
( Pρmax
2
ρ
A
A
q
2
h 6 ( Pρmax
) αAA − r22
A
1
TBS coverage probability
Coverage probability
1
h=200 m
0.8
0.6
h=500 m
0.4
0.2
0
-20
(13)
TBS coverage probability
ABS coverage probability
Simulation (ATG)
-15
-10
-5
SINR threshold
0
T
,
A
5
(dB)
Pmax=10 dBm
0.7
0.6
0.5
0.4
d=200 m
d=250 m
d=300 m
Pmax=20 dBm
h*=631 m
h*=251 m
300
400
500
600
700
800
ABS height h (m)
Fig. 3. Coverage probability of the TBS versus the height of the ABS
with different distance between the center of the stadium and the TBS
and different maximum transmit power for the AsUE.
Theorem 2: Based on the system model in Section II, the
coverage probability of the ABS is given as (13) at the top of
this page, where
mX
AA −1
(−s)n
exp(−sσ 2 )
n!
n=0
n
X
n
dk
(−σ 2 )n−k k LIA (s),
×
k
ds
0.8
0.3
150
10
Fig. 2. Coverage probability of the TBS and the ABS versus the SINR
threshold γT and γA with different height of the ABS and simulation
with ATG aerial link model.
PA
cov (Pa |z) =
0.9
(14)
k=0
αAA
s = mAA γpA z
and LIA (s) is given by Lemma 3. The pdf
of the distance between the AsUE and the ABS fZA (z) is
provided in Lemma 2.
Proof: See Appendix 5.
Combining Lemma 3, 4, and 5 with Theorem 2, we can
calculate the coverage probability of the ABS.
IV. R ESULTS
In this section, we first validate the analytical results and
then discuss the design insights. The simulation results are
generated by averaging over 106 Monte Carlo simulation runs.
In the simulations, the terrestrial links follow the power-loss
path-loss model described in Section II. However, the aerial
links follow the air-to-ground (ATG) channel model in [15]
with C = 4.88, B = 0.43 and η = −20 dB. Unless stated
otherwise, we set the parameters as follows: r1 = 500 m,
r2 = 100 m, d = 200 m, Pmax = 20 dBm, ρT = −75 dBm,
ρA = −50 dBm, αAA = 2.5, αTA = 3, αTBS = 4, mAA = 5,
mTA = 3, σ 2 = −100 dBm, γA = 0 dB and γT = 0 dB.
Fig. 2 plots the coverage probability of the TBS and the
ABS against the SINR threshold with an ABS height of 200
m and 500 m. The simulation results match very well with the
analytical results. This is because that the ABS is assumed to
be placed directly above the stadium in our model and the
likelihood of the aerial links being LOS is very high. This
validates the use of the tractable power-law path-loss model
for aerial links in our analysis. From this figure, we can see
that if an ABS is placed at a higher altitude, the coverage
probability of the ABS is higher, but the coverage probability
of the TBS is lower. This is discussed in detail next.
Impact of ABS height: Fig. 3 plots the coverage probability
of the TBS against the height of the ABS with different d (i.e.,
distance between the center of the stadium and the TBS) and
different Pmax (i.e., maximum transmit power for the AsUE).
From the figure, we can see that the coverage probability of
the TBS first decreases as the ABS height increase. This is
because the transmit power of the AsUE increases with the
ABS’s height, whereby the interference at the TBS increases.
After a certain ABS height, the coverage probability of the
TBS stays as a constant. This is due to the fact that the AsUE
has reached its maximum transmit power and the interference
generated at the TBS keeps the same on average. Note that
the height where the coverage probability of the TBS starts to
level off is independent of the projection distance between the
center of the stadium and the TBS, but the height increases
with the AsUE maximum transmit power.
Fig. 4 plots the coverage probability of the ABS against the
height of the ABS with different maximum transmit power
of the AsUE. We can see that the simulation results using
the ATG channel model for the aerial links match very well
with the analytical results3 . The figure shows that there is an
3 The analytical performance with the ATG channel model in [15] is the
subject of our ongoing work.
Maximum ABS coverage probability
1
ABS coverage probability
(630,0.9805)
0.9
(394,0.8775)
0.8
(240,0.7363)
0.7
0.6
0.5
0.4
150
Pmax=10dBm
Pmax=15dBm
Pmax=20dBm
Simulation (ATG)
300 400 500 600 700 800 900 1000
1
0.9
h=631 m
h=568 m
h=631 m
99% TBS coverage
90% TBS coverage
80% TBS coverage
h=492 m
h=451 m
h=342 m
h=356 m
0.8
h=174 m
h=242 m
h=131 m
h=234 m
0.7
h=149 m
h=125 m
0.6
150
200
250
300
350
400
ABS height h (m)
Distance between center of stadium and TBS d (m)
Fig. 4. Coverage probability of the ABS versus the ABS height with
different AsUE maximum transmit power and simulation with ATG
aerial link model. The triangle markers denote the optimal ABS height.
Fig. 5. Maximum coverage probability of the ABS versus the distance
between the center of the stadium and the TBS d with different TBS
coverage requirements.
optimal height which can maximize the coverage probability
of the ABS. This optimal height increases with the maximum
transmit power of the AsUE Pmax . From the figure, we find
that the curves with different Pmax are overlapped when h is
small. This comes from the fact that in this certain range, the
transmit power of the AsUE is below Pmax and the AsUE to
ABS link is under full channel inversion. Under full channel
inversion, the average desired signal power received at the
ABS stays the same and the interference power drops as the
height increases. Therefore, the coverage probability increases.
When the ABS height is above the optimal height, the AsUE
transmits with its maximum power Pmax and the received
power of the desired signal reduces as the height increases
further. Hence, the coverage probability starts to drop (e.g.,
the red and green curves in Fig. 4). Furthermore, a larger
AsUE maximum transmit power generally results in a higher
coverage probability of the ABS when the height increases.
Feasibility: From the previous figures, we can see that there
is a trade-off between the coverage probability of the TBS and
the ABS. Fig. 5 plots the maximum coverage probability that
can be achieved at the ABS versus the distance between the
center of the stadium and the TBS d, if 99%, 90% or 80%
coverage probability of the TBS is required. The maximum
achievable ABS coverage probability increases with d and then
becomes almost flat after a certain point.The optimal height
of the ABS cannot be achieved, when d is small. When d
is sufficiently large, the ABS can be placed at its optimal
height, so the optimal coverage probability of the ABS can be
achieved. For instance, from Fig. 5, we can find that for 90%
coverage probability of the TBS, 85% coverage probability of
the ABS can be achieved by placing the ABS at a height of
342 m if the center of the stadium is 300 m apart from the
TBS. Therefore, ABS underlay cellular network is feasible
under practical network setup.
V. C ONCLUSIONS
In this paper, the uplink communication in a two-cell
network with a TBS and an ABS was considered. We derived
the exact expressions for uplink coverage probability of the
TBS and the ABS in terms of the Laplace transforms of
the interference power distribution at the TBS and the ABS
and the distance distribution between the ABS and an i.u.d.
AsUE and between the ABS and an i.u.d. TsUE. Our results
demonstrated the feasibility of establishing an underlay ABS
to provide uplink coverage for temporary events. Future work
can consider the impact of beamforming at the ABS and user
scheduling if there are multiple UEs per channel.
A PPENDIX
1) Proof of Lemma 1: Following the definition of the Laplace
transform, the Laplace transform of the interference power
distribution at the TBS is expressed as
−αTBS
LIT (s) = EIT [exp(−sIT)] = EPa ,h,d [exp(−sPa HA dA
)]
"
#
1
= EPa ,d
,
(15)
TBS
1 + sPa d−α
A
where the third step comes from the fact that HA follows
exponential distribution with unit mean. Conditioned on the
value of h, there are three possible cases for LIT (s). When
q
2
1
( Pρmax
) αAA − r22 < h < ( Pρmax
) αAA , the Laplace transform
A
A
of the interference power distribution at the TBS equals to
1
max ) αAA
( Pρ
Z
LIT (s) =
fZA (z)dz
TBS
1+sρD z αAA d−α
A
"
#
1
Ed
fZA (z)dz
TBS
1 + sPmax d−α
A
Z √h2 +r2
2
max
( Pρ
A
1
) αAA
=
1
Z
max ) αAA
( Pρ
A
Eθ
h
1
1+
sρD z αAA
z 2−h2+d2−2
Z √h2+r2
2
+
max
( Pρ
A
#
1
Ed
h
+
"
A
√
αTBS
2
z 2−h2 d cos Θ
Eθ
1
) αAA
fZ (z)dz
A
1
1+
z 2−h2+d2−2
√
sPmax
− αTBS
2
z 2−h2 d cos Θ
fZ (z)dz
A
(16a)
Z
=
max
( Pρ
A
1
) αAA
h
Z √h2 +r2 Z
2
+
1
max
( Pρ
A
) αAA
Z
2π
LIT (s, ρA z αAA |θ, z)
0
2π
LIT(s,Pmax |θ, z)
0
1
fZ (z)dθdz
2π A
1
fZ (z)dθdz,
2π A
(16b)
where dA is expressed in terms of z, h, d, and θ by cosine
rule in (16a) and (16b) is obtained by taking the expectation
1
over Θ, which
has a conditional pdf as fΘ (θ|z) = 2π
for
p
h 6 z 6 r22 + h2 .
Following similar steps, we can work out the Laplace
transform of the interference power distribution at the TBS
1
for the other two cases, i.e., when h > ( Pρmax
) αAA and
A
q
2
) αAA − r22 .
h 6 ( Pρmax
A
2) Proof of Lemma 2: The relation between the length of the
AsUE to ABS link Z
pA with its projection distance on the
2 + h2 . The distance distribution of
ground rA is ZA = rA
the projection distance on the ground is frA (r) = 2r
[20].
r22
Thus, we can get the pdf of ZA as
√
p
d( z 2 − h2 )
z 2 − h2
frA
dz
√
z
2 z 2 − h2
2z
= √
= 2.
r22
r2
z 2 − h2
fZA (z) =
(17)
3) Proof of Theorem 1: The coverage probability of the TBS
is given by
PT
cov = Pr(SINRT > γT )
γT
2
TBS
= Pr HT >
+
σ
)
(Pa HA d−α
A
ρT
γT
γT 2
2
= E exp − (IT +σ ) = exp − σ LIT (s),
ρT
ρT
(18)
where in the third step we use the fact that the link between
the TsUE and the TBS experiences Rayleigh fading with a pdf
of fHA (h) = exp(−h).
4) Proof of Lemma 4: From the proof of Lemma 2, we know
that in order to find the distribution of the distance ZT between
the TsUE and the ABS, the distribution of its projection
distance rT is needed. Using similar approach in [20, 21], the
distance distribution of rT is derived as
frT (r)=
22r 2 ,
r −r
1
2
22r 2 arcsec
πr1 −πr2
r2 6 r 6 r1 −d
2dr
2
d2 +r 2 −r1
,
r1 −d 6 r 6 r1 +d
. (19)
Using the pdf of the auxiliary random variable rT in (19),
we can obtain Lemma 4.
5) Proof of Theorem 2: The coverage probability of the ABS
is given by
PA
cov = Pr(SINRA > γA )
"
!#
γA
−αTA
2
= EZA Pr GA >
+ σ )|z
−αAA (Pt GT ZT
Pa ZA
"m −1
!
#
n
AA
k
X
X
(−s)n
n
2
2 n−k d
= EZA
exp(−sσ )
(−σ )
LI (s)
n!
k
dsk A
n=0
k=0
Z √ 2 2
h +r2
=
PA
cov (Pa |z)fZA (z)dz,
(20)
h
where the third step comes from the fact that GA follows
Gamma distribution. The transmit power of the AsUE Pa is
given in (1).
R EFERENCES
[1] I. Bor-Yaliniz and H. Yanikomeroglu, “The new frontier in RAN
heterogeneity: Multi-tier drone-cells,” IEEE Commun. Mag., vol. 54,
no. 11, pp. 48–55, Nov. 2016.
[2] S. Chandrasekharan, K. Gomez, A. Al-Hourani, S. Kandeepan,
T. Rasheed, L. Goratti, L. Reynaud, D. Grace, I. Bucaille, T. Wirth, and
S. Allsopp, “Designing and implementing future aerial communication
networks,” IEEE Commun. Mag., vol. 54, no. 5, pp. 26–34, May 2016.
[3] R. Amorim, H. Nguyen, P. Mogensen, I. Z. Kovcs, J. Wigard, and T. B.
Srensen, “Radio channel modeling for UAV communication over cellular
networks,” IEEE Wireless Commun. Lett., vol. 6, no. 4, pp. 514–517,
Aug. 2017.
[4] R. Amorim, P. Mogensen, T. Sorensen, I. Z. Kovacs, and J. Wigard,
“Pathloss measurements and modeling for UAVs connected to cellular
networks,” in Proc. IEEE VTC Spring, Jun. 2017.
[5] A. Al-Hourani and K. Gomez, “Modeling cellular-to-UAV path-loss for
suburban environments,” to appear in IEEE Wireless Commun. Lett.,
2017.
[6] A. Al-Hourani, S. Kandeepan, and A. Jamalipour, “Modeling air-toground path loss for low altitude platforms in urban environments,” in
Proc. IEEE Globecom, Dec. 2014.
[7] A. Fotouhi, M. Ding, and M. Hassan. (2017) “Dronecells: Improving
5G spectral efficiency using drone-mounted flying base stations”.
[Online]. Available: https://arxiv.org/abs/1707.02041
[8] Y. Zeng, R. Zhang, and T. J. Lim, “Throughput maximization for
UAV-enabled mobile relaying systems,” IEEE Trans. Commun., vol. 64,
no. 12, pp. 4983–4996, Dec. 2016.
[9] M. Alzenad, A. El-Keyi, F. Lagum, and H. Yanikomeroglu, “3-D placement of an unmanned aerial vehicle base station (UAV-BS) for energyefficient maximal coverage,” IEEE Wireless Commun. Lett., vol. 6, no. 4,
pp. 434–437, Aug. 2017.
[10] H. He, S. Zhang, Y. Zeng, and R. Zhang, “Joint altitude and beamwidth
optimization for UAV-enabled multiuser communications,” to appear in
IEEE Commun. Lett., 2017.
[11] C. Zhang and W. Zhang, “Spectrum sharing for drone networks,” IEEE
J. Sel. Areas Commun., vol. 35, no. 1, pp. 136–144, Jan. 2017.
[12] F. Lagum, I. Bor-Yaliniz, and H. Yanikomeroglu, “Strategic densification with UAV-BSs in cellular networks,” to appear in IEEE Wireless
Commun. Lett., 2017.
[13] Q. Wu, Y. Zeng, and R. Zhang, “Joint trajectory and communication
design for multi-UAV enabled wireless networks,” to appear in IEEE
Trans. Wireless Commun., 2018.
[14] J. Lyu, Y. Zeng, and R. Zhang, “Spectrum sharing and cyclical multiple
access in UAV-aided cellular offloading,” in Proc. IEEE Globecom, Dec.
2017.
[15] M. Mozaffari, W. Saad, M. Bennis, and M. Debbah, “Unmanned aerial
vehicle with underlaid device-to-device communications: Performance
and tradeoffs,” IEEE Trans. Wireless Commun., vol. 15, no. 6, pp. 3949–
3963, Jun. 2016.
[16] V. V. Chetlur and H. S. Dhillon, “Downlink coverage analysis for a
finite 3-D wireless network of unmanned aerial vehicles,” IEEE Trans.
Commun., vol. 65, no. 10, pp. 4543–4558, Oct. 2017.
[17] M. M. Azari, Y. Murillo, O. Amin, F. Rosas, M.-S. Alouini, and
S. Pollin. (2017) “Coverage maximization for a poisson field of drone
cells”. [Online]. Available: https://arxiv.org/abs/1708.06598
[18] B. Galkin, J. Kibilda, and L. A. DaSilva. (2017) “Coverage analysis
for low-altitude UAV networks in urban environments”. [Online].
Available: https://arxiv.org/abs/1704.06214
[19] M. M. Azari, F. Rosas, A. Chiumento, and S. Pollin, “Coexistence of
terrestrial and aerial users in cellular networks,” in Proc. IEEE Globecom
Workshops, Dec. 2017.
[20] J. Guo, S. Durrani, and X. Zhou, “Outage probability in arbitrarilyshaped finite wireless networks,” IEEE Trans. Commun., vol. 62, no. 2,
pp. 699–712, Feb. 2014.
[21] Z. Khalid and S. Durrani, “Distance distributions in regular polygons,”
IEEE Trans. Veh. Technol., vol. 62, no. 5, pp. 2363–2368, Jun. 2013.
| 7cs.IT
|
Subsets and Splits
No saved queries yet
Save your SQL queries to embed, download, and access them later. Queries will appear here once saved.